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 /*
23 * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
24 */
25
26 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
27 /* All Rights Reserved */
28
29 /*
30 * University Copyright- Copyright (c) 1982, 1986, 1988
31 * The Regents of the University of California
32 * All Rights Reserved
33 *
34 * University Acknowledgment- Portions of this document are derived from
35 * software developed by the University of California, Berkeley, and its
36 * contributors.
37 */
38
39 /*
40 * Inter-Process Communication Shared Memory Facility.
41 *
42 * See os/ipc.c for a description of common IPC functionality.
43 *
44 * Resource controls
45 * -----------------
46 *
47 * Control: zone.max-shm-ids (rc_zone_shmmni)
48 * Description: Maximum number of shared memory ids allowed a zone.
49 *
50 * When shmget() is used to allocate a shared memory segment, one id
51 * is allocated. If the id allocation doesn't succeed, shmget()
52 * fails and errno is set to ENOSPC. Upon successful shmctl(,
53 * IPC_RMID) the id is deallocated.
54 *
55 * Control: project.max-shm-ids (rc_project_shmmni)
56 * Description: Maximum number of shared memory ids allowed a project.
57 *
58 * When shmget() is used to allocate a shared memory segment, one id
59 * is allocated. If the id allocation doesn't succeed, shmget()
60 * fails and errno is set to ENOSPC. Upon successful shmctl(,
61 * IPC_RMID) the id is deallocated.
62 *
63 * Control: zone.max-shm-memory (rc_zone_shmmax)
64 * Description: Total amount of shared memory allowed a zone.
65 *
66 * When shmget() is used to allocate a shared memory segment, the
67 * segment's size is allocated against this limit. If the space
68 * allocation doesn't succeed, shmget() fails and errno is set to
69 * EINVAL. The size will be deallocated once the last process has
70 * detached the segment and the segment has been successfully
71 * shmctl(, IPC_RMID)ed.
72 *
73 * Control: project.max-shm-memory (rc_project_shmmax)
74 * Description: Total amount of shared memory allowed a project.
75 *
76 * When shmget() is used to allocate a shared memory segment, the
77 * segment's size is allocated against this limit. If the space
78 * allocation doesn't succeed, shmget() fails and errno is set to
79 * EINVAL. The size will be deallocated once the last process has
80 * detached the segment and the segment has been successfully
81 * shmctl(, IPC_RMID)ed.
82 */
83
84 #include <sys/types.h>
85 #include <sys/param.h>
86 #include <sys/cred.h>
87 #include <sys/errno.h>
88 #include <sys/time.h>
89 #include <sys/kmem.h>
90 #include <sys/user.h>
91 #include <sys/proc.h>
92 #include <sys/systm.h>
93 #include <sys/prsystm.h>
94 #include <sys/sysmacros.h>
95 #include <sys/tuneable.h>
96 #include <sys/vm.h>
97 #include <sys/mman.h>
98 #include <sys/swap.h>
99 #include <sys/cmn_err.h>
100 #include <sys/debug.h>
101 #include <sys/lwpchan_impl.h>
102 #include <sys/avl.h>
103 #include <sys/modctl.h>
104 #include <sys/syscall.h>
105 #include <sys/task.h>
106 #include <sys/project.h>
107 #include <sys/policy.h>
108 #include <sys/zone.h>
109 #include <sys/rctl.h>
110
111 #include <sys/ipc.h>
112 #include <sys/ipc_impl.h>
113 #include <sys/shm.h>
114 #include <sys/shm_impl.h>
115
116 #include <vm/hat.h>
117 #include <vm/seg.h>
118 #include <vm/as.h>
119 #include <vm/seg_vn.h>
120 #include <vm/anon.h>
121 #include <vm/page.h>
122 #include <vm/vpage.h>
123 #include <vm/seg_spt.h>
124
125 #include <c2/audit.h>
126
127 static int shmem_lock(kshmid_t *sp, struct anon_map *amp);
128 static void shmem_unlock(kshmid_t *sp, struct anon_map *amp);
129 static void sa_add(struct proc *pp, caddr_t addr, size_t len, ulong_t flags,
130 kshmid_t *id);
131 static void shm_rm_amp(kshmid_t *sp);
132 static void shm_dtor(kipc_perm_t *);
133 static void shm_rmid(kipc_perm_t *);
134 static void shm_remove_zone(zoneid_t, void *);
135
136 /*
137 * Semantics for share_page_table and ism_off:
138 *
139 * These are hooks in /etc/system - only for internal testing purpose.
140 *
141 * Setting share_page_table automatically turns on the SHM_SHARE_MMU (ISM) flag
142 * in a call to shmat(2). In other words, with share_page_table set, you always
143 * get ISM, even if say, DISM is specified. It should really be called "ism_on".
144 *
145 * Setting ism_off turns off the SHM_SHARE_MMU flag from the flags passed to
146 * shmat(2).
147 *
148 * If both share_page_table and ism_off are set, share_page_table prevails.
149 *
150 * Although these tunables should probably be removed, they do have some
151 * external exposure; as long as they exist, they should at least work sensibly.
152 */
153
154 int share_page_table;
155 int ism_off;
156
157 /*
158 * The following tunables are obsolete. Though for compatibility we
159 * still read and interpret shminfo_shmmax and shminfo_shmmni (see
160 * os/project.c), the preferred mechanism for administrating the IPC
161 * Shared Memory facility is through the resource controls described at
162 * the top of this file.
163 */
164 size_t shminfo_shmmax = 0x800000; /* (obsolete) */
165 int shminfo_shmmni = 100; /* (obsolete) */
166 size_t shminfo_shmmin = 1; /* (obsolete) */
167 int shminfo_shmseg = 6; /* (obsolete) */
168
169 extern rctl_hndl_t rc_zone_shmmax;
170 extern rctl_hndl_t rc_zone_shmmni;
171 extern rctl_hndl_t rc_project_shmmax;
172 extern rctl_hndl_t rc_project_shmmni;
173 static ipc_service_t *shm_svc;
174 static zone_key_t shm_zone_key;
175
176 /*
177 * Module linkage information for the kernel.
178 */
179 static uintptr_t shmsys(int, uintptr_t, uintptr_t, uintptr_t);
180
181 static struct sysent ipcshm_sysent = {
182 4,
183 #ifdef _SYSCALL32_IMPL
184 SE_ARGC | SE_NOUNLOAD | SE_64RVAL,
185 #else /* _SYSCALL32_IMPL */
186 SE_ARGC | SE_NOUNLOAD | SE_32RVAL1,
187 #endif /* _SYSCALL32_IMPL */
188 (int (*)())shmsys
189 };
190
191 #ifdef _SYSCALL32_IMPL
192 static struct sysent ipcshm_sysent32 = {
193 4,
194 SE_ARGC | SE_NOUNLOAD | SE_32RVAL1,
195 (int (*)())shmsys
196 };
197 #endif /* _SYSCALL32_IMPL */
198
199 static struct modlsys modlsys = {
200 &mod_syscallops, "System V shared memory", &ipcshm_sysent
201 };
202
203 #ifdef _SYSCALL32_IMPL
204 static struct modlsys modlsys32 = {
205 &mod_syscallops32, "32-bit System V shared memory", &ipcshm_sysent32
206 };
207 #endif /* _SYSCALL32_IMPL */
208
209 static struct modlinkage modlinkage = {
210 MODREV_1,
211 &modlsys,
212 #ifdef _SYSCALL32_IMPL
213 &modlsys32,
214 #endif
215 NULL
216 };
217
218
219 int
220 _init(void)
221 {
222 int result;
223
224 shm_svc = ipcs_create("shmids", rc_project_shmmni, rc_zone_shmmni,
225 sizeof (kshmid_t), shm_dtor, shm_rmid, AT_IPC_SHM,
226 offsetof(ipc_rqty_t, ipcq_shmmni));
227 zone_key_create(&shm_zone_key, NULL, shm_remove_zone, NULL);
228
229 if ((result = mod_install(&modlinkage)) == 0)
230 return (0);
231
232 (void) zone_key_delete(shm_zone_key);
233 ipcs_destroy(shm_svc);
234
235 return (result);
236 }
237
238 int
239 _fini(void)
240 {
241 return (EBUSY);
242 }
243
244 int
245 _info(struct modinfo *modinfop)
246 {
247 return (mod_info(&modlinkage, modinfop));
248 }
249
250 /*
251 * Shmat (attach shared segment) system call.
252 */
253 static int
254 shmat(int shmid, caddr_t uaddr, int uflags, uintptr_t *rvp)
255 {
256 kshmid_t *sp; /* shared memory header ptr */
257 size_t size;
258 int error = 0;
259 proc_t *pp = curproc;
260 struct as *as = pp->p_as;
261 struct segvn_crargs crargs; /* segvn create arguments */
262 kmutex_t *lock;
263 struct seg *segspt = NULL;
264 caddr_t addr = uaddr;
265 int flags = (uflags & SHMAT_VALID_FLAGS_MASK);
266 int useISM;
267 uchar_t prot = PROT_ALL;
268 int result;
269
270 if ((lock = ipc_lookup(shm_svc, shmid, (kipc_perm_t **)&sp)) == NULL)
271 return (EINVAL);
272 if (error = ipcperm_access(&sp->shm_perm, SHM_R, CRED()))
273 goto errret;
274 if ((flags & SHM_RDONLY) == 0 &&
275 (error = ipcperm_access(&sp->shm_perm, SHM_W, CRED())))
276 goto errret;
277 if (spt_invalid(flags)) {
278 error = EINVAL;
279 goto errret;
280 }
281 if (ism_off)
282 flags = flags & ~SHM_SHARE_MMU;
283 if (share_page_table) {
284 flags = flags & ~SHM_PAGEABLE;
285 flags = flags | SHM_SHARE_MMU;
286 }
287 useISM = (spt_locked(flags) || spt_pageable(flags));
288 if (useISM && (error = ipcperm_access(&sp->shm_perm, SHM_W, CRED())))
289 goto errret;
290 if (useISM && isspt(sp)) {
291 uint_t newsptflags = flags | spt_flags(sp->shm_sptseg);
292 /*
293 * If trying to change an existing {D}ISM segment from ISM
294 * to DISM or vice versa, return error. Note that this
295 * validation of flags needs to be done after the effect of
296 * tunables such as ism_off and share_page_table, for
297 * semantics that are consistent with the tunables' settings.
298 */
299 if (spt_invalid(newsptflags)) {
300 error = EINVAL;
301 goto errret;
302 }
303 }
304 ANON_LOCK_ENTER(&sp->shm_amp->a_rwlock, RW_WRITER);
305 size = sp->shm_amp->size;
306 ANON_LOCK_EXIT(&sp->shm_amp->a_rwlock);
307
308 /* somewhere to record spt info for final detach */
309 if (sp->shm_sptinfo == NULL)
310 sp->shm_sptinfo = kmem_zalloc(sizeof (sptinfo_t), KM_SLEEP);
311
312 as_rangelock(as);
313
314 if (useISM) {
315 /*
316 * Handle ISM
317 */
318 uint_t share_szc;
319 size_t share_size;
320 struct shm_data ssd;
321 uintptr_t align_hint;
322
323 /*
324 * Pick a share pagesize to use, if (!isspt(sp)).
325 * Otherwise use the already chosen page size.
326 *
327 * For the initial shmat (!isspt(sp)), where sptcreate is
328 * called, map_pgsz is called to recommend a [D]ISM pagesize,
329 * important for systems which offer more than one potential
330 * [D]ISM pagesize.
331 * If the shmat is just to attach to an already created
332 * [D]ISM segment, then use the previously selected page size.
333 */
334 if (!isspt(sp)) {
335 share_size = map_pgsz(MAPPGSZ_ISM, pp, addr, size, 0);
336 if (share_size == 0) {
337 as_rangeunlock(as);
338 error = EINVAL;
339 goto errret;
340 }
341 share_szc = page_szc(share_size);
342 } else {
343 share_szc = sp->shm_sptseg->s_szc;
344 share_size = page_get_pagesize(share_szc);
345 }
346 size = P2ROUNDUP(size, share_size);
347
348 align_hint = share_size;
349 #if defined(__i386) || defined(__amd64)
350 /*
351 * For x86, we want to share as much of the page table tree
352 * as possible. We use a large align_hint at first, but
353 * if that fails, then the code below retries with align_hint
354 * set to share_size.
355 *
356 * The explicit extern here is due to the difficulties
357 * of getting to platform dependent includes. When/if the
358 * platform dependent bits of this function are cleaned up,
359 * another way of doing this should found.
360 */
361 {
362 extern uint_t ptes_per_table;
363
364 while (size >= ptes_per_table * (uint64_t)align_hint)
365 align_hint *= ptes_per_table;
366 }
367 #endif /* __i386 || __amd64 */
368
369 #if defined(__sparcv9)
370 if (addr == 0 &&
371 pp->p_model == DATAMODEL_LP64 && AS_TYPE_64BIT(as)) {
372 /*
373 * If no address has been passed in, and this is a
374 * 64-bit process, we'll try to find an address
375 * in the predict-ISM zone.
376 */
377 caddr_t predbase = (caddr_t)PREDISM_1T_BASE;
378 size_t len = PREDISM_BOUND - PREDISM_1T_BASE;
379
380 as_purge(as);
381 if (as_gap(as, size + share_size, &predbase, &len,
382 AH_LO, (caddr_t)NULL) != -1) {
383 /*
384 * We found an address which looks like a
385 * candidate. We want to round it up, and
386 * then check that it's a valid user range.
387 * This assures that we won't fail below.
388 */
389 addr = (caddr_t)P2ROUNDUP((uintptr_t)predbase,
390 share_size);
391
392 if (valid_usr_range(addr, size, prot,
393 as, as->a_userlimit) != RANGE_OKAY) {
394 addr = 0;
395 }
396 }
397 }
398 #endif /* __sparcv9 */
399
400 if (addr == 0) {
401 for (;;) {
402 addr = (caddr_t)align_hint;
403 map_addr(&addr, size, 0ll, 1, MAP_ALIGN);
404 if (addr != NULL || align_hint == share_size)
405 break;
406 align_hint = share_size;
407 }
408 if (addr == NULL) {
409 as_rangeunlock(as);
410 error = ENOMEM;
411 goto errret;
412 }
413 ASSERT(((uintptr_t)addr & (align_hint - 1)) == 0);
414 } else {
415 /* Use the user-supplied attach address */
416 caddr_t base;
417 size_t len;
418
419 /*
420 * Check that the address range
421 * 1) is properly aligned
422 * 2) is correct in unix terms
423 * 3) is within an unmapped address segment
424 */
425 base = addr;
426 len = size; /* use spt aligned size */
427 /* XXX - in SunOS, is sp->shm_segsz */
428 if ((uintptr_t)base & (share_size - 1)) {
429 error = EINVAL;
430 as_rangeunlock(as);
431 goto errret;
432 }
433 result = valid_usr_range(base, len, prot, as,
434 as->a_userlimit);
435 if (result == RANGE_BADPROT) {
436 /*
437 * We try to accomodate processors which
438 * may not support execute permissions on
439 * all ISM segments by trying the check
440 * again but without PROT_EXEC.
441 */
442 prot &= ~PROT_EXEC;
443 result = valid_usr_range(base, len, prot, as,
444 as->a_userlimit);
445 }
446 as_purge(as);
447 if (result != RANGE_OKAY ||
448 as_gap(as, len, &base, &len, AH_LO,
449 (caddr_t)NULL) != 0) {
450 error = EINVAL;
451 as_rangeunlock(as);
452 goto errret;
453 }
454 }
455
456 if (!isspt(sp)) {
457 error = sptcreate(size, &segspt, sp->shm_amp, prot,
458 flags, share_szc);
459 if (error) {
460 as_rangeunlock(as);
461 goto errret;
462 }
463 sp->shm_sptinfo->sptas = segspt->s_as;
464 sp->shm_sptseg = segspt;
465 sp->shm_sptprot = prot;
466 } else if ((prot & sp->shm_sptprot) != sp->shm_sptprot) {
467 /*
468 * Ensure we're attaching to an ISM segment with
469 * fewer or equal permissions than what we're
470 * allowed. Fail if the segment has more
471 * permissions than what we're allowed.
472 */
473 error = EACCES;
474 as_rangeunlock(as);
475 goto errret;
476 }
477
478 ssd.shm_sptseg = sp->shm_sptseg;
479 ssd.shm_sptas = sp->shm_sptinfo->sptas;
480 ssd.shm_amp = sp->shm_amp;
481 error = as_map(as, addr, size, segspt_shmattach, &ssd);
482 if (error == 0)
483 sp->shm_ismattch++; /* keep count of ISM attaches */
484 } else {
485
486 /*
487 * Normal case.
488 */
489 if (flags & SHM_RDONLY)
490 prot &= ~PROT_WRITE;
491
492 if (addr == 0) {
493 /* Let the system pick the attach address */
494 map_addr(&addr, size, 0ll, 1, 0);
495 if (addr == NULL) {
496 as_rangeunlock(as);
497 error = ENOMEM;
498 goto errret;
499 }
500 } else {
501 /* Use the user-supplied attach address */
502 caddr_t base;
503 size_t len;
504
505 if (flags & SHM_RND)
506 addr = (caddr_t)((uintptr_t)addr &
507 ~(SHMLBA - 1));
508 /*
509 * Check that the address range
510 * 1) is properly aligned
511 * 2) is correct in unix terms
512 * 3) is within an unmapped address segment
513 */
514 base = addr;
515 len = size; /* use aligned size */
516 /* XXX - in SunOS, is sp->shm_segsz */
517 if ((uintptr_t)base & PAGEOFFSET) {
518 error = EINVAL;
519 as_rangeunlock(as);
520 goto errret;
521 }
522 result = valid_usr_range(base, len, prot, as,
523 as->a_userlimit);
524 if (result == RANGE_BADPROT) {
525 prot &= ~PROT_EXEC;
526 result = valid_usr_range(base, len, prot, as,
527 as->a_userlimit);
528 }
529 as_purge(as);
530 if (result != RANGE_OKAY ||
531 as_gap(as, len, &base, &len,
532 AH_LO, (caddr_t)NULL) != 0) {
533 error = EINVAL;
534 as_rangeunlock(as);
535 goto errret;
536 }
537 }
538
539 /* Initialize the create arguments and map the segment */
540 crargs = *(struct segvn_crargs *)zfod_argsp;
541 crargs.offset = 0;
542 crargs.type = MAP_SHARED;
543 crargs.amp = sp->shm_amp;
544 crargs.prot = prot;
545 crargs.maxprot = crargs.prot;
546 crargs.flags = 0;
547
548 error = as_map(as, addr, size, segvn_create, &crargs);
549 }
550
551 as_rangeunlock(as);
552 if (error)
553 goto errret;
554
555 /* record shmem range for the detach */
556 sa_add(pp, addr, (size_t)size, useISM ? SHMSA_ISM : 0, sp);
557 *rvp = (uintptr_t)addr;
558
559 sp->shm_atime = gethrestime_sec();
560 sp->shm_lpid = pp->p_pid;
561 ipc_hold(shm_svc, (kipc_perm_t *)sp);
562
563 /*
564 * Tell machine specific code that lwp has mapped shared memory
565 */
566 LWP_MMODEL_SHARED_AS(addr, size);
567
568 errret:
569 mutex_exit(lock);
570 return (error);
571 }
572
573 static void
574 shm_dtor(kipc_perm_t *perm)
575 {
576 kshmid_t *sp = (kshmid_t *)perm;
577 uint_t cnt;
578 size_t rsize;
579
580 ANON_LOCK_ENTER(&sp->shm_amp->a_rwlock, RW_WRITER);
581 anonmap_purge(sp->shm_amp);
582 ANON_LOCK_EXIT(&sp->shm_amp->a_rwlock);
583
584 if (sp->shm_sptinfo) {
585 if (isspt(sp)) {
586 sptdestroy(sp->shm_sptinfo->sptas, sp->shm_amp);
587 sp->shm_lkcnt = 0;
588 }
589 kmem_free(sp->shm_sptinfo, sizeof (sptinfo_t));
590 }
591
592 if (sp->shm_lkcnt > 0) {
593 shmem_unlock(sp, sp->shm_amp);
594 sp->shm_lkcnt = 0;
595 }
596
597 ANON_LOCK_ENTER(&sp->shm_amp->a_rwlock, RW_WRITER);
598 cnt = --sp->shm_amp->refcnt;
599 ANON_LOCK_EXIT(&sp->shm_amp->a_rwlock);
600 ASSERT(cnt == 0);
601 shm_rm_amp(sp);
602
603 if (sp->shm_perm.ipc_id != IPC_ID_INVAL) {
604 rsize = ptob(btopr(sp->shm_segsz));
605 ipcs_lock(shm_svc);
606 sp->shm_perm.ipc_proj->kpj_data.kpd_shmmax -= rsize;
607 sp->shm_perm.ipc_zone_ref.zref_zone->zone_shmmax -= rsize;
608 ipcs_unlock(shm_svc);
609 }
610 }
611
612 /* ARGSUSED */
613 static void
614 shm_rmid(kipc_perm_t *perm)
615 {
616 /* nothing to do */
617 }
618
619 /*
620 * Shmctl system call.
621 */
622 /* ARGSUSED */
623 static int
624 shmctl(int shmid, int cmd, void *arg)
625 {
626 kshmid_t *sp; /* shared memory header ptr */
627 STRUCT_DECL(shmid_ds, ds); /* for SVR4 IPC_SET */
628 int error = 0;
629 struct cred *cr = CRED();
630 kmutex_t *lock;
631 model_t mdl = get_udatamodel();
632 struct shmid_ds64 ds64;
633 shmatt_t nattch;
634
635 STRUCT_INIT(ds, mdl);
636
637 /*
638 * Perform pre- or non-lookup actions (e.g. copyins, RMID).
639 */
640 switch (cmd) {
641 case IPC_SET:
642 if (copyin(arg, STRUCT_BUF(ds), STRUCT_SIZE(ds)))
643 return (EFAULT);
644 break;
645
646 case IPC_SET64:
647 if (copyin(arg, &ds64, sizeof (struct shmid_ds64)))
648 return (EFAULT);
649 break;
650
651 case IPC_RMID:
652 return (ipc_rmid(shm_svc, shmid, cr));
653 }
654
655 if ((lock = ipc_lookup(shm_svc, shmid, (kipc_perm_t **)&sp)) == NULL)
656 return (EINVAL);
657
658 switch (cmd) {
659 /* Set ownership and permissions. */
660 case IPC_SET:
661 if (error = ipcperm_set(shm_svc, cr, &sp->shm_perm,
662 &STRUCT_BUF(ds)->shm_perm, mdl))
663 break;
664 sp->shm_ctime = gethrestime_sec();
665 break;
666
667 case IPC_STAT:
668 if (error = ipcperm_access(&sp->shm_perm, SHM_R, cr))
669 break;
670
671 nattch = sp->shm_perm.ipc_ref - 1;
672
673 ipcperm_stat(&STRUCT_BUF(ds)->shm_perm, &sp->shm_perm, mdl);
674 STRUCT_FSET(ds, shm_segsz, sp->shm_segsz);
675 STRUCT_FSETP(ds, shm_amp, NULL); /* kernel addr */
676 STRUCT_FSET(ds, shm_lkcnt, sp->shm_lkcnt);
677 STRUCT_FSET(ds, shm_lpid, sp->shm_lpid);
678 STRUCT_FSET(ds, shm_cpid, sp->shm_cpid);
679 STRUCT_FSET(ds, shm_nattch, nattch);
680 STRUCT_FSET(ds, shm_cnattch, sp->shm_ismattch);
681 STRUCT_FSET(ds, shm_atime, sp->shm_atime);
682 STRUCT_FSET(ds, shm_dtime, sp->shm_dtime);
683 STRUCT_FSET(ds, shm_ctime, sp->shm_ctime);
684
685 mutex_exit(lock);
686 if (copyout(STRUCT_BUF(ds), arg, STRUCT_SIZE(ds)))
687 return (EFAULT);
688
689 return (0);
690
691 case IPC_SET64:
692 if (error = ipcperm_set64(shm_svc, cr,
693 &sp->shm_perm, &ds64.shmx_perm))
694 break;
695 sp->shm_ctime = gethrestime_sec();
696 break;
697
698 case IPC_STAT64:
699 nattch = sp->shm_perm.ipc_ref - 1;
700
701 ipcperm_stat64(&ds64.shmx_perm, &sp->shm_perm);
702 ds64.shmx_segsz = sp->shm_segsz;
703 ds64.shmx_lkcnt = sp->shm_lkcnt;
704 ds64.shmx_lpid = sp->shm_lpid;
705 ds64.shmx_cpid = sp->shm_cpid;
706 ds64.shmx_nattch = nattch;
707 ds64.shmx_cnattch = sp->shm_ismattch;
708 ds64.shmx_atime = sp->shm_atime;
709 ds64.shmx_dtime = sp->shm_dtime;
710 ds64.shmx_ctime = sp->shm_ctime;
711
712 mutex_exit(lock);
713 if (copyout(&ds64, arg, sizeof (struct shmid_ds64)))
714 return (EFAULT);
715
716 return (0);
717
718 /* Lock segment in memory */
719 case SHM_LOCK:
720 if ((error = secpolicy_lock_memory(cr)) != 0)
721 break;
722
723 /* protect against overflow */
724 if (sp->shm_lkcnt >= USHRT_MAX) {
725 error = ENOMEM;
726 break;
727 }
728 if (!isspt(sp) && (sp->shm_lkcnt++ == 0)) {
729 if (error = shmem_lock(sp, sp->shm_amp)) {
730 ANON_LOCK_ENTER(&sp->shm_amp->a_rwlock,
731 RW_WRITER);
732 cmn_err(CE_NOTE, "shmctl - couldn't lock %ld"
733 " pages into memory", sp->shm_amp->size);
734 ANON_LOCK_EXIT(&sp->shm_amp->a_rwlock);
735 error = ENOMEM;
736 sp->shm_lkcnt--;
737 }
738 }
739 break;
740
741 /* Unlock segment */
742 case SHM_UNLOCK:
743 if ((error = secpolicy_lock_memory(cr)) != 0)
744 break;
745
746 if (sp->shm_lkcnt && (--sp->shm_lkcnt == 0)) {
747 shmem_unlock(sp, sp->shm_amp);
748 }
749 break;
750
751 default:
752 error = EINVAL;
753 break;
754 }
755 mutex_exit(lock);
756 return (error);
757 }
758
759 static void
760 shm_detach(proc_t *pp, segacct_t *sap)
761 {
762 kshmid_t *sp = sap->sa_id;
763 size_t len = sap->sa_len;
764 caddr_t addr = sap->sa_addr;
765
766 /*
767 * Discard lwpchan mappings.
768 */
769 if (pp->p_lcp != NULL)
770 lwpchan_delete_mapping(pp, addr, addr + len);
771 (void) as_unmap(pp->p_as, addr, len);
772
773 /*
774 * Perform some detach-time accounting.
775 */
776 (void) ipc_lock(shm_svc, sp->shm_perm.ipc_id);
777 if (sap->sa_flags & SHMSA_ISM)
778 sp->shm_ismattch--;
779 sp->shm_dtime = gethrestime_sec();
780 sp->shm_lpid = pp->p_pid;
781 ipc_rele(shm_svc, (kipc_perm_t *)sp); /* Drops lock */
782
783 kmem_free(sap, sizeof (segacct_t));
784 }
785
786 static int
787 shmdt(caddr_t addr)
788 {
789 proc_t *pp = curproc;
790 segacct_t *sap, template;
791
792 mutex_enter(&pp->p_lock);
793 prbarrier(pp); /* block /proc. See shmgetid(). */
794
795 template.sa_addr = addr;
796 template.sa_len = 0;
797 if ((pp->p_segacct == NULL) ||
798 ((sap = avl_find(pp->p_segacct, &template, NULL)) == NULL)) {
799 mutex_exit(&pp->p_lock);
800 return (EINVAL);
801 }
802 if (sap->sa_addr != addr) {
803 mutex_exit(&pp->p_lock);
804 return (EINVAL);
805 }
806 avl_remove(pp->p_segacct, sap);
807 mutex_exit(&pp->p_lock);
808
809 shm_detach(pp, sap);
810
811 return (0);
812 }
813
814 /*
815 * Remove all shared memory segments associated with a given zone.
816 * Called by zone_shutdown when the zone is halted.
817 */
818 /*ARGSUSED1*/
819 static void
820 shm_remove_zone(zoneid_t zoneid, void *arg)
821 {
822 ipc_remove_zone(shm_svc, zoneid);
823 }
824
825 /*
826 * Shmget (create new shmem) system call.
827 */
828 static int
829 shmget(key_t key, size_t size, int shmflg, uintptr_t *rvp)
830 {
831 proc_t *pp = curproc;
832 kshmid_t *sp;
833 kmutex_t *lock;
834 int error;
835
836 top:
837 if (error = ipc_get(shm_svc, key, shmflg, (kipc_perm_t **)&sp, &lock))
838 return (error);
839
840 if (!IPC_FREE(&sp->shm_perm)) {
841 /*
842 * A segment with the requested key exists.
843 */
844 if (size > sp->shm_segsz) {
845 mutex_exit(lock);
846 return (EINVAL);
847 }
848 } else {
849 /*
850 * A new segment should be created.
851 */
852 size_t npages = btopr(size);
853 size_t rsize = ptob(npages);
854
855 /*
856 * Check rsize and the per-project and per-zone limit on
857 * shared memory. Checking rsize handles both the size == 0
858 * case and the size < ULONG_MAX & PAGEMASK case (i.e.
859 * rounding up wraps a size_t).
860 */
861 if (rsize == 0 ||
862 (rctl_test(rc_project_shmmax,
863 pp->p_task->tk_proj->kpj_rctls, pp, rsize,
864 RCA_SAFE) & RCT_DENY) ||
865 (rctl_test(rc_zone_shmmax,
866 pp->p_zone->zone_rctls, pp, rsize,
867 RCA_SAFE) & RCT_DENY)) {
868
869 mutex_exit(&pp->p_lock);
870 mutex_exit(lock);
871 ipc_cleanup(shm_svc, (kipc_perm_t *)sp);
872 return (EINVAL);
873 }
874 mutex_exit(&pp->p_lock);
875 mutex_exit(lock);
876
877 if (anon_resv(rsize) == 0) {
878 ipc_cleanup(shm_svc, (kipc_perm_t *)sp);
879 return (ENOMEM);
880 }
881
882 /*
883 * If any new failure points are introduced between the
884 * the above anon_resv() and the below ipc_commit_begin(),
885 * these failure points will need to unreserve the anon
886 * reserved using anon_unresv().
887 *
888 * Once ipc_commit_begin() is called, the anon reserved
889 * above will be automatically unreserved by future calls to
890 * ipcs_cleanup() -> shm_dtor() -> shm_rm_amp(). If
891 * ipc_commit_begin() fails, it internally calls shm_dtor(),
892 * unreserving the above anon, and freeing the below amp.
893 */
894
895 sp->shm_amp = anonmap_alloc(rsize, rsize, ANON_SLEEP);
896 sp->shm_amp->a_sp = sp;
897 /*
898 * Store the original user's requested size, in bytes,
899 * rather than the page-aligned size. The former is
900 * used for IPC_STAT and shmget() lookups. The latter
901 * is saved in the anon_map structure and is used for
902 * calls to the vm layer.
903 */
904 sp->shm_segsz = size;
905 sp->shm_atime = sp->shm_dtime = 0;
906 sp->shm_ctime = gethrestime_sec();
907 sp->shm_lpid = (pid_t)0;
908 sp->shm_cpid = curproc->p_pid;
909 sp->shm_ismattch = 0;
910 sp->shm_sptinfo = NULL;
911 /*
912 * Check limits one last time, push id into global
913 * visibility, and update resource usage counts.
914 */
915 if (error = ipc_commit_begin(shm_svc, key, shmflg,
916 (kipc_perm_t *)sp)) {
917 if (error == EAGAIN)
918 goto top;
919 return (error);
920 }
921
922 if ((rctl_test(rc_project_shmmax,
923 sp->shm_perm.ipc_proj->kpj_rctls, pp, rsize,
924 RCA_SAFE) & RCT_DENY) ||
925 (rctl_test(rc_zone_shmmax,
926 sp->shm_perm.ipc_zone_ref.zref_zone->zone_rctls, pp, rsize,
927 RCA_SAFE) & RCT_DENY)) {
928 ipc_cleanup(shm_svc, (kipc_perm_t *)sp);
929 return (EINVAL);
930 }
931 sp->shm_perm.ipc_proj->kpj_data.kpd_shmmax += rsize;
932 sp->shm_perm.ipc_zone_ref.zref_zone->zone_shmmax += rsize;
933
934 lock = ipc_commit_end(shm_svc, &sp->shm_perm);
935 }
936
937 if (AU_AUDITING())
938 audit_ipcget(AT_IPC_SHM, (void *)sp);
939
940 *rvp = (uintptr_t)(sp->shm_perm.ipc_id);
941
942 mutex_exit(lock);
943 return (0);
944 }
945
946 /*
947 * shmids system call.
948 */
949 static int
950 shmids(int *buf, uint_t nids, uint_t *pnids)
951 {
952 return (ipc_ids(shm_svc, buf, nids, pnids));
953 }
954
955 /*
956 * System entry point for shmat, shmctl, shmdt, and shmget system calls.
957 */
958 static uintptr_t
959 shmsys(int opcode, uintptr_t a0, uintptr_t a1, uintptr_t a2)
960 {
961 int error;
962 uintptr_t r_val = 0;
963
964 switch (opcode) {
965 case SHMAT:
966 error = shmat((int)a0, (caddr_t)a1, (int)a2, &r_val);
967 break;
968 case SHMCTL:
969 error = shmctl((int)a0, (int)a1, (void *)a2);
970 break;
971 case SHMDT:
972 error = shmdt((caddr_t)a0);
973 break;
974 case SHMGET:
975 error = shmget((key_t)a0, (size_t)a1, (int)a2, &r_val);
976 break;
977 case SHMIDS:
978 error = shmids((int *)a0, (uint_t)a1, (uint_t *)a2);
979 break;
980 default:
981 error = EINVAL;
982 break;
983 }
984
985 if (error)
986 return ((uintptr_t)set_errno(error));
987
988 return (r_val);
989 }
990
991 /*
992 * segacct_t comparator
993 * This works as expected, with one minor change: the first of two real
994 * segments with equal addresses is considered to be 'greater than' the
995 * second. We only return equal when searching using a template, in
996 * which case we explicitly set the template segment's length to 0
997 * (which is invalid for a real segment).
998 */
999 static int
1000 shm_sacompar(const void *x, const void *y)
1001 {
1002 segacct_t *sa1 = (segacct_t *)x;
1003 segacct_t *sa2 = (segacct_t *)y;
1004
1005 if (sa1->sa_addr < sa2->sa_addr) {
1006 return (-1);
1007 } else if (sa2->sa_len != 0) {
1008 if (sa1->sa_addr >= sa2->sa_addr + sa2->sa_len) {
1009 return (1);
1010 } else if (sa1->sa_len != 0) {
1011 return (1);
1012 } else {
1013 return (0);
1014 }
1015 } else if (sa1->sa_addr > sa2->sa_addr) {
1016 return (1);
1017 } else {
1018 return (0);
1019 }
1020 }
1021
1022 /*
1023 * add this record to the segacct list.
1024 */
1025 static void
1026 sa_add(struct proc *pp, caddr_t addr, size_t len, ulong_t flags, kshmid_t *id)
1027 {
1028 segacct_t *nsap;
1029 avl_tree_t *tree = NULL;
1030 avl_index_t where;
1031
1032 nsap = kmem_alloc(sizeof (segacct_t), KM_SLEEP);
1033 nsap->sa_addr = addr;
1034 nsap->sa_len = len;
1035 nsap->sa_flags = flags;
1036 nsap->sa_id = id;
1037
1038 if (pp->p_segacct == NULL)
1039 tree = kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
1040
1041 mutex_enter(&pp->p_lock);
1042 prbarrier(pp); /* block /proc. See shmgetid(). */
1043
1044 if (pp->p_segacct == NULL) {
1045 avl_create(tree, shm_sacompar, sizeof (segacct_t),
1046 offsetof(segacct_t, sa_tree));
1047 pp->p_segacct = tree;
1048 } else if (tree) {
1049 kmem_free(tree, sizeof (avl_tree_t));
1050 }
1051
1052 /*
1053 * We can ignore the result of avl_find, as the comparator will
1054 * never return equal for segments with non-zero length. This
1055 * is a necessary hack to get around the fact that we do, in
1056 * fact, have duplicate keys.
1057 */
1058 (void) avl_find(pp->p_segacct, nsap, &where);
1059 avl_insert(pp->p_segacct, nsap, where);
1060
1061 mutex_exit(&pp->p_lock);
1062 }
1063
1064 /*
1065 * Duplicate parent's segacct records in child.
1066 */
1067 void
1068 shmfork(struct proc *ppp, struct proc *cpp)
1069 {
1070 segacct_t *sap;
1071 kshmid_t *sp;
1072 kmutex_t *mp;
1073
1074 ASSERT(ppp->p_segacct != NULL);
1075
1076 /*
1077 * We are the only lwp running in the parent so nobody can
1078 * mess with our p_segacct list. Thus it is safe to traverse
1079 * the list without holding p_lock. This is essential because
1080 * we can't hold p_lock during a KM_SLEEP allocation.
1081 */
1082 for (sap = (segacct_t *)avl_first(ppp->p_segacct); sap != NULL;
1083 sap = (segacct_t *)AVL_NEXT(ppp->p_segacct, sap)) {
1084 sa_add(cpp, sap->sa_addr, sap->sa_len, sap->sa_flags,
1085 sap->sa_id);
1086 sp = sap->sa_id;
1087 mp = ipc_lock(shm_svc, sp->shm_perm.ipc_id);
1088 if (sap->sa_flags & SHMSA_ISM)
1089 sp->shm_ismattch++;
1090 ipc_hold(shm_svc, (kipc_perm_t *)sp);
1091 mutex_exit(mp);
1092 }
1093 }
1094
1095 /*
1096 * Detach shared memory segments from exiting process.
1097 */
1098 void
1099 shmexit(struct proc *pp)
1100 {
1101 segacct_t *sap;
1102 avl_tree_t *tree;
1103 void *cookie = NULL;
1104
1105 ASSERT(pp->p_segacct != NULL);
1106
1107 mutex_enter(&pp->p_lock);
1108 prbarrier(pp);
1109 tree = pp->p_segacct;
1110 pp->p_segacct = NULL;
1111 mutex_exit(&pp->p_lock);
1112
1113 while ((sap = avl_destroy_nodes(tree, &cookie)) != NULL)
1114 (void) shm_detach(pp, sap);
1115
1116 avl_destroy(tree);
1117 kmem_free(tree, sizeof (avl_tree_t));
1118 }
1119
1120 /*
1121 * At this time pages should be in memory, so just lock them.
1122 */
1123 static void
1124 lock_again(size_t npages, kshmid_t *sp, struct anon_map *amp)
1125 {
1126 struct anon *ap;
1127 struct page *pp;
1128 struct vnode *vp;
1129 u_offset_t off;
1130 ulong_t anon_idx;
1131 anon_sync_obj_t cookie;
1132
1133 mutex_enter(&sp->shm_mlock);
1134 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
1135 for (anon_idx = 0; npages != 0; anon_idx++, npages--) {
1136
1137 anon_array_enter(amp, anon_idx, &cookie);
1138 ap = anon_get_ptr(amp->ahp, anon_idx);
1139 ASSERT(ap != NULL);
1140 swap_xlate(ap, &vp, &off);
1141 anon_array_exit(&cookie);
1142
1143 pp = page_lookup(vp, off, SE_SHARED);
1144 if (pp == NULL) {
1145 panic("lock_again: page not in the system");
1146 /*NOTREACHED*/
1147 }
1148 /* page should already be locked by caller */
1149 ASSERT(pp->p_lckcnt > 0);
1150 (void) page_pp_lock(pp, 0, 0);
1151 page_unlock(pp);
1152 }
1153 ANON_LOCK_EXIT(&->a_rwlock);
1154 mutex_exit(&sp->shm_mlock);
1155 }
1156
1157 /*
1158 * Attach the shared memory segment to the process
1159 * address space and lock the pages.
1160 */
1161 static int
1162 shmem_lock(kshmid_t *sp, struct anon_map *amp)
1163 {
1164 size_t npages = btopr(amp->size);
1165 struct as *as;
1166 struct segvn_crargs crargs;
1167 uint_t error;
1168
1169 /*
1170 * A later ISM/DISM attach may increase the size of the amp, so
1171 * cache the number of pages locked for the future shmem_unlock()
1172 */
1173 sp->shm_lkpages = npages;
1174
1175 as = as_alloc();
1176 /* Initialize the create arguments and map the segment */
1177 crargs = *(struct segvn_crargs *)zfod_argsp; /* structure copy */
1178 crargs.offset = (u_offset_t)0;
1179 crargs.type = MAP_SHARED;
1180 crargs.amp = amp;
1181 crargs.prot = PROT_ALL;
1182 crargs.maxprot = crargs.prot;
1183 crargs.flags = 0;
1184 error = as_map(as, 0x0, amp->size, segvn_create, &crargs);
1185 if (!error) {
1186 if ((error = as_ctl(as, 0x0, amp->size, MC_LOCK, 0, 0,
1187 NULL, 0)) == 0) {
1188 lock_again(npages, sp, amp);
1189 }
1190 (void) as_unmap(as, 0x0, amp->size);
1191 }
1192 as_free(as);
1193 return (error);
1194 }
1195
1196
1197 /*
1198 * Unlock shared memory
1199 */
1200 static void
1201 shmem_unlock(kshmid_t *sp, struct anon_map *amp)
1202 {
1203 struct anon *ap;
1204 pgcnt_t npages = sp->shm_lkpages;
1205 struct vnode *vp;
1206 struct page *pp;
1207 u_offset_t off;
1208 ulong_t anon_idx;
1209 size_t unlocked_bytes = 0;
1210 kproject_t *proj;
1211 anon_sync_obj_t cookie;
1212
1213 proj = sp->shm_perm.ipc_proj;
1214 mutex_enter(&sp->shm_mlock);
1215 ANON_LOCK_ENTER(&->a_rwlock, RW_READER);
1216 for (anon_idx = 0; anon_idx < npages; anon_idx++) {
1217
1218 anon_array_enter(amp, anon_idx, &cookie);
1219 if ((ap = anon_get_ptr(amp->ahp, anon_idx)) == NULL) {
1220 panic("shmem_unlock: null app");
1221 /*NOTREACHED*/
1222 }
1223 swap_xlate(ap, &vp, &off);
1224 anon_array_exit(&cookie);
1225 pp = page_lookup(vp, off, SE_SHARED);
1226 if (pp == NULL) {
1227 panic("shmem_unlock: page not in the system");
1228 /*NOTREACHED*/
1229 }
1230 /*
1231 * Page should at least have once lock from previous
1232 * shmem_lock
1233 */
1234 ASSERT(pp->p_lckcnt > 0);
1235 page_pp_unlock(pp, 0, 0);
1236 if (pp->p_lckcnt == 0)
1237 unlocked_bytes += PAGESIZE;
1238
1239 page_unlock(pp);
1240 }
1241
1242 if (unlocked_bytes > 0) {
1243 rctl_decr_locked_mem(NULL, proj, unlocked_bytes, 0);
1244 }
1245
1246 ANON_LOCK_EXIT(&->a_rwlock);
1247 mutex_exit(&sp->shm_mlock);
1248 }
1249
1250 /*
1251 * We call this routine when we have removed all references to this
1252 * amp. This means all shmdt()s and the IPC_RMID have been done.
1253 */
1254 static void
1255 shm_rm_amp(kshmid_t *sp)
1256 {
1257 struct anon_map *amp = sp->shm_amp;
1258 zone_t *zone;
1259
1260 zone = sp->shm_perm.ipc_zone_ref.zref_zone;
1261 ASSERT(zone != NULL);
1262 /*
1263 * Free up the anon_map.
1264 */
1265 lgrp_shm_policy_fini(amp, NULL);
1266 ANON_LOCK_ENTER(&->a_rwlock, RW_WRITER);
1267 if (amp->a_szc != 0) {
1268 anon_shmap_free_pages(amp, 0, amp->size);
1269 } else {
1270 anon_free(amp->ahp, 0, amp->size);
1271 }
1272 ANON_LOCK_EXIT(&->a_rwlock);
1273 anon_unresv_zone(amp->swresv, zone);
1274 anonmap_free(amp);
1275 }
1276
1277 /*
1278 * Return the shared memory id for the process's virtual address.
1279 * Return SHMID_NONE if addr is not within a SysV shared memory segment.
1280 * Return SHMID_FREE if addr's SysV shared memory segment's id has been freed.
1281 *
1282 * shmgetid() is called from code in /proc with the process locked but
1283 * with pp->p_lock not held. The address space lock is held, so we
1284 * cannot grab pp->p_lock here due to lock-ordering constraints.
1285 * Because of all this, modifications to the p_segacct list must only
1286 * be made after calling prbarrier() to ensure the process is not locked.
1287 * See shmdt() and sa_add(), above. shmgetid() may also be called on a
1288 * thread's own process without the process locked.
1289 */
1290 int
1291 shmgetid(proc_t *pp, caddr_t addr)
1292 {
1293 segacct_t *sap, template;
1294
1295 ASSERT(MUTEX_NOT_HELD(&pp->p_lock));
1296 ASSERT((pp->p_proc_flag & P_PR_LOCK) || pp == curproc);
1297
1298 if (pp->p_segacct == NULL)
1299 return (SHMID_NONE);
1300
1301 template.sa_addr = addr;
1302 template.sa_len = 0;
1303 if ((sap = avl_find(pp->p_segacct, &template, NULL)) == NULL)
1304 return (SHMID_NONE);
1305
1306 if (IPC_FREE(&sap->sa_id->shm_perm))
1307 return (SHMID_FREE);
1308
1309 return (sap->sa_id->shm_perm.ipc_id);
1310 }