1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2015, Joyent, Inc. All rights reserved.
24 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
25 */
26
27 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
28 /* All Rights Reserved */
29
30 /*
31 * University Copyright- Copyright (c) 1982, 1986, 1988
32 * The Regents of the University of California
33 * All Rights Reserved
34 *
35 * University Acknowledgment- Portions of this document are derived from
36 * software developed by the University of California, Berkeley, and its
37 * contributors.
38 */
39
40 /*
41 * VM - shared or copy-on-write from a vnode/anonymous memory.
42 */
43
44 #include <sys/types.h>
45 #include <sys/param.h>
46 #include <sys/t_lock.h>
47 #include <sys/errno.h>
48 #include <sys/systm.h>
49 #include <sys/mman.h>
50 #include <sys/debug.h>
51 #include <sys/cred.h>
52 #include <sys/vmsystm.h>
53 #include <sys/tuneable.h>
54 #include <sys/bitmap.h>
55 #include <sys/swap.h>
56 #include <sys/kmem.h>
57 #include <sys/sysmacros.h>
58 #include <sys/vtrace.h>
59 #include <sys/cmn_err.h>
60 #include <sys/callb.h>
61 #include <sys/vm.h>
62 #include <sys/dumphdr.h>
63 #include <sys/lgrp.h>
64
65 #include <vm/hat.h>
66 #include <vm/as.h>
67 #include <vm/seg.h>
68 #include <vm/seg_vn.h>
69 #include <vm/pvn.h>
70 #include <vm/anon.h>
71 #include <vm/page.h>
72 #include <vm/vpage.h>
73 #include <sys/proc.h>
74 #include <sys/task.h>
75 #include <sys/project.h>
76 #include <sys/zone.h>
77 #include <sys/shm_impl.h>
78
79 /*
80 * segvn_fault needs a temporary page list array. To avoid calling kmem all
81 * the time, it creates a small (PVN_GETPAGE_NUM entry) array and uses it if
82 * it can. In the rare case when this page list is not large enough, it
83 * goes and gets a large enough array from kmem.
84 *
85 * This small page list array covers either 8 pages or 64kB worth of pages -
86 * whichever is smaller.
87 */
88 #define PVN_MAX_GETPAGE_SZ 0x10000
89 #define PVN_MAX_GETPAGE_NUM 0x8
90
91 #if PVN_MAX_GETPAGE_SZ > PVN_MAX_GETPAGE_NUM * PAGESIZE
92 #define PVN_GETPAGE_SZ ptob(PVN_MAX_GETPAGE_NUM)
93 #define PVN_GETPAGE_NUM PVN_MAX_GETPAGE_NUM
94 #else
95 #define PVN_GETPAGE_SZ PVN_MAX_GETPAGE_SZ
96 #define PVN_GETPAGE_NUM btop(PVN_MAX_GETPAGE_SZ)
97 #endif
98
99 /*
100 * Private seg op routines.
101 */
102 static int segvn_dup(struct seg *seg, struct seg *newseg);
103 static int segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
104 static void segvn_free(struct seg *seg);
105 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
106 caddr_t addr, size_t len, enum fault_type type,
107 enum seg_rw rw);
108 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
109 static int segvn_setprot(struct seg *seg, caddr_t addr,
110 size_t len, uint_t prot);
111 static int segvn_checkprot(struct seg *seg, caddr_t addr,
112 size_t len, uint_t prot);
113 static int segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
114 static int segvn_sync(struct seg *seg, caddr_t addr, size_t len,
115 int attr, uint_t flags);
116 static size_t segvn_incore(struct seg *seg, caddr_t addr, size_t len,
117 char *vec);
118 static int segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
119 int attr, int op, ulong_t *lockmap, size_t pos);
120 static int segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
121 uint_t *protv);
122 static u_offset_t segvn_getoffset(struct seg *seg, caddr_t addr);
123 static int segvn_gettype(struct seg *seg, caddr_t addr);
124 static int segvn_getvp(struct seg *seg, caddr_t addr,
125 struct vnode **vpp);
126 static int segvn_advise(struct seg *seg, caddr_t addr, size_t len,
127 uint_t behav);
128 static void segvn_dump(struct seg *seg);
129 static int segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
130 struct page ***ppp, enum lock_type type, enum seg_rw rw);
131 static int segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
132 uint_t szc);
133 static int segvn_getmemid(struct seg *seg, caddr_t addr,
134 memid_t *memidp);
135 static lgrp_mem_policy_info_t *segvn_getpolicy(struct seg *, caddr_t);
136 static int segvn_capable(struct seg *seg, segcapability_t capable);
137 static int segvn_inherit(struct seg *, caddr_t, size_t, uint_t);
138
139 struct seg_ops segvn_ops = {
140 segvn_dup,
141 segvn_unmap,
142 segvn_free,
143 segvn_fault,
144 segvn_faulta,
145 segvn_setprot,
146 segvn_checkprot,
147 segvn_kluster,
148 segvn_sync,
149 segvn_incore,
150 segvn_lockop,
151 segvn_getprot,
152 segvn_getoffset,
153 segvn_gettype,
154 segvn_getvp,
155 segvn_advise,
156 segvn_dump,
157 segvn_pagelock,
158 segvn_setpagesize,
159 segvn_getmemid,
160 segvn_getpolicy,
161 segvn_capable,
162 segvn_inherit
163 };
164
165 /*
166 * Common zfod structures, provided as a shorthand for others to use.
167 */
168 static segvn_crargs_t zfod_segvn_crargs =
169 SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
170 static segvn_crargs_t kzfod_segvn_crargs =
171 SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
172 PROT_ALL & ~PROT_USER);
173 static segvn_crargs_t stack_noexec_crargs =
174 SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
175
176 caddr_t zfod_argsp = (caddr_t)&zfod_segvn_crargs; /* user zfod argsp */
177 caddr_t kzfod_argsp = (caddr_t)&kzfod_segvn_crargs; /* kernel zfod argsp */
178 caddr_t stack_exec_argsp = (caddr_t)&zfod_segvn_crargs; /* executable stack */
179 caddr_t stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
180
181 #define vpgtob(n) ((n) * sizeof (struct vpage)) /* For brevity */
182
183 size_t segvn_comb_thrshld = UINT_MAX; /* patchable -- see 1196681 */
184
185 size_t segvn_pglock_comb_thrshld = (1UL << 16); /* 64K */
186 size_t segvn_pglock_comb_balign = (1UL << 16); /* 64K */
187 uint_t segvn_pglock_comb_bshift;
188 size_t segvn_pglock_comb_palign;
189
190 static int segvn_concat(struct seg *, struct seg *, int);
191 static int segvn_extend_prev(struct seg *, struct seg *,
192 struct segvn_crargs *, size_t);
193 static int segvn_extend_next(struct seg *, struct seg *,
194 struct segvn_crargs *, size_t);
195 static void segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
196 static void segvn_pagelist_rele(page_t **);
197 static void segvn_setvnode_mpss(vnode_t *);
198 static void segvn_relocate_pages(page_t **, page_t *);
199 static int segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
200 static int segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
201 uint_t, page_t **, page_t **, uint_t *, int *);
202 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
203 caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
204 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
205 caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
206 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
207 u_offset_t, struct vpage *, page_t **, uint_t,
208 enum fault_type, enum seg_rw, int);
209 static void segvn_vpage(struct seg *);
210 static size_t segvn_count_swap_by_vpages(struct seg *);
211
212 static void segvn_purge(struct seg *seg);
213 static int segvn_reclaim(void *, caddr_t, size_t, struct page **,
214 enum seg_rw, int);
215 static int shamp_reclaim(void *, caddr_t, size_t, struct page **,
216 enum seg_rw, int);
217
218 static int sameprot(struct seg *, caddr_t, size_t);
219
220 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
221 static int segvn_clrszc(struct seg *);
222 static struct seg *segvn_split_seg(struct seg *, caddr_t);
223 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
224 ulong_t, uint_t);
225
226 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
227 size_t, void *, u_offset_t);
228
229 static struct kmem_cache *segvn_cache;
230 static struct kmem_cache **segvn_szc_cache;
231
232 #ifdef VM_STATS
233 static struct segvnvmstats_str {
234 ulong_t fill_vp_pages[31];
235 ulong_t fltvnpages[49];
236 ulong_t fullszcpages[10];
237 ulong_t relocatepages[3];
238 ulong_t fltanpages[17];
239 ulong_t pagelock[2];
240 ulong_t demoterange[3];
241 } segvnvmstats;
242 #endif /* VM_STATS */
243
244 #define SDR_RANGE 1 /* demote entire range */
245 #define SDR_END 2 /* demote non aligned ends only */
246
247 #define CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) { \
248 if ((len) != 0) { \
249 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz); \
250 ASSERT(lpgaddr >= (seg)->s_base); \
251 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) + \
252 (len)), pgsz); \
253 ASSERT(lpgeaddr > lpgaddr); \
254 ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size); \
255 } else { \
256 lpgeaddr = lpgaddr = (addr); \
257 } \
258 }
259
260 /*ARGSUSED*/
261 static int
262 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
263 {
264 struct segvn_data *svd = buf;
265
266 rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
267 mutex_init(&svd->segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
268 svd->svn_trnext = svd->svn_trprev = NULL;
269 return (0);
270 }
271
272 /*ARGSUSED1*/
273 static void
274 segvn_cache_destructor(void *buf, void *cdrarg)
275 {
276 struct segvn_data *svd = buf;
277
278 rw_destroy(&svd->lock);
279 mutex_destroy(&svd->segfree_syncmtx);
280 }
281
282 /*ARGSUSED*/
283 static int
284 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
285 {
286 bzero(buf, sizeof (svntr_t));
287 return (0);
288 }
289
290 /*
291 * Patching this variable to non-zero allows the system to run with
292 * stacks marked as "not executable". It's a bit of a kludge, but is
293 * provided as a tweakable for platforms that export those ABIs
294 * (e.g. sparc V8) that have executable stacks enabled by default.
295 * There are also some restrictions for platforms that don't actually
296 * implement 'noexec' protections.
297 *
298 * Once enabled, the system is (therefore) unable to provide a fully
299 * ABI-compliant execution environment, though practically speaking,
300 * most everything works. The exceptions are generally some interpreters
301 * and debuggers that create executable code on the stack and jump
302 * into it (without explicitly mprotecting the address range to include
303 * PROT_EXEC).
304 *
305 * One important class of applications that are disabled are those
306 * that have been transformed into malicious agents using one of the
307 * numerous "buffer overflow" attacks. See 4007890.
308 */
309 int noexec_user_stack = 0;
310 int noexec_user_stack_log = 1;
311
312 int segvn_lpg_disable = 0;
313 uint_t segvn_maxpgszc = 0;
314
315 ulong_t segvn_vmpss_clrszc_cnt;
316 ulong_t segvn_vmpss_clrszc_err;
317 ulong_t segvn_fltvnpages_clrszc_cnt;
318 ulong_t segvn_fltvnpages_clrszc_err;
319 ulong_t segvn_setpgsz_align_err;
320 ulong_t segvn_setpgsz_anon_align_err;
321 ulong_t segvn_setpgsz_getattr_err;
322 ulong_t segvn_setpgsz_eof_err;
323 ulong_t segvn_faultvnmpss_align_err1;
324 ulong_t segvn_faultvnmpss_align_err2;
325 ulong_t segvn_faultvnmpss_align_err3;
326 ulong_t segvn_faultvnmpss_align_err4;
327 ulong_t segvn_faultvnmpss_align_err5;
328 ulong_t segvn_vmpss_pageio_deadlk_err;
329
330 int segvn_use_regions = 1;
331
332 /*
333 * Segvn supports text replication optimization for NUMA platforms. Text
334 * replica's are represented by anon maps (amp). There's one amp per text file
335 * region per lgroup. A process chooses the amp for each of its text mappings
336 * based on the lgroup assignment of its main thread (t_tid = 1). All
337 * processes that want a replica on a particular lgroup for the same text file
338 * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
339 * with vp,off,size,szc used as a key. Text replication segments are read only
340 * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
341 * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
342 * pages. Replication amp is assigned to a segment when it gets its first
343 * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
344 * rechecks periodically if the process still maps an amp local to the main
345 * thread. If not async thread forces process to remap to an amp in the new
346 * home lgroup of the main thread. Current text replication implementation
347 * only provides the benefit to workloads that do most of their work in the
348 * main thread of a process or all the threads of a process run in the same
349 * lgroup. To extend text replication benefit to different types of
350 * multithreaded workloads further work would be needed in the hat layer to
351 * allow the same virtual address in the same hat to simultaneously map
352 * different physical addresses (i.e. page table replication would be needed
353 * for x86).
354 *
355 * amp pages are used instead of vnode pages as long as segment has a very
356 * simple life cycle. It's created via segvn_create(), handles S_EXEC
357 * (S_READ) pagefaults and is fully unmapped. If anything more complicated
358 * happens such as protection is changed, real COW fault happens, pagesize is
359 * changed, MC_LOCK is requested or segment is partially unmapped we turn off
360 * text replication by converting the segment back to vnode only segment
361 * (unmap segment's address range and set svd->amp to NULL).
362 *
363 * The original file can be changed after amp is inserted into
364 * svntr_hashtab. Processes that are launched after the file is already
365 * changed can't use the replica's created prior to the file change. To
366 * implement this functionality hash entries are timestamped. Replica's can
367 * only be used if current file modification time is the same as the timestamp
368 * saved when hash entry was created. However just timestamps alone are not
369 * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
370 * deal with file changes via MAP_SHARED mappings differently. When writable
371 * MAP_SHARED mappings are created to vnodes marked as executable we mark all
372 * existing replica's for this vnode as not usable for future text
373 * mappings. And we don't create new replica's for files that currently have
374 * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
375 * true).
376 */
377
378 #define SEGVN_TEXTREPL_MAXBYTES_FACTOR (20)
379 size_t segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
380
381 static ulong_t svntr_hashtab_sz = 512;
382 static svntr_bucket_t *svntr_hashtab = NULL;
383 static struct kmem_cache *svntr_cache;
384 static svntr_stats_t *segvn_textrepl_stats;
385 static ksema_t segvn_trasync_sem;
386
387 int segvn_disable_textrepl = 1;
388 size_t textrepl_size_thresh = (size_t)-1;
389 size_t segvn_textrepl_bytes = 0;
390 size_t segvn_textrepl_max_bytes = 0;
391 clock_t segvn_update_textrepl_interval = 0;
392 int segvn_update_tr_time = 10;
393 int segvn_disable_textrepl_update = 0;
394
395 static void segvn_textrepl(struct seg *);
396 static void segvn_textunrepl(struct seg *, int);
397 static void segvn_inval_trcache(vnode_t *);
398 static void segvn_trasync_thread(void);
399 static void segvn_trupdate_wakeup(void *);
400 static void segvn_trupdate(void);
401 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
402 ulong_t);
403
404 /*
405 * Initialize segvn data structures
406 */
407 void
408 segvn_init(void)
409 {
410 uint_t maxszc;
411 uint_t szc;
412 size_t pgsz;
413
414 segvn_cache = kmem_cache_create("segvn_cache",
415 sizeof (struct segvn_data), 0,
416 segvn_cache_constructor, segvn_cache_destructor, NULL,
417 NULL, NULL, 0);
418
419 if (segvn_lpg_disable == 0) {
420 szc = maxszc = page_num_pagesizes() - 1;
421 if (szc == 0) {
422 segvn_lpg_disable = 1;
423 }
424 if (page_get_pagesize(0) != PAGESIZE) {
425 panic("segvn_init: bad szc 0");
426 /*NOTREACHED*/
427 }
428 while (szc != 0) {
429 pgsz = page_get_pagesize(szc);
430 if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
431 panic("segvn_init: bad szc %d", szc);
432 /*NOTREACHED*/
433 }
434 szc--;
435 }
436 if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
437 segvn_maxpgszc = maxszc;
438 }
439
440 if (segvn_maxpgszc) {
441 segvn_szc_cache = (struct kmem_cache **)kmem_alloc(
442 (segvn_maxpgszc + 1) * sizeof (struct kmem_cache *),
443 KM_SLEEP);
444 }
445
446 for (szc = 1; szc <= segvn_maxpgszc; szc++) {
447 char str[32];
448
449 (void) sprintf(str, "segvn_szc_cache%d", szc);
450 segvn_szc_cache[szc] = kmem_cache_create(str,
451 page_get_pagecnt(szc) * sizeof (page_t *), 0,
452 NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
453 }
454
455
456 if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
457 segvn_use_regions = 0;
458
459 /*
460 * For now shared regions and text replication segvn support
461 * are mutually exclusive. This is acceptable because
462 * currently significant benefit from text replication was
463 * only observed on AMD64 NUMA platforms (due to relatively
464 * small L2$ size) and currently we don't support shared
465 * regions on x86.
466 */
467 if (segvn_use_regions && !segvn_disable_textrepl) {
468 segvn_disable_textrepl = 1;
469 }
470
471 #if defined(_LP64)
472 if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
473 !segvn_disable_textrepl) {
474 ulong_t i;
475 size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
476
477 svntr_cache = kmem_cache_create("svntr_cache",
478 sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
479 NULL, NULL, NULL, 0);
480 svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
481 for (i = 0; i < svntr_hashtab_sz; i++) {
482 mutex_init(&svntr_hashtab[i].tr_lock, NULL,
483 MUTEX_DEFAULT, NULL);
484 }
485 segvn_textrepl_max_bytes = ptob(physmem) /
486 segvn_textrepl_max_bytes_factor;
487 segvn_textrepl_stats = kmem_zalloc(NCPU *
488 sizeof (svntr_stats_t), KM_SLEEP);
489 sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
490 (void) thread_create(NULL, 0, segvn_trasync_thread,
491 NULL, 0, &p0, TS_RUN, minclsyspri);
492 }
493 #endif
494
495 if (!ISP2(segvn_pglock_comb_balign) ||
496 segvn_pglock_comb_balign < PAGESIZE) {
497 segvn_pglock_comb_balign = 1UL << 16; /* 64K */
498 }
499 segvn_pglock_comb_bshift = highbit(segvn_pglock_comb_balign) - 1;
500 segvn_pglock_comb_palign = btop(segvn_pglock_comb_balign);
501 }
502
503 #define SEGVN_PAGEIO ((void *)0x1)
504 #define SEGVN_NOPAGEIO ((void *)0x2)
505
506 static void
507 segvn_setvnode_mpss(vnode_t *vp)
508 {
509 int err;
510
511 ASSERT(vp->v_mpssdata == NULL ||
512 vp->v_mpssdata == SEGVN_PAGEIO ||
513 vp->v_mpssdata == SEGVN_NOPAGEIO);
514
515 if (vp->v_mpssdata == NULL) {
516 if (vn_vmpss_usepageio(vp)) {
517 err = VOP_PAGEIO(vp, (page_t *)NULL,
518 (u_offset_t)0, 0, 0, CRED(), NULL);
519 } else {
520 err = ENOSYS;
521 }
522 /*
523 * set v_mpssdata just once per vnode life
524 * so that it never changes.
525 */
526 mutex_enter(&vp->v_lock);
527 if (vp->v_mpssdata == NULL) {
528 if (err == EINVAL) {
529 vp->v_mpssdata = SEGVN_PAGEIO;
530 } else {
531 vp->v_mpssdata = SEGVN_NOPAGEIO;
532 }
533 }
534 mutex_exit(&vp->v_lock);
535 }
536 }
537
538 int
539 segvn_create(struct seg *seg, void *argsp)
540 {
541 struct segvn_crargs *a = (struct segvn_crargs *)argsp;
542 struct segvn_data *svd;
543 size_t swresv = 0;
544 struct cred *cred;
545 struct anon_map *amp;
546 int error = 0;
547 size_t pgsz;
548 lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
549 int use_rgn = 0;
550 int trok = 0;
551
552 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
553
554 if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
555 panic("segvn_create type");
556 /*NOTREACHED*/
557 }
558
559 /*
560 * Check arguments. If a shared anon structure is given then
561 * it is illegal to also specify a vp.
562 */
563 if (a->amp != NULL && a->vp != NULL) {
564 panic("segvn_create anon_map");
565 /*NOTREACHED*/
566 }
567
568 if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
569 a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
570 segvn_use_regions) {
571 use_rgn = 1;
572 }
573
574 /* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
575 if (a->type == MAP_SHARED)
576 a->flags &= ~MAP_NORESERVE;
577
578 if (a->szc != 0) {
579 if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
580 (a->amp != NULL && a->type == MAP_PRIVATE) ||
581 (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
582 a->szc = 0;
583 } else {
584 if (a->szc > segvn_maxpgszc)
585 a->szc = segvn_maxpgszc;
586 pgsz = page_get_pagesize(a->szc);
587 if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
588 !IS_P2ALIGNED(seg->s_size, pgsz)) {
589 a->szc = 0;
590 } else if (a->vp != NULL) {
591 if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
592 /*
593 * paranoid check.
594 * hat_page_demote() is not supported
595 * on swapfs pages.
596 */
597 a->szc = 0;
598 } else if (map_addr_vacalign_check(seg->s_base,
599 a->offset & PAGEMASK)) {
600 a->szc = 0;
601 }
602 } else if (a->amp != NULL) {
603 pgcnt_t anum = btopr(a->offset);
604 pgcnt_t pgcnt = page_get_pagecnt(a->szc);
605 if (!IS_P2ALIGNED(anum, pgcnt)) {
606 a->szc = 0;
607 }
608 }
609 }
610 }
611
612 /*
613 * If segment may need private pages, reserve them now.
614 */
615 if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
616 (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
617 if (anon_resv_zone(seg->s_size,
618 seg->s_as->a_proc->p_zone) == 0)
619 return (EAGAIN);
620 swresv = seg->s_size;
621 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
622 seg, swresv, 1);
623 }
624
625 /*
626 * Reserve any mapping structures that may be required.
627 *
628 * Don't do it for segments that may use regions. It's currently a
629 * noop in the hat implementations anyway.
630 */
631 if (!use_rgn) {
632 hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
633 }
634
635 if (a->cred) {
636 cred = a->cred;
637 crhold(cred);
638 } else {
639 crhold(cred = CRED());
640 }
641
642 /* Inform the vnode of the new mapping */
643 if (a->vp != NULL) {
644 error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
645 seg->s_as, seg->s_base, seg->s_size, a->prot,
646 a->maxprot, a->type, cred, NULL);
647 if (error) {
648 if (swresv != 0) {
649 anon_unresv_zone(swresv,
650 seg->s_as->a_proc->p_zone);
651 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
652 "anon proc:%p %lu %u", seg, swresv, 0);
653 }
654 crfree(cred);
655 if (!use_rgn) {
656 hat_unload(seg->s_as->a_hat, seg->s_base,
657 seg->s_size, HAT_UNLOAD_UNMAP);
658 }
659 return (error);
660 }
661 /*
662 * svntr_hashtab will be NULL if we support shared regions.
663 */
664 trok = ((a->flags & MAP_TEXT) &&
665 (seg->s_size > textrepl_size_thresh ||
666 (a->flags & _MAP_TEXTREPL)) &&
667 lgrp_optimizations() && svntr_hashtab != NULL &&
668 a->type == MAP_PRIVATE && swresv == 0 &&
669 !(a->flags & MAP_NORESERVE) &&
670 seg->s_as != &kas && a->vp->v_type == VREG);
671
672 ASSERT(!trok || !use_rgn);
673 }
674
675 /*
676 * MAP_NORESERVE mappings don't count towards the VSZ of a process
677 * until we fault the pages in.
678 */
679 if ((a->vp == NULL || a->vp->v_type != VREG) &&
680 a->flags & MAP_NORESERVE) {
681 seg->s_as->a_resvsize -= seg->s_size;
682 }
683
684 /*
685 * If more than one segment in the address space, and they're adjacent
686 * virtually, try to concatenate them. Don't concatenate if an
687 * explicit anon_map structure was supplied (e.g., SystemV shared
688 * memory) or if we'll use text replication for this segment.
689 */
690 if (a->amp == NULL && !use_rgn && !trok) {
691 struct seg *pseg, *nseg;
692 struct segvn_data *psvd, *nsvd;
693 lgrp_mem_policy_t ppolicy, npolicy;
694 uint_t lgrp_mem_policy_flags = 0;
695 extern lgrp_mem_policy_t lgrp_mem_default_policy;
696
697 /*
698 * Memory policy flags (lgrp_mem_policy_flags) is valid when
699 * extending stack/heap segments.
700 */
701 if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
702 !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
703 lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
704 } else {
705 /*
706 * Get policy when not extending it from another segment
707 */
708 mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
709 }
710
711 /*
712 * First, try to concatenate the previous and new segments
713 */
714 pseg = AS_SEGPREV(seg->s_as, seg);
715 if (pseg != NULL &&
716 pseg->s_base + pseg->s_size == seg->s_base &&
717 pseg->s_ops == &segvn_ops) {
718 /*
719 * Get memory allocation policy from previous segment.
720 * When extension is specified (e.g. for heap) apply
721 * this policy to the new segment regardless of the
722 * outcome of segment concatenation. Extension occurs
723 * for non-default policy otherwise default policy is
724 * used and is based on extended segment size.
725 */
726 psvd = (struct segvn_data *)pseg->s_data;
727 ppolicy = psvd->policy_info.mem_policy;
728 if (lgrp_mem_policy_flags ==
729 LGRP_MP_FLAG_EXTEND_UP) {
730 if (ppolicy != lgrp_mem_default_policy) {
731 mpolicy = ppolicy;
732 } else {
733 mpolicy = lgrp_mem_policy_default(
734 pseg->s_size + seg->s_size,
735 a->type);
736 }
737 }
738
739 if (mpolicy == ppolicy &&
740 (pseg->s_size + seg->s_size <=
741 segvn_comb_thrshld || psvd->amp == NULL) &&
742 segvn_extend_prev(pseg, seg, a, swresv) == 0) {
743 /*
744 * success! now try to concatenate
745 * with following seg
746 */
747 crfree(cred);
748 nseg = AS_SEGNEXT(pseg->s_as, pseg);
749 if (nseg != NULL &&
750 nseg != pseg &&
751 nseg->s_ops == &segvn_ops &&
752 pseg->s_base + pseg->s_size ==
753 nseg->s_base)
754 (void) segvn_concat(pseg, nseg, 0);
755 ASSERT(pseg->s_szc == 0 ||
756 (a->szc == pseg->s_szc &&
757 IS_P2ALIGNED(pseg->s_base, pgsz) &&
758 IS_P2ALIGNED(pseg->s_size, pgsz)));
759 return (0);
760 }
761 }
762
763 /*
764 * Failed, so try to concatenate with following seg
765 */
766 nseg = AS_SEGNEXT(seg->s_as, seg);
767 if (nseg != NULL &&
768 seg->s_base + seg->s_size == nseg->s_base &&
769 nseg->s_ops == &segvn_ops) {
770 /*
771 * Get memory allocation policy from next segment.
772 * When extension is specified (e.g. for stack) apply
773 * this policy to the new segment regardless of the
774 * outcome of segment concatenation. Extension occurs
775 * for non-default policy otherwise default policy is
776 * used and is based on extended segment size.
777 */
778 nsvd = (struct segvn_data *)nseg->s_data;
779 npolicy = nsvd->policy_info.mem_policy;
780 if (lgrp_mem_policy_flags ==
781 LGRP_MP_FLAG_EXTEND_DOWN) {
782 if (npolicy != lgrp_mem_default_policy) {
783 mpolicy = npolicy;
784 } else {
785 mpolicy = lgrp_mem_policy_default(
786 nseg->s_size + seg->s_size,
787 a->type);
788 }
789 }
790
791 if (mpolicy == npolicy &&
792 segvn_extend_next(seg, nseg, a, swresv) == 0) {
793 crfree(cred);
794 ASSERT(nseg->s_szc == 0 ||
795 (a->szc == nseg->s_szc &&
796 IS_P2ALIGNED(nseg->s_base, pgsz) &&
797 IS_P2ALIGNED(nseg->s_size, pgsz)));
798 return (0);
799 }
800 }
801 }
802
803 if (a->vp != NULL) {
804 VN_HOLD(a->vp);
805 if (a->type == MAP_SHARED)
806 lgrp_shm_policy_init(NULL, a->vp);
807 }
808 svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
809
810 seg->s_ops = &segvn_ops;
811 seg->s_data = (void *)svd;
812 seg->s_szc = a->szc;
813
814 svd->seg = seg;
815 svd->vp = a->vp;
816 /*
817 * Anonymous mappings have no backing file so the offset is meaningless.
818 */
819 svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
820 svd->prot = a->prot;
821 svd->maxprot = a->maxprot;
822 svd->pageprot = 0;
823 svd->type = a->type;
824 svd->vpage = NULL;
825 svd->cred = cred;
826 svd->advice = MADV_NORMAL;
827 svd->pageadvice = 0;
828 svd->flags = (ushort_t)a->flags;
829 svd->softlockcnt = 0;
830 svd->softlockcnt_sbase = 0;
831 svd->softlockcnt_send = 0;
832 svd->svn_inz = 0;
833 svd->rcookie = HAT_INVALID_REGION_COOKIE;
834 svd->pageswap = 0;
835
836 if (a->szc != 0 && a->vp != NULL) {
837 segvn_setvnode_mpss(a->vp);
838 }
839 if (svd->type == MAP_SHARED && svd->vp != NULL &&
840 (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
841 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
842 segvn_inval_trcache(svd->vp);
843 }
844
845 amp = a->amp;
846 if ((svd->amp = amp) == NULL) {
847 svd->anon_index = 0;
848 if (svd->type == MAP_SHARED) {
849 svd->swresv = 0;
850 /*
851 * Shared mappings to a vp need no other setup.
852 * If we have a shared mapping to an anon_map object
853 * which hasn't been allocated yet, allocate the
854 * struct now so that it will be properly shared
855 * by remembering the swap reservation there.
856 */
857 if (a->vp == NULL) {
858 svd->amp = anonmap_alloc(seg->s_size, swresv,
859 ANON_SLEEP);
860 svd->amp->a_szc = seg->s_szc;
861 }
862 } else {
863 /*
864 * Private mapping (with or without a vp).
865 * Allocate anon_map when needed.
866 */
867 svd->swresv = swresv;
868 }
869 } else {
870 pgcnt_t anon_num;
871
872 /*
873 * Mapping to an existing anon_map structure without a vp.
874 * For now we will insure that the segment size isn't larger
875 * than the size - offset gives us. Later on we may wish to
876 * have the anon array dynamically allocated itself so that
877 * we don't always have to allocate all the anon pointer slots.
878 * This of course involves adding extra code to check that we
879 * aren't trying to use an anon pointer slot beyond the end
880 * of the currently allocated anon array.
881 */
882 if ((amp->size - a->offset) < seg->s_size) {
883 panic("segvn_create anon_map size");
884 /*NOTREACHED*/
885 }
886
887 anon_num = btopr(a->offset);
888
889 if (a->type == MAP_SHARED) {
890 /*
891 * SHARED mapping to a given anon_map.
892 */
893 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
894 amp->refcnt++;
895 if (a->szc > amp->a_szc) {
896 amp->a_szc = a->szc;
897 }
898 ANON_LOCK_EXIT(&->a_rwlock);
899 svd->anon_index = anon_num;
900 svd->swresv = 0;
901 } else {
902 /*
903 * PRIVATE mapping to a given anon_map.
904 * Make sure that all the needed anon
905 * structures are created (so that we will
906 * share the underlying pages if nothing
907 * is written by this mapping) and then
908 * duplicate the anon array as is done
909 * when a privately mapped segment is dup'ed.
910 */
911 struct anon *ap;
912 caddr_t addr;
913 caddr_t eaddr;
914 ulong_t anon_idx;
915 int hat_flag = HAT_LOAD;
916
917 if (svd->flags & MAP_TEXT) {
918 hat_flag |= HAT_LOAD_TEXT;
919 }
920
921 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
922 svd->amp->a_szc = seg->s_szc;
923 svd->anon_index = 0;
924 svd->swresv = swresv;
925
926 /*
927 * Prevent 2 threads from allocating anon
928 * slots simultaneously.
929 */
930 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
931 eaddr = seg->s_base + seg->s_size;
932
933 for (anon_idx = anon_num, addr = seg->s_base;
934 addr < eaddr; addr += PAGESIZE, anon_idx++) {
935 page_t *pp;
936
937 if ((ap = anon_get_ptr(amp->ahp,
938 anon_idx)) != NULL)
939 continue;
940
941 /*
942 * Allocate the anon struct now.
943 * Might as well load up translation
944 * to the page while we're at it...
945 */
946 pp = anon_zero(seg, addr, &ap, cred);
947 if (ap == NULL || pp == NULL) {
948 panic("segvn_create anon_zero");
949 /*NOTREACHED*/
950 }
951
952 /*
953 * Re-acquire the anon_map lock and
954 * initialize the anon array entry.
955 */
956 ASSERT(anon_get_ptr(amp->ahp,
957 anon_idx) == NULL);
958 (void) anon_set_ptr(amp->ahp, anon_idx, ap,
959 ANON_SLEEP);
960
961 ASSERT(seg->s_szc == 0);
962 ASSERT(!IS_VMODSORT(pp->p_vnode));
963
964 ASSERT(use_rgn == 0);
965 hat_memload(seg->s_as->a_hat, addr, pp,
966 svd->prot & ~PROT_WRITE, hat_flag);
967
968 page_unlock(pp);
969 }
970 ASSERT(seg->s_szc == 0);
971 anon_dup(amp->ahp, anon_num, svd->amp->ahp,
972 0, seg->s_size);
973 ANON_LOCK_EXIT(&->a_rwlock);
974 }
975 }
976
977 /*
978 * Set default memory allocation policy for segment
979 *
980 * Always set policy for private memory at least for initialization
981 * even if this is a shared memory segment
982 */
983 (void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
984
985 if (svd->type == MAP_SHARED)
986 (void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
987 svd->vp, svd->offset, seg->s_size);
988
989 if (use_rgn) {
990 ASSERT(!trok);
991 ASSERT(svd->amp == NULL);
992 svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
993 seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
994 (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
995 HAT_REGION_TEXT);
996 }
997
998 ASSERT(!trok || !(svd->prot & PROT_WRITE));
999 svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
1000
1001 return (0);
1002 }
1003
1004 /*
1005 * Concatenate two existing segments, if possible.
1006 * Return 0 on success, -1 if two segments are not compatible
1007 * or -2 on memory allocation failure.
1008 * If amp_cat == 1 then try and concat segments with anon maps
1009 */
1010 static int
1011 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
1012 {
1013 struct segvn_data *svd1 = seg1->s_data;
1014 struct segvn_data *svd2 = seg2->s_data;
1015 struct anon_map *amp1 = svd1->amp;
1016 struct anon_map *amp2 = svd2->amp;
1017 struct vpage *vpage1 = svd1->vpage;
1018 struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
1019 size_t size, nvpsize;
1020 pgcnt_t npages1, npages2;
1021
1022 ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
1023 ASSERT(AS_WRITE_HELD(seg1->s_as));
1024 ASSERT(seg1->s_ops == seg2->s_ops);
1025
1026 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1027 HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1028 return (-1);
1029 }
1030
1031 /* both segments exist, try to merge them */
1032 #define incompat(x) (svd1->x != svd2->x)
1033 if (incompat(vp) || incompat(maxprot) ||
1034 (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1035 (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1036 incompat(type) || incompat(cred) || incompat(flags) ||
1037 seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1038 (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1039 return (-1);
1040 #undef incompat
1041
1042 /*
1043 * vp == NULL implies zfod, offset doesn't matter
1044 */
1045 if (svd1->vp != NULL &&
1046 svd1->offset + seg1->s_size != svd2->offset) {
1047 return (-1);
1048 }
1049
1050 /*
1051 * Don't concatenate if either segment uses text replication.
1052 */
1053 if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1054 return (-1);
1055 }
1056
1057 /*
1058 * Fail early if we're not supposed to concatenate
1059 * segments with non NULL amp.
1060 */
1061 if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1062 return (-1);
1063 }
1064
1065 if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1066 if (amp1 != amp2) {
1067 return (-1);
1068 }
1069 if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1070 svd2->anon_index) {
1071 return (-1);
1072 }
1073 ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1074 }
1075
1076 /*
1077 * If either seg has vpages, create a new merged vpage array.
1078 */
1079 if (vpage1 != NULL || vpage2 != NULL) {
1080 struct vpage *vp, *evp;
1081
1082 npages1 = seg_pages(seg1);
1083 npages2 = seg_pages(seg2);
1084 nvpsize = vpgtob(npages1 + npages2);
1085
1086 if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1087 return (-2);
1088 }
1089
1090 if (vpage1 != NULL) {
1091 bcopy(vpage1, nvpage, vpgtob(npages1));
1092 } else {
1093 evp = nvpage + npages1;
1094 for (vp = nvpage; vp < evp; vp++) {
1095 VPP_SETPROT(vp, svd1->prot);
1096 VPP_SETADVICE(vp, svd1->advice);
1097 }
1098 }
1099
1100 if (vpage2 != NULL) {
1101 bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1102 } else {
1103 evp = nvpage + npages1 + npages2;
1104 for (vp = nvpage + npages1; vp < evp; vp++) {
1105 VPP_SETPROT(vp, svd2->prot);
1106 VPP_SETADVICE(vp, svd2->advice);
1107 }
1108 }
1109
1110 if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1111 ASSERT(svd1->swresv == seg1->s_size);
1112 ASSERT(!(svd1->flags & MAP_NORESERVE));
1113 ASSERT(!(svd2->flags & MAP_NORESERVE));
1114 evp = nvpage + npages1;
1115 for (vp = nvpage; vp < evp; vp++) {
1116 VPP_SETSWAPRES(vp);
1117 }
1118 }
1119
1120 if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1121 ASSERT(svd2->swresv == seg2->s_size);
1122 ASSERT(!(svd1->flags & MAP_NORESERVE));
1123 ASSERT(!(svd2->flags & MAP_NORESERVE));
1124 vp = nvpage + npages1;
1125 evp = vp + npages2;
1126 for (; vp < evp; vp++) {
1127 VPP_SETSWAPRES(vp);
1128 }
1129 }
1130 }
1131 ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1132 (svd1->pageswap == 0 && svd2->pageswap == 0));
1133
1134 /*
1135 * If either segment has private pages, create a new merged anon
1136 * array. If mergeing shared anon segments just decrement anon map's
1137 * refcnt.
1138 */
1139 if (amp1 != NULL && svd1->type == MAP_SHARED) {
1140 ASSERT(amp1 == amp2 && svd1->vp == NULL);
1141 ANON_LOCK_ENTER(&1->a_rwlock, RW_WRITER);
1142 ASSERT(amp1->refcnt >= 2);
1143 amp1->refcnt--;
1144 ANON_LOCK_EXIT(&1->a_rwlock);
1145 svd2->amp = NULL;
1146 } else if (amp1 != NULL || amp2 != NULL) {
1147 struct anon_hdr *nahp;
1148 struct anon_map *namp = NULL;
1149 size_t asize;
1150
1151 ASSERT(svd1->type == MAP_PRIVATE);
1152
1153 asize = seg1->s_size + seg2->s_size;
1154 if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1155 if (nvpage != NULL) {
1156 kmem_free(nvpage, nvpsize);
1157 }
1158 return (-2);
1159 }
1160 if (amp1 != NULL) {
1161 /*
1162 * XXX anon rwlock is not really needed because
1163 * this is a private segment and we are writers.
1164 */
1165 ANON_LOCK_ENTER(&1->a_rwlock, RW_WRITER);
1166 ASSERT(amp1->refcnt == 1);
1167 if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1168 nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1169 anon_release(nahp, btop(asize));
1170 ANON_LOCK_EXIT(&1->a_rwlock);
1171 if (nvpage != NULL) {
1172 kmem_free(nvpage, nvpsize);
1173 }
1174 return (-2);
1175 }
1176 }
1177 if (amp2 != NULL) {
1178 ANON_LOCK_ENTER(&2->a_rwlock, RW_WRITER);
1179 ASSERT(amp2->refcnt == 1);
1180 if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1181 nahp, btop(seg1->s_size), btop(seg2->s_size),
1182 ANON_NOSLEEP)) {
1183 anon_release(nahp, btop(asize));
1184 ANON_LOCK_EXIT(&2->a_rwlock);
1185 if (amp1 != NULL) {
1186 ANON_LOCK_EXIT(&1->a_rwlock);
1187 }
1188 if (nvpage != NULL) {
1189 kmem_free(nvpage, nvpsize);
1190 }
1191 return (-2);
1192 }
1193 }
1194 if (amp1 != NULL) {
1195 namp = amp1;
1196 anon_release(amp1->ahp, btop(amp1->size));
1197 }
1198 if (amp2 != NULL) {
1199 if (namp == NULL) {
1200 ASSERT(amp1 == NULL);
1201 namp = amp2;
1202 anon_release(amp2->ahp, btop(amp2->size));
1203 } else {
1204 amp2->refcnt--;
1205 ANON_LOCK_EXIT(&2->a_rwlock);
1206 anonmap_free(amp2);
1207 }
1208 svd2->amp = NULL; /* needed for seg_free */
1209 }
1210 namp->ahp = nahp;
1211 namp->size = asize;
1212 svd1->amp = namp;
1213 svd1->anon_index = 0;
1214 ANON_LOCK_EXIT(&namp->a_rwlock);
1215 }
1216 /*
1217 * Now free the old vpage structures.
1218 */
1219 if (nvpage != NULL) {
1220 if (vpage1 != NULL) {
1221 kmem_free(vpage1, vpgtob(npages1));
1222 }
1223 if (vpage2 != NULL) {
1224 svd2->vpage = NULL;
1225 kmem_free(vpage2, vpgtob(npages2));
1226 }
1227 if (svd2->pageprot) {
1228 svd1->pageprot = 1;
1229 }
1230 if (svd2->pageadvice) {
1231 svd1->pageadvice = 1;
1232 }
1233 if (svd2->pageswap) {
1234 svd1->pageswap = 1;
1235 }
1236 svd1->vpage = nvpage;
1237 }
1238
1239 /* all looks ok, merge segments */
1240 svd1->swresv += svd2->swresv;
1241 svd2->swresv = 0; /* so seg_free doesn't release swap space */
1242 size = seg2->s_size;
1243 seg_free(seg2);
1244 seg1->s_size += size;
1245 return (0);
1246 }
1247
1248 /*
1249 * Extend the previous segment (seg1) to include the
1250 * new segment (seg2 + a), if possible.
1251 * Return 0 on success.
1252 */
1253 static int
1254 segvn_extend_prev(seg1, seg2, a, swresv)
1255 struct seg *seg1, *seg2;
1256 struct segvn_crargs *a;
1257 size_t swresv;
1258 {
1259 struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1260 size_t size;
1261 struct anon_map *amp1;
1262 struct vpage *new_vpage;
1263
1264 /*
1265 * We don't need any segment level locks for "segvn" data
1266 * since the address space is "write" locked.
1267 */
1268 ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as));
1269
1270 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1271 return (-1);
1272 }
1273
1274 /* second segment is new, try to extend first */
1275 /* XXX - should also check cred */
1276 if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1277 (!svd1->pageprot && (svd1->prot != a->prot)) ||
1278 svd1->type != a->type || svd1->flags != a->flags ||
1279 seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1280 return (-1);
1281
1282 /* vp == NULL implies zfod, offset doesn't matter */
1283 if (svd1->vp != NULL &&
1284 svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1285 return (-1);
1286
1287 if (svd1->tr_state != SEGVN_TR_OFF) {
1288 return (-1);
1289 }
1290
1291 amp1 = svd1->amp;
1292 if (amp1) {
1293 pgcnt_t newpgs;
1294
1295 /*
1296 * Segment has private pages, can data structures
1297 * be expanded?
1298 *
1299 * Acquire the anon_map lock to prevent it from changing,
1300 * if it is shared. This ensures that the anon_map
1301 * will not change while a thread which has a read/write
1302 * lock on an address space references it.
1303 * XXX - Don't need the anon_map lock at all if "refcnt"
1304 * is 1.
1305 *
1306 * Can't grow a MAP_SHARED segment with an anonmap because
1307 * there may be existing anon slots where we want to extend
1308 * the segment and we wouldn't know what to do with them
1309 * (e.g., for tmpfs right thing is to just leave them there,
1310 * for /dev/zero they should be cleared out).
1311 */
1312 if (svd1->type == MAP_SHARED)
1313 return (-1);
1314
1315 ANON_LOCK_ENTER(&1->a_rwlock, RW_WRITER);
1316 if (amp1->refcnt > 1) {
1317 ANON_LOCK_EXIT(&1->a_rwlock);
1318 return (-1);
1319 }
1320 newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1321 btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1322
1323 if (newpgs == 0) {
1324 ANON_LOCK_EXIT(&1->a_rwlock);
1325 return (-1);
1326 }
1327 amp1->size = ptob(newpgs);
1328 ANON_LOCK_EXIT(&1->a_rwlock);
1329 }
1330 if (svd1->vpage != NULL) {
1331 struct vpage *vp, *evp;
1332 new_vpage =
1333 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1334 KM_NOSLEEP);
1335 if (new_vpage == NULL)
1336 return (-1);
1337 bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1338 kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1339 svd1->vpage = new_vpage;
1340
1341 vp = new_vpage + seg_pages(seg1);
1342 evp = vp + seg_pages(seg2);
1343 for (; vp < evp; vp++)
1344 VPP_SETPROT(vp, a->prot);
1345 if (svd1->pageswap && swresv) {
1346 ASSERT(!(svd1->flags & MAP_NORESERVE));
1347 ASSERT(swresv == seg2->s_size);
1348 vp = new_vpage + seg_pages(seg1);
1349 for (; vp < evp; vp++) {
1350 VPP_SETSWAPRES(vp);
1351 }
1352 }
1353 }
1354 ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1355 size = seg2->s_size;
1356 seg_free(seg2);
1357 seg1->s_size += size;
1358 svd1->swresv += swresv;
1359 if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1360 svd1->type == MAP_SHARED && svd1->vp != NULL &&
1361 (svd1->vp->v_flag & VVMEXEC)) {
1362 ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1363 segvn_inval_trcache(svd1->vp);
1364 }
1365 return (0);
1366 }
1367
1368 /*
1369 * Extend the next segment (seg2) to include the
1370 * new segment (seg1 + a), if possible.
1371 * Return 0 on success.
1372 */
1373 static int
1374 segvn_extend_next(
1375 struct seg *seg1,
1376 struct seg *seg2,
1377 struct segvn_crargs *a,
1378 size_t swresv)
1379 {
1380 struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1381 size_t size;
1382 struct anon_map *amp2;
1383 struct vpage *new_vpage;
1384
1385 /*
1386 * We don't need any segment level locks for "segvn" data
1387 * since the address space is "write" locked.
1388 */
1389 ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as));
1390
1391 if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1392 return (-1);
1393 }
1394
1395 /* first segment is new, try to extend second */
1396 /* XXX - should also check cred */
1397 if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1398 (!svd2->pageprot && (svd2->prot != a->prot)) ||
1399 svd2->type != a->type || svd2->flags != a->flags ||
1400 seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1401 return (-1);
1402 /* vp == NULL implies zfod, offset doesn't matter */
1403 if (svd2->vp != NULL &&
1404 (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1405 return (-1);
1406
1407 if (svd2->tr_state != SEGVN_TR_OFF) {
1408 return (-1);
1409 }
1410
1411 amp2 = svd2->amp;
1412 if (amp2) {
1413 pgcnt_t newpgs;
1414
1415 /*
1416 * Segment has private pages, can data structures
1417 * be expanded?
1418 *
1419 * Acquire the anon_map lock to prevent it from changing,
1420 * if it is shared. This ensures that the anon_map
1421 * will not change while a thread which has a read/write
1422 * lock on an address space references it.
1423 *
1424 * XXX - Don't need the anon_map lock at all if "refcnt"
1425 * is 1.
1426 */
1427 if (svd2->type == MAP_SHARED)
1428 return (-1);
1429
1430 ANON_LOCK_ENTER(&2->a_rwlock, RW_WRITER);
1431 if (amp2->refcnt > 1) {
1432 ANON_LOCK_EXIT(&2->a_rwlock);
1433 return (-1);
1434 }
1435 newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1436 btop(seg2->s_size), btop(seg1->s_size),
1437 ANON_NOSLEEP | ANON_GROWDOWN);
1438
1439 if (newpgs == 0) {
1440 ANON_LOCK_EXIT(&2->a_rwlock);
1441 return (-1);
1442 }
1443 amp2->size = ptob(newpgs);
1444 ANON_LOCK_EXIT(&2->a_rwlock);
1445 }
1446 if (svd2->vpage != NULL) {
1447 struct vpage *vp, *evp;
1448 new_vpage =
1449 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1450 KM_NOSLEEP);
1451 if (new_vpage == NULL) {
1452 /* Not merging segments so adjust anon_index back */
1453 if (amp2)
1454 svd2->anon_index += seg_pages(seg1);
1455 return (-1);
1456 }
1457 bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1458 vpgtob(seg_pages(seg2)));
1459 kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1460 svd2->vpage = new_vpage;
1461
1462 vp = new_vpage;
1463 evp = vp + seg_pages(seg1);
1464 for (; vp < evp; vp++)
1465 VPP_SETPROT(vp, a->prot);
1466 if (svd2->pageswap && swresv) {
1467 ASSERT(!(svd2->flags & MAP_NORESERVE));
1468 ASSERT(swresv == seg1->s_size);
1469 vp = new_vpage;
1470 for (; vp < evp; vp++) {
1471 VPP_SETSWAPRES(vp);
1472 }
1473 }
1474 }
1475 ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1476 size = seg1->s_size;
1477 seg_free(seg1);
1478 seg2->s_size += size;
1479 seg2->s_base -= size;
1480 svd2->offset -= size;
1481 svd2->swresv += swresv;
1482 if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1483 svd2->type == MAP_SHARED && svd2->vp != NULL &&
1484 (svd2->vp->v_flag & VVMEXEC)) {
1485 ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1486 segvn_inval_trcache(svd2->vp);
1487 }
1488 return (0);
1489 }
1490
1491 /*
1492 * Duplicate all the pages in the segment. This may break COW sharing for a
1493 * given page. If the page is marked with inherit zero set, then instead of
1494 * duplicating the page, we zero the page.
1495 */
1496 static int
1497 segvn_dup_pages(struct seg *seg, struct seg *newseg)
1498 {
1499 int error;
1500 uint_t prot;
1501 page_t *pp;
1502 struct anon *ap, *newap;
1503 size_t i;
1504 caddr_t addr;
1505
1506 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1507 struct segvn_data *newsvd = (struct segvn_data *)newseg->s_data;
1508 ulong_t old_idx = svd->anon_index;
1509 ulong_t new_idx = 0;
1510
1511 i = btopr(seg->s_size);
1512 addr = seg->s_base;
1513
1514 /*
1515 * XXX break cow sharing using PAGESIZE
1516 * pages. They will be relocated into larger
1517 * pages at fault time.
1518 */
1519 while (i-- > 0) {
1520 if ((ap = anon_get_ptr(svd->amp->ahp, old_idx)) != NULL) {
1521 struct vpage *vpp;
1522
1523 vpp = &svd->vpage[seg_page(seg, addr)];
1524
1525 /*
1526 * prot need not be computed below 'cause anon_private
1527 * is going to ignore it anyway as child doesn't inherit
1528 * pagelock from parent.
1529 */
1530 prot = svd->pageprot ? VPP_PROT(vpp) : svd->prot;
1531
1532 /*
1533 * Check whether we should zero this or dup it.
1534 */
1535 if (svd->svn_inz == SEGVN_INZ_ALL ||
1536 (svd->svn_inz == SEGVN_INZ_VPP &&
1537 VPP_ISINHZERO(vpp))) {
1538 pp = anon_zero(newseg, addr, &newap,
1539 newsvd->cred);
1540 } else {
1541 page_t *anon_pl[1+1];
1542 uint_t vpprot;
1543 error = anon_getpage(&ap, &vpprot, anon_pl,
1544 PAGESIZE, seg, addr, S_READ, svd->cred);
1545 if (error != 0)
1546 return (error);
1547
1548 pp = anon_private(&newap, newseg, addr, prot,
1549 anon_pl[0], 0, newsvd->cred);
1550 }
1551 if (pp == NULL) {
1552 return (ENOMEM);
1553 }
1554 (void) anon_set_ptr(newsvd->amp->ahp, new_idx, newap,
1555 ANON_SLEEP);
1556 page_unlock(pp);
1557 }
1558 addr += PAGESIZE;
1559 old_idx++;
1560 new_idx++;
1561 }
1562
1563 return (0);
1564 }
1565
1566 static int
1567 segvn_dup(struct seg *seg, struct seg *newseg)
1568 {
1569 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1570 struct segvn_data *newsvd;
1571 pgcnt_t npages = seg_pages(seg);
1572 int error = 0;
1573 size_t len;
1574 struct anon_map *amp;
1575
1576 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1577 ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1578
1579 /*
1580 * If segment has anon reserved, reserve more for the new seg.
1581 * For a MAP_NORESERVE segment swresv will be a count of all the
1582 * allocated anon slots; thus we reserve for the child as many slots
1583 * as the parent has allocated. This semantic prevents the child or
1584 * parent from dieing during a copy-on-write fault caused by trying
1585 * to write a shared pre-existing anon page.
1586 */
1587 if ((len = svd->swresv) != 0) {
1588 if (anon_resv(svd->swresv) == 0)
1589 return (ENOMEM);
1590
1591 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1592 seg, len, 0);
1593 }
1594
1595 newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1596
1597 newseg->s_ops = &segvn_ops;
1598 newseg->s_data = (void *)newsvd;
1599 newseg->s_szc = seg->s_szc;
1600
1601 newsvd->seg = newseg;
1602 if ((newsvd->vp = svd->vp) != NULL) {
1603 VN_HOLD(svd->vp);
1604 if (svd->type == MAP_SHARED)
1605 lgrp_shm_policy_init(NULL, svd->vp);
1606 }
1607 newsvd->offset = svd->offset;
1608 newsvd->prot = svd->prot;
1609 newsvd->maxprot = svd->maxprot;
1610 newsvd->pageprot = svd->pageprot;
1611 newsvd->type = svd->type;
1612 newsvd->cred = svd->cred;
1613 crhold(newsvd->cred);
1614 newsvd->advice = svd->advice;
1615 newsvd->pageadvice = svd->pageadvice;
1616 newsvd->svn_inz = svd->svn_inz;
1617 newsvd->swresv = svd->swresv;
1618 newsvd->pageswap = svd->pageswap;
1619 newsvd->flags = svd->flags;
1620 newsvd->softlockcnt = 0;
1621 newsvd->softlockcnt_sbase = 0;
1622 newsvd->softlockcnt_send = 0;
1623 newsvd->policy_info = svd->policy_info;
1624 newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1625
1626 if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1627 /*
1628 * Not attaching to a shared anon object.
1629 */
1630 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1631 svd->tr_state == SEGVN_TR_OFF);
1632 if (svd->tr_state == SEGVN_TR_ON) {
1633 ASSERT(newsvd->vp != NULL && amp != NULL);
1634 newsvd->tr_state = SEGVN_TR_INIT;
1635 } else {
1636 newsvd->tr_state = svd->tr_state;
1637 }
1638 newsvd->amp = NULL;
1639 newsvd->anon_index = 0;
1640 } else {
1641 /* regions for now are only used on pure vnode segments */
1642 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1643 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1644 newsvd->tr_state = SEGVN_TR_OFF;
1645 if (svd->type == MAP_SHARED) {
1646 ASSERT(svd->svn_inz == SEGVN_INZ_NONE);
1647 newsvd->amp = amp;
1648 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
1649 amp->refcnt++;
1650 ANON_LOCK_EXIT(&->a_rwlock);
1651 newsvd->anon_index = svd->anon_index;
1652 } else {
1653 int reclaim = 1;
1654
1655 /*
1656 * Allocate and initialize new anon_map structure.
1657 */
1658 newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1659 ANON_SLEEP);
1660 newsvd->amp->a_szc = newseg->s_szc;
1661 newsvd->anon_index = 0;
1662 ASSERT(svd->svn_inz == SEGVN_INZ_NONE ||
1663 svd->svn_inz == SEGVN_INZ_ALL ||
1664 svd->svn_inz == SEGVN_INZ_VPP);
1665
1666 /*
1667 * We don't have to acquire the anon_map lock
1668 * for the new segment (since it belongs to an
1669 * address space that is still not associated
1670 * with any process), or the segment in the old
1671 * address space (since all threads in it
1672 * are stopped while duplicating the address space).
1673 */
1674
1675 /*
1676 * The goal of the following code is to make sure that
1677 * softlocked pages do not end up as copy on write
1678 * pages. This would cause problems where one
1679 * thread writes to a page that is COW and a different
1680 * thread in the same process has softlocked it. The
1681 * softlock lock would move away from this process
1682 * because the write would cause this process to get
1683 * a copy (without the softlock).
1684 *
1685 * The strategy here is to just break the
1686 * sharing on pages that could possibly be
1687 * softlocked.
1688 *
1689 * In addition, if any pages have been marked that they
1690 * should be inherited as zero, then we immediately go
1691 * ahead and break COW and zero them. In the case of a
1692 * softlocked page that should be inherited zero, we
1693 * break COW and just get a zero page.
1694 */
1695 retry:
1696 if (svd->softlockcnt ||
1697 svd->svn_inz != SEGVN_INZ_NONE) {
1698 /*
1699 * The softlock count might be non zero
1700 * because some pages are still stuck in the
1701 * cache for lazy reclaim. Flush the cache
1702 * now. This should drop the count to zero.
1703 * [or there is really I/O going on to these
1704 * pages]. Note, we have the writers lock so
1705 * nothing gets inserted during the flush.
1706 */
1707 if (svd->softlockcnt && reclaim == 1) {
1708 segvn_purge(seg);
1709 reclaim = 0;
1710 goto retry;
1711 }
1712
1713 error = segvn_dup_pages(seg, newseg);
1714 if (error != 0) {
1715 newsvd->vpage = NULL;
1716 goto out;
1717 }
1718 } else { /* common case */
1719 if (seg->s_szc != 0) {
1720 /*
1721 * If at least one of anon slots of a
1722 * large page exists then make sure
1723 * all anon slots of a large page
1724 * exist to avoid partial cow sharing
1725 * of a large page in the future.
1726 */
1727 anon_dup_fill_holes(amp->ahp,
1728 svd->anon_index, newsvd->amp->ahp,
1729 0, seg->s_size, seg->s_szc,
1730 svd->vp != NULL);
1731 } else {
1732 anon_dup(amp->ahp, svd->anon_index,
1733 newsvd->amp->ahp, 0, seg->s_size);
1734 }
1735
1736 hat_clrattr(seg->s_as->a_hat, seg->s_base,
1737 seg->s_size, PROT_WRITE);
1738 }
1739 }
1740 }
1741 /*
1742 * If necessary, create a vpage structure for the new segment.
1743 * Do not copy any page lock indications.
1744 */
1745 if (svd->vpage != NULL) {
1746 uint_t i;
1747 struct vpage *ovp = svd->vpage;
1748 struct vpage *nvp;
1749
1750 nvp = newsvd->vpage =
1751 kmem_alloc(vpgtob(npages), KM_SLEEP);
1752 for (i = 0; i < npages; i++) {
1753 *nvp = *ovp++;
1754 VPP_CLRPPLOCK(nvp++);
1755 }
1756 } else
1757 newsvd->vpage = NULL;
1758
1759 /* Inform the vnode of the new mapping */
1760 if (newsvd->vp != NULL) {
1761 error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1762 newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1763 newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1764 }
1765 out:
1766 if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1767 ASSERT(newsvd->amp == NULL);
1768 ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1769 newsvd->rcookie = svd->rcookie;
1770 hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1771 }
1772 return (error);
1773 }
1774
1775
1776 /*
1777 * callback function to invoke free_vp_pages() for only those pages actually
1778 * processed by the HAT when a shared region is destroyed.
1779 */
1780 extern int free_pages;
1781
1782 static void
1783 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1784 size_t r_size, void *r_obj, u_offset_t r_objoff)
1785 {
1786 u_offset_t off;
1787 size_t len;
1788 vnode_t *vp = (vnode_t *)r_obj;
1789
1790 ASSERT(eaddr > saddr);
1791 ASSERT(saddr >= r_saddr);
1792 ASSERT(saddr < r_saddr + r_size);
1793 ASSERT(eaddr > r_saddr);
1794 ASSERT(eaddr <= r_saddr + r_size);
1795 ASSERT(vp != NULL);
1796
1797 if (!free_pages) {
1798 return;
1799 }
1800
1801 len = eaddr - saddr;
1802 off = (saddr - r_saddr) + r_objoff;
1803 free_vp_pages(vp, off, len);
1804 }
1805
1806 /*
1807 * callback function used by segvn_unmap to invoke free_vp_pages() for only
1808 * those pages actually processed by the HAT
1809 */
1810 static void
1811 segvn_hat_unload_callback(hat_callback_t *cb)
1812 {
1813 struct seg *seg = cb->hcb_data;
1814 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1815 size_t len;
1816 u_offset_t off;
1817
1818 ASSERT(svd->vp != NULL);
1819 ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1820 ASSERT(cb->hcb_start_addr >= seg->s_base);
1821
1822 len = cb->hcb_end_addr - cb->hcb_start_addr;
1823 off = cb->hcb_start_addr - seg->s_base;
1824 free_vp_pages(svd->vp, svd->offset + off, len);
1825 }
1826
1827 /*
1828 * This function determines the number of bytes of swap reserved by
1829 * a segment for which per-page accounting is present. It is used to
1830 * calculate the correct value of a segvn_data's swresv.
1831 */
1832 static size_t
1833 segvn_count_swap_by_vpages(struct seg *seg)
1834 {
1835 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1836 struct vpage *vp, *evp;
1837 size_t nswappages = 0;
1838
1839 ASSERT(svd->pageswap);
1840 ASSERT(svd->vpage != NULL);
1841
1842 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1843
1844 for (vp = svd->vpage; vp < evp; vp++) {
1845 if (VPP_ISSWAPRES(vp))
1846 nswappages++;
1847 }
1848
1849 return (nswappages << PAGESHIFT);
1850 }
1851
1852 static int
1853 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1854 {
1855 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1856 struct segvn_data *nsvd;
1857 struct seg *nseg;
1858 struct anon_map *amp;
1859 pgcnt_t opages; /* old segment size in pages */
1860 pgcnt_t npages; /* new segment size in pages */
1861 pgcnt_t dpages; /* pages being deleted (unmapped) */
1862 hat_callback_t callback; /* used for free_vp_pages() */
1863 hat_callback_t *cbp = NULL;
1864 caddr_t nbase;
1865 size_t nsize;
1866 size_t oswresv;
1867 int reclaim = 1;
1868
1869 /*
1870 * We don't need any segment level locks for "segvn" data
1871 * since the address space is "write" locked.
1872 */
1873 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1874
1875 /*
1876 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1877 * softlockcnt is protected from change by the as write lock.
1878 */
1879 retry:
1880 if (svd->softlockcnt > 0) {
1881 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1882
1883 /*
1884 * If this is shared segment non 0 softlockcnt
1885 * means locked pages are still in use.
1886 */
1887 if (svd->type == MAP_SHARED) {
1888 return (EAGAIN);
1889 }
1890
1891 /*
1892 * since we do have the writers lock nobody can fill
1893 * the cache during the purge. The flush either succeeds
1894 * or we still have pending I/Os.
1895 */
1896 if (reclaim == 1) {
1897 segvn_purge(seg);
1898 reclaim = 0;
1899 goto retry;
1900 }
1901 return (EAGAIN);
1902 }
1903
1904 /*
1905 * Check for bad sizes
1906 */
1907 if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1908 (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1909 panic("segvn_unmap");
1910 /*NOTREACHED*/
1911 }
1912
1913 if (seg->s_szc != 0) {
1914 size_t pgsz = page_get_pagesize(seg->s_szc);
1915 int err;
1916 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1917 ASSERT(seg->s_base != addr || seg->s_size != len);
1918 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1919 ASSERT(svd->amp == NULL);
1920 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1921 hat_leave_region(seg->s_as->a_hat,
1922 svd->rcookie, HAT_REGION_TEXT);
1923 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1924 /*
1925 * could pass a flag to segvn_demote_range()
1926 * below to tell it not to do any unloads but
1927 * this case is rare enough to not bother for
1928 * now.
1929 */
1930 } else if (svd->tr_state == SEGVN_TR_INIT) {
1931 svd->tr_state = SEGVN_TR_OFF;
1932 } else if (svd->tr_state == SEGVN_TR_ON) {
1933 ASSERT(svd->amp != NULL);
1934 segvn_textunrepl(seg, 1);
1935 ASSERT(svd->amp == NULL);
1936 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1937 }
1938 VM_STAT_ADD(segvnvmstats.demoterange[0]);
1939 err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1940 if (err == 0) {
1941 return (IE_RETRY);
1942 }
1943 return (err);
1944 }
1945 }
1946
1947 /* Inform the vnode of the unmapping. */
1948 if (svd->vp) {
1949 int error;
1950
1951 error = VOP_DELMAP(svd->vp,
1952 (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1953 seg->s_as, addr, len, svd->prot, svd->maxprot,
1954 svd->type, svd->cred, NULL);
1955
1956 if (error == EAGAIN)
1957 return (error);
1958 }
1959
1960 /*
1961 * Remove any page locks set through this mapping.
1962 * If text replication is not off no page locks could have been
1963 * established via this mapping.
1964 */
1965 if (svd->tr_state == SEGVN_TR_OFF) {
1966 (void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1967 }
1968
1969 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1970 ASSERT(svd->amp == NULL);
1971 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1972 ASSERT(svd->type == MAP_PRIVATE);
1973 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1974 HAT_REGION_TEXT);
1975 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1976 } else if (svd->tr_state == SEGVN_TR_ON) {
1977 ASSERT(svd->amp != NULL);
1978 ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1979 segvn_textunrepl(seg, 1);
1980 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1981 } else {
1982 if (svd->tr_state != SEGVN_TR_OFF) {
1983 ASSERT(svd->tr_state == SEGVN_TR_INIT);
1984 svd->tr_state = SEGVN_TR_OFF;
1985 }
1986 /*
1987 * Unload any hardware translations in the range to be taken
1988 * out. Use a callback to invoke free_vp_pages() effectively.
1989 */
1990 if (svd->vp != NULL && free_pages != 0) {
1991 callback.hcb_data = seg;
1992 callback.hcb_function = segvn_hat_unload_callback;
1993 cbp = &callback;
1994 }
1995 hat_unload_callback(seg->s_as->a_hat, addr, len,
1996 HAT_UNLOAD_UNMAP, cbp);
1997
1998 if (svd->type == MAP_SHARED && svd->vp != NULL &&
1999 (svd->vp->v_flag & VVMEXEC) &&
2000 ((svd->prot & PROT_WRITE) || svd->pageprot)) {
2001 segvn_inval_trcache(svd->vp);
2002 }
2003 }
2004
2005 /*
2006 * Check for entire segment
2007 */
2008 if (addr == seg->s_base && len == seg->s_size) {
2009 seg_free(seg);
2010 return (0);
2011 }
2012
2013 opages = seg_pages(seg);
2014 dpages = btop(len);
2015 npages = opages - dpages;
2016 amp = svd->amp;
2017 ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
2018
2019 /*
2020 * Check for beginning of segment
2021 */
2022 if (addr == seg->s_base) {
2023 if (svd->vpage != NULL) {
2024 size_t nbytes;
2025 struct vpage *ovpage;
2026
2027 ovpage = svd->vpage; /* keep pointer to vpage */
2028
2029 nbytes = vpgtob(npages);
2030 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2031 bcopy(&ovpage[dpages], svd->vpage, nbytes);
2032
2033 /* free up old vpage */
2034 kmem_free(ovpage, vpgtob(opages));
2035 }
2036 if (amp != NULL) {
2037 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
2038 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2039 /*
2040 * Shared anon map is no longer in use. Before
2041 * freeing its pages purge all entries from
2042 * pcache that belong to this amp.
2043 */
2044 if (svd->type == MAP_SHARED) {
2045 ASSERT(amp->refcnt == 1);
2046 ASSERT(svd->softlockcnt == 0);
2047 anonmap_purge(amp);
2048 }
2049 /*
2050 * Free up now unused parts of anon_map array.
2051 */
2052 if (amp->a_szc == seg->s_szc) {
2053 if (seg->s_szc != 0) {
2054 anon_free_pages(amp->ahp,
2055 svd->anon_index, len,
2056 seg->s_szc);
2057 } else {
2058 anon_free(amp->ahp,
2059 svd->anon_index,
2060 len);
2061 }
2062 } else {
2063 ASSERT(svd->type == MAP_SHARED);
2064 ASSERT(amp->a_szc > seg->s_szc);
2065 anon_shmap_free_pages(amp,
2066 svd->anon_index, len);
2067 }
2068
2069 /*
2070 * Unreserve swap space for the
2071 * unmapped chunk of this segment in
2072 * case it's MAP_SHARED
2073 */
2074 if (svd->type == MAP_SHARED) {
2075 anon_unresv_zone(len,
2076 seg->s_as->a_proc->p_zone);
2077 amp->swresv -= len;
2078 }
2079 }
2080 ANON_LOCK_EXIT(&->a_rwlock);
2081 svd->anon_index += dpages;
2082 }
2083 if (svd->vp != NULL)
2084 svd->offset += len;
2085
2086 seg->s_base += len;
2087 seg->s_size -= len;
2088
2089 if (svd->swresv) {
2090 if (svd->flags & MAP_NORESERVE) {
2091 ASSERT(amp);
2092 oswresv = svd->swresv;
2093
2094 svd->swresv = ptob(anon_pages(amp->ahp,
2095 svd->anon_index, npages));
2096 anon_unresv_zone(oswresv - svd->swresv,
2097 seg->s_as->a_proc->p_zone);
2098 if (SEG_IS_PARTIAL_RESV(seg))
2099 seg->s_as->a_resvsize -= oswresv -
2100 svd->swresv;
2101 } else {
2102 size_t unlen;
2103
2104 if (svd->pageswap) {
2105 oswresv = svd->swresv;
2106 svd->swresv =
2107 segvn_count_swap_by_vpages(seg);
2108 ASSERT(oswresv >= svd->swresv);
2109 unlen = oswresv - svd->swresv;
2110 } else {
2111 svd->swresv -= len;
2112 ASSERT(svd->swresv == seg->s_size);
2113 unlen = len;
2114 }
2115 anon_unresv_zone(unlen,
2116 seg->s_as->a_proc->p_zone);
2117 }
2118 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2119 seg, len, 0);
2120 }
2121
2122 return (0);
2123 }
2124
2125 /*
2126 * Check for end of segment
2127 */
2128 if (addr + len == seg->s_base + seg->s_size) {
2129 if (svd->vpage != NULL) {
2130 size_t nbytes;
2131 struct vpage *ovpage;
2132
2133 ovpage = svd->vpage; /* keep pointer to vpage */
2134
2135 nbytes = vpgtob(npages);
2136 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2137 bcopy(ovpage, svd->vpage, nbytes);
2138
2139 /* free up old vpage */
2140 kmem_free(ovpage, vpgtob(opages));
2141
2142 }
2143 if (amp != NULL) {
2144 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
2145 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2146 /*
2147 * Free up now unused parts of anon_map array.
2148 */
2149 ulong_t an_idx = svd->anon_index + npages;
2150
2151 /*
2152 * Shared anon map is no longer in use. Before
2153 * freeing its pages purge all entries from
2154 * pcache that belong to this amp.
2155 */
2156 if (svd->type == MAP_SHARED) {
2157 ASSERT(amp->refcnt == 1);
2158 ASSERT(svd->softlockcnt == 0);
2159 anonmap_purge(amp);
2160 }
2161
2162 if (amp->a_szc == seg->s_szc) {
2163 if (seg->s_szc != 0) {
2164 anon_free_pages(amp->ahp,
2165 an_idx, len,
2166 seg->s_szc);
2167 } else {
2168 anon_free(amp->ahp, an_idx,
2169 len);
2170 }
2171 } else {
2172 ASSERT(svd->type == MAP_SHARED);
2173 ASSERT(amp->a_szc > seg->s_szc);
2174 anon_shmap_free_pages(amp,
2175 an_idx, len);
2176 }
2177
2178 /*
2179 * Unreserve swap space for the
2180 * unmapped chunk of this segment in
2181 * case it's MAP_SHARED
2182 */
2183 if (svd->type == MAP_SHARED) {
2184 anon_unresv_zone(len,
2185 seg->s_as->a_proc->p_zone);
2186 amp->swresv -= len;
2187 }
2188 }
2189 ANON_LOCK_EXIT(&->a_rwlock);
2190 }
2191
2192 seg->s_size -= len;
2193
2194 if (svd->swresv) {
2195 if (svd->flags & MAP_NORESERVE) {
2196 ASSERT(amp);
2197 oswresv = svd->swresv;
2198 svd->swresv = ptob(anon_pages(amp->ahp,
2199 svd->anon_index, npages));
2200 anon_unresv_zone(oswresv - svd->swresv,
2201 seg->s_as->a_proc->p_zone);
2202 if (SEG_IS_PARTIAL_RESV(seg))
2203 seg->s_as->a_resvsize -= oswresv -
2204 svd->swresv;
2205 } else {
2206 size_t unlen;
2207
2208 if (svd->pageswap) {
2209 oswresv = svd->swresv;
2210 svd->swresv =
2211 segvn_count_swap_by_vpages(seg);
2212 ASSERT(oswresv >= svd->swresv);
2213 unlen = oswresv - svd->swresv;
2214 } else {
2215 svd->swresv -= len;
2216 ASSERT(svd->swresv == seg->s_size);
2217 unlen = len;
2218 }
2219 anon_unresv_zone(unlen,
2220 seg->s_as->a_proc->p_zone);
2221 }
2222 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2223 "anon proc:%p %lu %u", seg, len, 0);
2224 }
2225
2226 return (0);
2227 }
2228
2229 /*
2230 * The section to go is in the middle of the segment,
2231 * have to make it into two segments. nseg is made for
2232 * the high end while seg is cut down at the low end.
2233 */
2234 nbase = addr + len; /* new seg base */
2235 nsize = (seg->s_base + seg->s_size) - nbase; /* new seg size */
2236 seg->s_size = addr - seg->s_base; /* shrink old seg */
2237 nseg = seg_alloc(seg->s_as, nbase, nsize);
2238 if (nseg == NULL) {
2239 panic("segvn_unmap seg_alloc");
2240 /*NOTREACHED*/
2241 }
2242 nseg->s_ops = seg->s_ops;
2243 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2244 nseg->s_data = (void *)nsvd;
2245 nseg->s_szc = seg->s_szc;
2246 *nsvd = *svd;
2247 nsvd->seg = nseg;
2248 nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2249 nsvd->swresv = 0;
2250 nsvd->softlockcnt = 0;
2251 nsvd->softlockcnt_sbase = 0;
2252 nsvd->softlockcnt_send = 0;
2253 nsvd->svn_inz = svd->svn_inz;
2254 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2255
2256 if (svd->vp != NULL) {
2257 VN_HOLD(nsvd->vp);
2258 if (nsvd->type == MAP_SHARED)
2259 lgrp_shm_policy_init(NULL, nsvd->vp);
2260 }
2261 crhold(svd->cred);
2262
2263 if (svd->vpage == NULL) {
2264 nsvd->vpage = NULL;
2265 } else {
2266 /* need to split vpage into two arrays */
2267 size_t nbytes;
2268 struct vpage *ovpage;
2269
2270 ovpage = svd->vpage; /* keep pointer to vpage */
2271
2272 npages = seg_pages(seg); /* seg has shrunk */
2273 nbytes = vpgtob(npages);
2274 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2275
2276 bcopy(ovpage, svd->vpage, nbytes);
2277
2278 npages = seg_pages(nseg);
2279 nbytes = vpgtob(npages);
2280 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2281
2282 bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2283
2284 /* free up old vpage */
2285 kmem_free(ovpage, vpgtob(opages));
2286 }
2287
2288 if (amp == NULL) {
2289 nsvd->amp = NULL;
2290 nsvd->anon_index = 0;
2291 } else {
2292 /*
2293 * Need to create a new anon map for the new segment.
2294 * We'll also allocate a new smaller array for the old
2295 * smaller segment to save space.
2296 */
2297 opages = btop((uintptr_t)(addr - seg->s_base));
2298 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
2299 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2300 /*
2301 * Free up now unused parts of anon_map array.
2302 */
2303 ulong_t an_idx = svd->anon_index + opages;
2304
2305 /*
2306 * Shared anon map is no longer in use. Before
2307 * freeing its pages purge all entries from
2308 * pcache that belong to this amp.
2309 */
2310 if (svd->type == MAP_SHARED) {
2311 ASSERT(amp->refcnt == 1);
2312 ASSERT(svd->softlockcnt == 0);
2313 anonmap_purge(amp);
2314 }
2315
2316 if (amp->a_szc == seg->s_szc) {
2317 if (seg->s_szc != 0) {
2318 anon_free_pages(amp->ahp, an_idx, len,
2319 seg->s_szc);
2320 } else {
2321 anon_free(amp->ahp, an_idx,
2322 len);
2323 }
2324 } else {
2325 ASSERT(svd->type == MAP_SHARED);
2326 ASSERT(amp->a_szc > seg->s_szc);
2327 anon_shmap_free_pages(amp, an_idx, len);
2328 }
2329
2330 /*
2331 * Unreserve swap space for the
2332 * unmapped chunk of this segment in
2333 * case it's MAP_SHARED
2334 */
2335 if (svd->type == MAP_SHARED) {
2336 anon_unresv_zone(len,
2337 seg->s_as->a_proc->p_zone);
2338 amp->swresv -= len;
2339 }
2340 }
2341 nsvd->anon_index = svd->anon_index +
2342 btop((uintptr_t)(nseg->s_base - seg->s_base));
2343 if (svd->type == MAP_SHARED) {
2344 amp->refcnt++;
2345 nsvd->amp = amp;
2346 } else {
2347 struct anon_map *namp;
2348 struct anon_hdr *nahp;
2349
2350 ASSERT(svd->type == MAP_PRIVATE);
2351 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2352 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2353 namp->a_szc = seg->s_szc;
2354 (void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2355 0, btop(seg->s_size), ANON_SLEEP);
2356 (void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2357 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2358 anon_release(amp->ahp, btop(amp->size));
2359 svd->anon_index = 0;
2360 nsvd->anon_index = 0;
2361 amp->ahp = nahp;
2362 amp->size = seg->s_size;
2363 nsvd->amp = namp;
2364 }
2365 ANON_LOCK_EXIT(&->a_rwlock);
2366 }
2367 if (svd->swresv) {
2368 if (svd->flags & MAP_NORESERVE) {
2369 ASSERT(amp);
2370 oswresv = svd->swresv;
2371 svd->swresv = ptob(anon_pages(amp->ahp,
2372 svd->anon_index, btop(seg->s_size)));
2373 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2374 nsvd->anon_index, btop(nseg->s_size)));
2375 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2376 anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2377 seg->s_as->a_proc->p_zone);
2378 if (SEG_IS_PARTIAL_RESV(seg))
2379 seg->s_as->a_resvsize -= oswresv -
2380 (svd->swresv + nsvd->swresv);
2381 } else {
2382 size_t unlen;
2383
2384 if (svd->pageswap) {
2385 oswresv = svd->swresv;
2386 svd->swresv = segvn_count_swap_by_vpages(seg);
2387 nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2388 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2389 unlen = oswresv - (svd->swresv + nsvd->swresv);
2390 } else {
2391 if (seg->s_size + nseg->s_size + len !=
2392 svd->swresv) {
2393 panic("segvn_unmap: cannot split "
2394 "swap reservation");
2395 /*NOTREACHED*/
2396 }
2397 svd->swresv = seg->s_size;
2398 nsvd->swresv = nseg->s_size;
2399 unlen = len;
2400 }
2401 anon_unresv_zone(unlen,
2402 seg->s_as->a_proc->p_zone);
2403 }
2404 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2405 seg, len, 0);
2406 }
2407
2408 return (0); /* I'm glad that's all over with! */
2409 }
2410
2411 static void
2412 segvn_free(struct seg *seg)
2413 {
2414 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2415 pgcnt_t npages = seg_pages(seg);
2416 struct anon_map *amp;
2417 size_t len;
2418
2419 /*
2420 * We don't need any segment level locks for "segvn" data
2421 * since the address space is "write" locked.
2422 */
2423 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
2424 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2425
2426 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2427
2428 /*
2429 * Be sure to unlock pages. XXX Why do things get free'ed instead
2430 * of unmapped? XXX
2431 */
2432 (void) segvn_lockop(seg, seg->s_base, seg->s_size,
2433 0, MC_UNLOCK, NULL, 0);
2434
2435 /*
2436 * Deallocate the vpage and anon pointers if necessary and possible.
2437 */
2438 if (svd->vpage != NULL) {
2439 kmem_free(svd->vpage, vpgtob(npages));
2440 svd->vpage = NULL;
2441 }
2442 if ((amp = svd->amp) != NULL) {
2443 /*
2444 * If there are no more references to this anon_map
2445 * structure, then deallocate the structure after freeing
2446 * up all the anon slot pointers that we can.
2447 */
2448 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
2449 ASSERT(amp->a_szc >= seg->s_szc);
2450 if (--amp->refcnt == 0) {
2451 if (svd->type == MAP_PRIVATE) {
2452 /*
2453 * Private - we only need to anon_free
2454 * the part that this segment refers to.
2455 */
2456 if (seg->s_szc != 0) {
2457 anon_free_pages(amp->ahp,
2458 svd->anon_index, seg->s_size,
2459 seg->s_szc);
2460 } else {
2461 anon_free(amp->ahp, svd->anon_index,
2462 seg->s_size);
2463 }
2464 } else {
2465
2466 /*
2467 * Shared anon map is no longer in use. Before
2468 * freeing its pages purge all entries from
2469 * pcache that belong to this amp.
2470 */
2471 ASSERT(svd->softlockcnt == 0);
2472 anonmap_purge(amp);
2473
2474 /*
2475 * Shared - anon_free the entire
2476 * anon_map's worth of stuff and
2477 * release any swap reservation.
2478 */
2479 if (amp->a_szc != 0) {
2480 anon_shmap_free_pages(amp, 0,
2481 amp->size);
2482 } else {
2483 anon_free(amp->ahp, 0, amp->size);
2484 }
2485 if ((len = amp->swresv) != 0) {
2486 anon_unresv_zone(len,
2487 seg->s_as->a_proc->p_zone);
2488 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2489 "anon proc:%p %lu %u", seg, len, 0);
2490 }
2491 }
2492 svd->amp = NULL;
2493 ANON_LOCK_EXIT(&->a_rwlock);
2494 anonmap_free(amp);
2495 } else if (svd->type == MAP_PRIVATE) {
2496 /*
2497 * We had a private mapping which still has
2498 * a held anon_map so just free up all the
2499 * anon slot pointers that we were using.
2500 */
2501 if (seg->s_szc != 0) {
2502 anon_free_pages(amp->ahp, svd->anon_index,
2503 seg->s_size, seg->s_szc);
2504 } else {
2505 anon_free(amp->ahp, svd->anon_index,
2506 seg->s_size);
2507 }
2508 ANON_LOCK_EXIT(&->a_rwlock);
2509 } else {
2510 ANON_LOCK_EXIT(&->a_rwlock);
2511 }
2512 }
2513
2514 /*
2515 * Release swap reservation.
2516 */
2517 if ((len = svd->swresv) != 0) {
2518 anon_unresv_zone(svd->swresv,
2519 seg->s_as->a_proc->p_zone);
2520 TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2521 seg, len, 0);
2522 if (SEG_IS_PARTIAL_RESV(seg))
2523 seg->s_as->a_resvsize -= svd->swresv;
2524 svd->swresv = 0;
2525 }
2526 /*
2527 * Release claim on vnode, credentials, and finally free the
2528 * private data.
2529 */
2530 if (svd->vp != NULL) {
2531 if (svd->type == MAP_SHARED)
2532 lgrp_shm_policy_fini(NULL, svd->vp);
2533 VN_RELE(svd->vp);
2534 svd->vp = NULL;
2535 }
2536 crfree(svd->cred);
2537 svd->pageprot = 0;
2538 svd->pageadvice = 0;
2539 svd->pageswap = 0;
2540 svd->cred = NULL;
2541
2542 /*
2543 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2544 * still working with this segment without holding as lock (in case
2545 * it's called by pcache async thread).
2546 */
2547 ASSERT(svd->softlockcnt == 0);
2548 mutex_enter(&svd->segfree_syncmtx);
2549 mutex_exit(&svd->segfree_syncmtx);
2550
2551 seg->s_data = NULL;
2552 kmem_cache_free(segvn_cache, svd);
2553 }
2554
2555 /*
2556 * Do a F_SOFTUNLOCK call over the range requested. The range must have
2557 * already been F_SOFTLOCK'ed.
2558 * Caller must always match addr and len of a softunlock with a previous
2559 * softlock with exactly the same addr and len.
2560 */
2561 static void
2562 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2563 {
2564 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2565 page_t *pp;
2566 caddr_t adr;
2567 struct vnode *vp;
2568 u_offset_t offset;
2569 ulong_t anon_index;
2570 struct anon_map *amp;
2571 struct anon *ap = NULL;
2572
2573 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2574 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2575
2576 if ((amp = svd->amp) != NULL)
2577 anon_index = svd->anon_index + seg_page(seg, addr);
2578
2579 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2580 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2581 hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2582 } else {
2583 hat_unlock(seg->s_as->a_hat, addr, len);
2584 }
2585 for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2586 if (amp != NULL) {
2587 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
2588 if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2589 != NULL) {
2590 swap_xlate(ap, &vp, &offset);
2591 } else {
2592 vp = svd->vp;
2593 offset = svd->offset +
2594 (uintptr_t)(adr - seg->s_base);
2595 }
2596 ANON_LOCK_EXIT(&->a_rwlock);
2597 } else {
2598 vp = svd->vp;
2599 offset = svd->offset +
2600 (uintptr_t)(adr - seg->s_base);
2601 }
2602
2603 /*
2604 * Use page_find() instead of page_lookup() to
2605 * find the page since we know that it is locked.
2606 */
2607 pp = page_find(vp, offset);
2608 if (pp == NULL) {
2609 panic(
2610 "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2611 (void *)adr, (void *)ap, (void *)vp, offset);
2612 /*NOTREACHED*/
2613 }
2614
2615 if (rw == S_WRITE) {
2616 hat_setrefmod(pp);
2617 if (seg->s_as->a_vbits)
2618 hat_setstat(seg->s_as, adr, PAGESIZE,
2619 P_REF | P_MOD);
2620 } else if (rw != S_OTHER) {
2621 hat_setref(pp);
2622 if (seg->s_as->a_vbits)
2623 hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2624 }
2625 TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2626 "segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2627 page_unlock(pp);
2628 }
2629 ASSERT(svd->softlockcnt >= btop(len));
2630 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2631 /*
2632 * All SOFTLOCKS are gone. Wakeup any waiting
2633 * unmappers so they can try again to unmap.
2634 * Check for waiters first without the mutex
2635 * held so we don't always grab the mutex on
2636 * softunlocks.
2637 */
2638 if (AS_ISUNMAPWAIT(seg->s_as)) {
2639 mutex_enter(&seg->s_as->a_contents);
2640 if (AS_ISUNMAPWAIT(seg->s_as)) {
2641 AS_CLRUNMAPWAIT(seg->s_as);
2642 cv_broadcast(&seg->s_as->a_cv);
2643 }
2644 mutex_exit(&seg->s_as->a_contents);
2645 }
2646 }
2647 }
2648
2649 #define PAGE_HANDLED ((page_t *)-1)
2650
2651 /*
2652 * Release all the pages in the NULL terminated ppp list
2653 * which haven't already been converted to PAGE_HANDLED.
2654 */
2655 static void
2656 segvn_pagelist_rele(page_t **ppp)
2657 {
2658 for (; *ppp != NULL; ppp++) {
2659 if (*ppp != PAGE_HANDLED)
2660 page_unlock(*ppp);
2661 }
2662 }
2663
2664 static int stealcow = 1;
2665
2666 /*
2667 * Workaround for viking chip bug. See bug id 1220902.
2668 * To fix this down in pagefault() would require importing so
2669 * much as and segvn code as to be unmaintainable.
2670 */
2671 int enable_mbit_wa = 0;
2672
2673 /*
2674 * Handles all the dirty work of getting the right
2675 * anonymous pages and loading up the translations.
2676 * This routine is called only from segvn_fault()
2677 * when looping over the range of addresses requested.
2678 *
2679 * The basic algorithm here is:
2680 * If this is an anon_zero case
2681 * Call anon_zero to allocate page
2682 * Load up translation
2683 * Return
2684 * endif
2685 * If this is an anon page
2686 * Use anon_getpage to get the page
2687 * else
2688 * Find page in pl[] list passed in
2689 * endif
2690 * If not a cow
2691 * Load up the translation to the page
2692 * return
2693 * endif
2694 * Call anon_private to handle cow
2695 * Load up (writable) translation to new page
2696 */
2697 static faultcode_t
2698 segvn_faultpage(
2699 struct hat *hat, /* the hat to use for mapping */
2700 struct seg *seg, /* seg_vn of interest */
2701 caddr_t addr, /* address in as */
2702 u_offset_t off, /* offset in vp */
2703 struct vpage *vpage, /* pointer to vpage for vp, off */
2704 page_t *pl[], /* object source page pointer */
2705 uint_t vpprot, /* access allowed to object pages */
2706 enum fault_type type, /* type of fault */
2707 enum seg_rw rw, /* type of access at fault */
2708 int brkcow) /* we may need to break cow */
2709 {
2710 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2711 page_t *pp, **ppp;
2712 uint_t pageflags = 0;
2713 page_t *anon_pl[1 + 1];
2714 page_t *opp = NULL; /* original page */
2715 uint_t prot;
2716 int err;
2717 int cow;
2718 int claim;
2719 int steal = 0;
2720 ulong_t anon_index;
2721 struct anon *ap, *oldap;
2722 struct anon_map *amp;
2723 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2724 int anon_lock = 0;
2725 anon_sync_obj_t cookie;
2726
2727 if (svd->flags & MAP_TEXT) {
2728 hat_flag |= HAT_LOAD_TEXT;
2729 }
2730
2731 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2732 ASSERT(seg->s_szc == 0);
2733 ASSERT(svd->tr_state != SEGVN_TR_INIT);
2734
2735 /*
2736 * Initialize protection value for this page.
2737 * If we have per page protection values check it now.
2738 */
2739 if (svd->pageprot) {
2740 uint_t protchk;
2741
2742 switch (rw) {
2743 case S_READ:
2744 protchk = PROT_READ;
2745 break;
2746 case S_WRITE:
2747 protchk = PROT_WRITE;
2748 break;
2749 case S_EXEC:
2750 protchk = PROT_EXEC;
2751 break;
2752 case S_OTHER:
2753 default:
2754 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2755 break;
2756 }
2757
2758 prot = VPP_PROT(vpage);
2759 if ((prot & protchk) == 0)
2760 return (FC_PROT); /* illegal access type */
2761 } else {
2762 prot = svd->prot;
2763 }
2764
2765 if (type == F_SOFTLOCK) {
2766 atomic_inc_ulong((ulong_t *)&svd->softlockcnt);
2767 }
2768
2769 /*
2770 * Always acquire the anon array lock to prevent 2 threads from
2771 * allocating separate anon slots for the same "addr".
2772 */
2773
2774 if ((amp = svd->amp) != NULL) {
2775 ASSERT(RW_READ_HELD(&->a_rwlock));
2776 anon_index = svd->anon_index + seg_page(seg, addr);
2777 anon_array_enter(amp, anon_index, &cookie);
2778 anon_lock = 1;
2779 }
2780
2781 if (svd->vp == NULL && amp != NULL) {
2782 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2783 /*
2784 * Allocate a (normally) writable anonymous page of
2785 * zeroes. If no advance reservations, reserve now.
2786 */
2787 if (svd->flags & MAP_NORESERVE) {
2788 if (anon_resv_zone(ptob(1),
2789 seg->s_as->a_proc->p_zone)) {
2790 atomic_add_long(&svd->swresv, ptob(1));
2791 atomic_add_long(&seg->s_as->a_resvsize,
2792 ptob(1));
2793 } else {
2794 err = ENOMEM;
2795 goto out;
2796 }
2797 }
2798 if ((pp = anon_zero(seg, addr, &ap,
2799 svd->cred)) == NULL) {
2800 err = ENOMEM;
2801 goto out; /* out of swap space */
2802 }
2803 /*
2804 * Re-acquire the anon_map lock and
2805 * initialize the anon array entry.
2806 */
2807 (void) anon_set_ptr(amp->ahp, anon_index, ap,
2808 ANON_SLEEP);
2809
2810 ASSERT(pp->p_szc == 0);
2811
2812 /*
2813 * Handle pages that have been marked for migration
2814 */
2815 if (lgrp_optimizations())
2816 page_migrate(seg, addr, &pp, 1);
2817
2818 if (enable_mbit_wa) {
2819 if (rw == S_WRITE)
2820 hat_setmod(pp);
2821 else if (!hat_ismod(pp))
2822 prot &= ~PROT_WRITE;
2823 }
2824 /*
2825 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2826 * with MC_LOCKAS, MCL_FUTURE) and this is a
2827 * MAP_NORESERVE segment, we may need to
2828 * permanently lock the page as it is being faulted
2829 * for the first time. The following text applies
2830 * only to MAP_NORESERVE segments:
2831 *
2832 * As per memcntl(2), if this segment was created
2833 * after MCL_FUTURE was applied (a "future"
2834 * segment), its pages must be locked. If this
2835 * segment existed at MCL_FUTURE application (a
2836 * "past" segment), the interface is unclear.
2837 *
2838 * We decide to lock only if vpage is present:
2839 *
2840 * - "future" segments will have a vpage array (see
2841 * as_map), and so will be locked as required
2842 *
2843 * - "past" segments may not have a vpage array,
2844 * depending on whether events (such as
2845 * mprotect) have occurred. Locking if vpage
2846 * exists will preserve legacy behavior. Not
2847 * locking if vpage is absent, will not break
2848 * the interface or legacy behavior. Note that
2849 * allocating vpage here if it's absent requires
2850 * upgrading the segvn reader lock, the cost of
2851 * which does not seem worthwhile.
2852 *
2853 * Usually testing and setting VPP_ISPPLOCK and
2854 * VPP_SETPPLOCK requires holding the segvn lock as
2855 * writer, but in this case all readers are
2856 * serializing on the anon array lock.
2857 */
2858 if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2859 (svd->flags & MAP_NORESERVE) &&
2860 !VPP_ISPPLOCK(vpage)) {
2861 proc_t *p = seg->s_as->a_proc;
2862 ASSERT(svd->type == MAP_PRIVATE);
2863 mutex_enter(&p->p_lock);
2864 if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2865 1) == 0) {
2866 claim = VPP_PROT(vpage) & PROT_WRITE;
2867 if (page_pp_lock(pp, claim, 0)) {
2868 VPP_SETPPLOCK(vpage);
2869 } else {
2870 rctl_decr_locked_mem(p, NULL,
2871 PAGESIZE, 1);
2872 }
2873 }
2874 mutex_exit(&p->p_lock);
2875 }
2876
2877 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2878 hat_memload(hat, addr, pp, prot, hat_flag);
2879
2880 if (!(hat_flag & HAT_LOAD_LOCK))
2881 page_unlock(pp);
2882
2883 anon_array_exit(&cookie);
2884 return (0);
2885 }
2886 }
2887
2888 /*
2889 * Obtain the page structure via anon_getpage() if it is
2890 * a private copy of an object (the result of a previous
2891 * copy-on-write).
2892 */
2893 if (amp != NULL) {
2894 if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2895 err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2896 seg, addr, rw, svd->cred);
2897 if (err)
2898 goto out;
2899
2900 if (svd->type == MAP_SHARED) {
2901 /*
2902 * If this is a shared mapping to an
2903 * anon_map, then ignore the write
2904 * permissions returned by anon_getpage().
2905 * They apply to the private mappings
2906 * of this anon_map.
2907 */
2908 vpprot |= PROT_WRITE;
2909 }
2910 opp = anon_pl[0];
2911 }
2912 }
2913
2914 /*
2915 * Search the pl[] list passed in if it is from the
2916 * original object (i.e., not a private copy).
2917 */
2918 if (opp == NULL) {
2919 /*
2920 * Find original page. We must be bringing it in
2921 * from the list in pl[].
2922 */
2923 for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2924 if (opp == PAGE_HANDLED)
2925 continue;
2926 ASSERT(opp->p_vnode == svd->vp); /* XXX */
2927 if (opp->p_offset == off)
2928 break;
2929 }
2930 if (opp == NULL) {
2931 panic("segvn_faultpage not found");
2932 /*NOTREACHED*/
2933 }
2934 *ppp = PAGE_HANDLED;
2935
2936 }
2937
2938 ASSERT(PAGE_LOCKED(opp));
2939
2940 TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2941 "segvn_fault:pp %p vp %p offset %llx", opp, NULL, 0);
2942
2943 /*
2944 * The fault is treated as a copy-on-write fault if a
2945 * write occurs on a private segment and the object
2946 * page (i.e., mapping) is write protected. We assume
2947 * that fatal protection checks have already been made.
2948 */
2949
2950 if (brkcow) {
2951 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2952 cow = !(vpprot & PROT_WRITE);
2953 } else if (svd->tr_state == SEGVN_TR_ON) {
2954 /*
2955 * If we are doing text replication COW on first touch.
2956 */
2957 ASSERT(amp != NULL);
2958 ASSERT(svd->vp != NULL);
2959 ASSERT(rw != S_WRITE);
2960 cow = (ap == NULL);
2961 } else {
2962 cow = 0;
2963 }
2964
2965 /*
2966 * If not a copy-on-write case load the translation
2967 * and return.
2968 */
2969 if (cow == 0) {
2970
2971 /*
2972 * Handle pages that have been marked for migration
2973 */
2974 if (lgrp_optimizations())
2975 page_migrate(seg, addr, &opp, 1);
2976
2977 if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2978 if (rw == S_WRITE)
2979 hat_setmod(opp);
2980 else if (rw != S_OTHER && !hat_ismod(opp))
2981 prot &= ~PROT_WRITE;
2982 }
2983
2984 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2985 (!svd->pageprot && svd->prot == (prot & vpprot)));
2986 ASSERT(amp == NULL ||
2987 svd->rcookie == HAT_INVALID_REGION_COOKIE);
2988 hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2989 svd->rcookie);
2990
2991 if (!(hat_flag & HAT_LOAD_LOCK))
2992 page_unlock(opp);
2993
2994 if (anon_lock) {
2995 anon_array_exit(&cookie);
2996 }
2997 return (0);
2998 }
2999
3000 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3001
3002 hat_setref(opp);
3003
3004 ASSERT(amp != NULL && anon_lock);
3005
3006 /*
3007 * Steal the page only if it isn't a private page
3008 * since stealing a private page is not worth the effort.
3009 */
3010 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
3011 steal = 1;
3012
3013 /*
3014 * Steal the original page if the following conditions are true:
3015 *
3016 * We are low on memory, the page is not private, page is not large,
3017 * not shared, not modified, not `locked' or if we have it `locked'
3018 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
3019 * that the page is not shared) and if it doesn't have any
3020 * translations. page_struct_lock isn't needed to look at p_cowcnt
3021 * and p_lckcnt because we first get exclusive lock on page.
3022 */
3023 (void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
3024
3025 if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
3026 page_tryupgrade(opp) && !hat_ismod(opp) &&
3027 ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
3028 (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
3029 vpage != NULL && VPP_ISPPLOCK(vpage)))) {
3030 /*
3031 * Check if this page has other translations
3032 * after unloading our translation.
3033 */
3034 if (hat_page_is_mapped(opp)) {
3035 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3036 hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
3037 HAT_UNLOAD);
3038 }
3039
3040 /*
3041 * hat_unload() might sync back someone else's recent
3042 * modification, so check again.
3043 */
3044 if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
3045 pageflags |= STEAL_PAGE;
3046 }
3047
3048 /*
3049 * If we have a vpage pointer, see if it indicates that we have
3050 * ``locked'' the page we map -- if so, tell anon_private to
3051 * transfer the locking resource to the new page.
3052 *
3053 * See Statement at the beginning of segvn_lockop regarding
3054 * the way lockcnts/cowcnts are handled during COW.
3055 *
3056 */
3057 if (vpage != NULL && VPP_ISPPLOCK(vpage))
3058 pageflags |= LOCK_PAGE;
3059
3060 /*
3061 * Allocate a private page and perform the copy.
3062 * For MAP_NORESERVE reserve swap space now, unless this
3063 * is a cow fault on an existing anon page in which case
3064 * MAP_NORESERVE will have made advance reservations.
3065 */
3066 if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3067 if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3068 atomic_add_long(&svd->swresv, ptob(1));
3069 atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3070 } else {
3071 page_unlock(opp);
3072 err = ENOMEM;
3073 goto out;
3074 }
3075 }
3076 oldap = ap;
3077 pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3078 if (pp == NULL) {
3079 err = ENOMEM; /* out of swap space */
3080 goto out;
3081 }
3082
3083 /*
3084 * If we copied away from an anonymous page, then
3085 * we are one step closer to freeing up an anon slot.
3086 *
3087 * NOTE: The original anon slot must be released while
3088 * holding the "anon_map" lock. This is necessary to prevent
3089 * other threads from obtaining a pointer to the anon slot
3090 * which may be freed if its "refcnt" is 1.
3091 */
3092 if (oldap != NULL)
3093 anon_decref(oldap);
3094
3095 (void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3096
3097 /*
3098 * Handle pages that have been marked for migration
3099 */
3100 if (lgrp_optimizations())
3101 page_migrate(seg, addr, &pp, 1);
3102
3103 ASSERT(pp->p_szc == 0);
3104
3105 ASSERT(!IS_VMODSORT(pp->p_vnode));
3106 if (enable_mbit_wa) {
3107 if (rw == S_WRITE)
3108 hat_setmod(pp);
3109 else if (!hat_ismod(pp))
3110 prot &= ~PROT_WRITE;
3111 }
3112
3113 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3114 hat_memload(hat, addr, pp, prot, hat_flag);
3115
3116 if (!(hat_flag & HAT_LOAD_LOCK))
3117 page_unlock(pp);
3118
3119 ASSERT(anon_lock);
3120 anon_array_exit(&cookie);
3121 return (0);
3122 out:
3123 if (anon_lock)
3124 anon_array_exit(&cookie);
3125
3126 if (type == F_SOFTLOCK) {
3127 atomic_dec_ulong((ulong_t *)&svd->softlockcnt);
3128 }
3129 return (FC_MAKE_ERR(err));
3130 }
3131
3132 /*
3133 * relocate a bunch of smaller targ pages into one large repl page. all targ
3134 * pages must be complete pages smaller than replacement pages.
3135 * it's assumed that no page's szc can change since they are all PAGESIZE or
3136 * complete large pages locked SHARED.
3137 */
3138 static void
3139 segvn_relocate_pages(page_t **targ, page_t *replacement)
3140 {
3141 page_t *pp;
3142 pgcnt_t repl_npgs, curnpgs;
3143 pgcnt_t i;
3144 uint_t repl_szc = replacement->p_szc;
3145 page_t *first_repl = replacement;
3146 page_t *repl;
3147 spgcnt_t npgs;
3148
3149 VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3150
3151 ASSERT(repl_szc != 0);
3152 npgs = repl_npgs = page_get_pagecnt(repl_szc);
3153
3154 i = 0;
3155 while (repl_npgs) {
3156 spgcnt_t nreloc;
3157 int err;
3158 ASSERT(replacement != NULL);
3159 pp = targ[i];
3160 ASSERT(pp->p_szc < repl_szc);
3161 ASSERT(PAGE_EXCL(pp));
3162 ASSERT(!PP_ISFREE(pp));
3163 curnpgs = page_get_pagecnt(pp->p_szc);
3164 if (curnpgs == 1) {
3165 VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3166 repl = replacement;
3167 page_sub(&replacement, repl);
3168 ASSERT(PAGE_EXCL(repl));
3169 ASSERT(!PP_ISFREE(repl));
3170 ASSERT(repl->p_szc == repl_szc);
3171 } else {
3172 page_t *repl_savepp;
3173 int j;
3174 VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3175 repl_savepp = replacement;
3176 for (j = 0; j < curnpgs; j++) {
3177 repl = replacement;
3178 page_sub(&replacement, repl);
3179 ASSERT(PAGE_EXCL(repl));
3180 ASSERT(!PP_ISFREE(repl));
3181 ASSERT(repl->p_szc == repl_szc);
3182 ASSERT(page_pptonum(targ[i + j]) ==
3183 page_pptonum(targ[i]) + j);
3184 }
3185 repl = repl_savepp;
3186 ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3187 }
3188 err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3189 if (err || nreloc != curnpgs) {
3190 panic("segvn_relocate_pages: "
3191 "page_relocate failed err=%d curnpgs=%ld "
3192 "nreloc=%ld", err, curnpgs, nreloc);
3193 }
3194 ASSERT(curnpgs <= repl_npgs);
3195 repl_npgs -= curnpgs;
3196 i += curnpgs;
3197 }
3198 ASSERT(replacement == NULL);
3199
3200 repl = first_repl;
3201 repl_npgs = npgs;
3202 for (i = 0; i < repl_npgs; i++) {
3203 ASSERT(PAGE_EXCL(repl));
3204 ASSERT(!PP_ISFREE(repl));
3205 targ[i] = repl;
3206 page_downgrade(targ[i]);
3207 repl++;
3208 }
3209 }
3210
3211 /*
3212 * Check if all pages in ppa array are complete smaller than szc pages and
3213 * their roots will still be aligned relative to their current size if the
3214 * entire ppa array is relocated into one szc page. If these conditions are
3215 * not met return 0.
3216 *
3217 * If all pages are properly aligned attempt to upgrade their locks
3218 * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3219 * upgrdfail was set to 0 by caller.
3220 *
3221 * Return 1 if all pages are aligned and locked exclusively.
3222 *
3223 * If all pages in ppa array happen to be physically contiguous to make one
3224 * szc page and all exclusive locks are successfully obtained promote the page
3225 * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3226 */
3227 static int
3228 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3229 {
3230 page_t *pp;
3231 pfn_t pfn;
3232 pgcnt_t totnpgs = page_get_pagecnt(szc);
3233 pfn_t first_pfn;
3234 int contig = 1;
3235 pgcnt_t i;
3236 pgcnt_t j;
3237 uint_t curszc;
3238 pgcnt_t curnpgs;
3239 int root = 0;
3240
3241 ASSERT(szc > 0);
3242
3243 VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3244
3245 for (i = 0; i < totnpgs; i++) {
3246 pp = ppa[i];
3247 ASSERT(PAGE_SHARED(pp));
3248 ASSERT(!PP_ISFREE(pp));
3249 pfn = page_pptonum(pp);
3250 if (i == 0) {
3251 if (!IS_P2ALIGNED(pfn, totnpgs)) {
3252 contig = 0;
3253 } else {
3254 first_pfn = pfn;
3255 }
3256 } else if (contig && pfn != first_pfn + i) {
3257 contig = 0;
3258 }
3259 if (pp->p_szc == 0) {
3260 if (root) {
3261 VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3262 return (0);
3263 }
3264 } else if (!root) {
3265 if ((curszc = pp->p_szc) >= szc) {
3266 VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3267 return (0);
3268 }
3269 if (curszc == 0) {
3270 /*
3271 * p_szc changed means we don't have all pages
3272 * locked. return failure.
3273 */
3274 VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3275 return (0);
3276 }
3277 curnpgs = page_get_pagecnt(curszc);
3278 if (!IS_P2ALIGNED(pfn, curnpgs) ||
3279 !IS_P2ALIGNED(i, curnpgs)) {
3280 VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3281 return (0);
3282 }
3283 root = 1;
3284 } else {
3285 ASSERT(i > 0);
3286 VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3287 if (pp->p_szc != curszc) {
3288 VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3289 return (0);
3290 }
3291 if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3292 panic("segvn_full_szcpages: "
3293 "large page not physically contiguous");
3294 }
3295 if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3296 root = 0;
3297 }
3298 }
3299 }
3300
3301 for (i = 0; i < totnpgs; i++) {
3302 ASSERT(ppa[i]->p_szc < szc);
3303 if (!page_tryupgrade(ppa[i])) {
3304 for (j = 0; j < i; j++) {
3305 page_downgrade(ppa[j]);
3306 }
3307 *pszc = ppa[i]->p_szc;
3308 *upgrdfail = 1;
3309 VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3310 return (0);
3311 }
3312 }
3313
3314 /*
3315 * When a page is put a free cachelist its szc is set to 0. if file
3316 * system reclaimed pages from cachelist targ pages will be physically
3317 * contiguous with 0 p_szc. in this case just upgrade szc of targ
3318 * pages without any relocations.
3319 * To avoid any hat issues with previous small mappings
3320 * hat_pageunload() the target pages first.
3321 */
3322 if (contig) {
3323 VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3324 for (i = 0; i < totnpgs; i++) {
3325 (void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3326 }
3327 for (i = 0; i < totnpgs; i++) {
3328 ppa[i]->p_szc = szc;
3329 }
3330 for (i = 0; i < totnpgs; i++) {
3331 ASSERT(PAGE_EXCL(ppa[i]));
3332 page_downgrade(ppa[i]);
3333 }
3334 if (pszc != NULL) {
3335 *pszc = szc;
3336 }
3337 }
3338 VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3339 return (1);
3340 }
3341
3342 /*
3343 * Create physically contiguous pages for [vp, off] - [vp, off +
3344 * page_size(szc)) range and for private segment return them in ppa array.
3345 * Pages are created either via IO or relocations.
3346 *
3347 * Return 1 on success and 0 on failure.
3348 *
3349 * If physically contiguous pages already exist for this range return 1 without
3350 * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3351 * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3352 */
3353
3354 static int
3355 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3356 uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3357 int *downsize)
3358
3359 {
3360 page_t *pplist = *ppplist;
3361 size_t pgsz = page_get_pagesize(szc);
3362 pgcnt_t pages = btop(pgsz);
3363 ulong_t start_off = off;
3364 u_offset_t eoff = off + pgsz;
3365 spgcnt_t nreloc;
3366 u_offset_t io_off = off;
3367 size_t io_len;
3368 page_t *io_pplist = NULL;
3369 page_t *done_pplist = NULL;
3370 pgcnt_t pgidx = 0;
3371 page_t *pp;
3372 page_t *newpp;
3373 page_t *targpp;
3374 int io_err = 0;
3375 int i;
3376 pfn_t pfn;
3377 ulong_t ppages;
3378 page_t *targ_pplist = NULL;
3379 page_t *repl_pplist = NULL;
3380 page_t *tmp_pplist;
3381 int nios = 0;
3382 uint_t pszc;
3383 struct vattr va;
3384
3385 VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3386
3387 ASSERT(szc != 0);
3388 ASSERT(pplist->p_szc == szc);
3389
3390 /*
3391 * downsize will be set to 1 only if we fail to lock pages. this will
3392 * allow subsequent faults to try to relocate the page again. If we
3393 * fail due to misalignment don't downsize and let the caller map the
3394 * whole region with small mappings to avoid more faults into the area
3395 * where we can't get large pages anyway.
3396 */
3397 *downsize = 0;
3398
3399 while (off < eoff) {
3400 newpp = pplist;
3401 ASSERT(newpp != NULL);
3402 ASSERT(PAGE_EXCL(newpp));
3403 ASSERT(!PP_ISFREE(newpp));
3404 /*
3405 * we pass NULL for nrelocp to page_lookup_create()
3406 * so that it doesn't relocate. We relocate here
3407 * later only after we make sure we can lock all
3408 * pages in the range we handle and they are all
3409 * aligned.
3410 */
3411 pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3412 ASSERT(pp != NULL);
3413 ASSERT(!PP_ISFREE(pp));
3414 ASSERT(pp->p_vnode == vp);
3415 ASSERT(pp->p_offset == off);
3416 if (pp == newpp) {
3417 VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3418 page_sub(&pplist, pp);
3419 ASSERT(PAGE_EXCL(pp));
3420 ASSERT(page_iolock_assert(pp));
3421 page_list_concat(&io_pplist, &pp);
3422 off += PAGESIZE;
3423 continue;
3424 }
3425 VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3426 pfn = page_pptonum(pp);
3427 pszc = pp->p_szc;
3428 if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3429 IS_P2ALIGNED(pfn, pages)) {
3430 ASSERT(repl_pplist == NULL);
3431 ASSERT(done_pplist == NULL);
3432 ASSERT(pplist == *ppplist);
3433 page_unlock(pp);
3434 page_free_replacement_page(pplist);
3435 page_create_putback(pages);
3436 *ppplist = NULL;
3437 VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3438 return (1);
3439 }
3440 if (pszc >= szc) {
3441 page_unlock(pp);
3442 segvn_faultvnmpss_align_err1++;
3443 goto out;
3444 }
3445 ppages = page_get_pagecnt(pszc);
3446 if (!IS_P2ALIGNED(pfn, ppages)) {
3447 ASSERT(pszc > 0);
3448 /*
3449 * sizing down to pszc won't help.
3450 */
3451 page_unlock(pp);
3452 segvn_faultvnmpss_align_err2++;
3453 goto out;
3454 }
3455 pfn = page_pptonum(newpp);
3456 if (!IS_P2ALIGNED(pfn, ppages)) {
3457 ASSERT(pszc > 0);
3458 /*
3459 * sizing down to pszc won't help.
3460 */
3461 page_unlock(pp);
3462 segvn_faultvnmpss_align_err3++;
3463 goto out;
3464 }
3465 if (!PAGE_EXCL(pp)) {
3466 VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3467 page_unlock(pp);
3468 *downsize = 1;
3469 *ret_pszc = pp->p_szc;
3470 goto out;
3471 }
3472 targpp = pp;
3473 if (io_pplist != NULL) {
3474 VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3475 io_len = off - io_off;
3476 /*
3477 * Some file systems like NFS don't check EOF
3478 * conditions in VOP_PAGEIO(). Check it here
3479 * now that pages are locked SE_EXCL. Any file
3480 * truncation will wait until the pages are
3481 * unlocked so no need to worry that file will
3482 * be truncated after we check its size here.
3483 * XXX fix NFS to remove this check.
3484 */
3485 va.va_mask = AT_SIZE;
3486 if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3487 VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3488 page_unlock(targpp);
3489 goto out;
3490 }
3491 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3492 VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3493 *downsize = 1;
3494 *ret_pszc = 0;
3495 page_unlock(targpp);
3496 goto out;
3497 }
3498 io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3499 B_READ, svd->cred, NULL);
3500 if (io_err) {
3501 VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3502 page_unlock(targpp);
3503 if (io_err == EDEADLK) {
3504 segvn_vmpss_pageio_deadlk_err++;
3505 }
3506 goto out;
3507 }
3508 nios++;
3509 VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3510 while (io_pplist != NULL) {
3511 pp = io_pplist;
3512 page_sub(&io_pplist, pp);
3513 ASSERT(page_iolock_assert(pp));
3514 page_io_unlock(pp);
3515 pgidx = (pp->p_offset - start_off) >>
3516 PAGESHIFT;
3517 ASSERT(pgidx < pages);
3518 ppa[pgidx] = pp;
3519 page_list_concat(&done_pplist, &pp);
3520 }
3521 }
3522 pp = targpp;
3523 ASSERT(PAGE_EXCL(pp));
3524 ASSERT(pp->p_szc <= pszc);
3525 if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3526 VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3527 page_unlock(pp);
3528 *downsize = 1;
3529 *ret_pszc = pp->p_szc;
3530 goto out;
3531 }
3532 VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3533 /*
3534 * page szc chould have changed before the entire group was
3535 * locked. reread page szc.
3536 */
3537 pszc = pp->p_szc;
3538 ppages = page_get_pagecnt(pszc);
3539
3540 /* link just the roots */
3541 page_list_concat(&targ_pplist, &pp);
3542 page_sub(&pplist, newpp);
3543 page_list_concat(&repl_pplist, &newpp);
3544 off += PAGESIZE;
3545 while (--ppages != 0) {
3546 newpp = pplist;
3547 page_sub(&pplist, newpp);
3548 off += PAGESIZE;
3549 }
3550 io_off = off;
3551 }
3552 if (io_pplist != NULL) {
3553 VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3554 io_len = eoff - io_off;
3555 va.va_mask = AT_SIZE;
3556 if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3557 VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3558 goto out;
3559 }
3560 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3561 VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3562 *downsize = 1;
3563 *ret_pszc = 0;
3564 goto out;
3565 }
3566 io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3567 B_READ, svd->cred, NULL);
3568 if (io_err) {
3569 VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3570 if (io_err == EDEADLK) {
3571 segvn_vmpss_pageio_deadlk_err++;
3572 }
3573 goto out;
3574 }
3575 nios++;
3576 while (io_pplist != NULL) {
3577 pp = io_pplist;
3578 page_sub(&io_pplist, pp);
3579 ASSERT(page_iolock_assert(pp));
3580 page_io_unlock(pp);
3581 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3582 ASSERT(pgidx < pages);
3583 ppa[pgidx] = pp;
3584 }
3585 }
3586 /*
3587 * we're now bound to succeed or panic.
3588 * remove pages from done_pplist. it's not needed anymore.
3589 */
3590 while (done_pplist != NULL) {
3591 pp = done_pplist;
3592 page_sub(&done_pplist, pp);
3593 }
3594 VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3595 ASSERT(pplist == NULL);
3596 *ppplist = NULL;
3597 while (targ_pplist != NULL) {
3598 int ret;
3599 VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3600 ASSERT(repl_pplist);
3601 pp = targ_pplist;
3602 page_sub(&targ_pplist, pp);
3603 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3604 newpp = repl_pplist;
3605 page_sub(&repl_pplist, newpp);
3606 #ifdef DEBUG
3607 pfn = page_pptonum(pp);
3608 pszc = pp->p_szc;
3609 ppages = page_get_pagecnt(pszc);
3610 ASSERT(IS_P2ALIGNED(pfn, ppages));
3611 pfn = page_pptonum(newpp);
3612 ASSERT(IS_P2ALIGNED(pfn, ppages));
3613 ASSERT(P2PHASE(pfn, pages) == pgidx);
3614 #endif
3615 nreloc = 0;
3616 ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3617 if (ret != 0 || nreloc == 0) {
3618 panic("segvn_fill_vp_pages: "
3619 "page_relocate failed");
3620 }
3621 pp = newpp;
3622 while (nreloc-- != 0) {
3623 ASSERT(PAGE_EXCL(pp));
3624 ASSERT(pp->p_vnode == vp);
3625 ASSERT(pgidx ==
3626 ((pp->p_offset - start_off) >> PAGESHIFT));
3627 ppa[pgidx++] = pp;
3628 pp++;
3629 }
3630 }
3631
3632 if (svd->type == MAP_PRIVATE) {
3633 VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3634 for (i = 0; i < pages; i++) {
3635 ASSERT(ppa[i] != NULL);
3636 ASSERT(PAGE_EXCL(ppa[i]));
3637 ASSERT(ppa[i]->p_vnode == vp);
3638 ASSERT(ppa[i]->p_offset ==
3639 start_off + (i << PAGESHIFT));
3640 page_downgrade(ppa[i]);
3641 }
3642 ppa[pages] = NULL;
3643 } else {
3644 VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3645 /*
3646 * the caller will still call VOP_GETPAGE() for shared segments
3647 * to check FS write permissions. For private segments we map
3648 * file read only anyway. so no VOP_GETPAGE is needed.
3649 */
3650 for (i = 0; i < pages; i++) {
3651 ASSERT(ppa[i] != NULL);
3652 ASSERT(PAGE_EXCL(ppa[i]));
3653 ASSERT(ppa[i]->p_vnode == vp);
3654 ASSERT(ppa[i]->p_offset ==
3655 start_off + (i << PAGESHIFT));
3656 page_unlock(ppa[i]);
3657 }
3658 ppa[0] = NULL;
3659 }
3660
3661 return (1);
3662 out:
3663 /*
3664 * Do the cleanup. Unlock target pages we didn't relocate. They are
3665 * linked on targ_pplist by root pages. reassemble unused replacement
3666 * and io pages back to pplist.
3667 */
3668 if (io_pplist != NULL) {
3669 VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3670 pp = io_pplist;
3671 do {
3672 ASSERT(pp->p_vnode == vp);
3673 ASSERT(pp->p_offset == io_off);
3674 ASSERT(page_iolock_assert(pp));
3675 page_io_unlock(pp);
3676 page_hashout(pp, NULL);
3677 io_off += PAGESIZE;
3678 } while ((pp = pp->p_next) != io_pplist);
3679 page_list_concat(&io_pplist, &pplist);
3680 pplist = io_pplist;
3681 }
3682 tmp_pplist = NULL;
3683 while (targ_pplist != NULL) {
3684 VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3685 pp = targ_pplist;
3686 ASSERT(PAGE_EXCL(pp));
3687 page_sub(&targ_pplist, pp);
3688
3689 pszc = pp->p_szc;
3690 ppages = page_get_pagecnt(pszc);
3691 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3692
3693 if (pszc != 0) {
3694 group_page_unlock(pp);
3695 }
3696 page_unlock(pp);
3697
3698 pp = repl_pplist;
3699 ASSERT(pp != NULL);
3700 ASSERT(PAGE_EXCL(pp));
3701 ASSERT(pp->p_szc == szc);
3702 page_sub(&repl_pplist, pp);
3703
3704 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3705
3706 /* relink replacement page */
3707 page_list_concat(&tmp_pplist, &pp);
3708 while (--ppages != 0) {
3709 VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3710 pp++;
3711 ASSERT(PAGE_EXCL(pp));
3712 ASSERT(pp->p_szc == szc);
3713 page_list_concat(&tmp_pplist, &pp);
3714 }
3715 }
3716 if (tmp_pplist != NULL) {
3717 VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3718 page_list_concat(&tmp_pplist, &pplist);
3719 pplist = tmp_pplist;
3720 }
3721 /*
3722 * at this point all pages are either on done_pplist or
3723 * pplist. They can't be all on done_pplist otherwise
3724 * we'd've been done.
3725 */
3726 ASSERT(pplist != NULL);
3727 if (nios != 0) {
3728 VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3729 pp = pplist;
3730 do {
3731 VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3732 ASSERT(pp->p_szc == szc);
3733 ASSERT(PAGE_EXCL(pp));
3734 ASSERT(pp->p_vnode != vp);
3735 pp->p_szc = 0;
3736 } while ((pp = pp->p_next) != pplist);
3737
3738 pp = done_pplist;
3739 do {
3740 VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3741 ASSERT(pp->p_szc == szc);
3742 ASSERT(PAGE_EXCL(pp));
3743 ASSERT(pp->p_vnode == vp);
3744 pp->p_szc = 0;
3745 } while ((pp = pp->p_next) != done_pplist);
3746
3747 while (pplist != NULL) {
3748 VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3749 pp = pplist;
3750 page_sub(&pplist, pp);
3751 page_free(pp, 0);
3752 }
3753
3754 while (done_pplist != NULL) {
3755 VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3756 pp = done_pplist;
3757 page_sub(&done_pplist, pp);
3758 page_unlock(pp);
3759 }
3760 *ppplist = NULL;
3761 return (0);
3762 }
3763 ASSERT(pplist == *ppplist);
3764 if (io_err) {
3765 VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3766 /*
3767 * don't downsize on io error.
3768 * see if vop_getpage succeeds.
3769 * pplist may still be used in this case
3770 * for relocations.
3771 */
3772 return (0);
3773 }
3774 VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3775 page_free_replacement_page(pplist);
3776 page_create_putback(pages);
3777 *ppplist = NULL;
3778 return (0);
3779 }
3780
3781 int segvn_anypgsz = 0;
3782
3783 #define SEGVN_RESTORE_SOFTLOCK_VP(type, pages) \
3784 if ((type) == F_SOFTLOCK) { \
3785 atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3786 -(pages)); \
3787 }
3788
3789 #define SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot) \
3790 if (IS_VMODSORT((ppa)[0]->p_vnode)) { \
3791 if ((rw) == S_WRITE) { \
3792 for (i = 0; i < (pages); i++) { \
3793 ASSERT((ppa)[i]->p_vnode == \
3794 (ppa)[0]->p_vnode); \
3795 hat_setmod((ppa)[i]); \
3796 } \
3797 } else if ((rw) != S_OTHER && \
3798 ((prot) & (vpprot) & PROT_WRITE)) { \
3799 for (i = 0; i < (pages); i++) { \
3800 ASSERT((ppa)[i]->p_vnode == \
3801 (ppa)[0]->p_vnode); \
3802 if (!hat_ismod((ppa)[i])) { \
3803 prot &= ~PROT_WRITE; \
3804 break; \
3805 } \
3806 } \
3807 } \
3808 }
3809
3810 #ifdef VM_STATS
3811
3812 #define SEGVN_VMSTAT_FLTVNPAGES(idx) \
3813 VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3814
3815 #else /* VM_STATS */
3816
3817 #define SEGVN_VMSTAT_FLTVNPAGES(idx)
3818
3819 #endif
3820
3821 static faultcode_t
3822 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3823 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3824 caddr_t eaddr, int brkcow)
3825 {
3826 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3827 struct anon_map *amp = svd->amp;
3828 uchar_t segtype = svd->type;
3829 uint_t szc = seg->s_szc;
3830 size_t pgsz = page_get_pagesize(szc);
3831 size_t maxpgsz = pgsz;
3832 pgcnt_t pages = btop(pgsz);
3833 pgcnt_t maxpages = pages;
3834 size_t ppasize = (pages + 1) * sizeof (page_t *);
3835 caddr_t a = lpgaddr;
3836 caddr_t maxlpgeaddr = lpgeaddr;
3837 u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3838 ulong_t aindx = svd->anon_index + seg_page(seg, a);
3839 struct vpage *vpage = (svd->vpage != NULL) ?
3840 &svd->vpage[seg_page(seg, a)] : NULL;
3841 vnode_t *vp = svd->vp;
3842 page_t **ppa;
3843 uint_t pszc;
3844 size_t ppgsz;
3845 pgcnt_t ppages;
3846 faultcode_t err = 0;
3847 int ierr;
3848 int vop_size_err = 0;
3849 uint_t protchk, prot, vpprot;
3850 ulong_t i;
3851 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3852 anon_sync_obj_t an_cookie;
3853 enum seg_rw arw;
3854 int alloc_failed = 0;
3855 int adjszc_chk;
3856 struct vattr va;
3857 page_t *pplist;
3858 pfn_t pfn;
3859 int physcontig;
3860 int upgrdfail;
3861 int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3862 int tron = (svd->tr_state == SEGVN_TR_ON);
3863
3864 ASSERT(szc != 0);
3865 ASSERT(vp != NULL);
3866 ASSERT(brkcow == 0 || amp != NULL);
3867 ASSERT(tron == 0 || amp != NULL);
3868 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3869 ASSERT(!(svd->flags & MAP_NORESERVE));
3870 ASSERT(type != F_SOFTUNLOCK);
3871 ASSERT(IS_P2ALIGNED(a, maxpgsz));
3872 ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3873 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3874 ASSERT(seg->s_szc < NBBY * sizeof (int));
3875 ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3876 ASSERT(svd->tr_state != SEGVN_TR_INIT);
3877
3878 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3879 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3880
3881 if (svd->flags & MAP_TEXT) {
3882 hat_flag |= HAT_LOAD_TEXT;
3883 }
3884
3885 if (svd->pageprot) {
3886 switch (rw) {
3887 case S_READ:
3888 protchk = PROT_READ;
3889 break;
3890 case S_WRITE:
3891 protchk = PROT_WRITE;
3892 break;
3893 case S_EXEC:
3894 protchk = PROT_EXEC;
3895 break;
3896 case S_OTHER:
3897 default:
3898 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3899 break;
3900 }
3901 } else {
3902 prot = svd->prot;
3903 /* caller has already done segment level protection check. */
3904 }
3905
3906 if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3907 SEGVN_VMSTAT_FLTVNPAGES(2);
3908 arw = S_READ;
3909 } else {
3910 arw = rw;
3911 }
3912
3913 ppa = kmem_alloc(ppasize, KM_SLEEP);
3914
3915 VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3916
3917 for (;;) {
3918 adjszc_chk = 0;
3919 for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3920 if (adjszc_chk) {
3921 while (szc < seg->s_szc) {
3922 uintptr_t e;
3923 uint_t tszc;
3924 tszc = segvn_anypgsz_vnode ? szc + 1 :
3925 seg->s_szc;
3926 ppgsz = page_get_pagesize(tszc);
3927 if (!IS_P2ALIGNED(a, ppgsz) ||
3928 ((alloc_failed >> tszc) & 0x1)) {
3929 break;
3930 }
3931 SEGVN_VMSTAT_FLTVNPAGES(4);
3932 szc = tszc;
3933 pgsz = ppgsz;
3934 pages = btop(pgsz);
3935 e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3936 lpgeaddr = (caddr_t)e;
3937 }
3938 }
3939
3940 again:
3941 if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3942 ASSERT(IS_P2ALIGNED(aindx, maxpages));
3943 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
3944 anon_array_enter(amp, aindx, &an_cookie);
3945 if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3946 SEGVN_VMSTAT_FLTVNPAGES(5);
3947 ASSERT(anon_pages(amp->ahp, aindx,
3948 maxpages) == maxpages);
3949 anon_array_exit(&an_cookie);
3950 ANON_LOCK_EXIT(&->a_rwlock);
3951 err = segvn_fault_anonpages(hat, seg,
3952 a, a + maxpgsz, type, rw,
3953 MAX(a, addr),
3954 MIN(a + maxpgsz, eaddr), brkcow);
3955 if (err != 0) {
3956 SEGVN_VMSTAT_FLTVNPAGES(6);
3957 goto out;
3958 }
3959 if (szc < seg->s_szc) {
3960 szc = seg->s_szc;
3961 pgsz = maxpgsz;
3962 pages = maxpages;
3963 lpgeaddr = maxlpgeaddr;
3964 }
3965 goto next;
3966 } else {
3967 ASSERT(anon_pages(amp->ahp, aindx,
3968 maxpages) == 0);
3969 SEGVN_VMSTAT_FLTVNPAGES(7);
3970 anon_array_exit(&an_cookie);
3971 ANON_LOCK_EXIT(&->a_rwlock);
3972 }
3973 }
3974 ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3975 ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3976
3977 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3978 ASSERT(vpage != NULL);
3979 prot = VPP_PROT(vpage);
3980 ASSERT(sameprot(seg, a, maxpgsz));
3981 if ((prot & protchk) == 0) {
3982 SEGVN_VMSTAT_FLTVNPAGES(8);
3983 err = FC_PROT;
3984 goto out;
3985 }
3986 }
3987 if (type == F_SOFTLOCK) {
3988 atomic_add_long((ulong_t *)&svd->softlockcnt,
3989 pages);
3990 }
3991
3992 pplist = NULL;
3993 physcontig = 0;
3994 ppa[0] = NULL;
3995 if (!brkcow && !tron && szc &&
3996 !page_exists_physcontig(vp, off, szc,
3997 segtype == MAP_PRIVATE ? ppa : NULL)) {
3998 SEGVN_VMSTAT_FLTVNPAGES(9);
3999 if (page_alloc_pages(vp, seg, a, &pplist, NULL,
4000 szc, 0, 0) && type != F_SOFTLOCK) {
4001 SEGVN_VMSTAT_FLTVNPAGES(10);
4002 pszc = 0;
4003 ierr = -1;
4004 alloc_failed |= (1 << szc);
4005 break;
4006 }
4007 if (pplist != NULL &&
4008 vp->v_mpssdata == SEGVN_PAGEIO) {
4009 int downsize;
4010 SEGVN_VMSTAT_FLTVNPAGES(11);
4011 physcontig = segvn_fill_vp_pages(svd,
4012 vp, off, szc, ppa, &pplist,
4013 &pszc, &downsize);
4014 ASSERT(!physcontig || pplist == NULL);
4015 if (!physcontig && downsize &&
4016 type != F_SOFTLOCK) {
4017 ASSERT(pplist == NULL);
4018 SEGVN_VMSTAT_FLTVNPAGES(12);
4019 ierr = -1;
4020 break;
4021 }
4022 ASSERT(!physcontig ||
4023 segtype == MAP_PRIVATE ||
4024 ppa[0] == NULL);
4025 if (physcontig && ppa[0] == NULL) {
4026 physcontig = 0;
4027 }
4028 }
4029 } else if (!brkcow && !tron && szc && ppa[0] != NULL) {
4030 SEGVN_VMSTAT_FLTVNPAGES(13);
4031 ASSERT(segtype == MAP_PRIVATE);
4032 physcontig = 1;
4033 }
4034
4035 if (!physcontig) {
4036 SEGVN_VMSTAT_FLTVNPAGES(14);
4037 ppa[0] = NULL;
4038 ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
4039 &vpprot, ppa, pgsz, seg, a, arw,
4040 svd->cred, NULL);
4041 #ifdef DEBUG
4042 if (ierr == 0) {
4043 for (i = 0; i < pages; i++) {
4044 ASSERT(PAGE_LOCKED(ppa[i]));
4045 ASSERT(!PP_ISFREE(ppa[i]));
4046 ASSERT(ppa[i]->p_vnode == vp);
4047 ASSERT(ppa[i]->p_offset ==
4048 off + (i << PAGESHIFT));
4049 }
4050 }
4051 #endif /* DEBUG */
4052 if (segtype == MAP_PRIVATE) {
4053 SEGVN_VMSTAT_FLTVNPAGES(15);
4054 vpprot &= ~PROT_WRITE;
4055 }
4056 } else {
4057 ASSERT(segtype == MAP_PRIVATE);
4058 SEGVN_VMSTAT_FLTVNPAGES(16);
4059 vpprot = PROT_ALL & ~PROT_WRITE;
4060 ierr = 0;
4061 }
4062
4063 if (ierr != 0) {
4064 SEGVN_VMSTAT_FLTVNPAGES(17);
4065 if (pplist != NULL) {
4066 SEGVN_VMSTAT_FLTVNPAGES(18);
4067 page_free_replacement_page(pplist);
4068 page_create_putback(pages);
4069 }
4070 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4071 if (a + pgsz <= eaddr) {
4072 SEGVN_VMSTAT_FLTVNPAGES(19);
4073 err = FC_MAKE_ERR(ierr);
4074 goto out;
4075 }
4076 va.va_mask = AT_SIZE;
4077 if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
4078 SEGVN_VMSTAT_FLTVNPAGES(20);
4079 err = FC_MAKE_ERR(EIO);
4080 goto out;
4081 }
4082 if (btopr(va.va_size) >= btopr(off + pgsz)) {
4083 SEGVN_VMSTAT_FLTVNPAGES(21);
4084 err = FC_MAKE_ERR(ierr);
4085 goto out;
4086 }
4087 if (btopr(va.va_size) <
4088 btopr(off + (eaddr - a))) {
4089 SEGVN_VMSTAT_FLTVNPAGES(22);
4090 err = FC_MAKE_ERR(ierr);
4091 goto out;
4092 }
4093 if (brkcow || tron || type == F_SOFTLOCK) {
4094 /* can't reduce map area */
4095 SEGVN_VMSTAT_FLTVNPAGES(23);
4096 vop_size_err = 1;
4097 goto out;
4098 }
4099 SEGVN_VMSTAT_FLTVNPAGES(24);
4100 ASSERT(szc != 0);
4101 pszc = 0;
4102 ierr = -1;
4103 break;
4104 }
4105
4106 if (amp != NULL) {
4107 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
4108 anon_array_enter(amp, aindx, &an_cookie);
4109 }
4110 if (amp != NULL &&
4111 anon_get_ptr(amp->ahp, aindx) != NULL) {
4112 ulong_t taindx = P2ALIGN(aindx, maxpages);
4113
4114 SEGVN_VMSTAT_FLTVNPAGES(25);
4115 ASSERT(anon_pages(amp->ahp, taindx,
4116 maxpages) == maxpages);
4117 for (i = 0; i < pages; i++) {
4118 page_unlock(ppa[i]);
4119 }
4120 anon_array_exit(&an_cookie);
4121 ANON_LOCK_EXIT(&->a_rwlock);
4122 if (pplist != NULL) {
4123 page_free_replacement_page(pplist);
4124 page_create_putback(pages);
4125 }
4126 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4127 if (szc < seg->s_szc) {
4128 SEGVN_VMSTAT_FLTVNPAGES(26);
4129 /*
4130 * For private segments SOFTLOCK
4131 * either always breaks cow (any rw
4132 * type except S_READ_NOCOW) or
4133 * address space is locked as writer
4134 * (S_READ_NOCOW case) and anon slots
4135 * can't show up on second check.
4136 * Therefore if we are here for
4137 * SOFTLOCK case it must be a cow
4138 * break but cow break never reduces
4139 * szc. text replication (tron) in
4140 * this case works as cow break.
4141 * Thus the assert below.
4142 */
4143 ASSERT(!brkcow && !tron &&
4144 type != F_SOFTLOCK);
4145 pszc = seg->s_szc;
4146 ierr = -2;
4147 break;
4148 }
4149 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4150 goto again;
4151 }
4152 #ifdef DEBUG
4153 if (amp != NULL) {
4154 ulong_t taindx = P2ALIGN(aindx, maxpages);
4155 ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4156 }
4157 #endif /* DEBUG */
4158
4159 if (brkcow || tron) {
4160 ASSERT(amp != NULL);
4161 ASSERT(pplist == NULL);
4162 ASSERT(szc == seg->s_szc);
4163 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4164 ASSERT(IS_P2ALIGNED(aindx, maxpages));
4165 SEGVN_VMSTAT_FLTVNPAGES(27);
4166 ierr = anon_map_privatepages(amp, aindx, szc,
4167 seg, a, prot, ppa, vpage, segvn_anypgsz,
4168 tron ? PG_LOCAL : 0, svd->cred);
4169 if (ierr != 0) {
4170 SEGVN_VMSTAT_FLTVNPAGES(28);
4171 anon_array_exit(&an_cookie);
4172 ANON_LOCK_EXIT(&->a_rwlock);
4173 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4174 err = FC_MAKE_ERR(ierr);
4175 goto out;
4176 }
4177
4178 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4179 /*
4180 * p_szc can't be changed for locked
4181 * swapfs pages.
4182 */
4183 ASSERT(svd->rcookie ==
4184 HAT_INVALID_REGION_COOKIE);
4185 hat_memload_array(hat, a, pgsz, ppa, prot,
4186 hat_flag);
4187
4188 if (!(hat_flag & HAT_LOAD_LOCK)) {
4189 SEGVN_VMSTAT_FLTVNPAGES(29);
4190 for (i = 0; i < pages; i++) {
4191 page_unlock(ppa[i]);
4192 }
4193 }
4194 anon_array_exit(&an_cookie);
4195 ANON_LOCK_EXIT(&->a_rwlock);
4196 goto next;
4197 }
4198
4199 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4200 (!svd->pageprot && svd->prot == (prot & vpprot)));
4201
4202 pfn = page_pptonum(ppa[0]);
4203 /*
4204 * hat_page_demote() needs an SE_EXCL lock on one of
4205 * constituent page_t's and it decreases root's p_szc
4206 * last. This means if root's p_szc is equal szc and
4207 * all its constituent pages are locked
4208 * hat_page_demote() that could have changed p_szc to
4209 * szc is already done and no new have page_demote()
4210 * can start for this large page.
4211 */
4212
4213 /*
4214 * we need to make sure same mapping size is used for
4215 * the same address range if there's a possibility the
4216 * adddress is already mapped because hat layer panics
4217 * when translation is loaded for the range already
4218 * mapped with a different page size. We achieve it
4219 * by always using largest page size possible subject
4220 * to the constraints of page size, segment page size
4221 * and page alignment. Since mappings are invalidated
4222 * when those constraints change and make it
4223 * impossible to use previously used mapping size no
4224 * mapping size conflicts should happen.
4225 */
4226
4227 chkszc:
4228 if ((pszc = ppa[0]->p_szc) == szc &&
4229 IS_P2ALIGNED(pfn, pages)) {
4230
4231 SEGVN_VMSTAT_FLTVNPAGES(30);
4232 #ifdef DEBUG
4233 for (i = 0; i < pages; i++) {
4234 ASSERT(PAGE_LOCKED(ppa[i]));
4235 ASSERT(!PP_ISFREE(ppa[i]));
4236 ASSERT(page_pptonum(ppa[i]) ==
4237 pfn + i);
4238 ASSERT(ppa[i]->p_szc == szc);
4239 ASSERT(ppa[i]->p_vnode == vp);
4240 ASSERT(ppa[i]->p_offset ==
4241 off + (i << PAGESHIFT));
4242 }
4243 #endif /* DEBUG */
4244 /*
4245 * All pages are of szc we need and they are
4246 * all locked so they can't change szc. load
4247 * translations.
4248 *
4249 * if page got promoted since last check
4250 * we don't need pplist.
4251 */
4252 if (pplist != NULL) {
4253 page_free_replacement_page(pplist);
4254 page_create_putback(pages);
4255 }
4256 if (PP_ISMIGRATE(ppa[0])) {
4257 page_migrate(seg, a, ppa, pages);
4258 }
4259 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4260 prot, vpprot);
4261 hat_memload_array_region(hat, a, pgsz,
4262 ppa, prot & vpprot, hat_flag,
4263 svd->rcookie);
4264
4265 if (!(hat_flag & HAT_LOAD_LOCK)) {
4266 for (i = 0; i < pages; i++) {
4267 page_unlock(ppa[i]);
4268 }
4269 }
4270 if (amp != NULL) {
4271 anon_array_exit(&an_cookie);
4272 ANON_LOCK_EXIT(&->a_rwlock);
4273 }
4274 goto next;
4275 }
4276
4277 /*
4278 * See if upsize is possible.
4279 */
4280 if (pszc > szc && szc < seg->s_szc &&
4281 (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4282 pgcnt_t aphase;
4283 uint_t pszc1 = MIN(pszc, seg->s_szc);
4284 ppgsz = page_get_pagesize(pszc1);
4285 ppages = btop(ppgsz);
4286 aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4287
4288 ASSERT(type != F_SOFTLOCK);
4289
4290 SEGVN_VMSTAT_FLTVNPAGES(31);
4291 if (aphase != P2PHASE(pfn, ppages)) {
4292 segvn_faultvnmpss_align_err4++;
4293 } else {
4294 SEGVN_VMSTAT_FLTVNPAGES(32);
4295 if (pplist != NULL) {
4296 page_t *pl = pplist;
4297 page_free_replacement_page(pl);
4298 page_create_putback(pages);
4299 }
4300 for (i = 0; i < pages; i++) {
4301 page_unlock(ppa[i]);
4302 }
4303 if (amp != NULL) {
4304 anon_array_exit(&an_cookie);
4305 ANON_LOCK_EXIT(&->a_rwlock);
4306 }
4307 pszc = pszc1;
4308 ierr = -2;
4309 break;
4310 }
4311 }
4312
4313 /*
4314 * check if we should use smallest mapping size.
4315 */
4316 upgrdfail = 0;
4317 if (szc == 0 ||
4318 (pszc >= szc &&
4319 !IS_P2ALIGNED(pfn, pages)) ||
4320 (pszc < szc &&
4321 !segvn_full_szcpages(ppa, szc, &upgrdfail,
4322 &pszc))) {
4323
4324 if (upgrdfail && type != F_SOFTLOCK) {
4325 /*
4326 * segvn_full_szcpages failed to lock
4327 * all pages EXCL. Size down.
4328 */
4329 ASSERT(pszc < szc);
4330
4331 SEGVN_VMSTAT_FLTVNPAGES(33);
4332
4333 if (pplist != NULL) {
4334 page_t *pl = pplist;
4335 page_free_replacement_page(pl);
4336 page_create_putback(pages);
4337 }
4338
4339 for (i = 0; i < pages; i++) {
4340 page_unlock(ppa[i]);
4341 }
4342 if (amp != NULL) {
4343 anon_array_exit(&an_cookie);
4344 ANON_LOCK_EXIT(&->a_rwlock);
4345 }
4346 ierr = -1;
4347 break;
4348 }
4349 if (szc != 0 && !upgrdfail) {
4350 segvn_faultvnmpss_align_err5++;
4351 }
4352 SEGVN_VMSTAT_FLTVNPAGES(34);
4353 if (pplist != NULL) {
4354 page_free_replacement_page(pplist);
4355 page_create_putback(pages);
4356 }
4357 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4358 prot, vpprot);
4359 if (upgrdfail && segvn_anypgsz_vnode) {
4360 /* SOFTLOCK case */
4361 hat_memload_array_region(hat, a, pgsz,
4362 ppa, prot & vpprot, hat_flag,
4363 svd->rcookie);
4364 } else {
4365 for (i = 0; i < pages; i++) {
4366 hat_memload_region(hat,
4367 a + (i << PAGESHIFT),
4368 ppa[i], prot & vpprot,
4369 hat_flag, svd->rcookie);
4370 }
4371 }
4372 if (!(hat_flag & HAT_LOAD_LOCK)) {
4373 for (i = 0; i < pages; i++) {
4374 page_unlock(ppa[i]);
4375 }
4376 }
4377 if (amp != NULL) {
4378 anon_array_exit(&an_cookie);
4379 ANON_LOCK_EXIT(&->a_rwlock);
4380 }
4381 goto next;
4382 }
4383
4384 if (pszc == szc) {
4385 /*
4386 * segvn_full_szcpages() upgraded pages szc.
4387 */
4388 ASSERT(pszc == ppa[0]->p_szc);
4389 ASSERT(IS_P2ALIGNED(pfn, pages));
4390 goto chkszc;
4391 }
4392
4393 if (pszc > szc) {
4394 kmutex_t *szcmtx;
4395 SEGVN_VMSTAT_FLTVNPAGES(35);
4396 /*
4397 * p_szc of ppa[0] can change since we haven't
4398 * locked all constituent pages. Call
4399 * page_lock_szc() to prevent szc changes.
4400 * This should be a rare case that happens when
4401 * multiple segments use a different page size
4402 * to map the same file offsets.
4403 */
4404 szcmtx = page_szc_lock(ppa[0]);
4405 pszc = ppa[0]->p_szc;
4406 ASSERT(szcmtx != NULL || pszc == 0);
4407 ASSERT(ppa[0]->p_szc <= pszc);
4408 if (pszc <= szc) {
4409 SEGVN_VMSTAT_FLTVNPAGES(36);
4410 if (szcmtx != NULL) {
4411 mutex_exit(szcmtx);
4412 }
4413 goto chkszc;
4414 }
4415 if (pplist != NULL) {
4416 /*
4417 * page got promoted since last check.
4418 * we don't need preaalocated large
4419 * page.
4420 */
4421 SEGVN_VMSTAT_FLTVNPAGES(37);
4422 page_free_replacement_page(pplist);
4423 page_create_putback(pages);
4424 }
4425 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4426 prot, vpprot);
4427 hat_memload_array_region(hat, a, pgsz, ppa,
4428 prot & vpprot, hat_flag, svd->rcookie);
4429 mutex_exit(szcmtx);
4430 if (!(hat_flag & HAT_LOAD_LOCK)) {
4431 for (i = 0; i < pages; i++) {
4432 page_unlock(ppa[i]);
4433 }
4434 }
4435 if (amp != NULL) {
4436 anon_array_exit(&an_cookie);
4437 ANON_LOCK_EXIT(&->a_rwlock);
4438 }
4439 goto next;
4440 }
4441
4442 /*
4443 * if page got demoted since last check
4444 * we could have not allocated larger page.
4445 * allocate now.
4446 */
4447 if (pplist == NULL &&
4448 page_alloc_pages(vp, seg, a, &pplist, NULL,
4449 szc, 0, 0) && type != F_SOFTLOCK) {
4450 SEGVN_VMSTAT_FLTVNPAGES(38);
4451 for (i = 0; i < pages; i++) {
4452 page_unlock(ppa[i]);
4453 }
4454 if (amp != NULL) {
4455 anon_array_exit(&an_cookie);
4456 ANON_LOCK_EXIT(&->a_rwlock);
4457 }
4458 ierr = -1;
4459 alloc_failed |= (1 << szc);
4460 break;
4461 }
4462
4463 SEGVN_VMSTAT_FLTVNPAGES(39);
4464
4465 if (pplist != NULL) {
4466 segvn_relocate_pages(ppa, pplist);
4467 #ifdef DEBUG
4468 } else {
4469 ASSERT(type == F_SOFTLOCK);
4470 SEGVN_VMSTAT_FLTVNPAGES(40);
4471 #endif /* DEBUG */
4472 }
4473
4474 SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4475
4476 if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4477 ASSERT(type == F_SOFTLOCK);
4478 for (i = 0; i < pages; i++) {
4479 ASSERT(ppa[i]->p_szc < szc);
4480 hat_memload_region(hat,
4481 a + (i << PAGESHIFT),
4482 ppa[i], prot & vpprot, hat_flag,
4483 svd->rcookie);
4484 }
4485 } else {
4486 ASSERT(pplist != NULL || type == F_SOFTLOCK);
4487 hat_memload_array_region(hat, a, pgsz, ppa,
4488 prot & vpprot, hat_flag, svd->rcookie);
4489 }
4490 if (!(hat_flag & HAT_LOAD_LOCK)) {
4491 for (i = 0; i < pages; i++) {
4492 ASSERT(PAGE_SHARED(ppa[i]));
4493 page_unlock(ppa[i]);
4494 }
4495 }
4496 if (amp != NULL) {
4497 anon_array_exit(&an_cookie);
4498 ANON_LOCK_EXIT(&->a_rwlock);
4499 }
4500
4501 next:
4502 if (vpage != NULL) {
4503 vpage += pages;
4504 }
4505 adjszc_chk = 1;
4506 }
4507 if (a == lpgeaddr)
4508 break;
4509 ASSERT(a < lpgeaddr);
4510
4511 ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4512
4513 /*
4514 * ierr == -1 means we failed to map with a large page.
4515 * (either due to allocation/relocation failures or
4516 * misalignment with other mappings to this file.
4517 *
4518 * ierr == -2 means some other thread allocated a large page
4519 * after we gave up tp map with a large page. retry with
4520 * larger mapping.
4521 */
4522 ASSERT(ierr == -1 || ierr == -2);
4523 ASSERT(ierr == -2 || szc != 0);
4524 ASSERT(ierr == -1 || szc < seg->s_szc);
4525 if (ierr == -2) {
4526 SEGVN_VMSTAT_FLTVNPAGES(41);
4527 ASSERT(pszc > szc && pszc <= seg->s_szc);
4528 szc = pszc;
4529 } else if (segvn_anypgsz_vnode) {
4530 SEGVN_VMSTAT_FLTVNPAGES(42);
4531 szc--;
4532 } else {
4533 SEGVN_VMSTAT_FLTVNPAGES(43);
4534 ASSERT(pszc < szc);
4535 /*
4536 * other process created pszc large page.
4537 * but we still have to drop to 0 szc.
4538 */
4539 szc = 0;
4540 }
4541
4542 pgsz = page_get_pagesize(szc);
4543 pages = btop(pgsz);
4544 if (ierr == -2) {
4545 /*
4546 * Size up case. Note lpgaddr may only be needed for
4547 * softlock case so we don't adjust it here.
4548 */
4549 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4550 ASSERT(a >= lpgaddr);
4551 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4552 off = svd->offset + (uintptr_t)(a - seg->s_base);
4553 aindx = svd->anon_index + seg_page(seg, a);
4554 vpage = (svd->vpage != NULL) ?
4555 &svd->vpage[seg_page(seg, a)] : NULL;
4556 } else {
4557 /*
4558 * Size down case. Note lpgaddr may only be needed for
4559 * softlock case so we don't adjust it here.
4560 */
4561 ASSERT(IS_P2ALIGNED(a, pgsz));
4562 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4563 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4564 ASSERT(a < lpgeaddr);
4565 if (a < addr) {
4566 SEGVN_VMSTAT_FLTVNPAGES(44);
4567 /*
4568 * The beginning of the large page region can
4569 * be pulled to the right to make a smaller
4570 * region. We haven't yet faulted a single
4571 * page.
4572 */
4573 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4574 ASSERT(a >= lpgaddr);
4575 off = svd->offset +
4576 (uintptr_t)(a - seg->s_base);
4577 aindx = svd->anon_index + seg_page(seg, a);
4578 vpage = (svd->vpage != NULL) ?
4579 &svd->vpage[seg_page(seg, a)] : NULL;
4580 }
4581 }
4582 }
4583 out:
4584 kmem_free(ppa, ppasize);
4585 if (!err && !vop_size_err) {
4586 SEGVN_VMSTAT_FLTVNPAGES(45);
4587 return (0);
4588 }
4589 if (type == F_SOFTLOCK && a > lpgaddr) {
4590 SEGVN_VMSTAT_FLTVNPAGES(46);
4591 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4592 }
4593 if (!vop_size_err) {
4594 SEGVN_VMSTAT_FLTVNPAGES(47);
4595 return (err);
4596 }
4597 ASSERT(brkcow || tron || type == F_SOFTLOCK);
4598 /*
4599 * Large page end is mapped beyond the end of file and it's a cow
4600 * fault (can be a text replication induced cow) or softlock so we can't
4601 * reduce the map area. For now just demote the segment. This should
4602 * really only happen if the end of the file changed after the mapping
4603 * was established since when large page segments are created we make
4604 * sure they don't extend beyond the end of the file.
4605 */
4606 SEGVN_VMSTAT_FLTVNPAGES(48);
4607
4608 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4609 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4610 err = 0;
4611 if (seg->s_szc != 0) {
4612 segvn_fltvnpages_clrszc_cnt++;
4613 ASSERT(svd->softlockcnt == 0);
4614 err = segvn_clrszc(seg);
4615 if (err != 0) {
4616 segvn_fltvnpages_clrszc_err++;
4617 }
4618 }
4619 ASSERT(err || seg->s_szc == 0);
4620 SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4621 /* segvn_fault will do its job as if szc had been zero to begin with */
4622 return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4623 }
4624
4625 /*
4626 * This routine will attempt to fault in one large page.
4627 * it will use smaller pages if that fails.
4628 * It should only be called for pure anonymous segments.
4629 */
4630 static faultcode_t
4631 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4632 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4633 caddr_t eaddr, int brkcow)
4634 {
4635 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4636 struct anon_map *amp = svd->amp;
4637 uchar_t segtype = svd->type;
4638 uint_t szc = seg->s_szc;
4639 size_t pgsz = page_get_pagesize(szc);
4640 size_t maxpgsz = pgsz;
4641 pgcnt_t pages = btop(pgsz);
4642 uint_t ppaszc = szc;
4643 caddr_t a = lpgaddr;
4644 ulong_t aindx = svd->anon_index + seg_page(seg, a);
4645 struct vpage *vpage = (svd->vpage != NULL) ?
4646 &svd->vpage[seg_page(seg, a)] : NULL;
4647 page_t **ppa;
4648 uint_t ppa_szc;
4649 faultcode_t err;
4650 int ierr;
4651 uint_t protchk, prot, vpprot;
4652 ulong_t i;
4653 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4654 anon_sync_obj_t cookie;
4655 int adjszc_chk;
4656 int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4657
4658 ASSERT(szc != 0);
4659 ASSERT(amp != NULL);
4660 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4661 ASSERT(!(svd->flags & MAP_NORESERVE));
4662 ASSERT(type != F_SOFTUNLOCK);
4663 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4664 ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4665 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4666
4667 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4668
4669 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4670 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4671
4672 if (svd->flags & MAP_TEXT) {
4673 hat_flag |= HAT_LOAD_TEXT;
4674 }
4675
4676 if (svd->pageprot) {
4677 switch (rw) {
4678 case S_READ:
4679 protchk = PROT_READ;
4680 break;
4681 case S_WRITE:
4682 protchk = PROT_WRITE;
4683 break;
4684 case S_EXEC:
4685 protchk = PROT_EXEC;
4686 break;
4687 case S_OTHER:
4688 default:
4689 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4690 break;
4691 }
4692 VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4693 } else {
4694 prot = svd->prot;
4695 /* caller has already done segment level protection check. */
4696 }
4697
4698 ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4699 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
4700 for (;;) {
4701 adjszc_chk = 0;
4702 for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4703 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4704 VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4705 ASSERT(vpage != NULL);
4706 prot = VPP_PROT(vpage);
4707 ASSERT(sameprot(seg, a, maxpgsz));
4708 if ((prot & protchk) == 0) {
4709 err = FC_PROT;
4710 goto error;
4711 }
4712 }
4713 if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4714 pgsz < maxpgsz) {
4715 ASSERT(a > lpgaddr);
4716 szc = seg->s_szc;
4717 pgsz = maxpgsz;
4718 pages = btop(pgsz);
4719 ASSERT(IS_P2ALIGNED(aindx, pages));
4720 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4721 pgsz);
4722 }
4723 if (type == F_SOFTLOCK) {
4724 atomic_add_long((ulong_t *)&svd->softlockcnt,
4725 pages);
4726 }
4727 anon_array_enter(amp, aindx, &cookie);
4728 ppa_szc = (uint_t)-1;
4729 ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4730 prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4731 segvn_anypgsz, pgflags, svd->cred);
4732 if (ierr != 0) {
4733 anon_array_exit(&cookie);
4734 VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4735 if (type == F_SOFTLOCK) {
4736 atomic_add_long(
4737 (ulong_t *)&svd->softlockcnt,
4738 -pages);
4739 }
4740 if (ierr > 0) {
4741 VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4742 err = FC_MAKE_ERR(ierr);
4743 goto error;
4744 }
4745 break;
4746 }
4747
4748 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4749
4750 ASSERT(segtype == MAP_SHARED ||
4751 ppa[0]->p_szc <= szc);
4752 ASSERT(segtype == MAP_PRIVATE ||
4753 ppa[0]->p_szc >= szc);
4754
4755 /*
4756 * Handle pages that have been marked for migration
4757 */
4758 if (lgrp_optimizations())
4759 page_migrate(seg, a, ppa, pages);
4760
4761 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4762
4763 if (segtype == MAP_SHARED) {
4764 vpprot |= PROT_WRITE;
4765 }
4766
4767 hat_memload_array(hat, a, pgsz, ppa,
4768 prot & vpprot, hat_flag);
4769
4770 if (hat_flag & HAT_LOAD_LOCK) {
4771 VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4772 } else {
4773 VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4774 for (i = 0; i < pages; i++)
4775 page_unlock(ppa[i]);
4776 }
4777 if (vpage != NULL)
4778 vpage += pages;
4779
4780 anon_array_exit(&cookie);
4781 adjszc_chk = 1;
4782 }
4783 if (a == lpgeaddr)
4784 break;
4785 ASSERT(a < lpgeaddr);
4786 /*
4787 * ierr == -1 means we failed to allocate a large page.
4788 * so do a size down operation.
4789 *
4790 * ierr == -2 means some other process that privately shares
4791 * pages with this process has allocated a larger page and we
4792 * need to retry with larger pages. So do a size up
4793 * operation. This relies on the fact that large pages are
4794 * never partially shared i.e. if we share any constituent
4795 * page of a large page with another process we must share the
4796 * entire large page. Note this cannot happen for SOFTLOCK
4797 * case, unless current address (a) is at the beginning of the
4798 * next page size boundary because the other process couldn't
4799 * have relocated locked pages.
4800 */
4801 ASSERT(ierr == -1 || ierr == -2);
4802
4803 if (segvn_anypgsz) {
4804 ASSERT(ierr == -2 || szc != 0);
4805 ASSERT(ierr == -1 || szc < seg->s_szc);
4806 szc = (ierr == -1) ? szc - 1 : szc + 1;
4807 } else {
4808 /*
4809 * For non COW faults and segvn_anypgsz == 0
4810 * we need to be careful not to loop forever
4811 * if existing page is found with szc other
4812 * than 0 or seg->s_szc. This could be due
4813 * to page relocations on behalf of DR or
4814 * more likely large page creation. For this
4815 * case simply re-size to existing page's szc
4816 * if returned by anon_map_getpages().
4817 */
4818 if (ppa_szc == (uint_t)-1) {
4819 szc = (ierr == -1) ? 0 : seg->s_szc;
4820 } else {
4821 ASSERT(ppa_szc <= seg->s_szc);
4822 ASSERT(ierr == -2 || ppa_szc < szc);
4823 ASSERT(ierr == -1 || ppa_szc > szc);
4824 szc = ppa_szc;
4825 }
4826 }
4827
4828 pgsz = page_get_pagesize(szc);
4829 pages = btop(pgsz);
4830 ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4831 (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4832 if (type == F_SOFTLOCK) {
4833 /*
4834 * For softlocks we cannot reduce the fault area
4835 * (calculated based on the largest page size for this
4836 * segment) for size down and a is already next
4837 * page size aligned as assertted above for size
4838 * ups. Therefore just continue in case of softlock.
4839 */
4840 VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4841 continue; /* keep lint happy */
4842 } else if (ierr == -2) {
4843
4844 /*
4845 * Size up case. Note lpgaddr may only be needed for
4846 * softlock case so we don't adjust it here.
4847 */
4848 VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4849 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4850 ASSERT(a >= lpgaddr);
4851 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4852 aindx = svd->anon_index + seg_page(seg, a);
4853 vpage = (svd->vpage != NULL) ?
4854 &svd->vpage[seg_page(seg, a)] : NULL;
4855 } else {
4856 /*
4857 * Size down case. Note lpgaddr may only be needed for
4858 * softlock case so we don't adjust it here.
4859 */
4860 VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4861 ASSERT(IS_P2ALIGNED(a, pgsz));
4862 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4863 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4864 ASSERT(a < lpgeaddr);
4865 if (a < addr) {
4866 /*
4867 * The beginning of the large page region can
4868 * be pulled to the right to make a smaller
4869 * region. We haven't yet faulted a single
4870 * page.
4871 */
4872 VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4873 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4874 ASSERT(a >= lpgaddr);
4875 aindx = svd->anon_index + seg_page(seg, a);
4876 vpage = (svd->vpage != NULL) ?
4877 &svd->vpage[seg_page(seg, a)] : NULL;
4878 }
4879 }
4880 }
4881 VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4882 ANON_LOCK_EXIT(&->a_rwlock);
4883 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4884 return (0);
4885 error:
4886 VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4887 ANON_LOCK_EXIT(&->a_rwlock);
4888 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4889 if (type == F_SOFTLOCK && a > lpgaddr) {
4890 VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4891 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4892 }
4893 return (err);
4894 }
4895
4896 int fltadvice = 1; /* set to free behind pages for sequential access */
4897
4898 /*
4899 * This routine is called via a machine specific fault handling routine.
4900 * It is also called by software routines wishing to lock or unlock
4901 * a range of addresses.
4902 *
4903 * Here is the basic algorithm:
4904 * If unlocking
4905 * Call segvn_softunlock
4906 * Return
4907 * endif
4908 * Checking and set up work
4909 * If we will need some non-anonymous pages
4910 * Call VOP_GETPAGE over the range of non-anonymous pages
4911 * endif
4912 * Loop over all addresses requested
4913 * Call segvn_faultpage passing in page list
4914 * to load up translations and handle anonymous pages
4915 * endloop
4916 * Load up translation to any additional pages in page list not
4917 * already handled that fit into this segment
4918 */
4919 static faultcode_t
4920 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4921 enum fault_type type, enum seg_rw rw)
4922 {
4923 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4924 page_t **plp, **ppp, *pp;
4925 u_offset_t off;
4926 caddr_t a;
4927 struct vpage *vpage;
4928 uint_t vpprot, prot;
4929 int err;
4930 page_t *pl[PVN_GETPAGE_NUM + 1];
4931 size_t plsz, pl_alloc_sz;
4932 size_t page;
4933 ulong_t anon_index;
4934 struct anon_map *amp;
4935 int dogetpage = 0;
4936 caddr_t lpgaddr, lpgeaddr;
4937 size_t pgsz;
4938 anon_sync_obj_t cookie;
4939 int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4940
4941 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
4942 ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4943
4944 /*
4945 * First handle the easy stuff
4946 */
4947 if (type == F_SOFTUNLOCK) {
4948 if (rw == S_READ_NOCOW) {
4949 rw = S_READ;
4950 ASSERT(AS_WRITE_HELD(seg->s_as));
4951 }
4952 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4953 pgsz = (seg->s_szc == 0) ? PAGESIZE :
4954 page_get_pagesize(seg->s_szc);
4955 VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4956 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4957 segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4958 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4959 return (0);
4960 }
4961
4962 ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4963 !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4964 if (brkcow == 0) {
4965 if (svd->tr_state == SEGVN_TR_INIT) {
4966 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4967 if (svd->tr_state == SEGVN_TR_INIT) {
4968 ASSERT(svd->vp != NULL && svd->amp == NULL);
4969 ASSERT(svd->flags & MAP_TEXT);
4970 ASSERT(svd->type == MAP_PRIVATE);
4971 segvn_textrepl(seg);
4972 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4973 ASSERT(svd->tr_state != SEGVN_TR_ON ||
4974 svd->amp != NULL);
4975 }
4976 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4977 }
4978 } else if (svd->tr_state != SEGVN_TR_OFF) {
4979 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4980
4981 if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4982 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4983 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4984 return (FC_PROT);
4985 }
4986
4987 if (svd->tr_state == SEGVN_TR_ON) {
4988 ASSERT(svd->vp != NULL && svd->amp != NULL);
4989 segvn_textunrepl(seg, 0);
4990 ASSERT(svd->amp == NULL &&
4991 svd->tr_state == SEGVN_TR_OFF);
4992 } else if (svd->tr_state != SEGVN_TR_OFF) {
4993 svd->tr_state = SEGVN_TR_OFF;
4994 }
4995 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4996 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4997 }
4998
4999 top:
5000 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5001
5002 /*
5003 * If we have the same protections for the entire segment,
5004 * insure that the access being attempted is legitimate.
5005 */
5006
5007 if (svd->pageprot == 0) {
5008 uint_t protchk;
5009
5010 switch (rw) {
5011 case S_READ:
5012 case S_READ_NOCOW:
5013 protchk = PROT_READ;
5014 break;
5015 case S_WRITE:
5016 protchk = PROT_WRITE;
5017 break;
5018 case S_EXEC:
5019 protchk = PROT_EXEC;
5020 break;
5021 case S_OTHER:
5022 default:
5023 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
5024 break;
5025 }
5026
5027 if ((svd->prot & protchk) == 0) {
5028 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5029 return (FC_PROT); /* illegal access type */
5030 }
5031 }
5032
5033 if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5034 /* this must be SOFTLOCK S_READ fault */
5035 ASSERT(svd->amp == NULL);
5036 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5037 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5038 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5039 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5040 /*
5041 * this must be the first ever non S_READ_NOCOW
5042 * softlock for this segment.
5043 */
5044 ASSERT(svd->softlockcnt == 0);
5045 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5046 HAT_REGION_TEXT);
5047 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5048 }
5049 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5050 goto top;
5051 }
5052
5053 /*
5054 * We can't allow the long term use of softlocks for vmpss segments,
5055 * because in some file truncation cases we should be able to demote
5056 * the segment, which requires that there are no softlocks. The
5057 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5058 * segment is S_READ_NOCOW, where the caller holds the address space
5059 * locked as writer and calls softunlock before dropping the as lock.
5060 * S_READ_NOCOW is used by /proc to read memory from another user.
5061 *
5062 * Another deadlock between SOFTLOCK and file truncation can happen
5063 * because segvn_fault_vnodepages() calls the FS one pagesize at
5064 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5065 * can cause a deadlock because the first set of page_t's remain
5066 * locked SE_SHARED. To avoid this, we demote segments on a first
5067 * SOFTLOCK if they have a length greater than the segment's
5068 * page size.
5069 *
5070 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5071 * the access type is S_READ_NOCOW and the fault length is less than
5072 * or equal to the segment's page size. While this is quite restrictive,
5073 * it should be the most common case of SOFTLOCK against a vmpss
5074 * segment.
5075 *
5076 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5077 * caller makes sure no COW will be caused by another thread for a
5078 * softlocked page.
5079 */
5080 if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5081 int demote = 0;
5082
5083 if (rw != S_READ_NOCOW) {
5084 demote = 1;
5085 }
5086 if (!demote && len > PAGESIZE) {
5087 pgsz = page_get_pagesize(seg->s_szc);
5088 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5089 lpgeaddr);
5090 if (lpgeaddr - lpgaddr > pgsz) {
5091 demote = 1;
5092 }
5093 }
5094
5095 ASSERT(demote || AS_WRITE_HELD(seg->s_as));
5096
5097 if (demote) {
5098 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5099 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5100 if (seg->s_szc != 0) {
5101 segvn_vmpss_clrszc_cnt++;
5102 ASSERT(svd->softlockcnt == 0);
5103 err = segvn_clrszc(seg);
5104 if (err) {
5105 segvn_vmpss_clrszc_err++;
5106 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5107 return (FC_MAKE_ERR(err));
5108 }
5109 }
5110 ASSERT(seg->s_szc == 0);
5111 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5112 goto top;
5113 }
5114 }
5115
5116 /*
5117 * Check to see if we need to allocate an anon_map structure.
5118 */
5119 if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5120 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5121 /*
5122 * Drop the "read" lock on the segment and acquire
5123 * the "write" version since we have to allocate the
5124 * anon_map.
5125 */
5126 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5127 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5128
5129 if (svd->amp == NULL) {
5130 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5131 svd->amp->a_szc = seg->s_szc;
5132 }
5133 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5134
5135 /*
5136 * Start all over again since segment protections
5137 * may have changed after we dropped the "read" lock.
5138 */
5139 goto top;
5140 }
5141
5142 /*
5143 * S_READ_NOCOW vs S_READ distinction was
5144 * only needed for the code above. After
5145 * that we treat it as S_READ.
5146 */
5147 if (rw == S_READ_NOCOW) {
5148 ASSERT(type == F_SOFTLOCK);
5149 ASSERT(AS_WRITE_HELD(seg->s_as));
5150 rw = S_READ;
5151 }
5152
5153 amp = svd->amp;
5154
5155 /*
5156 * MADV_SEQUENTIAL work is ignored for large page segments.
5157 */
5158 if (seg->s_szc != 0) {
5159 pgsz = page_get_pagesize(seg->s_szc);
5160 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5161 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5162 if (svd->vp == NULL) {
5163 err = segvn_fault_anonpages(hat, seg, lpgaddr,
5164 lpgeaddr, type, rw, addr, addr + len, brkcow);
5165 } else {
5166 err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5167 lpgeaddr, type, rw, addr, addr + len, brkcow);
5168 if (err == IE_RETRY) {
5169 ASSERT(seg->s_szc == 0);
5170 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5171 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5172 goto top;
5173 }
5174 }
5175 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5176 return (err);
5177 }
5178
5179 page = seg_page(seg, addr);
5180 if (amp != NULL) {
5181 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5182 anon_index = svd->anon_index + page;
5183
5184 if (type == F_PROT && rw == S_READ &&
5185 svd->tr_state == SEGVN_TR_OFF &&
5186 svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5187 size_t index = anon_index;
5188 struct anon *ap;
5189
5190 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5191 /*
5192 * The fast path could apply to S_WRITE also, except
5193 * that the protection fault could be caused by lazy
5194 * tlb flush when ro->rw. In this case, the pte is
5195 * RW already. But RO in the other cpu's tlb causes
5196 * the fault. Since hat_chgprot won't do anything if
5197 * pte doesn't change, we may end up faulting
5198 * indefinitely until the RO tlb entry gets replaced.
5199 */
5200 for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5201 anon_array_enter(amp, index, &cookie);
5202 ap = anon_get_ptr(amp->ahp, index);
5203 anon_array_exit(&cookie);
5204 if ((ap == NULL) || (ap->an_refcnt != 1)) {
5205 ANON_LOCK_EXIT(&->a_rwlock);
5206 goto slow;
5207 }
5208 }
5209 hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5210 ANON_LOCK_EXIT(&->a_rwlock);
5211 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5212 return (0);
5213 }
5214 }
5215 slow:
5216
5217 if (svd->vpage == NULL)
5218 vpage = NULL;
5219 else
5220 vpage = &svd->vpage[page];
5221
5222 off = svd->offset + (uintptr_t)(addr - seg->s_base);
5223
5224 /*
5225 * If MADV_SEQUENTIAL has been set for the particular page we
5226 * are faulting on, free behind all pages in the segment and put
5227 * them on the free list.
5228 */
5229
5230 if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5231 struct vpage *vpp;
5232 ulong_t fanon_index;
5233 size_t fpage;
5234 u_offset_t pgoff, fpgoff;
5235 struct vnode *fvp;
5236 struct anon *fap = NULL;
5237
5238 if (svd->advice == MADV_SEQUENTIAL ||
5239 (svd->pageadvice &&
5240 VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5241 pgoff = off - PAGESIZE;
5242 fpage = page - 1;
5243 if (vpage != NULL)
5244 vpp = &svd->vpage[fpage];
5245 if (amp != NULL)
5246 fanon_index = svd->anon_index + fpage;
5247
5248 while (pgoff > svd->offset) {
5249 if (svd->advice != MADV_SEQUENTIAL &&
5250 (!svd->pageadvice || (vpage &&
5251 VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5252 break;
5253
5254 /*
5255 * If this is an anon page, we must find the
5256 * correct <vp, offset> for it
5257 */
5258 fap = NULL;
5259 if (amp != NULL) {
5260 ANON_LOCK_ENTER(&->a_rwlock,
5261 RW_READER);
5262 anon_array_enter(amp, fanon_index,
5263 &cookie);
5264 fap = anon_get_ptr(amp->ahp,
5265 fanon_index);
5266 if (fap != NULL) {
5267 swap_xlate(fap, &fvp, &fpgoff);
5268 } else {
5269 fpgoff = pgoff;
5270 fvp = svd->vp;
5271 }
5272 anon_array_exit(&cookie);
5273 ANON_LOCK_EXIT(&->a_rwlock);
5274 } else {
5275 fpgoff = pgoff;
5276 fvp = svd->vp;
5277 }
5278 if (fvp == NULL)
5279 break; /* XXX */
5280 /*
5281 * Skip pages that are free or have an
5282 * "exclusive" lock.
5283 */
5284 pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5285 if (pp == NULL)
5286 break;
5287 /*
5288 * We don't need the page_struct_lock to test
5289 * as this is only advisory; even if we
5290 * acquire it someone might race in and lock
5291 * the page after we unlock and before the
5292 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5293 */
5294 if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5295 /*
5296 * Hold the vnode before releasing
5297 * the page lock to prevent it from
5298 * being freed and re-used by some
5299 * other thread.
5300 */
5301 VN_HOLD(fvp);
5302 page_unlock(pp);
5303 /*
5304 * We should build a page list
5305 * to kluster putpages XXX
5306 */
5307 (void) VOP_PUTPAGE(fvp,
5308 (offset_t)fpgoff, PAGESIZE,
5309 (B_DONTNEED|B_FREE|B_ASYNC),
5310 svd->cred, NULL);
5311 VN_RELE(fvp);
5312 } else {
5313 /*
5314 * XXX - Should the loop terminate if
5315 * the page is `locked'?
5316 */
5317 page_unlock(pp);
5318 }
5319 --vpp;
5320 --fanon_index;
5321 pgoff -= PAGESIZE;
5322 }
5323 }
5324 }
5325
5326 plp = pl;
5327 *plp = NULL;
5328 pl_alloc_sz = 0;
5329
5330 /*
5331 * See if we need to call VOP_GETPAGE for
5332 * *any* of the range being faulted on.
5333 * We can skip all of this work if there
5334 * was no original vnode.
5335 */
5336 if (svd->vp != NULL) {
5337 u_offset_t vp_off;
5338 size_t vp_len;
5339 struct anon *ap;
5340 vnode_t *vp;
5341
5342 vp_off = off;
5343 vp_len = len;
5344
5345 if (amp == NULL)
5346 dogetpage = 1;
5347 else {
5348 /*
5349 * Only acquire reader lock to prevent amp->ahp
5350 * from being changed. It's ok to miss pages,
5351 * hence we don't do anon_array_enter
5352 */
5353 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5354 ap = anon_get_ptr(amp->ahp, anon_index);
5355
5356 if (len <= PAGESIZE)
5357 /* inline non_anon() */
5358 dogetpage = (ap == NULL);
5359 else
5360 dogetpage = non_anon(amp->ahp, anon_index,
5361 &vp_off, &vp_len);
5362 ANON_LOCK_EXIT(&->a_rwlock);
5363 }
5364
5365 if (dogetpage) {
5366 enum seg_rw arw;
5367 struct as *as = seg->s_as;
5368
5369 if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5370 /*
5371 * Page list won't fit in local array,
5372 * allocate one of the needed size.
5373 */
5374 pl_alloc_sz =
5375 (btop(len) + 1) * sizeof (page_t *);
5376 plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5377 plp[0] = NULL;
5378 plsz = len;
5379 } else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5380 svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5381 (((size_t)(addr + PAGESIZE) <
5382 (size_t)(seg->s_base + seg->s_size)) &&
5383 hat_probe(as->a_hat, addr + PAGESIZE))) {
5384 /*
5385 * Ask VOP_GETPAGE to return the exact number
5386 * of pages if
5387 * (a) this is a COW fault, or
5388 * (b) this is a software fault, or
5389 * (c) next page is already mapped.
5390 */
5391 plsz = len;
5392 } else {
5393 /*
5394 * Ask VOP_GETPAGE to return adjacent pages
5395 * within the segment.
5396 */
5397 plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5398 ((seg->s_base + seg->s_size) - addr));
5399 ASSERT((addr + plsz) <=
5400 (seg->s_base + seg->s_size));
5401 }
5402
5403 /*
5404 * Need to get some non-anonymous pages.
5405 * We need to make only one call to GETPAGE to do
5406 * this to prevent certain deadlocking conditions
5407 * when we are doing locking. In this case
5408 * non_anon() should have picked up the smallest
5409 * range which includes all the non-anonymous
5410 * pages in the requested range. We have to
5411 * be careful regarding which rw flag to pass in
5412 * because on a private mapping, the underlying
5413 * object is never allowed to be written.
5414 */
5415 if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5416 arw = S_READ;
5417 } else {
5418 arw = rw;
5419 }
5420 vp = svd->vp;
5421 TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5422 "segvn_getpage:seg %p addr %p vp %p",
5423 seg, addr, vp);
5424 err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5425 &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5426 svd->cred, NULL);
5427 if (err) {
5428 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5429 segvn_pagelist_rele(plp);
5430 if (pl_alloc_sz)
5431 kmem_free(plp, pl_alloc_sz);
5432 return (FC_MAKE_ERR(err));
5433 }
5434 if (svd->type == MAP_PRIVATE)
5435 vpprot &= ~PROT_WRITE;
5436 }
5437 }
5438
5439 /*
5440 * N.B. at this time the plp array has all the needed non-anon
5441 * pages in addition to (possibly) having some adjacent pages.
5442 */
5443
5444 /*
5445 * Always acquire the anon_array_lock to prevent
5446 * 2 threads from allocating separate anon slots for
5447 * the same "addr".
5448 *
5449 * If this is a copy-on-write fault and we don't already
5450 * have the anon_array_lock, acquire it to prevent the
5451 * fault routine from handling multiple copy-on-write faults
5452 * on the same "addr" in the same address space.
5453 *
5454 * Only one thread should deal with the fault since after
5455 * it is handled, the other threads can acquire a translation
5456 * to the newly created private page. This prevents two or
5457 * more threads from creating different private pages for the
5458 * same fault.
5459 *
5460 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5461 * to prevent deadlock between this thread and another thread
5462 * which has soft-locked this page and wants to acquire serial_lock.
5463 * ( bug 4026339 )
5464 *
5465 * The fix for bug 4026339 becomes unnecessary when using the
5466 * locking scheme with per amp rwlock and a global set of hash
5467 * lock, anon_array_lock. If we steal a vnode page when low
5468 * on memory and upgrad the page lock through page_rename,
5469 * then the page is PAGE_HANDLED, nothing needs to be done
5470 * for this page after returning from segvn_faultpage.
5471 *
5472 * But really, the page lock should be downgraded after
5473 * the stolen page is page_rename'd.
5474 */
5475
5476 if (amp != NULL)
5477 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5478
5479 /*
5480 * Ok, now loop over the address range and handle faults
5481 */
5482 for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5483 err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5484 type, rw, brkcow);
5485 if (err) {
5486 if (amp != NULL)
5487 ANON_LOCK_EXIT(&->a_rwlock);
5488 if (type == F_SOFTLOCK && a > addr) {
5489 segvn_softunlock(seg, addr, (a - addr),
5490 S_OTHER);
5491 }
5492 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5493 segvn_pagelist_rele(plp);
5494 if (pl_alloc_sz)
5495 kmem_free(plp, pl_alloc_sz);
5496 return (err);
5497 }
5498 if (vpage) {
5499 vpage++;
5500 } else if (svd->vpage) {
5501 page = seg_page(seg, addr);
5502 vpage = &svd->vpage[++page];
5503 }
5504 }
5505
5506 /* Didn't get pages from the underlying fs so we're done */
5507 if (!dogetpage)
5508 goto done;
5509
5510 /*
5511 * Now handle any other pages in the list returned.
5512 * If the page can be used, load up the translations now.
5513 * Note that the for loop will only be entered if "plp"
5514 * is pointing to a non-NULL page pointer which means that
5515 * VOP_GETPAGE() was called and vpprot has been initialized.
5516 */
5517 if (svd->pageprot == 0)
5518 prot = svd->prot & vpprot;
5519
5520
5521 /*
5522 * Large Files: diff should be unsigned value because we started
5523 * supporting > 2GB segment sizes from 2.5.1 and when a
5524 * large file of size > 2GB gets mapped to address space
5525 * the diff value can be > 2GB.
5526 */
5527
5528 for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5529 size_t diff;
5530 struct anon *ap;
5531 int anon_index;
5532 anon_sync_obj_t cookie;
5533 int hat_flag = HAT_LOAD_ADV;
5534
5535 if (svd->flags & MAP_TEXT) {
5536 hat_flag |= HAT_LOAD_TEXT;
5537 }
5538
5539 if (pp == PAGE_HANDLED)
5540 continue;
5541
5542 if (svd->tr_state != SEGVN_TR_ON &&
5543 pp->p_offset >= svd->offset &&
5544 pp->p_offset < svd->offset + seg->s_size) {
5545
5546 diff = pp->p_offset - svd->offset;
5547
5548 /*
5549 * Large Files: Following is the assertion
5550 * validating the above cast.
5551 */
5552 ASSERT(svd->vp == pp->p_vnode);
5553
5554 page = btop(diff);
5555 if (svd->pageprot)
5556 prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5557
5558 /*
5559 * Prevent other threads in the address space from
5560 * creating private pages (i.e., allocating anon slots)
5561 * while we are in the process of loading translations
5562 * to additional pages returned by the underlying
5563 * object.
5564 */
5565 if (amp != NULL) {
5566 anon_index = svd->anon_index + page;
5567 anon_array_enter(amp, anon_index, &cookie);
5568 ap = anon_get_ptr(amp->ahp, anon_index);
5569 }
5570 if ((amp == NULL) || (ap == NULL)) {
5571 if (IS_VMODSORT(pp->p_vnode) ||
5572 enable_mbit_wa) {
5573 if (rw == S_WRITE)
5574 hat_setmod(pp);
5575 else if (rw != S_OTHER &&
5576 !hat_ismod(pp))
5577 prot &= ~PROT_WRITE;
5578 }
5579 /*
5580 * Skip mapping read ahead pages marked
5581 * for migration, so they will get migrated
5582 * properly on fault
5583 */
5584 ASSERT(amp == NULL ||
5585 svd->rcookie == HAT_INVALID_REGION_COOKIE);
5586 if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5587 hat_memload_region(hat,
5588 seg->s_base + diff,
5589 pp, prot, hat_flag,
5590 svd->rcookie);
5591 }
5592 }
5593 if (amp != NULL)
5594 anon_array_exit(&cookie);
5595 }
5596 page_unlock(pp);
5597 }
5598 done:
5599 if (amp != NULL)
5600 ANON_LOCK_EXIT(&->a_rwlock);
5601 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5602 if (pl_alloc_sz)
5603 kmem_free(plp, pl_alloc_sz);
5604 return (0);
5605 }
5606
5607 /*
5608 * This routine is used to start I/O on pages asynchronously. XXX it will
5609 * only create PAGESIZE pages. At fault time they will be relocated into
5610 * larger pages.
5611 */
5612 static faultcode_t
5613 segvn_faulta(struct seg *seg, caddr_t addr)
5614 {
5615 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5616 int err;
5617 struct anon_map *amp;
5618 vnode_t *vp;
5619
5620 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5621
5622 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5623 if ((amp = svd->amp) != NULL) {
5624 struct anon *ap;
5625
5626 /*
5627 * Reader lock to prevent amp->ahp from being changed.
5628 * This is advisory, it's ok to miss a page, so
5629 * we don't do anon_array_enter lock.
5630 */
5631 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5632 if ((ap = anon_get_ptr(amp->ahp,
5633 svd->anon_index + seg_page(seg, addr))) != NULL) {
5634
5635 err = anon_getpage(&ap, NULL, NULL,
5636 0, seg, addr, S_READ, svd->cred);
5637
5638 ANON_LOCK_EXIT(&->a_rwlock);
5639 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5640 if (err)
5641 return (FC_MAKE_ERR(err));
5642 return (0);
5643 }
5644 ANON_LOCK_EXIT(&->a_rwlock);
5645 }
5646
5647 if (svd->vp == NULL) {
5648 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5649 return (0); /* zfod page - do nothing now */
5650 }
5651
5652 vp = svd->vp;
5653 TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5654 "segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5655 err = VOP_GETPAGE(vp,
5656 (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5657 PAGESIZE, NULL, NULL, 0, seg, addr,
5658 S_OTHER, svd->cred, NULL);
5659
5660 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5661 if (err)
5662 return (FC_MAKE_ERR(err));
5663 return (0);
5664 }
5665
5666 static int
5667 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5668 {
5669 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5670 struct vpage *cvp, *svp, *evp;
5671 struct vnode *vp;
5672 size_t pgsz;
5673 pgcnt_t pgcnt;
5674 anon_sync_obj_t cookie;
5675 int unload_done = 0;
5676
5677 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5678
5679 if ((svd->maxprot & prot) != prot)
5680 return (EACCES); /* violated maxprot */
5681
5682 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5683
5684 /* return if prot is the same */
5685 if (!svd->pageprot && svd->prot == prot) {
5686 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5687 return (0);
5688 }
5689
5690 /*
5691 * Since we change protections we first have to flush the cache.
5692 * This makes sure all the pagelock calls have to recheck
5693 * protections.
5694 */
5695 if (svd->softlockcnt > 0) {
5696 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5697
5698 /*
5699 * If this is shared segment non 0 softlockcnt
5700 * means locked pages are still in use.
5701 */
5702 if (svd->type == MAP_SHARED) {
5703 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5704 return (EAGAIN);
5705 }
5706
5707 /*
5708 * Since we do have the segvn writers lock nobody can fill
5709 * the cache with entries belonging to this seg during
5710 * the purge. The flush either succeeds or we still have
5711 * pending I/Os.
5712 */
5713 segvn_purge(seg);
5714 if (svd->softlockcnt > 0) {
5715 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5716 return (EAGAIN);
5717 }
5718 }
5719
5720 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5721 ASSERT(svd->amp == NULL);
5722 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5723 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5724 HAT_REGION_TEXT);
5725 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5726 unload_done = 1;
5727 } else if (svd->tr_state == SEGVN_TR_INIT) {
5728 svd->tr_state = SEGVN_TR_OFF;
5729 } else if (svd->tr_state == SEGVN_TR_ON) {
5730 ASSERT(svd->amp != NULL);
5731 segvn_textunrepl(seg, 0);
5732 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5733 unload_done = 1;
5734 }
5735
5736 if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5737 svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5738 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5739 segvn_inval_trcache(svd->vp);
5740 }
5741 if (seg->s_szc != 0) {
5742 int err;
5743 pgsz = page_get_pagesize(seg->s_szc);
5744 pgcnt = pgsz >> PAGESHIFT;
5745 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5746 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5747 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5748 ASSERT(seg->s_base != addr || seg->s_size != len);
5749 /*
5750 * If we are holding the as lock as a reader then
5751 * we need to return IE_RETRY and let the as
5752 * layer drop and re-acquire the lock as a writer.
5753 */
5754 if (AS_READ_HELD(seg->s_as))
5755 return (IE_RETRY);
5756 VM_STAT_ADD(segvnvmstats.demoterange[1]);
5757 if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5758 err = segvn_demote_range(seg, addr, len,
5759 SDR_END, 0);
5760 } else {
5761 uint_t szcvec = map_pgszcvec(seg->s_base,
5762 pgsz, (uintptr_t)seg->s_base,
5763 (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5764 err = segvn_demote_range(seg, addr, len,
5765 SDR_END, szcvec);
5766 }
5767 if (err == 0)
5768 return (IE_RETRY);
5769 if (err == ENOMEM)
5770 return (IE_NOMEM);
5771 return (err);
5772 }
5773 }
5774
5775
5776 /*
5777 * If it's a private mapping and we're making it writable then we
5778 * may have to reserve the additional swap space now. If we are
5779 * making writable only a part of the segment then we use its vpage
5780 * array to keep a record of the pages for which we have reserved
5781 * swap. In this case we set the pageswap field in the segment's
5782 * segvn structure to record this.
5783 *
5784 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5785 * removing write permission on the entire segment and we haven't
5786 * modified any pages, we can release the swap space.
5787 */
5788 if (svd->type == MAP_PRIVATE) {
5789 if (prot & PROT_WRITE) {
5790 if (!(svd->flags & MAP_NORESERVE) &&
5791 !(svd->swresv && svd->pageswap == 0)) {
5792 size_t sz = 0;
5793
5794 /*
5795 * Start by determining how much swap
5796 * space is required.
5797 */
5798 if (addr == seg->s_base &&
5799 len == seg->s_size &&
5800 svd->pageswap == 0) {
5801 /* The whole segment */
5802 sz = seg->s_size;
5803 } else {
5804 /*
5805 * Make sure that the vpage array
5806 * exists, and make a note of the
5807 * range of elements corresponding
5808 * to len.
5809 */
5810 segvn_vpage(seg);
5811 if (svd->vpage == NULL) {
5812 SEGVN_LOCK_EXIT(seg->s_as,
5813 &svd->lock);
5814 return (ENOMEM);
5815 }
5816 svp = &svd->vpage[seg_page(seg, addr)];
5817 evp = &svd->vpage[seg_page(seg,
5818 addr + len)];
5819
5820 if (svd->pageswap == 0) {
5821 /*
5822 * This is the first time we've
5823 * asked for a part of this
5824 * segment, so we need to
5825 * reserve everything we've
5826 * been asked for.
5827 */
5828 sz = len;
5829 } else {
5830 /*
5831 * We have to count the number
5832 * of pages required.
5833 */
5834 for (cvp = svp; cvp < evp;
5835 cvp++) {
5836 if (!VPP_ISSWAPRES(cvp))
5837 sz++;
5838 }
5839 sz <<= PAGESHIFT;
5840 }
5841 }
5842
5843 /* Try to reserve the necessary swap. */
5844 if (anon_resv_zone(sz,
5845 seg->s_as->a_proc->p_zone) == 0) {
5846 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5847 return (IE_NOMEM);
5848 }
5849
5850 /*
5851 * Make a note of how much swap space
5852 * we've reserved.
5853 */
5854 if (svd->pageswap == 0 && sz == seg->s_size) {
5855 svd->swresv = sz;
5856 } else {
5857 ASSERT(svd->vpage != NULL);
5858 svd->swresv += sz;
5859 svd->pageswap = 1;
5860 for (cvp = svp; cvp < evp; cvp++) {
5861 if (!VPP_ISSWAPRES(cvp))
5862 VPP_SETSWAPRES(cvp);
5863 }
5864 }
5865 }
5866 } else {
5867 /*
5868 * Swap space is released only if this segment
5869 * does not map anonymous memory, since read faults
5870 * on such segments still need an anon slot to read
5871 * in the data.
5872 */
5873 if (svd->swresv != 0 && svd->vp != NULL &&
5874 svd->amp == NULL && addr == seg->s_base &&
5875 len == seg->s_size && svd->pageprot == 0) {
5876 ASSERT(svd->pageswap == 0);
5877 anon_unresv_zone(svd->swresv,
5878 seg->s_as->a_proc->p_zone);
5879 svd->swresv = 0;
5880 TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5881 "anon proc:%p %lu %u", seg, 0, 0);
5882 }
5883 }
5884 }
5885
5886 if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5887 if (svd->prot == prot) {
5888 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5889 return (0); /* all done */
5890 }
5891 svd->prot = (uchar_t)prot;
5892 } else if (svd->type == MAP_PRIVATE) {
5893 struct anon *ap = NULL;
5894 page_t *pp;
5895 u_offset_t offset, off;
5896 struct anon_map *amp;
5897 ulong_t anon_idx = 0;
5898
5899 /*
5900 * A vpage structure exists or else the change does not
5901 * involve the entire segment. Establish a vpage structure
5902 * if none is there. Then, for each page in the range,
5903 * adjust its individual permissions. Note that write-
5904 * enabling a MAP_PRIVATE page can affect the claims for
5905 * locked down memory. Overcommitting memory terminates
5906 * the operation.
5907 */
5908 segvn_vpage(seg);
5909 if (svd->vpage == NULL) {
5910 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5911 return (ENOMEM);
5912 }
5913 svd->pageprot = 1;
5914 if ((amp = svd->amp) != NULL) {
5915 anon_idx = svd->anon_index + seg_page(seg, addr);
5916 ASSERT(seg->s_szc == 0 ||
5917 IS_P2ALIGNED(anon_idx, pgcnt));
5918 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
5919 }
5920
5921 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5922 evp = &svd->vpage[seg_page(seg, addr + len)];
5923
5924 /*
5925 * See Statement at the beginning of segvn_lockop regarding
5926 * the way cowcnts and lckcnts are handled.
5927 */
5928 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5929
5930 if (seg->s_szc != 0) {
5931 if (amp != NULL) {
5932 anon_array_enter(amp, anon_idx,
5933 &cookie);
5934 }
5935 if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5936 !segvn_claim_pages(seg, svp, offset,
5937 anon_idx, prot)) {
5938 if (amp != NULL) {
5939 anon_array_exit(&cookie);
5940 }
5941 break;
5942 }
5943 if (amp != NULL) {
5944 anon_array_exit(&cookie);
5945 }
5946 anon_idx++;
5947 } else {
5948 if (amp != NULL) {
5949 anon_array_enter(amp, anon_idx,
5950 &cookie);
5951 ap = anon_get_ptr(amp->ahp, anon_idx++);
5952 }
5953
5954 if (VPP_ISPPLOCK(svp) &&
5955 VPP_PROT(svp) != prot) {
5956
5957 if (amp == NULL || ap == NULL) {
5958 vp = svd->vp;
5959 off = offset;
5960 } else
5961 swap_xlate(ap, &vp, &off);
5962 if (amp != NULL)
5963 anon_array_exit(&cookie);
5964
5965 if ((pp = page_lookup(vp, off,
5966 SE_SHARED)) == NULL) {
5967 panic("segvn_setprot: no page");
5968 /*NOTREACHED*/
5969 }
5970 ASSERT(seg->s_szc == 0);
5971 if ((VPP_PROT(svp) ^ prot) &
5972 PROT_WRITE) {
5973 if (prot & PROT_WRITE) {
5974 if (!page_addclaim(
5975 pp)) {
5976 page_unlock(pp);
5977 break;
5978 }
5979 } else {
5980 if (!page_subclaim(
5981 pp)) {
5982 page_unlock(pp);
5983 break;
5984 }
5985 }
5986 }
5987 page_unlock(pp);
5988 } else if (amp != NULL)
5989 anon_array_exit(&cookie);
5990 }
5991 VPP_SETPROT(svp, prot);
5992 offset += PAGESIZE;
5993 }
5994 if (amp != NULL)
5995 ANON_LOCK_EXIT(&->a_rwlock);
5996
5997 /*
5998 * Did we terminate prematurely? If so, simply unload
5999 * the translations to the things we've updated so far.
6000 */
6001 if (svp != evp) {
6002 if (unload_done) {
6003 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6004 return (IE_NOMEM);
6005 }
6006 len = (svp - &svd->vpage[seg_page(seg, addr)]) *
6007 PAGESIZE;
6008 ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
6009 if (len != 0)
6010 hat_unload(seg->s_as->a_hat, addr,
6011 len, HAT_UNLOAD);
6012 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6013 return (IE_NOMEM);
6014 }
6015 } else {
6016 segvn_vpage(seg);
6017 if (svd->vpage == NULL) {
6018 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6019 return (ENOMEM);
6020 }
6021 svd->pageprot = 1;
6022 evp = &svd->vpage[seg_page(seg, addr + len)];
6023 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
6024 VPP_SETPROT(svp, prot);
6025 }
6026 }
6027
6028 if (unload_done) {
6029 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6030 return (0);
6031 }
6032
6033 if (((prot & PROT_WRITE) != 0 &&
6034 (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
6035 (prot & ~PROT_USER) == PROT_NONE) {
6036 /*
6037 * Either private or shared data with write access (in
6038 * which case we need to throw out all former translations
6039 * so that we get the right translations set up on fault
6040 * and we don't allow write access to any copy-on-write pages
6041 * that might be around or to prevent write access to pages
6042 * representing holes in a file), or we don't have permission
6043 * to access the memory at all (in which case we have to
6044 * unload any current translations that might exist).
6045 */
6046 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
6047 } else {
6048 /*
6049 * A shared mapping or a private mapping in which write
6050 * protection is going to be denied - just change all the
6051 * protections over the range of addresses in question.
6052 * segvn does not support any other attributes other
6053 * than prot so we can use hat_chgattr.
6054 */
6055 hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6056 }
6057
6058 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6059
6060 return (0);
6061 }
6062
6063 /*
6064 * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
6065 * to determine if the seg is capable of mapping the requested szc.
6066 */
6067 static int
6068 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6069 {
6070 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6071 struct segvn_data *nsvd;
6072 struct anon_map *amp = svd->amp;
6073 struct seg *nseg;
6074 caddr_t eaddr = addr + len, a;
6075 size_t pgsz = page_get_pagesize(szc);
6076 pgcnt_t pgcnt = page_get_pagecnt(szc);
6077 int err;
6078 u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6079
6080 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
6081 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6082
6083 if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6084 return (0);
6085 }
6086
6087 /*
6088 * addr should always be pgsz aligned but eaddr may be misaligned if
6089 * it's at the end of the segment.
6090 *
6091 * XXX we should assert this condition since as_setpagesize() logic
6092 * guarantees it.
6093 */
6094 if (!IS_P2ALIGNED(addr, pgsz) ||
6095 (!IS_P2ALIGNED(eaddr, pgsz) &&
6096 eaddr != seg->s_base + seg->s_size)) {
6097
6098 segvn_setpgsz_align_err++;
6099 return (EINVAL);
6100 }
6101
6102 if (amp != NULL && svd->type == MAP_SHARED) {
6103 ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6104 if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6105
6106 segvn_setpgsz_anon_align_err++;
6107 return (EINVAL);
6108 }
6109 }
6110
6111 if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6112 szc > segvn_maxpgszc) {
6113 return (EINVAL);
6114 }
6115
6116 /* paranoid check */
6117 if (svd->vp != NULL &&
6118 (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6119 return (EINVAL);
6120 }
6121
6122 if (seg->s_szc == 0 && svd->vp != NULL &&
6123 map_addr_vacalign_check(addr, off)) {
6124 return (EINVAL);
6125 }
6126
6127 /*
6128 * Check that protections are the same within new page
6129 * size boundaries.
6130 */
6131 if (svd->pageprot) {
6132 for (a = addr; a < eaddr; a += pgsz) {
6133 if ((a + pgsz) > eaddr) {
6134 if (!sameprot(seg, a, eaddr - a)) {
6135 return (EINVAL);
6136 }
6137 } else {
6138 if (!sameprot(seg, a, pgsz)) {
6139 return (EINVAL);
6140 }
6141 }
6142 }
6143 }
6144
6145 /*
6146 * Since we are changing page size we first have to flush
6147 * the cache. This makes sure all the pagelock calls have
6148 * to recheck protections.
6149 */
6150 if (svd->softlockcnt > 0) {
6151 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6152
6153 /*
6154 * If this is shared segment non 0 softlockcnt
6155 * means locked pages are still in use.
6156 */
6157 if (svd->type == MAP_SHARED) {
6158 return (EAGAIN);
6159 }
6160
6161 /*
6162 * Since we do have the segvn writers lock nobody can fill
6163 * the cache with entries belonging to this seg during
6164 * the purge. The flush either succeeds or we still have
6165 * pending I/Os.
6166 */
6167 segvn_purge(seg);
6168 if (svd->softlockcnt > 0) {
6169 return (EAGAIN);
6170 }
6171 }
6172
6173 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6174 ASSERT(svd->amp == NULL);
6175 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6176 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6177 HAT_REGION_TEXT);
6178 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6179 } else if (svd->tr_state == SEGVN_TR_INIT) {
6180 svd->tr_state = SEGVN_TR_OFF;
6181 } else if (svd->tr_state == SEGVN_TR_ON) {
6182 ASSERT(svd->amp != NULL);
6183 segvn_textunrepl(seg, 1);
6184 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6185 amp = NULL;
6186 }
6187
6188 /*
6189 * Operation for sub range of existing segment.
6190 */
6191 if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6192 if (szc < seg->s_szc) {
6193 VM_STAT_ADD(segvnvmstats.demoterange[2]);
6194 err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6195 if (err == 0) {
6196 return (IE_RETRY);
6197 }
6198 if (err == ENOMEM) {
6199 return (IE_NOMEM);
6200 }
6201 return (err);
6202 }
6203 if (addr != seg->s_base) {
6204 nseg = segvn_split_seg(seg, addr);
6205 if (eaddr != (nseg->s_base + nseg->s_size)) {
6206 /* eaddr is szc aligned */
6207 (void) segvn_split_seg(nseg, eaddr);
6208 }
6209 return (IE_RETRY);
6210 }
6211 if (eaddr != (seg->s_base + seg->s_size)) {
6212 /* eaddr is szc aligned */
6213 (void) segvn_split_seg(seg, eaddr);
6214 }
6215 return (IE_RETRY);
6216 }
6217
6218 /*
6219 * Break any low level sharing and reset seg->s_szc to 0.
6220 */
6221 if ((err = segvn_clrszc(seg)) != 0) {
6222 if (err == ENOMEM) {
6223 err = IE_NOMEM;
6224 }
6225 return (err);
6226 }
6227 ASSERT(seg->s_szc == 0);
6228
6229 /*
6230 * If the end of the current segment is not pgsz aligned
6231 * then attempt to concatenate with the next segment.
6232 */
6233 if (!IS_P2ALIGNED(eaddr, pgsz)) {
6234 nseg = AS_SEGNEXT(seg->s_as, seg);
6235 if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6236 return (ENOMEM);
6237 }
6238 if (nseg->s_ops != &segvn_ops) {
6239 return (EINVAL);
6240 }
6241 nsvd = (struct segvn_data *)nseg->s_data;
6242 if (nsvd->softlockcnt > 0) {
6243 /*
6244 * If this is shared segment non 0 softlockcnt
6245 * means locked pages are still in use.
6246 */
6247 if (nsvd->type == MAP_SHARED) {
6248 return (EAGAIN);
6249 }
6250 segvn_purge(nseg);
6251 if (nsvd->softlockcnt > 0) {
6252 return (EAGAIN);
6253 }
6254 }
6255 err = segvn_clrszc(nseg);
6256 if (err == ENOMEM) {
6257 err = IE_NOMEM;
6258 }
6259 if (err != 0) {
6260 return (err);
6261 }
6262 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6263 err = segvn_concat(seg, nseg, 1);
6264 if (err == -1) {
6265 return (EINVAL);
6266 }
6267 if (err == -2) {
6268 return (IE_NOMEM);
6269 }
6270 return (IE_RETRY);
6271 }
6272
6273 /*
6274 * May need to re-align anon array to
6275 * new szc.
6276 */
6277 if (amp != NULL) {
6278 if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6279 struct anon_hdr *nahp;
6280
6281 ASSERT(svd->type == MAP_PRIVATE);
6282
6283 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
6284 ASSERT(amp->refcnt == 1);
6285 nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6286 if (nahp == NULL) {
6287 ANON_LOCK_EXIT(&->a_rwlock);
6288 return (IE_NOMEM);
6289 }
6290 if (anon_copy_ptr(amp->ahp, svd->anon_index,
6291 nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6292 anon_release(nahp, btop(amp->size));
6293 ANON_LOCK_EXIT(&->a_rwlock);
6294 return (IE_NOMEM);
6295 }
6296 anon_release(amp->ahp, btop(amp->size));
6297 amp->ahp = nahp;
6298 svd->anon_index = 0;
6299 ANON_LOCK_EXIT(&->a_rwlock);
6300 }
6301 }
6302 if (svd->vp != NULL && szc != 0) {
6303 struct vattr va;
6304 u_offset_t eoffpage = svd->offset;
6305 va.va_mask = AT_SIZE;
6306 eoffpage += seg->s_size;
6307 eoffpage = btopr(eoffpage);
6308 if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6309 segvn_setpgsz_getattr_err++;
6310 return (EINVAL);
6311 }
6312 if (btopr(va.va_size) < eoffpage) {
6313 segvn_setpgsz_eof_err++;
6314 return (EINVAL);
6315 }
6316 if (amp != NULL) {
6317 /*
6318 * anon_fill_cow_holes() may call VOP_GETPAGE().
6319 * don't take anon map lock here to avoid holding it
6320 * across VOP_GETPAGE() calls that may call back into
6321 * segvn for klsutering checks. We don't really need
6322 * anon map lock here since it's a private segment and
6323 * we hold as level lock as writers.
6324 */
6325 if ((err = anon_fill_cow_holes(seg, seg->s_base,
6326 amp->ahp, svd->anon_index, svd->vp, svd->offset,
6327 seg->s_size, szc, svd->prot, svd->vpage,
6328 svd->cred)) != 0) {
6329 return (EINVAL);
6330 }
6331 }
6332 segvn_setvnode_mpss(svd->vp);
6333 }
6334
6335 if (amp != NULL) {
6336 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
6337 if (svd->type == MAP_PRIVATE) {
6338 amp->a_szc = szc;
6339 } else if (szc > amp->a_szc) {
6340 amp->a_szc = szc;
6341 }
6342 ANON_LOCK_EXIT(&->a_rwlock);
6343 }
6344
6345 seg->s_szc = szc;
6346
6347 return (0);
6348 }
6349
6350 static int
6351 segvn_clrszc(struct seg *seg)
6352 {
6353 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6354 struct anon_map *amp = svd->amp;
6355 size_t pgsz;
6356 pgcnt_t pages;
6357 int err = 0;
6358 caddr_t a = seg->s_base;
6359 caddr_t ea = a + seg->s_size;
6360 ulong_t an_idx = svd->anon_index;
6361 vnode_t *vp = svd->vp;
6362 struct vpage *vpage = svd->vpage;
6363 page_t *anon_pl[1 + 1], *pp;
6364 struct anon *ap, *oldap;
6365 uint_t prot = svd->prot, vpprot;
6366 int pageflag = 0;
6367
6368 ASSERT(AS_WRITE_HELD(seg->s_as) ||
6369 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6370 ASSERT(svd->softlockcnt == 0);
6371
6372 if (vp == NULL && amp == NULL) {
6373 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6374 seg->s_szc = 0;
6375 return (0);
6376 }
6377
6378 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6379 ASSERT(svd->amp == NULL);
6380 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6381 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6382 HAT_REGION_TEXT);
6383 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6384 } else if (svd->tr_state == SEGVN_TR_ON) {
6385 ASSERT(svd->amp != NULL);
6386 segvn_textunrepl(seg, 1);
6387 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6388 amp = NULL;
6389 } else {
6390 if (svd->tr_state != SEGVN_TR_OFF) {
6391 ASSERT(svd->tr_state == SEGVN_TR_INIT);
6392 svd->tr_state = SEGVN_TR_OFF;
6393 }
6394
6395 /*
6396 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6397 * unload argument is 0 when we are freeing the segment
6398 * and unload was already done.
6399 */
6400 hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6401 HAT_UNLOAD_UNMAP);
6402 }
6403
6404 if (amp == NULL || svd->type == MAP_SHARED) {
6405 seg->s_szc = 0;
6406 return (0);
6407 }
6408
6409 pgsz = page_get_pagesize(seg->s_szc);
6410 pages = btop(pgsz);
6411
6412 /*
6413 * XXX anon rwlock is not really needed because this is a
6414 * private segment and we are writers.
6415 */
6416 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
6417
6418 for (; a < ea; a += pgsz, an_idx += pages) {
6419 if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6420 ASSERT(vpage != NULL || svd->pageprot == 0);
6421 if (vpage != NULL) {
6422 ASSERT(sameprot(seg, a, pgsz));
6423 prot = VPP_PROT(vpage);
6424 pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6425 }
6426 if (seg->s_szc != 0) {
6427 ASSERT(vp == NULL || anon_pages(amp->ahp,
6428 an_idx, pages) == pages);
6429 if ((err = anon_map_demotepages(amp, an_idx,
6430 seg, a, prot, vpage, svd->cred)) != 0) {
6431 goto out;
6432 }
6433 } else {
6434 if (oldap->an_refcnt == 1) {
6435 continue;
6436 }
6437 if ((err = anon_getpage(&oldap, &vpprot,
6438 anon_pl, PAGESIZE, seg, a, S_READ,
6439 svd->cred))) {
6440 goto out;
6441 }
6442 if ((pp = anon_private(&ap, seg, a, prot,
6443 anon_pl[0], pageflag, svd->cred)) == NULL) {
6444 err = ENOMEM;
6445 goto out;
6446 }
6447 anon_decref(oldap);
6448 (void) anon_set_ptr(amp->ahp, an_idx, ap,
6449 ANON_SLEEP);
6450 page_unlock(pp);
6451 }
6452 }
6453 vpage = (vpage == NULL) ? NULL : vpage + pages;
6454 }
6455
6456 amp->a_szc = 0;
6457 seg->s_szc = 0;
6458 out:
6459 ANON_LOCK_EXIT(&->a_rwlock);
6460 return (err);
6461 }
6462
6463 static int
6464 segvn_claim_pages(
6465 struct seg *seg,
6466 struct vpage *svp,
6467 u_offset_t off,
6468 ulong_t anon_idx,
6469 uint_t prot)
6470 {
6471 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6472 size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6473 page_t **ppa;
6474 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6475 struct anon_map *amp = svd->amp;
6476 struct vpage *evp = svp + pgcnt;
6477 caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6478 + seg->s_base;
6479 struct anon *ap;
6480 struct vnode *vp = svd->vp;
6481 page_t *pp;
6482 pgcnt_t pg_idx, i;
6483 int err = 0;
6484 anoff_t aoff;
6485 int anon = (amp != NULL) ? 1 : 0;
6486
6487 ASSERT(svd->type == MAP_PRIVATE);
6488 ASSERT(svd->vpage != NULL);
6489 ASSERT(seg->s_szc != 0);
6490 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6491 ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6492 ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6493
6494 if (VPP_PROT(svp) == prot)
6495 return (1);
6496 if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6497 return (1);
6498
6499 ppa = kmem_alloc(ppasize, KM_SLEEP);
6500 if (anon && vp != NULL) {
6501 if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6502 anon = 0;
6503 ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6504 }
6505 ASSERT(!anon ||
6506 anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6507 }
6508
6509 for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6510 if (!VPP_ISPPLOCK(svp))
6511 continue;
6512 if (anon) {
6513 ap = anon_get_ptr(amp->ahp, anon_idx);
6514 if (ap == NULL) {
6515 panic("segvn_claim_pages: no anon slot");
6516 }
6517 swap_xlate(ap, &vp, &aoff);
6518 off = (u_offset_t)aoff;
6519 }
6520 ASSERT(vp != NULL);
6521 if ((pp = page_lookup(vp,
6522 (u_offset_t)off, SE_SHARED)) == NULL) {
6523 panic("segvn_claim_pages: no page");
6524 }
6525 ppa[pg_idx++] = pp;
6526 off += PAGESIZE;
6527 }
6528
6529 if (ppa[0] == NULL) {
6530 kmem_free(ppa, ppasize);
6531 return (1);
6532 }
6533
6534 ASSERT(pg_idx <= pgcnt);
6535 ppa[pg_idx] = NULL;
6536
6537
6538 /* Find each large page within ppa, and adjust its claim */
6539
6540 /* Does ppa cover a single large page? */
6541 if (ppa[0]->p_szc == seg->s_szc) {
6542 if (prot & PROT_WRITE)
6543 err = page_addclaim_pages(ppa);
6544 else
6545 err = page_subclaim_pages(ppa);
6546 } else {
6547 for (i = 0; ppa[i]; i += pgcnt) {
6548 ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6549 if (prot & PROT_WRITE)
6550 err = page_addclaim_pages(&ppa[i]);
6551 else
6552 err = page_subclaim_pages(&ppa[i]);
6553 if (err == 0)
6554 break;
6555 }
6556 }
6557
6558 for (i = 0; i < pg_idx; i++) {
6559 ASSERT(ppa[i] != NULL);
6560 page_unlock(ppa[i]);
6561 }
6562
6563 kmem_free(ppa, ppasize);
6564 return (err);
6565 }
6566
6567 /*
6568 * Returns right (upper address) segment if split occurred.
6569 * If the address is equal to the beginning or end of its segment it returns
6570 * the current segment.
6571 */
6572 static struct seg *
6573 segvn_split_seg(struct seg *seg, caddr_t addr)
6574 {
6575 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6576 struct seg *nseg;
6577 size_t nsize;
6578 struct segvn_data *nsvd;
6579
6580 ASSERT(AS_WRITE_HELD(seg->s_as));
6581 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6582
6583 ASSERT(addr >= seg->s_base);
6584 ASSERT(addr <= seg->s_base + seg->s_size);
6585 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6586
6587 if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6588 return (seg);
6589
6590 nsize = seg->s_base + seg->s_size - addr;
6591 seg->s_size = addr - seg->s_base;
6592 nseg = seg_alloc(seg->s_as, addr, nsize);
6593 ASSERT(nseg != NULL);
6594 nseg->s_ops = seg->s_ops;
6595 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6596 nseg->s_data = (void *)nsvd;
6597 nseg->s_szc = seg->s_szc;
6598 *nsvd = *svd;
6599 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6600 nsvd->seg = nseg;
6601 rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6602
6603 if (nsvd->vp != NULL) {
6604 VN_HOLD(nsvd->vp);
6605 nsvd->offset = svd->offset +
6606 (uintptr_t)(nseg->s_base - seg->s_base);
6607 if (nsvd->type == MAP_SHARED)
6608 lgrp_shm_policy_init(NULL, nsvd->vp);
6609 } else {
6610 /*
6611 * The offset for an anonymous segment has no signifigance in
6612 * terms of an offset into a file. If we were to use the above
6613 * calculation instead, the structures read out of
6614 * /proc/<pid>/xmap would be more difficult to decipher since
6615 * it would be unclear whether two seemingly contiguous
6616 * prxmap_t structures represented different segments or a
6617 * single segment that had been split up into multiple prxmap_t
6618 * structures (e.g. if some part of the segment had not yet
6619 * been faulted in).
6620 */
6621 nsvd->offset = 0;
6622 }
6623
6624 ASSERT(svd->softlockcnt == 0);
6625 ASSERT(svd->softlockcnt_sbase == 0);
6626 ASSERT(svd->softlockcnt_send == 0);
6627 crhold(svd->cred);
6628
6629 if (svd->vpage != NULL) {
6630 size_t bytes = vpgtob(seg_pages(seg));
6631 size_t nbytes = vpgtob(seg_pages(nseg));
6632 struct vpage *ovpage = svd->vpage;
6633
6634 svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6635 bcopy(ovpage, svd->vpage, bytes);
6636 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6637 bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6638 kmem_free(ovpage, bytes + nbytes);
6639 }
6640 if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6641 struct anon_map *oamp = svd->amp, *namp;
6642 struct anon_hdr *nahp;
6643
6644 ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6645 ASSERT(oamp->refcnt == 1);
6646 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6647 (void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6648 nahp, 0, btop(seg->s_size), ANON_SLEEP);
6649
6650 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6651 namp->a_szc = nseg->s_szc;
6652 (void) anon_copy_ptr(oamp->ahp,
6653 svd->anon_index + btop(seg->s_size),
6654 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6655 anon_release(oamp->ahp, btop(oamp->size));
6656 oamp->ahp = nahp;
6657 oamp->size = seg->s_size;
6658 svd->anon_index = 0;
6659 nsvd->amp = namp;
6660 nsvd->anon_index = 0;
6661 ANON_LOCK_EXIT(&oamp->a_rwlock);
6662 } else if (svd->amp != NULL) {
6663 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6664 ASSERT(svd->amp == nsvd->amp);
6665 ASSERT(seg->s_szc <= svd->amp->a_szc);
6666 nsvd->anon_index = svd->anon_index + seg_pages(seg);
6667 ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6668 ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6669 svd->amp->refcnt++;
6670 ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6671 }
6672
6673 /*
6674 * Split the amount of swap reserved.
6675 */
6676 if (svd->swresv) {
6677 /*
6678 * For MAP_NORESERVE, only allocate swap reserve for pages
6679 * being used. Other segments get enough to cover whole
6680 * segment.
6681 */
6682 if (svd->flags & MAP_NORESERVE) {
6683 size_t oswresv;
6684
6685 ASSERT(svd->amp);
6686 oswresv = svd->swresv;
6687 svd->swresv = ptob(anon_pages(svd->amp->ahp,
6688 svd->anon_index, btop(seg->s_size)));
6689 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6690 nsvd->anon_index, btop(nseg->s_size)));
6691 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6692 } else {
6693 if (svd->pageswap) {
6694 svd->swresv = segvn_count_swap_by_vpages(seg);
6695 ASSERT(nsvd->swresv >= svd->swresv);
6696 nsvd->swresv -= svd->swresv;
6697 } else {
6698 ASSERT(svd->swresv == seg->s_size +
6699 nseg->s_size);
6700 svd->swresv = seg->s_size;
6701 nsvd->swresv = nseg->s_size;
6702 }
6703 }
6704 }
6705
6706 return (nseg);
6707 }
6708
6709 /*
6710 * called on memory operations (unmap, setprot, setpagesize) for a subset
6711 * of a large page segment to either demote the memory range (SDR_RANGE)
6712 * or the ends (SDR_END) by addr/len.
6713 *
6714 * returns 0 on success. returns errno, including ENOMEM, on failure.
6715 */
6716 static int
6717 segvn_demote_range(
6718 struct seg *seg,
6719 caddr_t addr,
6720 size_t len,
6721 int flag,
6722 uint_t szcvec)
6723 {
6724 caddr_t eaddr = addr + len;
6725 caddr_t lpgaddr, lpgeaddr;
6726 struct seg *nseg;
6727 struct seg *badseg1 = NULL;
6728 struct seg *badseg2 = NULL;
6729 size_t pgsz;
6730 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6731 int err;
6732 uint_t szc = seg->s_szc;
6733 uint_t tszcvec;
6734
6735 ASSERT(AS_WRITE_HELD(seg->s_as));
6736 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6737 ASSERT(szc != 0);
6738 pgsz = page_get_pagesize(szc);
6739 ASSERT(seg->s_base != addr || seg->s_size != len);
6740 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6741 ASSERT(svd->softlockcnt == 0);
6742 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6743 ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6744
6745 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6746 ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6747 if (flag == SDR_RANGE) {
6748 /* demote entire range */
6749 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6750 (void) segvn_split_seg(nseg, lpgeaddr);
6751 ASSERT(badseg1->s_base == lpgaddr);
6752 ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6753 } else if (addr != lpgaddr) {
6754 ASSERT(flag == SDR_END);
6755 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6756 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6757 eaddr < lpgaddr + 2 * pgsz) {
6758 (void) segvn_split_seg(nseg, lpgeaddr);
6759 ASSERT(badseg1->s_base == lpgaddr);
6760 ASSERT(badseg1->s_size == 2 * pgsz);
6761 } else {
6762 nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6763 ASSERT(badseg1->s_base == lpgaddr);
6764 ASSERT(badseg1->s_size == pgsz);
6765 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6766 ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6767 nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6768 badseg2 = nseg;
6769 (void) segvn_split_seg(nseg, lpgeaddr);
6770 ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6771 ASSERT(badseg2->s_size == pgsz);
6772 }
6773 }
6774 } else {
6775 ASSERT(flag == SDR_END);
6776 ASSERT(eaddr < lpgeaddr);
6777 badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6778 (void) segvn_split_seg(nseg, lpgeaddr);
6779 ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6780 ASSERT(badseg1->s_size == pgsz);
6781 }
6782
6783 ASSERT(badseg1 != NULL);
6784 ASSERT(badseg1->s_szc == szc);
6785 ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6786 badseg1->s_size == 2 * pgsz);
6787 ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6788 ASSERT(badseg1->s_size == pgsz ||
6789 sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6790 if (err = segvn_clrszc(badseg1)) {
6791 return (err);
6792 }
6793 ASSERT(badseg1->s_szc == 0);
6794
6795 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6796 uint_t tszc = highbit(tszcvec) - 1;
6797 caddr_t ta = MAX(addr, badseg1->s_base);
6798 caddr_t te;
6799 size_t tpgsz = page_get_pagesize(tszc);
6800
6801 ASSERT(svd->type == MAP_SHARED);
6802 ASSERT(flag == SDR_END);
6803 ASSERT(tszc < szc && tszc > 0);
6804
6805 if (eaddr > badseg1->s_base + badseg1->s_size) {
6806 te = badseg1->s_base + badseg1->s_size;
6807 } else {
6808 te = eaddr;
6809 }
6810
6811 ASSERT(ta <= te);
6812 badseg1->s_szc = tszc;
6813 if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6814 if (badseg2 != NULL) {
6815 err = segvn_demote_range(badseg1, ta, te - ta,
6816 SDR_END, tszcvec);
6817 if (err != 0) {
6818 return (err);
6819 }
6820 } else {
6821 return (segvn_demote_range(badseg1, ta,
6822 te - ta, SDR_END, tszcvec));
6823 }
6824 }
6825 }
6826
6827 if (badseg2 == NULL)
6828 return (0);
6829 ASSERT(badseg2->s_szc == szc);
6830 ASSERT(badseg2->s_size == pgsz);
6831 ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6832 if (err = segvn_clrszc(badseg2)) {
6833 return (err);
6834 }
6835 ASSERT(badseg2->s_szc == 0);
6836
6837 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6838 uint_t tszc = highbit(tszcvec) - 1;
6839 size_t tpgsz = page_get_pagesize(tszc);
6840
6841 ASSERT(svd->type == MAP_SHARED);
6842 ASSERT(flag == SDR_END);
6843 ASSERT(tszc < szc && tszc > 0);
6844 ASSERT(badseg2->s_base > addr);
6845 ASSERT(eaddr > badseg2->s_base);
6846 ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6847
6848 badseg2->s_szc = tszc;
6849 if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6850 return (segvn_demote_range(badseg2, badseg2->s_base,
6851 eaddr - badseg2->s_base, SDR_END, tszcvec));
6852 }
6853 }
6854
6855 return (0);
6856 }
6857
6858 static int
6859 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6860 {
6861 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6862 struct vpage *vp, *evp;
6863
6864 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6865
6866 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6867 /*
6868 * If segment protection can be used, simply check against them.
6869 */
6870 if (svd->pageprot == 0) {
6871 int err;
6872
6873 err = ((svd->prot & prot) != prot) ? EACCES : 0;
6874 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6875 return (err);
6876 }
6877
6878 /*
6879 * Have to check down to the vpage level.
6880 */
6881 evp = &svd->vpage[seg_page(seg, addr + len)];
6882 for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6883 if ((VPP_PROT(vp) & prot) != prot) {
6884 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6885 return (EACCES);
6886 }
6887 }
6888 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6889 return (0);
6890 }
6891
6892 static int
6893 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6894 {
6895 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6896 size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6897
6898 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6899
6900 if (pgno != 0) {
6901 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6902 if (svd->pageprot == 0) {
6903 do {
6904 protv[--pgno] = svd->prot;
6905 } while (pgno != 0);
6906 } else {
6907 size_t pgoff = seg_page(seg, addr);
6908
6909 do {
6910 pgno--;
6911 protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6912 } while (pgno != 0);
6913 }
6914 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6915 }
6916 return (0);
6917 }
6918
6919 static u_offset_t
6920 segvn_getoffset(struct seg *seg, caddr_t addr)
6921 {
6922 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6923
6924 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6925
6926 return (svd->offset + (uintptr_t)(addr - seg->s_base));
6927 }
6928
6929 /*ARGSUSED*/
6930 static int
6931 segvn_gettype(struct seg *seg, caddr_t addr)
6932 {
6933 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6934
6935 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6936
6937 return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6938 MAP_INITDATA)));
6939 }
6940
6941 /*ARGSUSED*/
6942 static int
6943 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6944 {
6945 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6946
6947 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6948
6949 *vpp = svd->vp;
6950 return (0);
6951 }
6952
6953 /*
6954 * Check to see if it makes sense to do kluster/read ahead to
6955 * addr + delta relative to the mapping at addr. We assume here
6956 * that delta is a signed PAGESIZE'd multiple (which can be negative).
6957 *
6958 * For segvn, we currently "approve" of the action if we are
6959 * still in the segment and it maps from the same vp/off,
6960 * or if the advice stored in segvn_data or vpages allows it.
6961 * Currently, klustering is not allowed only if MADV_RANDOM is set.
6962 */
6963 static int
6964 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6965 {
6966 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6967 struct anon *oap, *ap;
6968 ssize_t pd;
6969 size_t page;
6970 struct vnode *vp1, *vp2;
6971 u_offset_t off1, off2;
6972 struct anon_map *amp;
6973
6974 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6975 ASSERT(AS_WRITE_HELD(seg->s_as) ||
6976 SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6977
6978 if (addr + delta < seg->s_base ||
6979 addr + delta >= (seg->s_base + seg->s_size))
6980 return (-1); /* exceeded segment bounds */
6981
6982 pd = delta / (ssize_t)PAGESIZE; /* divide to preserve sign bit */
6983 page = seg_page(seg, addr);
6984
6985 /*
6986 * Check to see if either of the pages addr or addr + delta
6987 * have advice set that prevents klustering (if MADV_RANDOM advice
6988 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6989 * is negative).
6990 */
6991 if (svd->advice == MADV_RANDOM ||
6992 svd->advice == MADV_SEQUENTIAL && delta < 0)
6993 return (-1);
6994 else if (svd->pageadvice && svd->vpage) {
6995 struct vpage *bvpp, *evpp;
6996
6997 bvpp = &svd->vpage[page];
6998 evpp = &svd->vpage[page + pd];
6999 if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
7000 VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
7001 return (-1);
7002 if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
7003 VPP_ADVICE(evpp) == MADV_RANDOM)
7004 return (-1);
7005 }
7006
7007 if (svd->type == MAP_SHARED)
7008 return (0); /* shared mapping - all ok */
7009
7010 if ((amp = svd->amp) == NULL)
7011 return (0); /* off original vnode */
7012
7013 page += svd->anon_index;
7014
7015 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7016
7017 oap = anon_get_ptr(amp->ahp, page);
7018 ap = anon_get_ptr(amp->ahp, page + pd);
7019
7020 ANON_LOCK_EXIT(&->a_rwlock);
7021
7022 if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
7023 return (-1); /* one with and one without an anon */
7024 }
7025
7026 if (oap == NULL) { /* implies that ap == NULL */
7027 return (0); /* off original vnode */
7028 }
7029
7030 /*
7031 * Now we know we have two anon pointers - check to
7032 * see if they happen to be properly allocated.
7033 */
7034
7035 /*
7036 * XXX We cheat here and don't lock the anon slots. We can't because
7037 * we may have been called from the anon layer which might already
7038 * have locked them. We are holding a refcnt on the slots so they
7039 * can't disappear. The worst that will happen is we'll get the wrong
7040 * names (vp, off) for the slots and make a poor klustering decision.
7041 */
7042 swap_xlate(ap, &vp1, &off1);
7043 swap_xlate(oap, &vp2, &off2);
7044
7045
7046 if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
7047 return (-1);
7048 return (0);
7049 }
7050
7051 /*
7052 * Synchronize primary storage cache with real object in virtual memory.
7053 *
7054 * XXX - Anonymous pages should not be sync'ed out at all.
7055 */
7056 static int
7057 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7058 {
7059 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7060 struct vpage *vpp;
7061 page_t *pp;
7062 u_offset_t offset;
7063 struct vnode *vp;
7064 u_offset_t off;
7065 caddr_t eaddr;
7066 int bflags;
7067 int err = 0;
7068 int segtype;
7069 int pageprot;
7070 int prot;
7071 ulong_t anon_index;
7072 struct anon_map *amp;
7073 struct anon *ap;
7074 anon_sync_obj_t cookie;
7075
7076 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7077
7078 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7079
7080 if (svd->softlockcnt > 0) {
7081 /*
7082 * If this is shared segment non 0 softlockcnt
7083 * means locked pages are still in use.
7084 */
7085 if (svd->type == MAP_SHARED) {
7086 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7087 return (EAGAIN);
7088 }
7089
7090 /*
7091 * flush all pages from seg cache
7092 * otherwise we may deadlock in swap_putpage
7093 * for B_INVAL page (4175402).
7094 *
7095 * Even if we grab segvn WRITER's lock
7096 * here, there might be another thread which could've
7097 * successfully performed lookup/insert just before
7098 * we acquired the lock here. So, grabbing either
7099 * lock here is of not much use. Until we devise
7100 * a strategy at upper layers to solve the
7101 * synchronization issues completely, we expect
7102 * applications to handle this appropriately.
7103 */
7104 segvn_purge(seg);
7105 if (svd->softlockcnt > 0) {
7106 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7107 return (EAGAIN);
7108 }
7109 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7110 svd->amp->a_softlockcnt > 0) {
7111 /*
7112 * Try to purge this amp's entries from pcache. It will
7113 * succeed only if other segments that share the amp have no
7114 * outstanding softlock's.
7115 */
7116 segvn_purge(seg);
7117 if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7118 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7119 return (EAGAIN);
7120 }
7121 }
7122
7123 vpp = svd->vpage;
7124 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7125 bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7126 ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7127
7128 if (attr) {
7129 pageprot = attr & ~(SHARED|PRIVATE);
7130 segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7131
7132 /*
7133 * We are done if the segment types don't match
7134 * or if we have segment level protections and
7135 * they don't match.
7136 */
7137 if (svd->type != segtype) {
7138 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7139 return (0);
7140 }
7141 if (vpp == NULL) {
7142 if (svd->prot != pageprot) {
7143 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7144 return (0);
7145 }
7146 prot = svd->prot;
7147 } else
7148 vpp = &svd->vpage[seg_page(seg, addr)];
7149
7150 } else if (svd->vp && svd->amp == NULL &&
7151 (flags & MS_INVALIDATE) == 0) {
7152
7153 /*
7154 * No attributes, no anonymous pages and MS_INVALIDATE flag
7155 * is not on, just use one big request.
7156 */
7157 err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7158 bflags, svd->cred, NULL);
7159 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7160 return (err);
7161 }
7162
7163 if ((amp = svd->amp) != NULL)
7164 anon_index = svd->anon_index + seg_page(seg, addr);
7165
7166 for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7167 ap = NULL;
7168 if (amp != NULL) {
7169 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7170 anon_array_enter(amp, anon_index, &cookie);
7171 ap = anon_get_ptr(amp->ahp, anon_index++);
7172 if (ap != NULL) {
7173 swap_xlate(ap, &vp, &off);
7174 } else {
7175 vp = svd->vp;
7176 off = offset;
7177 }
7178 anon_array_exit(&cookie);
7179 ANON_LOCK_EXIT(&->a_rwlock);
7180 } else {
7181 vp = svd->vp;
7182 off = offset;
7183 }
7184 offset += PAGESIZE;
7185
7186 if (vp == NULL) /* untouched zfod page */
7187 continue;
7188
7189 if (attr) {
7190 if (vpp) {
7191 prot = VPP_PROT(vpp);
7192 vpp++;
7193 }
7194 if (prot != pageprot) {
7195 continue;
7196 }
7197 }
7198
7199 /*
7200 * See if any of these pages are locked -- if so, then we
7201 * will have to truncate an invalidate request at the first
7202 * locked one. We don't need the page_struct_lock to test
7203 * as this is only advisory; even if we acquire it someone
7204 * might race in and lock the page after we unlock and before
7205 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7206 */
7207 if (flags & MS_INVALIDATE) {
7208 if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7209 if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7210 page_unlock(pp);
7211 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7212 return (EBUSY);
7213 }
7214 if (ap != NULL && pp->p_szc != 0 &&
7215 page_tryupgrade(pp)) {
7216 if (pp->p_lckcnt == 0 &&
7217 pp->p_cowcnt == 0) {
7218 /*
7219 * swapfs VN_DISPOSE() won't
7220 * invalidate large pages.
7221 * Attempt to demote.
7222 * XXX can't help it if it
7223 * fails. But for swapfs
7224 * pages it is no big deal.
7225 */
7226 (void) page_try_demote_pages(
7227 pp);
7228 }
7229 }
7230 page_unlock(pp);
7231 }
7232 } else if (svd->type == MAP_SHARED && amp != NULL) {
7233 /*
7234 * Avoid writing out to disk ISM's large pages
7235 * because segspt_free_pages() relies on NULL an_pvp
7236 * of anon slots of such pages.
7237 */
7238
7239 ASSERT(svd->vp == NULL);
7240 /*
7241 * swapfs uses page_lookup_nowait if not freeing or
7242 * invalidating and skips a page if
7243 * page_lookup_nowait returns NULL.
7244 */
7245 pp = page_lookup_nowait(vp, off, SE_SHARED);
7246 if (pp == NULL) {
7247 continue;
7248 }
7249 if (pp->p_szc != 0) {
7250 page_unlock(pp);
7251 continue;
7252 }
7253
7254 /*
7255 * Note ISM pages are created large so (vp, off)'s
7256 * page cannot suddenly become large after we unlock
7257 * pp.
7258 */
7259 page_unlock(pp);
7260 }
7261 /*
7262 * XXX - Should ultimately try to kluster
7263 * calls to VOP_PUTPAGE() for performance.
7264 */
7265 VN_HOLD(vp);
7266 err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7267 (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7268 svd->cred, NULL);
7269
7270 VN_RELE(vp);
7271 if (err)
7272 break;
7273 }
7274 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7275 return (err);
7276 }
7277
7278 /*
7279 * Determine if we have data corresponding to pages in the
7280 * primary storage virtual memory cache (i.e., "in core").
7281 */
7282 static size_t
7283 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7284 {
7285 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7286 struct vnode *vp, *avp;
7287 u_offset_t offset, aoffset;
7288 size_t p, ep;
7289 int ret;
7290 struct vpage *vpp;
7291 page_t *pp;
7292 uint_t start;
7293 struct anon_map *amp; /* XXX - for locknest */
7294 struct anon *ap;
7295 uint_t attr;
7296 anon_sync_obj_t cookie;
7297
7298 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7299
7300 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7301 if (svd->amp == NULL && svd->vp == NULL) {
7302 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7303 bzero(vec, btopr(len));
7304 return (len); /* no anonymous pages created yet */
7305 }
7306
7307 p = seg_page(seg, addr);
7308 ep = seg_page(seg, addr + len);
7309 start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7310
7311 amp = svd->amp;
7312 for (; p < ep; p++, addr += PAGESIZE) {
7313 vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7314 ret = start;
7315 ap = NULL;
7316 avp = NULL;
7317 /* Grab the vnode/offset for the anon slot */
7318 if (amp != NULL) {
7319 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7320 anon_array_enter(amp, svd->anon_index + p, &cookie);
7321 ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7322 if (ap != NULL) {
7323 swap_xlate(ap, &avp, &aoffset);
7324 }
7325 anon_array_exit(&cookie);
7326 ANON_LOCK_EXIT(&->a_rwlock);
7327 }
7328 if ((avp != NULL) && page_exists(avp, aoffset)) {
7329 /* A page exists for the anon slot */
7330 ret |= SEG_PAGE_INCORE;
7331
7332 /*
7333 * If page is mapped and writable
7334 */
7335 attr = (uint_t)0;
7336 if ((hat_getattr(seg->s_as->a_hat, addr,
7337 &attr) != -1) && (attr & PROT_WRITE)) {
7338 ret |= SEG_PAGE_ANON;
7339 }
7340 /*
7341 * Don't get page_struct lock for lckcnt and cowcnt,
7342 * since this is purely advisory.
7343 */
7344 if ((pp = page_lookup_nowait(avp, aoffset,
7345 SE_SHARED)) != NULL) {
7346 if (pp->p_lckcnt)
7347 ret |= SEG_PAGE_SOFTLOCK;
7348 if (pp->p_cowcnt)
7349 ret |= SEG_PAGE_HASCOW;
7350 page_unlock(pp);
7351 }
7352 }
7353
7354 /* Gather vnode statistics */
7355 vp = svd->vp;
7356 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7357
7358 if (vp != NULL) {
7359 /*
7360 * Try to obtain a "shared" lock on the page
7361 * without blocking. If this fails, determine
7362 * if the page is in memory.
7363 */
7364 pp = page_lookup_nowait(vp, offset, SE_SHARED);
7365 if ((pp == NULL) && (page_exists(vp, offset))) {
7366 /* Page is incore, and is named */
7367 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7368 }
7369 /*
7370 * Don't get page_struct lock for lckcnt and cowcnt,
7371 * since this is purely advisory.
7372 */
7373 if (pp != NULL) {
7374 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7375 if (pp->p_lckcnt)
7376 ret |= SEG_PAGE_SOFTLOCK;
7377 if (pp->p_cowcnt)
7378 ret |= SEG_PAGE_HASCOW;
7379 page_unlock(pp);
7380 }
7381 }
7382
7383 /* Gather virtual page information */
7384 if (vpp) {
7385 if (VPP_ISPPLOCK(vpp))
7386 ret |= SEG_PAGE_LOCKED;
7387 vpp++;
7388 }
7389
7390 *vec++ = (char)ret;
7391 }
7392 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7393 return (len);
7394 }
7395
7396 /*
7397 * Statement for p_cowcnts/p_lckcnts.
7398 *
7399 * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7400 * irrespective of the following factors or anything else:
7401 *
7402 * (1) anon slots are populated or not
7403 * (2) cow is broken or not
7404 * (3) refcnt on ap is 1 or greater than 1
7405 *
7406 * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7407 * and munlock.
7408 *
7409 *
7410 * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7411 *
7412 * if vpage has PROT_WRITE
7413 * transfer cowcnt on the oldpage -> cowcnt on the newpage
7414 * else
7415 * transfer lckcnt on the oldpage -> lckcnt on the newpage
7416 *
7417 * During copy-on-write, decrement p_cowcnt on the oldpage and increment
7418 * p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7419 *
7420 * We may also break COW if softlocking on read access in the physio case.
7421 * In this case, vpage may not have PROT_WRITE. So, we need to decrement
7422 * p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7423 * vpage doesn't have PROT_WRITE.
7424 *
7425 *
7426 * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7427 *
7428 * If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7429 * increment p_lckcnt by calling page_subclaim() which takes care of
7430 * availrmem accounting and p_lckcnt overflow.
7431 *
7432 * If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7433 * increment p_cowcnt by calling page_addclaim() which takes care of
7434 * availrmem availability and p_cowcnt overflow.
7435 */
7436
7437 /*
7438 * Lock down (or unlock) pages mapped by this segment.
7439 *
7440 * XXX only creates PAGESIZE pages if anon slots are not initialized.
7441 * At fault time they will be relocated into larger pages.
7442 */
7443 static int
7444 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7445 int attr, int op, ulong_t *lockmap, size_t pos)
7446 {
7447 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7448 struct vpage *vpp;
7449 struct vpage *evp;
7450 page_t *pp;
7451 u_offset_t offset;
7452 u_offset_t off;
7453 int segtype;
7454 int pageprot;
7455 int claim;
7456 struct vnode *vp;
7457 ulong_t anon_index;
7458 struct anon_map *amp;
7459 struct anon *ap;
7460 struct vattr va;
7461 anon_sync_obj_t cookie;
7462 struct kshmid *sp = NULL;
7463 struct proc *p = curproc;
7464 kproject_t *proj = NULL;
7465 int chargeproc = 1;
7466 size_t locked_bytes = 0;
7467 size_t unlocked_bytes = 0;
7468 int err = 0;
7469
7470 /*
7471 * Hold write lock on address space because may split or concatenate
7472 * segments
7473 */
7474 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7475
7476 /*
7477 * If this is a shm, use shm's project and zone, else use
7478 * project and zone of calling process
7479 */
7480
7481 /* Determine if this segment backs a sysV shm */
7482 if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7483 ASSERT(svd->type == MAP_SHARED);
7484 ASSERT(svd->tr_state == SEGVN_TR_OFF);
7485 sp = svd->amp->a_sp;
7486 proj = sp->shm_perm.ipc_proj;
7487 chargeproc = 0;
7488 }
7489
7490 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7491 if (attr) {
7492 pageprot = attr & ~(SHARED|PRIVATE);
7493 segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7494
7495 /*
7496 * We are done if the segment types don't match
7497 * or if we have segment level protections and
7498 * they don't match.
7499 */
7500 if (svd->type != segtype) {
7501 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7502 return (0);
7503 }
7504 if (svd->pageprot == 0 && svd->prot != pageprot) {
7505 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7506 return (0);
7507 }
7508 }
7509
7510 if (op == MC_LOCK) {
7511 if (svd->tr_state == SEGVN_TR_INIT) {
7512 svd->tr_state = SEGVN_TR_OFF;
7513 } else if (svd->tr_state == SEGVN_TR_ON) {
7514 ASSERT(svd->amp != NULL);
7515 segvn_textunrepl(seg, 0);
7516 ASSERT(svd->amp == NULL &&
7517 svd->tr_state == SEGVN_TR_OFF);
7518 }
7519 }
7520
7521 /*
7522 * If we're locking, then we must create a vpage structure if
7523 * none exists. If we're unlocking, then check to see if there
7524 * is a vpage -- if not, then we could not have locked anything.
7525 */
7526
7527 if ((vpp = svd->vpage) == NULL) {
7528 if (op == MC_LOCK) {
7529 segvn_vpage(seg);
7530 if (svd->vpage == NULL) {
7531 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7532 return (ENOMEM);
7533 }
7534 } else {
7535 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7536 return (0);
7537 }
7538 }
7539
7540 /*
7541 * The anonymous data vector (i.e., previously
7542 * unreferenced mapping to swap space) can be allocated
7543 * by lazily testing for its existence.
7544 */
7545 if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7546 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7547 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7548 svd->amp->a_szc = seg->s_szc;
7549 }
7550
7551 if ((amp = svd->amp) != NULL) {
7552 anon_index = svd->anon_index + seg_page(seg, addr);
7553 }
7554
7555 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7556 evp = &svd->vpage[seg_page(seg, addr + len)];
7557
7558 if (sp != NULL)
7559 mutex_enter(&sp->shm_mlock);
7560
7561 /* determine number of unlocked bytes in range for lock operation */
7562 if (op == MC_LOCK) {
7563
7564 if (sp == NULL) {
7565 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7566 vpp++) {
7567 if (!VPP_ISPPLOCK(vpp))
7568 unlocked_bytes += PAGESIZE;
7569 }
7570 } else {
7571 ulong_t i_idx, i_edx;
7572 anon_sync_obj_t i_cookie;
7573 struct anon *i_ap;
7574 struct vnode *i_vp;
7575 u_offset_t i_off;
7576
7577 /* Only count sysV pages once for locked memory */
7578 i_edx = svd->anon_index + seg_page(seg, addr + len);
7579 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7580 for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7581 anon_array_enter(amp, i_idx, &i_cookie);
7582 i_ap = anon_get_ptr(amp->ahp, i_idx);
7583 if (i_ap == NULL) {
7584 unlocked_bytes += PAGESIZE;
7585 anon_array_exit(&i_cookie);
7586 continue;
7587 }
7588 swap_xlate(i_ap, &i_vp, &i_off);
7589 anon_array_exit(&i_cookie);
7590 pp = page_lookup(i_vp, i_off, SE_SHARED);
7591 if (pp == NULL) {
7592 unlocked_bytes += PAGESIZE;
7593 continue;
7594 } else if (pp->p_lckcnt == 0)
7595 unlocked_bytes += PAGESIZE;
7596 page_unlock(pp);
7597 }
7598 ANON_LOCK_EXIT(&->a_rwlock);
7599 }
7600
7601 mutex_enter(&p->p_lock);
7602 err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7603 chargeproc);
7604 mutex_exit(&p->p_lock);
7605
7606 if (err) {
7607 if (sp != NULL)
7608 mutex_exit(&sp->shm_mlock);
7609 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7610 return (err);
7611 }
7612 }
7613 /*
7614 * Loop over all pages in the range. Process if we're locking and
7615 * page has not already been locked in this mapping; or if we're
7616 * unlocking and the page has been locked.
7617 */
7618 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7619 vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7620 if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7621 ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7622 (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7623
7624 if (amp != NULL)
7625 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7626 /*
7627 * If this isn't a MAP_NORESERVE segment and
7628 * we're locking, allocate anon slots if they
7629 * don't exist. The page is brought in later on.
7630 */
7631 if (op == MC_LOCK && svd->vp == NULL &&
7632 ((svd->flags & MAP_NORESERVE) == 0) &&
7633 amp != NULL &&
7634 ((ap = anon_get_ptr(amp->ahp, anon_index))
7635 == NULL)) {
7636 anon_array_enter(amp, anon_index, &cookie);
7637
7638 if ((ap = anon_get_ptr(amp->ahp,
7639 anon_index)) == NULL) {
7640 pp = anon_zero(seg, addr, &ap,
7641 svd->cred);
7642 if (pp == NULL) {
7643 anon_array_exit(&cookie);
7644 ANON_LOCK_EXIT(&->a_rwlock);
7645 err = ENOMEM;
7646 goto out;
7647 }
7648 ASSERT(anon_get_ptr(amp->ahp,
7649 anon_index) == NULL);
7650 (void) anon_set_ptr(amp->ahp,
7651 anon_index, ap, ANON_SLEEP);
7652 page_unlock(pp);
7653 }
7654 anon_array_exit(&cookie);
7655 }
7656
7657 /*
7658 * Get name for page, accounting for
7659 * existence of private copy.
7660 */
7661 ap = NULL;
7662 if (amp != NULL) {
7663 anon_array_enter(amp, anon_index, &cookie);
7664 ap = anon_get_ptr(amp->ahp, anon_index);
7665 if (ap != NULL) {
7666 swap_xlate(ap, &vp, &off);
7667 } else {
7668 if (svd->vp == NULL &&
7669 (svd->flags & MAP_NORESERVE)) {
7670 anon_array_exit(&cookie);
7671 ANON_LOCK_EXIT(&->a_rwlock);
7672 continue;
7673 }
7674 vp = svd->vp;
7675 off = offset;
7676 }
7677 if (op != MC_LOCK || ap == NULL) {
7678 anon_array_exit(&cookie);
7679 ANON_LOCK_EXIT(&->a_rwlock);
7680 }
7681 } else {
7682 vp = svd->vp;
7683 off = offset;
7684 }
7685
7686 /*
7687 * Get page frame. It's ok if the page is
7688 * not available when we're unlocking, as this
7689 * may simply mean that a page we locked got
7690 * truncated out of existence after we locked it.
7691 *
7692 * Invoke VOP_GETPAGE() to obtain the page struct
7693 * since we may need to read it from disk if its
7694 * been paged out.
7695 */
7696 if (op != MC_LOCK)
7697 pp = page_lookup(vp, off, SE_SHARED);
7698 else {
7699 page_t *pl[1 + 1];
7700 int error;
7701
7702 ASSERT(vp != NULL);
7703
7704 error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7705 (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7706 S_OTHER, svd->cred, NULL);
7707
7708 if (error && ap != NULL) {
7709 anon_array_exit(&cookie);
7710 ANON_LOCK_EXIT(&->a_rwlock);
7711 }
7712
7713 /*
7714 * If the error is EDEADLK then we must bounce
7715 * up and drop all vm subsystem locks and then
7716 * retry the operation later
7717 * This behavior is a temporary measure because
7718 * ufs/sds logging is badly designed and will
7719 * deadlock if we don't allow this bounce to
7720 * happen. The real solution is to re-design
7721 * the logging code to work properly. See bug
7722 * 4125102 for details of the problem.
7723 */
7724 if (error == EDEADLK) {
7725 err = error;
7726 goto out;
7727 }
7728 /*
7729 * Quit if we fail to fault in the page. Treat
7730 * the failure as an error, unless the addr
7731 * is mapped beyond the end of a file.
7732 */
7733 if (error && svd->vp) {
7734 va.va_mask = AT_SIZE;
7735 if (VOP_GETATTR(svd->vp, &va, 0,
7736 svd->cred, NULL) != 0) {
7737 err = EIO;
7738 goto out;
7739 }
7740 if (btopr(va.va_size) >=
7741 btopr(off + 1)) {
7742 err = EIO;
7743 goto out;
7744 }
7745 goto out;
7746
7747 } else if (error) {
7748 err = EIO;
7749 goto out;
7750 }
7751 pp = pl[0];
7752 ASSERT(pp != NULL);
7753 }
7754
7755 /*
7756 * See Statement at the beginning of this routine.
7757 *
7758 * claim is always set if MAP_PRIVATE and PROT_WRITE
7759 * irrespective of following factors:
7760 *
7761 * (1) anon slots are populated or not
7762 * (2) cow is broken or not
7763 * (3) refcnt on ap is 1 or greater than 1
7764 *
7765 * See 4140683 for details
7766 */
7767 claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7768 (svd->type == MAP_PRIVATE));
7769
7770 /*
7771 * Perform page-level operation appropriate to
7772 * operation. If locking, undo the SOFTLOCK
7773 * performed to bring the page into memory
7774 * after setting the lock. If unlocking,
7775 * and no page was found, account for the claim
7776 * separately.
7777 */
7778 if (op == MC_LOCK) {
7779 int ret = 1; /* Assume success */
7780
7781 ASSERT(!VPP_ISPPLOCK(vpp));
7782
7783 ret = page_pp_lock(pp, claim, 0);
7784 if (ap != NULL) {
7785 if (ap->an_pvp != NULL) {
7786 anon_swap_free(ap, pp);
7787 }
7788 anon_array_exit(&cookie);
7789 ANON_LOCK_EXIT(&->a_rwlock);
7790 }
7791 if (ret == 0) {
7792 /* locking page failed */
7793 page_unlock(pp);
7794 err = EAGAIN;
7795 goto out;
7796 }
7797 VPP_SETPPLOCK(vpp);
7798 if (sp != NULL) {
7799 if (pp->p_lckcnt == 1)
7800 locked_bytes += PAGESIZE;
7801 } else
7802 locked_bytes += PAGESIZE;
7803
7804 if (lockmap != (ulong_t *)NULL)
7805 BT_SET(lockmap, pos);
7806
7807 page_unlock(pp);
7808 } else {
7809 ASSERT(VPP_ISPPLOCK(vpp));
7810 if (pp != NULL) {
7811 /* sysV pages should be locked */
7812 ASSERT(sp == NULL || pp->p_lckcnt > 0);
7813 page_pp_unlock(pp, claim, 0);
7814 if (sp != NULL) {
7815 if (pp->p_lckcnt == 0)
7816 unlocked_bytes
7817 += PAGESIZE;
7818 } else
7819 unlocked_bytes += PAGESIZE;
7820 page_unlock(pp);
7821 } else {
7822 ASSERT(sp == NULL);
7823 unlocked_bytes += PAGESIZE;
7824 }
7825 VPP_CLRPPLOCK(vpp);
7826 }
7827 }
7828 }
7829 out:
7830 if (op == MC_LOCK) {
7831 /* Credit back bytes that did not get locked */
7832 if ((unlocked_bytes - locked_bytes) > 0) {
7833 if (proj == NULL)
7834 mutex_enter(&p->p_lock);
7835 rctl_decr_locked_mem(p, proj,
7836 (unlocked_bytes - locked_bytes), chargeproc);
7837 if (proj == NULL)
7838 mutex_exit(&p->p_lock);
7839 }
7840
7841 } else {
7842 /* Account bytes that were unlocked */
7843 if (unlocked_bytes > 0) {
7844 if (proj == NULL)
7845 mutex_enter(&p->p_lock);
7846 rctl_decr_locked_mem(p, proj, unlocked_bytes,
7847 chargeproc);
7848 if (proj == NULL)
7849 mutex_exit(&p->p_lock);
7850 }
7851 }
7852 if (sp != NULL)
7853 mutex_exit(&sp->shm_mlock);
7854 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7855
7856 return (err);
7857 }
7858
7859 /*
7860 * Set advice from user for specified pages
7861 * There are 9 types of advice:
7862 * MADV_NORMAL - Normal (default) behavior (whatever that is)
7863 * MADV_RANDOM - Random page references
7864 * do not allow readahead or 'klustering'
7865 * MADV_SEQUENTIAL - Sequential page references
7866 * Pages previous to the one currently being
7867 * accessed (determined by fault) are 'not needed'
7868 * and are freed immediately
7869 * MADV_WILLNEED - Pages are likely to be used (fault ahead in mctl)
7870 * MADV_DONTNEED - Pages are not needed (synced out in mctl)
7871 * MADV_FREE - Contents can be discarded
7872 * MADV_ACCESS_DEFAULT- Default access
7873 * MADV_ACCESS_LWP - Next LWP will access heavily
7874 * MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7875 */
7876 static int
7877 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7878 {
7879 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7880 size_t page;
7881 int err = 0;
7882 int already_set;
7883 struct anon_map *amp;
7884 ulong_t anon_index;
7885 struct seg *next;
7886 lgrp_mem_policy_t policy;
7887 struct seg *prev;
7888 struct vnode *vp;
7889
7890 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7891
7892 /*
7893 * In case of MADV_FREE, we won't be modifying any segment private
7894 * data structures; so, we only need to grab READER's lock
7895 */
7896 if (behav != MADV_FREE) {
7897 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7898 if (svd->tr_state != SEGVN_TR_OFF) {
7899 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7900 return (0);
7901 }
7902 } else {
7903 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7904 }
7905
7906 /*
7907 * Large pages are assumed to be only turned on when accesses to the
7908 * segment's address range have spatial and temporal locality. That
7909 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7910 * Also, ignore advice affecting lgroup memory allocation
7911 * if don't need to do lgroup optimizations on this system
7912 */
7913
7914 if ((behav == MADV_SEQUENTIAL &&
7915 (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7916 (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7917 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7918 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7919 return (0);
7920 }
7921
7922 if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
7923 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
7924 /*
7925 * Since we are going to unload hat mappings
7926 * we first have to flush the cache. Otherwise
7927 * this might lead to system panic if another
7928 * thread is doing physio on the range whose
7929 * mappings are unloaded by madvise(3C).
7930 */
7931 if (svd->softlockcnt > 0) {
7932 /*
7933 * If this is shared segment non 0 softlockcnt
7934 * means locked pages are still in use.
7935 */
7936 if (svd->type == MAP_SHARED) {
7937 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7938 return (EAGAIN);
7939 }
7940 /*
7941 * Since we do have the segvn writers lock
7942 * nobody can fill the cache with entries
7943 * belonging to this seg during the purge.
7944 * The flush either succeeds or we still
7945 * have pending I/Os. In the later case,
7946 * madvise(3C) fails.
7947 */
7948 segvn_purge(seg);
7949 if (svd->softlockcnt > 0) {
7950 /*
7951 * Since madvise(3C) is advisory and
7952 * it's not part of UNIX98, madvise(3C)
7953 * failure here doesn't cause any hardship.
7954 * Note that we don't block in "as" layer.
7955 */
7956 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7957 return (EAGAIN);
7958 }
7959 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7960 svd->amp->a_softlockcnt > 0) {
7961 /*
7962 * Try to purge this amp's entries from pcache. It
7963 * will succeed only if other segments that share the
7964 * amp have no outstanding softlock's.
7965 */
7966 segvn_purge(seg);
7967 }
7968 }
7969
7970 amp = svd->amp;
7971 vp = svd->vp;
7972 if (behav == MADV_FREE) {
7973 /*
7974 * MADV_FREE is not supported for segments with
7975 * underlying object; if anonmap is NULL, anon slots
7976 * are not yet populated and there is nothing for
7977 * us to do. As MADV_FREE is advisory, we don't
7978 * return error in either case.
7979 */
7980 if (vp != NULL || amp == NULL) {
7981 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7982 return (0);
7983 }
7984
7985 segvn_purge(seg);
7986
7987 page = seg_page(seg, addr);
7988 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
7989 anon_disclaim(amp, svd->anon_index + page, len);
7990 ANON_LOCK_EXIT(&->a_rwlock);
7991 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7992 return (0);
7993 }
7994
7995 /*
7996 * If advice is to be applied to entire segment,
7997 * use advice field in seg_data structure
7998 * otherwise use appropriate vpage entry.
7999 */
8000 if ((addr == seg->s_base) && (len == seg->s_size)) {
8001 switch (behav) {
8002 case MADV_ACCESS_LWP:
8003 case MADV_ACCESS_MANY:
8004 case MADV_ACCESS_DEFAULT:
8005 /*
8006 * Set memory allocation policy for this segment
8007 */
8008 policy = lgrp_madv_to_policy(behav, len, svd->type);
8009 if (svd->type == MAP_SHARED)
8010 already_set = lgrp_shm_policy_set(policy, amp,
8011 svd->anon_index, vp, svd->offset, len);
8012 else {
8013 /*
8014 * For private memory, need writers lock on
8015 * address space because the segment may be
8016 * split or concatenated when changing policy
8017 */
8018 if (AS_READ_HELD(seg->s_as)) {
8019 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8020 return (IE_RETRY);
8021 }
8022
8023 already_set = lgrp_privm_policy_set(policy,
8024 &svd->policy_info, len);
8025 }
8026
8027 /*
8028 * If policy set already and it shouldn't be reapplied,
8029 * don't do anything.
8030 */
8031 if (already_set &&
8032 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8033 break;
8034
8035 /*
8036 * Mark any existing pages in given range for
8037 * migration
8038 */
8039 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8040 vp, svd->offset, 1);
8041
8042 /*
8043 * If same policy set already or this is a shared
8044 * memory segment, don't need to try to concatenate
8045 * segment with adjacent ones.
8046 */
8047 if (already_set || svd->type == MAP_SHARED)
8048 break;
8049
8050 /*
8051 * Try to concatenate this segment with previous
8052 * one and next one, since we changed policy for
8053 * this one and it may be compatible with adjacent
8054 * ones now.
8055 */
8056 prev = AS_SEGPREV(seg->s_as, seg);
8057 next = AS_SEGNEXT(seg->s_as, seg);
8058
8059 if (next && next->s_ops == &segvn_ops &&
8060 addr + len == next->s_base)
8061 (void) segvn_concat(seg, next, 1);
8062
8063 if (prev && prev->s_ops == &segvn_ops &&
8064 addr == prev->s_base + prev->s_size) {
8065 /*
8066 * Drop lock for private data of current
8067 * segment before concatenating (deleting) it
8068 * and return IE_REATTACH to tell as_ctl() that
8069 * current segment has changed
8070 */
8071 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8072 if (!segvn_concat(prev, seg, 1))
8073 err = IE_REATTACH;
8074
8075 return (err);
8076 }
8077 break;
8078
8079 case MADV_SEQUENTIAL:
8080 /*
8081 * unloading mapping guarantees
8082 * detection in segvn_fault
8083 */
8084 ASSERT(seg->s_szc == 0);
8085 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8086 hat_unload(seg->s_as->a_hat, addr, len,
8087 HAT_UNLOAD);
8088 /* FALLTHROUGH */
8089 case MADV_NORMAL:
8090 case MADV_RANDOM:
8091 svd->advice = (uchar_t)behav;
8092 svd->pageadvice = 0;
8093 break;
8094 case MADV_WILLNEED: /* handled in memcntl */
8095 case MADV_DONTNEED: /* handled in memcntl */
8096 case MADV_FREE: /* handled above */
8097 break;
8098 default:
8099 err = EINVAL;
8100 }
8101 } else {
8102 caddr_t eaddr;
8103 struct seg *new_seg;
8104 struct segvn_data *new_svd;
8105 u_offset_t off;
8106 caddr_t oldeaddr;
8107
8108 page = seg_page(seg, addr);
8109
8110 segvn_vpage(seg);
8111 if (svd->vpage == NULL) {
8112 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8113 return (ENOMEM);
8114 }
8115
8116 switch (behav) {
8117 struct vpage *bvpp, *evpp;
8118
8119 case MADV_ACCESS_LWP:
8120 case MADV_ACCESS_MANY:
8121 case MADV_ACCESS_DEFAULT:
8122 /*
8123 * Set memory allocation policy for portion of this
8124 * segment
8125 */
8126
8127 /*
8128 * Align address and length of advice to page
8129 * boundaries for large pages
8130 */
8131 if (seg->s_szc != 0) {
8132 size_t pgsz;
8133
8134 pgsz = page_get_pagesize(seg->s_szc);
8135 addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8136 len = P2ROUNDUP(len, pgsz);
8137 }
8138
8139 /*
8140 * Check to see whether policy is set already
8141 */
8142 policy = lgrp_madv_to_policy(behav, len, svd->type);
8143
8144 anon_index = svd->anon_index + page;
8145 off = svd->offset + (uintptr_t)(addr - seg->s_base);
8146
8147 if (svd->type == MAP_SHARED)
8148 already_set = lgrp_shm_policy_set(policy, amp,
8149 anon_index, vp, off, len);
8150 else
8151 already_set =
8152 (policy == svd->policy_info.mem_policy);
8153
8154 /*
8155 * If policy set already and it shouldn't be reapplied,
8156 * don't do anything.
8157 */
8158 if (already_set &&
8159 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8160 break;
8161
8162 /*
8163 * For private memory, need writers lock on
8164 * address space because the segment may be
8165 * split or concatenated when changing policy
8166 */
8167 if (svd->type == MAP_PRIVATE &&
8168 AS_READ_HELD(seg->s_as)) {
8169 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8170 return (IE_RETRY);
8171 }
8172
8173 /*
8174 * Mark any existing pages in given range for
8175 * migration
8176 */
8177 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8178 vp, svd->offset, 1);
8179
8180 /*
8181 * Don't need to try to split or concatenate
8182 * segments, since policy is same or this is a shared
8183 * memory segment
8184 */
8185 if (already_set || svd->type == MAP_SHARED)
8186 break;
8187
8188 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8189 ASSERT(svd->amp == NULL);
8190 ASSERT(svd->tr_state == SEGVN_TR_OFF);
8191 ASSERT(svd->softlockcnt == 0);
8192 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8193 HAT_REGION_TEXT);
8194 svd->rcookie = HAT_INVALID_REGION_COOKIE;
8195 }
8196
8197 /*
8198 * Split off new segment if advice only applies to a
8199 * portion of existing segment starting in middle
8200 */
8201 new_seg = NULL;
8202 eaddr = addr + len;
8203 oldeaddr = seg->s_base + seg->s_size;
8204 if (addr > seg->s_base) {
8205 /*
8206 * Must flush I/O page cache
8207 * before splitting segment
8208 */
8209 if (svd->softlockcnt > 0)
8210 segvn_purge(seg);
8211
8212 /*
8213 * Split segment and return IE_REATTACH to tell
8214 * as_ctl() that current segment changed
8215 */
8216 new_seg = segvn_split_seg(seg, addr);
8217 new_svd = (struct segvn_data *)new_seg->s_data;
8218 err = IE_REATTACH;
8219
8220 /*
8221 * If new segment ends where old one
8222 * did, try to concatenate the new
8223 * segment with next one.
8224 */
8225 if (eaddr == oldeaddr) {
8226 /*
8227 * Set policy for new segment
8228 */
8229 (void) lgrp_privm_policy_set(policy,
8230 &new_svd->policy_info,
8231 new_seg->s_size);
8232
8233 next = AS_SEGNEXT(new_seg->s_as,
8234 new_seg);
8235
8236 if (next &&
8237 next->s_ops == &segvn_ops &&
8238 eaddr == next->s_base)
8239 (void) segvn_concat(new_seg,
8240 next, 1);
8241 }
8242 }
8243
8244 /*
8245 * Split off end of existing segment if advice only
8246 * applies to a portion of segment ending before
8247 * end of the existing segment
8248 */
8249 if (eaddr < oldeaddr) {
8250 /*
8251 * Must flush I/O page cache
8252 * before splitting segment
8253 */
8254 if (svd->softlockcnt > 0)
8255 segvn_purge(seg);
8256
8257 /*
8258 * If beginning of old segment was already
8259 * split off, use new segment to split end off
8260 * from.
8261 */
8262 if (new_seg != NULL && new_seg != seg) {
8263 /*
8264 * Split segment
8265 */
8266 (void) segvn_split_seg(new_seg, eaddr);
8267
8268 /*
8269 * Set policy for new segment
8270 */
8271 (void) lgrp_privm_policy_set(policy,
8272 &new_svd->policy_info,
8273 new_seg->s_size);
8274 } else {
8275 /*
8276 * Split segment and return IE_REATTACH
8277 * to tell as_ctl() that current
8278 * segment changed
8279 */
8280 (void) segvn_split_seg(seg, eaddr);
8281 err = IE_REATTACH;
8282
8283 (void) lgrp_privm_policy_set(policy,
8284 &svd->policy_info, seg->s_size);
8285
8286 /*
8287 * If new segment starts where old one
8288 * did, try to concatenate it with
8289 * previous segment.
8290 */
8291 if (addr == seg->s_base) {
8292 prev = AS_SEGPREV(seg->s_as,
8293 seg);
8294
8295 /*
8296 * Drop lock for private data
8297 * of current segment before
8298 * concatenating (deleting) it
8299 */
8300 if (prev &&
8301 prev->s_ops ==
8302 &segvn_ops &&
8303 addr == prev->s_base +
8304 prev->s_size) {
8305 SEGVN_LOCK_EXIT(
8306 seg->s_as,
8307 &svd->lock);
8308 (void) segvn_concat(
8309 prev, seg, 1);
8310 return (err);
8311 }
8312 }
8313 }
8314 }
8315 break;
8316 case MADV_SEQUENTIAL:
8317 ASSERT(seg->s_szc == 0);
8318 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8319 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8320 /* FALLTHROUGH */
8321 case MADV_NORMAL:
8322 case MADV_RANDOM:
8323 bvpp = &svd->vpage[page];
8324 evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8325 for (; bvpp < evpp; bvpp++)
8326 VPP_SETADVICE(bvpp, behav);
8327 svd->advice = MADV_NORMAL;
8328 break;
8329 case MADV_WILLNEED: /* handled in memcntl */
8330 case MADV_DONTNEED: /* handled in memcntl */
8331 case MADV_FREE: /* handled above */
8332 break;
8333 default:
8334 err = EINVAL;
8335 }
8336 }
8337 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8338 return (err);
8339 }
8340
8341 /*
8342 * There is one kind of inheritance that can be specified for pages:
8343 *
8344 * SEGP_INH_ZERO - Pages should be zeroed in the child
8345 */
8346 static int
8347 segvn_inherit(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8348 {
8349 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8350 struct vpage *bvpp, *evpp;
8351 size_t page;
8352 int ret = 0;
8353
8354 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8355
8356 /* Can't support something we don't know about */
8357 if (behav != SEGP_INH_ZERO)
8358 return (ENOTSUP);
8359
8360 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8361
8362 /*
8363 * This must be a straightforward anonymous segment that is mapped
8364 * privately and is not backed by a vnode.
8365 */
8366 if (svd->tr_state != SEGVN_TR_OFF ||
8367 svd->type != MAP_PRIVATE ||
8368 svd->vp != NULL) {
8369 ret = EINVAL;
8370 goto out;
8371 }
8372
8373 /*
8374 * If the entire segment has been marked as inherit zero, then no reason
8375 * to do anything else.
8376 */
8377 if (svd->svn_inz == SEGVN_INZ_ALL) {
8378 ret = 0;
8379 goto out;
8380 }
8381
8382 /*
8383 * If this applies to the entire segment, simply mark it and we're done.
8384 */
8385 if ((addr == seg->s_base) && (len == seg->s_size)) {
8386 svd->svn_inz = SEGVN_INZ_ALL;
8387 ret = 0;
8388 goto out;
8389 }
8390
8391 /*
8392 * We've been asked to mark a subset of this segment as inherit zero,
8393 * therefore we need to mainpulate its vpages.
8394 */
8395 if (svd->vpage == NULL) {
8396 segvn_vpage(seg);
8397 if (svd->vpage == NULL) {
8398 ret = ENOMEM;
8399 goto out;
8400 }
8401 }
8402
8403 svd->svn_inz = SEGVN_INZ_VPP;
8404 page = seg_page(seg, addr);
8405 bvpp = &svd->vpage[page];
8406 evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8407 for (; bvpp < evpp; bvpp++)
8408 VPP_SETINHZERO(bvpp);
8409 ret = 0;
8410
8411 out:
8412 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8413 return (ret);
8414 }
8415
8416 /*
8417 * Create a vpage structure for this seg.
8418 */
8419 static void
8420 segvn_vpage(struct seg *seg)
8421 {
8422 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8423 struct vpage *vp, *evp;
8424 static pgcnt_t page_limit = 0;
8425
8426 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8427
8428 /*
8429 * If no vpage structure exists, allocate one. Copy the protections
8430 * and the advice from the segment itself to the individual pages.
8431 */
8432 if (svd->vpage == NULL) {
8433 /*
8434 * Start by calculating the number of pages we must allocate to
8435 * track the per-page vpage structs needs for this entire
8436 * segment. If we know now that it will require more than our
8437 * heuristic for the maximum amount of kmem we can consume then
8438 * fail. We do this here, instead of trying to detect this deep
8439 * in page_resv and propagating the error up, since the entire
8440 * memory allocation stack is not amenable to passing this
8441 * back. Instead, it wants to keep trying.
8442 *
8443 * As a heuristic we set a page limit of 5/8s of total_pages
8444 * for this allocation. We use shifts so that no floating
8445 * point conversion takes place and only need to do the
8446 * calculation once.
8447 */
8448 ulong_t mem_needed = seg_pages(seg) * sizeof (struct vpage);
8449 pgcnt_t npages = mem_needed >> PAGESHIFT;
8450
8451 if (page_limit == 0)
8452 page_limit = (total_pages >> 1) + (total_pages >> 3);
8453
8454 if (npages > page_limit)
8455 return;
8456
8457 svd->pageadvice = 1;
8458 svd->vpage = kmem_zalloc(mem_needed, KM_SLEEP);
8459 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8460 for (vp = svd->vpage; vp < evp; vp++) {
8461 VPP_SETPROT(vp, svd->prot);
8462 VPP_SETADVICE(vp, svd->advice);
8463 }
8464 }
8465 }
8466
8467 /*
8468 * Dump the pages belonging to this segvn segment.
8469 */
8470 static void
8471 segvn_dump(struct seg *seg)
8472 {
8473 struct segvn_data *svd;
8474 page_t *pp;
8475 struct anon_map *amp;
8476 ulong_t anon_index;
8477 struct vnode *vp;
8478 u_offset_t off, offset;
8479 pfn_t pfn;
8480 pgcnt_t page, npages;
8481 caddr_t addr;
8482
8483 npages = seg_pages(seg);
8484 svd = (struct segvn_data *)seg->s_data;
8485 vp = svd->vp;
8486 off = offset = svd->offset;
8487 addr = seg->s_base;
8488
8489 if ((amp = svd->amp) != NULL) {
8490 anon_index = svd->anon_index;
8491 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
8492 }
8493
8494 for (page = 0; page < npages; page++, offset += PAGESIZE) {
8495 struct anon *ap;
8496 int we_own_it = 0;
8497
8498 if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8499 swap_xlate_nopanic(ap, &vp, &off);
8500 } else {
8501 vp = svd->vp;
8502 off = offset;
8503 }
8504
8505 /*
8506 * If pp == NULL, the page either does not exist
8507 * or is exclusively locked. So determine if it
8508 * exists before searching for it.
8509 */
8510
8511 if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8512 we_own_it = 1;
8513 else
8514 pp = page_exists(vp, off);
8515
8516 if (pp) {
8517 pfn = page_pptonum(pp);
8518 dump_addpage(seg->s_as, addr, pfn);
8519 if (we_own_it)
8520 page_unlock(pp);
8521 }
8522 addr += PAGESIZE;
8523 dump_timeleft = dump_timeout;
8524 }
8525
8526 if (amp != NULL)
8527 ANON_LOCK_EXIT(&->a_rwlock);
8528 }
8529
8530 #ifdef DEBUG
8531 static uint32_t segvn_pglock_mtbf = 0;
8532 #endif
8533
8534 #define PCACHE_SHWLIST ((page_t *)-2)
8535 #define NOPCACHE_SHWLIST ((page_t *)-1)
8536
8537 /*
8538 * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8539 * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8540 * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8541 * the same parts of the segment. Currently shadow list creation is only
8542 * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8543 * tagged with segment pointer, starting virtual address and length. This
8544 * approach for MAP_SHARED segments may add many pcache entries for the same
8545 * set of pages and lead to long hash chains that decrease pcache lookup
8546 * performance. To avoid this issue for shared segments shared anon map and
8547 * starting anon index are used for pcache entry tagging. This allows all
8548 * segments to share pcache entries for the same anon range and reduces pcache
8549 * chain's length as well as memory overhead from duplicate shadow lists and
8550 * pcache entries.
8551 *
8552 * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8553 * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8554 * part of softlockcnt accounting is done differently for private and shared
8555 * segments. In private segment case softlock is only incremented when a new
8556 * shadow list is created but not when an existing one is found via
8557 * seg_plookup(). pcache entries have reference count incremented/decremented
8558 * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8559 * reference count can be purged (and purging is needed before segment can be
8560 * freed). When a private segment pcache entry is purged segvn_reclaim() will
8561 * decrement softlockcnt. Since in private segment case each of its pcache
8562 * entries only belongs to this segment we can expect that when
8563 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8564 * segment purge will succeed and softlockcnt will drop to 0. In shared
8565 * segment case reference count in pcache entry counts active locks from many
8566 * different segments so we can't expect segment purging to succeed even when
8567 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8568 * segment. To be able to determine when there're no pending pagelocks in
8569 * shared segment case we don't rely on purging to make softlockcnt drop to 0
8570 * but instead softlockcnt is incremented and decremented for every
8571 * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8572 * list was created or an existing one was found. When softlockcnt drops to 0
8573 * this segment no longer has any claims for pcached shadow lists and the
8574 * segment can be freed even if there're still active pcache entries
8575 * shared by this segment anon map. Shared segment pcache entries belong to
8576 * anon map and are typically removed when anon map is freed after all
8577 * processes destroy the segments that use this anon map.
8578 */
8579 static int
8580 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8581 enum lock_type type, enum seg_rw rw)
8582 {
8583 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8584 size_t np;
8585 pgcnt_t adjustpages;
8586 pgcnt_t npages;
8587 ulong_t anon_index;
8588 uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8589 uint_t error;
8590 struct anon_map *amp;
8591 pgcnt_t anpgcnt;
8592 struct page **pplist, **pl, *pp;
8593 caddr_t a;
8594 size_t page;
8595 caddr_t lpgaddr, lpgeaddr;
8596 anon_sync_obj_t cookie;
8597 int anlock;
8598 struct anon_map *pamp;
8599 caddr_t paddr;
8600 seg_preclaim_cbfunc_t preclaim_callback;
8601 size_t pgsz;
8602 int use_pcache;
8603 size_t wlen;
8604 uint_t pflags = 0;
8605 int sftlck_sbase = 0;
8606 int sftlck_send = 0;
8607
8608 #ifdef DEBUG
8609 if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8610 hrtime_t ts = gethrtime();
8611 if ((ts % segvn_pglock_mtbf) == 0) {
8612 return (ENOTSUP);
8613 }
8614 if ((ts % segvn_pglock_mtbf) == 1) {
8615 return (EFAULT);
8616 }
8617 }
8618 #endif
8619
8620 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8621 "segvn_pagelock: start seg %p addr %p", seg, addr);
8622
8623 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8624 ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8625
8626 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8627
8628 /*
8629 * for now we only support pagelock to anon memory. We would have to
8630 * check protections for vnode objects and call into the vnode driver.
8631 * That's too much for a fast path. Let the fault entry point handle
8632 * it.
8633 */
8634 if (svd->vp != NULL) {
8635 if (type == L_PAGELOCK) {
8636 error = ENOTSUP;
8637 goto out;
8638 }
8639 panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8640 }
8641 if ((amp = svd->amp) == NULL) {
8642 if (type == L_PAGELOCK) {
8643 error = EFAULT;
8644 goto out;
8645 }
8646 panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8647 }
8648 if (rw != S_READ && rw != S_WRITE) {
8649 if (type == L_PAGELOCK) {
8650 error = ENOTSUP;
8651 goto out;
8652 }
8653 panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8654 }
8655
8656 if (seg->s_szc != 0) {
8657 /*
8658 * We are adjusting the pagelock region to the large page size
8659 * boundary because the unlocked part of a large page cannot
8660 * be freed anyway unless all constituent pages of a large
8661 * page are locked. Bigger regions reduce pcache chain length
8662 * and improve lookup performance. The tradeoff is that the
8663 * very first segvn_pagelock() call for a given page is more
8664 * expensive if only 1 page_t is needed for IO. This is only
8665 * an issue if pcache entry doesn't get reused by several
8666 * subsequent calls. We optimize here for the case when pcache
8667 * is heavily used by repeated IOs to the same address range.
8668 *
8669 * Note segment's page size cannot change while we are holding
8670 * as lock. And then it cannot change while softlockcnt is
8671 * not 0. This will allow us to correctly recalculate large
8672 * page size region for the matching pageunlock/reclaim call
8673 * since as_pageunlock() caller must always match
8674 * as_pagelock() call's addr and len.
8675 *
8676 * For pageunlock *ppp points to the pointer of page_t that
8677 * corresponds to the real unadjusted start address. Similar
8678 * for pagelock *ppp must point to the pointer of page_t that
8679 * corresponds to the real unadjusted start address.
8680 */
8681 pgsz = page_get_pagesize(seg->s_szc);
8682 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8683 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8684 } else if (len < segvn_pglock_comb_thrshld) {
8685 lpgaddr = addr;
8686 lpgeaddr = addr + len;
8687 adjustpages = 0;
8688 pgsz = PAGESIZE;
8689 } else {
8690 /*
8691 * Align the address range of large enough requests to allow
8692 * combining of different shadow lists into 1 to reduce memory
8693 * overhead from potentially overlapping large shadow lists
8694 * (worst case is we have a 1MB IO into buffers with start
8695 * addresses separated by 4K). Alignment is only possible if
8696 * padded chunks have sufficient access permissions. Note
8697 * permissions won't change between L_PAGELOCK and
8698 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8699 * segvn_setprot() to wait until softlockcnt drops to 0. This
8700 * allows us to determine in L_PAGEUNLOCK the same range we
8701 * computed in L_PAGELOCK.
8702 *
8703 * If alignment is limited by segment ends set
8704 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8705 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8706 * per segment counters. In L_PAGEUNLOCK case decrease
8707 * softlockcnt_sbase/softlockcnt_send counters if
8708 * sftlck_sbase/sftlck_send flags are set. When
8709 * softlockcnt_sbase/softlockcnt_send are non 0
8710 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8711 * won't merge the segments. This restriction combined with
8712 * restriction on segment unmapping and splitting for segments
8713 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8714 * correctly determine the same range that was previously
8715 * locked by matching L_PAGELOCK.
8716 */
8717 pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8718 pgsz = PAGESIZE;
8719 if (svd->type == MAP_PRIVATE) {
8720 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8721 segvn_pglock_comb_balign);
8722 if (lpgaddr < seg->s_base) {
8723 lpgaddr = seg->s_base;
8724 sftlck_sbase = 1;
8725 }
8726 } else {
8727 ulong_t aix = svd->anon_index + seg_page(seg, addr);
8728 ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8729 if (aaix < svd->anon_index) {
8730 lpgaddr = seg->s_base;
8731 sftlck_sbase = 1;
8732 } else {
8733 lpgaddr = addr - ptob(aix - aaix);
8734 ASSERT(lpgaddr >= seg->s_base);
8735 }
8736 }
8737 if (svd->pageprot && lpgaddr != addr) {
8738 struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8739 struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8740 while (vp < evp) {
8741 if ((VPP_PROT(vp) & protchk) == 0) {
8742 break;
8743 }
8744 vp++;
8745 }
8746 if (vp < evp) {
8747 lpgaddr = addr;
8748 pflags = 0;
8749 }
8750 }
8751 lpgeaddr = addr + len;
8752 if (pflags) {
8753 if (svd->type == MAP_PRIVATE) {
8754 lpgeaddr = (caddr_t)P2ROUNDUP(
8755 (uintptr_t)lpgeaddr,
8756 segvn_pglock_comb_balign);
8757 } else {
8758 ulong_t aix = svd->anon_index +
8759 seg_page(seg, lpgeaddr);
8760 ulong_t aaix = P2ROUNDUP(aix,
8761 segvn_pglock_comb_palign);
8762 if (aaix < aix) {
8763 lpgeaddr = 0;
8764 } else {
8765 lpgeaddr += ptob(aaix - aix);
8766 }
8767 }
8768 if (lpgeaddr == 0 ||
8769 lpgeaddr > seg->s_base + seg->s_size) {
8770 lpgeaddr = seg->s_base + seg->s_size;
8771 sftlck_send = 1;
8772 }
8773 }
8774 if (svd->pageprot && lpgeaddr != addr + len) {
8775 struct vpage *vp;
8776 struct vpage *evp;
8777
8778 vp = &svd->vpage[seg_page(seg, addr + len)];
8779 evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8780
8781 while (vp < evp) {
8782 if ((VPP_PROT(vp) & protchk) == 0) {
8783 break;
8784 }
8785 vp++;
8786 }
8787 if (vp < evp) {
8788 lpgeaddr = addr + len;
8789 }
8790 }
8791 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8792 }
8793
8794 /*
8795 * For MAP_SHARED segments we create pcache entries tagged by amp and
8796 * anon index so that we can share pcache entries with other segments
8797 * that map this amp. For private segments pcache entries are tagged
8798 * with segment and virtual address.
8799 */
8800 if (svd->type == MAP_SHARED) {
8801 pamp = amp;
8802 paddr = (caddr_t)((lpgaddr - seg->s_base) +
8803 ptob(svd->anon_index));
8804 preclaim_callback = shamp_reclaim;
8805 } else {
8806 pamp = NULL;
8807 paddr = lpgaddr;
8808 preclaim_callback = segvn_reclaim;
8809 }
8810
8811 if (type == L_PAGEUNLOCK) {
8812 VM_STAT_ADD(segvnvmstats.pagelock[0]);
8813
8814 /*
8815 * update hat ref bits for /proc. We need to make sure
8816 * that threads tracing the ref and mod bits of the
8817 * address space get the right data.
8818 * Note: page ref and mod bits are updated at reclaim time
8819 */
8820 if (seg->s_as->a_vbits) {
8821 for (a = addr; a < addr + len; a += PAGESIZE) {
8822 if (rw == S_WRITE) {
8823 hat_setstat(seg->s_as, a,
8824 PAGESIZE, P_REF | P_MOD);
8825 } else {
8826 hat_setstat(seg->s_as, a,
8827 PAGESIZE, P_REF);
8828 }
8829 }
8830 }
8831
8832 /*
8833 * Check the shadow list entry after the last page used in
8834 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8835 * was not inserted into pcache and is not large page
8836 * adjusted. In this case call reclaim callback directly and
8837 * don't adjust the shadow list start and size for large
8838 * pages.
8839 */
8840 npages = btop(len);
8841 if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
8842 void *ptag;
8843 if (pamp != NULL) {
8844 ASSERT(svd->type == MAP_SHARED);
8845 ptag = (void *)pamp;
8846 paddr = (caddr_t)((addr - seg->s_base) +
8847 ptob(svd->anon_index));
8848 } else {
8849 ptag = (void *)seg;
8850 paddr = addr;
8851 }
8852 (*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
8853 } else {
8854 ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
8855 IS_SWAPFSVP((*ppp)[npages]->p_vnode));
8856 len = lpgeaddr - lpgaddr;
8857 npages = btop(len);
8858 seg_pinactive(seg, pamp, paddr, len,
8859 *ppp - adjustpages, rw, pflags, preclaim_callback);
8860 }
8861
8862 if (pamp != NULL) {
8863 ASSERT(svd->type == MAP_SHARED);
8864 ASSERT(svd->softlockcnt >= npages);
8865 atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
8866 }
8867
8868 if (sftlck_sbase) {
8869 ASSERT(svd->softlockcnt_sbase > 0);
8870 atomic_dec_ulong((ulong_t *)&svd->softlockcnt_sbase);
8871 }
8872 if (sftlck_send) {
8873 ASSERT(svd->softlockcnt_send > 0);
8874 atomic_dec_ulong((ulong_t *)&svd->softlockcnt_send);
8875 }
8876
8877 /*
8878 * If someone is blocked while unmapping, we purge
8879 * segment page cache and thus reclaim pplist synchronously
8880 * without waiting for seg_pasync_thread. This speeds up
8881 * unmapping in cases where munmap(2) is called, while
8882 * raw async i/o is still in progress or where a thread
8883 * exits on data fault in a multithreaded application.
8884 */
8885 if (AS_ISUNMAPWAIT(seg->s_as)) {
8886 if (svd->softlockcnt == 0) {
8887 mutex_enter(&seg->s_as->a_contents);
8888 if (AS_ISUNMAPWAIT(seg->s_as)) {
8889 AS_CLRUNMAPWAIT(seg->s_as);
8890 cv_broadcast(&seg->s_as->a_cv);
8891 }
8892 mutex_exit(&seg->s_as->a_contents);
8893 } else if (pamp == NULL) {
8894 /*
8895 * softlockcnt is not 0 and this is a
8896 * MAP_PRIVATE segment. Try to purge its
8897 * pcache entries to reduce softlockcnt.
8898 * If it drops to 0 segvn_reclaim()
8899 * will wake up a thread waiting on
8900 * unmapwait flag.
8901 *
8902 * We don't purge MAP_SHARED segments with non
8903 * 0 softlockcnt since IO is still in progress
8904 * for such segments.
8905 */
8906 ASSERT(svd->type == MAP_PRIVATE);
8907 segvn_purge(seg);
8908 }
8909 }
8910 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8911 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8912 "segvn_pagelock: unlock seg %p addr %p", seg, addr);
8913 return (0);
8914 }
8915
8916 /* The L_PAGELOCK case ... */
8917
8918 VM_STAT_ADD(segvnvmstats.pagelock[1]);
8919
8920 /*
8921 * For MAP_SHARED segments we have to check protections before
8922 * seg_plookup() since pcache entries may be shared by many segments
8923 * with potentially different page protections.
8924 */
8925 if (pamp != NULL) {
8926 ASSERT(svd->type == MAP_SHARED);
8927 if (svd->pageprot == 0) {
8928 if ((svd->prot & protchk) == 0) {
8929 error = EACCES;
8930 goto out;
8931 }
8932 } else {
8933 /*
8934 * check page protections
8935 */
8936 caddr_t ea;
8937
8938 if (seg->s_szc) {
8939 a = lpgaddr;
8940 ea = lpgeaddr;
8941 } else {
8942 a = addr;
8943 ea = addr + len;
8944 }
8945 for (; a < ea; a += pgsz) {
8946 struct vpage *vp;
8947
8948 ASSERT(seg->s_szc == 0 ||
8949 sameprot(seg, a, pgsz));
8950 vp = &svd->vpage[seg_page(seg, a)];
8951 if ((VPP_PROT(vp) & protchk) == 0) {
8952 error = EACCES;
8953 goto out;
8954 }
8955 }
8956 }
8957 }
8958
8959 /*
8960 * try to find pages in segment page cache
8961 */
8962 pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
8963 if (pplist != NULL) {
8964 if (pamp != NULL) {
8965 npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
8966 ASSERT(svd->type == MAP_SHARED);
8967 atomic_add_long((ulong_t *)&svd->softlockcnt,
8968 npages);
8969 }
8970 if (sftlck_sbase) {
8971 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
8972 }
8973 if (sftlck_send) {
8974 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
8975 }
8976 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8977 *ppp = pplist + adjustpages;
8978 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
8979 "segvn_pagelock: cache hit seg %p addr %p", seg, addr);
8980 return (0);
8981 }
8982
8983 /*
8984 * For MAP_SHARED segments we already verified above that segment
8985 * protections allow this pagelock operation.
8986 */
8987 if (pamp == NULL) {
8988 ASSERT(svd->type == MAP_PRIVATE);
8989 if (svd->pageprot == 0) {
8990 if ((svd->prot & protchk) == 0) {
8991 error = EACCES;
8992 goto out;
8993 }
8994 if (svd->prot & PROT_WRITE) {
8995 wlen = lpgeaddr - lpgaddr;
8996 } else {
8997 wlen = 0;
8998 ASSERT(rw == S_READ);
8999 }
9000 } else {
9001 int wcont = 1;
9002 /*
9003 * check page protections
9004 */
9005 for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
9006 struct vpage *vp;
9007
9008 ASSERT(seg->s_szc == 0 ||
9009 sameprot(seg, a, pgsz));
9010 vp = &svd->vpage[seg_page(seg, a)];
9011 if ((VPP_PROT(vp) & protchk) == 0) {
9012 error = EACCES;
9013 goto out;
9014 }
9015 if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
9016 wlen += pgsz;
9017 } else {
9018 wcont = 0;
9019 ASSERT(rw == S_READ);
9020 }
9021 }
9022 }
9023 ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
9024 ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9025 }
9026
9027 /*
9028 * Only build large page adjusted shadow list if we expect to insert
9029 * it into pcache. For large enough pages it's a big overhead to
9030 * create a shadow list of the entire large page. But this overhead
9031 * should be amortized over repeated pcache hits on subsequent reuse
9032 * of this shadow list (IO into any range within this shadow list will
9033 * find it in pcache since we large page align the request for pcache
9034 * lookups). pcache performance is improved with bigger shadow lists
9035 * as it reduces the time to pcache the entire big segment and reduces
9036 * pcache chain length.
9037 */
9038 if (seg_pinsert_check(seg, pamp, paddr,
9039 lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9040 addr = lpgaddr;
9041 len = lpgeaddr - lpgaddr;
9042 use_pcache = 1;
9043 } else {
9044 use_pcache = 0;
9045 /*
9046 * Since this entry will not be inserted into the pcache, we
9047 * will not do any adjustments to the starting address or
9048 * size of the memory to be locked.
9049 */
9050 adjustpages = 0;
9051 }
9052 npages = btop(len);
9053
9054 pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9055 pl = pplist;
9056 *ppp = pplist + adjustpages;
9057 /*
9058 * If use_pcache is 0 this shadow list is not large page adjusted.
9059 * Record this info in the last entry of shadow array so that
9060 * L_PAGEUNLOCK can determine if it should large page adjust the
9061 * address range to find the real range that was locked.
9062 */
9063 pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9064
9065 page = seg_page(seg, addr);
9066 anon_index = svd->anon_index + page;
9067
9068 anlock = 0;
9069 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
9070 ASSERT(amp->a_szc >= seg->s_szc);
9071 anpgcnt = page_get_pagecnt(amp->a_szc);
9072 for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9073 struct anon *ap;
9074 struct vnode *vp;
9075 u_offset_t off;
9076
9077 /*
9078 * Lock and unlock anon array only once per large page.
9079 * anon_array_enter() locks the root anon slot according to
9080 * a_szc which can't change while anon map is locked. We lock
9081 * anon the first time through this loop and each time we
9082 * reach anon index that corresponds to a root of a large
9083 * page.
9084 */
9085 if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9086 ASSERT(anlock == 0);
9087 anon_array_enter(amp, anon_index, &cookie);
9088 anlock = 1;
9089 }
9090 ap = anon_get_ptr(amp->ahp, anon_index);
9091
9092 /*
9093 * We must never use seg_pcache for COW pages
9094 * because we might end up with original page still
9095 * lying in seg_pcache even after private page is
9096 * created. This leads to data corruption as
9097 * aio_write refers to the page still in cache
9098 * while all other accesses refer to the private
9099 * page.
9100 */
9101 if (ap == NULL || ap->an_refcnt != 1) {
9102 struct vpage *vpage;
9103
9104 if (seg->s_szc) {
9105 error = EFAULT;
9106 break;
9107 }
9108 if (svd->vpage != NULL) {
9109 vpage = &svd->vpage[seg_page(seg, a)];
9110 } else {
9111 vpage = NULL;
9112 }
9113 ASSERT(anlock);
9114 anon_array_exit(&cookie);
9115 anlock = 0;
9116 pp = NULL;
9117 error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9118 vpage, &pp, 0, F_INVAL, rw, 1);
9119 if (error) {
9120 error = fc_decode(error);
9121 break;
9122 }
9123 anon_array_enter(amp, anon_index, &cookie);
9124 anlock = 1;
9125 ap = anon_get_ptr(amp->ahp, anon_index);
9126 if (ap == NULL || ap->an_refcnt != 1) {
9127 error = EFAULT;
9128 break;
9129 }
9130 }
9131 swap_xlate(ap, &vp, &off);
9132 pp = page_lookup_nowait(vp, off, SE_SHARED);
9133 if (pp == NULL) {
9134 error = EFAULT;
9135 break;
9136 }
9137 if (ap->an_pvp != NULL) {
9138 anon_swap_free(ap, pp);
9139 }
9140 /*
9141 * Unlock anon if this is the last slot in a large page.
9142 */
9143 if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9144 ASSERT(anlock);
9145 anon_array_exit(&cookie);
9146 anlock = 0;
9147 }
9148 *pplist++ = pp;
9149 }
9150 if (anlock) { /* Ensure the lock is dropped */
9151 anon_array_exit(&cookie);
9152 }
9153 ANON_LOCK_EXIT(&->a_rwlock);
9154
9155 if (a >= addr + len) {
9156 atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9157 if (pamp != NULL) {
9158 ASSERT(svd->type == MAP_SHARED);
9159 atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9160 npages);
9161 wlen = len;
9162 }
9163 if (sftlck_sbase) {
9164 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9165 }
9166 if (sftlck_send) {
9167 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9168 }
9169 if (use_pcache) {
9170 (void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9171 rw, pflags, preclaim_callback);
9172 }
9173 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9174 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
9175 "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
9176 return (0);
9177 }
9178
9179 pplist = pl;
9180 np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9181 while (np > (uint_t)0) {
9182 ASSERT(PAGE_LOCKED(*pplist));
9183 page_unlock(*pplist);
9184 np--;
9185 pplist++;
9186 }
9187 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9188 out:
9189 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9190 *ppp = NULL;
9191 TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
9192 "segvn_pagelock: cache miss seg %p addr %p", seg, addr);
9193 return (error);
9194 }
9195
9196 /*
9197 * purge any cached pages in the I/O page cache
9198 */
9199 static void
9200 segvn_purge(struct seg *seg)
9201 {
9202 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9203
9204 /*
9205 * pcache is only used by pure anon segments.
9206 */
9207 if (svd->amp == NULL || svd->vp != NULL) {
9208 return;
9209 }
9210
9211 /*
9212 * For MAP_SHARED segments non 0 segment's softlockcnt means
9213 * active IO is still in progress via this segment. So we only
9214 * purge MAP_SHARED segments when their softlockcnt is 0.
9215 */
9216 if (svd->type == MAP_PRIVATE) {
9217 if (svd->softlockcnt) {
9218 seg_ppurge(seg, NULL, 0);
9219 }
9220 } else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9221 seg_ppurge(seg, svd->amp, 0);
9222 }
9223 }
9224
9225 /*
9226 * If async argument is not 0 we are called from pcache async thread and don't
9227 * hold AS lock.
9228 */
9229
9230 /*ARGSUSED*/
9231 static int
9232 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9233 enum seg_rw rw, int async)
9234 {
9235 struct seg *seg = (struct seg *)ptag;
9236 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9237 pgcnt_t np, npages;
9238 struct page **pl;
9239
9240 npages = np = btop(len);
9241 ASSERT(npages);
9242
9243 ASSERT(svd->vp == NULL && svd->amp != NULL);
9244 ASSERT(svd->softlockcnt >= npages);
9245 ASSERT(async || AS_LOCK_HELD(seg->s_as));
9246
9247 pl = pplist;
9248
9249 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9250 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9251
9252 while (np > (uint_t)0) {
9253 if (rw == S_WRITE) {
9254 hat_setrefmod(*pplist);
9255 } else {
9256 hat_setref(*pplist);
9257 }
9258 page_unlock(*pplist);
9259 np--;
9260 pplist++;
9261 }
9262
9263 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9264
9265 /*
9266 * If we are pcache async thread we don't hold AS lock. This means if
9267 * softlockcnt drops to 0 after the decrement below address space may
9268 * get freed. We can't allow it since after softlock derement to 0 we
9269 * still need to access as structure for possible wakeup of unmap
9270 * waiters. To prevent the disappearance of as we take this segment
9271 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9272 * make sure this routine completes before segment is freed.
9273 *
9274 * The second complication we have to deal with in async case is a
9275 * possibility of missed wake up of unmap wait thread. When we don't
9276 * hold as lock here we may take a_contents lock before unmap wait
9277 * thread that was first to see softlockcnt was still not 0. As a
9278 * result we'll fail to wake up an unmap wait thread. To avoid this
9279 * race we set nounmapwait flag in as structure if we drop softlockcnt
9280 * to 0 when we were called by pcache async thread. unmapwait thread
9281 * will not block if this flag is set.
9282 */
9283 if (async) {
9284 mutex_enter(&svd->segfree_syncmtx);
9285 }
9286
9287 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9288 if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9289 mutex_enter(&seg->s_as->a_contents);
9290 if (async) {
9291 AS_SETNOUNMAPWAIT(seg->s_as);
9292 }
9293 if (AS_ISUNMAPWAIT(seg->s_as)) {
9294 AS_CLRUNMAPWAIT(seg->s_as);
9295 cv_broadcast(&seg->s_as->a_cv);
9296 }
9297 mutex_exit(&seg->s_as->a_contents);
9298 }
9299 }
9300
9301 if (async) {
9302 mutex_exit(&svd->segfree_syncmtx);
9303 }
9304 return (0);
9305 }
9306
9307 /*ARGSUSED*/
9308 static int
9309 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9310 enum seg_rw rw, int async)
9311 {
9312 amp_t *amp = (amp_t *)ptag;
9313 pgcnt_t np, npages;
9314 struct page **pl;
9315
9316 npages = np = btop(len);
9317 ASSERT(npages);
9318 ASSERT(amp->a_softlockcnt >= npages);
9319
9320 pl = pplist;
9321
9322 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9323 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9324
9325 while (np > (uint_t)0) {
9326 if (rw == S_WRITE) {
9327 hat_setrefmod(*pplist);
9328 } else {
9329 hat_setref(*pplist);
9330 }
9331 page_unlock(*pplist);
9332 np--;
9333 pplist++;
9334 }
9335
9336 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9337
9338 /*
9339 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9340 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9341 * and anonmap_purge() acquires a_purgemtx.
9342 */
9343 mutex_enter(&->a_purgemtx);
9344 if (!atomic_add_long_nv((ulong_t *)&->a_softlockcnt, -npages) &&
9345 amp->a_purgewait) {
9346 amp->a_purgewait = 0;
9347 cv_broadcast(&->a_purgecv);
9348 }
9349 mutex_exit(&->a_purgemtx);
9350 return (0);
9351 }
9352
9353 /*
9354 * get a memory ID for an addr in a given segment
9355 *
9356 * XXX only creates PAGESIZE pages if anon slots are not initialized.
9357 * At fault time they will be relocated into larger pages.
9358 */
9359 static int
9360 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9361 {
9362 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9363 struct anon *ap = NULL;
9364 ulong_t anon_index;
9365 struct anon_map *amp;
9366 anon_sync_obj_t cookie;
9367
9368 if (svd->type == MAP_PRIVATE) {
9369 memidp->val[0] = (uintptr_t)seg->s_as;
9370 memidp->val[1] = (uintptr_t)addr;
9371 return (0);
9372 }
9373
9374 if (svd->type == MAP_SHARED) {
9375 if (svd->vp) {
9376 memidp->val[0] = (uintptr_t)svd->vp;
9377 memidp->val[1] = (u_longlong_t)svd->offset +
9378 (uintptr_t)(addr - seg->s_base);
9379 return (0);
9380 } else {
9381
9382 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9383 if ((amp = svd->amp) != NULL) {
9384 anon_index = svd->anon_index +
9385 seg_page(seg, addr);
9386 }
9387 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9388
9389 ASSERT(amp != NULL);
9390
9391 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
9392 anon_array_enter(amp, anon_index, &cookie);
9393 ap = anon_get_ptr(amp->ahp, anon_index);
9394 if (ap == NULL) {
9395 page_t *pp;
9396
9397 pp = anon_zero(seg, addr, &ap, svd->cred);
9398 if (pp == NULL) {
9399 anon_array_exit(&cookie);
9400 ANON_LOCK_EXIT(&->a_rwlock);
9401 return (ENOMEM);
9402 }
9403 ASSERT(anon_get_ptr(amp->ahp, anon_index)
9404 == NULL);
9405 (void) anon_set_ptr(amp->ahp, anon_index,
9406 ap, ANON_SLEEP);
9407 page_unlock(pp);
9408 }
9409
9410 anon_array_exit(&cookie);
9411 ANON_LOCK_EXIT(&->a_rwlock);
9412
9413 memidp->val[0] = (uintptr_t)ap;
9414 memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9415 return (0);
9416 }
9417 }
9418 return (EINVAL);
9419 }
9420
9421 static int
9422 sameprot(struct seg *seg, caddr_t a, size_t len)
9423 {
9424 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9425 struct vpage *vpage;
9426 spgcnt_t pages = btop(len);
9427 uint_t prot;
9428
9429 if (svd->pageprot == 0)
9430 return (1);
9431
9432 ASSERT(svd->vpage != NULL);
9433
9434 vpage = &svd->vpage[seg_page(seg, a)];
9435 prot = VPP_PROT(vpage);
9436 vpage++;
9437 pages--;
9438 while (pages-- > 0) {
9439 if (prot != VPP_PROT(vpage))
9440 return (0);
9441 vpage++;
9442 }
9443 return (1);
9444 }
9445
9446 /*
9447 * Get memory allocation policy info for specified address in given segment
9448 */
9449 static lgrp_mem_policy_info_t *
9450 segvn_getpolicy(struct seg *seg, caddr_t addr)
9451 {
9452 struct anon_map *amp;
9453 ulong_t anon_index;
9454 lgrp_mem_policy_info_t *policy_info;
9455 struct segvn_data *svn_data;
9456 u_offset_t vn_off;
9457 vnode_t *vp;
9458
9459 ASSERT(seg != NULL);
9460
9461 svn_data = (struct segvn_data *)seg->s_data;
9462 if (svn_data == NULL)
9463 return (NULL);
9464
9465 /*
9466 * Get policy info for private or shared memory
9467 */
9468 if (svn_data->type != MAP_SHARED) {
9469 if (svn_data->tr_state != SEGVN_TR_ON) {
9470 policy_info = &svn_data->policy_info;
9471 } else {
9472 policy_info = &svn_data->tr_policy_info;
9473 ASSERT(policy_info->mem_policy ==
9474 LGRP_MEM_POLICY_NEXT_SEG);
9475 }
9476 } else {
9477 amp = svn_data->amp;
9478 anon_index = svn_data->anon_index + seg_page(seg, addr);
9479 vp = svn_data->vp;
9480 vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9481 policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9482 }
9483
9484 return (policy_info);
9485 }
9486
9487 /*ARGSUSED*/
9488 static int
9489 segvn_capable(struct seg *seg, segcapability_t capability)
9490 {
9491 return (0);
9492 }
9493
9494 /*
9495 * Bind text vnode segment to an amp. If we bind successfully mappings will be
9496 * established to per vnode mapping per lgroup amp pages instead of to vnode
9497 * pages. There's one amp per vnode text mapping per lgroup. Many processes
9498 * may share the same text replication amp. If a suitable amp doesn't already
9499 * exist in svntr hash table create a new one. We may fail to bind to amp if
9500 * segment is not eligible for text replication. Code below first checks for
9501 * these conditions. If binding is successful segment tr_state is set to on
9502 * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9503 * svd->amp remains as NULL.
9504 */
9505 static void
9506 segvn_textrepl(struct seg *seg)
9507 {
9508 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9509 vnode_t *vp = svd->vp;
9510 u_offset_t off = svd->offset;
9511 size_t size = seg->s_size;
9512 u_offset_t eoff = off + size;
9513 uint_t szc = seg->s_szc;
9514 ulong_t hash = SVNTR_HASH_FUNC(vp);
9515 svntr_t *svntrp;
9516 struct vattr va;
9517 proc_t *p = seg->s_as->a_proc;
9518 lgrp_id_t lgrp_id;
9519 lgrp_id_t olid;
9520 int first;
9521 struct anon_map *amp;
9522
9523 ASSERT(AS_LOCK_HELD(seg->s_as));
9524 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9525 ASSERT(p != NULL);
9526 ASSERT(svd->tr_state == SEGVN_TR_INIT);
9527 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9528 ASSERT(svd->flags & MAP_TEXT);
9529 ASSERT(svd->type == MAP_PRIVATE);
9530 ASSERT(vp != NULL && svd->amp == NULL);
9531 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9532 ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9533 ASSERT(seg->s_as != &kas);
9534 ASSERT(off < eoff);
9535 ASSERT(svntr_hashtab != NULL);
9536
9537 /*
9538 * If numa optimizations are no longer desired bail out.
9539 */
9540 if (!lgrp_optimizations()) {
9541 svd->tr_state = SEGVN_TR_OFF;
9542 return;
9543 }
9544
9545 /*
9546 * Avoid creating anon maps with size bigger than the file size.
9547 * If VOP_GETATTR() call fails bail out.
9548 */
9549 va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9550 if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9551 svd->tr_state = SEGVN_TR_OFF;
9552 SEGVN_TR_ADDSTAT(gaerr);
9553 return;
9554 }
9555 if (btopr(va.va_size) < btopr(eoff)) {
9556 svd->tr_state = SEGVN_TR_OFF;
9557 SEGVN_TR_ADDSTAT(overmap);
9558 return;
9559 }
9560
9561 /*
9562 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9563 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9564 * mapping that checks if trcache for this vnode needs to be
9565 * invalidated can't miss us.
9566 */
9567 if (!(vp->v_flag & VVMEXEC)) {
9568 mutex_enter(&vp->v_lock);
9569 vp->v_flag |= VVMEXEC;
9570 mutex_exit(&vp->v_lock);
9571 }
9572 mutex_enter(&svntr_hashtab[hash].tr_lock);
9573 /*
9574 * Bail out if potentially MAP_SHARED writable mappings exist to this
9575 * vnode. We don't want to use old file contents from existing
9576 * replicas if this mapping was established after the original file
9577 * was changed.
9578 */
9579 if (vn_is_mapped(vp, V_WRITE)) {
9580 mutex_exit(&svntr_hashtab[hash].tr_lock);
9581 svd->tr_state = SEGVN_TR_OFF;
9582 SEGVN_TR_ADDSTAT(wrcnt);
9583 return;
9584 }
9585 svntrp = svntr_hashtab[hash].tr_head;
9586 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9587 ASSERT(svntrp->tr_refcnt != 0);
9588 if (svntrp->tr_vp != vp) {
9589 continue;
9590 }
9591
9592 /*
9593 * Bail out if the file or its attributes were changed after
9594 * this replication entry was created since we need to use the
9595 * latest file contents. Note that mtime test alone is not
9596 * sufficient because a user can explicitly change mtime via
9597 * utimes(2) interfaces back to the old value after modifiying
9598 * the file contents. To detect this case we also have to test
9599 * ctime which among other things records the time of the last
9600 * mtime change by utimes(2). ctime is not changed when the file
9601 * is only read or executed so we expect that typically existing
9602 * replication amp's can be used most of the time.
9603 */
9604 if (!svntrp->tr_valid ||
9605 svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9606 svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9607 svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9608 svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9609 mutex_exit(&svntr_hashtab[hash].tr_lock);
9610 svd->tr_state = SEGVN_TR_OFF;
9611 SEGVN_TR_ADDSTAT(stale);
9612 return;
9613 }
9614 /*
9615 * if off, eoff and szc match current segment we found the
9616 * existing entry we can use.
9617 */
9618 if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9619 svntrp->tr_szc == szc) {
9620 break;
9621 }
9622 /*
9623 * Don't create different but overlapping in file offsets
9624 * entries to avoid replication of the same file pages more
9625 * than once per lgroup.
9626 */
9627 if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9628 (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9629 mutex_exit(&svntr_hashtab[hash].tr_lock);
9630 svd->tr_state = SEGVN_TR_OFF;
9631 SEGVN_TR_ADDSTAT(overlap);
9632 return;
9633 }
9634 }
9635 /*
9636 * If we didn't find existing entry create a new one.
9637 */
9638 if (svntrp == NULL) {
9639 svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9640 if (svntrp == NULL) {
9641 mutex_exit(&svntr_hashtab[hash].tr_lock);
9642 svd->tr_state = SEGVN_TR_OFF;
9643 SEGVN_TR_ADDSTAT(nokmem);
9644 return;
9645 }
9646 #ifdef DEBUG
9647 {
9648 lgrp_id_t i;
9649 for (i = 0; i < NLGRPS_MAX; i++) {
9650 ASSERT(svntrp->tr_amp[i] == NULL);
9651 }
9652 }
9653 #endif /* DEBUG */
9654 svntrp->tr_vp = vp;
9655 svntrp->tr_off = off;
9656 svntrp->tr_eoff = eoff;
9657 svntrp->tr_szc = szc;
9658 svntrp->tr_valid = 1;
9659 svntrp->tr_mtime = va.va_mtime;
9660 svntrp->tr_ctime = va.va_ctime;
9661 svntrp->tr_refcnt = 0;
9662 svntrp->tr_next = svntr_hashtab[hash].tr_head;
9663 svntr_hashtab[hash].tr_head = svntrp;
9664 }
9665 first = 1;
9666 again:
9667 /*
9668 * We want to pick a replica with pages on main thread's (t_tid = 1,
9669 * aka T1) lgrp. Currently text replication is only optimized for
9670 * workloads that either have all threads of a process on the same
9671 * lgrp or execute their large text primarily on main thread.
9672 */
9673 lgrp_id = p->p_t1_lgrpid;
9674 if (lgrp_id == LGRP_NONE) {
9675 /*
9676 * In case exec() prefaults text on non main thread use
9677 * current thread lgrpid. It will become main thread anyway
9678 * soon.
9679 */
9680 lgrp_id = lgrp_home_id(curthread);
9681 }
9682 /*
9683 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet. Otherwise
9684 * just set it to NLGRPS_MAX if it's different from current process T1
9685 * home lgrp. p_tr_lgrpid is used to detect if process uses text
9686 * replication and T1 new home is different from lgrp used for text
9687 * replication. When this happens asyncronous segvn thread rechecks if
9688 * segments should change lgrps used for text replication. If we fail
9689 * to set p_tr_lgrpid with atomic_cas_32 then set it to NLGRPS_MAX
9690 * without cas if it's not already NLGRPS_MAX and not equal lgrp_id
9691 * we want to use. We don't need to use cas in this case because
9692 * another thread that races in between our non atomic check and set
9693 * may only change p_tr_lgrpid to NLGRPS_MAX at this point.
9694 */
9695 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9696 olid = p->p_tr_lgrpid;
9697 if (lgrp_id != olid && olid != NLGRPS_MAX) {
9698 lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9699 if (atomic_cas_32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) !=
9700 olid) {
9701 olid = p->p_tr_lgrpid;
9702 ASSERT(olid != LGRP_NONE);
9703 if (olid != lgrp_id && olid != NLGRPS_MAX) {
9704 p->p_tr_lgrpid = NLGRPS_MAX;
9705 }
9706 }
9707 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9708 membar_producer();
9709 /*
9710 * lgrp_move_thread() won't schedule async recheck after
9711 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9712 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9713 * is not LGRP_NONE.
9714 */
9715 if (first && p->p_t1_lgrpid != LGRP_NONE &&
9716 p->p_t1_lgrpid != lgrp_id) {
9717 first = 0;
9718 goto again;
9719 }
9720 }
9721 /*
9722 * If no amp was created yet for lgrp_id create a new one as long as
9723 * we have enough memory to afford it.
9724 */
9725 if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9726 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9727 if (trmem > segvn_textrepl_max_bytes) {
9728 SEGVN_TR_ADDSTAT(normem);
9729 goto fail;
9730 }
9731 if (anon_try_resv_zone(size, NULL) == 0) {
9732 SEGVN_TR_ADDSTAT(noanon);
9733 goto fail;
9734 }
9735 amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9736 if (amp == NULL) {
9737 anon_unresv_zone(size, NULL);
9738 SEGVN_TR_ADDSTAT(nokmem);
9739 goto fail;
9740 }
9741 ASSERT(amp->refcnt == 1);
9742 amp->a_szc = szc;
9743 svntrp->tr_amp[lgrp_id] = amp;
9744 SEGVN_TR_ADDSTAT(newamp);
9745 }
9746 svntrp->tr_refcnt++;
9747 ASSERT(svd->svn_trnext == NULL);
9748 ASSERT(svd->svn_trprev == NULL);
9749 svd->svn_trnext = svntrp->tr_svnhead;
9750 svd->svn_trprev = NULL;
9751 if (svntrp->tr_svnhead != NULL) {
9752 svntrp->tr_svnhead->svn_trprev = svd;
9753 }
9754 svntrp->tr_svnhead = svd;
9755 ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9756 ASSERT(amp->refcnt >= 1);
9757 svd->amp = amp;
9758 svd->anon_index = 0;
9759 svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9760 svd->tr_policy_info.mem_lgrpid = lgrp_id;
9761 svd->tr_state = SEGVN_TR_ON;
9762 mutex_exit(&svntr_hashtab[hash].tr_lock);
9763 SEGVN_TR_ADDSTAT(repl);
9764 return;
9765 fail:
9766 ASSERT(segvn_textrepl_bytes >= size);
9767 atomic_add_long(&segvn_textrepl_bytes, -size);
9768 ASSERT(svntrp != NULL);
9769 ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9770 if (svntrp->tr_refcnt == 0) {
9771 ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9772 svntr_hashtab[hash].tr_head = svntrp->tr_next;
9773 mutex_exit(&svntr_hashtab[hash].tr_lock);
9774 kmem_cache_free(svntr_cache, svntrp);
9775 } else {
9776 mutex_exit(&svntr_hashtab[hash].tr_lock);
9777 }
9778 svd->tr_state = SEGVN_TR_OFF;
9779 }
9780
9781 /*
9782 * Convert seg back to regular vnode mapping seg by unbinding it from its text
9783 * replication amp. This routine is most typically called when segment is
9784 * unmapped but can also be called when segment no longer qualifies for text
9785 * replication (e.g. due to protection changes). If unload_unmap is set use
9786 * HAT_UNLOAD_UNMAP flag in hat_unload_callback(). If we are the last user of
9787 * svntr free all its anon maps and remove it from the hash table.
9788 */
9789 static void
9790 segvn_textunrepl(struct seg *seg, int unload_unmap)
9791 {
9792 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9793 vnode_t *vp = svd->vp;
9794 u_offset_t off = svd->offset;
9795 size_t size = seg->s_size;
9796 u_offset_t eoff = off + size;
9797 uint_t szc = seg->s_szc;
9798 ulong_t hash = SVNTR_HASH_FUNC(vp);
9799 svntr_t *svntrp;
9800 svntr_t **prv_svntrp;
9801 lgrp_id_t lgrp_id = svd->tr_policy_info.mem_lgrpid;
9802 lgrp_id_t i;
9803
9804 ASSERT(AS_LOCK_HELD(seg->s_as));
9805 ASSERT(AS_WRITE_HELD(seg->s_as) ||
9806 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9807 ASSERT(svd->tr_state == SEGVN_TR_ON);
9808 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9809 ASSERT(svd->amp != NULL);
9810 ASSERT(svd->amp->refcnt >= 1);
9811 ASSERT(svd->anon_index == 0);
9812 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9813 ASSERT(svntr_hashtab != NULL);
9814
9815 mutex_enter(&svntr_hashtab[hash].tr_lock);
9816 prv_svntrp = &svntr_hashtab[hash].tr_head;
9817 for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9818 ASSERT(svntrp->tr_refcnt != 0);
9819 if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9820 svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9821 break;
9822 }
9823 }
9824 if (svntrp == NULL) {
9825 panic("segvn_textunrepl: svntr record not found");
9826 }
9827 if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9828 panic("segvn_textunrepl: amp mismatch");
9829 }
9830 svd->tr_state = SEGVN_TR_OFF;
9831 svd->amp = NULL;
9832 if (svd->svn_trprev == NULL) {
9833 ASSERT(svntrp->tr_svnhead == svd);
9834 svntrp->tr_svnhead = svd->svn_trnext;
9835 if (svntrp->tr_svnhead != NULL) {
9836 svntrp->tr_svnhead->svn_trprev = NULL;
9837 }
9838 svd->svn_trnext = NULL;
9839 } else {
9840 svd->svn_trprev->svn_trnext = svd->svn_trnext;
9841 if (svd->svn_trnext != NULL) {
9842 svd->svn_trnext->svn_trprev = svd->svn_trprev;
9843 svd->svn_trnext = NULL;
9844 }
9845 svd->svn_trprev = NULL;
9846 }
9847 if (--svntrp->tr_refcnt) {
9848 mutex_exit(&svntr_hashtab[hash].tr_lock);
9849 goto done;
9850 }
9851 *prv_svntrp = svntrp->tr_next;
9852 mutex_exit(&svntr_hashtab[hash].tr_lock);
9853 for (i = 0; i < NLGRPS_MAX; i++) {
9854 struct anon_map *amp = svntrp->tr_amp[i];
9855 if (amp == NULL) {
9856 continue;
9857 }
9858 ASSERT(amp->refcnt == 1);
9859 ASSERT(amp->swresv == size);
9860 ASSERT(amp->size == size);
9861 ASSERT(amp->a_szc == szc);
9862 if (amp->a_szc != 0) {
9863 anon_free_pages(amp->ahp, 0, size, szc);
9864 } else {
9865 anon_free(amp->ahp, 0, size);
9866 }
9867 svntrp->tr_amp[i] = NULL;
9868 ASSERT(segvn_textrepl_bytes >= size);
9869 atomic_add_long(&segvn_textrepl_bytes, -size);
9870 anon_unresv_zone(amp->swresv, NULL);
9871 amp->refcnt = 0;
9872 anonmap_free(amp);
9873 }
9874 kmem_cache_free(svntr_cache, svntrp);
9875 done:
9876 hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9877 unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9878 }
9879
9880 /*
9881 * This is called when a MAP_SHARED writable mapping is created to a vnode
9882 * that is currently used for execution (VVMEXEC flag is set). In this case we
9883 * need to prevent further use of existing replicas.
9884 */
9885 static void
9886 segvn_inval_trcache(vnode_t *vp)
9887 {
9888 ulong_t hash = SVNTR_HASH_FUNC(vp);
9889 svntr_t *svntrp;
9890
9891 ASSERT(vp->v_flag & VVMEXEC);
9892
9893 if (svntr_hashtab == NULL) {
9894 return;
9895 }
9896
9897 mutex_enter(&svntr_hashtab[hash].tr_lock);
9898 svntrp = svntr_hashtab[hash].tr_head;
9899 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9900 ASSERT(svntrp->tr_refcnt != 0);
9901 if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9902 svntrp->tr_valid = 0;
9903 }
9904 }
9905 mutex_exit(&svntr_hashtab[hash].tr_lock);
9906 }
9907
9908 static void
9909 segvn_trasync_thread(void)
9910 {
9911 callb_cpr_t cpr_info;
9912 kmutex_t cpr_lock; /* just for CPR stuff */
9913
9914 mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9915
9916 CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9917 callb_generic_cpr, "segvn_async");
9918
9919 if (segvn_update_textrepl_interval == 0) {
9920 segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9921 } else {
9922 segvn_update_textrepl_interval *= hz;
9923 }
9924 (void) timeout(segvn_trupdate_wakeup, NULL,
9925 segvn_update_textrepl_interval);
9926
9927 for (;;) {
9928 mutex_enter(&cpr_lock);
9929 CALLB_CPR_SAFE_BEGIN(&cpr_info);
9930 mutex_exit(&cpr_lock);
9931 sema_p(&segvn_trasync_sem);
9932 mutex_enter(&cpr_lock);
9933 CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9934 mutex_exit(&cpr_lock);
9935 segvn_trupdate();
9936 }
9937 }
9938
9939 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9940
9941 static void
9942 segvn_trupdate_wakeup(void *dummy)
9943 {
9944 uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9945
9946 if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9947 segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9948 sema_v(&segvn_trasync_sem);
9949 }
9950
9951 if (!segvn_disable_textrepl_update &&
9952 segvn_update_textrepl_interval != 0) {
9953 (void) timeout(segvn_trupdate_wakeup, dummy,
9954 segvn_update_textrepl_interval);
9955 }
9956 }
9957
9958 static void
9959 segvn_trupdate(void)
9960 {
9961 ulong_t hash;
9962 svntr_t *svntrp;
9963 segvn_data_t *svd;
9964
9965 ASSERT(svntr_hashtab != NULL);
9966
9967 for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9968 mutex_enter(&svntr_hashtab[hash].tr_lock);
9969 svntrp = svntr_hashtab[hash].tr_head;
9970 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9971 ASSERT(svntrp->tr_refcnt != 0);
9972 svd = svntrp->tr_svnhead;
9973 for (; svd != NULL; svd = svd->svn_trnext) {
9974 segvn_trupdate_seg(svd->seg, svd, svntrp,
9975 hash);
9976 }
9977 }
9978 mutex_exit(&svntr_hashtab[hash].tr_lock);
9979 }
9980 }
9981
9982 static void
9983 segvn_trupdate_seg(struct seg *seg,
9984 segvn_data_t *svd,
9985 svntr_t *svntrp,
9986 ulong_t hash)
9987 {
9988 proc_t *p;
9989 lgrp_id_t lgrp_id;
9990 struct as *as;
9991 size_t size;
9992 struct anon_map *amp;
9993
9994 ASSERT(svd->vp != NULL);
9995 ASSERT(svd->vp == svntrp->tr_vp);
9996 ASSERT(svd->offset == svntrp->tr_off);
9997 ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9998 ASSERT(seg != NULL);
9999 ASSERT(svd->seg == seg);
10000 ASSERT(seg->s_data == (void *)svd);
10001 ASSERT(seg->s_szc == svntrp->tr_szc);
10002 ASSERT(svd->tr_state == SEGVN_TR_ON);
10003 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
10004 ASSERT(svd->amp != NULL);
10005 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10006 ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
10007 ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
10008 ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
10009 ASSERT(svntrp->tr_refcnt != 0);
10010 ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
10011
10012 as = seg->s_as;
10013 ASSERT(as != NULL && as != &kas);
10014 p = as->a_proc;
10015 ASSERT(p != NULL);
10016 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
10017 lgrp_id = p->p_t1_lgrpid;
10018 if (lgrp_id == LGRP_NONE) {
10019 return;
10020 }
10021 ASSERT(lgrp_id < NLGRPS_MAX);
10022 if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
10023 return;
10024 }
10025
10026 /*
10027 * Use tryenter locking since we are locking as/seg and svntr hash
10028 * lock in reverse from syncrounous thread order.
10029 */
10030 if (!AS_LOCK_TRYENTER(as, RW_READER)) {
10031 SEGVN_TR_ADDSTAT(nolock);
10032 if (segvn_lgrp_trthr_migrs_snpsht) {
10033 segvn_lgrp_trthr_migrs_snpsht = 0;
10034 }
10035 return;
10036 }
10037 if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10038 AS_LOCK_EXIT(as);
10039 SEGVN_TR_ADDSTAT(nolock);
10040 if (segvn_lgrp_trthr_migrs_snpsht) {
10041 segvn_lgrp_trthr_migrs_snpsht = 0;
10042 }
10043 return;
10044 }
10045 size = seg->s_size;
10046 if (svntrp->tr_amp[lgrp_id] == NULL) {
10047 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10048 if (trmem > segvn_textrepl_max_bytes) {
10049 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10050 AS_LOCK_EXIT(as);
10051 atomic_add_long(&segvn_textrepl_bytes, -size);
10052 SEGVN_TR_ADDSTAT(normem);
10053 return;
10054 }
10055 if (anon_try_resv_zone(size, NULL) == 0) {
10056 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10057 AS_LOCK_EXIT(as);
10058 atomic_add_long(&segvn_textrepl_bytes, -size);
10059 SEGVN_TR_ADDSTAT(noanon);
10060 return;
10061 }
10062 amp = anonmap_alloc(size, size, KM_NOSLEEP);
10063 if (amp == NULL) {
10064 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10065 AS_LOCK_EXIT(as);
10066 atomic_add_long(&segvn_textrepl_bytes, -size);
10067 anon_unresv_zone(size, NULL);
10068 SEGVN_TR_ADDSTAT(nokmem);
10069 return;
10070 }
10071 ASSERT(amp->refcnt == 1);
10072 amp->a_szc = seg->s_szc;
10073 svntrp->tr_amp[lgrp_id] = amp;
10074 }
10075 /*
10076 * We don't need to drop the bucket lock but here we give other
10077 * threads a chance. svntr and svd can't be unlinked as long as
10078 * segment lock is held as a writer and AS held as well. After we
10079 * retake bucket lock we'll continue from where we left. We'll be able
10080 * to reach the end of either list since new entries are always added
10081 * to the beginning of the lists.
10082 */
10083 mutex_exit(&svntr_hashtab[hash].tr_lock);
10084 hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10085 mutex_enter(&svntr_hashtab[hash].tr_lock);
10086
10087 ASSERT(svd->tr_state == SEGVN_TR_ON);
10088 ASSERT(svd->amp != NULL);
10089 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10090 ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10091 ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10092
10093 svd->tr_policy_info.mem_lgrpid = lgrp_id;
10094 svd->amp = svntrp->tr_amp[lgrp_id];
10095 p->p_tr_lgrpid = NLGRPS_MAX;
10096 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10097 AS_LOCK_EXIT(as);
10098
10099 ASSERT(svntrp->tr_refcnt != 0);
10100 ASSERT(svd->vp == svntrp->tr_vp);
10101 ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10102 ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10103 ASSERT(svd->seg == seg);
10104 ASSERT(svd->tr_state == SEGVN_TR_ON);
10105
10106 SEGVN_TR_ADDSTAT(asyncrepl);
10107 }