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, Version 1.0 only
6 * (the "License"). You may not use this file except in compliance
7 * with the License.
8 *
9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 * or http://www.opensolaris.org/os/licensing.
11 * See the License for the specific language governing permissions
12 * and limitations under the License.
13 *
14 * When distributing Covered Code, include this CDDL HEADER in each
15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 * If applicable, add the following below this CDDL HEADER, with the
17 * fields enclosed by brackets "[]" replaced with your own identifying
18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 *
20 * CDDL HEADER END
21 */
22 /*
23 * Copyright 2006 Sun Microsystems, Inc. All rights reserved.
24 * Use is subject to license terms.
25 */
26
27 /*
28 * Kernel Physical Mapping (kpm) segment driver (segkpm).
29 *
30 * This driver delivers along with the hat_kpm* interfaces an alternative
31 * mechanism for kernel mappings within the 64-bit Solaris operating system,
32 * which allows the mapping of all physical memory into the kernel address
33 * space at once. This is feasible in 64 bit kernels, e.g. for Ultrasparc II
34 * and beyond processors, since the available VA range is much larger than
35 * possible physical memory. Momentarily all physical memory is supported,
36 * that is represented by the list of memory segments (memsegs).
37 *
38 * Segkpm mappings have also very low overhead and large pages are used
39 * (when possible) to minimize the TLB and TSB footprint. It is also
40 * extentable for other than Sparc architectures (e.g. AMD64). Main
41 * advantage is the avoidance of the TLB-shootdown X-calls, which are
42 * normally needed when a kernel (global) mapping has to be removed.
43 *
44 * First example of a kernel facility that uses the segkpm mapping scheme
45 * is seg_map, where it is used as an alternative to hat_memload().
46 * See also hat layer for more information about the hat_kpm* routines.
47 * The kpm facilty can be turned off at boot time (e.g. /etc/system).
48 */
49
50 #include <sys/types.h>
51 #include <sys/param.h>
52 #include <sys/sysmacros.h>
53 #include <sys/systm.h>
54 #include <sys/vnode.h>
55 #include <sys/cmn_err.h>
56 #include <sys/debug.h>
57 #include <sys/thread.h>
58 #include <sys/cpuvar.h>
59 #include <sys/bitmap.h>
60 #include <sys/atomic.h>
61 #include <sys/lgrp.h>
62
63 #include <vm/seg_kmem.h>
64 #include <vm/seg_kpm.h>
65 #include <vm/hat.h>
66 #include <vm/as.h>
67 #include <vm/seg.h>
68 #include <vm/page.h>
69
70 /*
71 * Global kpm controls.
72 * See also platform and mmu specific controls.
73 *
74 * kpm_enable -- global on/off switch for segkpm.
75 * . Set by default on 64bit platforms that have kpm support.
76 * . Will be disabled from platform layer if not supported.
77 * . Can be disabled via /etc/system.
78 *
79 * kpm_smallpages -- use only regular/system pagesize for kpm mappings.
80 * . Can be useful for critical debugging of kpm clients.
81 * . Set to zero by default for platforms that support kpm large pages.
82 * The use of kpm large pages reduces the footprint of kpm meta data
83 * and has all the other advantages of using large pages (e.g TLB
84 * miss reduction).
85 * . Set by default for platforms that don't support kpm large pages or
86 * where large pages cannot be used for other reasons (e.g. there are
87 * only few full associative TLB entries available for large pages).
88 *
89 * segmap_kpm -- separate on/off switch for segmap using segkpm:
90 * . Set by default.
91 * . Will be disabled when kpm_enable is zero.
92 * . Will be disabled when MAXBSIZE != PAGESIZE.
93 * . Can be disabled via /etc/system.
94 *
95 */
96 int kpm_enable = 1;
97 int kpm_smallpages = 0;
98 int segmap_kpm = 1;
99
100 /*
101 * Private seg op routines.
102 */
103 faultcode_t segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr,
104 size_t len, enum fault_type type, enum seg_rw rw);
105 static void segkpm_dump(struct seg *);
106 static void segkpm_badop(void);
107 static int segkpm_notsup(void);
108 static int segkpm_capable(struct seg *, segcapability_t);
109
110 #define SEGKPM_BADOP(t) (t(*)())segkpm_badop
111 #define SEGKPM_NOTSUP (int(*)())segkpm_notsup
112
113 static struct seg_ops segkpm_ops = {
114 SEGKPM_BADOP(int), /* dup */
115 SEGKPM_BADOP(int), /* unmap */
116 SEGKPM_BADOP(void), /* free */
117 segkpm_fault,
118 SEGKPM_BADOP(int), /* faulta */
119 SEGKPM_BADOP(int), /* setprot */
120 SEGKPM_BADOP(int), /* checkprot */
121 SEGKPM_BADOP(int), /* kluster */
122 SEGKPM_BADOP(int), /* sync */
123 SEGKPM_BADOP(size_t), /* incore */
124 SEGKPM_BADOP(int), /* lockop */
125 SEGKPM_BADOP(int), /* getprot */
126 SEGKPM_BADOP(u_offset_t), /* getoffset */
127 SEGKPM_BADOP(int), /* gettype */
128 SEGKPM_BADOP(int), /* getvp */
129 SEGKPM_BADOP(int), /* advise */
130 segkpm_dump, /* dump */
131 SEGKPM_NOTSUP, /* pagelock */
132 SEGKPM_BADOP(int), /* setpgsz */
133 SEGKPM_BADOP(int), /* getmemid */
134 SEGKPM_BADOP(lgrp_mem_policy_info_t *), /* getpolicy */
135 segkpm_capable, /* capable */
136 };
137
138 /*
139 * kpm_pgsz and kpm_pgshft are set by platform layer.
140 */
141 size_t kpm_pgsz; /* kpm page size */
142 uint_t kpm_pgshft; /* kpm page shift */
143 u_offset_t kpm_pgoff; /* kpm page offset mask */
144 uint_t kpmp2pshft; /* kpm page to page shift */
145 pgcnt_t kpmpnpgs; /* how many pages per kpm page */
146
147
148 #ifdef SEGKPM_SUPPORT
149
150 int
151 segkpm_create(struct seg *seg, void *argsp)
152 {
153 struct segkpm_data *skd;
154 struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
155 ushort_t *p;
156 int i, j;
157
158 ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
159 ASSERT(btokpmp(seg->s_size) >= 1 &&
160 kpmpageoff((uintptr_t)seg->s_base) == 0 &&
161 kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
162
163 skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
164
165 seg->s_data = (void *)skd;
166 seg->s_ops = &segkpm_ops;
167 skd->skd_prot = b->prot;
168
169 /*
170 * (1) Segkpm virtual addresses are based on physical adresses.
171 * From this and in opposite to other segment drivers it is
172 * often required to allocate a page first to be able to
173 * calculate the final segkpm virtual address.
174 * (2) Page allocation is done by calling page_create_va(),
175 * one important input argument is a virtual address (also
176 * expressed by the "va" in the function name). This function
177 * is highly optimized to select the right page for an optimal
178 * processor and platform support (e.g. virtual addressed
179 * caches (VAC), physical addressed caches, NUMA).
180 *
181 * Because of (1) the approach is to generate a faked virtual
182 * address for calling page_create_va(). In order to exploit
183 * the abilities of (2), especially to utilize the cache
184 * hierarchy (3) and to avoid VAC alias conflicts (4) the
185 * selection has to be done carefully. For each virtual color
186 * a separate counter is provided (4). The count values are
187 * used for the utilization of all cache lines (3) and are
188 * corresponding to the cache bins.
189 */
190 skd->skd_nvcolors = b->nvcolors;
191
192 p = skd->skd_va_select =
193 kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
194
195 for (i = 0; i < NCPU; i++)
196 for (j = 0; j < b->nvcolors; j++, p++)
197 *p = j;
198
199 return (0);
200 }
201
202 /*
203 * This routine is called via a machine specific fault handling
204 * routine.
205 */
206 /* ARGSUSED */
207 faultcode_t
208 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
209 enum fault_type type, enum seg_rw rw)
210 {
211 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
212
213 switch (type) {
214 case F_INVAL:
215 return (hat_kpm_fault(hat, addr));
216 case F_SOFTLOCK:
217 case F_SOFTUNLOCK:
218 return (0);
219 default:
220 return (FC_NOSUPPORT);
221 }
222 /*NOTREACHED*/
223 }
224
225 #define addr_to_vcolor(addr, vcolors) \
226 ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
227
228 /*
229 * Create a virtual address that can be used for invocations of
230 * page_create_va. Goal is to utilize the cache hierarchy (round
231 * robin bins) and to select the right color for virtual indexed
232 * caches. It isn't exact since we also increment the bin counter
233 * when the caller uses VOP_GETPAGE and gets a hit in the page
234 * cache, but we keep the bins turning for cache distribution
235 * (see also segkpm_create block comment).
236 */
237 caddr_t
238 segkpm_create_va(u_offset_t off)
239 {
240 int vcolor;
241 ushort_t *p;
242 struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
243 int nvcolors = skd->skd_nvcolors;
244 caddr_t va;
245
246 vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
247 p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
248 va = (caddr_t)ptob(*p);
249
250 atomic_add_16(p, nvcolors);
251
252 return (va);
253 }
254
255 /*
256 * Unload mapping if the instance has an active kpm mapping.
257 */
258 void
259 segkpm_mapout_validkpme(struct kpme *kpme)
260 {
261 caddr_t vaddr;
262 page_t *pp;
263
264 retry:
265 if ((pp = kpme->kpe_page) == NULL) {
266 return;
267 }
268
269 if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
270 goto retry;
271
272 /*
273 * Check if segkpm mapping is not unloaded in the meantime
274 */
275 if (kpme->kpe_page == NULL) {
276 page_unlock(pp);
277 return;
278 }
279
280 vaddr = hat_kpm_page2va(pp, 1);
281 hat_kpm_mapout(pp, kpme, vaddr);
282 page_unlock(pp);
283 }
284
285 static void
286 segkpm_badop()
287 {
288 panic("segkpm_badop");
289 }
290
291 #else /* SEGKPM_SUPPORT */
292
293 /* segkpm stubs */
294
295 /*ARGSUSED*/
296 int segkpm_create(struct seg *seg, void *argsp) { return (0); }
297
298 /* ARGSUSED */
299 faultcode_t
300 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
301 enum fault_type type, enum seg_rw rw)
302 {
303 return ((faultcode_t)0);
304 }
305
306 /* ARGSUSED */
307 caddr_t segkpm_create_va(u_offset_t off) { return (NULL); }
308
309 /* ARGSUSED */
310 void segkpm_mapout_validkpme(struct kpme *kpme) {}
311
312 static void
313 segkpm_badop() {}
314
315 #endif /* SEGKPM_SUPPORT */
316
317 static int
318 segkpm_notsup()
319 {
320 return (ENOTSUP);
321 }
322
323 /*
324 * segkpm pages are not dumped, so we just return
325 */
326 /*ARGSUSED*/
327 static void
328 segkpm_dump(struct seg *seg)
329 {}
330
331 /*
332 * We claim to have no special capabilities.
333 */
334 /*ARGSUSED*/
335 static int
336 segkpm_capable(struct seg *seg, segcapability_t capability)
337 {
338 return (0);
339 }