]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/kern/uipc_shm.c
Implement process-shared locks support for libthr.so.3, without
[FreeBSD/FreeBSD.git] / sys / kern / uipc_shm.c
1 /*-
2  * Copyright (c) 2006, 2011 Robert N. M. Watson
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26
27 /*
28  * Support for shared swap-backed anonymous memory objects via
29  * shm_open(2) and shm_unlink(2).  While most of the implementation is
30  * here, vm_mmap.c contains mapping logic changes.
31  *
32  * TODO:
33  *
34  * (1) Need to export data to a userland tool via a sysctl.  Should ipcs(1)
35  *     and ipcrm(1) be expanded or should new tools to manage both POSIX
36  *     kernel semaphores and POSIX shared memory be written?
37  *
38  * (2) Add support for this file type to fstat(1).
39  *
40  * (3) Resource limits?  Does this need its own resource limits or are the
41  *     existing limits in mmap(2) sufficient?
42  */
43
44 #include <sys/cdefs.h>
45 __FBSDID("$FreeBSD$");
46
47 #include "opt_capsicum.h"
48 #include "opt_ktrace.h"
49
50 #include <sys/param.h>
51 #include <sys/capsicum.h>
52 #include <sys/conf.h>
53 #include <sys/fcntl.h>
54 #include <sys/file.h>
55 #include <sys/filedesc.h>
56 #include <sys/fnv_hash.h>
57 #include <sys/kernel.h>
58 #include <sys/uio.h>
59 #include <sys/signal.h>
60 #include <sys/ktrace.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mman.h>
64 #include <sys/mutex.h>
65 #include <sys/priv.h>
66 #include <sys/proc.h>
67 #include <sys/refcount.h>
68 #include <sys/resourcevar.h>
69 #include <sys/rwlock.h>
70 #include <sys/stat.h>
71 #include <sys/syscallsubr.h>
72 #include <sys/sysctl.h>
73 #include <sys/sysproto.h>
74 #include <sys/systm.h>
75 #include <sys/sx.h>
76 #include <sys/time.h>
77 #include <sys/vnode.h>
78 #include <sys/unistd.h>
79 #include <sys/user.h>
80
81 #include <security/mac/mac_framework.h>
82
83 #include <vm/vm.h>
84 #include <vm/vm_param.h>
85 #include <vm/pmap.h>
86 #include <vm/vm_extern.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94
95 struct shm_mapping {
96         char            *sm_path;
97         Fnv32_t         sm_fnv;
98         struct shmfd    *sm_shmfd;
99         LIST_ENTRY(shm_mapping) sm_link;
100 };
101
102 static MALLOC_DEFINE(M_SHMFD, "shmfd", "shared memory file descriptor");
103 static LIST_HEAD(, shm_mapping) *shm_dictionary;
104 static struct sx shm_dict_lock;
105 static struct mtx shm_timestamp_lock;
106 static u_long shm_hash;
107 static struct unrhdr *shm_ino_unr;
108 static dev_t shm_dev_ino;
109
110 #define SHM_HASH(fnv)   (&shm_dictionary[(fnv) & shm_hash])
111
112 static void     shm_init(void *arg);
113 static void     shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd);
114 static struct shmfd *shm_lookup(char *path, Fnv32_t fnv);
115 static int      shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred);
116
117 static fo_rdwr_t        shm_read;
118 static fo_rdwr_t        shm_write;
119 static fo_truncate_t    shm_truncate;
120 static fo_stat_t        shm_stat;
121 static fo_close_t       shm_close;
122 static fo_chmod_t       shm_chmod;
123 static fo_chown_t       shm_chown;
124 static fo_seek_t        shm_seek;
125 static fo_fill_kinfo_t  shm_fill_kinfo;
126 static fo_mmap_t        shm_mmap;
127
128 /* File descriptor operations. */
129 struct fileops shm_ops = {
130         .fo_read = shm_read,
131         .fo_write = shm_write,
132         .fo_truncate = shm_truncate,
133         .fo_ioctl = invfo_ioctl,
134         .fo_poll = invfo_poll,
135         .fo_kqfilter = invfo_kqfilter,
136         .fo_stat = shm_stat,
137         .fo_close = shm_close,
138         .fo_chmod = shm_chmod,
139         .fo_chown = shm_chown,
140         .fo_sendfile = vn_sendfile,
141         .fo_seek = shm_seek,
142         .fo_fill_kinfo = shm_fill_kinfo,
143         .fo_mmap = shm_mmap,
144         .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
145 };
146
147 FEATURE(posix_shm, "POSIX shared memory");
148
149 static int
150 uiomove_object_page(vm_object_t obj, size_t len, struct uio *uio)
151 {
152         vm_page_t m;
153         vm_pindex_t idx;
154         size_t tlen;
155         int error, offset, rv;
156
157         idx = OFF_TO_IDX(uio->uio_offset);
158         offset = uio->uio_offset & PAGE_MASK;
159         tlen = MIN(PAGE_SIZE - offset, len);
160
161         VM_OBJECT_WLOCK(obj);
162
163         /*
164          * Read I/O without either a corresponding resident page or swap
165          * page: use zero_region.  This is intended to avoid instantiating
166          * pages on read from a sparse region.
167          */
168         if (uio->uio_rw == UIO_READ && vm_page_lookup(obj, idx) == NULL &&
169             !vm_pager_has_page(obj, idx, NULL, NULL)) {
170                 VM_OBJECT_WUNLOCK(obj);
171                 return (uiomove(__DECONST(void *, zero_region), tlen, uio));
172         }
173
174         /*
175          * Parallel reads of the page content from disk are prevented
176          * by exclusive busy.
177          *
178          * Although the tmpfs vnode lock is held here, it is
179          * nonetheless safe to sleep waiting for a free page.  The
180          * pageout daemon does not need to acquire the tmpfs vnode
181          * lock to page out tobj's pages because tobj is a OBJT_SWAP
182          * type object.
183          */
184         m = vm_page_grab(obj, idx, VM_ALLOC_NORMAL);
185         if (m->valid != VM_PAGE_BITS_ALL) {
186                 if (vm_pager_has_page(obj, idx, NULL, NULL)) {
187                         rv = vm_pager_get_pages(obj, &m, 1, NULL, NULL);
188                         if (rv != VM_PAGER_OK) {
189                                 printf(
190             "uiomove_object: vm_obj %p idx %jd valid %x pager error %d\n",
191                                     obj, idx, m->valid, rv);
192                                 vm_page_lock(m);
193                                 vm_page_free(m);
194                                 vm_page_unlock(m);
195                                 VM_OBJECT_WUNLOCK(obj);
196                                 return (EIO);
197                         }
198                 } else
199                         vm_page_zero_invalid(m, TRUE);
200         }
201         vm_page_xunbusy(m);
202         vm_page_lock(m);
203         vm_page_hold(m);
204         if (m->queue == PQ_NONE) {
205                 vm_page_deactivate(m);
206         } else {
207                 /* Requeue to maintain LRU ordering. */
208                 vm_page_requeue(m);
209         }
210         vm_page_unlock(m);
211         VM_OBJECT_WUNLOCK(obj);
212         error = uiomove_fromphys(&m, offset, tlen, uio);
213         if (uio->uio_rw == UIO_WRITE && error == 0) {
214                 VM_OBJECT_WLOCK(obj);
215                 vm_page_dirty(m);
216                 vm_pager_page_unswapped(m);
217                 VM_OBJECT_WUNLOCK(obj);
218         }
219         vm_page_lock(m);
220         vm_page_unhold(m);
221         vm_page_unlock(m);
222
223         return (error);
224 }
225
226 int
227 uiomove_object(vm_object_t obj, off_t obj_size, struct uio *uio)
228 {
229         ssize_t resid;
230         size_t len;
231         int error;
232
233         error = 0;
234         while ((resid = uio->uio_resid) > 0) {
235                 if (obj_size <= uio->uio_offset)
236                         break;
237                 len = MIN(obj_size - uio->uio_offset, resid);
238                 if (len == 0)
239                         break;
240                 error = uiomove_object_page(obj, len, uio);
241                 if (error != 0 || resid == uio->uio_resid)
242                         break;
243         }
244         return (error);
245 }
246
247 static int
248 shm_seek(struct file *fp, off_t offset, int whence, struct thread *td)
249 {
250         struct shmfd *shmfd;
251         off_t foffset;
252         int error;
253
254         shmfd = fp->f_data;
255         foffset = foffset_lock(fp, 0);
256         error = 0;
257         switch (whence) {
258         case L_INCR:
259                 if (foffset < 0 ||
260                     (offset > 0 && foffset > OFF_MAX - offset)) {
261                         error = EOVERFLOW;
262                         break;
263                 }
264                 offset += foffset;
265                 break;
266         case L_XTND:
267                 if (offset > 0 && shmfd->shm_size > OFF_MAX - offset) {
268                         error = EOVERFLOW;
269                         break;
270                 }
271                 offset += shmfd->shm_size;
272                 break;
273         case L_SET:
274                 break;
275         default:
276                 error = EINVAL;
277         }
278         if (error == 0) {
279                 if (offset < 0 || offset > shmfd->shm_size)
280                         error = EINVAL;
281                 else
282                         td->td_uretoff.tdu_off = offset;
283         }
284         foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
285         return (error);
286 }
287
288 static int
289 shm_read(struct file *fp, struct uio *uio, struct ucred *active_cred,
290     int flags, struct thread *td)
291 {
292         struct shmfd *shmfd;
293         void *rl_cookie;
294         int error;
295
296         shmfd = fp->f_data;
297         foffset_lock_uio(fp, uio, flags);
298         rl_cookie = rangelock_rlock(&shmfd->shm_rl, uio->uio_offset,
299             uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
300 #ifdef MAC
301         error = mac_posixshm_check_read(active_cred, fp->f_cred, shmfd);
302         if (error)
303                 return (error);
304 #endif
305         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
306         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
307         foffset_unlock_uio(fp, uio, flags);
308         return (error);
309 }
310
311 static int
312 shm_write(struct file *fp, struct uio *uio, struct ucred *active_cred,
313     int flags, struct thread *td)
314 {
315         struct shmfd *shmfd;
316         void *rl_cookie;
317         int error;
318
319         shmfd = fp->f_data;
320 #ifdef MAC
321         error = mac_posixshm_check_write(active_cred, fp->f_cred, shmfd);
322         if (error)
323                 return (error);
324 #endif
325         foffset_lock_uio(fp, uio, flags);
326         if ((flags & FOF_OFFSET) == 0) {
327                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, 0, OFF_MAX,
328                     &shmfd->shm_mtx);
329         } else {
330                 rl_cookie = rangelock_wlock(&shmfd->shm_rl, uio->uio_offset,
331                     uio->uio_offset + uio->uio_resid, &shmfd->shm_mtx);
332         }
333
334         error = uiomove_object(shmfd->shm_object, shmfd->shm_size, uio);
335         rangelock_unlock(&shmfd->shm_rl, rl_cookie, &shmfd->shm_mtx);
336         foffset_unlock_uio(fp, uio, flags);
337         return (error);
338 }
339
340 static int
341 shm_truncate(struct file *fp, off_t length, struct ucred *active_cred,
342     struct thread *td)
343 {
344         struct shmfd *shmfd;
345 #ifdef MAC
346         int error;
347 #endif
348
349         shmfd = fp->f_data;
350 #ifdef MAC
351         error = mac_posixshm_check_truncate(active_cred, fp->f_cred, shmfd);
352         if (error)
353                 return (error);
354 #endif
355         return (shm_dotruncate(shmfd, length));
356 }
357
358 static int
359 shm_stat(struct file *fp, struct stat *sb, struct ucred *active_cred,
360     struct thread *td)
361 {
362         struct shmfd *shmfd;
363 #ifdef MAC
364         int error;
365 #endif
366
367         shmfd = fp->f_data;
368
369 #ifdef MAC
370         error = mac_posixshm_check_stat(active_cred, fp->f_cred, shmfd);
371         if (error)
372                 return (error);
373 #endif
374         
375         /*
376          * Attempt to return sanish values for fstat() on a memory file
377          * descriptor.
378          */
379         bzero(sb, sizeof(*sb));
380         sb->st_blksize = PAGE_SIZE;
381         sb->st_size = shmfd->shm_size;
382         sb->st_blocks = (sb->st_size + sb->st_blksize - 1) / sb->st_blksize;
383         mtx_lock(&shm_timestamp_lock);
384         sb->st_atim = shmfd->shm_atime;
385         sb->st_ctim = shmfd->shm_ctime;
386         sb->st_mtim = shmfd->shm_mtime;
387         sb->st_birthtim = shmfd->shm_birthtime;
388         sb->st_mode = S_IFREG | shmfd->shm_mode;                /* XXX */
389         sb->st_uid = shmfd->shm_uid;
390         sb->st_gid = shmfd->shm_gid;
391         mtx_unlock(&shm_timestamp_lock);
392         sb->st_dev = shm_dev_ino;
393         sb->st_ino = shmfd->shm_ino;
394
395         return (0);
396 }
397
398 static int
399 shm_close(struct file *fp, struct thread *td)
400 {
401         struct shmfd *shmfd;
402
403         shmfd = fp->f_data;
404         fp->f_data = NULL;
405         shm_drop(shmfd);
406
407         return (0);
408 }
409
410 int
411 shm_dotruncate(struct shmfd *shmfd, off_t length)
412 {
413         vm_object_t object;
414         vm_page_t m;
415         vm_pindex_t idx, nobjsize;
416         vm_ooffset_t delta;
417         int base, rv;
418
419         object = shmfd->shm_object;
420         VM_OBJECT_WLOCK(object);
421         if (length == shmfd->shm_size) {
422                 VM_OBJECT_WUNLOCK(object);
423                 return (0);
424         }
425         nobjsize = OFF_TO_IDX(length + PAGE_MASK);
426
427         /* Are we shrinking?  If so, trim the end. */
428         if (length < shmfd->shm_size) {
429                 /*
430                  * Disallow any requests to shrink the size if this
431                  * object is mapped into the kernel.
432                  */
433                 if (shmfd->shm_kmappings > 0) {
434                         VM_OBJECT_WUNLOCK(object);
435                         return (EBUSY);
436                 }
437
438                 /*
439                  * Zero the truncated part of the last page.
440                  */
441                 base = length & PAGE_MASK;
442                 if (base != 0) {
443                         idx = OFF_TO_IDX(length);
444 retry:
445                         m = vm_page_lookup(object, idx);
446                         if (m != NULL) {
447                                 if (vm_page_sleep_if_busy(m, "shmtrc"))
448                                         goto retry;
449                         } else if (vm_pager_has_page(object, idx, NULL, NULL)) {
450                                 m = vm_page_alloc(object, idx, VM_ALLOC_NORMAL);
451                                 if (m == NULL) {
452                                         VM_OBJECT_WUNLOCK(object);
453                                         VM_WAIT;
454                                         VM_OBJECT_WLOCK(object);
455                                         goto retry;
456                                 } else if (m->valid != VM_PAGE_BITS_ALL)
457                                         rv = vm_pager_get_pages(object, &m, 1,
458                                             NULL, NULL);
459                                 else
460                                         /* A cached page was reactivated. */
461                                         rv = VM_PAGER_OK;
462                                 vm_page_lock(m);
463                                 if (rv == VM_PAGER_OK) {
464                                         vm_page_deactivate(m);
465                                         vm_page_unlock(m);
466                                         vm_page_xunbusy(m);
467                                 } else {
468                                         vm_page_free(m);
469                                         vm_page_unlock(m);
470                                         VM_OBJECT_WUNLOCK(object);
471                                         return (EIO);
472                                 }
473                         }
474                         if (m != NULL) {
475                                 pmap_zero_page_area(m, base, PAGE_SIZE - base);
476                                 KASSERT(m->valid == VM_PAGE_BITS_ALL,
477                                     ("shm_dotruncate: page %p is invalid", m));
478                                 vm_page_dirty(m);
479                                 vm_pager_page_unswapped(m);
480                         }
481                 }
482                 delta = ptoa(object->size - nobjsize);
483
484                 /* Toss in memory pages. */
485                 if (nobjsize < object->size)
486                         vm_object_page_remove(object, nobjsize, object->size,
487                             0);
488
489                 /* Toss pages from swap. */
490                 if (object->type == OBJT_SWAP)
491                         swap_pager_freespace(object, nobjsize, delta);
492
493                 /* Free the swap accounted for shm */
494                 swap_release_by_cred(delta, object->cred);
495                 object->charge -= delta;
496         } else {
497                 /* Attempt to reserve the swap */
498                 delta = ptoa(nobjsize - object->size);
499                 if (!swap_reserve_by_cred(delta, object->cred)) {
500                         VM_OBJECT_WUNLOCK(object);
501                         return (ENOMEM);
502                 }
503                 object->charge += delta;
504         }
505         shmfd->shm_size = length;
506         mtx_lock(&shm_timestamp_lock);
507         vfs_timestamp(&shmfd->shm_ctime);
508         shmfd->shm_mtime = shmfd->shm_ctime;
509         mtx_unlock(&shm_timestamp_lock);
510         object->size = nobjsize;
511         VM_OBJECT_WUNLOCK(object);
512         return (0);
513 }
514
515 /*
516  * shmfd object management including creation and reference counting
517  * routines.
518  */
519 struct shmfd *
520 shm_alloc(struct ucred *ucred, mode_t mode)
521 {
522         struct shmfd *shmfd;
523         int ino;
524
525         shmfd = malloc(sizeof(*shmfd), M_SHMFD, M_WAITOK | M_ZERO);
526         shmfd->shm_size = 0;
527         shmfd->shm_uid = ucred->cr_uid;
528         shmfd->shm_gid = ucred->cr_gid;
529         shmfd->shm_mode = mode;
530         shmfd->shm_object = vm_pager_allocate(OBJT_DEFAULT, NULL,
531             shmfd->shm_size, VM_PROT_DEFAULT, 0, ucred);
532         KASSERT(shmfd->shm_object != NULL, ("shm_create: vm_pager_allocate"));
533         shmfd->shm_object->pg_color = 0;
534         VM_OBJECT_WLOCK(shmfd->shm_object);
535         vm_object_clear_flag(shmfd->shm_object, OBJ_ONEMAPPING);
536         vm_object_set_flag(shmfd->shm_object, OBJ_COLORED | OBJ_NOSPLIT);
537         VM_OBJECT_WUNLOCK(shmfd->shm_object);
538         vfs_timestamp(&shmfd->shm_birthtime);
539         shmfd->shm_atime = shmfd->shm_mtime = shmfd->shm_ctime =
540             shmfd->shm_birthtime;
541         ino = alloc_unr(shm_ino_unr);
542         if (ino == -1)
543                 shmfd->shm_ino = 0;
544         else
545                 shmfd->shm_ino = ino;
546         refcount_init(&shmfd->shm_refs, 1);
547         mtx_init(&shmfd->shm_mtx, "shmrl", NULL, MTX_DEF);
548         rangelock_init(&shmfd->shm_rl);
549 #ifdef MAC
550         mac_posixshm_init(shmfd);
551         mac_posixshm_create(ucred, shmfd);
552 #endif
553
554         return (shmfd);
555 }
556
557 struct shmfd *
558 shm_hold(struct shmfd *shmfd)
559 {
560
561         refcount_acquire(&shmfd->shm_refs);
562         return (shmfd);
563 }
564
565 void
566 shm_drop(struct shmfd *shmfd)
567 {
568
569         if (refcount_release(&shmfd->shm_refs)) {
570 #ifdef MAC
571                 mac_posixshm_destroy(shmfd);
572 #endif
573                 rangelock_destroy(&shmfd->shm_rl);
574                 mtx_destroy(&shmfd->shm_mtx);
575                 vm_object_deallocate(shmfd->shm_object);
576                 if (shmfd->shm_ino != 0)
577                         free_unr(shm_ino_unr, shmfd->shm_ino);
578                 free(shmfd, M_SHMFD);
579         }
580 }
581
582 /*
583  * Determine if the credentials have sufficient permissions for a
584  * specified combination of FREAD and FWRITE.
585  */
586 int
587 shm_access(struct shmfd *shmfd, struct ucred *ucred, int flags)
588 {
589         accmode_t accmode;
590         int error;
591
592         accmode = 0;
593         if (flags & FREAD)
594                 accmode |= VREAD;
595         if (flags & FWRITE)
596                 accmode |= VWRITE;
597         mtx_lock(&shm_timestamp_lock);
598         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid, shmfd->shm_gid,
599             accmode, ucred, NULL);
600         mtx_unlock(&shm_timestamp_lock);
601         return (error);
602 }
603
604 /*
605  * Dictionary management.  We maintain an in-kernel dictionary to map
606  * paths to shmfd objects.  We use the FNV hash on the path to store
607  * the mappings in a hash table.
608  */
609 static void
610 shm_init(void *arg)
611 {
612
613         mtx_init(&shm_timestamp_lock, "shm timestamps", NULL, MTX_DEF);
614         sx_init(&shm_dict_lock, "shm dictionary");
615         shm_dictionary = hashinit(1024, M_SHMFD, &shm_hash);
616         shm_ino_unr = new_unrhdr(1, INT32_MAX, NULL);
617         KASSERT(shm_ino_unr != NULL, ("shm fake inodes not initialized"));
618         shm_dev_ino = devfs_alloc_cdp_inode();
619         KASSERT(shm_dev_ino > 0, ("shm dev inode not initialized"));
620 }
621 SYSINIT(shm_init, SI_SUB_SYSV_SHM, SI_ORDER_ANY, shm_init, NULL);
622
623 static struct shmfd *
624 shm_lookup(char *path, Fnv32_t fnv)
625 {
626         struct shm_mapping *map;
627
628         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
629                 if (map->sm_fnv != fnv)
630                         continue;
631                 if (strcmp(map->sm_path, path) == 0)
632                         return (map->sm_shmfd);
633         }
634
635         return (NULL);
636 }
637
638 static void
639 shm_insert(char *path, Fnv32_t fnv, struct shmfd *shmfd)
640 {
641         struct shm_mapping *map;
642
643         map = malloc(sizeof(struct shm_mapping), M_SHMFD, M_WAITOK);
644         map->sm_path = path;
645         map->sm_fnv = fnv;
646         map->sm_shmfd = shm_hold(shmfd);
647         shmfd->shm_path = path;
648         LIST_INSERT_HEAD(SHM_HASH(fnv), map, sm_link);
649 }
650
651 static int
652 shm_remove(char *path, Fnv32_t fnv, struct ucred *ucred)
653 {
654         struct shm_mapping *map;
655         int error;
656
657         LIST_FOREACH(map, SHM_HASH(fnv), sm_link) {
658                 if (map->sm_fnv != fnv)
659                         continue;
660                 if (strcmp(map->sm_path, path) == 0) {
661 #ifdef MAC
662                         error = mac_posixshm_check_unlink(ucred, map->sm_shmfd);
663                         if (error)
664                                 return (error);
665 #endif
666                         error = shm_access(map->sm_shmfd, ucred,
667                             FREAD | FWRITE);
668                         if (error)
669                                 return (error);
670                         map->sm_shmfd->shm_path = NULL;
671                         LIST_REMOVE(map, sm_link);
672                         shm_drop(map->sm_shmfd);
673                         free(map->sm_path, M_SHMFD);
674                         free(map, M_SHMFD);
675                         return (0);
676                 }
677         }
678
679         return (ENOENT);
680 }
681
682 int
683 kern_shm_open(struct thread *td, const char *userpath, int flags, mode_t mode,
684     struct filecaps *fcaps)
685 {
686         struct filedesc *fdp;
687         struct shmfd *shmfd;
688         struct file *fp;
689         char *path;
690         Fnv32_t fnv;
691         mode_t cmode;
692         int fd, error;
693
694 #ifdef CAPABILITY_MODE
695         /*
696          * shm_open(2) is only allowed for anonymous objects.
697          */
698         if (IN_CAPABILITY_MODE(td) && (userpath != SHM_ANON))
699                 return (ECAPMODE);
700 #endif
701
702         if ((flags & O_ACCMODE) != O_RDONLY && (flags & O_ACCMODE) != O_RDWR)
703                 return (EINVAL);
704
705         if ((flags & ~(O_ACCMODE | O_CREAT | O_EXCL | O_TRUNC | O_CLOEXEC)) != 0)
706                 return (EINVAL);
707
708         fdp = td->td_proc->p_fd;
709         cmode = (mode & ~fdp->fd_cmask) & ACCESSPERMS;
710
711         error = falloc_caps(td, &fp, &fd, O_CLOEXEC, fcaps);
712         if (error)
713                 return (error);
714
715         /* A SHM_ANON path pointer creates an anonymous object. */
716         if (userpath == SHM_ANON) {
717                 /* A read-only anonymous object is pointless. */
718                 if ((flags & O_ACCMODE) == O_RDONLY) {
719                         fdclose(td, fp, fd);
720                         fdrop(fp, td);
721                         return (EINVAL);
722                 }
723                 shmfd = shm_alloc(td->td_ucred, cmode);
724         } else {
725                 path = malloc(MAXPATHLEN, M_SHMFD, M_WAITOK);
726                 error = copyinstr(userpath, path, MAXPATHLEN, NULL);
727 #ifdef KTRACE
728                 if (error == 0 && KTRPOINT(curthread, KTR_NAMEI))
729                         ktrnamei(path);
730 #endif
731                 /* Require paths to start with a '/' character. */
732                 if (error == 0 && path[0] != '/')
733                         error = EINVAL;
734                 if (error) {
735                         fdclose(td, fp, fd);
736                         fdrop(fp, td);
737                         free(path, M_SHMFD);
738                         return (error);
739                 }
740
741                 fnv = fnv_32_str(path, FNV1_32_INIT);
742                 sx_xlock(&shm_dict_lock);
743                 shmfd = shm_lookup(path, fnv);
744                 if (shmfd == NULL) {
745                         /* Object does not yet exist, create it if requested. */
746                         if (flags & O_CREAT) {
747 #ifdef MAC
748                                 error = mac_posixshm_check_create(td->td_ucred,
749                                     path);
750                                 if (error == 0) {
751 #endif
752                                         shmfd = shm_alloc(td->td_ucred, cmode);
753                                         shm_insert(path, fnv, shmfd);
754 #ifdef MAC
755                                 }
756 #endif
757                         } else {
758                                 free(path, M_SHMFD);
759                                 error = ENOENT;
760                         }
761                 } else {
762                         /*
763                          * Object already exists, obtain a new
764                          * reference if requested and permitted.
765                          */
766                         free(path, M_SHMFD);
767                         if ((flags & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL))
768                                 error = EEXIST;
769                         else {
770 #ifdef MAC
771                                 error = mac_posixshm_check_open(td->td_ucred,
772                                     shmfd, FFLAGS(flags & O_ACCMODE));
773                                 if (error == 0)
774 #endif
775                                 error = shm_access(shmfd, td->td_ucred,
776                                     FFLAGS(flags & O_ACCMODE));
777                         }
778
779                         /*
780                          * Truncate the file back to zero length if
781                          * O_TRUNC was specified and the object was
782                          * opened with read/write.
783                          */
784                         if (error == 0 &&
785                             (flags & (O_ACCMODE | O_TRUNC)) ==
786                             (O_RDWR | O_TRUNC)) {
787 #ifdef MAC
788                                 error = mac_posixshm_check_truncate(
789                                         td->td_ucred, fp->f_cred, shmfd);
790                                 if (error == 0)
791 #endif
792                                         shm_dotruncate(shmfd, 0);
793                         }
794                         if (error == 0)
795                                 shm_hold(shmfd);
796                 }
797                 sx_xunlock(&shm_dict_lock);
798
799                 if (error) {
800                         fdclose(td, fp, fd);
801                         fdrop(fp, td);
802                         return (error);
803                 }
804         }
805
806         finit(fp, FFLAGS(flags & O_ACCMODE), DTYPE_SHM, shmfd, &shm_ops);
807
808         td->td_retval[0] = fd;
809         fdrop(fp, td);
810
811         return (0);
812 }
813
814 /* System calls. */
815 int
816 sys_shm_open(struct thread *td, struct shm_open_args *uap)
817 {
818
819         return (kern_shm_open(td, uap->path, uap->flags, uap->mode, NULL));
820 }
821
822 int
823 sys_shm_unlink(struct thread *td, struct shm_unlink_args *uap)
824 {
825         char *path;
826         Fnv32_t fnv;
827         int error;
828
829         path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
830         error = copyinstr(uap->path, path, MAXPATHLEN, NULL);
831         if (error) {
832                 free(path, M_TEMP);
833                 return (error);
834         }
835 #ifdef KTRACE
836         if (KTRPOINT(curthread, KTR_NAMEI))
837                 ktrnamei(path);
838 #endif
839         fnv = fnv_32_str(path, FNV1_32_INIT);
840         sx_xlock(&shm_dict_lock);
841         error = shm_remove(path, fnv, td->td_ucred);
842         sx_xunlock(&shm_dict_lock);
843         free(path, M_TEMP);
844
845         return (error);
846 }
847
848 int
849 shm_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t objsize,
850     vm_prot_t prot, vm_prot_t cap_maxprot, int flags,
851     vm_ooffset_t foff, struct thread *td)
852 {
853         struct shmfd *shmfd;
854         vm_prot_t maxprot;
855         int error;
856
857         shmfd = fp->f_data;
858         maxprot = VM_PROT_NONE;
859
860         /* FREAD should always be set. */
861         if ((fp->f_flag & FREAD) != 0)
862                 maxprot |= VM_PROT_EXECUTE | VM_PROT_READ;
863         if ((fp->f_flag & FWRITE) != 0)
864                 maxprot |= VM_PROT_WRITE;
865
866         /* Don't permit shared writable mappings on read-only descriptors. */
867         if ((flags & MAP_SHARED) != 0 &&
868             (maxprot & VM_PROT_WRITE) == 0 &&
869             (prot & VM_PROT_WRITE) != 0)
870                 return (EACCES);
871         maxprot &= cap_maxprot;
872
873 #ifdef MAC
874         error = mac_posixshm_check_mmap(td->td_ucred, shmfd, prot, flags);
875         if (error != 0)
876                 return (error);
877 #endif
878         
879         /*
880          * XXXRW: This validation is probably insufficient, and subject to
881          * sign errors.  It should be fixed.
882          */
883         if (foff >= shmfd->shm_size ||
884             foff + objsize > round_page(shmfd->shm_size))
885                 return (EINVAL);
886
887         mtx_lock(&shm_timestamp_lock);
888         vfs_timestamp(&shmfd->shm_atime);
889         mtx_unlock(&shm_timestamp_lock);
890         vm_object_reference(shmfd->shm_object);
891
892         error = vm_mmap_object(map, addr, objsize, prot, maxprot, flags,
893             shmfd->shm_object, foff, FALSE, td);
894         if (error != 0)
895                 vm_object_deallocate(shmfd->shm_object);
896         return (0);
897 }
898
899 static int
900 shm_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
901     struct thread *td)
902 {
903         struct shmfd *shmfd;
904         int error;
905
906         error = 0;
907         shmfd = fp->f_data;
908         mtx_lock(&shm_timestamp_lock);
909         /*
910          * SUSv4 says that x bits of permission need not be affected.
911          * Be consistent with our shm_open there.
912          */
913 #ifdef MAC
914         error = mac_posixshm_check_setmode(active_cred, shmfd, mode);
915         if (error != 0)
916                 goto out;
917 #endif
918         error = vaccess(VREG, shmfd->shm_mode, shmfd->shm_uid,
919             shmfd->shm_gid, VADMIN, active_cred, NULL);
920         if (error != 0)
921                 goto out;
922         shmfd->shm_mode = mode & ACCESSPERMS;
923 out:
924         mtx_unlock(&shm_timestamp_lock);
925         return (error);
926 }
927
928 static int
929 shm_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
930     struct thread *td)
931 {
932         struct shmfd *shmfd;
933         int error;
934
935         error = 0;
936         shmfd = fp->f_data;
937         mtx_lock(&shm_timestamp_lock);
938 #ifdef MAC
939         error = mac_posixshm_check_setowner(active_cred, shmfd, uid, gid);
940         if (error != 0)
941                 goto out;
942 #endif
943         if (uid == (uid_t)-1)
944                 uid = shmfd->shm_uid;
945         if (gid == (gid_t)-1)
946                  gid = shmfd->shm_gid;
947         if (((uid != shmfd->shm_uid && uid != active_cred->cr_uid) ||
948             (gid != shmfd->shm_gid && !groupmember(gid, active_cred))) &&
949             (error = priv_check_cred(active_cred, PRIV_VFS_CHOWN, 0)))
950                 goto out;
951         shmfd->shm_uid = uid;
952         shmfd->shm_gid = gid;
953 out:
954         mtx_unlock(&shm_timestamp_lock);
955         return (error);
956 }
957
958 /*
959  * Helper routines to allow the backing object of a shared memory file
960  * descriptor to be mapped in the kernel.
961  */
962 int
963 shm_map(struct file *fp, size_t size, off_t offset, void **memp)
964 {
965         struct shmfd *shmfd;
966         vm_offset_t kva, ofs;
967         vm_object_t obj;
968         int rv;
969
970         if (fp->f_type != DTYPE_SHM)
971                 return (EINVAL);
972         shmfd = fp->f_data;
973         obj = shmfd->shm_object;
974         VM_OBJECT_WLOCK(obj);
975         /*
976          * XXXRW: This validation is probably insufficient, and subject to
977          * sign errors.  It should be fixed.
978          */
979         if (offset >= shmfd->shm_size ||
980             offset + size > round_page(shmfd->shm_size)) {
981                 VM_OBJECT_WUNLOCK(obj);
982                 return (EINVAL);
983         }
984
985         shmfd->shm_kmappings++;
986         vm_object_reference_locked(obj);
987         VM_OBJECT_WUNLOCK(obj);
988
989         /* Map the object into the kernel_map and wire it. */
990         kva = vm_map_min(kernel_map);
991         ofs = offset & PAGE_MASK;
992         offset = trunc_page(offset);
993         size = round_page(size + ofs);
994         rv = vm_map_find(kernel_map, obj, offset, &kva, size, 0,
995             VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
996             VM_PROT_READ | VM_PROT_WRITE, 0);
997         if (rv == KERN_SUCCESS) {
998                 rv = vm_map_wire(kernel_map, kva, kva + size,
999                     VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
1000                 if (rv == KERN_SUCCESS) {
1001                         *memp = (void *)(kva + ofs);
1002                         return (0);
1003                 }
1004                 vm_map_remove(kernel_map, kva, kva + size);
1005         } else
1006                 vm_object_deallocate(obj);
1007
1008         /* On failure, drop our mapping reference. */
1009         VM_OBJECT_WLOCK(obj);
1010         shmfd->shm_kmappings--;
1011         VM_OBJECT_WUNLOCK(obj);
1012
1013         return (vm_mmap_to_errno(rv));
1014 }
1015
1016 /*
1017  * We require the caller to unmap the entire entry.  This allows us to
1018  * safely decrement shm_kmappings when a mapping is removed.
1019  */
1020 int
1021 shm_unmap(struct file *fp, void *mem, size_t size)
1022 {
1023         struct shmfd *shmfd;
1024         vm_map_entry_t entry;
1025         vm_offset_t kva, ofs;
1026         vm_object_t obj;
1027         vm_pindex_t pindex;
1028         vm_prot_t prot;
1029         boolean_t wired;
1030         vm_map_t map;
1031         int rv;
1032
1033         if (fp->f_type != DTYPE_SHM)
1034                 return (EINVAL);
1035         shmfd = fp->f_data;
1036         kva = (vm_offset_t)mem;
1037         ofs = kva & PAGE_MASK;
1038         kva = trunc_page(kva);
1039         size = round_page(size + ofs);
1040         map = kernel_map;
1041         rv = vm_map_lookup(&map, kva, VM_PROT_READ | VM_PROT_WRITE, &entry,
1042             &obj, &pindex, &prot, &wired);
1043         if (rv != KERN_SUCCESS)
1044                 return (EINVAL);
1045         if (entry->start != kva || entry->end != kva + size) {
1046                 vm_map_lookup_done(map, entry);
1047                 return (EINVAL);
1048         }
1049         vm_map_lookup_done(map, entry);
1050         if (obj != shmfd->shm_object)
1051                 return (EINVAL);
1052         vm_map_remove(map, kva, kva + size);
1053         VM_OBJECT_WLOCK(obj);
1054         KASSERT(shmfd->shm_kmappings > 0, ("shm_unmap: object not mapped"));
1055         shmfd->shm_kmappings--;
1056         VM_OBJECT_WUNLOCK(obj);
1057         return (0);
1058 }
1059
1060 static int
1061 shm_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
1062 {
1063         struct shmfd *shmfd;
1064
1065         kif->kf_type = KF_TYPE_SHM;
1066         shmfd = fp->f_data;
1067
1068         mtx_lock(&shm_timestamp_lock);
1069         kif->kf_un.kf_file.kf_file_mode = S_IFREG | shmfd->shm_mode;    /* XXX */
1070         mtx_unlock(&shm_timestamp_lock);
1071         kif->kf_un.kf_file.kf_file_size = shmfd->shm_size;
1072         if (shmfd->shm_path != NULL) {
1073                 sx_slock(&shm_dict_lock);
1074                 if (shmfd->shm_path != NULL)
1075                         strlcpy(kif->kf_path, shmfd->shm_path,
1076                             sizeof(kif->kf_path));
1077                 sx_sunlock(&shm_dict_lock);
1078         }
1079         return (0);
1080 }