Print this page
const-ify make segment ops structures
There is no reason to keep the segment ops structures writable.
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/vm/seg_kp.c
+++ new/usr/src/uts/common/vm/seg_kp.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 (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 */
24 24
25 25 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
26 26 /* All Rights Reserved */
27 27
28 28 /*
29 29 * Portions of this source code were derived from Berkeley 4.3 BSD
30 30 * under license from the Regents of the University of California.
31 31 */
32 32
33 33 /*
34 34 * segkp is a segment driver that administers the allocation and deallocation
35 35 * of pageable variable size chunks of kernel virtual address space. Each
36 36 * allocated resource is page-aligned.
37 37 *
38 38 * The user may specify whether the resource should be initialized to 0,
39 39 * include a redzone, or locked in memory.
40 40 */
41 41
42 42 #include <sys/types.h>
43 43 #include <sys/t_lock.h>
44 44 #include <sys/thread.h>
45 45 #include <sys/param.h>
46 46 #include <sys/errno.h>
47 47 #include <sys/sysmacros.h>
48 48 #include <sys/systm.h>
49 49 #include <sys/buf.h>
50 50 #include <sys/mman.h>
51 51 #include <sys/vnode.h>
52 52 #include <sys/cmn_err.h>
53 53 #include <sys/swap.h>
54 54 #include <sys/tuneable.h>
55 55 #include <sys/kmem.h>
56 56 #include <sys/vmem.h>
57 57 #include <sys/cred.h>
58 58 #include <sys/dumphdr.h>
59 59 #include <sys/debug.h>
60 60 #include <sys/vtrace.h>
61 61 #include <sys/stack.h>
62 62 #include <sys/atomic.h>
63 63 #include <sys/archsystm.h>
64 64 #include <sys/lgrp.h>
65 65
66 66 #include <vm/as.h>
67 67 #include <vm/seg.h>
68 68 #include <vm/seg_kp.h>
69 69 #include <vm/seg_kmem.h>
70 70 #include <vm/anon.h>
71 71 #include <vm/page.h>
72 72 #include <vm/hat.h>
73 73 #include <sys/bitmap.h>
74 74
75 75 /*
76 76 * Private seg op routines
77 77 */
78 78 static void segkp_dump(struct seg *seg);
79 79 static int segkp_checkprot(struct seg *seg, caddr_t addr, size_t len,
80 80 uint_t prot);
81 81 static int segkp_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
82 82 static int segkp_pagelock(struct seg *seg, caddr_t addr, size_t len,
83 83 struct page ***page, enum lock_type type,
84 84 enum seg_rw rw);
85 85 static void segkp_insert(struct seg *seg, struct segkp_data *kpd);
86 86 static void segkp_delete(struct seg *seg, struct segkp_data *kpd);
87 87 static caddr_t segkp_get_internal(struct seg *seg, size_t len, uint_t flags,
88 88 struct segkp_data **tkpd, struct anon_map *amp);
89 89 static void segkp_release_internal(struct seg *seg,
90 90 struct segkp_data *kpd, size_t len);
91 91 static int segkp_unlock(struct hat *hat, struct seg *seg, caddr_t vaddr,
92 92 size_t len, struct segkp_data *kpd, uint_t flags);
93 93 static int segkp_load(struct hat *hat, struct seg *seg, caddr_t vaddr,
94 94 size_t len, struct segkp_data *kpd, uint_t flags);
95 95 static struct segkp_data *segkp_find(struct seg *seg, caddr_t vaddr);
96 96
97 97 /*
98 98 * Lock used to protect the hash table(s) and caches.
99 99 */
100 100 static kmutex_t segkp_lock;
101 101
102 102 /*
103 103 * The segkp caches
104 104 */
105 105 static struct segkp_cache segkp_cache[SEGKP_MAX_CACHE];
106 106
107 107 /*
108 108 * When there are fewer than red_minavail bytes left on the stack,
109 109 * segkp_map_red() will map in the redzone (if called). 5000 seems
110 110 * to work reasonably well...
111 111 */
112 112 long red_minavail = 5000;
113 113
114 114 /*
115 115 * will be set to 1 for 32 bit x86 systems only, in startup.c
116 116 */
117 117 int segkp_fromheap = 0;
118 118 ulong_t *segkp_bitmap;
119 119
120 120 /*
121 121 * If segkp_map_red() is called with the redzone already mapped and
122 122 * with less than RED_DEEP_THRESHOLD bytes available on the stack,
123 123 * then the stack situation has become quite serious; if much more stack
124 124 * is consumed, we have the potential of scrogging the next thread/LWP
125 125 * structure. To help debug the "can't happen" panics which may
126 126 * result from this condition, we record hrestime and the calling thread
127 127 * in red_deep_hires and red_deep_thread respectively.
128 128 */
129 129 #define RED_DEEP_THRESHOLD 2000
130 130
|
↓ open down ↓ |
130 lines elided |
↑ open up ↑ |
131 131 hrtime_t red_deep_hires;
132 132 kthread_t *red_deep_thread;
133 133
134 134 uint32_t red_nmapped;
135 135 uint32_t red_closest = UINT_MAX;
136 136 uint32_t red_ndoubles;
137 137
138 138 pgcnt_t anon_segkp_pages_locked; /* See vm/anon.h */
139 139 pgcnt_t anon_segkp_pages_resv; /* anon reserved by seg_kp */
140 140
141 -static struct seg_ops segkp_ops = {
141 +static const struct seg_ops segkp_ops = {
142 142 .fault = segkp_fault,
143 143 .checkprot = segkp_checkprot,
144 144 .kluster = segkp_kluster,
145 145 .dump = segkp_dump,
146 146 .pagelock = segkp_pagelock,
147 147 };
148 148
149 149
150 150 static void segkpinit_mem_config(struct seg *);
151 151
152 152 static uint32_t segkp_indel;
153 153
154 154 /*
155 155 * Allocate the segment specific private data struct and fill it in
156 156 * with the per kp segment mutex, anon ptr. array and hash table.
157 157 */
158 158 int
159 159 segkp_create(struct seg *seg)
160 160 {
161 161 struct segkp_segdata *kpsd;
162 162 size_t np;
163 163
164 164 ASSERT(seg != NULL && seg->s_as == &kas);
165 165 ASSERT(RW_WRITE_HELD(&seg->s_as->a_lock));
166 166
167 167 if (seg->s_size & PAGEOFFSET) {
168 168 panic("Bad segkp size");
169 169 /*NOTREACHED*/
170 170 }
171 171
172 172 kpsd = kmem_zalloc(sizeof (struct segkp_segdata), KM_SLEEP);
173 173
174 174 /*
175 175 * Allocate the virtual memory for segkp and initialize it
176 176 */
177 177 if (segkp_fromheap) {
178 178 np = btop(kvseg.s_size);
179 179 segkp_bitmap = kmem_zalloc(BT_SIZEOFMAP(np), KM_SLEEP);
180 180 kpsd->kpsd_arena = vmem_create("segkp", NULL, 0, PAGESIZE,
181 181 vmem_alloc, vmem_free, heap_arena, 5 * PAGESIZE, VM_SLEEP);
182 182 } else {
183 183 segkp_bitmap = NULL;
184 184 np = btop(seg->s_size);
185 185 kpsd->kpsd_arena = vmem_create("segkp", seg->s_base,
186 186 seg->s_size, PAGESIZE, NULL, NULL, NULL, 5 * PAGESIZE,
187 187 VM_SLEEP);
188 188 }
189 189
190 190 kpsd->kpsd_anon = anon_create(np, ANON_SLEEP | ANON_ALLOC_FORCE);
191 191
192 192 kpsd->kpsd_hash = kmem_zalloc(SEGKP_HASHSZ * sizeof (struct segkp *),
193 193 KM_SLEEP);
194 194 seg->s_data = (void *)kpsd;
195 195 seg->s_ops = &segkp_ops;
196 196 segkpinit_mem_config(seg);
197 197 return (0);
198 198 }
199 199
200 200
201 201 /*
202 202 * Find a free 'freelist' and initialize it with the appropriate attributes
203 203 */
204 204 void *
205 205 segkp_cache_init(struct seg *seg, int maxsize, size_t len, uint_t flags)
206 206 {
207 207 int i;
208 208
209 209 if ((flags & KPD_NO_ANON) && !(flags & KPD_LOCKED))
210 210 return ((void *)-1);
211 211
212 212 mutex_enter(&segkp_lock);
213 213 for (i = 0; i < SEGKP_MAX_CACHE; i++) {
214 214 if (segkp_cache[i].kpf_inuse)
215 215 continue;
216 216 segkp_cache[i].kpf_inuse = 1;
217 217 segkp_cache[i].kpf_max = maxsize;
218 218 segkp_cache[i].kpf_flags = flags;
219 219 segkp_cache[i].kpf_seg = seg;
220 220 segkp_cache[i].kpf_len = len;
221 221 mutex_exit(&segkp_lock);
222 222 return ((void *)(uintptr_t)i);
223 223 }
224 224 mutex_exit(&segkp_lock);
225 225 return ((void *)-1);
226 226 }
227 227
228 228 /*
229 229 * Free all the cache resources.
230 230 */
231 231 void
232 232 segkp_cache_free(void)
233 233 {
234 234 struct segkp_data *kpd;
235 235 struct seg *seg;
236 236 int i;
237 237
238 238 mutex_enter(&segkp_lock);
239 239 for (i = 0; i < SEGKP_MAX_CACHE; i++) {
240 240 if (!segkp_cache[i].kpf_inuse)
241 241 continue;
242 242 /*
243 243 * Disconnect the freelist and process each element
244 244 */
245 245 kpd = segkp_cache[i].kpf_list;
246 246 seg = segkp_cache[i].kpf_seg;
247 247 segkp_cache[i].kpf_list = NULL;
248 248 segkp_cache[i].kpf_count = 0;
249 249 mutex_exit(&segkp_lock);
250 250
251 251 while (kpd != NULL) {
252 252 struct segkp_data *next;
253 253
254 254 next = kpd->kp_next;
255 255 segkp_release_internal(seg, kpd, kpd->kp_len);
256 256 kpd = next;
257 257 }
258 258 mutex_enter(&segkp_lock);
259 259 }
260 260 mutex_exit(&segkp_lock);
261 261 }
262 262
263 263 /*
264 264 * There are 2 entries into segkp_get_internal. The first includes a cookie
265 265 * used to access a pool of cached segkp resources. The second does not
266 266 * use the cache.
267 267 */
268 268 caddr_t
269 269 segkp_get(struct seg *seg, size_t len, uint_t flags)
270 270 {
271 271 struct segkp_data *kpd = NULL;
272 272
273 273 if (segkp_get_internal(seg, len, flags, &kpd, NULL) != NULL) {
274 274 kpd->kp_cookie = -1;
275 275 return (stom(kpd->kp_base, flags));
276 276 }
277 277 return (NULL);
278 278 }
279 279
280 280 /*
281 281 * Return a 'cached' segkp address
282 282 */
283 283 caddr_t
284 284 segkp_cache_get(void *cookie)
285 285 {
286 286 struct segkp_cache *freelist = NULL;
287 287 struct segkp_data *kpd = NULL;
288 288 int index = (int)(uintptr_t)cookie;
289 289 struct seg *seg;
290 290 size_t len;
291 291 uint_t flags;
292 292
293 293 if (index < 0 || index >= SEGKP_MAX_CACHE)
294 294 return (NULL);
295 295 freelist = &segkp_cache[index];
296 296
297 297 mutex_enter(&segkp_lock);
298 298 seg = freelist->kpf_seg;
299 299 flags = freelist->kpf_flags;
300 300 if (freelist->kpf_list != NULL) {
301 301 kpd = freelist->kpf_list;
302 302 freelist->kpf_list = kpd->kp_next;
303 303 freelist->kpf_count--;
304 304 mutex_exit(&segkp_lock);
305 305 kpd->kp_next = NULL;
306 306 segkp_insert(seg, kpd);
307 307 return (stom(kpd->kp_base, flags));
308 308 }
309 309 len = freelist->kpf_len;
310 310 mutex_exit(&segkp_lock);
311 311 if (segkp_get_internal(seg, len, flags, &kpd, NULL) != NULL) {
312 312 kpd->kp_cookie = index;
313 313 return (stom(kpd->kp_base, flags));
314 314 }
315 315 return (NULL);
316 316 }
317 317
318 318 caddr_t
319 319 segkp_get_withanonmap(
320 320 struct seg *seg,
321 321 size_t len,
322 322 uint_t flags,
323 323 struct anon_map *amp)
324 324 {
325 325 struct segkp_data *kpd = NULL;
326 326
327 327 ASSERT(amp != NULL);
328 328 flags |= KPD_HASAMP;
329 329 if (segkp_get_internal(seg, len, flags, &kpd, amp) != NULL) {
330 330 kpd->kp_cookie = -1;
331 331 return (stom(kpd->kp_base, flags));
332 332 }
333 333 return (NULL);
334 334 }
335 335
336 336 /*
337 337 * This does the real work of segkp allocation.
338 338 * Return to client base addr. len must be page-aligned. A null value is
339 339 * returned if there are no more vm resources (e.g. pages, swap). The len
340 340 * and base recorded in the private data structure include the redzone
341 341 * and the redzone length (if applicable). If the user requests a redzone
342 342 * either the first or last page is left unmapped depending whether stacks
343 343 * grow to low or high memory.
344 344 *
345 345 * The client may also specify a no-wait flag. If that is set then the
346 346 * request will choose a non-blocking path when requesting resources.
347 347 * The default is make the client wait.
348 348 */
349 349 static caddr_t
350 350 segkp_get_internal(
351 351 struct seg *seg,
352 352 size_t len,
353 353 uint_t flags,
354 354 struct segkp_data **tkpd,
355 355 struct anon_map *amp)
356 356 {
357 357 struct segkp_segdata *kpsd = (struct segkp_segdata *)seg->s_data;
358 358 struct segkp_data *kpd;
359 359 caddr_t vbase = NULL; /* always first virtual, may not be mapped */
360 360 pgcnt_t np = 0; /* number of pages in the resource */
361 361 pgcnt_t segkpindex;
362 362 long i;
363 363 caddr_t va;
364 364 pgcnt_t pages = 0;
365 365 ulong_t anon_idx = 0;
366 366 int kmflag = (flags & KPD_NOWAIT) ? KM_NOSLEEP : KM_SLEEP;
367 367 caddr_t s_base = (segkp_fromheap) ? kvseg.s_base : seg->s_base;
368 368
369 369 if (len & PAGEOFFSET) {
370 370 panic("segkp_get: len is not page-aligned");
371 371 /*NOTREACHED*/
372 372 }
373 373
374 374 ASSERT(((flags & KPD_HASAMP) == 0) == (amp == NULL));
375 375
376 376 /* Only allow KPD_NO_ANON if we are going to lock it down */
377 377 if ((flags & (KPD_LOCKED|KPD_NO_ANON)) == KPD_NO_ANON)
378 378 return (NULL);
379 379
380 380 if ((kpd = kmem_zalloc(sizeof (struct segkp_data), kmflag)) == NULL)
381 381 return (NULL);
382 382 /*
383 383 * Fix up the len to reflect the REDZONE if applicable
384 384 */
385 385 if (flags & KPD_HASREDZONE)
386 386 len += PAGESIZE;
387 387 np = btop(len);
388 388
389 389 vbase = vmem_alloc(SEGKP_VMEM(seg), len, kmflag | VM_BESTFIT);
390 390 if (vbase == NULL) {
391 391 kmem_free(kpd, sizeof (struct segkp_data));
392 392 return (NULL);
393 393 }
394 394
395 395 /* If locking, reserve physical memory */
396 396 if (flags & KPD_LOCKED) {
397 397 pages = btop(SEGKP_MAPLEN(len, flags));
398 398 if (page_resv(pages, kmflag) == 0) {
399 399 vmem_free(SEGKP_VMEM(seg), vbase, len);
400 400 kmem_free(kpd, sizeof (struct segkp_data));
401 401 return (NULL);
402 402 }
403 403 if ((flags & KPD_NO_ANON) == 0)
404 404 atomic_add_long(&anon_segkp_pages_locked, pages);
405 405 }
406 406
407 407 /*
408 408 * Reserve sufficient swap space for this vm resource. We'll
409 409 * actually allocate it in the loop below, but reserving it
410 410 * here allows us to back out more gracefully than if we
411 411 * had an allocation failure in the body of the loop.
412 412 *
413 413 * Note that we don't need swap space for the red zone page.
414 414 */
415 415 if (amp != NULL) {
416 416 /*
417 417 * The swap reservation has been done, if required, and the
418 418 * anon_hdr is separate.
419 419 */
420 420 anon_idx = 0;
421 421 kpd->kp_anon_idx = anon_idx;
422 422 kpd->kp_anon = amp->ahp;
423 423
424 424 TRACE_5(TR_FAC_VM, TR_ANON_SEGKP, "anon segkp:%p %p %lu %u %u",
425 425 kpd, vbase, len, flags, 1);
426 426
427 427 } else if ((flags & KPD_NO_ANON) == 0) {
428 428 if (anon_resv_zone(SEGKP_MAPLEN(len, flags), NULL) == 0) {
429 429 if (flags & KPD_LOCKED) {
430 430 atomic_add_long(&anon_segkp_pages_locked,
431 431 -pages);
432 432 page_unresv(pages);
433 433 }
434 434 vmem_free(SEGKP_VMEM(seg), vbase, len);
435 435 kmem_free(kpd, sizeof (struct segkp_data));
436 436 return (NULL);
437 437 }
438 438 atomic_add_long(&anon_segkp_pages_resv,
439 439 btop(SEGKP_MAPLEN(len, flags)));
440 440 anon_idx = ((uintptr_t)(vbase - s_base)) >> PAGESHIFT;
441 441 kpd->kp_anon_idx = anon_idx;
442 442 kpd->kp_anon = kpsd->kpsd_anon;
443 443
444 444 TRACE_5(TR_FAC_VM, TR_ANON_SEGKP, "anon segkp:%p %p %lu %u %u",
445 445 kpd, vbase, len, flags, 1);
446 446 } else {
447 447 kpd->kp_anon = NULL;
448 448 kpd->kp_anon_idx = 0;
449 449 }
450 450
451 451 /*
452 452 * Allocate page and anon resources for the virtual address range
453 453 * except the redzone
454 454 */
455 455 if (segkp_fromheap)
456 456 segkpindex = btop((uintptr_t)(vbase - kvseg.s_base));
457 457 for (i = 0, va = vbase; i < np; i++, va += PAGESIZE) {
458 458 page_t *pl[2];
459 459 struct vnode *vp;
460 460 anoff_t off;
461 461 int err;
462 462 page_t *pp = NULL;
463 463
464 464 /*
465 465 * Mark this page to be a segkp page in the bitmap.
466 466 */
467 467 if (segkp_fromheap) {
468 468 BT_ATOMIC_SET(segkp_bitmap, segkpindex);
469 469 segkpindex++;
470 470 }
471 471
472 472 /*
473 473 * If this page is the red zone page, we don't need swap
474 474 * space for it. Note that we skip over the code that
475 475 * establishes MMU mappings, so that the page remains
476 476 * invalid.
477 477 */
478 478 if ((flags & KPD_HASREDZONE) && KPD_REDZONE(kpd) == i)
479 479 continue;
480 480
481 481 if (kpd->kp_anon != NULL) {
482 482 struct anon *ap;
483 483
484 484 ASSERT(anon_get_ptr(kpd->kp_anon, anon_idx + i)
485 485 == NULL);
486 486 /*
487 487 * Determine the "vp" and "off" of the anon slot.
488 488 */
489 489 ap = anon_alloc(NULL, 0);
490 490 if (amp != NULL)
491 491 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
492 492 (void) anon_set_ptr(kpd->kp_anon, anon_idx + i,
493 493 ap, ANON_SLEEP);
494 494 if (amp != NULL)
495 495 ANON_LOCK_EXIT(&->a_rwlock);
496 496 swap_xlate(ap, &vp, &off);
497 497
498 498 /*
499 499 * Create a page with the specified identity. The
500 500 * page is returned with the "shared" lock held.
501 501 */
502 502 err = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
503 503 NULL, pl, PAGESIZE, seg, va, S_CREATE,
504 504 kcred, NULL);
505 505 if (err) {
506 506 /*
507 507 * XXX - This should not fail.
508 508 */
509 509 panic("segkp_get: no pages");
510 510 /*NOTREACHED*/
511 511 }
512 512 pp = pl[0];
513 513 } else {
514 514 ASSERT(page_exists(&kvp,
515 515 (u_offset_t)(uintptr_t)va) == NULL);
516 516
517 517 if ((pp = page_create_va(&kvp,
518 518 (u_offset_t)(uintptr_t)va, PAGESIZE,
519 519 (flags & KPD_NOWAIT ? 0 : PG_WAIT) | PG_EXCL |
520 520 PG_NORELOC, seg, va)) == NULL) {
521 521 /*
522 522 * Legitimize resource; then destroy it.
523 523 * Easier than trying to unwind here.
524 524 */
525 525 kpd->kp_flags = flags;
526 526 kpd->kp_base = vbase;
527 527 kpd->kp_len = len;
528 528 segkp_release_internal(seg, kpd, va - vbase);
529 529 return (NULL);
530 530 }
531 531 page_io_unlock(pp);
532 532 }
533 533
534 534 if (flags & KPD_ZERO)
535 535 pagezero(pp, 0, PAGESIZE);
536 536
537 537 /*
538 538 * Load and lock an MMU translation for the page.
539 539 */
540 540 hat_memload(seg->s_as->a_hat, va, pp, (PROT_READ|PROT_WRITE),
541 541 ((flags & KPD_LOCKED) ? HAT_LOAD_LOCK : HAT_LOAD));
542 542
543 543 /*
544 544 * Now, release lock on the page.
545 545 */
546 546 if (flags & KPD_LOCKED) {
547 547 /*
548 548 * Indicate to page_retire framework that this
549 549 * page can only be retired when it is freed.
550 550 */
551 551 PP_SETRAF(pp);
552 552 page_downgrade(pp);
553 553 } else
554 554 page_unlock(pp);
555 555 }
556 556
557 557 kpd->kp_flags = flags;
558 558 kpd->kp_base = vbase;
559 559 kpd->kp_len = len;
560 560 segkp_insert(seg, kpd);
561 561 *tkpd = kpd;
562 562 return (stom(kpd->kp_base, flags));
563 563 }
564 564
565 565 /*
566 566 * Release the resource to cache if the pool(designate by the cookie)
567 567 * has less than the maximum allowable. If inserted in cache,
568 568 * segkp_delete insures element is taken off of active list.
569 569 */
570 570 void
571 571 segkp_release(struct seg *seg, caddr_t vaddr)
572 572 {
573 573 struct segkp_cache *freelist;
574 574 struct segkp_data *kpd = NULL;
575 575
576 576 if ((kpd = segkp_find(seg, vaddr)) == NULL) {
577 577 panic("segkp_release: null kpd");
578 578 /*NOTREACHED*/
579 579 }
580 580
581 581 if (kpd->kp_cookie != -1) {
582 582 freelist = &segkp_cache[kpd->kp_cookie];
583 583 mutex_enter(&segkp_lock);
584 584 if (!segkp_indel && freelist->kpf_count < freelist->kpf_max) {
585 585 segkp_delete(seg, kpd);
586 586 kpd->kp_next = freelist->kpf_list;
587 587 freelist->kpf_list = kpd;
588 588 freelist->kpf_count++;
589 589 mutex_exit(&segkp_lock);
590 590 return;
591 591 } else {
592 592 mutex_exit(&segkp_lock);
593 593 kpd->kp_cookie = -1;
594 594 }
595 595 }
596 596 segkp_release_internal(seg, kpd, kpd->kp_len);
597 597 }
598 598
599 599 /*
600 600 * Free the entire resource. segkp_unlock gets called with the start of the
601 601 * mapped portion of the resource. The length is the size of the mapped
602 602 * portion
603 603 */
604 604 static void
605 605 segkp_release_internal(struct seg *seg, struct segkp_data *kpd, size_t len)
606 606 {
607 607 caddr_t va;
608 608 long i;
609 609 long redzone;
610 610 size_t np;
611 611 page_t *pp;
612 612 struct vnode *vp;
613 613 anoff_t off;
614 614 struct anon *ap;
615 615 pgcnt_t segkpindex;
616 616
617 617 ASSERT(kpd != NULL);
618 618 ASSERT((kpd->kp_flags & KPD_HASAMP) == 0 || kpd->kp_cookie == -1);
619 619 np = btop(len);
620 620
621 621 /* Remove from active hash list */
622 622 if (kpd->kp_cookie == -1) {
623 623 mutex_enter(&segkp_lock);
624 624 segkp_delete(seg, kpd);
625 625 mutex_exit(&segkp_lock);
626 626 }
627 627
628 628 /*
629 629 * Precompute redzone page index.
630 630 */
631 631 redzone = -1;
632 632 if (kpd->kp_flags & KPD_HASREDZONE)
633 633 redzone = KPD_REDZONE(kpd);
634 634
635 635
636 636 va = kpd->kp_base;
637 637
638 638 hat_unload(seg->s_as->a_hat, va, (np << PAGESHIFT),
639 639 ((kpd->kp_flags & KPD_LOCKED) ? HAT_UNLOAD_UNLOCK : HAT_UNLOAD));
640 640 /*
641 641 * Free up those anon resources that are quiescent.
642 642 */
643 643 if (segkp_fromheap)
644 644 segkpindex = btop((uintptr_t)(va - kvseg.s_base));
645 645 for (i = 0; i < np; i++, va += PAGESIZE) {
646 646
647 647 /*
648 648 * Clear the bit for this page from the bitmap.
649 649 */
650 650 if (segkp_fromheap) {
651 651 BT_ATOMIC_CLEAR(segkp_bitmap, segkpindex);
652 652 segkpindex++;
653 653 }
654 654
655 655 if (i == redzone)
656 656 continue;
657 657 if (kpd->kp_anon) {
658 658 /*
659 659 * Free up anon resources and destroy the
660 660 * associated pages.
661 661 *
662 662 * Release the lock if there is one. Have to get the
663 663 * page to do this, unfortunately.
664 664 */
665 665 if (kpd->kp_flags & KPD_LOCKED) {
666 666 ap = anon_get_ptr(kpd->kp_anon,
667 667 kpd->kp_anon_idx + i);
668 668 swap_xlate(ap, &vp, &off);
669 669 /* Find the shared-locked page. */
670 670 pp = page_find(vp, (u_offset_t)off);
671 671 if (pp == NULL) {
672 672 panic("segkp_release: "
673 673 "kp_anon: no page to unlock ");
674 674 /*NOTREACHED*/
675 675 }
676 676 if (PP_ISRAF(pp))
677 677 PP_CLRRAF(pp);
678 678
679 679 page_unlock(pp);
680 680 }
681 681 if ((kpd->kp_flags & KPD_HASAMP) == 0) {
682 682 anon_free(kpd->kp_anon, kpd->kp_anon_idx + i,
683 683 PAGESIZE);
684 684 anon_unresv_zone(PAGESIZE, NULL);
685 685 atomic_dec_ulong(&anon_segkp_pages_resv);
686 686 }
687 687 TRACE_5(TR_FAC_VM,
688 688 TR_ANON_SEGKP, "anon segkp:%p %p %lu %u %u",
689 689 kpd, va, PAGESIZE, 0, 0);
690 690 } else {
691 691 if (kpd->kp_flags & KPD_LOCKED) {
692 692 pp = page_find(&kvp, (u_offset_t)(uintptr_t)va);
693 693 if (pp == NULL) {
694 694 panic("segkp_release: "
695 695 "no page to unlock");
696 696 /*NOTREACHED*/
697 697 }
698 698 if (PP_ISRAF(pp))
699 699 PP_CLRRAF(pp);
700 700 /*
701 701 * We should just upgrade the lock here
702 702 * but there is no upgrade that waits.
703 703 */
704 704 page_unlock(pp);
705 705 }
706 706 pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)va,
707 707 SE_EXCL);
708 708 if (pp != NULL)
709 709 page_destroy(pp, 0);
710 710 }
711 711 }
712 712
713 713 /* If locked, release physical memory reservation */
714 714 if (kpd->kp_flags & KPD_LOCKED) {
715 715 pgcnt_t pages = btop(SEGKP_MAPLEN(kpd->kp_len, kpd->kp_flags));
716 716 if ((kpd->kp_flags & KPD_NO_ANON) == 0)
717 717 atomic_add_long(&anon_segkp_pages_locked, -pages);
718 718 page_unresv(pages);
719 719 }
720 720
721 721 vmem_free(SEGKP_VMEM(seg), kpd->kp_base, kpd->kp_len);
722 722 kmem_free(kpd, sizeof (struct segkp_data));
723 723 }
724 724
725 725 /*
726 726 * segkp_map_red() will check the current frame pointer against the
727 727 * stack base. If the amount of stack remaining is questionable
728 728 * (less than red_minavail), then segkp_map_red() will map in the redzone
729 729 * and return 1. Otherwise, it will return 0. segkp_map_red() can
730 730 * _only_ be called when it is safe to sleep on page_create_va().
731 731 *
732 732 * It is up to the caller to remember whether segkp_map_red() successfully
733 733 * mapped the redzone, and, if so, to call segkp_unmap_red() at a later
734 734 * time.
735 735 *
736 736 * Currently, this routine is only called from pagefault() (which necessarily
737 737 * satisfies the above conditions).
738 738 */
739 739 #if defined(STACK_GROWTH_DOWN)
740 740 int
741 741 segkp_map_red(void)
742 742 {
743 743 uintptr_t fp = STACK_BIAS + (uintptr_t)getfp();
744 744 #ifndef _LP64
745 745 caddr_t stkbase;
746 746 #endif
747 747
748 748 /*
749 749 * Optimize for the common case where we simply return.
750 750 */
751 751 if ((curthread->t_red_pp == NULL) &&
752 752 (fp - (uintptr_t)curthread->t_stkbase >= red_minavail))
753 753 return (0);
754 754
755 755 #if defined(_LP64)
756 756 /*
757 757 * XXX We probably need something better than this.
758 758 */
759 759 panic("kernel stack overflow");
760 760 /*NOTREACHED*/
761 761 #else /* _LP64 */
762 762 if (curthread->t_red_pp == NULL) {
763 763 page_t *red_pp;
764 764 struct seg kseg;
765 765
766 766 caddr_t red_va = (caddr_t)
767 767 (((uintptr_t)curthread->t_stkbase & (uintptr_t)PAGEMASK) -
768 768 PAGESIZE);
769 769
770 770 ASSERT(page_exists(&kvp, (u_offset_t)(uintptr_t)red_va) ==
771 771 NULL);
772 772
773 773 /*
774 774 * Allocate the physical for the red page.
775 775 */
776 776 /*
777 777 * No PG_NORELOC here to avoid waits. Unlikely to get
778 778 * a relocate happening in the short time the page exists
779 779 * and it will be OK anyway.
780 780 */
781 781
782 782 kseg.s_as = &kas;
783 783 red_pp = page_create_va(&kvp, (u_offset_t)(uintptr_t)red_va,
784 784 PAGESIZE, PG_WAIT | PG_EXCL, &kseg, red_va);
785 785 ASSERT(red_pp != NULL);
786 786
787 787 /*
788 788 * So we now have a page to jam into the redzone...
789 789 */
790 790 page_io_unlock(red_pp);
791 791
792 792 hat_memload(kas.a_hat, red_va, red_pp,
793 793 (PROT_READ|PROT_WRITE), HAT_LOAD_LOCK);
794 794 page_downgrade(red_pp);
795 795
796 796 /*
797 797 * The page is left SE_SHARED locked so we can hold on to
798 798 * the page_t pointer.
799 799 */
800 800 curthread->t_red_pp = red_pp;
801 801
802 802 atomic_inc_32(&red_nmapped);
803 803 while (fp - (uintptr_t)curthread->t_stkbase < red_closest) {
804 804 (void) atomic_cas_32(&red_closest, red_closest,
805 805 (uint32_t)(fp - (uintptr_t)curthread->t_stkbase));
806 806 }
807 807 return (1);
808 808 }
809 809
810 810 stkbase = (caddr_t)(((uintptr_t)curthread->t_stkbase &
811 811 (uintptr_t)PAGEMASK) - PAGESIZE);
812 812
813 813 atomic_inc_32(&red_ndoubles);
814 814
815 815 if (fp - (uintptr_t)stkbase < RED_DEEP_THRESHOLD) {
816 816 /*
817 817 * Oh boy. We're already deep within the mapped-in
818 818 * redzone page, and the caller is trying to prepare
819 819 * for a deep stack run. We're running without a
820 820 * redzone right now: if the caller plows off the
821 821 * end of the stack, it'll plow another thread or
822 822 * LWP structure. That situation could result in
823 823 * a very hard-to-debug panic, so, in the spirit of
824 824 * recording the name of one's killer in one's own
825 825 * blood, we're going to record hrestime and the calling
826 826 * thread.
827 827 */
828 828 red_deep_hires = hrestime.tv_nsec;
829 829 red_deep_thread = curthread;
830 830 }
831 831
832 832 /*
833 833 * If this is a DEBUG kernel, and we've run too deep for comfort, toss.
834 834 */
835 835 ASSERT(fp - (uintptr_t)stkbase >= RED_DEEP_THRESHOLD);
836 836 return (0);
837 837 #endif /* _LP64 */
838 838 }
839 839
840 840 void
841 841 segkp_unmap_red(void)
842 842 {
843 843 page_t *pp;
844 844 caddr_t red_va = (caddr_t)(((uintptr_t)curthread->t_stkbase &
845 845 (uintptr_t)PAGEMASK) - PAGESIZE);
846 846
847 847 ASSERT(curthread->t_red_pp != NULL);
848 848
849 849 /*
850 850 * Because we locked the mapping down, we can't simply rely
851 851 * on page_destroy() to clean everything up; we need to call
852 852 * hat_unload() to explicitly unlock the mapping resources.
853 853 */
854 854 hat_unload(kas.a_hat, red_va, PAGESIZE, HAT_UNLOAD_UNLOCK);
855 855
856 856 pp = curthread->t_red_pp;
857 857
858 858 ASSERT(pp == page_find(&kvp, (u_offset_t)(uintptr_t)red_va));
859 859
860 860 /*
861 861 * Need to upgrade the SE_SHARED lock to SE_EXCL.
862 862 */
863 863 if (!page_tryupgrade(pp)) {
864 864 /*
865 865 * As there is now wait for upgrade, release the
866 866 * SE_SHARED lock and wait for SE_EXCL.
867 867 */
868 868 page_unlock(pp);
869 869 pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)red_va, SE_EXCL);
870 870 /* pp may be NULL here, hence the test below */
871 871 }
872 872
873 873 /*
874 874 * Destroy the page, with dontfree set to zero (i.e. free it).
875 875 */
876 876 if (pp != NULL)
877 877 page_destroy(pp, 0);
878 878 curthread->t_red_pp = NULL;
879 879 }
880 880 #else
881 881 #error Red stacks only supported with downwards stack growth.
882 882 #endif
883 883
884 884 /*
885 885 * Handle a fault on an address corresponding to one of the
886 886 * resources in the segkp segment.
887 887 */
888 888 faultcode_t
889 889 segkp_fault(
890 890 struct hat *hat,
891 891 struct seg *seg,
892 892 caddr_t vaddr,
893 893 size_t len,
894 894 enum fault_type type,
895 895 enum seg_rw rw)
896 896 {
897 897 struct segkp_data *kpd = NULL;
898 898 int err;
899 899
900 900 ASSERT(seg->s_as == &kas && RW_READ_HELD(&seg->s_as->a_lock));
901 901
902 902 /*
903 903 * Sanity checks.
904 904 */
905 905 if (type == F_PROT) {
906 906 panic("segkp_fault: unexpected F_PROT fault");
907 907 /*NOTREACHED*/
908 908 }
909 909
910 910 if ((kpd = segkp_find(seg, vaddr)) == NULL)
911 911 return (FC_NOMAP);
912 912
913 913 mutex_enter(&kpd->kp_lock);
914 914
915 915 if (type == F_SOFTLOCK) {
916 916 ASSERT(!(kpd->kp_flags & KPD_LOCKED));
917 917 /*
918 918 * The F_SOFTLOCK case has more stringent
919 919 * range requirements: the given range must exactly coincide
920 920 * with the resource's mapped portion. Note reference to
921 921 * redzone is handled since vaddr would not equal base
922 922 */
923 923 if (vaddr != stom(kpd->kp_base, kpd->kp_flags) ||
924 924 len != SEGKP_MAPLEN(kpd->kp_len, kpd->kp_flags)) {
925 925 mutex_exit(&kpd->kp_lock);
926 926 return (FC_MAKE_ERR(EFAULT));
927 927 }
928 928
929 929 if ((err = segkp_load(hat, seg, vaddr, len, kpd, KPD_LOCKED))) {
930 930 mutex_exit(&kpd->kp_lock);
931 931 return (FC_MAKE_ERR(err));
932 932 }
933 933 kpd->kp_flags |= KPD_LOCKED;
934 934 mutex_exit(&kpd->kp_lock);
935 935 return (0);
936 936 }
937 937
938 938 if (type == F_INVAL) {
939 939 ASSERT(!(kpd->kp_flags & KPD_NO_ANON));
940 940
941 941 /*
942 942 * Check if we touched the redzone. Somewhat optimistic
943 943 * here if we are touching the redzone of our own stack
944 944 * since we wouldn't have a stack to get this far...
945 945 */
946 946 if ((kpd->kp_flags & KPD_HASREDZONE) &&
947 947 btop((uintptr_t)(vaddr - kpd->kp_base)) == KPD_REDZONE(kpd))
948 948 panic("segkp_fault: accessing redzone");
949 949
950 950 /*
951 951 * This fault may occur while the page is being F_SOFTLOCK'ed.
952 952 * Return since a 2nd segkp_load is unnecessary and also would
953 953 * result in the page being locked twice and eventually
954 954 * hang the thread_reaper thread.
955 955 */
956 956 if (kpd->kp_flags & KPD_LOCKED) {
957 957 mutex_exit(&kpd->kp_lock);
958 958 return (0);
959 959 }
960 960
961 961 err = segkp_load(hat, seg, vaddr, len, kpd, kpd->kp_flags);
962 962 mutex_exit(&kpd->kp_lock);
963 963 return (err ? FC_MAKE_ERR(err) : 0);
964 964 }
965 965
966 966 if (type == F_SOFTUNLOCK) {
967 967 uint_t flags;
968 968
969 969 /*
970 970 * Make sure the addr is LOCKED and it has anon backing
971 971 * before unlocking
972 972 */
973 973 if ((kpd->kp_flags & (KPD_LOCKED|KPD_NO_ANON)) != KPD_LOCKED) {
974 974 panic("segkp_fault: bad unlock");
975 975 /*NOTREACHED*/
976 976 }
977 977
978 978 if (vaddr != stom(kpd->kp_base, kpd->kp_flags) ||
979 979 len != SEGKP_MAPLEN(kpd->kp_len, kpd->kp_flags)) {
980 980 panic("segkp_fault: bad range");
981 981 /*NOTREACHED*/
982 982 }
983 983
984 984 if (rw == S_WRITE)
985 985 flags = kpd->kp_flags | KPD_WRITEDIRTY;
986 986 else
987 987 flags = kpd->kp_flags;
988 988 err = segkp_unlock(hat, seg, vaddr, len, kpd, flags);
989 989 kpd->kp_flags &= ~KPD_LOCKED;
990 990 mutex_exit(&kpd->kp_lock);
991 991 return (err ? FC_MAKE_ERR(err) : 0);
992 992 }
993 993 mutex_exit(&kpd->kp_lock);
994 994 panic("segkp_fault: bogus fault type: %d\n", type);
995 995 /*NOTREACHED*/
996 996 }
997 997
998 998 /*
999 999 * Check that the given protections suffice over the range specified by
1000 1000 * vaddr and len. For this segment type, the only issue is whether or
1001 1001 * not the range lies completely within the mapped part of an allocated
1002 1002 * resource.
1003 1003 */
1004 1004 /* ARGSUSED */
1005 1005 static int
1006 1006 segkp_checkprot(struct seg *seg, caddr_t vaddr, size_t len, uint_t prot)
1007 1007 {
1008 1008 struct segkp_data *kpd = NULL;
1009 1009 caddr_t mbase;
1010 1010 size_t mlen;
1011 1011
1012 1012 if ((kpd = segkp_find(seg, vaddr)) == NULL)
1013 1013 return (EACCES);
1014 1014
1015 1015 mutex_enter(&kpd->kp_lock);
1016 1016 mbase = stom(kpd->kp_base, kpd->kp_flags);
1017 1017 mlen = SEGKP_MAPLEN(kpd->kp_len, kpd->kp_flags);
1018 1018 if (len > mlen || vaddr < mbase ||
1019 1019 ((vaddr + len) > (mbase + mlen))) {
1020 1020 mutex_exit(&kpd->kp_lock);
1021 1021 return (EACCES);
1022 1022 }
1023 1023 mutex_exit(&kpd->kp_lock);
1024 1024 return (0);
1025 1025 }
1026 1026
1027 1027
1028 1028 /*
1029 1029 * Check to see if it makes sense to do kluster/read ahead to
1030 1030 * addr + delta relative to the mapping at addr. We assume here
1031 1031 * that delta is a signed PAGESIZE'd multiple (which can be negative).
1032 1032 *
1033 1033 * For seg_u we always "approve" of this action from our standpoint.
1034 1034 */
1035 1035 /*ARGSUSED*/
1036 1036 static int
1037 1037 segkp_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
1038 1038 {
1039 1039 return (0);
1040 1040 }
1041 1041
1042 1042 /*
1043 1043 * Load and possibly lock intra-slot resources in the range given by
1044 1044 * vaddr and len.
1045 1045 */
1046 1046 static int
1047 1047 segkp_load(
1048 1048 struct hat *hat,
1049 1049 struct seg *seg,
1050 1050 caddr_t vaddr,
1051 1051 size_t len,
1052 1052 struct segkp_data *kpd,
1053 1053 uint_t flags)
1054 1054 {
1055 1055 caddr_t va;
1056 1056 caddr_t vlim;
1057 1057 ulong_t i;
1058 1058 uint_t lock;
1059 1059
1060 1060 ASSERT(MUTEX_HELD(&kpd->kp_lock));
1061 1061
1062 1062 len = P2ROUNDUP(len, PAGESIZE);
1063 1063
1064 1064 /* If locking, reserve physical memory */
1065 1065 if (flags & KPD_LOCKED) {
1066 1066 pgcnt_t pages = btop(len);
1067 1067 if ((kpd->kp_flags & KPD_NO_ANON) == 0)
1068 1068 atomic_add_long(&anon_segkp_pages_locked, pages);
1069 1069 (void) page_resv(pages, KM_SLEEP);
1070 1070 }
1071 1071
1072 1072 /*
1073 1073 * Loop through the pages in the given range.
1074 1074 */
1075 1075 va = (caddr_t)((uintptr_t)vaddr & (uintptr_t)PAGEMASK);
1076 1076 vaddr = va;
1077 1077 vlim = va + len;
1078 1078 lock = flags & KPD_LOCKED;
1079 1079 i = ((uintptr_t)(va - kpd->kp_base)) >> PAGESHIFT;
1080 1080 for (; va < vlim; va += PAGESIZE, i++) {
1081 1081 page_t *pl[2]; /* second element NULL terminator */
1082 1082 struct vnode *vp;
1083 1083 anoff_t off;
1084 1084 int err;
1085 1085 struct anon *ap;
1086 1086
1087 1087 /*
1088 1088 * Summon the page. If it's not resident, arrange
1089 1089 * for synchronous i/o to pull it in.
1090 1090 */
1091 1091 ap = anon_get_ptr(kpd->kp_anon, kpd->kp_anon_idx + i);
1092 1092 swap_xlate(ap, &vp, &off);
1093 1093
1094 1094 /*
1095 1095 * The returned page list will have exactly one entry,
1096 1096 * which is returned to us already kept.
1097 1097 */
1098 1098 err = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE, NULL,
1099 1099 pl, PAGESIZE, seg, va, S_READ, kcred, NULL);
1100 1100
1101 1101 if (err) {
1102 1102 /*
1103 1103 * Back out of what we've done so far.
1104 1104 */
1105 1105 (void) segkp_unlock(hat, seg, vaddr,
1106 1106 (va - vaddr), kpd, flags);
1107 1107 return (err);
1108 1108 }
1109 1109
1110 1110 /*
1111 1111 * Load an MMU translation for the page.
1112 1112 */
1113 1113 hat_memload(hat, va, pl[0], (PROT_READ|PROT_WRITE),
1114 1114 lock ? HAT_LOAD_LOCK : HAT_LOAD);
1115 1115
1116 1116 if (!lock) {
1117 1117 /*
1118 1118 * Now, release "shared" lock on the page.
1119 1119 */
1120 1120 page_unlock(pl[0]);
1121 1121 }
1122 1122 }
1123 1123 return (0);
1124 1124 }
1125 1125
1126 1126 /*
1127 1127 * At the very least unload the mmu-translations and unlock the range if locked
1128 1128 * Can be called with the following flag value KPD_WRITEDIRTY which specifies
1129 1129 * any dirty pages should be written to disk.
1130 1130 */
1131 1131 static int
1132 1132 segkp_unlock(
1133 1133 struct hat *hat,
1134 1134 struct seg *seg,
1135 1135 caddr_t vaddr,
1136 1136 size_t len,
1137 1137 struct segkp_data *kpd,
1138 1138 uint_t flags)
1139 1139 {
1140 1140 caddr_t va;
1141 1141 caddr_t vlim;
1142 1142 ulong_t i;
1143 1143 struct page *pp;
1144 1144 struct vnode *vp;
1145 1145 anoff_t off;
1146 1146 struct anon *ap;
1147 1147
1148 1148 #ifdef lint
1149 1149 seg = seg;
1150 1150 #endif /* lint */
1151 1151
1152 1152 ASSERT(MUTEX_HELD(&kpd->kp_lock));
1153 1153
1154 1154 /*
1155 1155 * Loop through the pages in the given range. It is assumed
1156 1156 * segkp_unlock is called with page aligned base
1157 1157 */
1158 1158 va = vaddr;
1159 1159 vlim = va + len;
1160 1160 i = ((uintptr_t)(va - kpd->kp_base)) >> PAGESHIFT;
1161 1161 hat_unload(hat, va, len,
1162 1162 ((flags & KPD_LOCKED) ? HAT_UNLOAD_UNLOCK : HAT_UNLOAD));
1163 1163 for (; va < vlim; va += PAGESIZE, i++) {
1164 1164 /*
1165 1165 * Find the page associated with this part of the
1166 1166 * slot, tracking it down through its associated swap
1167 1167 * space.
1168 1168 */
1169 1169 ap = anon_get_ptr(kpd->kp_anon, kpd->kp_anon_idx + i);
1170 1170 swap_xlate(ap, &vp, &off);
1171 1171
1172 1172 if (flags & KPD_LOCKED) {
1173 1173 if ((pp = page_find(vp, off)) == NULL) {
1174 1174 if (flags & KPD_LOCKED) {
1175 1175 panic("segkp_softunlock: missing page");
1176 1176 /*NOTREACHED*/
1177 1177 }
1178 1178 }
1179 1179 } else {
1180 1180 /*
1181 1181 * Nothing to do if the slot is not locked and the
1182 1182 * page doesn't exist.
1183 1183 */
1184 1184 if ((pp = page_lookup(vp, off, SE_SHARED)) == NULL)
1185 1185 continue;
1186 1186 }
1187 1187
1188 1188 /*
1189 1189 * If the page doesn't have any translations, is
1190 1190 * dirty and not being shared, then push it out
1191 1191 * asynchronously and avoid waiting for the
1192 1192 * pageout daemon to do it for us.
1193 1193 *
1194 1194 * XXX - Do we really need to get the "exclusive"
1195 1195 * lock via an upgrade?
1196 1196 */
1197 1197 if ((flags & KPD_WRITEDIRTY) && !hat_page_is_mapped(pp) &&
1198 1198 hat_ismod(pp) && page_tryupgrade(pp)) {
1199 1199 /*
1200 1200 * Hold the vnode before releasing the page lock to
1201 1201 * prevent it from being freed and re-used by some
1202 1202 * other thread.
1203 1203 */
1204 1204 VN_HOLD(vp);
1205 1205 page_unlock(pp);
1206 1206
1207 1207 /*
1208 1208 * Want most powerful credentials we can get so
1209 1209 * use kcred.
1210 1210 */
1211 1211 (void) VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
1212 1212 B_ASYNC | B_FREE, kcred, NULL);
1213 1213 VN_RELE(vp);
1214 1214 } else {
1215 1215 page_unlock(pp);
1216 1216 }
1217 1217 }
1218 1218
1219 1219 /* If unlocking, release physical memory */
1220 1220 if (flags & KPD_LOCKED) {
1221 1221 pgcnt_t pages = btopr(len);
1222 1222 if ((kpd->kp_flags & KPD_NO_ANON) == 0)
1223 1223 atomic_add_long(&anon_segkp_pages_locked, -pages);
1224 1224 page_unresv(pages);
1225 1225 }
1226 1226 return (0);
1227 1227 }
1228 1228
1229 1229 /*
1230 1230 * Insert the kpd in the hash table.
1231 1231 */
1232 1232 static void
1233 1233 segkp_insert(struct seg *seg, struct segkp_data *kpd)
1234 1234 {
1235 1235 struct segkp_segdata *kpsd = (struct segkp_segdata *)seg->s_data;
1236 1236 int index;
1237 1237
1238 1238 /*
1239 1239 * Insert the kpd based on the address that will be returned
1240 1240 * via segkp_release.
1241 1241 */
1242 1242 index = SEGKP_HASH(stom(kpd->kp_base, kpd->kp_flags));
1243 1243 mutex_enter(&segkp_lock);
1244 1244 kpd->kp_next = kpsd->kpsd_hash[index];
1245 1245 kpsd->kpsd_hash[index] = kpd;
1246 1246 mutex_exit(&segkp_lock);
1247 1247 }
1248 1248
1249 1249 /*
1250 1250 * Remove kpd from the hash table.
1251 1251 */
1252 1252 static void
1253 1253 segkp_delete(struct seg *seg, struct segkp_data *kpd)
1254 1254 {
1255 1255 struct segkp_segdata *kpsd = (struct segkp_segdata *)seg->s_data;
1256 1256 struct segkp_data **kpp;
1257 1257 int index;
1258 1258
1259 1259 ASSERT(MUTEX_HELD(&segkp_lock));
1260 1260
1261 1261 index = SEGKP_HASH(stom(kpd->kp_base, kpd->kp_flags));
1262 1262 for (kpp = &kpsd->kpsd_hash[index];
1263 1263 *kpp != NULL; kpp = &((*kpp)->kp_next)) {
1264 1264 if (*kpp == kpd) {
1265 1265 *kpp = kpd->kp_next;
1266 1266 return;
1267 1267 }
1268 1268 }
1269 1269 panic("segkp_delete: unable to find element to delete");
1270 1270 /*NOTREACHED*/
1271 1271 }
1272 1272
1273 1273 /*
1274 1274 * Find the kpd associated with a vaddr.
1275 1275 *
1276 1276 * Most of the callers of segkp_find will pass the vaddr that
1277 1277 * hashes to the desired index, but there are cases where
1278 1278 * this is not true in which case we have to (potentially) scan
1279 1279 * the whole table looking for it. This should be very rare
1280 1280 * (e.g. a segkp_fault(F_INVAL) on an address somewhere in the
1281 1281 * middle of the segkp_data region).
1282 1282 */
1283 1283 static struct segkp_data *
1284 1284 segkp_find(struct seg *seg, caddr_t vaddr)
1285 1285 {
1286 1286 struct segkp_segdata *kpsd = (struct segkp_segdata *)seg->s_data;
1287 1287 struct segkp_data *kpd;
1288 1288 int i;
1289 1289 int stop;
1290 1290
1291 1291 i = stop = SEGKP_HASH(vaddr);
1292 1292 mutex_enter(&segkp_lock);
1293 1293 do {
1294 1294 for (kpd = kpsd->kpsd_hash[i]; kpd != NULL;
1295 1295 kpd = kpd->kp_next) {
1296 1296 if (vaddr >= kpd->kp_base &&
1297 1297 vaddr < kpd->kp_base + kpd->kp_len) {
1298 1298 mutex_exit(&segkp_lock);
1299 1299 return (kpd);
1300 1300 }
1301 1301 }
1302 1302 if (--i < 0)
1303 1303 i = SEGKP_HASHSZ - 1; /* Wrap */
1304 1304 } while (i != stop);
1305 1305 mutex_exit(&segkp_lock);
1306 1306 return (NULL); /* Not found */
1307 1307 }
1308 1308
1309 1309 /*
1310 1310 * returns size of swappable area.
1311 1311 */
1312 1312 size_t
1313 1313 swapsize(caddr_t v)
1314 1314 {
1315 1315 struct segkp_data *kpd;
1316 1316
1317 1317 if ((kpd = segkp_find(segkp, v)) != NULL)
1318 1318 return (SEGKP_MAPLEN(kpd->kp_len, kpd->kp_flags));
1319 1319 else
1320 1320 return (NULL);
1321 1321 }
1322 1322
1323 1323 /*
1324 1324 * Dump out all the active segkp pages
1325 1325 */
1326 1326 static void
1327 1327 segkp_dump(struct seg *seg)
1328 1328 {
1329 1329 int i;
1330 1330 struct segkp_data *kpd;
1331 1331 struct segkp_segdata *kpsd = (struct segkp_segdata *)seg->s_data;
1332 1332
1333 1333 for (i = 0; i < SEGKP_HASHSZ; i++) {
1334 1334 for (kpd = kpsd->kpsd_hash[i];
1335 1335 kpd != NULL; kpd = kpd->kp_next) {
1336 1336 pfn_t pfn;
1337 1337 caddr_t addr;
1338 1338 caddr_t eaddr;
1339 1339
1340 1340 addr = kpd->kp_base;
1341 1341 eaddr = addr + kpd->kp_len;
1342 1342 while (addr < eaddr) {
1343 1343 ASSERT(seg->s_as == &kas);
1344 1344 pfn = hat_getpfnum(seg->s_as->a_hat, addr);
1345 1345 if (pfn != PFN_INVALID)
1346 1346 dump_addpage(seg->s_as, addr, pfn);
1347 1347 addr += PAGESIZE;
1348 1348 dump_timeleft = dump_timeout;
1349 1349 }
1350 1350 }
1351 1351 }
1352 1352 }
1353 1353
1354 1354 /*ARGSUSED*/
1355 1355 static int
1356 1356 segkp_pagelock(struct seg *seg, caddr_t addr, size_t len,
1357 1357 struct page ***ppp, enum lock_type type, enum seg_rw rw)
1358 1358 {
1359 1359 return (ENOTSUP);
1360 1360 }
1361 1361
1362 1362 #include <sys/mem_config.h>
1363 1363
1364 1364 /*ARGSUSED*/
1365 1365 static void
1366 1366 segkp_mem_config_post_add(void *arg, pgcnt_t delta_pages)
1367 1367 {}
1368 1368
1369 1369 /*
1370 1370 * During memory delete, turn off caches so that pages are not held.
1371 1371 * A better solution may be to unlock the pages while they are
1372 1372 * in the cache so that they may be collected naturally.
1373 1373 */
1374 1374
1375 1375 /*ARGSUSED*/
1376 1376 static int
1377 1377 segkp_mem_config_pre_del(void *arg, pgcnt_t delta_pages)
1378 1378 {
1379 1379 atomic_inc_32(&segkp_indel);
1380 1380 segkp_cache_free();
1381 1381 return (0);
1382 1382 }
1383 1383
1384 1384 /*ARGSUSED*/
1385 1385 static void
1386 1386 segkp_mem_config_post_del(void *arg, pgcnt_t delta_pages, int cancelled)
1387 1387 {
1388 1388 atomic_dec_32(&segkp_indel);
1389 1389 }
1390 1390
1391 1391 static kphysm_setup_vector_t segkp_mem_config_vec = {
1392 1392 KPHYSM_SETUP_VECTOR_VERSION,
1393 1393 segkp_mem_config_post_add,
1394 1394 segkp_mem_config_pre_del,
1395 1395 segkp_mem_config_post_del,
1396 1396 };
1397 1397
1398 1398 static void
1399 1399 segkpinit_mem_config(struct seg *seg)
1400 1400 {
1401 1401 int ret;
1402 1402
1403 1403 ret = kphysm_setup_func_register(&segkp_mem_config_vec, (void *)seg);
1404 1404 ASSERT(ret == 0);
1405 1405 }
|
↓ open down ↓ |
1254 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX