0acf2abccd433cf64d8835a33311cfcc714231a0
[dragonfly.git] / sys / kern / sysv_shm.c
1 /*
2  * Copyright (c) 1994 Adam Glass and Charles Hannum.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  * 3. All advertising materials mentioning features or use of this software
13  *    must display the following acknowledgement:
14  *      This product includes software developed by Adam Glass and Charles
15  *      Hannum.
16  * 4. The names of the authors may not be used to endorse or promote products
17  *    derived from this software without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR
20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22  * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT,
23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30
31 #include "opt_sysvipc.h"
32
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/sysproto.h>
36 #include <sys/kernel.h>
37 #include <sys/sysctl.h>
38 #include <sys/shm.h>
39 #include <sys/proc.h>
40 #include <sys/malloc.h>
41 #include <sys/mman.h>
42 #include <sys/stat.h>
43 #include <sys/sysent.h>
44 #include <sys/jail.h>
45
46 #include <vm/vm.h>
47 #include <vm/vm_param.h>
48 #include <sys/lock.h>
49 #include <vm/pmap.h>
50 #include <vm/vm_object.h>
51 #include <vm/vm_map.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_pager.h>
54
55 static MALLOC_DEFINE(M_SHM, "shm", "SVID compatible shared memory segments");
56
57 static int shmget_allocate_segment (struct proc *p, struct shmget_args *uap, int mode);
58 static int shmget_existing (struct proc *p, struct shmget_args *uap, int mode, int segnum);
59
60 #define SHMSEG_FREE             0x0200
61 #define SHMSEG_REMOVED          0x0400
62 #define SHMSEG_ALLOCATED        0x0800
63 #define SHMSEG_WANTED           0x1000
64
65 static int shm_last_free, shm_committed, shmalloced;
66 int shm_nused;
67 static struct shmid_ds  *shmsegs;
68 static struct lwkt_token shm_token = LWKT_TOKEN_INITIALIZER(shm_token);
69
70 struct shm_handle {
71         /* vm_offset_t kva; */
72         vm_object_t shm_object;
73 };
74
75 struct shmmap_state {
76         vm_offset_t va;
77         int shmid;
78         int reserved;
79 };
80
81 static void shm_deallocate_segment (struct shmid_ds *);
82 static int shm_find_segment_by_key (key_t);
83 static struct shmid_ds *shm_find_segment_by_shmid (int);
84 static int shm_delete_mapping (struct vmspace *vm, struct shmmap_state *);
85 static void shmrealloc (void);
86 static void shminit (void *);
87
88 /*
89  * Tuneable values
90  */
91 #ifndef SHMMIN
92 #define SHMMIN  1
93 #endif
94 #ifndef SHMMNI
95 #define SHMMNI  512
96 #endif
97 #ifndef SHMSEG
98 #define SHMSEG  1024
99 #endif
100
101 struct  shminfo shminfo = {
102         0,
103         SHMMIN,
104         SHMMNI,
105         SHMSEG,
106         0
107 };
108
109 /*
110  * allow-removed    Allow a shared memory segment to be attached by its shmid
111  *                  even after it has been deleted, as long as it was still
112  *                  being referenced by someone.  This is a trick used by
113  *                  chrome and other applications to avoid leaving shm
114  *                  segments hanging around after the application is killed
115  *                  or seg-faults unexpectedly.
116  *
117  * use-phys         Shared memory segments are to use physical memory by
118  *                  default, which may allow the kernel to better-optimize
119  *                  the pmap and reduce overhead.  The pages are effectively
120  *                  wired.
121  */
122 static int shm_allow_removed = 1;
123 static int shm_use_phys = 1;
124
125 TUNABLE_LONG("kern.ipc.shmmin", &shminfo.shmmin);
126 TUNABLE_LONG("kern.ipc.shmmni", &shminfo.shmmni);
127 TUNABLE_LONG("kern.ipc.shmseg", &shminfo.shmseg);
128 TUNABLE_LONG("kern.ipc.shmmaxpgs", &shminfo.shmall);
129 TUNABLE_INT("kern.ipc.shm_use_phys", &shm_use_phys);
130
131 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmax, CTLFLAG_RW, &shminfo.shmmax, 0,
132     "Max shared memory segment size");
133 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmin, CTLFLAG_RW, &shminfo.shmmin, 0,
134     "Min shared memory segment size");
135 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmni, CTLFLAG_RD, &shminfo.shmmni, 0,
136     "Max number of shared memory identifiers");
137 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmseg, CTLFLAG_RW, &shminfo.shmseg, 0,
138     "Max shared memory segments per process");
139 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmall, CTLFLAG_RW, &shminfo.shmall, 0,
140     "Max pages of shared memory");
141 SYSCTL_INT(_kern_ipc, OID_AUTO, shm_use_phys, CTLFLAG_RW, &shm_use_phys, 0,
142     "Use phys pager allocation instead of swap pager allocation");
143 SYSCTL_INT(_kern_ipc, OID_AUTO, shm_allow_removed, CTLFLAG_RW,
144     &shm_allow_removed, 0,
145     "Enable/Disable attachment to attached segments marked for removal");
146
147 static int
148 shm_find_segment_by_key(key_t key)
149 {
150         int i;
151
152         for (i = 0; i < shmalloced; i++) {
153                 if ((shmsegs[i].shm_perm.mode & SHMSEG_ALLOCATED) &&
154                     shmsegs[i].shm_perm.key == key)
155                         return i;
156         }
157         return -1;
158 }
159
160 static struct shmid_ds *
161 shm_find_segment_by_shmid(int shmid)
162 {
163         int segnum;
164         struct shmid_ds *shmseg;
165
166         segnum = IPCID_TO_IX(shmid);
167         if (segnum < 0 || segnum >= shmalloced)
168                 return NULL;
169         shmseg = &shmsegs[segnum];
170         if ((shmseg->shm_perm.mode & SHMSEG_ALLOCATED) == 0 ||
171             (!shm_allow_removed &&
172             (shmseg->shm_perm.mode & SHMSEG_REMOVED) != 0) ||
173             shmseg->shm_perm.seq != IPCID_TO_SEQ(shmid)) {
174                 return NULL;
175         }
176         return shmseg;
177 }
178
179 static void
180 shm_deallocate_segment(struct shmid_ds *shmseg)
181 {
182         struct shm_handle *shm_handle;
183         size_t size;
184
185         shm_handle = shmseg->shm_internal;
186         vm_object_deallocate(shm_handle->shm_object);
187         kfree((caddr_t)shm_handle, M_SHM);
188         shmseg->shm_internal = NULL;
189         size = round_page(shmseg->shm_segsz);
190         shm_committed -= btoc(size);
191         shm_nused--;
192         shmseg->shm_perm.mode = SHMSEG_FREE;
193 }
194
195 static int
196 shm_delete_mapping(struct vmspace *vm, struct shmmap_state *shmmap_s)
197 {
198         struct shmid_ds *shmseg;
199         int segnum, result;
200         size_t size;
201
202         segnum = IPCID_TO_IX(shmmap_s->shmid);
203         shmseg = &shmsegs[segnum];
204         size = round_page(shmseg->shm_segsz);
205         result = vm_map_remove(&vm->vm_map, shmmap_s->va, shmmap_s->va + size);
206         if (result != KERN_SUCCESS)
207                 return EINVAL;
208         shmmap_s->shmid = -1;
209         shmseg->shm_dtime = time_second;
210         if ((--shmseg->shm_nattch <= 0) &&
211             (shmseg->shm_perm.mode & SHMSEG_REMOVED)) {
212                 shm_deallocate_segment(shmseg);
213                 shm_last_free = segnum;
214         }
215         return 0;
216 }
217
218 /*
219  * MPALMOSTSAFE
220  */
221 int
222 sys_shmdt(struct shmdt_args *uap)
223 {
224         struct thread *td = curthread;
225         struct proc *p = td->td_proc;
226         struct shmmap_state *shmmap_s;
227         long i;
228         int error;
229
230         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
231                 return (ENOSYS);
232
233         lwkt_gettoken(&shm_token);
234         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
235         if (shmmap_s == NULL) {
236                 error = EINVAL;
237                 goto done;
238         }
239         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
240                 if (shmmap_s->shmid != -1 &&
241                     shmmap_s->va == (vm_offset_t)uap->shmaddr)
242                         break;
243         }
244         if (i == shminfo.shmseg)
245                 error = EINVAL;
246         else
247                 error = shm_delete_mapping(p->p_vmspace, shmmap_s);
248 done:
249         lwkt_reltoken(&shm_token);
250
251         return (error);
252 }
253
254 /*
255  * MPALMOSTSAFE
256  */
257 int
258 sys_shmat(struct shmat_args *uap)
259 {
260         struct thread *td = curthread;
261         struct proc *p = td->td_proc;
262         int error, flags;
263         long i;
264         struct shmid_ds *shmseg;
265         struct shmmap_state *shmmap_s = NULL;
266         struct shm_handle *shm_handle;
267         vm_offset_t attach_va;
268         vm_prot_t prot;
269         vm_size_t size;
270         vm_size_t align;
271         int rv;
272
273         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
274                 return (ENOSYS);
275
276         lwkt_gettoken(&shm_token);
277 again:
278         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
279         if (shmmap_s == NULL) {
280                 size = shminfo.shmseg * sizeof(struct shmmap_state);
281                 shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
282                 for (i = 0; i < shminfo.shmseg; i++) {
283                         shmmap_s[i].shmid = -1;
284                         shmmap_s[i].reserved = 0;
285                 }
286                 if (p->p_vmspace->vm_shm != NULL) {
287                         kfree(shmmap_s, M_SHM);
288                         goto again;
289                 }
290                 p->p_vmspace->vm_shm = (caddr_t)shmmap_s;
291         }
292         shmseg = shm_find_segment_by_shmid(uap->shmid);
293         if (shmseg == NULL) {
294                 error = EINVAL;
295                 goto done;
296         }
297         error = ipcperm(p, &shmseg->shm_perm,
298                         (uap->shmflg & SHM_RDONLY) ? IPC_R : IPC_R|IPC_W);
299         if (error)
300                 goto done;
301
302         /*
303          * Find a free element and mark reserved.  This fixes races
304          * against concurrent allocations due to the token being
305          * interrupted by blocking operations.  The shmmap_s reservation
306          * will be cleared upon completion or error.
307          */
308         for (i = 0; i < shminfo.shmseg; i++) {
309                 if (shmmap_s->shmid == -1 && shmmap_s->reserved == 0) {
310                         shmmap_s->reserved = 1;
311                         break;
312                 }
313                 shmmap_s++;
314         }
315         if (i >= shminfo.shmseg) {
316                 error = EMFILE;
317                 goto done;
318         }
319         size = round_page(shmseg->shm_segsz);
320 #ifdef VM_PROT_READ_IS_EXEC
321         prot = VM_PROT_READ | VM_PROT_EXECUTE;
322 #else
323         prot = VM_PROT_READ;
324 #endif
325         if ((uap->shmflg & SHM_RDONLY) == 0)
326                 prot |= VM_PROT_WRITE;
327         flags = MAP_ANON | MAP_SHARED;
328         if (uap->shmaddr) {
329                 flags |= MAP_FIXED;
330                 if (uap->shmflg & SHM_RND) {
331                         attach_va = (vm_offset_t)uap->shmaddr & ~(SHMLBA-1);
332                 } else if (((vm_offset_t)uap->shmaddr & (SHMLBA-1)) == 0) {
333                         attach_va = (vm_offset_t)uap->shmaddr;
334                 } else {
335                         error = EINVAL;
336                         shmmap_s->reserved = 0;
337                         goto done;
338                 }
339         } else {
340                 /*
341                  * This is just a hint to vm_map_find() about where to put it.
342                  */
343                 attach_va = round_page((vm_offset_t)p->p_vmspace->vm_taddr +
344                                        maxtsiz + maxdsiz);
345         }
346
347         /*
348          * Handle alignment.  For large memory maps it is possible
349          * that the MMU can optimize the page table so align anything
350          * that is a multiple of SEG_SIZE to SEG_SIZE.
351          */
352         if ((flags & MAP_FIXED) == 0 && (size & SEG_MASK) == 0)
353                 align = SEG_SIZE;
354         else
355                 align = PAGE_SIZE;
356
357         shm_handle = shmseg->shm_internal;
358         vm_object_hold(shm_handle->shm_object);
359         vm_object_reference_locked(shm_handle->shm_object);
360         rv = vm_map_find(&p->p_vmspace->vm_map, 
361                          shm_handle->shm_object, NULL,
362                          0, &attach_va, size,
363                          align,
364                          ((flags & MAP_FIXED) ? 0 : 1), 
365                          VM_MAPTYPE_NORMAL, VM_SUBSYS_SHMEM,
366                          prot, prot, 0);
367         vm_object_drop(shm_handle->shm_object);
368         if (rv != KERN_SUCCESS) {
369                 vm_object_deallocate(shm_handle->shm_object);
370                 shmmap_s->reserved = 0;
371                 error = ENOMEM;
372                 goto done;
373         }
374         vm_map_inherit(&p->p_vmspace->vm_map,
375                        attach_va, attach_va + size, VM_INHERIT_SHARE);
376
377         KKASSERT(shmmap_s->shmid == -1);
378         shmmap_s->va = attach_va;
379         shmmap_s->shmid = uap->shmid;
380         shmmap_s->reserved = 0;
381         shmseg->shm_lpid = p->p_pid;
382         shmseg->shm_atime = time_second;
383         shmseg->shm_nattch++;
384         uap->sysmsg_resultp = (void *)attach_va;
385         error = 0;
386 done:
387         lwkt_reltoken(&shm_token);
388
389         return error;
390 }
391
392 /*
393  * MPALMOSTSAFE
394  */
395 int
396 sys_shmctl(struct shmctl_args *uap)
397 {
398         struct thread *td = curthread;
399         struct proc *p = td->td_proc;
400         int error;
401         struct shmid_ds inbuf;
402         struct shmid_ds *shmseg;
403
404         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
405                 return (ENOSYS);
406
407         lwkt_gettoken(&shm_token);
408         shmseg = shm_find_segment_by_shmid(uap->shmid);
409         if (shmseg == NULL) {
410                 error = EINVAL;
411                 goto done;
412         }
413
414         switch (uap->cmd) {
415         case IPC_STAT:
416                 error = ipcperm(p, &shmseg->shm_perm, IPC_R);
417                 if (error == 0)
418                         error = copyout(shmseg, uap->buf, sizeof(inbuf));
419                 break;
420         case IPC_SET:
421                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
422                 if (error == 0)
423                         error = copyin(uap->buf, &inbuf, sizeof(inbuf));
424                 if (error == 0) {
425                         shmseg->shm_perm.uid = inbuf.shm_perm.uid;
426                         shmseg->shm_perm.gid = inbuf.shm_perm.gid;
427                         shmseg->shm_perm.mode =
428                             (shmseg->shm_perm.mode & ~ACCESSPERMS) |
429                             (inbuf.shm_perm.mode & ACCESSPERMS);
430                         shmseg->shm_ctime = time_second;
431                 }
432                 break;
433         case IPC_RMID:
434                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
435                 if (error == 0) {
436                         shmseg->shm_perm.key = IPC_PRIVATE;
437                         shmseg->shm_perm.mode |= SHMSEG_REMOVED;
438                         if (shmseg->shm_nattch <= 0) {
439                                 shm_deallocate_segment(shmseg);
440                                 shm_last_free = IPCID_TO_IX(uap->shmid);
441                         }
442                 }
443                 break;
444 #if 0
445         case SHM_LOCK:
446         case SHM_UNLOCK:
447 #endif
448         default:
449                 error = EINVAL;
450                 break;
451         }
452 done:
453         lwkt_reltoken(&shm_token);
454
455         return error;
456 }
457
458 static int
459 shmget_existing(struct proc *p, struct shmget_args *uap, int mode, int segnum)
460 {
461         struct shmid_ds *shmseg;
462         int error;
463
464         shmseg = &shmsegs[segnum];
465         if (shmseg->shm_perm.mode & SHMSEG_REMOVED) {
466                 /*
467                  * This segment is in the process of being allocated.  Wait
468                  * until it's done, and look the key up again (in case the
469                  * allocation failed or it was freed).
470                  */
471                 shmseg->shm_perm.mode |= SHMSEG_WANTED;
472                 error = tsleep((caddr_t)shmseg, PCATCH, "shmget", 0);
473                 if (error)
474                         return error;
475                 return EAGAIN;
476         }
477         if ((uap->shmflg & (IPC_CREAT | IPC_EXCL)) == (IPC_CREAT | IPC_EXCL))
478                 return EEXIST;
479         error = ipcperm(p, &shmseg->shm_perm, mode);
480         if (error)
481                 return error;
482         if (uap->size && uap->size > shmseg->shm_segsz)
483                 return EINVAL;
484         uap->sysmsg_result = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
485         return 0;
486 }
487
488 static int
489 shmget_allocate_segment(struct proc *p, struct shmget_args *uap, int mode)
490 {
491         int i, segnum, shmid;
492         size_t size;
493         struct ucred *cred = p->p_ucred;
494         struct shmid_ds *shmseg;
495         struct shm_handle *shm_handle;
496
497         if (uap->size < shminfo.shmmin || uap->size > shminfo.shmmax)
498                 return EINVAL;
499         if (shm_nused >= shminfo.shmmni) /* any shmids left? */
500                 return ENOSPC;
501         size = round_page(uap->size);
502         if (shm_committed + btoc(size) > shminfo.shmall)
503                 return ENOMEM;
504         if (shm_last_free < 0) {
505                 shmrealloc();   /* maybe expand the shmsegs[] array */
506                 for (i = 0; i < shmalloced; i++) {
507                         if (shmsegs[i].shm_perm.mode & SHMSEG_FREE)
508                                 break;
509                 }
510                 if (i == shmalloced)
511                         return ENOSPC;
512                 segnum = i;
513         } else  {
514                 segnum = shm_last_free;
515                 shm_last_free = -1;
516         }
517         shmseg = &shmsegs[segnum];
518         /*
519          * In case we sleep in malloc(), mark the segment present but deleted
520          * so that noone else tries to create the same key.
521          */
522         shmseg->shm_perm.mode = SHMSEG_ALLOCATED | SHMSEG_REMOVED;
523         shmseg->shm_perm.key = uap->key;
524         shmseg->shm_perm.seq = (shmseg->shm_perm.seq + 1) & 0x7fff;
525         shm_handle = kmalloc(sizeof(struct shm_handle), M_SHM, M_WAITOK);
526         shmid = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
527         
528         /*
529          * We make sure that we have allocated a pager before we need
530          * to.
531          */
532         if (shm_use_phys) {
533                 shm_handle->shm_object =
534                    phys_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
535         } else {
536                 shm_handle->shm_object =
537                    swap_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
538         }
539         vm_object_clear_flag(shm_handle->shm_object, OBJ_ONEMAPPING);
540         vm_object_set_flag(shm_handle->shm_object, OBJ_NOSPLIT);
541
542         shmseg->shm_internal = shm_handle;
543         shmseg->shm_perm.cuid = shmseg->shm_perm.uid = cred->cr_uid;
544         shmseg->shm_perm.cgid = shmseg->shm_perm.gid = cred->cr_gid;
545         shmseg->shm_perm.mode = (shmseg->shm_perm.mode & SHMSEG_WANTED) |
546             (mode & ACCESSPERMS) | SHMSEG_ALLOCATED;
547         shmseg->shm_segsz = uap->size;
548         shmseg->shm_cpid = p->p_pid;
549         shmseg->shm_lpid = shmseg->shm_nattch = 0;
550         shmseg->shm_atime = shmseg->shm_dtime = 0;
551         shmseg->shm_ctime = time_second;
552         shm_committed += btoc(size);
553         shm_nused++;
554
555         /*
556          * If a physical mapping is desired and we have a ton of free pages
557          * we pre-allocate the pages here in order to avoid on-the-fly
558          * allocation later.  This has a big effect on database warm-up
559          * times since DFly supports concurrent page faults coming from the
560          * same VM object for pages which already exist.
561          *
562          * This can hang the kernel for a while so only do it if shm_use_phys
563          * is set to 2 or higher.
564          */
565         if (shm_use_phys > 1) {
566                 vm_pindex_t pi, pmax;
567                 vm_page_t m;
568
569                 pmax = round_page(shmseg->shm_segsz) >> PAGE_SHIFT;
570                 vm_object_hold(shm_handle->shm_object);
571                 if (pmax > vmstats.v_free_count)
572                         pmax = vmstats.v_free_count;
573                 for (pi = 0; pi < pmax; ++pi) {
574                         m = vm_page_grab(shm_handle->shm_object, pi,
575                                          VM_ALLOC_SYSTEM | VM_ALLOC_NULL_OK |
576                                          VM_ALLOC_ZERO);
577                         if (m == NULL)
578                                 break;
579                         vm_pager_get_page(shm_handle->shm_object, &m, 1);
580                         vm_page_activate(m);
581                         vm_page_wakeup(m);
582                         lwkt_yield();
583                 }
584                 vm_object_drop(shm_handle->shm_object);
585         }
586
587         if (shmseg->shm_perm.mode & SHMSEG_WANTED) {
588                 /*
589                  * Somebody else wanted this key while we were asleep.  Wake
590                  * them up now.
591                  */
592                 shmseg->shm_perm.mode &= ~SHMSEG_WANTED;
593                 wakeup((caddr_t)shmseg);
594         }
595         uap->sysmsg_result = shmid;
596         return 0;
597 }
598
599 /*
600  * MPALMOSTSAFE
601  */
602 int
603 sys_shmget(struct shmget_args *uap)
604 {
605         struct thread *td = curthread;
606         struct proc *p = td->td_proc;
607         int segnum, mode, error;
608
609         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
610                 return (ENOSYS);
611
612         mode = uap->shmflg & ACCESSPERMS;
613
614         lwkt_gettoken(&shm_token);
615
616         if (uap->key != IPC_PRIVATE) {
617         again:
618                 segnum = shm_find_segment_by_key(uap->key);
619                 if (segnum >= 0) {
620                         error = shmget_existing(p, uap, mode, segnum);
621                         if (error == EAGAIN)
622                                 goto again;
623                         goto done;
624                 }
625                 if ((uap->shmflg & IPC_CREAT) == 0) {
626                         error = ENOENT;
627                         goto done;
628                 }
629         }
630         error = shmget_allocate_segment(p, uap, mode);
631 done:
632         lwkt_reltoken(&shm_token);
633
634         return (error);
635 }
636
637 void
638 shmfork(struct proc *p1, struct proc *p2)
639 {
640         struct shmmap_state *shmmap_s;
641         size_t size;
642         int i;
643
644         lwkt_gettoken(&shm_token);
645         size = shminfo.shmseg * sizeof(struct shmmap_state);
646         shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
647         bcopy((caddr_t)p1->p_vmspace->vm_shm, (caddr_t)shmmap_s, size);
648         p2->p_vmspace->vm_shm = (caddr_t)shmmap_s;
649         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
650                 if (shmmap_s->shmid != -1)
651                         shmsegs[IPCID_TO_IX(shmmap_s->shmid)].shm_nattch++;
652         }
653         lwkt_reltoken(&shm_token);
654 }
655
656 void
657 shmexit(struct vmspace *vm)
658 {
659         struct shmmap_state *base, *shm;
660         int i;
661
662         if ((base = (struct shmmap_state *)vm->vm_shm) != NULL) {
663                 vm->vm_shm = NULL;
664                 lwkt_gettoken(&shm_token);
665                 for (i = 0, shm = base; i < shminfo.shmseg; i++, shm++) {
666                         if (shm->shmid != -1)
667                                 shm_delete_mapping(vm, shm);
668                 }
669                 kfree(base, M_SHM);
670                 lwkt_reltoken(&shm_token);
671         }
672 }
673
674 static void
675 shmrealloc(void)
676 {
677         int i;
678         struct shmid_ds *newsegs;
679
680         if (shmalloced >= shminfo.shmmni)
681                 return;
682
683         newsegs = kmalloc(shminfo.shmmni * sizeof(*newsegs), M_SHM, M_WAITOK);
684         for (i = 0; i < shmalloced; i++)
685                 bcopy(&shmsegs[i], &newsegs[i], sizeof(newsegs[0]));
686         for (; i < shminfo.shmmni; i++) {
687                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
688                 shmsegs[i].shm_perm.seq = 0;
689         }
690         kfree(shmsegs, M_SHM);
691         shmsegs = newsegs;
692         shmalloced = shminfo.shmmni;
693 }
694
695 static void
696 shminit(void *dummy)
697 {
698         int i;
699
700         /*
701          * If not overridden by a tunable set the maximum shm to
702          * 2/3 of main memory.
703          */
704         if (shminfo.shmall == 0)
705                 shminfo.shmall = (size_t)vmstats.v_page_count * 2 / 3;
706
707         shminfo.shmmax = shminfo.shmall * PAGE_SIZE;
708         shmalloced = shminfo.shmmni;
709         shmsegs = kmalloc(shmalloced * sizeof(shmsegs[0]), M_SHM, M_WAITOK);
710         for (i = 0; i < shmalloced; i++) {
711                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
712                 shmsegs[i].shm_perm.seq = 0;
713         }
714         shm_last_free = 0;
715         shm_nused = 0;
716         shm_committed = 0;
717 }
718 SYSINIT(sysv_shm, SI_SUB_SYSV_SHM, SI_ORDER_FIRST, shminit, NULL);