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 };
  79
  80 static void shm_deallocate_segment (struct shmid_ds *);
  81 static int shm_find_segment_by_key (key_t);
  82 static struct shmid_ds *shm_find_segment_by_shmid (int);
  83 static int shm_delete_mapping (struct vmspace *vm, struct shmmap_state *);
  84 static void shmrealloc (void);
  85 static void shminit (void *);
  86
  87 /*
  88  * Tuneable values
  89  */
  90 #ifndef SHMMIN
  91 #define SHMMIN  1
  92 #endif
  93 #ifndef SHMMNI
  94 #define SHMMNI  512
  95 #endif
  96 #ifndef SHMSEG
  97 #define SHMSEG  1024
  98 #endif
  99
 100 struct  shminfo shminfo = {
 101         0,
 102         SHMMIN,
 103         SHMMNI,
 104         SHMSEG,
 105         0
 106 };
 107
 108 /*
 109  * allow-removed    Allow a shared memory segment to be attached by its shmid
 110  *                  even after it has been deleted, as long as it was still
 111  *                  being referenced by someone.  This is a trick used by
 112  *                  chrome and other applications to avoid leaving shm
 113  *                  segments hanging around after the application is killed
 114  *                  or seg-faults unexpectedly.
 115  *
 116  * use-phys         Shared memory segments are to use physical memory by
 117  *                  default, which allows the kernel to optimize (remove)
 118  *                  pv_entry management structures for the related PTEs and
 119  *                  prevents paging.  This has distinctly different and
 120  *                  usually desireable characteristics verses mmap()ing
 121  *                  anonymous memory.
 122  */
 123 static int shm_allow_removed = 1;
 124 static int shm_use_phys = 1;
 125
 126 TUNABLE_LONG("kern.ipc.shmmin", &shminfo.shmmin);
 127 TUNABLE_LONG("kern.ipc.shmmni", &shminfo.shmmni);
 128 TUNABLE_LONG("kern.ipc.shmseg", &shminfo.shmseg);
 129 TUNABLE_LONG("kern.ipc.shmmaxpgs", &shminfo.shmall);
 130 TUNABLE_INT("kern.ipc.shm_use_phys", &shm_use_phys);
 131
 132 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmax, CTLFLAG_RW, &shminfo.shmmax, 0,
 133     "Max shared memory segment size");
 134 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmin, CTLFLAG_RW, &shminfo.shmmin, 0,
 135     "Min shared memory segment size");
 136 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmni, CTLFLAG_RD, &shminfo.shmmni, 0,
 137     "Max number of shared memory identifiers");
 138 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmseg, CTLFLAG_RW, &shminfo.shmseg, 0,
 139     "Max shared memory segments per process");
 140 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmall, CTLFLAG_RW, &shminfo.shmall, 0,
 141     "Max pages of shared memory");
 142 SYSCTL_INT(_kern_ipc, OID_AUTO, shm_use_phys, CTLFLAG_RW, &shm_use_phys, 0,
 143     "Use phys pager allocation instead of swap pager allocation");
 144 SYSCTL_INT(_kern_ipc, OID_AUTO, shm_allow_removed, CTLFLAG_RW,
 145     &shm_allow_removed, 0,
 146     "Enable/Disable attachment to attached segments marked for removal");
 147
 148 static int
 149 shm_find_segment_by_key(key_t key)
 150 {
 151         int i;
 152
 153         for (i = 0; i < shmalloced; i++) {
 154                 if ((shmsegs[i].shm_perm.mode & SHMSEG_ALLOCATED) &&
 155                     shmsegs[i].shm_perm.key == key)
 156                         return i;
 157         }
 158         return -1;
 159 }
 160
 161 static struct shmid_ds *
 162 shm_find_segment_by_shmid(int shmid)
 163 {
 164         int segnum;
 165         struct shmid_ds *shmseg;
 166
 167         segnum = IPCID_TO_IX(shmid);
 168         if (segnum < 0 || segnum >= shmalloced)
 169                 return NULL;
 170         shmseg = &shmsegs[segnum];
 171         if ((shmseg->shm_perm.mode & SHMSEG_ALLOCATED) == 0 ||
 172             (!shm_allow_removed &&
 173             (shmseg->shm_perm.mode & SHMSEG_REMOVED) != 0) ||
 174             shmseg->shm_perm.seq != IPCID_TO_SEQ(shmid)) {
 175                 return NULL;
 176         }
 177         return shmseg;
 178 }
 179
 180 static void
 181 shm_deallocate_segment(struct shmid_ds *shmseg)
 182 {
 183         struct shm_handle *shm_handle;
 184         size_t size;
 185
 186         shm_handle = shmseg->shm_internal;
 187         vm_object_deallocate(shm_handle->shm_object);
 188         kfree((caddr_t)shm_handle, M_SHM);
 189         shmseg->shm_internal = NULL;
 190         size = round_page(shmseg->shm_segsz);
 191         shm_committed -= btoc(size);
 192         shm_nused--;
 193         shmseg->shm_perm.mode = SHMSEG_FREE;
 194 }
 195
 196 static int
 197 shm_delete_mapping(struct vmspace *vm, struct shmmap_state *shmmap_s)
 198 {
 199         struct shmid_ds *shmseg;
 200         int segnum, result;
 201         size_t size;
 202
 203         segnum = IPCID_TO_IX(shmmap_s->shmid);
 204         shmseg = &shmsegs[segnum];
 205         size = round_page(shmseg->shm_segsz);
 206         result = vm_map_remove(&vm->vm_map, shmmap_s->va, shmmap_s->va + size);
 207         if (result != KERN_SUCCESS)
 208                 return EINVAL;
 209         shmmap_s->shmid = -1;
 210         shmseg->shm_dtime = time_second;
 211         if ((--shmseg->shm_nattch <= 0) &&
 212             (shmseg->shm_perm.mode & SHMSEG_REMOVED)) {
 213                 shm_deallocate_segment(shmseg);
 214                 shm_last_free = segnum;
 215         }
 216         return 0;
 217 }
 218
 219 /*
 220  * MPALMOSTSAFE
 221  */
 222 int
 223 sys_shmdt(struct shmdt_args *uap)
 224 {
 225         struct thread *td = curthread;
 226         struct proc *p = td->td_proc;
 227         struct shmmap_state *shmmap_s;
 228         long i;
 229         int error;
 230
 231         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 232                 return (ENOSYS);
 233
 234         lwkt_gettoken(&shm_token);
 235         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 236         if (shmmap_s == NULL) {
 237                 error = EINVAL;
 238                 goto done;
 239         }
 240         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 241                 if (shmmap_s->shmid != -1 &&
 242                     shmmap_s->va == (vm_offset_t)uap->shmaddr)
 243                         break;
 244         }
 245         if (i == shminfo.shmseg)
 246                 error = EINVAL;
 247         else
 248                 error = shm_delete_mapping(p->p_vmspace, shmmap_s);
 249 done:
 250         lwkt_reltoken(&shm_token);
 251
 252         return (error);
 253 }
 254
 255 /*
 256  * MPALMOSTSAFE
 257  */
 258 int
 259 sys_shmat(struct shmat_args *uap)
 260 {
 261         struct thread *td = curthread;
 262         struct proc *p = td->td_proc;
 263         int error, flags;
 264         long i;
 265         struct shmid_ds *shmseg;
 266         struct shmmap_state *shmmap_s = NULL;
 267         struct shm_handle *shm_handle;
 268         vm_offset_t attach_va;
 269         vm_prot_t prot;
 270         vm_size_t size;
 271         vm_size_t align;
 272         int rv;
 273
 274         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 275                 return (ENOSYS);
 276
 277         lwkt_gettoken(&shm_token);
 278 again:
 279         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 280         if (shmmap_s == NULL) {
 281                 size = shminfo.shmseg * sizeof(struct shmmap_state);
 282                 shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 283                 for (i = 0; i < shminfo.shmseg; i++)
 284                         shmmap_s[i].shmid = -1;
 285                 if (p->p_vmspace->vm_shm != NULL) {
 286                         kfree(shmmap_s, M_SHM);
 287                         goto again;
 288                 }
 289                 p->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 290         }
 291         shmseg = shm_find_segment_by_shmid(uap->shmid);
 292         if (shmseg == NULL) {
 293                 error = EINVAL;
 294                 goto done;
 295         }
 296         error = ipcperm(p, &shmseg->shm_perm,
 297                         (uap->shmflg & SHM_RDONLY) ? IPC_R : IPC_R|IPC_W);
 298         if (error)
 299                 goto done;
 300         for (i = 0; i < shminfo.shmseg; i++) {
 301                 if (shmmap_s->shmid == -1)
 302                         break;
 303                 shmmap_s++;
 304         }
 305         if (i >= shminfo.shmseg) {
 306                 error = EMFILE;
 307                 goto done;
 308         }
 309         size = round_page(shmseg->shm_segsz);
 310 #ifdef VM_PROT_READ_IS_EXEC
 311         prot = VM_PROT_READ | VM_PROT_EXECUTE;
 312 #else
 313         prot = VM_PROT_READ;
 314 #endif
 315         if ((uap->shmflg & SHM_RDONLY) == 0)
 316                 prot |= VM_PROT_WRITE;
 317         flags = MAP_ANON | MAP_SHARED;
 318         if (uap->shmaddr) {
 319                 flags |= MAP_FIXED;
 320                 if (uap->shmflg & SHM_RND) {
 321                         attach_va = (vm_offset_t)uap->shmaddr & ~(SHMLBA-1);
 322                 } else if (((vm_offset_t)uap->shmaddr & (SHMLBA-1)) == 0) {
 323                         attach_va = (vm_offset_t)uap->shmaddr;
 324                 } else {
 325                         error = EINVAL;
 326                         goto done;
 327                 }
 328         } else {
 329                 /*
 330                  * This is just a hint to vm_map_find() about where to put it.
 331                  */
 332                 attach_va = round_page((vm_offset_t)p->p_vmspace->vm_taddr +
 333                                        maxtsiz + maxdsiz);
 334         }
 335
 336         /*
 337          * Handle alignment.  For large memory maps it is possible
 338          * that the MMU can optimize the page table so align anything
 339          * that is a multiple of SEG_SIZE to SEG_SIZE.
 340          */
 341         if ((flags & MAP_FIXED) == 0 && (size & SEG_MASK) == 0)
 342                 align = SEG_SIZE;
 343         else
 344                 align = PAGE_SIZE;
 345
 346         shm_handle = shmseg->shm_internal;
 347         vm_object_hold(shm_handle->shm_object);
 348         vm_object_chain_wait(shm_handle->shm_object, 0);
 349         vm_object_reference_locked(shm_handle->shm_object);
 350         rv = vm_map_find(&p->p_vmspace->vm_map,
 351                          shm_handle->shm_object, NULL,
 352                          0, &attach_va, size,
 353                          align,
 354                          ((flags & MAP_FIXED) ? 0 : 1),
 355                          VM_MAPTYPE_NORMAL, VM_SUBSYS_SHMEM,
 356                          prot, prot, 0);
 357         vm_object_drop(shm_handle->shm_object);
 358         if (rv != KERN_SUCCESS) {
 359                 vm_object_deallocate(shm_handle->shm_object);
 360                 error = ENOMEM;
 361                 goto done;
 362         }
 363         vm_map_inherit(&p->p_vmspace->vm_map,
 364                        attach_va, attach_va + size, VM_INHERIT_SHARE);
 365
 366         KKASSERT(shmmap_s->shmid == -1);
 367         shmmap_s->va = attach_va;
 368         shmmap_s->shmid = uap->shmid;
 369         shmseg->shm_lpid = p->p_pid;
 370         shmseg->shm_atime = time_second;
 371         shmseg->shm_nattch++;
 372         uap->sysmsg_resultp = (void *)attach_va;
 373         error = 0;
 374 done:
 375         lwkt_reltoken(&shm_token);
 376
 377         return error;
 378 }
 379
 380 /*
 381  * MPALMOSTSAFE
 382  */
 383 int
 384 sys_shmctl(struct shmctl_args *uap)
 385 {
 386         struct thread *td = curthread;
 387         struct proc *p = td->td_proc;
 388         int error;
 389         struct shmid_ds inbuf;
 390         struct shmid_ds *shmseg;
 391
 392         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 393                 return (ENOSYS);
 394
 395         lwkt_gettoken(&shm_token);
 396         shmseg = shm_find_segment_by_shmid(uap->shmid);
 397         if (shmseg == NULL) {
 398                 error = EINVAL;
 399                 goto done;
 400         }
 401
 402         switch (uap->cmd) {
 403         case IPC_STAT:
 404                 error = ipcperm(p, &shmseg->shm_perm, IPC_R);
 405                 if (error == 0)
 406                         error = copyout(shmseg, uap->buf, sizeof(inbuf));
 407                 break;
 408         case IPC_SET:
 409                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 410                 if (error == 0)
 411                         error = copyin(uap->buf, &inbuf, sizeof(inbuf));
 412                 if (error == 0) {
 413                         shmseg->shm_perm.uid = inbuf.shm_perm.uid;
 414                         shmseg->shm_perm.gid = inbuf.shm_perm.gid;
 415                         shmseg->shm_perm.mode =
 416                             (shmseg->shm_perm.mode & ~ACCESSPERMS) |
 417                             (inbuf.shm_perm.mode & ACCESSPERMS);
 418                         shmseg->shm_ctime = time_second;
 419                 }
 420                 break;
 421         case IPC_RMID:
 422                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 423                 if (error == 0) {
 424                         shmseg->shm_perm.key = IPC_PRIVATE;
 425                         shmseg->shm_perm.mode |= SHMSEG_REMOVED;
 426                         if (shmseg->shm_nattch <= 0) {
 427                                 shm_deallocate_segment(shmseg);
 428                                 shm_last_free = IPCID_TO_IX(uap->shmid);
 429                         }
 430                 }
 431                 break;
 432 #if 0
 433         case SHM_LOCK:
 434         case SHM_UNLOCK:
 435 #endif
 436         default:
 437                 error = EINVAL;
 438                 break;
 439         }
 440 done:
 441         lwkt_reltoken(&shm_token);
 442
 443         return error;
 444 }
 445
 446 static int
 447 shmget_existing(struct proc *p, struct shmget_args *uap, int mode, int segnum)
 448 {
 449         struct shmid_ds *shmseg;
 450         int error;
 451
 452         shmseg = &shmsegs[segnum];
 453         if (shmseg->shm_perm.mode & SHMSEG_REMOVED) {
 454                 /*
 455                  * This segment is in the process of being allocated.  Wait
 456                  * until it's done, and look the key up again (in case the
 457                  * allocation failed or it was freed).
 458                  */
 459                 shmseg->shm_perm.mode |= SHMSEG_WANTED;
 460                 error = tsleep((caddr_t)shmseg, PCATCH, "shmget", 0);
 461                 if (error)
 462                         return error;
 463                 return EAGAIN;
 464         }
 465         if ((uap->shmflg & (IPC_CREAT | IPC_EXCL)) == (IPC_CREAT | IPC_EXCL))
 466                 return EEXIST;
 467         error = ipcperm(p, &shmseg->shm_perm, mode);
 468         if (error)
 469                 return error;
 470         if (uap->size && uap->size > shmseg->shm_segsz)
 471                 return EINVAL;
 472         uap->sysmsg_result = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 473         return 0;
 474 }
 475
 476 static int
 477 shmget_allocate_segment(struct proc *p, struct shmget_args *uap, int mode)
 478 {
 479         int i, segnum, shmid;
 480         size_t size;
 481         struct ucred *cred = p->p_ucred;
 482         struct shmid_ds *shmseg;
 483         struct shm_handle *shm_handle;
 484
 485         if (uap->size < shminfo.shmmin || uap->size > shminfo.shmmax)
 486                 return EINVAL;
 487         if (shm_nused >= shminfo.shmmni) /* any shmids left? */
 488                 return ENOSPC;
 489         size = round_page(uap->size);
 490         if (shm_committed + btoc(size) > shminfo.shmall)
 491                 return ENOMEM;
 492         if (shm_last_free < 0) {
 493                 shmrealloc();   /* maybe expand the shmsegs[] array */
 494                 for (i = 0; i < shmalloced; i++) {
 495                         if (shmsegs[i].shm_perm.mode & SHMSEG_FREE)
 496                                 break;
 497                 }
 498                 if (i == shmalloced)
 499                         return ENOSPC;
 500                 segnum = i;
 501         } else  {
 502                 segnum = shm_last_free;
 503                 shm_last_free = -1;
 504         }
 505         shmseg = &shmsegs[segnum];
 506         /*
 507          * In case we sleep in malloc(), mark the segment present but deleted
 508          * so that noone else tries to create the same key.
 509          */
 510         shmseg->shm_perm.mode = SHMSEG_ALLOCATED | SHMSEG_REMOVED;
 511         shmseg->shm_perm.key = uap->key;
 512         shmseg->shm_perm.seq = (shmseg->shm_perm.seq + 1) & 0x7fff;
 513         shm_handle = kmalloc(sizeof(struct shm_handle), M_SHM, M_WAITOK);
 514         shmid = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 515
 516         /*
 517          * We make sure that we have allocated a pager before we need
 518          * to.
 519          */
 520         if (shm_use_phys) {
 521                 shm_handle->shm_object =
 522                    phys_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 523         } else {
 524                 shm_handle->shm_object =
 525                    swap_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 526         }
 527         vm_object_clear_flag(shm_handle->shm_object, OBJ_ONEMAPPING);
 528         vm_object_set_flag(shm_handle->shm_object, OBJ_NOSPLIT);
 529
 530         shmseg->shm_internal = shm_handle;
 531         shmseg->shm_perm.cuid = shmseg->shm_perm.uid = cred->cr_uid;
 532         shmseg->shm_perm.cgid = shmseg->shm_perm.gid = cred->cr_gid;
 533         shmseg->shm_perm.mode = (shmseg->shm_perm.mode & SHMSEG_WANTED) |
 534             (mode & ACCESSPERMS) | SHMSEG_ALLOCATED;
 535         shmseg->shm_segsz = uap->size;
 536         shmseg->shm_cpid = p->p_pid;
 537         shmseg->shm_lpid = shmseg->shm_nattch = 0;
 538         shmseg->shm_atime = shmseg->shm_dtime = 0;
 539         shmseg->shm_ctime = time_second;
 540         shm_committed += btoc(size);
 541         shm_nused++;
 542
 543         /*
 544          * If a physical mapping is desired and we have a ton of free pages
 545          * we pre-allocate the pages here in order to avoid on-the-fly
 546          * allocation later.  This has a big effect on database warm-up
 547          * times since DFly supports concurrent page faults coming from the
 548          * same VM object for pages which already exist.
 549          *
 550          * This can hang the kernel for a while so only do it if shm_use_phys
 551          * is set to 2 or higher.
 552          */
 553         if (shm_use_phys > 1) {
 554                 vm_pindex_t pi, pmax;
 555                 vm_page_t m;
 556
 557                 pmax = round_page(shmseg->shm_segsz) >> PAGE_SHIFT;
 558                 vm_object_hold(shm_handle->shm_object);
 559                 if (pmax > vmstats.v_free_count)
 560                         pmax = vmstats.v_free_count;
 561                 for (pi = 0; pi < pmax; ++pi) {
 562                         m = vm_page_grab(shm_handle->shm_object, pi,
 563                                          VM_ALLOC_SYSTEM | VM_ALLOC_NULL_OK |
 564                                          VM_ALLOC_ZERO);
 565                         if (m == NULL)
 566                                 break;
 567                         vm_pager_get_page(shm_handle->shm_object, &m, 1);
 568                         vm_page_activate(m);
 569                         vm_page_wakeup(m);
 570                         lwkt_yield();
 571                 }
 572                 vm_object_drop(shm_handle->shm_object);
 573         }
 574
 575         if (shmseg->shm_perm.mode & SHMSEG_WANTED) {
 576                 /*
 577                  * Somebody else wanted this key while we were asleep.  Wake
 578                  * them up now.
 579                  */
 580                 shmseg->shm_perm.mode &= ~SHMSEG_WANTED;
 581                 wakeup((caddr_t)shmseg);
 582         }
 583         uap->sysmsg_result = shmid;
 584         return 0;
 585 }
 586
 587 /*
 588  * MPALMOSTSAFE
 589  */
 590 int
 591 sys_shmget(struct shmget_args *uap)
 592 {
 593         struct thread *td = curthread;
 594         struct proc *p = td->td_proc;
 595         int segnum, mode, error;
 596
 597         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 598                 return (ENOSYS);
 599
 600         mode = uap->shmflg & ACCESSPERMS;
 601
 602         lwkt_gettoken(&shm_token);
 603
 604         if (uap->key != IPC_PRIVATE) {
 605         again:
 606                 segnum = shm_find_segment_by_key(uap->key);
 607                 if (segnum >= 0) {
 608                         error = shmget_existing(p, uap, mode, segnum);
 609                         if (error == EAGAIN)
 610                                 goto again;
 611                         goto done;
 612                 }
 613                 if ((uap->shmflg & IPC_CREAT) == 0) {
 614                         error = ENOENT;
 615                         goto done;
 616                 }
 617         }
 618         error = shmget_allocate_segment(p, uap, mode);
 619 done:
 620         lwkt_reltoken(&shm_token);
 621
 622         return (error);
 623 }
 624
 625 void
 626 shmfork(struct proc *p1, struct proc *p2)
 627 {
 628         struct shmmap_state *shmmap_s;
 629         size_t size;
 630         int i;
 631
 632         lwkt_gettoken(&shm_token);
 633         size = shminfo.shmseg * sizeof(struct shmmap_state);
 634         shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 635         bcopy((caddr_t)p1->p_vmspace->vm_shm, (caddr_t)shmmap_s, size);
 636         p2->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 637         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 638                 if (shmmap_s->shmid != -1)
 639                         shmsegs[IPCID_TO_IX(shmmap_s->shmid)].shm_nattch++;
 640         }
 641         lwkt_reltoken(&shm_token);
 642 }
 643
 644 void
 645 shmexit(struct vmspace *vm)
 646 {
 647         struct shmmap_state *base, *shm;
 648         int i;
 649
 650         if ((base = (struct shmmap_state *)vm->vm_shm) != NULL) {
 651                 vm->vm_shm = NULL;
 652                 lwkt_gettoken(&shm_token);
 653                 for (i = 0, shm = base; i < shminfo.shmseg; i++, shm++) {
 654                         if (shm->shmid != -1)
 655                                 shm_delete_mapping(vm, shm);
 656                 }
 657                 kfree(base, M_SHM);
 658                 lwkt_reltoken(&shm_token);
 659         }
 660 }
 661
 662 static void
 663 shmrealloc(void)
 664 {
 665         int i;
 666         struct shmid_ds *newsegs;
 667
 668         if (shmalloced >= shminfo.shmmni)
 669                 return;
 670
 671         newsegs = kmalloc(shminfo.shmmni * sizeof(*newsegs), M_SHM, M_WAITOK);
 672         for (i = 0; i < shmalloced; i++)
 673                 bcopy(&shmsegs[i], &newsegs[i], sizeof(newsegs[0]));
 674         for (; i < shminfo.shmmni; i++) {
 675                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 676                 shmsegs[i].shm_perm.seq = 0;
 677         }
 678         kfree(shmsegs, M_SHM);
 679         shmsegs = newsegs;
 680         shmalloced = shminfo.shmmni;
 681 }
 682
 683 static void
 684 shminit(void *dummy)
 685 {
 686         int i;
 687
 688         /*
 689          * If not overridden by a tunable set the maximum shm to
 690          * 2/3 of main memory.
 691          */
 692         if (shminfo.shmall == 0)
 693                 shminfo.shmall = (size_t)vmstats.v_page_count * 2 / 3;
 694
 695         shminfo.shmmax = shminfo.shmall * PAGE_SIZE;
 696         shmalloced = shminfo.shmmni;
 697         shmsegs = kmalloc(shmalloced * sizeof(shmsegs[0]), M_SHM, M_WAITOK);
 698         for (i = 0; i < shmalloced; i++) {
 699                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 700                 shmsegs[i].shm_perm.seq = 0;
 701         }
 702         shm_last_free = 0;
 703         shm_nused = 0;
 704         shm_committed = 0;
 705 }
 706 SYSINIT(sysv_shm, SI_SUB_SYSV_SHM, SI_ORDER_FIRST, shminit, NULL);