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_compat.h"
  32 #include "opt_sysvipc.h"
  33
  34 #include <sys/param.h>
  35 #include <sys/systm.h>
  36 #include <sys/sysproto.h>
  37 #include <sys/kernel.h>
  38 #include <sys/sysctl.h>
  39 #include <sys/shm.h>
  40 #include <sys/proc.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mman.h>
  43 #include <sys/stat.h>
  44 #include <sys/sysent.h>
  45 #include <sys/jail.h>
  46
  47 #include <sys/mplock2.h>
  48
  49 #include <vm/vm.h>
  50 #include <vm/vm_param.h>
  51 #include <sys/lock.h>
  52 #include <vm/pmap.h>
  53 #include <vm/vm_object.h>
  54 #include <vm/vm_map.h>
  55 #include <vm/vm_page.h>
  56 #include <vm/vm_pager.h>
  57
  58 static MALLOC_DEFINE(M_SHM, "shm", "SVID compatible shared memory segments");
  59
  60 static int shmget_allocate_segment (struct proc *p, struct shmget_args *uap, int mode);
  61 static int shmget_existing (struct proc *p, struct shmget_args *uap, int mode, int segnum);
  62
  63 #define SHMSEG_FREE             0x0200
  64 #define SHMSEG_REMOVED          0x0400
  65 #define SHMSEG_ALLOCATED        0x0800
  66 #define SHMSEG_WANTED           0x1000
  67
  68 static int shm_last_free, shm_committed, shmalloced;
  69 int shm_nused;
  70 static struct shmid_ds  *shmsegs;
  71
  72 struct shm_handle {
  73         /* vm_offset_t kva; */
  74         vm_object_t shm_object;
  75 };
  76
  77 struct shmmap_state {
  78         vm_offset_t va;
  79         int shmid;
  80 };
  81
  82 static void shm_deallocate_segment (struct shmid_ds *);
  83 static int shm_find_segment_by_key (key_t);
  84 static struct shmid_ds *shm_find_segment_by_shmid (int);
  85 static int shm_delete_mapping (struct vmspace *vm, struct shmmap_state *);
  86 static void shmrealloc (void);
  87 static void shminit (void *);
  88
  89 /*
  90  * Tuneable values
  91  */
  92 #ifndef SHMMIN
  93 #define SHMMIN  1
  94 #endif
  95 #ifndef SHMMNI
  96 #define SHMMNI  512
  97 #endif
  98 #ifndef SHMSEG
  99 #define SHMSEG  1024
 100 #endif
 101
 102 struct  shminfo shminfo = {
 103         0,
 104         SHMMIN,
 105         SHMMNI,
 106         SHMSEG,
 107         0
 108 };
 109
 110 static int shm_use_phys = 1;
 111
 112 TUNABLE_LONG("kern.ipc.shmmin", &shminfo.shmmin);
 113 TUNABLE_LONG("kern.ipc.shmmni", &shminfo.shmmni);
 114 TUNABLE_LONG("kern.ipc.shmseg", &shminfo.shmseg);
 115 TUNABLE_LONG("kern.ipc.shmmaxpgs", &shminfo.shmall);
 116 TUNABLE_INT("kern.ipc.shm_use_phys", &shm_use_phys);
 117
 118 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmax, CTLFLAG_RW, &shminfo.shmmax, 0,
 119     "Max shared memory segment size");
 120 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmin, CTLFLAG_RW, &shminfo.shmmin, 0,
 121     "Min shared memory segment size");
 122 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmmni, CTLFLAG_RD, &shminfo.shmmni, 0,
 123     "Max number of shared memory identifiers");
 124 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmseg, CTLFLAG_RW, &shminfo.shmseg, 0,
 125     "Max shared memory segments per process");
 126 SYSCTL_LONG(_kern_ipc, OID_AUTO, shmall, CTLFLAG_RW, &shminfo.shmall, 0,
 127     "Max pages of shared memory");
 128 SYSCTL_INT(_kern_ipc, OID_AUTO, shm_use_phys, CTLFLAG_RW, &shm_use_phys, 0,
 129     "Use phys pager allocation instead of swap pager allocation");
 130
 131 static int
 132 shm_find_segment_by_key(key_t key)
 133 {
 134         int i;
 135
 136         for (i = 0; i < shmalloced; i++) {
 137                 if ((shmsegs[i].shm_perm.mode & SHMSEG_ALLOCATED) &&
 138                     shmsegs[i].shm_perm.key == key)
 139                         return i;
 140         }
 141         return -1;
 142 }
 143
 144 static struct shmid_ds *
 145 shm_find_segment_by_shmid(int shmid)
 146 {
 147         int segnum;
 148         struct shmid_ds *shmseg;
 149
 150         segnum = IPCID_TO_IX(shmid);
 151         if (segnum < 0 || segnum >= shmalloced)
 152                 return NULL;
 153         shmseg = &shmsegs[segnum];
 154         if ((shmseg->shm_perm.mode & (SHMSEG_ALLOCATED | SHMSEG_REMOVED))
 155             != SHMSEG_ALLOCATED ||
 156             shmseg->shm_perm.seq != IPCID_TO_SEQ(shmid)) {
 157                 return NULL;
 158         }
 159         return shmseg;
 160 }
 161
 162 static void
 163 shm_deallocate_segment(struct shmid_ds *shmseg)
 164 {
 165         struct shm_handle *shm_handle;
 166         size_t size;
 167
 168         shm_handle = shmseg->shm_internal;
 169         vm_object_deallocate(shm_handle->shm_object);
 170         kfree((caddr_t)shm_handle, M_SHM);
 171         shmseg->shm_internal = NULL;
 172         size = round_page(shmseg->shm_segsz);
 173         shm_committed -= btoc(size);
 174         shm_nused--;
 175         shmseg->shm_perm.mode = SHMSEG_FREE;
 176 }
 177
 178 static int
 179 shm_delete_mapping(struct vmspace *vm, struct shmmap_state *shmmap_s)
 180 {
 181         struct shmid_ds *shmseg;
 182         int segnum, result;
 183         size_t size;
 184
 185         segnum = IPCID_TO_IX(shmmap_s->shmid);
 186         shmseg = &shmsegs[segnum];
 187         size = round_page(shmseg->shm_segsz);
 188         result = vm_map_remove(&vm->vm_map, shmmap_s->va, shmmap_s->va + size);
 189         if (result != KERN_SUCCESS)
 190                 return EINVAL;
 191         shmmap_s->shmid = -1;
 192         shmseg->shm_dtime = time_second;
 193         if ((--shmseg->shm_nattch <= 0) &&
 194             (shmseg->shm_perm.mode & SHMSEG_REMOVED)) {
 195                 shm_deallocate_segment(shmseg);
 196                 shm_last_free = segnum;
 197         }
 198         return 0;
 199 }
 200
 201 /*
 202  * MPALMOSTSAFE
 203  */
 204 int
 205 sys_shmdt(struct shmdt_args *uap)
 206 {
 207         struct thread *td = curthread;
 208         struct proc *p = td->td_proc;
 209         struct shmmap_state *shmmap_s;
 210         long i;
 211         int error;
 212
 213         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 214                 return (ENOSYS);
 215
 216         get_mplock();
 217         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 218         if (shmmap_s == NULL) {
 219                 error = EINVAL;
 220                 goto done;
 221         }
 222         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 223                 if (shmmap_s->shmid != -1 &&
 224                     shmmap_s->va == (vm_offset_t)uap->shmaddr)
 225                         break;
 226         }
 227         if (i == shminfo.shmseg)
 228                 error = EINVAL;
 229         else
 230                 error = shm_delete_mapping(p->p_vmspace, shmmap_s);
 231 done:
 232         rel_mplock();
 233         return (error);
 234 }
 235
 236 /*
 237  * MPALMOSTSAFE
 238  */
 239 int
 240 sys_shmat(struct shmat_args *uap)
 241 {
 242         struct thread *td = curthread;
 243         struct proc *p = td->td_proc;
 244         int error, flags;
 245         long i;
 246         struct shmid_ds *shmseg;
 247         struct shmmap_state *shmmap_s = NULL;
 248         struct shm_handle *shm_handle;
 249         vm_offset_t attach_va;
 250         vm_prot_t prot;
 251         vm_size_t size;
 252         vm_size_t align;
 253         int rv;
 254
 255         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 256                 return (ENOSYS);
 257
 258         get_mplock();
 259 again:
 260         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 261         if (shmmap_s == NULL) {
 262                 size = shminfo.shmseg * sizeof(struct shmmap_state);
 263                 shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 264                 for (i = 0; i < shminfo.shmseg; i++)
 265                         shmmap_s[i].shmid = -1;
 266                 if (p->p_vmspace->vm_shm != NULL) {
 267                         kfree(shmmap_s, M_SHM);
 268                         goto again;
 269                 }
 270                 p->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 271         }
 272         shmseg = shm_find_segment_by_shmid(uap->shmid);
 273         if (shmseg == NULL) {
 274                 error = EINVAL;
 275                 goto done;
 276         }
 277         error = ipcperm(p, &shmseg->shm_perm,
 278                         (uap->shmflg & SHM_RDONLY) ? IPC_R : IPC_R|IPC_W);
 279         if (error)
 280                 goto done;
 281         for (i = 0; i < shminfo.shmseg; i++) {
 282                 if (shmmap_s->shmid == -1)
 283                         break;
 284                 shmmap_s++;
 285         }
 286         if (i >= shminfo.shmseg) {
 287                 error = EMFILE;
 288                 goto done;
 289         }
 290         size = round_page(shmseg->shm_segsz);
 291 #ifdef VM_PROT_READ_IS_EXEC
 292         prot = VM_PROT_READ | VM_PROT_EXECUTE;
 293 #else
 294         prot = VM_PROT_READ;
 295 #endif
 296         if ((uap->shmflg & SHM_RDONLY) == 0)
 297                 prot |= VM_PROT_WRITE;
 298         flags = MAP_ANON | MAP_SHARED;
 299         if (uap->shmaddr) {
 300                 flags |= MAP_FIXED;
 301                 if (uap->shmflg & SHM_RND) {
 302                         attach_va = (vm_offset_t)uap->shmaddr & ~(SHMLBA-1);
 303                 } else if (((vm_offset_t)uap->shmaddr & (SHMLBA-1)) == 0) {
 304                         attach_va = (vm_offset_t)uap->shmaddr;
 305                 } else {
 306                         error = EINVAL;
 307                         goto done;
 308                 }
 309         } else {
 310                 /*
 311                  * This is just a hint to vm_map_find() about where to put it.
 312                  */
 313                 attach_va = round_page((vm_offset_t)p->p_vmspace->vm_taddr +
 314                                        maxtsiz + maxdsiz);
 315         }
 316
 317         /*
 318          * Handle alignment.  For large memory maps it is possible
 319          * that the MMU can optimize the page table so align anything
 320          * that is a multiple of SEG_SIZE to SEG_SIZE.
 321          */
 322         if ((flags & MAP_FIXED) == 0 && (size & SEG_MASK) == 0)
 323                 align = SEG_SIZE;
 324         else
 325                 align = PAGE_SIZE;
 326
 327         shm_handle = shmseg->shm_internal;
 328         vm_object_hold(shm_handle->shm_object);
 329         vm_object_chain_wait(shm_handle->shm_object, 0);
 330         vm_object_reference_locked(shm_handle->shm_object);
 331         rv = vm_map_find(&p->p_vmspace->vm_map,
 332                          shm_handle->shm_object, NULL,
 333                          0, &attach_va, size,
 334                          align,
 335                          ((flags & MAP_FIXED) ? 0 : 1),
 336                          VM_MAPTYPE_NORMAL,
 337                          prot, prot, 0);
 338         vm_object_drop(shm_handle->shm_object);
 339         if (rv != KERN_SUCCESS) {
 340                 vm_object_deallocate(shm_handle->shm_object);
 341                 error = ENOMEM;
 342                 goto done;
 343         }
 344         vm_map_inherit(&p->p_vmspace->vm_map,
 345                        attach_va, attach_va + size, VM_INHERIT_SHARE);
 346
 347         KKASSERT(shmmap_s->shmid == -1);
 348         shmmap_s->va = attach_va;
 349         shmmap_s->shmid = uap->shmid;
 350         shmseg->shm_lpid = p->p_pid;
 351         shmseg->shm_atime = time_second;
 352         shmseg->shm_nattch++;
 353         uap->sysmsg_resultp = (void *)attach_va;
 354         error = 0;
 355 done:
 356         rel_mplock();
 357         return error;
 358 }
 359
 360 /*
 361  * MPALMOSTSAFE
 362  */
 363 int
 364 sys_shmctl(struct shmctl_args *uap)
 365 {
 366         struct thread *td = curthread;
 367         struct proc *p = td->td_proc;
 368         int error;
 369         struct shmid_ds inbuf;
 370         struct shmid_ds *shmseg;
 371
 372         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 373                 return (ENOSYS);
 374
 375         get_mplock();
 376         shmseg = shm_find_segment_by_shmid(uap->shmid);
 377         if (shmseg == NULL) {
 378                 error = EINVAL;
 379                 goto done;
 380         }
 381
 382         switch (uap->cmd) {
 383         case IPC_STAT:
 384                 error = ipcperm(p, &shmseg->shm_perm, IPC_R);
 385                 if (error == 0)
 386                         error = copyout(shmseg, uap->buf, sizeof(inbuf));
 387                 break;
 388         case IPC_SET:
 389                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 390                 if (error == 0)
 391                         error = copyin(uap->buf, &inbuf, sizeof(inbuf));
 392                 if (error == 0) {
 393                         shmseg->shm_perm.uid = inbuf.shm_perm.uid;
 394                         shmseg->shm_perm.gid = inbuf.shm_perm.gid;
 395                         shmseg->shm_perm.mode =
 396                             (shmseg->shm_perm.mode & ~ACCESSPERMS) |
 397                             (inbuf.shm_perm.mode & ACCESSPERMS);
 398                         shmseg->shm_ctime = time_second;
 399                 }
 400                 break;
 401         case IPC_RMID:
 402                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 403                 if (error == 0) {
 404                         shmseg->shm_perm.key = IPC_PRIVATE;
 405                         shmseg->shm_perm.mode |= SHMSEG_REMOVED;
 406                         if (shmseg->shm_nattch <= 0) {
 407                                 shm_deallocate_segment(shmseg);
 408                                 shm_last_free = IPCID_TO_IX(uap->shmid);
 409                         }
 410                 }
 411                 break;
 412 #if 0
 413         case SHM_LOCK:
 414         case SHM_UNLOCK:
 415 #endif
 416         default:
 417                 error = EINVAL;
 418                 break;
 419         }
 420 done:
 421         rel_mplock();
 422         return error;
 423 }
 424
 425 static int
 426 shmget_existing(struct proc *p, struct shmget_args *uap, int mode, int segnum)
 427 {
 428         struct shmid_ds *shmseg;
 429         int error;
 430
 431         shmseg = &shmsegs[segnum];
 432         if (shmseg->shm_perm.mode & SHMSEG_REMOVED) {
 433                 /*
 434                  * This segment is in the process of being allocated.  Wait
 435                  * until it's done, and look the key up again (in case the
 436                  * allocation failed or it was freed).
 437                  */
 438                 shmseg->shm_perm.mode |= SHMSEG_WANTED;
 439                 error = tsleep((caddr_t)shmseg, PCATCH, "shmget", 0);
 440                 if (error)
 441                         return error;
 442                 return EAGAIN;
 443         }
 444         if ((uap->shmflg & (IPC_CREAT | IPC_EXCL)) == (IPC_CREAT | IPC_EXCL))
 445                 return EEXIST;
 446         error = ipcperm(p, &shmseg->shm_perm, mode);
 447         if (error)
 448                 return error;
 449         if (uap->size && uap->size > shmseg->shm_segsz)
 450                 return EINVAL;
 451         uap->sysmsg_result = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 452         return 0;
 453 }
 454
 455 static int
 456 shmget_allocate_segment(struct proc *p, struct shmget_args *uap, int mode)
 457 {
 458         int i, segnum, shmid;
 459         size_t size;
 460         struct ucred *cred = p->p_ucred;
 461         struct shmid_ds *shmseg;
 462         struct shm_handle *shm_handle;
 463
 464         if (uap->size < shminfo.shmmin || uap->size > shminfo.shmmax)
 465                 return EINVAL;
 466         if (shm_nused >= shminfo.shmmni) /* any shmids left? */
 467                 return ENOSPC;
 468         size = round_page(uap->size);
 469         if (shm_committed + btoc(size) > shminfo.shmall)
 470                 return ENOMEM;
 471         if (shm_last_free < 0) {
 472                 shmrealloc();   /* maybe expand the shmsegs[] array */
 473                 for (i = 0; i < shmalloced; i++) {
 474                         if (shmsegs[i].shm_perm.mode & SHMSEG_FREE)
 475                                 break;
 476                 }
 477                 if (i == shmalloced)
 478                         return ENOSPC;
 479                 segnum = i;
 480         } else  {
 481                 segnum = shm_last_free;
 482                 shm_last_free = -1;
 483         }
 484         shmseg = &shmsegs[segnum];
 485         /*
 486          * In case we sleep in malloc(), mark the segment present but deleted
 487          * so that noone else tries to create the same key.
 488          */
 489         shmseg->shm_perm.mode = SHMSEG_ALLOCATED | SHMSEG_REMOVED;
 490         shmseg->shm_perm.key = uap->key;
 491         shmseg->shm_perm.seq = (shmseg->shm_perm.seq + 1) & 0x7fff;
 492         shm_handle = kmalloc(sizeof(struct shm_handle), M_SHM, M_WAITOK);
 493         shmid = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 494
 495         /*
 496          * We make sure that we have allocated a pager before we need
 497          * to.
 498          */
 499         if (shm_use_phys) {
 500                 shm_handle->shm_object =
 501                    phys_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 502         } else {
 503                 shm_handle->shm_object =
 504                    swap_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 505         }
 506         vm_object_clear_flag(shm_handle->shm_object, OBJ_ONEMAPPING);
 507         vm_object_set_flag(shm_handle->shm_object, OBJ_NOSPLIT);
 508
 509         shmseg->shm_internal = shm_handle;
 510         shmseg->shm_perm.cuid = shmseg->shm_perm.uid = cred->cr_uid;
 511         shmseg->shm_perm.cgid = shmseg->shm_perm.gid = cred->cr_gid;
 512         shmseg->shm_perm.mode = (shmseg->shm_perm.mode & SHMSEG_WANTED) |
 513             (mode & ACCESSPERMS) | SHMSEG_ALLOCATED;
 514         shmseg->shm_segsz = uap->size;
 515         shmseg->shm_cpid = p->p_pid;
 516         shmseg->shm_lpid = shmseg->shm_nattch = 0;
 517         shmseg->shm_atime = shmseg->shm_dtime = 0;
 518         shmseg->shm_ctime = time_second;
 519         shm_committed += btoc(size);
 520         shm_nused++;
 521
 522         /*
 523          * If a physical mapping is desired and we have a ton of free pages
 524          * we pre-allocate the pages here in order to avoid on-the-fly
 525          * allocation later.  This has a big effect on database warm-up
 526          * times since DFly supports concurrent page faults coming from the
 527          * same VM object for pages which already exist.
 528          *
 529          * This can hang the kernel for a while so only do it if shm_use_phys
 530          * is set to 2 or higher.
 531          */
 532         if (shm_use_phys > 1) {
 533                 vm_pindex_t pi, pmax;
 534                 vm_page_t m;
 535
 536                 pmax = round_page(shmseg->shm_segsz) >> PAGE_SHIFT;
 537                 vm_object_hold(shm_handle->shm_object);
 538                 if (pmax > vmstats.v_free_count)
 539                         pmax = vmstats.v_free_count;
 540                 for (pi = 0; pi < pmax; ++pi) {
 541                         m = vm_page_grab(shm_handle->shm_object, pi,
 542                                          VM_ALLOC_SYSTEM | VM_ALLOC_NULL_OK |
 543                                          VM_ALLOC_ZERO);
 544                         if (m == NULL)
 545                                 break;
 546                         vm_pager_get_page(shm_handle->shm_object, &m, 1);
 547                         vm_page_activate(m);
 548                         vm_page_wakeup(m);
 549                         lwkt_yield();
 550                 }
 551                 vm_object_drop(shm_handle->shm_object);
 552         }
 553
 554         if (shmseg->shm_perm.mode & SHMSEG_WANTED) {
 555                 /*
 556                  * Somebody else wanted this key while we were asleep.  Wake
 557                  * them up now.
 558                  */
 559                 shmseg->shm_perm.mode &= ~SHMSEG_WANTED;
 560                 wakeup((caddr_t)shmseg);
 561         }
 562         uap->sysmsg_result = shmid;
 563         return 0;
 564 }
 565
 566 /*
 567  * MPALMOSTSAFE
 568  */
 569 int
 570 sys_shmget(struct shmget_args *uap)
 571 {
 572         struct thread *td = curthread;
 573         struct proc *p = td->td_proc;
 574         int segnum, mode, error;
 575
 576         if (!jail_sysvipc_allowed && td->td_ucred->cr_prison != NULL)
 577                 return (ENOSYS);
 578
 579         mode = uap->shmflg & ACCESSPERMS;
 580         get_mplock();
 581
 582         if (uap->key != IPC_PRIVATE) {
 583         again:
 584                 segnum = shm_find_segment_by_key(uap->key);
 585                 if (segnum >= 0) {
 586                         error = shmget_existing(p, uap, mode, segnum);
 587                         if (error == EAGAIN)
 588                                 goto again;
 589                         goto done;
 590                 }
 591                 if ((uap->shmflg & IPC_CREAT) == 0) {
 592                         error = ENOENT;
 593                         goto done;
 594                 }
 595         }
 596         error = shmget_allocate_segment(p, uap, mode);
 597 done:
 598         rel_mplock();
 599         return (error);
 600 }
 601
 602 void
 603 shmfork(struct proc *p1, struct proc *p2)
 604 {
 605         struct shmmap_state *shmmap_s;
 606         size_t size;
 607         int i;
 608
 609         get_mplock();
 610         size = shminfo.shmseg * sizeof(struct shmmap_state);
 611         shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 612         bcopy((caddr_t)p1->p_vmspace->vm_shm, (caddr_t)shmmap_s, size);
 613         p2->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 614         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 615                 if (shmmap_s->shmid != -1)
 616                         shmsegs[IPCID_TO_IX(shmmap_s->shmid)].shm_nattch++;
 617         }
 618         rel_mplock();
 619 }
 620
 621 void
 622 shmexit(struct vmspace *vm)
 623 {
 624         struct shmmap_state *base, *shm;
 625         int i;
 626
 627         if ((base = (struct shmmap_state *)vm->vm_shm) != NULL) {
 628                 vm->vm_shm = NULL;
 629                 get_mplock();
 630                 for (i = 0, shm = base; i < shminfo.shmseg; i++, shm++) {
 631                         if (shm->shmid != -1)
 632                                 shm_delete_mapping(vm, shm);
 633                 }
 634                 kfree(base, M_SHM);
 635                 rel_mplock();
 636         }
 637 }
 638
 639 static void
 640 shmrealloc(void)
 641 {
 642         int i;
 643         struct shmid_ds *newsegs;
 644
 645         if (shmalloced >= shminfo.shmmni)
 646                 return;
 647
 648         newsegs = kmalloc(shminfo.shmmni * sizeof(*newsegs), M_SHM, M_WAITOK);
 649         for (i = 0; i < shmalloced; i++)
 650                 bcopy(&shmsegs[i], &newsegs[i], sizeof(newsegs[0]));
 651         for (; i < shminfo.shmmni; i++) {
 652                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 653                 shmsegs[i].shm_perm.seq = 0;
 654         }
 655         kfree(shmsegs, M_SHM);
 656         shmsegs = newsegs;
 657         shmalloced = shminfo.shmmni;
 658 }
 659
 660 static void
 661 shminit(void *dummy)
 662 {
 663         int i;
 664
 665         /*
 666          * If not overridden by a tunable set the maximum shm to
 667          * 2/3 of main memory.
 668          */
 669         if (shminfo.shmall == 0)
 670                 shminfo.shmall = (size_t)vmstats.v_page_count * 2 / 3;
 671
 672         shminfo.shmmax = shminfo.shmall * PAGE_SIZE;
 673         shmalloced = shminfo.shmmni;
 674         shmsegs = kmalloc(shmalloced * sizeof(shmsegs[0]), M_SHM, M_WAITOK);
 675         for (i = 0; i < shmalloced; i++) {
 676                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 677                 shmsegs[i].shm_perm.seq = 0;
 678         }
 679         shm_last_free = 0;
 680         shm_nused = 0;
 681         shm_committed = 0;
 682 }
 683 SYSINIT(sysv_shm, SI_SUB_SYSV_SHM, SI_ORDER_FIRST, shminit, NULL);