?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge commit 'ea01a15a654b9e1c7b37d958f4d1911882ed7781'
[unleashed.git] / kernel / vm / seg_spt.c
blobafd1ace4e4f627cedd956ec3ea3b5d6e794aca4d
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) 1993, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2018 Joyent, Inc.
24 * Copyright (c) 2016 by Delphix. All rights reserved.
25 */
26
27 #include <sys/param.h>
28 #include <sys/user.h>
29 #include <sys/mman.h>
30 #include <sys/kmem.h>
31 #include <sys/sysmacros.h>
32 #include <sys/cmn_err.h>
33 #include <sys/systm.h>
34 #include <sys/tuneable.h>
35 #include <vm/hat.h>
36 #include <vm/seg.h>
37 #include <vm/as.h>
38 #include <vm/anon.h>
39 #include <vm/page.h>
40 #include <sys/buf.h>
41 #include <sys/swap.h>
42 #include <sys/atomic.h>
43 #include <vm/seg_spt.h>
44 #include <sys/debug.h>
45 #include <sys/vtrace.h>
46 #include <sys/shm.h>
47 #include <sys/shm_impl.h>
48 #include <sys/lgrp.h>
49 #include <sys/vmsystm.h>
50 #include <sys/policy.h>
51 #include <sys/project.h>
52 #include <sys/tnf_probe.h>
53 #include <sys/zone.h>
54
55 #define SEGSPTADDR (caddr_t)0x0
56
57 /*
58 * # pages used for spt
59 */
60 size_t spt_used;
61
62 /*
63 * segspt_minfree is the memory left for system after ISM
64 * locked its pages; it is set up to 5% of availrmem in
65 * sptcreate when ISM is created. ISM should not use more
66 * than ~90% of availrmem; if it does, then the performance
67 * of the system may decrease. Machines with large memories may
68 * be able to use up more memory for ISM so we set the default
69 * segspt_minfree to 5% (which gives ISM max 95% of availrmem.
70 * If somebody wants even more memory for ISM (risking hanging
71 * the system) they can patch the segspt_minfree to smaller number.
72 */
73 pgcnt_t segspt_minfree = 0;
74
75 static int segspt_create(struct seg **segpp, void *argsp);
76 static int segspt_unmap(struct seg *seg, caddr_t raddr, size_t ssize);
77 static void segspt_free(struct seg *seg);
78 static void segspt_free_pages(struct seg *seg, caddr_t addr, size_t len);
79 static lgrp_mem_policy_info_t *segspt_getpolicy(struct seg *seg, caddr_t addr);
80
81 static void
82 segspt_badop()
83 {
84 panic("segspt_badop called");
85 /*NOTREACHED*/
86 }
87
88 #define SEGSPT_BADOP(t) (t(*)())segspt_badop
89
90 static const struct seg_ops segspt_ops = {
91 .dup = SEGSPT_BADOP(int),
92 .unmap = segspt_unmap,
93 .free = segspt_free,
94 .fault = SEGSPT_BADOP(int),
95 .faulta = SEGSPT_BADOP(faultcode_t),
96 .setprot = SEGSPT_BADOP(int),
97 .checkprot = SEGSPT_BADOP(int),
98 .kluster = SEGSPT_BADOP(int),
99 .sync = SEGSPT_BADOP(int),
100 .incore = SEGSPT_BADOP(size_t),
101 .lockop = SEGSPT_BADOP(int),
102 .getprot = SEGSPT_BADOP(int),
103 .getoffset = SEGSPT_BADOP(uoff_t),
104 .gettype = SEGSPT_BADOP(int),
105 .getvp = SEGSPT_BADOP(int),
106 .advise = SEGSPT_BADOP(int),
107 .dump = SEGSPT_BADOP(void),
108 .pagelock = SEGSPT_BADOP(int),
109 .setpagesize = SEGSPT_BADOP(int),
110 .getmemid = SEGSPT_BADOP(int),
111 .getpolicy = segspt_getpolicy,
112 .capable = SEGSPT_BADOP(int),
113 };
114
115 static int segspt_shmdup(struct seg *seg, struct seg *newseg);
116 static int segspt_shmunmap(struct seg *seg, caddr_t raddr, size_t ssize);
117 static void segspt_shmfree(struct seg *seg);
118 static faultcode_t segspt_shmfault(struct hat *hat, struct seg *seg,
119 caddr_t addr, size_t len, enum fault_type type, enum seg_rw rw);
120 static faultcode_t segspt_shmfaulta(struct seg *seg, caddr_t addr);
121 static int segspt_shmsetprot(register struct seg *seg, register caddr_t addr,
122 register size_t len, register uint_t prot);
123 static int segspt_shmcheckprot(struct seg *seg, caddr_t addr, size_t size,
124 uint_t prot);
125 static int segspt_shmkluster(struct seg *seg, caddr_t addr, ssize_t delta);
126 static size_t segspt_shmincore(struct seg *seg, caddr_t addr, size_t len,
127 register char *vec);
128 static int segspt_shmsync(struct seg *seg, register caddr_t addr, size_t len,
129 int attr, uint_t flags);
130 static int segspt_shmlockop(struct seg *seg, caddr_t addr, size_t len,
131 int attr, int op, ulong_t *lockmap, size_t pos);
132 static int segspt_shmgetprot(struct seg *seg, caddr_t addr, size_t len,
133 uint_t *protv);
134 static uoff_t segspt_shmgetoffset(struct seg *seg, caddr_t addr);
135 static int segspt_shmgettype(struct seg *seg, caddr_t addr);
136 static int segspt_shmgetvp(struct seg *seg, caddr_t addr, struct vnode **vpp);
137 static int segspt_shmadvise(struct seg *seg, caddr_t addr, size_t len,
138 uint_t behav);
139 static int segspt_shmpagelock(struct seg *, caddr_t, size_t,
140 struct page ***, enum lock_type, enum seg_rw);
141 static int segspt_shmgetmemid(struct seg *, caddr_t, memid_t *);
142 static lgrp_mem_policy_info_t *segspt_shmgetpolicy(struct seg *, caddr_t);
143
144 const struct seg_ops segspt_shmops = {
145 .dup = segspt_shmdup,
146 .unmap = segspt_shmunmap,
147 .free = segspt_shmfree,
148 .fault = segspt_shmfault,
149 .faulta = segspt_shmfaulta,
150 .setprot = segspt_shmsetprot,
151 .checkprot = segspt_shmcheckprot,
152 .kluster = segspt_shmkluster,
153 .sync = segspt_shmsync,
154 .incore = segspt_shmincore,
155 .lockop = segspt_shmlockop,
156 .getprot = segspt_shmgetprot,
157 .getoffset = segspt_shmgetoffset,
158 .gettype = segspt_shmgettype,
159 .getvp = segspt_shmgetvp,
160 .advise = segspt_shmadvise,
161 .pagelock = segspt_shmpagelock,
162 .getmemid = segspt_shmgetmemid,
163 .getpolicy = segspt_shmgetpolicy,
164 };
165
166 static void segspt_purge(struct seg *seg);
167 static int segspt_reclaim(void *, caddr_t, size_t, struct page **,
168 enum seg_rw, int);
169 static int spt_anon_getpages(struct seg *seg, caddr_t addr, size_t len,
170 page_t **ppa);
171
172
173
174 /*ARGSUSED*/
175 int
176 sptcreate(size_t size, struct seg **sptseg, struct anon_map *amp,
177 uint_t prot, uint_t flags, uint_t share_szc)
178 {
179 int err;
180 struct as *newas;
181 struct segspt_crargs sptcargs;
182
183 if (segspt_minfree == 0) /* leave min 5% of availrmem for */
184 segspt_minfree = availrmem/20; /* for the system */
185
186 if (!hat_supported(HAT_SHARED_PT, NULL))
187 return (EINVAL);
188
189 /*
190 * get a new as for this shared memory segment
191 */
192 newas = as_alloc();
193 newas->a_proc = NULL;
194 sptcargs.amp = amp;
195 sptcargs.prot = prot;
196 sptcargs.flags = flags;
197 sptcargs.szc = share_szc;
198 /*
199 * create a shared page table (spt) segment
200 */
201
202 if (err = as_map(newas, SEGSPTADDR, size, segspt_create, &sptcargs)) {
203 as_free(newas);
204 return (err);
205 }
206 *sptseg = sptcargs.seg_spt;
207 return (0);
208 }
209
210 void
211 sptdestroy(struct as *as, struct anon_map *amp)
212 {
213 (void) as_unmap(as, SEGSPTADDR, amp->size);
214 as_free(as);
215 }
216
217 /*
218 * called from seg_free().
219 * free (i.e., unlock, unmap, return to free list)
220 * all the pages in the given seg.
221 */
222 void
223 segspt_free(struct seg *seg)
224 {
225 struct spt_data *sptd = (struct spt_data *)seg->s_data;
226
227 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
228
229 if (sptd != NULL) {
230 if (sptd->spt_realsize)
231 segspt_free_pages(seg, seg->s_base, sptd->spt_realsize);
232
233 if (sptd->spt_ppa_lckcnt) {
234 kmem_free(sptd->spt_ppa_lckcnt,
235 sizeof (*sptd->spt_ppa_lckcnt)
236 * btopr(sptd->spt_amp->size));
237 }
238 kmem_free(sptd->spt_vp, sizeof (*sptd->spt_vp));
239 cv_destroy(&sptd->spt_cv);
240 mutex_destroy(&sptd->spt_lock);
241 kmem_free(sptd, sizeof (*sptd));
242 }
243 }
244
245 /*ARGSUSED*/
246 static int
247 segspt_shmsync(struct seg *seg, caddr_t addr, size_t len, int attr,
248 uint_t flags)
249 {
250 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
251
252 return (0);
253 }
254
255 /*ARGSUSED*/
256 static size_t
257 segspt_shmincore(struct seg *seg, caddr_t addr, size_t len, char *vec)
258 {
259 caddr_t eo_seg;
260 pgcnt_t npages;
261 struct shm_data *shmd = (struct shm_data *)seg->s_data;
262 struct seg *sptseg;
263 struct spt_data *sptd;
264
265 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
266 sptseg = shmd->shm_sptseg;
267 sptd = sptseg->s_data;
268
269 if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
270 eo_seg = addr + len;
271 while (addr < eo_seg) {
272 /* page exists, and it's locked. */
273 *vec++ = SEG_PAGE_INCORE | SEG_PAGE_LOCKED |
274 SEG_PAGE_ANON;
275 addr += PAGESIZE;
276 }
277 return (len);
278 } else {
279 struct anon_map *amp = shmd->shm_amp;
280 struct anon *ap;
281 page_t *pp;
282 pgcnt_t anon_index;
283 struct vnode *vp;
284 uoff_t off;
285 ulong_t i;
286 int ret;
287 anon_sync_obj_t cookie;
288
289 addr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
290 anon_index = seg_page(seg, addr);
291 npages = btopr(len);
292 if (anon_index + npages > btopr(shmd->shm_amp->size)) {
293 return (EINVAL);
294 }
295 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
296 for (i = 0; i < npages; i++, anon_index++) {
297 ret = 0;
298 anon_array_enter(amp, anon_index, &cookie);
299 ap = anon_get_ptr(amp->ahp, anon_index);
300 if (ap != NULL) {
301 swap_xlate(ap, &vp, &off);
302 anon_array_exit(&cookie);
303 pp = page_lookup_nowait(&vp->v_object, off,
304 SE_SHARED);
305 if (pp != NULL) {
306 ret |= SEG_PAGE_INCORE | SEG_PAGE_ANON;
307 page_unlock(pp);
308 }
309 } else {
310 anon_array_exit(&cookie);
311 }
312 if (shmd->shm_vpage[anon_index] & DISM_PG_LOCKED) {
313 ret |= SEG_PAGE_LOCKED;
314 }
315 *vec++ = (char)ret;
316 }
317 ANON_LOCK_EXIT(&amp->a_rwlock);
318 return (len);
319 }
320 }
321
322 static int
323 segspt_unmap(struct seg *seg, caddr_t raddr, size_t ssize)
324 {
325 size_t share_size;
326
327 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
328
329 /*
330 * seg.s_size may have been rounded up to the largest page size
331 * in shmat().
332 * XXX This should be cleanedup. sptdestroy should take a length
333 * argument which should be the same as sptcreate. Then
334 * this rounding would not be needed (or is done in shm.c)
335 * Only the check for full segment will be needed.
336 *
337 * XXX -- shouldn't raddr == 0 always? These tests don't seem
338 * to be useful at all.
339 */
340 share_size = page_get_pagesize(seg->s_szc);
341 ssize = P2ROUNDUP(ssize, share_size);
342
343 if (raddr == seg->s_base && ssize == seg->s_size) {
344 seg_free(seg);
345 return (0);
346 } else
347 return (EINVAL);
348 }
349
350 int
351 segspt_create(struct seg **segpp, void *argsp)
352 {
353 struct seg *seg = *segpp;
354 int err;
355 caddr_t addr = seg->s_base;
356 struct spt_data *sptd;
357 struct segspt_crargs *sptcargs = (struct segspt_crargs *)argsp;
358 struct anon_map *amp = sptcargs->amp;
359 struct kshmid *sp = amp->a_sp;
360 struct cred *cred = CRED();
361 ulong_t i, j, anon_index = 0;
362 pgcnt_t npages = btopr(amp->size);
363 struct vnode *vp;
364 page_t **ppa;
365 uint_t hat_flags;
366 size_t pgsz;
367 pgcnt_t pgcnt;
368 caddr_t a;
369 pgcnt_t pidx;
370 size_t sz;
371 proc_t *procp = curproc;
372 rctl_qty_t lockedbytes = 0;
373 kproject_t *proj;
374
375 /*
376 * We are holding the a_lock on the underlying dummy as,
377 * so we can make calls to the HAT layer.
378 */
379 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
380 ASSERT(sp != NULL);
381
382 if ((sptcargs->flags & SHM_PAGEABLE) == 0) {
383 if (err = anon_swap_adjust(npages))
384 return (err);
385 }
386 err = ENOMEM;
387
388 if ((sptd = kmem_zalloc(sizeof (*sptd), KM_NOSLEEP)) == NULL)
389 goto out1;
390
391 if ((sptcargs->flags & SHM_PAGEABLE) == 0) {
392 if ((ppa = kmem_zalloc(((sizeof (page_t *)) * npages),
393 KM_NOSLEEP)) == NULL)
394 goto out2;
395 }
396
397 mutex_init(&sptd->spt_lock, NULL, MUTEX_DEFAULT, NULL);
398
399 if ((vp = kmem_zalloc(sizeof (*vp), KM_NOSLEEP)) == NULL)
400 goto out3;
401
402 seg->s_ops = &segspt_ops;
403 sptd->spt_vp = vp;
404 sptd->spt_amp = amp;
405 sptd->spt_prot = sptcargs->prot;
406 sptd->spt_flags = sptcargs->flags;
407 seg->s_data = (caddr_t)sptd;
408 sptd->spt_ppa = NULL;
409 sptd->spt_ppa_lckcnt = NULL;
410 seg->s_szc = sptcargs->szc;
411 cv_init(&sptd->spt_cv, NULL, CV_DEFAULT, NULL);
412 sptd->spt_gen = 0;
413
414 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
415 if (seg->s_szc > amp->a_szc) {
416 amp->a_szc = seg->s_szc;
417 }
418 ANON_LOCK_EXIT(&amp->a_rwlock);
419
420 /*
421 * Set policy to affect initial allocation of pages in
422 * anon_map_createpages()
423 */
424 (void) lgrp_shm_policy_set(LGRP_MEM_POLICY_DEFAULT, amp, anon_index,
425 NULL, 0, ptob(npages));
426
427 if (sptcargs->flags & SHM_PAGEABLE) {
428 size_t share_sz;
429 pgcnt_t new_npgs, more_pgs;
430 struct anon_hdr *nahp;
431 zone_t *zone;
432
433 share_sz = page_get_pagesize(seg->s_szc);
434 if (!IS_P2ALIGNED(amp->size, share_sz)) {
435 /*
436 * We are rounding up the size of the anon array
437 * on 4 M boundary because we always create 4 M
438 * of page(s) when locking, faulting pages and we
439 * don't have to check for all corner cases e.g.
440 * if there is enough space to allocate 4 M
441 * page.
442 */
443 new_npgs = btop(P2ROUNDUP(amp->size, share_sz));
444 more_pgs = new_npgs - npages;
445
446 /*
447 * The zone will never be NULL, as a fully created
448 * shm always has an owning zone.
449 */
450 zone = sp->shm_perm.ipc_zone_ref.zref_zone;
451 ASSERT(zone != NULL);
452 if (anon_resv_zone(ptob(more_pgs), zone) == 0) {
453 err = ENOMEM;
454 goto out4;
455 }
456
457 nahp = anon_create(new_npgs, ANON_SLEEP);
458 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
459 (void) anon_copy_ptr(amp->ahp, 0, nahp, 0, npages,
460 ANON_SLEEP);
461 anon_release(amp->ahp, npages);
462 amp->ahp = nahp;
463 ASSERT(amp->swresv == ptob(npages));
464 amp->swresv = amp->size = ptob(new_npgs);
465 ANON_LOCK_EXIT(&amp->a_rwlock);
466 npages = new_npgs;
467 }
468
469 sptd->spt_ppa_lckcnt = kmem_zalloc(npages *
470 sizeof (*sptd->spt_ppa_lckcnt), KM_SLEEP);
471 sptd->spt_pcachecnt = 0;
472 sptd->spt_realsize = ptob(npages);
473 sptcargs->seg_spt = seg;
474 return (0);
475 }
476
477 /*
478 * get array of pages for each anon slot in amp
479 */
480 if ((err = anon_map_createpages(amp, anon_index, ptob(npages), ppa,
481 seg, addr, S_CREATE, cred)) != 0)
482 goto out4;
483
484 mutex_enter(&sp->shm_mlock);
485
486 /* May be partially locked, so, count bytes to charge for locking */
487 for (i = 0; i < npages; i++)
488 if (ppa[i]->p_lckcnt == 0)
489 lockedbytes += PAGESIZE;
490
491 proj = sp->shm_perm.ipc_proj;
492
493 if (lockedbytes > 0) {
494 mutex_enter(&procp->p_lock);
495 if (rctl_incr_locked_mem(procp, proj, lockedbytes, 0)) {
496 mutex_exit(&procp->p_lock);
497 mutex_exit(&sp->shm_mlock);
498 for (i = 0; i < npages; i++)
499 page_unlock(ppa[i]);
500 err = ENOMEM;
501 goto out4;
502 }
503 mutex_exit(&procp->p_lock);
504 }
505
506 /*
507 * addr is initial address corresponding to the first page on ppa list
508 */
509 for (i = 0; i < npages; i++) {
510 /* attempt to lock all pages */
511 if (page_pp_lock(ppa[i], 0, 1) == 0) {
512 /*
513 * if unable to lock any page, unlock all
514 * of them and return error
515 */
516 for (j = 0; j < i; j++)
517 page_pp_unlock(ppa[j], 0, 1);
518 for (i = 0; i < npages; i++)
519 page_unlock(ppa[i]);
520 rctl_decr_locked_mem(NULL, proj, lockedbytes, 0);
521 mutex_exit(&sp->shm_mlock);
522 err = ENOMEM;
523 goto out4;
524 }
525 }
526 mutex_exit(&sp->shm_mlock);
527
528 /*
529 * Some platforms assume that ISM mappings are HAT_LOAD_LOCK
530 * for the entire life of the segment. For example platforms
531 * that do not support Dynamic Reconfiguration.
532 */
533 hat_flags = HAT_LOAD_SHARE;
534 if (!hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL))
535 hat_flags |= HAT_LOAD_LOCK;
536
537 /*
538 * Load translations one lare page at a time
539 * to make sure we don't create mappings bigger than
540 * segment's size code in case underlying pages
541 * are shared with segvn's segment that uses bigger
542 * size code than we do.
543 */
544 pgsz = page_get_pagesize(seg->s_szc);
545 pgcnt = page_get_pagecnt(seg->s_szc);
546 for (a = addr, pidx = 0; pidx < npages; a += pgsz, pidx += pgcnt) {
547 sz = MIN(pgsz, ptob(npages - pidx));
548 hat_memload_array(seg->s_as->a_hat, a, sz,
549 &ppa[pidx], sptd->spt_prot, hat_flags);
550 }
551
552 /*
553 * On platforms that do not support HAT_DYNAMIC_ISM_UNMAP,
554 * we will leave the pages locked SE_SHARED for the life
555 * of the ISM segment. This will prevent any calls to
556 * hat_pageunload() on this ISM segment for those platforms.
557 */
558 if (!(hat_flags & HAT_LOAD_LOCK)) {
559 /*
560 * On platforms that support HAT_DYNAMIC_ISM_UNMAP,
561 * we no longer need to hold the SE_SHARED lock on the pages,
562 * since L_PAGELOCK and F_SOFTLOCK calls will grab the
563 * SE_SHARED lock on the pages as necessary.
564 */
565 for (i = 0; i < npages; i++)
566 page_unlock(ppa[i]);
567 }
568 sptd->spt_pcachecnt = 0;
569 kmem_free(ppa, ((sizeof (page_t *)) * npages));
570 sptd->spt_realsize = ptob(npages);
571 atomic_add_long(&spt_used, npages);
572 sptcargs->seg_spt = seg;
573 return (0);
574
575 out4:
576 seg->s_data = NULL;
577 kmem_free(vp, sizeof (*vp));
578 cv_destroy(&sptd->spt_cv);
579 out3:
580 mutex_destroy(&sptd->spt_lock);
581 if ((sptcargs->flags & SHM_PAGEABLE) == 0)
582 kmem_free(ppa, (sizeof (*ppa) * npages));
583 out2:
584 kmem_free(sptd, sizeof (*sptd));
585 out1:
586 if ((sptcargs->flags & SHM_PAGEABLE) == 0)
587 anon_swap_restore(npages);
588 return (err);
589 }
590
591 /*ARGSUSED*/
592 void
593 segspt_free_pages(struct seg *seg, caddr_t addr, size_t len)
594 {
595 struct page *pp;
596 struct spt_data *sptd = (struct spt_data *)seg->s_data;
597 pgcnt_t npages;
598 ulong_t anon_idx;
599 struct anon_map *amp;
600 struct anon *ap;
601 struct vnode *vp;
602 uoff_t off;
603 uint_t hat_flags;
604 int root = 0;
605 pgcnt_t pgs, curnpgs = 0;
606 page_t *rootpp;
607 rctl_qty_t unlocked_bytes = 0;
608 kproject_t *proj;
609 kshmid_t *sp;
610
611 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
612
613 len = P2ROUNDUP(len, PAGESIZE);
614
615 npages = btop(len);
616
617 hat_flags = HAT_UNLOAD_UNLOCK | HAT_UNLOAD_UNMAP;
618 if ((hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL)) ||
619 (sptd->spt_flags & SHM_PAGEABLE)) {
620 hat_flags = HAT_UNLOAD_UNMAP;
621 }
622
623 hat_unload(seg->s_as->a_hat, addr, len, hat_flags);
624
625 amp = sptd->spt_amp;
626 if (sptd->spt_flags & SHM_PAGEABLE)
627 npages = btop(amp->size);
628
629 ASSERT(amp != NULL);
630
631 if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
632 sp = amp->a_sp;
633 proj = sp->shm_perm.ipc_proj;
634 mutex_enter(&sp->shm_mlock);
635 }
636 for (anon_idx = 0; anon_idx < npages; anon_idx++) {
637 if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
638 if ((ap = anon_get_ptr(amp->ahp, anon_idx)) == NULL) {
639 panic("segspt_free_pages: null app");
640 /*NOTREACHED*/
641 }
642 } else {
643 if ((ap = anon_get_next_ptr(amp->ahp, &anon_idx))
644 == NULL)
645 continue;
646 }
647 ASSERT(ANON_ISBUSY(anon_get_slot(amp->ahp, anon_idx)) == 0);
648 swap_xlate(ap, &vp, &off);
649
650 /*
651 * If this platform supports HAT_DYNAMIC_ISM_UNMAP,
652 * the pages won't be having SE_SHARED lock at this
653 * point.
654 *
655 * On platforms that do not support HAT_DYNAMIC_ISM_UNMAP,
656 * the pages are still held SE_SHARED locked from the
657 * original segspt_create()
658 *
659 * Our goal is to get SE_EXCL lock on each page, remove
660 * permanent lock on it and invalidate the page.
661 */
662 if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
663 if (hat_flags == HAT_UNLOAD_UNMAP)
664 pp = page_lookup(&vp->v_object, off, SE_EXCL);
665 else {
666 if ((pp = page_find(&vp->v_object, off)) == NULL) {
667 panic("segspt_free_pages: "
668 "page not locked");
669 /*NOTREACHED*/
670 }
671 if (!page_tryupgrade(pp)) {
672 page_unlock(pp);
673 pp = page_lookup(&vp->v_object, off,
674 SE_EXCL);
675 }
676 }
677 if (pp == NULL) {
678 panic("segspt_free_pages: "
679 "page not in the system");
680 /*NOTREACHED*/
681 }
682 ASSERT(pp->p_lckcnt > 0);
683 page_pp_unlock(pp, 0, 1);
684 if (pp->p_lckcnt == 0)
685 unlocked_bytes += PAGESIZE;
686 } else {
687 if ((pp = page_lookup(&vp->v_object, off, SE_EXCL)) == NULL)
688 continue;
689 }
690 /*
691 * It's logical to invalidate the pages here as in most cases
692 * these were created by segspt.
693 */
694 if (pp->p_szc != 0) {
695 if (root == 0) {
696 ASSERT(curnpgs == 0);
697 root = 1;
698 rootpp = pp;
699 pgs = curnpgs = page_get_pagecnt(pp->p_szc);
700 ASSERT(pgs > 1);
701 ASSERT(IS_P2ALIGNED(pgs, pgs));
702 ASSERT(!(page_pptonum(pp) & (pgs - 1)));
703 curnpgs--;
704 } else if ((page_pptonum(pp) & (pgs - 1)) == pgs - 1) {
705 ASSERT(curnpgs == 1);
706 ASSERT(page_pptonum(pp) ==
707 page_pptonum(rootpp) + (pgs - 1));
708 page_destroy_pages(rootpp);
709 root = 0;
710 curnpgs = 0;
711 } else {
712 ASSERT(curnpgs > 1);
713 ASSERT(page_pptonum(pp) ==
714 page_pptonum(rootpp) + (pgs - curnpgs));
715 curnpgs--;
716 }
717 } else {
718 if (root != 0 || curnpgs != 0) {
719 panic("segspt_free_pages: bad large page");
720 /*NOTREACHED*/
721 }
722 /*
723 * Before destroying the pages, we need to take care
724 * of the rctl locked memory accounting. For that
725 * we need to calculte the unlocked_bytes.
726 */
727 if (pp->p_lckcnt > 0)
728 unlocked_bytes += PAGESIZE;
729
730 VN_DISPOSE(pp, B_INVAL, 0, kcred);
731 }
732 }
733 if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
734 if (unlocked_bytes > 0)
735 rctl_decr_locked_mem(NULL, proj, unlocked_bytes, 0);
736 mutex_exit(&sp->shm_mlock);
737 }
738 if (root != 0 || curnpgs != 0) {
739 panic("segspt_free_pages: bad large page");
740 /*NOTREACHED*/
741 }
742
743 /*
744 * mark that pages have been released
745 */
746 sptd->spt_realsize = 0;
747
748 if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
749 atomic_add_long(&spt_used, -npages);
750 anon_swap_restore(npages);
751 }
752 }
753
754 /*
755 * Get memory allocation policy info for specified address in given segment
756 */
757 static lgrp_mem_policy_info_t *
758 segspt_getpolicy(struct seg *seg, caddr_t addr)
759 {
760 struct anon_map *amp;
761 ulong_t anon_index;
762 lgrp_mem_policy_info_t *policy_info;
763 struct spt_data *spt_data;
764
765 ASSERT(seg != NULL);
766
767 /*
768 * Get anon_map from segspt
769 *
770 * Assume that no lock needs to be held on anon_map, since
771 * it should be protected by its reference count which must be
772 * nonzero for an existing segment
773 * Need to grab readers lock on policy tree though
774 */
775 spt_data = (struct spt_data *)seg->s_data;
776 if (spt_data == NULL)
777 return (NULL);
778 amp = spt_data->spt_amp;
779 ASSERT(amp->refcnt != 0);
780
781 /*
782 * Get policy info
783 *
784 * Assume starting anon index of 0
785 */
786 anon_index = seg_page(seg, addr);
787 policy_info = lgrp_shm_policy_get(amp, anon_index, NULL, 0);
788
789 return (policy_info);
790 }
791
792 /*
793 * DISM only.
794 * Return locked pages over a given range.
795 *
796 * We will cache all DISM locked pages and save the pplist for the
797 * entire segment in the ppa field of the underlying DISM segment structure.
798 * Later, during a call to segspt_reclaim() we will use this ppa array
799 * to page_unlock() all of the pages and then we will free this ppa list.
800 */
801 /*ARGSUSED*/
802 static int
803 segspt_dismpagelock(struct seg *seg, caddr_t addr, size_t len,
804 struct page ***ppp, enum lock_type type, enum seg_rw rw)
805 {
806 struct shm_data *shmd = (struct shm_data *)seg->s_data;
807 struct seg *sptseg = shmd->shm_sptseg;
808 struct spt_data *sptd = sptseg->s_data;
809 pgcnt_t pg_idx, npages, tot_npages, npgs;
810 struct page **pplist, **pl, **ppa, *pp;
811 struct anon_map *amp;
812 spgcnt_t an_idx;
813 int ret = ENOTSUP;
814 uint_t pl_built = 0;
815 struct anon *ap;
816 struct vnode *vp;
817 uoff_t off;
818 pgcnt_t claim_availrmem = 0;
819 uint_t szc;
820
821 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
822 ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
823
824 /*
825 * We want to lock/unlock the entire ISM segment. Therefore,
826 * we will be using the underlying sptseg and it's base address
827 * and length for the caching arguments.
828 */
829 ASSERT(sptseg);
830 ASSERT(sptd);
831
832 pg_idx = seg_page(seg, addr);
833 npages = btopr(len);
834
835 /*
836 * check if the request is larger than number of pages covered
837 * by amp
838 */
839 if (pg_idx + npages > btopr(sptd->spt_amp->size)) {
840 *ppp = NULL;
841 return (ENOTSUP);
842 }
843
844 if (type == L_PAGEUNLOCK) {
845 ASSERT(sptd->spt_ppa != NULL);
846
847 seg_pinactive(seg, NULL, seg->s_base, sptd->spt_amp->size,
848 sptd->spt_ppa, S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
849
850 /*
851 * If someone is blocked while unmapping, we purge
852 * segment page cache and thus reclaim pplist synchronously
853 * without waiting for seg_pasync_thread. This speeds up
854 * unmapping in cases where munmap(2) is called, while
855 * raw async i/o is still in progress or where a thread
856 * exits on data fault in a multithreaded application.
857 */
858 if ((sptd->spt_flags & DISM_PPA_CHANGED) ||
859 (AS_ISUNMAPWAIT(seg->s_as) &&
860 shmd->shm_softlockcnt > 0)) {
861 segspt_purge(seg);
862 }
863 return (0);
864 }
865
866 /* The L_PAGELOCK case ... */
867
868 if (sptd->spt_flags & DISM_PPA_CHANGED) {
869 segspt_purge(seg);
870 /*
871 * for DISM ppa needs to be rebuild since
872 * number of locked pages could be changed
873 */
874 *ppp = NULL;
875 return (ENOTSUP);
876 }
877
878 /*
879 * First try to find pages in segment page cache, without
880 * holding the segment lock.
881 */
882 pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
883 S_WRITE, SEGP_FORCE_WIRED);
884 if (pplist != NULL) {
885 ASSERT(sptd->spt_ppa != NULL);
886 ASSERT(sptd->spt_ppa == pplist);
887 ppa = sptd->spt_ppa;
888 for (an_idx = pg_idx; an_idx < pg_idx + npages; ) {
889 if (ppa[an_idx] == NULL) {
890 seg_pinactive(seg, NULL, seg->s_base,
891 sptd->spt_amp->size, ppa,
892 S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
893 *ppp = NULL;
894 return (ENOTSUP);
895 }
896 if ((szc = ppa[an_idx]->p_szc) != 0) {
897 npgs = page_get_pagecnt(szc);
898 an_idx = P2ROUNDUP(an_idx + 1, npgs);
899 } else {
900 an_idx++;
901 }
902 }
903 /*
904 * Since we cache the entire DISM segment, we want to
905 * set ppp to point to the first slot that corresponds
906 * to the requested addr, i.e. pg_idx.
907 */
908 *ppp = &(sptd->spt_ppa[pg_idx]);
909 return (0);
910 }
911
912 mutex_enter(&sptd->spt_lock);
913 /*
914 * try to find pages in segment page cache with mutex
915 */
916 pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
917 S_WRITE, SEGP_FORCE_WIRED);
918 if (pplist != NULL) {
919 ASSERT(sptd->spt_ppa != NULL);
920 ASSERT(sptd->spt_ppa == pplist);
921 ppa = sptd->spt_ppa;
922 for (an_idx = pg_idx; an_idx < pg_idx + npages; ) {
923 if (ppa[an_idx] == NULL) {
924 mutex_exit(&sptd->spt_lock);
925 seg_pinactive(seg, NULL, seg->s_base,
926 sptd->spt_amp->size, ppa,
927 S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
928 *ppp = NULL;
929 return (ENOTSUP);
930 }
931 if ((szc = ppa[an_idx]->p_szc) != 0) {
932 npgs = page_get_pagecnt(szc);
933 an_idx = P2ROUNDUP(an_idx + 1, npgs);
934 } else {
935 an_idx++;
936 }
937 }
938 /*
939 * Since we cache the entire DISM segment, we want to
940 * set ppp to point to the first slot that corresponds
941 * to the requested addr, i.e. pg_idx.
942 */
943 mutex_exit(&sptd->spt_lock);
944 *ppp = &(sptd->spt_ppa[pg_idx]);
945 return (0);
946 }
947 if (seg_pinsert_check(seg, NULL, seg->s_base, sptd->spt_amp->size,
948 SEGP_FORCE_WIRED) == SEGP_FAIL) {
949 mutex_exit(&sptd->spt_lock);
950 *ppp = NULL;
951 return (ENOTSUP);
952 }
953
954 /*
955 * No need to worry about protections because DISM pages are always rw.
956 */
957 pl = pplist = NULL;
958 amp = sptd->spt_amp;
959
960 /*
961 * Do we need to build the ppa array?
962 */
963 if (sptd->spt_ppa == NULL) {
964 pgcnt_t lpg_cnt = 0;
965
966 pl_built = 1;
967 tot_npages = btopr(sptd->spt_amp->size);
968
969 ASSERT(sptd->spt_pcachecnt == 0);
970 pplist = kmem_zalloc(sizeof (page_t *) * tot_npages, KM_SLEEP);
971 pl = pplist;
972
973 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
974 for (an_idx = 0; an_idx < tot_npages; ) {
975 ap = anon_get_ptr(amp->ahp, an_idx);
976 /*
977 * Cache only mlocked pages. For large pages
978 * if one (constituent) page is mlocked
979 * all pages for that large page
980 * are cached also. This is for quick
981 * lookups of ppa array;
982 */
983 if ((ap != NULL) && (lpg_cnt != 0 ||
984 (sptd->spt_ppa_lckcnt[an_idx] != 0))) {
985
986 swap_xlate(ap, &vp, &off);
987 pp = page_lookup(&vp->v_object, off,
988 SE_SHARED);
989 ASSERT(pp != NULL);
990 if (lpg_cnt == 0) {
991 lpg_cnt++;
992 /*
993 * For a small page, we are done --
994 * lpg_count is reset to 0 below.
995 *
996 * For a large page, we are guaranteed
997 * to find the anon structures of all
998 * constituent pages and a non-zero
999 * lpg_cnt ensures that we don't test
1000 * for mlock for these. We are done
1001 * when lpg_count reaches (npgs + 1).
1002 * If we are not the first constituent
1003 * page, restart at the first one.
1004 */
1005 npgs = page_get_pagecnt(pp->p_szc);
1006 if (!IS_P2ALIGNED(an_idx, npgs)) {
1007 an_idx = P2ALIGN(an_idx, npgs);
1008 page_unlock(pp);
1009 continue;
1010 }
1011 }
1012 if (++lpg_cnt > npgs)
1013 lpg_cnt = 0;
1014
1015 /*
1016 * availrmem is decremented only
1017 * for unlocked pages
1018 */
1019 if (sptd->spt_ppa_lckcnt[an_idx] == 0)
1020 claim_availrmem++;
1021 pplist[an_idx] = pp;
1022 }
1023 an_idx++;
1024 }
1025 ANON_LOCK_EXIT(&amp->a_rwlock);
1026
1027 if (claim_availrmem) {
1028 mutex_enter(&freemem_lock);
1029 if (availrmem < tune.t_minarmem + claim_availrmem) {
1030 mutex_exit(&freemem_lock);
1031 ret = ENOTSUP;
1032 claim_availrmem = 0;
1033 goto insert_fail;
1034 } else {
1035 availrmem -= claim_availrmem;
1036 }
1037 mutex_exit(&freemem_lock);
1038 }
1039
1040 sptd->spt_ppa = pl;
1041 } else {
1042 /*
1043 * We already have a valid ppa[].
1044 */
1045 pl = sptd->spt_ppa;
1046 }
1047
1048 ASSERT(pl != NULL);
1049
1050 ret = seg_pinsert(seg, NULL, seg->s_base, sptd->spt_amp->size,
1051 sptd->spt_amp->size, pl, S_WRITE, SEGP_FORCE_WIRED,
1052 segspt_reclaim);
1053 if (ret == SEGP_FAIL) {
1054 /*
1055 * seg_pinsert failed. We return
1056 * ENOTSUP, so that the as_pagelock() code will
1057 * then try the slower F_SOFTLOCK path.
1058 */
1059 if (pl_built) {
1060 /*
1061 * No one else has referenced the ppa[].
1062 * We created it and we need to destroy it.
1063 */
1064 sptd->spt_ppa = NULL;
1065 }
1066 ret = ENOTSUP;
1067 goto insert_fail;
1068 }
1069
1070 /*
1071 * In either case, we increment softlockcnt on the 'real' segment.
1072 */
1073 sptd->spt_pcachecnt++;
1074 atomic_inc_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
1075
1076 ppa = sptd->spt_ppa;
1077 for (an_idx = pg_idx; an_idx < pg_idx + npages; ) {
1078 if (ppa[an_idx] == NULL) {
1079 mutex_exit(&sptd->spt_lock);
1080 seg_pinactive(seg, NULL, seg->s_base,
1081 sptd->spt_amp->size,
1082 pl, S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
1083 *ppp = NULL;
1084 return (ENOTSUP);
1085 }
1086 if ((szc = ppa[an_idx]->p_szc) != 0) {
1087 npgs = page_get_pagecnt(szc);
1088 an_idx = P2ROUNDUP(an_idx + 1, npgs);
1089 } else {
1090 an_idx++;
1091 }
1092 }
1093 /*
1094 * We can now drop the sptd->spt_lock since the ppa[]
1095 * exists and we have incremented pacachecnt.
1096 */
1097 mutex_exit(&sptd->spt_lock);
1098
1099 /*
1100 * Since we cache the entire segment, we want to
1101 * set ppp to point to the first slot that corresponds
1102 * to the requested addr, i.e. pg_idx.
1103 */
1104 *ppp = &(sptd->spt_ppa[pg_idx]);
1105 return (0);
1106
1107 insert_fail:
1108 /*
1109 * We will only reach this code if we tried and failed.
1110 *
1111 * And we can drop the lock on the dummy seg, once we've failed
1112 * to set up a new ppa[].
1113 */
1114 mutex_exit(&sptd->spt_lock);
1115
1116 if (pl_built) {
1117 if (claim_availrmem) {
1118 mutex_enter(&freemem_lock);
1119 availrmem += claim_availrmem;
1120 mutex_exit(&freemem_lock);
1121 }
1122
1123 /*
1124 * We created pl and we need to destroy it.
1125 */
1126 pplist = pl;
1127 for (an_idx = 0; an_idx < tot_npages; an_idx++) {
1128 if (pplist[an_idx] != NULL)
1129 page_unlock(pplist[an_idx]);
1130 }
1131 kmem_free(pl, sizeof (page_t *) * tot_npages);
1132 }
1133
1134 if (shmd->shm_softlockcnt <= 0) {
1135 if (AS_ISUNMAPWAIT(seg->s_as)) {
1136 mutex_enter(&seg->s_as->a_contents);
1137 if (AS_ISUNMAPWAIT(seg->s_as)) {
1138 AS_CLRUNMAPWAIT(seg->s_as);
1139 cv_broadcast(&seg->s_as->a_cv);
1140 }
1141 mutex_exit(&seg->s_as->a_contents);
1142 }
1143 }
1144 *ppp = NULL;
1145 return (ret);
1146 }
1147
1148
1149
1150 /*
1151 * return locked pages over a given range.
1152 *
1153 * We will cache the entire ISM segment and save the pplist for the
1154 * entire segment in the ppa field of the underlying ISM segment structure.
1155 * Later, during a call to segspt_reclaim() we will use this ppa array
1156 * to page_unlock() all of the pages and then we will free this ppa list.
1157 */
1158 /*ARGSUSED*/
1159 static int
1160 segspt_shmpagelock(struct seg *seg, caddr_t addr, size_t len,
1161 struct page ***ppp, enum lock_type type, enum seg_rw rw)
1162 {
1163 struct shm_data *shmd = (struct shm_data *)seg->s_data;
1164 struct seg *sptseg = shmd->shm_sptseg;
1165 struct spt_data *sptd = sptseg->s_data;
1166 pgcnt_t np, page_index, npages;
1167 caddr_t a, spt_base;
1168 struct page **pplist, **pl, *pp;
1169 struct anon_map *amp;
1170 ulong_t anon_index;
1171 int ret = ENOTSUP;
1172 uint_t pl_built = 0;
1173 struct anon *ap;
1174 struct vnode *vp;
1175 uoff_t off;
1176
1177 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1178 ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
1179
1180
1181 /*
1182 * We want to lock/unlock the entire ISM segment. Therefore,
1183 * we will be using the underlying sptseg and it's base address
1184 * and length for the caching arguments.
1185 */
1186 ASSERT(sptseg);
1187 ASSERT(sptd);
1188
1189 if (sptd->spt_flags & SHM_PAGEABLE) {
1190 return (segspt_dismpagelock(seg, addr, len, ppp, type, rw));
1191 }
1192
1193 page_index = seg_page(seg, addr);
1194 npages = btopr(len);
1195
1196 /*
1197 * check if the request is larger than number of pages covered
1198 * by amp
1199 */
1200 if (page_index + npages > btopr(sptd->spt_amp->size)) {
1201 *ppp = NULL;
1202 return (ENOTSUP);
1203 }
1204
1205 if (type == L_PAGEUNLOCK) {
1206
1207 ASSERT(sptd->spt_ppa != NULL);
1208
1209 seg_pinactive(seg, NULL, seg->s_base, sptd->spt_amp->size,
1210 sptd->spt_ppa, S_WRITE, SEGP_FORCE_WIRED, segspt_reclaim);
1211
1212 /*
1213 * If someone is blocked while unmapping, we purge
1214 * segment page cache and thus reclaim pplist synchronously
1215 * without waiting for seg_pasync_thread. This speeds up
1216 * unmapping in cases where munmap(2) is called, while
1217 * raw async i/o is still in progress or where a thread
1218 * exits on data fault in a multithreaded application.
1219 */
1220 if (AS_ISUNMAPWAIT(seg->s_as) && (shmd->shm_softlockcnt > 0)) {
1221 segspt_purge(seg);
1222 }
1223 return (0);
1224 }
1225
1226 /* The L_PAGELOCK case... */
1227
1228 /*
1229 * First try to find pages in segment page cache, without
1230 * holding the segment lock.
1231 */
1232 pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
1233 S_WRITE, SEGP_FORCE_WIRED);
1234 if (pplist != NULL) {
1235 ASSERT(sptd->spt_ppa == pplist);
1236 ASSERT(sptd->spt_ppa[page_index]);
1237 /*
1238 * Since we cache the entire ISM segment, we want to
1239 * set ppp to point to the first slot that corresponds
1240 * to the requested addr, i.e. page_index.
1241 */
1242 *ppp = &(sptd->spt_ppa[page_index]);
1243 return (0);
1244 }
1245
1246 mutex_enter(&sptd->spt_lock);
1247
1248 /*
1249 * try to find pages in segment page cache
1250 */
1251 pplist = seg_plookup(seg, NULL, seg->s_base, sptd->spt_amp->size,
1252 S_WRITE, SEGP_FORCE_WIRED);
1253 if (pplist != NULL) {
1254 ASSERT(sptd->spt_ppa == pplist);
1255 /*
1256 * Since we cache the entire segment, we want to
1257 * set ppp to point to the first slot that corresponds
1258 * to the requested addr, i.e. page_index.
1259 */
1260 mutex_exit(&sptd->spt_lock);
1261 *ppp = &(sptd->spt_ppa[page_index]);
1262 return (0);
1263 }
1264
1265 if (seg_pinsert_check(seg, NULL, seg->s_base, sptd->spt_amp->size,
1266 SEGP_FORCE_WIRED) == SEGP_FAIL) {
1267 mutex_exit(&sptd->spt_lock);
1268 *ppp = NULL;
1269 return (ENOTSUP);
1270 }
1271
1272 /*
1273 * No need to worry about protections because ISM pages
1274 * are always rw.
1275 */
1276 pl = pplist = NULL;
1277
1278 /*
1279 * Do we need to build the ppa array?
1280 */
1281 if (sptd->spt_ppa == NULL) {
1282 ASSERT(sptd->spt_ppa == pplist);
1283
1284 spt_base = sptseg->s_base;
1285 pl_built = 1;
1286
1287 /*
1288 * availrmem is decremented once during anon_swap_adjust()
1289 * and is incremented during the anon_unresv(), which is
1290 * called from shm_rm_amp() when the segment is destroyed.
1291 */
1292 amp = sptd->spt_amp;
1293 ASSERT(amp != NULL);
1294
1295 /* pcachecnt is protected by sptd->spt_lock */
1296 ASSERT(sptd->spt_pcachecnt == 0);
1297 pplist = kmem_zalloc(sizeof (page_t *)
1298 * btopr(sptd->spt_amp->size), KM_SLEEP);
1299 pl = pplist;
1300
1301 anon_index = seg_page(sptseg, spt_base);
1302
1303 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1304 for (a = spt_base; a < (spt_base + sptd->spt_amp->size);
1305 a += PAGESIZE, anon_index++, pplist++) {
1306 ap = anon_get_ptr(amp->ahp, anon_index);
1307 ASSERT(ap != NULL);
1308 swap_xlate(ap, &vp, &off);
1309 pp = page_lookup(&vp->v_object, off, SE_SHARED);
1310 ASSERT(pp != NULL);
1311 *pplist = pp;
1312 }
1313 ANON_LOCK_EXIT(&amp->a_rwlock);
1314
1315 if (a < (spt_base + sptd->spt_amp->size)) {
1316 ret = ENOTSUP;
1317 goto insert_fail;
1318 }
1319 sptd->spt_ppa = pl;
1320 } else {
1321 /*
1322 * We already have a valid ppa[].
1323 */
1324 pl = sptd->spt_ppa;
1325 }
1326
1327 ASSERT(pl != NULL);
1328
1329 ret = seg_pinsert(seg, NULL, seg->s_base, sptd->spt_amp->size,
1330 sptd->spt_amp->size, pl, S_WRITE, SEGP_FORCE_WIRED,
1331 segspt_reclaim);
1332 if (ret == SEGP_FAIL) {
1333 /*
1334 * seg_pinsert failed. We return
1335 * ENOTSUP, so that the as_pagelock() code will
1336 * then try the slower F_SOFTLOCK path.
1337 */
1338 if (pl_built) {
1339 /*
1340 * No one else has referenced the ppa[].
1341 * We created it and we need to destroy it.
1342 */
1343 sptd->spt_ppa = NULL;
1344 }
1345 ret = ENOTSUP;
1346 goto insert_fail;
1347 }
1348
1349 /*
1350 * In either case, we increment softlockcnt on the 'real' segment.
1351 */
1352 sptd->spt_pcachecnt++;
1353 atomic_inc_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
1354
1355 /*
1356 * We can now drop the sptd->spt_lock since the ppa[]
1357 * exists and we have incremented pacachecnt.
1358 */
1359 mutex_exit(&sptd->spt_lock);
1360
1361 /*
1362 * Since we cache the entire segment, we want to
1363 * set ppp to point to the first slot that corresponds
1364 * to the requested addr, i.e. page_index.
1365 */
1366 *ppp = &(sptd->spt_ppa[page_index]);
1367 return (0);
1368
1369 insert_fail:
1370 /*
1371 * We will only reach this code if we tried and failed.
1372 *
1373 * And we can drop the lock on the dummy seg, once we've failed
1374 * to set up a new ppa[].
1375 */
1376 mutex_exit(&sptd->spt_lock);
1377
1378 if (pl_built) {
1379 /*
1380 * We created pl and we need to destroy it.
1381 */
1382 pplist = pl;
1383 np = (((uintptr_t)(a - spt_base)) >> PAGESHIFT);
1384 while (np) {
1385 page_unlock(*pplist);
1386 np--;
1387 pplist++;
1388 }
1389 kmem_free(pl, sizeof (page_t *) * btopr(sptd->spt_amp->size));
1390 }
1391 if (shmd->shm_softlockcnt <= 0) {
1392 if (AS_ISUNMAPWAIT(seg->s_as)) {
1393 mutex_enter(&seg->s_as->a_contents);
1394 if (AS_ISUNMAPWAIT(seg->s_as)) {
1395 AS_CLRUNMAPWAIT(seg->s_as);
1396 cv_broadcast(&seg->s_as->a_cv);
1397 }
1398 mutex_exit(&seg->s_as->a_contents);
1399 }
1400 }
1401 *ppp = NULL;
1402 return (ret);
1403 }
1404
1405 /*
1406 * purge any cached pages in the I/O page cache
1407 */
1408 static void
1409 segspt_purge(struct seg *seg)
1410 {
1411 seg_ppurge(seg, NULL, SEGP_FORCE_WIRED);
1412 }
1413
1414 static int
1415 segspt_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
1416 enum seg_rw rw, int async)
1417 {
1418 struct seg *seg = (struct seg *)ptag;
1419 struct shm_data *shmd = (struct shm_data *)seg->s_data;
1420 struct seg *sptseg;
1421 struct spt_data *sptd;
1422 pgcnt_t npages, i, free_availrmem = 0;
1423 int done = 0;
1424
1425 sptseg = shmd->shm_sptseg;
1426 sptd = sptseg->s_data;
1427 npages = (len >> PAGESHIFT);
1428 ASSERT(npages);
1429 ASSERT(sptd->spt_pcachecnt != 0);
1430 ASSERT(sptd->spt_ppa == pplist);
1431 ASSERT(npages == btopr(sptd->spt_amp->size));
1432 ASSERT(async || AS_LOCK_HELD(seg->s_as));
1433
1434 /*
1435 * Acquire the lock on the dummy seg and destroy the
1436 * ppa array IF this is the last pcachecnt.
1437 */
1438 mutex_enter(&sptd->spt_lock);
1439 if (--sptd->spt_pcachecnt == 0) {
1440 for (i = 0; i < npages; i++) {
1441 if (pplist[i] == NULL) {
1442 continue;
1443 }
1444 if (rw == S_WRITE) {
1445 hat_setrefmod(pplist[i]);
1446 } else {
1447 hat_setref(pplist[i]);
1448 }
1449 if ((sptd->spt_flags & SHM_PAGEABLE) &&
1450 (sptd->spt_ppa_lckcnt[i] == 0))
1451 free_availrmem++;
1452 page_unlock(pplist[i]);
1453 }
1454 if ((sptd->spt_flags & SHM_PAGEABLE) && free_availrmem) {
1455 mutex_enter(&freemem_lock);
1456 availrmem += free_availrmem;
1457 mutex_exit(&freemem_lock);
1458 }
1459 /*
1460 * Since we want to cach/uncache the entire ISM segment,
1461 * we will track the pplist in a segspt specific field
1462 * ppa, that is initialized at the time we add an entry to
1463 * the cache.
1464 */
1465 ASSERT(sptd->spt_pcachecnt == 0);
1466 kmem_free(pplist, sizeof (page_t *) * npages);
1467 sptd->spt_ppa = NULL;
1468 sptd->spt_flags &= ~DISM_PPA_CHANGED;
1469 sptd->spt_gen++;
1470 cv_broadcast(&sptd->spt_cv);
1471 done = 1;
1472 }
1473 mutex_exit(&sptd->spt_lock);
1474
1475 /*
1476 * If we are pcache async thread or called via seg_ppurge_wiredpp() we
1477 * may not hold AS lock (in this case async argument is not 0). This
1478 * means if softlockcnt drops to 0 after the decrement below address
1479 * space may get freed. We can't allow it since after softlock
1480 * derement to 0 we still need to access as structure for possible
1481 * wakeup of unmap waiters. To prevent the disappearance of as we take
1482 * this segment's shm_segfree_syncmtx. segspt_shmfree() also takes
1483 * this mutex as a barrier to make sure this routine completes before
1484 * segment is freed.
1485 *
1486 * The second complication we have to deal with in async case is a
1487 * possibility of missed wake up of unmap wait thread. When we don't
1488 * hold as lock here we may take a_contents lock before unmap wait
1489 * thread that was first to see softlockcnt was still not 0. As a
1490 * result we'll fail to wake up an unmap wait thread. To avoid this
1491 * race we set nounmapwait flag in as structure if we drop softlockcnt
1492 * to 0 if async is not 0. unmapwait thread
1493 * will not block if this flag is set.
1494 */
1495 if (async)
1496 mutex_enter(&shmd->shm_segfree_syncmtx);
1497
1498 /*
1499 * Now decrement softlockcnt.
1500 */
1501 ASSERT(shmd->shm_softlockcnt > 0);
1502 atomic_dec_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
1503
1504 if (shmd->shm_softlockcnt <= 0) {
1505 if (async || AS_ISUNMAPWAIT(seg->s_as)) {
1506 mutex_enter(&seg->s_as->a_contents);
1507 if (async)
1508 AS_SETNOUNMAPWAIT(seg->s_as);
1509 if (AS_ISUNMAPWAIT(seg->s_as)) {
1510 AS_CLRUNMAPWAIT(seg->s_as);
1511 cv_broadcast(&seg->s_as->a_cv);
1512 }
1513 mutex_exit(&seg->s_as->a_contents);
1514 }
1515 }
1516
1517 if (async)
1518 mutex_exit(&shmd->shm_segfree_syncmtx);
1519
1520 return (done);
1521 }
1522
1523 /*
1524 * Do a F_SOFTUNLOCK call over the range requested.
1525 * The range must have already been F_SOFTLOCK'ed.
1526 *
1527 * The calls to acquire and release the anon map lock mutex were
1528 * removed in order to avoid a deadly embrace during a DR
1529 * memory delete operation. (Eg. DR blocks while waiting for a
1530 * exclusive lock on a page that is being used for kaio; the
1531 * thread that will complete the kaio and call segspt_softunlock
1532 * blocks on the anon map lock; another thread holding the anon
1533 * map lock blocks on another page lock via the segspt_shmfault
1534 * -> page_lookup -> page_lookup_create -> page_lock_es code flow.)
1535 *
1536 * The appropriateness of the removal is based upon the following:
1537 * 1. If we are holding a segment's reader lock and the page is held
1538 * shared, then the corresponding element in anonmap which points to
1539 * anon struct cannot change and there is no need to acquire the
1540 * anonymous map lock.
1541 * 2. Threads in segspt_softunlock have a reader lock on the segment
1542 * and already have the shared page lock, so we are guaranteed that
1543 * the anon map slot cannot change and therefore can call anon_get_ptr()
1544 * without grabbing the anonymous map lock.
1545 * 3. Threads that softlock a shared page break copy-on-write, even if
1546 * its a read. Thus cow faults can be ignored with respect to soft
1547 * unlocking, since the breaking of cow means that the anon slot(s) will
1548 * not be shared.
1549 */
1550 static void
1551 segspt_softunlock(struct seg *seg, caddr_t sptseg_addr,
1552 size_t len, enum seg_rw rw)
1553 {
1554 struct shm_data *shmd = (struct shm_data *)seg->s_data;
1555 struct seg *sptseg;
1556 struct spt_data *sptd;
1557 page_t *pp;
1558 caddr_t adr;
1559 struct vnode *vp;
1560 uoff_t offset;
1561 ulong_t anon_index;
1562 struct anon_map *amp; /* XXX - for locknest */
1563 struct anon *ap = NULL;
1564 pgcnt_t npages;
1565
1566 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1567
1568 sptseg = shmd->shm_sptseg;
1569 sptd = sptseg->s_data;
1570
1571 /*
1572 * Some platforms assume that ISM mappings are HAT_LOAD_LOCK
1573 * and therefore their pages are SE_SHARED locked
1574 * for the entire life of the segment.
1575 */
1576 if ((!hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL)) &&
1577 ((sptd->spt_flags & SHM_PAGEABLE) == 0)) {
1578 goto softlock_decrement;
1579 }
1580
1581 /*
1582 * Any thread is free to do a page_find and
1583 * page_unlock() on the pages within this seg.
1584 *
1585 * We are already holding the as->a_lock on the user's
1586 * real segment, but we need to hold the a_lock on the
1587 * underlying dummy as. This is mostly to satisfy the
1588 * underlying HAT layer.
1589 */
1590 AS_LOCK_ENTER(sptseg->s_as, RW_READER);
1591 hat_unlock(sptseg->s_as->a_hat, sptseg_addr, len);
1592 AS_LOCK_EXIT(sptseg->s_as);
1593
1594 amp = sptd->spt_amp;
1595 ASSERT(amp != NULL);
1596 anon_index = seg_page(sptseg, sptseg_addr);
1597
1598 for (adr = sptseg_addr; adr < sptseg_addr + len; adr += PAGESIZE) {
1599 ap = anon_get_ptr(amp->ahp, anon_index++);
1600 ASSERT(ap != NULL);
1601 swap_xlate(ap, &vp, &offset);
1602
1603 /*
1604 * Use page_find() instead of page_lookup() to
1605 * find the page since we know that it has a
1606 * "shared" lock.
1607 */
1608 pp = page_find(&vp->v_object, offset);
1609 ASSERT(ap == anon_get_ptr(amp->ahp, anon_index - 1));
1610 if (pp == NULL) {
1611 panic("segspt_softunlock: "
1612 "addr %p, ap %p, vp %p, off %llx",
1613 (void *)adr, (void *)ap, (void *)vp, offset);
1614 /*NOTREACHED*/
1615 }
1616
1617 if (rw == S_WRITE) {
1618 hat_setrefmod(pp);
1619 } else if (rw != S_OTHER) {
1620 hat_setref(pp);
1621 }
1622 page_unlock(pp);
1623 }
1624
1625 softlock_decrement:
1626 npages = btopr(len);
1627 ASSERT(shmd->shm_softlockcnt >= npages);
1628 atomic_add_long((ulong_t *)(&(shmd->shm_softlockcnt)), -npages);
1629 if (shmd->shm_softlockcnt == 0) {
1630 /*
1631 * All SOFTLOCKS are gone. Wakeup any waiting
1632 * unmappers so they can try again to unmap.
1633 * Check for waiters first without the mutex
1634 * held so we don't always grab the mutex on
1635 * softunlocks.
1636 */
1637 if (AS_ISUNMAPWAIT(seg->s_as)) {
1638 mutex_enter(&seg->s_as->a_contents);
1639 if (AS_ISUNMAPWAIT(seg->s_as)) {
1640 AS_CLRUNMAPWAIT(seg->s_as);
1641 cv_broadcast(&seg->s_as->a_cv);
1642 }
1643 mutex_exit(&seg->s_as->a_contents);
1644 }
1645 }
1646 }
1647
1648 int
1649 segspt_shmattach(struct seg **segpp, void *argsp)
1650 {
1651 struct seg *seg = *segpp;
1652 struct shm_data *shmd_arg = (struct shm_data *)argsp;
1653 struct shm_data *shmd;
1654 struct anon_map *shm_amp = shmd_arg->shm_amp;
1655 struct spt_data *sptd;
1656 int error = 0;
1657
1658 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1659
1660 shmd = kmem_zalloc((sizeof (*shmd)), KM_NOSLEEP);
1661 if (shmd == NULL)
1662 return (ENOMEM);
1663
1664 shmd->shm_sptas = shmd_arg->shm_sptas;
1665 shmd->shm_amp = shm_amp;
1666 shmd->shm_sptseg = shmd_arg->shm_sptseg;
1667
1668 (void) lgrp_shm_policy_set(LGRP_MEM_POLICY_DEFAULT, shm_amp, 0,
1669 NULL, 0, seg->s_size);
1670
1671 mutex_init(&shmd->shm_segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
1672
1673 seg->s_data = (void *)shmd;
1674 seg->s_ops = &segspt_shmops;
1675 seg->s_szc = shmd->shm_sptseg->s_szc;
1676 sptd = shmd->shm_sptseg->s_data;
1677
1678 if (sptd->spt_flags & SHM_PAGEABLE) {
1679 if ((shmd->shm_vpage = kmem_zalloc(btopr(shm_amp->size),
1680 KM_NOSLEEP)) == NULL) {
1681 seg->s_data = NULL;
1682 kmem_free(shmd, (sizeof (*shmd)));
1683 return (ENOMEM);
1684 }
1685 shmd->shm_lckpgs = 0;
1686 if (hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL)) {
1687 if ((error = hat_share(seg->s_as->a_hat, seg->s_base,
1688 shmd_arg->shm_sptas->a_hat, SEGSPTADDR,
1689 seg->s_size, seg->s_szc)) != 0) {
1690 kmem_free(shmd->shm_vpage,
1691 btopr(shm_amp->size));
1692 }
1693 }
1694 } else {
1695 error = hat_share(seg->s_as->a_hat, seg->s_base,
1696 shmd_arg->shm_sptas->a_hat, SEGSPTADDR,
1697 seg->s_size, seg->s_szc);
1698 }
1699 if (error) {
1700 seg->s_szc = 0;
1701 seg->s_data = NULL;
1702 kmem_free(shmd, (sizeof (*shmd)));
1703 } else {
1704 ANON_LOCK_ENTER(&shm_amp->a_rwlock, RW_WRITER);
1705 shm_amp->refcnt++;
1706 ANON_LOCK_EXIT(&shm_amp->a_rwlock);
1707 }
1708 return (error);
1709 }
1710
1711 int
1712 segspt_shmunmap(struct seg *seg, caddr_t raddr, size_t ssize)
1713 {
1714 struct shm_data *shmd = (struct shm_data *)seg->s_data;
1715 int reclaim = 1;
1716
1717 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1718 retry:
1719 if (shmd->shm_softlockcnt > 0) {
1720 if (reclaim == 1) {
1721 segspt_purge(seg);
1722 reclaim = 0;
1723 goto retry;
1724 }
1725 return (EAGAIN);
1726 }
1727
1728 if (ssize != seg->s_size) {
1729 #ifdef DEBUG
1730 cmn_err(CE_WARN, "Incompatible ssize %lx s_size %lx\n",
1731 ssize, seg->s_size);
1732 #endif
1733 return (EINVAL);
1734 }
1735
1736 (void) segspt_shmlockop(seg, raddr, shmd->shm_amp->size, 0, MC_UNLOCK,
1737 NULL, 0);
1738 hat_unshare(seg->s_as->a_hat, raddr, ssize, seg->s_szc);
1739
1740 seg_free(seg);
1741
1742 return (0);
1743 }
1744
1745 void
1746 segspt_shmfree(struct seg *seg)
1747 {
1748 struct shm_data *shmd = (struct shm_data *)seg->s_data;
1749 struct anon_map *shm_amp = shmd->shm_amp;
1750
1751 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1752
1753 (void) segspt_shmlockop(seg, seg->s_base, shm_amp->size, 0,
1754 MC_UNLOCK, NULL, 0);
1755
1756 /*
1757 * Need to increment refcnt when attaching
1758 * and decrement when detaching because of dup().
1759 */
1760 ANON_LOCK_ENTER(&shm_amp->a_rwlock, RW_WRITER);
1761 shm_amp->refcnt--;
1762 ANON_LOCK_EXIT(&shm_amp->a_rwlock);
1763
1764 if (shmd->shm_vpage) { /* only for DISM */
1765 kmem_free(shmd->shm_vpage, btopr(shm_amp->size));
1766 shmd->shm_vpage = NULL;
1767 }
1768
1769 /*
1770 * Take shm_segfree_syncmtx lock to let segspt_reclaim() finish if it's
1771 * still working with this segment without holding as lock.
1772 */
1773 ASSERT(shmd->shm_softlockcnt == 0);
1774 mutex_enter(&shmd->shm_segfree_syncmtx);
1775 mutex_destroy(&shmd->shm_segfree_syncmtx);
1776
1777 kmem_free(shmd, sizeof (*shmd));
1778 }
1779
1780 /*ARGSUSED*/
1781 int
1782 segspt_shmsetprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
1783 {
1784 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1785
1786 /*
1787 * Shared page table is more than shared mapping.
1788 * Individual process sharing page tables can't change prot
1789 * because there is only one set of page tables.
1790 * This will be allowed after private page table is
1791 * supported.
1792 */
1793 /* need to return correct status error? */
1794 return (0);
1795 }
1796
1797
1798 faultcode_t
1799 segspt_dismfault(struct hat *hat, struct seg *seg, caddr_t addr,
1800 size_t len, enum fault_type type, enum seg_rw rw)
1801 {
1802 struct shm_data *shmd = (struct shm_data *)seg->s_data;
1803 struct seg *sptseg = shmd->shm_sptseg;
1804 struct as *curspt = shmd->shm_sptas;
1805 struct spt_data *sptd = sptseg->s_data;
1806 pgcnt_t npages;
1807 size_t size;
1808 caddr_t segspt_addr, shm_addr;
1809 page_t **ppa;
1810 int i;
1811 ulong_t an_idx = 0;
1812 int err = 0;
1813 int dyn_ism_unmap = hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL);
1814 size_t pgsz;
1815 pgcnt_t pgcnt;
1816 caddr_t a;
1817 pgcnt_t pidx;
1818
1819 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
1820
1821 /*
1822 * Because of the way spt is implemented
1823 * the realsize of the segment does not have to be
1824 * equal to the segment size itself. The segment size is
1825 * often in multiples of a page size larger than PAGESIZE.
1826 * The realsize is rounded up to the nearest PAGESIZE
1827 * based on what the user requested. This is a bit of
1828 * ungliness that is historical but not easily fixed
1829 * without re-designing the higher levels of ISM.
1830 */
1831 ASSERT(addr >= seg->s_base);
1832 if (((addr + len) - seg->s_base) > sptd->spt_realsize)
1833 return (FC_NOMAP);
1834 /*
1835 * For all of the following cases except F_PROT, we need to
1836 * make any necessary adjustments to addr and len
1837 * and get all of the necessary page_t's into an array called ppa[].
1838 *
1839 * The code in shmat() forces base addr and len of ISM segment
1840 * to be aligned to largest page size supported. Therefore,
1841 * we are able to handle F_SOFTLOCK and F_INVAL calls in "large
1842 * pagesize" chunks. We want to make sure that we HAT_LOAD_LOCK
1843 * in large pagesize chunks, or else we will screw up the HAT
1844 * layer by calling hat_memload_array() with differing page sizes
1845 * over a given virtual range.
1846 */
1847 pgsz = page_get_pagesize(sptseg->s_szc);
1848 pgcnt = page_get_pagecnt(sptseg->s_szc);
1849 shm_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);
1850 size = P2ROUNDUP((uintptr_t)(((addr + len) - shm_addr)), pgsz);
1851 npages = btopr(size);
1852
1853 /*
1854 * Now we need to convert from addr in segshm to addr in segspt.
1855 */
1856 an_idx = seg_page(seg, shm_addr);
1857 segspt_addr = sptseg->s_base + ptob(an_idx);
1858
1859 ASSERT((segspt_addr + ptob(npages)) <=
1860 (sptseg->s_base + sptd->spt_realsize));
1861 ASSERT(segspt_addr < (sptseg->s_base + sptseg->s_size));
1862
1863 switch (type) {
1864
1865 case F_SOFTLOCK:
1866
1867 atomic_add_long((ulong_t *)(&(shmd->shm_softlockcnt)), npages);
1868 /*
1869 * Fall through to the F_INVAL case to load up the hat layer
1870 * entries with the HAT_LOAD_LOCK flag.
1871 */
1872 /* FALLTHRU */
1873 case F_INVAL:
1874
1875 if ((rw == S_EXEC) && !(sptd->spt_prot & PROT_EXEC))
1876 return (FC_NOMAP);
1877
1878 ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
1879
1880 err = spt_anon_getpages(sptseg, segspt_addr, size, ppa);
1881 if (err != 0) {
1882 if (type == F_SOFTLOCK) {
1883 atomic_add_long((ulong_t *)(
1884 &(shmd->shm_softlockcnt)), -npages);
1885 }
1886 goto dism_err;
1887 }
1888 AS_LOCK_ENTER(sptseg->s_as, RW_READER);
1889 a = segspt_addr;
1890 pidx = 0;
1891 if (type == F_SOFTLOCK) {
1892
1893 /*
1894 * Load up the translation keeping it
1895 * locked and don't unlock the page.
1896 */
1897 for (; pidx < npages; a += pgsz, pidx += pgcnt) {
1898 hat_memload_array(sptseg->s_as->a_hat,
1899 a, pgsz, &ppa[pidx], sptd->spt_prot,
1900 HAT_LOAD_LOCK | HAT_LOAD_SHARE);
1901 }
1902 } else {
1903 /*
1904 * Migrate pages marked for migration
1905 */
1906 if (lgrp_optimizations())
1907 page_migrate(seg, shm_addr, ppa, npages);
1908
1909 for (; pidx < npages; a += pgsz, pidx += pgcnt) {
1910 hat_memload_array(sptseg->s_as->a_hat,
1911 a, pgsz, &ppa[pidx],
1912 sptd->spt_prot,
1913 HAT_LOAD_SHARE);
1914 }
1915
1916 /*
1917 * And now drop the SE_SHARED lock(s).
1918 */
1919 if (dyn_ism_unmap) {
1920 for (i = 0; i < npages; i++) {
1921 page_unlock(ppa[i]);
1922 }
1923 }
1924 }
1925
1926 if (!dyn_ism_unmap) {
1927 if (hat_share(seg->s_as->a_hat, shm_addr,
1928 curspt->a_hat, segspt_addr, ptob(npages),
1929 seg->s_szc) != 0) {
1930 panic("hat_share err in DISM fault");
1931 /* NOTREACHED */
1932 }
1933 if (type == F_INVAL) {
1934 for (i = 0; i < npages; i++) {
1935 page_unlock(ppa[i]);
1936 }
1937 }
1938 }
1939 AS_LOCK_EXIT(sptseg->s_as);
1940 dism_err:
1941 kmem_free(ppa, npages * sizeof (page_t *));
1942 return (err);
1943
1944 case F_SOFTUNLOCK:
1945
1946 /*
1947 * This is a bit ugly, we pass in the real seg pointer,
1948 * but the segspt_addr is the virtual address within the
1949 * dummy seg.
1950 */
1951 segspt_softunlock(seg, segspt_addr, size, rw);
1952 return (0);
1953
1954 case F_PROT:
1955
1956 /*
1957 * This takes care of the unusual case where a user
1958 * allocates a stack in shared memory and a register
1959 * window overflow is written to that stack page before
1960 * it is otherwise modified.
1961 *
1962 * We can get away with this because ISM segments are
1963 * always rw. Other than this unusual case, there
1964 * should be no instances of protection violations.
1965 */
1966 return (0);
1967
1968 default:
1969 #ifdef DEBUG
1970 panic("segspt_dismfault default type?");
1971 #else
1972 return (FC_NOMAP);
1973 #endif
1974 }
1975 }
1976
1977
1978 faultcode_t
1979 segspt_shmfault(struct hat *hat, struct seg *seg, caddr_t addr,
1980 size_t len, enum fault_type type, enum seg_rw rw)
1981 {
1982 struct shm_data *shmd = (struct shm_data *)seg->s_data;
1983 struct seg *sptseg = shmd->shm_sptseg;
1984 struct as *curspt = shmd->shm_sptas;
1985 struct spt_data *sptd = sptseg->s_data;
1986 pgcnt_t npages;
1987 size_t size;
1988 caddr_t sptseg_addr, shm_addr;
1989 page_t *pp, **ppa;
1990 int i;
1991 uoff_t offset;
1992 ulong_t anon_index = 0;
1993 struct vnode *vp;
1994 struct anon_map *amp; /* XXX - for locknest */
1995 struct anon *ap = NULL;
1996 size_t pgsz;
1997 pgcnt_t pgcnt;
1998 caddr_t a;
1999 pgcnt_t pidx;
2000 size_t sz;
2001
2002
2003 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2004
2005 if (sptd->spt_flags & SHM_PAGEABLE) {
2006 return (segspt_dismfault(hat, seg, addr, len, type, rw));
2007 }
2008
2009 /*
2010 * Because of the way spt is implemented
2011 * the realsize of the segment does not have to be
2012 * equal to the segment size itself. The segment size is
2013 * often in multiples of a page size larger than PAGESIZE.
2014 * The realsize is rounded up to the nearest PAGESIZE
2015 * based on what the user requested. This is a bit of
2016 * ungliness that is historical but not easily fixed
2017 * without re-designing the higher levels of ISM.
2018 */
2019 ASSERT(addr >= seg->s_base);
2020 if (((addr + len) - seg->s_base) > sptd->spt_realsize)
2021 return (FC_NOMAP);
2022 /*
2023 * For all of the following cases except F_PROT, we need to
2024 * make any necessary adjustments to addr and len
2025 * and get all of the necessary page_t's into an array called ppa[].
2026 *
2027 * The code in shmat() forces base addr and len of ISM segment
2028 * to be aligned to largest page size supported. Therefore,
2029 * we are able to handle F_SOFTLOCK and F_INVAL calls in "large
2030 * pagesize" chunks. We want to make sure that we HAT_LOAD_LOCK
2031 * in large pagesize chunks, or else we will screw up the HAT
2032 * layer by calling hat_memload_array() with differing page sizes
2033 * over a given virtual range.
2034 */
2035 pgsz = page_get_pagesize(sptseg->s_szc);
2036 pgcnt = page_get_pagecnt(sptseg->s_szc);
2037 shm_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);
2038 size = P2ROUNDUP((uintptr_t)(((addr + len) - shm_addr)), pgsz);
2039 npages = btopr(size);
2040
2041 /*
2042 * Now we need to convert from addr in segshm to addr in segspt.
2043 */
2044 anon_index = seg_page(seg, shm_addr);
2045 sptseg_addr = sptseg->s_base + ptob(anon_index);
2046
2047 /*
2048 * And now we may have to adjust npages downward if we have
2049 * exceeded the realsize of the segment or initial anon
2050 * allocations.
2051 */
2052 if ((sptseg_addr + ptob(npages)) >
2053 (sptseg->s_base + sptd->spt_realsize))
2054 size = (sptseg->s_base + sptd->spt_realsize) - sptseg_addr;
2055
2056 npages = btopr(size);
2057
2058 ASSERT(sptseg_addr < (sptseg->s_base + sptseg->s_size));
2059 ASSERT((sptd->spt_flags & SHM_PAGEABLE) == 0);
2060
2061 switch (type) {
2062
2063 case F_SOFTLOCK:
2064
2065 /*
2066 * availrmem is decremented once during anon_swap_adjust()
2067 * and is incremented during the anon_unresv(), which is
2068 * called from shm_rm_amp() when the segment is destroyed.
2069 */
2070 atomic_add_long((ulong_t *)(&(shmd->shm_softlockcnt)), npages);
2071 /*
2072 * Some platforms assume that ISM pages are SE_SHARED
2073 * locked for the entire life of the segment.
2074 */
2075 if (!hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL))
2076 return (0);
2077 /*
2078 * Fall through to the F_INVAL case to load up the hat layer
2079 * entries with the HAT_LOAD_LOCK flag.
2080 */
2081
2082 /* FALLTHRU */
2083 case F_INVAL:
2084
2085 if ((rw == S_EXEC) && !(sptd->spt_prot & PROT_EXEC))
2086 return (FC_NOMAP);
2087
2088 /*
2089 * Some platforms that do NOT support DYNAMIC_ISM_UNMAP
2090 * may still rely on this call to hat_share(). That
2091 * would imply that those hat's can fault on a
2092 * HAT_LOAD_LOCK translation, which would seem
2093 * contradictory.
2094 */
2095 if (!hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL)) {
2096 if (hat_share(seg->s_as->a_hat, seg->s_base,
2097 curspt->a_hat, sptseg->s_base,
2098 sptseg->s_size, sptseg->s_szc) != 0) {
2099 panic("hat_share error in ISM fault");
2100 /*NOTREACHED*/
2101 }
2102 return (0);
2103 }
2104 ppa = kmem_zalloc(sizeof (page_t *) * npages, KM_SLEEP);
2105
2106 /*
2107 * I see no need to lock the real seg,
2108 * here, because all of our work will be on the underlying
2109 * dummy seg.
2110 *
2111 * sptseg_addr and npages now account for large pages.
2112 */
2113 amp = sptd->spt_amp;
2114 ASSERT(amp != NULL);
2115 anon_index = seg_page(sptseg, sptseg_addr);
2116
2117 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2118 for (i = 0; i < npages; i++) {
2119 ap = anon_get_ptr(amp->ahp, anon_index++);
2120 ASSERT(ap != NULL);
2121 swap_xlate(ap, &vp, &offset);
2122 pp = page_lookup(&vp->v_object, offset, SE_SHARED);
2123 ASSERT(pp != NULL);
2124 ppa[i] = pp;
2125 }
2126 ANON_LOCK_EXIT(&amp->a_rwlock);
2127 ASSERT(i == npages);
2128
2129 /*
2130 * We are already holding the as->a_lock on the user's
2131 * real segment, but we need to hold the a_lock on the
2132 * underlying dummy as. This is mostly to satisfy the
2133 * underlying HAT layer.
2134 */
2135 AS_LOCK_ENTER(sptseg->s_as, RW_READER);
2136 a = sptseg_addr;
2137 pidx = 0;
2138 if (type == F_SOFTLOCK) {
2139 /*
2140 * Load up the translation keeping it
2141 * locked and don't unlock the page.
2142 */
2143 for (; pidx < npages; a += pgsz, pidx += pgcnt) {
2144 sz = MIN(pgsz, ptob(npages - pidx));
2145 hat_memload_array(sptseg->s_as->a_hat, a,
2146 sz, &ppa[pidx], sptd->spt_prot,
2147 HAT_LOAD_LOCK | HAT_LOAD_SHARE);
2148 }
2149 } else {
2150 /*
2151 * Migrate pages marked for migration.
2152 */
2153 if (lgrp_optimizations())
2154 page_migrate(seg, shm_addr, ppa, npages);
2155
2156 for (; pidx < npages; a += pgsz, pidx += pgcnt) {
2157 sz = MIN(pgsz, ptob(npages - pidx));
2158 hat_memload_array(sptseg->s_as->a_hat,
2159 a, sz, &ppa[pidx],
2160 sptd->spt_prot, HAT_LOAD_SHARE);
2161 }
2162
2163 /*
2164 * And now drop the SE_SHARED lock(s).
2165 */
2166 for (i = 0; i < npages; i++)
2167 page_unlock(ppa[i]);
2168 }
2169 AS_LOCK_EXIT(sptseg->s_as);
2170
2171 kmem_free(ppa, sizeof (page_t *) * npages);
2172 return (0);
2173 case F_SOFTUNLOCK:
2174
2175 /*
2176 * This is a bit ugly, we pass in the real seg pointer,
2177 * but the sptseg_addr is the virtual address within the
2178 * dummy seg.
2179 */
2180 segspt_softunlock(seg, sptseg_addr, ptob(npages), rw);
2181 return (0);
2182
2183 case F_PROT:
2184
2185 /*
2186 * This takes care of the unusual case where a user
2187 * allocates a stack in shared memory and a register
2188 * window overflow is written to that stack page before
2189 * it is otherwise modified.
2190 *
2191 * We can get away with this because ISM segments are
2192 * always rw. Other than this unusual case, there
2193 * should be no instances of protection violations.
2194 */
2195 return (0);
2196
2197 default:
2198 #ifdef DEBUG
2199 cmn_err(CE_WARN, "segspt_shmfault default type?");
2200 #endif
2201 return (FC_NOMAP);
2202 }
2203 }
2204
2205 /*ARGSUSED*/
2206 static faultcode_t
2207 segspt_shmfaulta(struct seg *seg, caddr_t addr)
2208 {
2209 return (0);
2210 }
2211
2212 /*ARGSUSED*/
2213 static int
2214 segspt_shmkluster(struct seg *seg, caddr_t addr, ssize_t delta)
2215 {
2216 return (0);
2217 }
2218
2219 /*
2220 * duplicate the shared page tables
2221 */
2222 int
2223 segspt_shmdup(struct seg *seg, struct seg *newseg)
2224 {
2225 struct shm_data *shmd = (struct shm_data *)seg->s_data;
2226 struct anon_map *amp = shmd->shm_amp;
2227 struct shm_data *shmd_new;
2228 struct seg *spt_seg = shmd->shm_sptseg;
2229 struct spt_data *sptd = spt_seg->s_data;
2230 int error = 0;
2231
2232 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
2233
2234 shmd_new = kmem_zalloc((sizeof (*shmd_new)), KM_SLEEP);
2235 newseg->s_data = (void *)shmd_new;
2236 shmd_new->shm_sptas = shmd->shm_sptas;
2237 shmd_new->shm_amp = amp;
2238 shmd_new->shm_sptseg = shmd->shm_sptseg;
2239 newseg->s_ops = &segspt_shmops;
2240 newseg->s_szc = seg->s_szc;
2241 ASSERT(seg->s_szc == shmd->shm_sptseg->s_szc);
2242
2243 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2244 amp->refcnt++;
2245 ANON_LOCK_EXIT(&amp->a_rwlock);
2246
2247 if (sptd->spt_flags & SHM_PAGEABLE) {
2248 shmd_new->shm_vpage = kmem_zalloc(btopr(amp->size), KM_SLEEP);
2249 shmd_new->shm_lckpgs = 0;
2250 if (hat_supported(HAT_DYNAMIC_ISM_UNMAP, NULL)) {
2251 if ((error = hat_share(newseg->s_as->a_hat,
2252 newseg->s_base, shmd->shm_sptas->a_hat, SEGSPTADDR,
2253 seg->s_size, seg->s_szc)) != 0) {
2254 kmem_free(shmd_new->shm_vpage,
2255 btopr(amp->size));
2256 }
2257 }
2258 return (error);
2259 } else {
2260 return (hat_share(newseg->s_as->a_hat, newseg->s_base,
2261 shmd->shm_sptas->a_hat, SEGSPTADDR, seg->s_size,
2262 seg->s_szc));
2263
2264 }
2265 }
2266
2267 /*ARGSUSED*/
2268 int
2269 segspt_shmcheckprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot)
2270 {
2271 struct shm_data *shmd = (struct shm_data *)seg->s_data;
2272 struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2273
2274 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2275
2276 /*
2277 * ISM segment is always rw.
2278 */
2279 return (((sptd->spt_prot & prot) != prot) ? EACCES : 0);
2280 }
2281
2282 /*
2283 * Return an array of locked large pages, for empty slots allocate
2284 * private zero-filled anon pages.
2285 */
2286 static int
2287 spt_anon_getpages(
2288 struct seg *sptseg,
2289 caddr_t sptaddr,
2290 size_t len,
2291 page_t *ppa[])
2292 {
2293 struct spt_data *sptd = sptseg->s_data;
2294 struct anon_map *amp = sptd->spt_amp;
2295 enum seg_rw rw = sptd->spt_prot;
2296 uint_t szc = sptseg->s_szc;
2297 size_t pg_sz, share_sz = page_get_pagesize(szc);
2298 pgcnt_t lp_npgs;
2299 caddr_t lp_addr, e_sptaddr;
2300 uint_t vpprot, ppa_szc = 0;
2301 struct vpage *vpage = NULL;
2302 ulong_t j, ppa_idx;
2303 int err, ierr = 0;
2304 pgcnt_t an_idx;
2305 anon_sync_obj_t cookie;
2306 int anon_locked = 0;
2307 pgcnt_t amp_pgs;
2308
2309
2310 ASSERT(IS_P2ALIGNED(sptaddr, share_sz) && IS_P2ALIGNED(len, share_sz));
2311 ASSERT(len != 0);
2312
2313 pg_sz = share_sz;
2314 lp_npgs = btop(pg_sz);
2315 lp_addr = sptaddr;
2316 e_sptaddr = sptaddr + len;
2317 an_idx = seg_page(sptseg, sptaddr);
2318 ppa_idx = 0;
2319
2320 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2321
2322 amp_pgs = page_get_pagecnt(amp->a_szc);
2323
2324 /*CONSTCOND*/
2325 while (1) {
2326 for (; lp_addr < e_sptaddr;
2327 an_idx += lp_npgs, lp_addr += pg_sz, ppa_idx += lp_npgs) {
2328
2329 /*
2330 * If we're currently locked, and we get to a new
2331 * page, unlock our current anon chunk.
2332 */
2333 if (anon_locked && P2PHASE(an_idx, amp_pgs) == 0) {
2334 anon_array_exit(&cookie);
2335 anon_locked = 0;
2336 }
2337 if (!anon_locked) {
2338 anon_array_enter(amp, an_idx, &cookie);
2339 anon_locked = 1;
2340 }
2341 ppa_szc = (uint_t)-1;
2342 ierr = anon_map_getpages(amp, an_idx, szc, sptseg,
2343 lp_addr, sptd->spt_prot, &vpprot, &ppa[ppa_idx],
2344 &ppa_szc, vpage, rw, 0, segvn_anypgsz, 0, kcred);
2345
2346 if (ierr != 0) {
2347 if (ierr > 0) {
2348 err = FC_MAKE_ERR(ierr);
2349 goto lpgs_err;
2350 }
2351 break;
2352 }
2353 }
2354 if (lp_addr == e_sptaddr) {
2355 break;
2356 }
2357 ASSERT(lp_addr < e_sptaddr);
2358
2359 /*
2360 * ierr == -1 means we failed to allocate a large page.
2361 * so do a size down operation.
2362 *
2363 * ierr == -2 means some other process that privately shares
2364 * pages with this process has allocated a larger page and we
2365 * need to retry with larger pages. So do a size up
2366 * operation. This relies on the fact that large pages are
2367 * never partially shared i.e. if we share any constituent
2368 * page of a large page with another process we must share the
2369 * entire large page. Note this cannot happen for SOFTLOCK
2370 * case, unless current address (lpaddr) is at the beginning
2371 * of the next page size boundary because the other process
2372 * couldn't have relocated locked pages.
2373 */
2374 ASSERT(ierr == -1 || ierr == -2);
2375 if (segvn_anypgsz) {
2376 ASSERT(ierr == -2 || szc != 0);
2377 ASSERT(ierr == -1 || szc < sptseg->s_szc);
2378 szc = (ierr == -1) ? szc - 1 : szc + 1;
2379 } else {
2380 /*
2381 * For faults and segvn_anypgsz == 0
2382 * we need to be careful not to loop forever
2383 * if existing page is found with szc other
2384 * than 0 or seg->s_szc. This could be due
2385 * to page relocations on behalf of DR or
2386 * more likely large page creation. For this
2387 * case simply re-size to existing page's szc
2388 * if returned by anon_map_getpages().
2389 */
2390 if (ppa_szc == (uint_t)-1) {
2391 szc = (ierr == -1) ? 0 : sptseg->s_szc;
2392 } else {
2393 ASSERT(ppa_szc <= sptseg->s_szc);
2394 ASSERT(ierr == -2 || ppa_szc < szc);
2395 ASSERT(ierr == -1 || ppa_szc > szc);
2396 szc = ppa_szc;
2397 }
2398 }
2399 pg_sz = page_get_pagesize(szc);
2400 lp_npgs = btop(pg_sz);
2401 ASSERT(IS_P2ALIGNED(lp_addr, pg_sz));
2402 }
2403 if (anon_locked) {
2404 anon_array_exit(&cookie);
2405 }
2406 ANON_LOCK_EXIT(&amp->a_rwlock);
2407 return (0);
2408
2409 lpgs_err:
2410 if (anon_locked) {
2411 anon_array_exit(&cookie);
2412 }
2413 ANON_LOCK_EXIT(&amp->a_rwlock);
2414 for (j = 0; j < ppa_idx; j++)
2415 page_unlock(ppa[j]);
2416 return (err);
2417 }
2418
2419 /*
2420 * count the number of bytes in a set of spt pages that are currently not
2421 * locked
2422 */
2423 static rctl_qty_t
2424 spt_unlockedbytes(pgcnt_t npages, page_t **ppa)
2425 {
2426 ulong_t i;
2427 rctl_qty_t unlocked = 0;
2428
2429 for (i = 0; i < npages; i++) {
2430 if (ppa[i]->p_lckcnt == 0)
2431 unlocked += PAGESIZE;
2432 }
2433 return (unlocked);
2434 }
2435
2436 extern u_longlong_t randtick(void);
2437 /* number of locks to reserve/skip by spt_lockpages() and spt_unlockpages() */
2438 #define NLCK (NCPU_P2)
2439 /* Random number with a range [0, n-1], n must be power of two */
2440 #define RAND_P2(n) \
2441 ((((long)curthread >> PTR24_LSB) ^ (long)randtick()) & ((n) - 1))
2442
2443 int
2444 spt_lockpages(struct seg *seg, pgcnt_t anon_index, pgcnt_t npages,
2445 page_t **ppa, ulong_t *lockmap, size_t pos,
2446 rctl_qty_t *locked)
2447 {
2448 struct shm_data *shmd = seg->s_data;
2449 struct spt_data *sptd = shmd->shm_sptseg->s_data;
2450 ulong_t i;
2451 int kernel;
2452 pgcnt_t nlck = 0;
2453 int rv = 0;
2454 int use_reserved = 1;
2455
2456 /* return the number of bytes actually locked */
2457 *locked = 0;
2458
2459 /*
2460 * To avoid contention on freemem_lock, availrmem and pages_locked
2461 * global counters are updated only every nlck locked pages instead of
2462 * every time. Reserve nlck locks up front and deduct from this
2463 * reservation for each page that requires a lock. When the reservation
2464 * is consumed, reserve again. nlck is randomized, so the competing
2465 * threads do not fall into a cyclic lock contention pattern. When
2466 * memory is low, the lock ahead is disabled, and instead page_pp_lock()
2467 * is used to lock pages.
2468 */
2469 for (i = 0; i < npages; anon_index++, pos++, i++) {
2470 if (nlck == 0 && use_reserved == 1) {
2471 nlck = NLCK + RAND_P2(NLCK);
2472 /* if fewer loops left, decrease nlck */
2473 nlck = MIN(nlck, npages - i);
2474 /*
2475 * Reserve nlck locks up front and deduct from this
2476 * reservation for each page that requires a lock. When
2477 * the reservation is consumed, reserve again.
2478 */
2479 mutex_enter(&freemem_lock);
2480 if ((availrmem - nlck) < pages_pp_maximum) {
2481 /* Do not do advance memory reserves */
2482 use_reserved = 0;
2483 } else {
2484 availrmem -= nlck;
2485 pages_locked += nlck;
2486 }
2487 mutex_exit(&freemem_lock);
2488 }
2489 if (!(shmd->shm_vpage[anon_index] & DISM_PG_LOCKED)) {
2490 if (sptd->spt_ppa_lckcnt[anon_index] <
2491 (ushort_t)DISM_LOCK_MAX) {
2492 if (++sptd->spt_ppa_lckcnt[anon_index] ==
2493 (ushort_t)DISM_LOCK_MAX) {
2494 cmn_err(CE_WARN,
2495 "DISM page lock limit "
2496 "reached on DISM offset 0x%lx\n",
2497 anon_index << PAGESHIFT);
2498 }
2499 kernel = (sptd->spt_ppa &&
2500 sptd->spt_ppa[anon_index]);
2501 if (!page_pp_lock(ppa[i], 0, kernel ||
2502 use_reserved)) {
2503 sptd->spt_ppa_lckcnt[anon_index]--;
2504 rv = EAGAIN;
2505 break;
2506 }
2507 /* if this is a newly locked page, count it */
2508 if (ppa[i]->p_lckcnt == 1) {
2509 if (kernel == 0 && use_reserved == 1)
2510 nlck--;
2511 *locked += PAGESIZE;
2512 }
2513 shmd->shm_lckpgs++;
2514 shmd->shm_vpage[anon_index] |= DISM_PG_LOCKED;
2515 if (lockmap != NULL)
2516 BT_SET(lockmap, pos);
2517 }
2518 }
2519 }
2520 /* Return unused lock reservation */
2521 if (nlck != 0 && use_reserved == 1) {
2522 mutex_enter(&freemem_lock);
2523 availrmem += nlck;
2524 pages_locked -= nlck;
2525 mutex_exit(&freemem_lock);
2526 }
2527
2528 return (rv);
2529 }
2530
2531 int
2532 spt_unlockpages(struct seg *seg, pgcnt_t anon_index, pgcnt_t npages,
2533 rctl_qty_t *unlocked)
2534 {
2535 struct shm_data *shmd = seg->s_data;
2536 struct spt_data *sptd = shmd->shm_sptseg->s_data;
2537 struct anon_map *amp = sptd->spt_amp;
2538 struct anon *ap;
2539 struct vnode *vp;
2540 uoff_t off;
2541 struct page *pp;
2542 int kernel;
2543 anon_sync_obj_t cookie;
2544 ulong_t i;
2545 pgcnt_t nlck = 0;
2546 pgcnt_t nlck_limit = NLCK;
2547
2548 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2549 for (i = 0; i < npages; i++, anon_index++) {
2550 if (shmd->shm_vpage[anon_index] & DISM_PG_LOCKED) {
2551 anon_array_enter(amp, anon_index, &cookie);
2552 ap = anon_get_ptr(amp->ahp, anon_index);
2553 ASSERT(ap);
2554
2555 swap_xlate(ap, &vp, &off);
2556 anon_array_exit(&cookie);
2557 pp = page_lookup(&vp->v_object, off, SE_SHARED);
2558 ASSERT(pp);
2559 /*
2560 * availrmem is decremented only for pages which are not
2561 * in seg pcache, for pages in seg pcache availrmem was
2562 * decremented in _dismpagelock()
2563 */
2564 kernel = (sptd->spt_ppa && sptd->spt_ppa[anon_index]);
2565 ASSERT(pp->p_lckcnt > 0);
2566
2567 /*
2568 * lock page but do not change availrmem, we do it
2569 * ourselves every nlck loops.
2570 */
2571 page_pp_unlock(pp, 0, 1);
2572 if (pp->p_lckcnt == 0) {
2573 if (kernel == 0)
2574 nlck++;
2575 *unlocked += PAGESIZE;
2576 }
2577 page_unlock(pp);
2578 shmd->shm_vpage[anon_index] &= ~DISM_PG_LOCKED;
2579 sptd->spt_ppa_lckcnt[anon_index]--;
2580 shmd->shm_lckpgs--;
2581 }
2582
2583 /*
2584 * To reduce freemem_lock contention, do not update availrmem
2585 * until at least NLCK pages have been unlocked.
2586 * 1. No need to update if nlck is zero
2587 * 2. Always update if the last iteration
2588 */
2589 if (nlck > 0 && (nlck == nlck_limit || i == npages - 1)) {
2590 mutex_enter(&freemem_lock);
2591 availrmem += nlck;
2592 pages_locked -= nlck;
2593 mutex_exit(&freemem_lock);
2594 nlck = 0;
2595 nlck_limit = NLCK + RAND_P2(NLCK);
2596 }
2597 }
2598 ANON_LOCK_EXIT(&amp->a_rwlock);
2599
2600 return (0);
2601 }
2602
2603 /*ARGSUSED*/
2604 static int
2605 segspt_shmlockop(struct seg *seg, caddr_t addr, size_t len,
2606 int attr, int op, ulong_t *lockmap, size_t pos)
2607 {
2608 struct shm_data *shmd = seg->s_data;
2609 struct seg *sptseg = shmd->shm_sptseg;
2610 struct spt_data *sptd = sptseg->s_data;
2611 struct kshmid *sp = sptd->spt_amp->a_sp;
2612 pgcnt_t npages, a_npages;
2613 page_t **ppa;
2614 pgcnt_t an_idx, a_an_idx, ppa_idx;
2615 caddr_t spt_addr, a_addr; /* spt and aligned address */
2616 size_t a_len; /* aligned len */
2617 size_t share_sz;
2618 ulong_t i;
2619 int sts = 0;
2620 rctl_qty_t unlocked = 0;
2621 rctl_qty_t locked = 0;
2622 struct proc *p = curproc;
2623 kproject_t *proj;
2624
2625 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2626 ASSERT(sp != NULL);
2627
2628 if ((sptd->spt_flags & SHM_PAGEABLE) == 0) {
2629 return (0);
2630 }
2631
2632 addr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
2633 an_idx = seg_page(seg, addr);
2634 npages = btopr(len);
2635
2636 if (an_idx + npages > btopr(shmd->shm_amp->size)) {
2637 return (ENOMEM);
2638 }
2639
2640 /*
2641 * A shm's project never changes, so no lock needed.
2642 * The shm has a hold on the project, so it will not go away.
2643 * Since we have a mapping to shm within this zone, we know
2644 * that the zone will not go away.
2645 */
2646 proj = sp->shm_perm.ipc_proj;
2647
2648 if (op == MC_LOCK) {
2649
2650 /*
2651 * Need to align addr and size request if they are not
2652 * aligned so we can always allocate large page(s) however
2653 * we only lock what was requested in initial request.
2654 */
2655 share_sz = page_get_pagesize(sptseg->s_szc);
2656 a_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), share_sz);
2657 a_len = P2ROUNDUP((uintptr_t)(((addr + len) - a_addr)),
2658 share_sz);
2659 a_npages = btop(a_len);
2660 a_an_idx = seg_page(seg, a_addr);
2661 spt_addr = sptseg->s_base + ptob(a_an_idx);
2662 ppa_idx = an_idx - a_an_idx;
2663
2664 if ((ppa = kmem_zalloc(((sizeof (page_t *)) * a_npages),
2665 KM_NOSLEEP)) == NULL) {
2666 return (ENOMEM);
2667 }
2668
2669 /*
2670 * Don't cache any new pages for IO and
2671 * flush any cached pages.
2672 */
2673 mutex_enter(&sptd->spt_lock);
2674 if (sptd->spt_ppa != NULL)
2675 sptd->spt_flags |= DISM_PPA_CHANGED;
2676
2677 sts = spt_anon_getpages(sptseg, spt_addr, a_len, ppa);
2678 if (sts != 0) {
2679 mutex_exit(&sptd->spt_lock);
2680 kmem_free(ppa, ((sizeof (page_t *)) * a_npages));
2681 return (sts);
2682 }
2683
2684 mutex_enter(&sp->shm_mlock);
2685 /* enforce locked memory rctl */
2686 unlocked = spt_unlockedbytes(npages, &ppa[ppa_idx]);
2687
2688 mutex_enter(&p->p_lock);
2689 if (rctl_incr_locked_mem(p, proj, unlocked, 0)) {
2690 mutex_exit(&p->p_lock);
2691 sts = EAGAIN;
2692 } else {
2693 mutex_exit(&p->p_lock);
2694 sts = spt_lockpages(seg, an_idx, npages,
2695 &ppa[ppa_idx], lockmap, pos, &locked);
2696
2697 /*
2698 * correct locked count if not all pages could be
2699 * locked
2700 */
2701 if ((unlocked - locked) > 0) {
2702 rctl_decr_locked_mem(NULL, proj,
2703 (unlocked - locked), 0);
2704 }
2705 }
2706 /*
2707 * unlock pages
2708 */
2709 for (i = 0; i < a_npages; i++)
2710 page_unlock(ppa[i]);
2711 if (sptd->spt_ppa != NULL)
2712 sptd->spt_flags |= DISM_PPA_CHANGED;
2713 mutex_exit(&sp->shm_mlock);
2714 mutex_exit(&sptd->spt_lock);
2715
2716 kmem_free(ppa, ((sizeof (page_t *)) * a_npages));
2717
2718 } else if (op == MC_UNLOCK) { /* unlock */
2719 page_t **ppa;
2720
2721 mutex_enter(&sptd->spt_lock);
2722 if (shmd->shm_lckpgs == 0) {
2723 mutex_exit(&sptd->spt_lock);
2724 return (0);
2725 }
2726 /*
2727 * Don't cache new IO pages.
2728 */
2729 if (sptd->spt_ppa != NULL)
2730 sptd->spt_flags |= DISM_PPA_CHANGED;
2731
2732 mutex_enter(&sp->shm_mlock);
2733 sts = spt_unlockpages(seg, an_idx, npages, &unlocked);
2734 if ((ppa = sptd->spt_ppa) != NULL)
2735 sptd->spt_flags |= DISM_PPA_CHANGED;
2736 mutex_exit(&sptd->spt_lock);
2737
2738 rctl_decr_locked_mem(NULL, proj, unlocked, 0);
2739 mutex_exit(&sp->shm_mlock);
2740
2741 if (ppa != NULL)
2742 seg_ppurge_wiredpp(ppa);
2743 }
2744 return (sts);
2745 }
2746
2747 /*ARGSUSED*/
2748 int
2749 segspt_shmgetprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
2750 {
2751 struct shm_data *shmd = (struct shm_data *)seg->s_data;
2752 struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2753 spgcnt_t pgno = seg_page(seg, addr+len) - seg_page(seg, addr) + 1;
2754
2755 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2756
2757 /*
2758 * ISM segment is always rw.
2759 */
2760 while (--pgno >= 0)
2761 *protv++ = sptd->spt_prot;
2762 return (0);
2763 }
2764
2765 /*ARGSUSED*/
2766 uoff_t
2767 segspt_shmgetoffset(struct seg *seg, caddr_t addr)
2768 {
2769 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2770
2771 /* Offset does not matter in ISM memory */
2772
2773 return (0);
2774 }
2775
2776 /* ARGSUSED */
2777 int
2778 segspt_shmgettype(struct seg *seg, caddr_t addr)
2779 {
2780 struct shm_data *shmd = (struct shm_data *)seg->s_data;
2781 struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2782
2783 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2784
2785 /*
2786 * The shared memory mapping is always MAP_SHARED, SWAP is only
2787 * reserved for DISM
2788 */
2789 return (MAP_SHARED |
2790 ((sptd->spt_flags & SHM_PAGEABLE) ? 0 : MAP_NORESERVE));
2791 }
2792
2793 /*ARGSUSED*/
2794 int
2795 segspt_shmgetvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
2796 {
2797 struct shm_data *shmd = (struct shm_data *)seg->s_data;
2798 struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2799
2800 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2801
2802 *vpp = sptd->spt_vp;
2803 return (0);
2804 }
2805
2806 /*
2807 * We need to wait for pending IO to complete to a DISM segment in order for
2808 * pages to get kicked out of the seg_pcache. 120 seconds should be more
2809 * than enough time to wait.
2810 */
2811 static clock_t spt_pcache_wait = 120;
2812
2813 /*ARGSUSED*/
2814 static int
2815 segspt_shmadvise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
2816 {
2817 struct shm_data *shmd = (struct shm_data *)seg->s_data;
2818 struct spt_data *sptd = (struct spt_data *)shmd->shm_sptseg->s_data;
2819 struct anon_map *amp;
2820 pgcnt_t pg_idx;
2821 ushort_t gen;
2822 clock_t end_lbolt;
2823 int writer;
2824 page_t **ppa;
2825
2826 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2827
2828 if (behav == MADV_FREE || behav == MADV_PURGE) {
2829 if ((sptd->spt_flags & SHM_PAGEABLE) == 0)
2830 return (0);
2831
2832 amp = sptd->spt_amp;
2833 pg_idx = seg_page(seg, addr);
2834
2835 mutex_enter(&sptd->spt_lock);
2836 if ((ppa = sptd->spt_ppa) == NULL) {
2837 mutex_exit(&sptd->spt_lock);
2838 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2839 (void) anon_disclaim(amp, pg_idx, len, behav, NULL);
2840 ANON_LOCK_EXIT(&amp->a_rwlock);
2841 return (0);
2842 }
2843
2844 sptd->spt_flags |= DISM_PPA_CHANGED;
2845 gen = sptd->spt_gen;
2846
2847 mutex_exit(&sptd->spt_lock);
2848
2849 /*
2850 * Purge all DISM cached pages
2851 */
2852 seg_ppurge_wiredpp(ppa);
2853
2854 /*
2855 * Drop the AS_LOCK so that other threads can grab it
2856 * in the as_pageunlock path and hopefully get the segment
2857 * kicked out of the seg_pcache. We bump the shm_softlockcnt
2858 * to keep this segment resident.
2859 */
2860 writer = AS_WRITE_HELD(seg->s_as);
2861 atomic_inc_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
2862 AS_LOCK_EXIT(seg->s_as);
2863
2864 mutex_enter(&sptd->spt_lock);
2865
2866 end_lbolt = ddi_get_lbolt() + (hz * spt_pcache_wait);
2867
2868 /*
2869 * Try to wait for pages to get kicked out of the seg_pcache.
2870 */
2871 while (sptd->spt_gen == gen &&
2872 (sptd->spt_flags & DISM_PPA_CHANGED) &&
2873 ddi_get_lbolt() < end_lbolt) {
2874 if (!cv_timedwait_sig(&sptd->spt_cv,
2875 &sptd->spt_lock, end_lbolt)) {
2876 break;
2877 }
2878 }
2879
2880 mutex_exit(&sptd->spt_lock);
2881
2882 /* Regrab the AS_LOCK and release our hold on the segment */
2883 AS_LOCK_ENTER(seg->s_as, writer ? RW_WRITER : RW_READER);
2884 atomic_dec_ulong((ulong_t *)(&(shmd->shm_softlockcnt)));
2885 if (shmd->shm_softlockcnt <= 0) {
2886 if (AS_ISUNMAPWAIT(seg->s_as)) {
2887 mutex_enter(&seg->s_as->a_contents);
2888 if (AS_ISUNMAPWAIT(seg->s_as)) {
2889 AS_CLRUNMAPWAIT(seg->s_as);
2890 cv_broadcast(&seg->s_as->a_cv);
2891 }
2892 mutex_exit(&seg->s_as->a_contents);
2893 }
2894 }
2895
2896 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2897 (void) anon_disclaim(amp, pg_idx, len, behav, NULL);
2898 ANON_LOCK_EXIT(&amp->a_rwlock);
2899 } else if (lgrp_optimizations() && (behav == MADV_ACCESS_LWP ||
2900 behav == MADV_ACCESS_MANY || behav == MADV_ACCESS_DEFAULT)) {
2901 int already_set;
2902 ulong_t anon_index;
2903 lgrp_mem_policy_t policy;
2904 caddr_t shm_addr;
2905 size_t share_size;
2906 size_t size;
2907 struct seg *sptseg = shmd->shm_sptseg;
2908 caddr_t sptseg_addr;
2909
2910 /*
2911 * Align address and length to page size of underlying segment
2912 */
2913 share_size = page_get_pagesize(shmd->shm_sptseg->s_szc);
2914 shm_addr = (caddr_t)P2ALIGN((uintptr_t)(addr), share_size);
2915 size = P2ROUNDUP((uintptr_t)(((addr + len) - shm_addr)),
2916 share_size);
2917
2918 amp = shmd->shm_amp;
2919 anon_index = seg_page(seg, shm_addr);
2920
2921 /*
2922 * And now we may have to adjust size downward if we have
2923 * exceeded the realsize of the segment or initial anon
2924 * allocations.
2925 */
2926 sptseg_addr = sptseg->s_base + ptob(anon_index);
2927 if ((sptseg_addr + size) >
2928 (sptseg->s_base + sptd->spt_realsize))
2929 size = (sptseg->s_base + sptd->spt_realsize) -
2930 sptseg_addr;
2931
2932 /*
2933 * Set memory allocation policy for this segment
2934 */
2935 policy = lgrp_madv_to_policy(behav, len, MAP_SHARED);
2936 already_set = lgrp_shm_policy_set(policy, amp, anon_index,
2937 NULL, 0, len);
2938
2939 /*
2940 * If random memory allocation policy set already,
2941 * don't bother reapplying it.
2942 */
2943 if (already_set && !LGRP_MEM_POLICY_REAPPLICABLE(policy))
2944 return (0);
2945
2946 /*
2947 * Mark any existing pages in the given range for
2948 * migration, flushing the I/O page cache, and using
2949 * underlying segment to calculate anon index and get
2950 * anonmap and vnode pointer from
2951 */
2952 if (shmd->shm_softlockcnt > 0)
2953 segspt_purge(seg);
2954
2955 page_mark_migrate(seg, shm_addr, size, amp, 0, NULL, 0, 0);
2956 }
2957
2958 return (0);
2959 }
2960
2961 /*
2962 * get a memory ID for an addr in a given segment
2963 */
2964 static int
2965 segspt_shmgetmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
2966 {
2967 struct shm_data *shmd = (struct shm_data *)seg->s_data;
2968 struct anon *ap;
2969 size_t anon_index;
2970 struct anon_map *amp = shmd->shm_amp;
2971 struct spt_data *sptd = shmd->shm_sptseg->s_data;
2972 struct seg *sptseg = shmd->shm_sptseg;
2973 anon_sync_obj_t cookie;
2974
2975 anon_index = seg_page(seg, addr);
2976
2977 if (addr > (seg->s_base + sptd->spt_realsize)) {
2978 return (EFAULT);
2979 }
2980
2981 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2982 anon_array_enter(amp, anon_index, &cookie);
2983 ap = anon_get_ptr(amp->ahp, anon_index);
2984 if (ap == NULL) {
2985 struct page *pp;
2986 caddr_t spt_addr = sptseg->s_base + ptob(anon_index);
2987
2988 pp = anon_zero(sptseg, spt_addr, &ap, kcred);
2989 if (pp == NULL) {
2990 anon_array_exit(&cookie);
2991 ANON_LOCK_EXIT(&amp->a_rwlock);
2992 return (ENOMEM);
2993 }
2994 (void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
2995 page_unlock(pp);
2996 }
2997 anon_array_exit(&cookie);
2998 ANON_LOCK_EXIT(&amp->a_rwlock);
2999 memidp->val[0] = (uintptr_t)ap;
3000 memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
3001 return (0);
3002 }
3003
3004 /*
3005 * Get memory allocation policy info for specified address in given segment
3006 */
3007 static lgrp_mem_policy_info_t *
3008 segspt_shmgetpolicy(struct seg *seg, caddr_t addr)
3009 {
3010 struct anon_map *amp;
3011 ulong_t anon_index;
3012 lgrp_mem_policy_info_t *policy_info;
3013 struct shm_data *shm_data;
3014
3015 ASSERT(seg != NULL);
3016
3017 /*
3018 * Get anon_map from segshm
3019 *
3020 * Assume that no lock needs to be held on anon_map, since
3021 * it should be protected by its reference count which must be
3022 * nonzero for an existing segment
3023 * Need to grab readers lock on policy tree though
3024 */
3025 shm_data = (struct shm_data *)seg->s_data;
3026 if (shm_data == NULL)
3027 return (NULL);
3028 amp = shm_data->shm_amp;
3029 ASSERT(amp->refcnt != 0);
3030
3031 /*
3032 * Get policy info
3033 *
3034 * Assume starting anon index of 0
3035 */
3036 anon_index = seg_page(seg, addr);
3037 policy_info = lgrp_shm_policy_get(amp, anon_index, NULL, 0);
3038
3039 return (policy_info);
3040 }
Unleashed OS