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         struct prison *pr = p->p_ucred->cr_prison;
 228
 229         long i;
 230         int error;
 231
 232         if (pr && !pr->pr_sysvipc_allowed)
 233                 return (ENOSYS);
 234
 235         lwkt_gettoken(&shm_token);
 236         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 237         if (shmmap_s == NULL) {
 238                 error = EINVAL;
 239                 goto done;
 240         }
 241         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 242                 if (shmmap_s->shmid != -1 &&
 243                     shmmap_s->va == (vm_offset_t)uap->shmaddr)
 244                         break;
 245         }
 246         if (i == shminfo.shmseg)
 247                 error = EINVAL;
 248         else
 249                 error = shm_delete_mapping(p->p_vmspace, shmmap_s);
 250 done:
 251         lwkt_reltoken(&shm_token);
 252
 253         return (error);
 254 }
 255
 256 /*
 257  * MPALMOSTSAFE
 258  */
 259 int
 260 sys_shmat(struct shmat_args *uap)
 261 {
 262         struct thread *td = curthread;
 263         struct proc *p = td->td_proc;
 264         struct prison *pr = p->p_ucred->cr_prison;
 265         int error, flags;
 266         long i;
 267         struct shmid_ds *shmseg;
 268         struct shmmap_state *shmmap_s = NULL;
 269         struct shm_handle *shm_handle;
 270         vm_offset_t attach_va;
 271         vm_prot_t prot;
 272         vm_size_t size;
 273         vm_size_t align;
 274         int rv;
 275
 276         if (pr && !pr->pr_sysvipc_allowed)
 277                 return (ENOSYS);
 278
 279         lwkt_gettoken(&shm_token);
 280 again:
 281         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 282         if (shmmap_s == NULL) {
 283                 size = shminfo.shmseg * sizeof(struct shmmap_state);
 284                 shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 285                 for (i = 0; i < shminfo.shmseg; i++) {
 286                         shmmap_s[i].shmid = -1;
 287                         shmmap_s[i].reserved = 0;
 288                 }
 289                 if (p->p_vmspace->vm_shm != NULL) {
 290                         kfree(shmmap_s, M_SHM);
 291                         goto again;
 292                 }
 293                 p->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 294         }
 295         shmseg = shm_find_segment_by_shmid(uap->shmid);
 296         if (shmseg == NULL) {
 297                 error = EINVAL;
 298                 goto done;
 299         }
 300         error = ipcperm(p, &shmseg->shm_perm,
 301                         (uap->shmflg & SHM_RDONLY) ? IPC_R : IPC_R|IPC_W);
 302         if (error)
 303                 goto done;
 304
 305         /*
 306          * Find a free element and mark reserved.  This fixes races
 307          * against concurrent allocations due to the token being
 308          * interrupted by blocking operations.  The shmmap_s reservation
 309          * will be cleared upon completion or error.
 310          */
 311         for (i = 0; i < shminfo.shmseg; i++) {
 312                 if (shmmap_s->shmid == -1 && shmmap_s->reserved == 0) {
 313                         shmmap_s->reserved = 1;
 314                         break;
 315                 }
 316                 shmmap_s++;
 317         }
 318         if (i >= shminfo.shmseg) {
 319                 error = EMFILE;
 320                 goto done;
 321         }
 322         size = round_page(shmseg->shm_segsz);
 323 #ifdef VM_PROT_READ_IS_EXEC
 324         prot = VM_PROT_READ | VM_PROT_EXECUTE;
 325 #else
 326         prot = VM_PROT_READ;
 327 #endif
 328         if ((uap->shmflg & SHM_RDONLY) == 0)
 329                 prot |= VM_PROT_WRITE;
 330         flags = MAP_ANON | MAP_SHARED;
 331         if (uap->shmaddr) {
 332                 flags |= MAP_FIXED;
 333                 if (uap->shmflg & SHM_RND) {
 334                         attach_va =
 335                             rounddown2((vm_offset_t)uap->shmaddr, SHMLBA);
 336                 } else if (((vm_offset_t)uap->shmaddr & (SHMLBA-1)) == 0) {
 337                         attach_va = (vm_offset_t)uap->shmaddr;
 338                 } else {
 339                         error = EINVAL;
 340                         shmmap_s->reserved = 0;
 341                         goto done;
 342                 }
 343         } else {
 344                 /*
 345                  * This is just a hint to vm_map_find() about where to put it.
 346                  */
 347                 attach_va = round_page((vm_offset_t)p->p_vmspace->vm_taddr +
 348                                        maxtsiz + maxdsiz);
 349         }
 350
 351         /*
 352          * Handle alignment.  For large memory maps it is possible
 353          * that the MMU can optimize the page table so align anything
 354          * that is a multiple of SEG_SIZE to SEG_SIZE.
 355          */
 356         if ((flags & MAP_FIXED) == 0 && (size & SEG_MASK) == 0)
 357                 align = SEG_SIZE;
 358         else
 359                 align = PAGE_SIZE;
 360
 361         shm_handle = shmseg->shm_internal;
 362         vm_object_hold(shm_handle->shm_object);
 363         vm_object_reference_locked(shm_handle->shm_object);
 364         rv = vm_map_find(&p->p_vmspace->vm_map,
 365                          shm_handle->shm_object, NULL,
 366                          0, &attach_va, size,
 367                          align,
 368                          ((flags & MAP_FIXED) ? 0 : 1),
 369                          VM_MAPTYPE_NORMAL, VM_SUBSYS_SHMEM,
 370                          prot, prot, 0);
 371         vm_object_drop(shm_handle->shm_object);
 372         if (rv != KERN_SUCCESS) {
 373                 vm_object_deallocate(shm_handle->shm_object);
 374                 shmmap_s->reserved = 0;
 375                 error = ENOMEM;
 376                 goto done;
 377         }
 378         vm_map_inherit(&p->p_vmspace->vm_map,
 379                        attach_va, attach_va + size, VM_INHERIT_SHARE);
 380
 381         KKASSERT(shmmap_s->shmid == -1);
 382         shmmap_s->va = attach_va;
 383         shmmap_s->shmid = uap->shmid;
 384         shmmap_s->reserved = 0;
 385         shmseg->shm_lpid = p->p_pid;
 386         shmseg->shm_atime = time_second;
 387         shmseg->shm_nattch++;
 388         uap->sysmsg_resultp = (void *)attach_va;
 389         error = 0;
 390 done:
 391         lwkt_reltoken(&shm_token);
 392
 393         return error;
 394 }
 395
 396 /*
 397  * MPALMOSTSAFE
 398  */
 399 int
 400 sys_shmctl(struct shmctl_args *uap)
 401 {
 402         struct thread *td = curthread;
 403         struct proc *p = td->td_proc;
 404         struct prison *pr = p->p_ucred->cr_prison;
 405         int error;
 406         struct shmid_ds inbuf;
 407         struct shmid_ds *shmseg;
 408
 409         if (pr && !pr->pr_sysvipc_allowed)
 410                 return (ENOSYS);
 411
 412         lwkt_gettoken(&shm_token);
 413         shmseg = shm_find_segment_by_shmid(uap->shmid);
 414         if (shmseg == NULL) {
 415                 error = EINVAL;
 416                 goto done;
 417         }
 418
 419         switch (uap->cmd) {
 420         case IPC_STAT:
 421                 error = ipcperm(p, &shmseg->shm_perm, IPC_R);
 422                 if (error == 0)
 423                         error = copyout(shmseg, uap->buf, sizeof(inbuf));
 424                 break;
 425         case IPC_SET:
 426                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 427                 if (error == 0)
 428                         error = copyin(uap->buf, &inbuf, sizeof(inbuf));
 429                 if (error == 0) {
 430                         shmseg->shm_perm.uid = inbuf.shm_perm.uid;
 431                         shmseg->shm_perm.gid = inbuf.shm_perm.gid;
 432                         shmseg->shm_perm.mode =
 433                             (shmseg->shm_perm.mode & ~ACCESSPERMS) |
 434                             (inbuf.shm_perm.mode & ACCESSPERMS);
 435                         shmseg->shm_ctime = time_second;
 436                 }
 437                 break;
 438         case IPC_RMID:
 439                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 440                 if (error == 0) {
 441                         shmseg->shm_perm.key = IPC_PRIVATE;
 442                         shmseg->shm_perm.mode |= SHMSEG_REMOVED;
 443                         if (shmseg->shm_nattch <= 0) {
 444                                 shm_deallocate_segment(shmseg);
 445                                 shm_last_free = IPCID_TO_IX(uap->shmid);
 446                         }
 447                 }
 448                 break;
 449 #if 0
 450         case SHM_LOCK:
 451         case SHM_UNLOCK:
 452 #endif
 453         default:
 454                 error = EINVAL;
 455                 break;
 456         }
 457 done:
 458         lwkt_reltoken(&shm_token);
 459
 460         return error;
 461 }
 462
 463 static int
 464 shmget_existing(struct proc *p, struct shmget_args *uap, int mode, int segnum)
 465 {
 466         struct shmid_ds *shmseg;
 467         int error;
 468
 469         shmseg = &shmsegs[segnum];
 470         if (shmseg->shm_perm.mode & SHMSEG_REMOVED) {
 471                 /*
 472                  * This segment is in the process of being allocated.  Wait
 473                  * until it's done, and look the key up again (in case the
 474                  * allocation failed or it was freed).
 475                  */
 476                 shmseg->shm_perm.mode |= SHMSEG_WANTED;
 477                 error = tsleep((caddr_t)shmseg, PCATCH, "shmget", 0);
 478                 if (error)
 479                         return error;
 480                 return EAGAIN;
 481         }
 482         if ((uap->shmflg & (IPC_CREAT | IPC_EXCL)) == (IPC_CREAT | IPC_EXCL))
 483                 return EEXIST;
 484         error = ipcperm(p, &shmseg->shm_perm, mode);
 485         if (error)
 486                 return error;
 487         if (uap->size && uap->size > shmseg->shm_segsz)
 488                 return EINVAL;
 489         uap->sysmsg_result = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 490         return 0;
 491 }
 492
 493 static int
 494 shmget_allocate_segment(struct proc *p, struct shmget_args *uap, int mode)
 495 {
 496         int i, segnum, shmid;
 497         size_t size;
 498         struct ucred *cred = p->p_ucred;
 499         struct shmid_ds *shmseg;
 500         struct shm_handle *shm_handle;
 501
 502         if (uap->size < shminfo.shmmin || uap->size > shminfo.shmmax)
 503                 return EINVAL;
 504         if (shm_nused >= shminfo.shmmni) /* any shmids left? */
 505                 return ENOSPC;
 506         size = round_page(uap->size);
 507         if (shm_committed + btoc(size) > shminfo.shmall)
 508                 return ENOMEM;
 509         if (shm_last_free < 0) {
 510                 shmrealloc();   /* maybe expand the shmsegs[] array */
 511                 for (i = 0; i < shmalloced; i++) {
 512                         if (shmsegs[i].shm_perm.mode & SHMSEG_FREE)
 513                                 break;
 514                 }
 515                 if (i == shmalloced)
 516                         return ENOSPC;
 517                 segnum = i;
 518         } else  {
 519                 segnum = shm_last_free;
 520                 shm_last_free = -1;
 521         }
 522         shmseg = &shmsegs[segnum];
 523         /*
 524          * In case we sleep in malloc(), mark the segment present but deleted
 525          * so that noone else tries to create the same key.
 526          */
 527         shmseg->shm_perm.mode = SHMSEG_ALLOCATED | SHMSEG_REMOVED;
 528         shmseg->shm_perm.key = uap->key;
 529         shmseg->shm_perm.seq = (shmseg->shm_perm.seq + 1) & 0x7fff;
 530         shm_handle = kmalloc(sizeof(struct shm_handle), M_SHM, M_WAITOK);
 531         shmid = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 532
 533         /*
 534          * We make sure that we have allocated a pager before we need
 535          * to.
 536          */
 537         if (shm_use_phys) {
 538                 shm_handle->shm_object =
 539                    phys_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 540         } else {
 541                 shm_handle->shm_object =
 542                    swap_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 543         }
 544         vm_object_clear_flag(shm_handle->shm_object, OBJ_ONEMAPPING);
 545         vm_object_set_flag(shm_handle->shm_object, OBJ_NOSPLIT);
 546
 547         shmseg->shm_internal = shm_handle;
 548         shmseg->shm_perm.cuid = shmseg->shm_perm.uid = cred->cr_uid;
 549         shmseg->shm_perm.cgid = shmseg->shm_perm.gid = cred->cr_gid;
 550         shmseg->shm_perm.mode = (shmseg->shm_perm.mode & SHMSEG_WANTED) |
 551             (mode & ACCESSPERMS) | SHMSEG_ALLOCATED;
 552         shmseg->shm_segsz = uap->size;
 553         shmseg->shm_cpid = p->p_pid;
 554         shmseg->shm_lpid = shmseg->shm_nattch = 0;
 555         shmseg->shm_atime = shmseg->shm_dtime = 0;
 556         shmseg->shm_ctime = time_second;
 557         shm_committed += btoc(size);
 558         shm_nused++;
 559
 560         /*
 561          * If a physical mapping is desired and we have a ton of free pages
 562          * we pre-allocate the pages here in order to avoid on-the-fly
 563          * allocation later.  This has a big effect on database warm-up
 564          * times since DFly supports concurrent page faults coming from the
 565          * same VM object for pages which already exist.
 566          *
 567          * This can hang the kernel for a while so only do it if shm_use_phys
 568          * is set to 2 or higher.
 569          */
 570         if (shm_use_phys > 1) {
 571                 vm_pindex_t pi, pmax;
 572                 vm_page_t m;
 573
 574                 pmax = round_page(shmseg->shm_segsz) >> PAGE_SHIFT;
 575                 vm_object_hold(shm_handle->shm_object);
 576                 if (pmax > vmstats.v_free_count)
 577                         pmax = vmstats.v_free_count;
 578                 for (pi = 0; pi < pmax; ++pi) {
 579                         m = vm_page_grab(shm_handle->shm_object, pi,
 580                                          VM_ALLOC_SYSTEM | VM_ALLOC_NULL_OK |
 581                                          VM_ALLOC_ZERO);
 582                         if (m == NULL)
 583                                 break;
 584                         vm_pager_get_page(shm_handle->shm_object, &m, 1);
 585                         vm_page_activate(m);
 586                         vm_page_wakeup(m);
 587                         lwkt_yield();
 588                 }
 589                 vm_object_drop(shm_handle->shm_object);
 590         }
 591
 592         if (shmseg->shm_perm.mode & SHMSEG_WANTED) {
 593                 /*
 594                  * Somebody else wanted this key while we were asleep.  Wake
 595                  * them up now.
 596                  */
 597                 shmseg->shm_perm.mode &= ~SHMSEG_WANTED;
 598                 wakeup((caddr_t)shmseg);
 599         }
 600         uap->sysmsg_result = shmid;
 601         return 0;
 602 }
 603
 604 /*
 605  * MPALMOSTSAFE
 606  */
 607 int
 608 sys_shmget(struct shmget_args *uap)
 609 {
 610         struct thread *td = curthread;
 611         struct proc *p = td->td_proc;
 612         struct prison *pr = p->p_ucred->cr_prison;
 613         int segnum, mode, error;
 614
 615         if (pr && !pr->pr_sysvipc_allowed)
 616                 return (ENOSYS);
 617
 618         mode = uap->shmflg & ACCESSPERMS;
 619
 620         lwkt_gettoken(&shm_token);
 621
 622         if (uap->key != IPC_PRIVATE) {
 623         again:
 624                 segnum = shm_find_segment_by_key(uap->key);
 625                 if (segnum >= 0) {
 626                         error = shmget_existing(p, uap, mode, segnum);
 627                         if (error == EAGAIN)
 628                                 goto again;
 629                         goto done;
 630                 }
 631                 if ((uap->shmflg & IPC_CREAT) == 0) {
 632                         error = ENOENT;
 633                         goto done;
 634                 }
 635         }
 636         error = shmget_allocate_segment(p, uap, mode);
 637 done:
 638         lwkt_reltoken(&shm_token);
 639
 640         return (error);
 641 }
 642
 643 void
 644 shmfork(struct proc *p1, struct proc *p2)
 645 {
 646         struct shmmap_state *shmmap_s;
 647         size_t size;
 648         int i;
 649
 650         lwkt_gettoken(&shm_token);
 651         size = shminfo.shmseg * sizeof(struct shmmap_state);
 652         shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 653         bcopy((caddr_t)p1->p_vmspace->vm_shm, (caddr_t)shmmap_s, size);
 654         p2->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 655         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 656                 if (shmmap_s->shmid != -1)
 657                         shmsegs[IPCID_TO_IX(shmmap_s->shmid)].shm_nattch++;
 658         }
 659         lwkt_reltoken(&shm_token);
 660 }
 661
 662 void
 663 shmexit(struct vmspace *vm)
 664 {
 665         struct shmmap_state *base, *shm;
 666         int i;
 667
 668         if ((base = (struct shmmap_state *)vm->vm_shm) != NULL) {
 669                 vm->vm_shm = NULL;
 670                 lwkt_gettoken(&shm_token);
 671                 for (i = 0, shm = base; i < shminfo.shmseg; i++, shm++) {
 672                         if (shm->shmid != -1)
 673                                 shm_delete_mapping(vm, shm);
 674                 }
 675                 kfree(base, M_SHM);
 676                 lwkt_reltoken(&shm_token);
 677         }
 678 }
 679
 680 static void
 681 shmrealloc(void)
 682 {
 683         int i;
 684         struct shmid_ds *newsegs;
 685
 686         if (shmalloced >= shminfo.shmmni)
 687                 return;
 688
 689         newsegs = kmalloc(shminfo.shmmni * sizeof(*newsegs), M_SHM, M_WAITOK);
 690         for (i = 0; i < shmalloced; i++)
 691                 bcopy(&shmsegs[i], &newsegs[i], sizeof(newsegs[0]));
 692         for (; i < shminfo.shmmni; i++) {
 693                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 694                 shmsegs[i].shm_perm.seq = 0;
 695         }
 696         kfree(shmsegs, M_SHM);
 697         shmsegs = newsegs;
 698         shmalloced = shminfo.shmmni;
 699 }
 700
 701 static void
 702 shminit(void *dummy)
 703 {
 704         int i;
 705
 706         /*
 707          * If not overridden by a tunable set the maximum shm to
 708          * 2/3 of main memory.
 709          */
 710         if (shminfo.shmall == 0)
 711                 shminfo.shmall = (size_t)vmstats.v_page_count * 2 / 3;
 712
 713         shminfo.shmmax = shminfo.shmall * PAGE_SIZE;
 714         shmalloced = shminfo.shmmni;
 715         shmsegs = kmalloc(shmalloced * sizeof(shmsegs[0]), M_SHM, M_WAITOK);
 716         for (i = 0; i < shmalloced; i++) {
 717                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 718                 shmsegs[i].shm_perm.seq = 0;
 719         }
 720         shm_last_free = 0;
 721         shm_nused = 0;
 722         shm_committed = 0;
 723 }
 724 SYSINIT(sysv_shm, SI_SUB_SYSV_SHM, SI_ORDER_FIRST, shminit, NULL);