Print this page
6146 seg_inherit_notsup is redundant
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/vm/seg_kpm.c
+++ new/usr/src/uts/common/vm/seg_kpm.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License, Version 1.0 only
6 6 * (the "License"). You may not use this file except in compliance
7 7 * with the License.
8 8 *
9 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 10 * or http://www.opensolaris.org/os/licensing.
11 11 * See the License for the specific language governing permissions
12 12 * and limitations under the License.
13 13 *
14 14 * When distributing Covered Code, include this CDDL HEADER in each
15 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 16 * If applicable, add the following below this CDDL HEADER, with the
17 17 * fields enclosed by brackets "[]" replaced with your own identifying
18 18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 19 *
20 20 * CDDL HEADER END
21 21 */
22 22 /*
23 23 * Copyright 2006 Sun Microsystems, Inc. All rights reserved.
24 24 * Use is subject to license terms.
25 25 */
26 26
27 27 /*
28 28 * Kernel Physical Mapping (kpm) segment driver (segkpm).
29 29 *
30 30 * This driver delivers along with the hat_kpm* interfaces an alternative
31 31 * mechanism for kernel mappings within the 64-bit Solaris operating system,
32 32 * which allows the mapping of all physical memory into the kernel address
33 33 * space at once. This is feasible in 64 bit kernels, e.g. for Ultrasparc II
34 34 * and beyond processors, since the available VA range is much larger than
35 35 * possible physical memory. Momentarily all physical memory is supported,
36 36 * that is represented by the list of memory segments (memsegs).
37 37 *
38 38 * Segkpm mappings have also very low overhead and large pages are used
39 39 * (when possible) to minimize the TLB and TSB footprint. It is also
40 40 * extentable for other than Sparc architectures (e.g. AMD64). Main
41 41 * advantage is the avoidance of the TLB-shootdown X-calls, which are
42 42 * normally needed when a kernel (global) mapping has to be removed.
43 43 *
44 44 * First example of a kernel facility that uses the segkpm mapping scheme
45 45 * is seg_map, where it is used as an alternative to hat_memload().
46 46 * See also hat layer for more information about the hat_kpm* routines.
47 47 * The kpm facilty can be turned off at boot time (e.g. /etc/system).
48 48 */
49 49
50 50 #include <sys/types.h>
51 51 #include <sys/param.h>
52 52 #include <sys/sysmacros.h>
53 53 #include <sys/systm.h>
54 54 #include <sys/vnode.h>
55 55 #include <sys/cmn_err.h>
56 56 #include <sys/debug.h>
57 57 #include <sys/thread.h>
58 58 #include <sys/cpuvar.h>
59 59 #include <sys/bitmap.h>
60 60 #include <sys/atomic.h>
61 61 #include <sys/lgrp.h>
62 62
63 63 #include <vm/seg_kmem.h>
64 64 #include <vm/seg_kpm.h>
65 65 #include <vm/hat.h>
66 66 #include <vm/as.h>
67 67 #include <vm/seg.h>
68 68 #include <vm/page.h>
69 69
70 70 /*
71 71 * Global kpm controls.
72 72 * See also platform and mmu specific controls.
73 73 *
74 74 * kpm_enable -- global on/off switch for segkpm.
75 75 * . Set by default on 64bit platforms that have kpm support.
76 76 * . Will be disabled from platform layer if not supported.
77 77 * . Can be disabled via /etc/system.
78 78 *
79 79 * kpm_smallpages -- use only regular/system pagesize for kpm mappings.
80 80 * . Can be useful for critical debugging of kpm clients.
81 81 * . Set to zero by default for platforms that support kpm large pages.
82 82 * The use of kpm large pages reduces the footprint of kpm meta data
83 83 * and has all the other advantages of using large pages (e.g TLB
84 84 * miss reduction).
85 85 * . Set by default for platforms that don't support kpm large pages or
86 86 * where large pages cannot be used for other reasons (e.g. there are
87 87 * only few full associative TLB entries available for large pages).
88 88 *
89 89 * segmap_kpm -- separate on/off switch for segmap using segkpm:
90 90 * . Set by default.
91 91 * . Will be disabled when kpm_enable is zero.
92 92 * . Will be disabled when MAXBSIZE != PAGESIZE.
93 93 * . Can be disabled via /etc/system.
94 94 *
95 95 */
96 96 int kpm_enable = 1;
97 97 int kpm_smallpages = 0;
98 98 int segmap_kpm = 1;
99 99
100 100 /*
101 101 * Private seg op routines.
102 102 */
103 103 faultcode_t segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr,
104 104 size_t len, enum fault_type type, enum seg_rw rw);
105 105 static void segkpm_dump(struct seg *);
106 106 static void segkpm_badop(void);
107 107 static int segkpm_notsup(void);
108 108 static int segkpm_capable(struct seg *, segcapability_t);
109 109
110 110 #define SEGKPM_BADOP(t) (t(*)())segkpm_badop
111 111 #define SEGKPM_NOTSUP (int(*)())segkpm_notsup
112 112
113 113 static struct seg_ops segkpm_ops = {
114 114 .dup = SEGKPM_BADOP(int),
115 115 .unmap = SEGKPM_BADOP(int),
116 116 .free = SEGKPM_BADOP(void),
117 117 .fault = segkpm_fault,
118 118 .faulta = SEGKPM_BADOP(int),
119 119 .setprot = SEGKPM_BADOP(int),
120 120 .checkprot = SEGKPM_BADOP(int),
121 121 .kluster = SEGKPM_BADOP(int),
122 122 .swapout = SEGKPM_BADOP(size_t),
123 123 .sync = SEGKPM_BADOP(int),
124 124 .incore = SEGKPM_BADOP(size_t),
125 125 .lockop = SEGKPM_BADOP(int),
126 126 .getprot = SEGKPM_BADOP(int),
|
↓ open down ↓ |
126 lines elided |
↑ open up ↑ |
127 127 .getoffset = SEGKPM_BADOP(u_offset_t),
128 128 .gettype = SEGKPM_BADOP(int),
129 129 .getvp = SEGKPM_BADOP(int),
130 130 .advise = SEGKPM_BADOP(int),
131 131 .dump = segkpm_dump,
132 132 .pagelock = SEGKPM_NOTSUP,
133 133 .setpagesize = SEGKPM_BADOP(int),
134 134 .getmemid = SEGKPM_BADOP(int),
135 135 .getpolicy = SEGKPM_BADOP(lgrp_mem_policy_info_t *),
136 136 .capable = segkpm_capable,
137 - .inherit = seg_inherit_notsup,
138 137 };
139 138
140 139 /*
141 140 * kpm_pgsz and kpm_pgshft are set by platform layer.
142 141 */
143 142 size_t kpm_pgsz; /* kpm page size */
144 143 uint_t kpm_pgshft; /* kpm page shift */
145 144 u_offset_t kpm_pgoff; /* kpm page offset mask */
146 145 uint_t kpmp2pshft; /* kpm page to page shift */
147 146 pgcnt_t kpmpnpgs; /* how many pages per kpm page */
148 147
149 148
150 149 #ifdef SEGKPM_SUPPORT
151 150
152 151 int
153 152 segkpm_create(struct seg *seg, void *argsp)
154 153 {
155 154 struct segkpm_data *skd;
156 155 struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
157 156 ushort_t *p;
158 157 int i, j;
159 158
160 159 ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
161 160 ASSERT(btokpmp(seg->s_size) >= 1 &&
162 161 kpmpageoff((uintptr_t)seg->s_base) == 0 &&
163 162 kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
164 163
165 164 skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
166 165
167 166 seg->s_data = (void *)skd;
168 167 seg->s_ops = &segkpm_ops;
169 168 skd->skd_prot = b->prot;
170 169
171 170 /*
172 171 * (1) Segkpm virtual addresses are based on physical adresses.
173 172 * From this and in opposite to other segment drivers it is
174 173 * often required to allocate a page first to be able to
175 174 * calculate the final segkpm virtual address.
176 175 * (2) Page allocation is done by calling page_create_va(),
177 176 * one important input argument is a virtual address (also
178 177 * expressed by the "va" in the function name). This function
179 178 * is highly optimized to select the right page for an optimal
180 179 * processor and platform support (e.g. virtual addressed
181 180 * caches (VAC), physical addressed caches, NUMA).
182 181 *
183 182 * Because of (1) the approach is to generate a faked virtual
184 183 * address for calling page_create_va(). In order to exploit
185 184 * the abilities of (2), especially to utilize the cache
186 185 * hierarchy (3) and to avoid VAC alias conflicts (4) the
187 186 * selection has to be done carefully. For each virtual color
188 187 * a separate counter is provided (4). The count values are
189 188 * used for the utilization of all cache lines (3) and are
190 189 * corresponding to the cache bins.
191 190 */
192 191 skd->skd_nvcolors = b->nvcolors;
193 192
194 193 p = skd->skd_va_select =
195 194 kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
196 195
197 196 for (i = 0; i < NCPU; i++)
198 197 for (j = 0; j < b->nvcolors; j++, p++)
199 198 *p = j;
200 199
201 200 return (0);
202 201 }
203 202
204 203 /*
205 204 * This routine is called via a machine specific fault handling
206 205 * routine.
207 206 */
208 207 /* ARGSUSED */
209 208 faultcode_t
210 209 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
211 210 enum fault_type type, enum seg_rw rw)
212 211 {
213 212 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
214 213
215 214 switch (type) {
216 215 case F_INVAL:
217 216 return (hat_kpm_fault(hat, addr));
218 217 case F_SOFTLOCK:
219 218 case F_SOFTUNLOCK:
220 219 return (0);
221 220 default:
222 221 return (FC_NOSUPPORT);
223 222 }
224 223 /*NOTREACHED*/
225 224 }
226 225
227 226 #define addr_to_vcolor(addr, vcolors) \
228 227 ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
229 228
230 229 /*
231 230 * Create a virtual address that can be used for invocations of
232 231 * page_create_va. Goal is to utilize the cache hierarchy (round
233 232 * robin bins) and to select the right color for virtual indexed
234 233 * caches. It isn't exact since we also increment the bin counter
235 234 * when the caller uses VOP_GETPAGE and gets a hit in the page
236 235 * cache, but we keep the bins turning for cache distribution
237 236 * (see also segkpm_create block comment).
238 237 */
239 238 caddr_t
240 239 segkpm_create_va(u_offset_t off)
241 240 {
242 241 int vcolor;
243 242 ushort_t *p;
244 243 struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
245 244 int nvcolors = skd->skd_nvcolors;
246 245 caddr_t va;
247 246
248 247 vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
249 248 p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
250 249 va = (caddr_t)ptob(*p);
251 250
252 251 atomic_add_16(p, nvcolors);
253 252
254 253 return (va);
255 254 }
256 255
257 256 /*
258 257 * Unload mapping if the instance has an active kpm mapping.
259 258 */
260 259 void
261 260 segkpm_mapout_validkpme(struct kpme *kpme)
262 261 {
263 262 caddr_t vaddr;
264 263 page_t *pp;
265 264
266 265 retry:
267 266 if ((pp = kpme->kpe_page) == NULL) {
268 267 return;
269 268 }
270 269
271 270 if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
272 271 goto retry;
273 272
274 273 /*
275 274 * Check if segkpm mapping is not unloaded in the meantime
276 275 */
277 276 if (kpme->kpe_page == NULL) {
278 277 page_unlock(pp);
279 278 return;
280 279 }
281 280
282 281 vaddr = hat_kpm_page2va(pp, 1);
283 282 hat_kpm_mapout(pp, kpme, vaddr);
284 283 page_unlock(pp);
285 284 }
286 285
287 286 static void
288 287 segkpm_badop()
289 288 {
290 289 panic("segkpm_badop");
291 290 }
292 291
293 292 #else /* SEGKPM_SUPPORT */
294 293
295 294 /* segkpm stubs */
296 295
297 296 /*ARGSUSED*/
298 297 int segkpm_create(struct seg *seg, void *argsp) { return (0); }
299 298
300 299 /* ARGSUSED */
301 300 faultcode_t
302 301 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
303 302 enum fault_type type, enum seg_rw rw)
304 303 {
305 304 return ((faultcode_t)0);
306 305 }
307 306
308 307 /* ARGSUSED */
309 308 caddr_t segkpm_create_va(u_offset_t off) { return (NULL); }
310 309
311 310 /* ARGSUSED */
312 311 void segkpm_mapout_validkpme(struct kpme *kpme) {}
313 312
314 313 static void
315 314 segkpm_badop() {}
316 315
317 316 #endif /* SEGKPM_SUPPORT */
318 317
319 318 static int
320 319 segkpm_notsup()
321 320 {
322 321 return (ENOTSUP);
323 322 }
324 323
325 324 /*
326 325 * segkpm pages are not dumped, so we just return
327 326 */
328 327 /*ARGSUSED*/
329 328 static void
330 329 segkpm_dump(struct seg *seg)
331 330 {}
332 331
333 332 /*
334 333 * We claim to have no special capabilities.
335 334 */
336 335 /*ARGSUSED*/
337 336 static int
338 337 segkpm_capable(struct seg *seg, segcapability_t capability)
339 338 {
340 339 return (0);
341 340 }
|
↓ open down ↓ |
194 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX