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/sysmsg.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 sysmsg *sysmsg,
  58                         const struct shmget_args *uap, int mode);
  59 static int shmget_existing(struct proc *p, struct sysmsg *sysmsg,
  60                         const struct shmget_args *uap, int mode, int segnum);
  61
  62 #define SHMSEG_FREE             0x0200
  63 #define SHMSEG_REMOVED          0x0400
  64 #define SHMSEG_ALLOCATED        0x0800
  65 #define SHMSEG_WANTED           0x1000
  66
  67 static int shm_last_free, shm_committed, shmalloced, shm_nused;
  68 static struct shmid_ds  *shmsegs;
  69 static struct lwkt_token shm_token = LWKT_TOKEN_INITIALIZER(shm_token);
  70
  71 struct shm_handle {
  72         /* vm_offset_t kva; */
  73         vm_object_t shm_object;
  74 };
  75
  76 struct shmmap_state {
  77         vm_offset_t va;
  78         int shmid;
  79         int reserved;
  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 /*
 111  * allow-removed    Allow a shared memory segment to be attached by its shmid
 112  *                  even after it has been deleted, as long as it was still
 113  *                  being referenced by someone.  This is a trick used by
 114  *                  chrome and other applications to avoid leaving shm
 115  *                  segments hanging around after the application is killed
 116  *                  or seg-faults unexpectedly.
 117  *
 118  * use-phys         Shared memory segments are to use physical memory by
 119  *                  default, which may allow the kernel to better-optimize
 120  *                  the pmap and reduce overhead.  The pages are effectively
 121  *                  wired.
 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 sysmsg *sysmsg, const struct shmdt_args *uap)
 224 {
 225         struct thread *td = curthread;
 226         struct proc *p = td->td_proc;
 227         struct shmmap_state *shmmap_s;
 228         struct prison *pr = p->p_ucred->cr_prison;
 229
 230         long i;
 231         int error;
 232
 233         if (pr && !PRISON_CAP_ISSET(pr->pr_caps, PRISON_CAP_SYS_SYSVIPC))
 234                 return (ENOSYS);
 235
 236         lwkt_gettoken(&shm_token);
 237         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 238         if (shmmap_s == NULL) {
 239                 error = EINVAL;
 240                 goto done;
 241         }
 242         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 243                 if (shmmap_s->shmid != -1 &&
 244                     shmmap_s->va == (vm_offset_t)uap->shmaddr)
 245                         break;
 246         }
 247         if (i == shminfo.shmseg)
 248                 error = EINVAL;
 249         else
 250                 error = shm_delete_mapping(p->p_vmspace, shmmap_s);
 251 done:
 252         lwkt_reltoken(&shm_token);
 253
 254         return (error);
 255 }
 256
 257 /*
 258  * MPALMOSTSAFE
 259  */
 260 int
 261 sys_shmat(struct sysmsg *sysmsg, const struct shmat_args *uap)
 262 {
 263         struct thread *td = curthread;
 264         struct proc *p = td->td_proc;
 265         struct prison *pr = p->p_ucred->cr_prison;
 266         int error, flags;
 267         long i;
 268         struct shmid_ds *shmseg;
 269         struct shmmap_state *shmmap_s = NULL;
 270         struct shm_handle *shm_handle;
 271         vm_offset_t attach_va;
 272         vm_prot_t prot;
 273         vm_size_t size;
 274         vm_size_t align;
 275         int rv;
 276
 277         if (pr && !PRISON_CAP_ISSET(pr->pr_caps, PRISON_CAP_SYS_SYSVIPC))
 278                 return (ENOSYS);
 279
 280         lwkt_gettoken(&shm_token);
 281 again:
 282         shmmap_s = (struct shmmap_state *)p->p_vmspace->vm_shm;
 283         if (shmmap_s == NULL) {
 284                 size = shminfo.shmseg * sizeof(struct shmmap_state);
 285                 shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 286                 for (i = 0; i < shminfo.shmseg; i++) {
 287                         shmmap_s[i].shmid = -1;
 288                         shmmap_s[i].reserved = 0;
 289                 }
 290                 if (p->p_vmspace->vm_shm != NULL) {
 291                         kfree(shmmap_s, M_SHM);
 292                         goto again;
 293                 }
 294                 p->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 295         }
 296         shmseg = shm_find_segment_by_shmid(uap->shmid);
 297         if (shmseg == NULL) {
 298                 error = EINVAL;
 299                 goto done;
 300         }
 301         error = ipcperm(p, &shmseg->shm_perm,
 302                         (uap->shmflg & SHM_RDONLY) ? IPC_R : IPC_R|IPC_W);
 303         if (error)
 304                 goto done;
 305
 306         /*
 307          * Find a free element and mark reserved.  This fixes races
 308          * against concurrent allocations due to the token being
 309          * interrupted by blocking operations.  The shmmap_s reservation
 310          * will be cleared upon completion or error.
 311          */
 312         for (i = 0; i < shminfo.shmseg; i++) {
 313                 if (shmmap_s->shmid == -1 && shmmap_s->reserved == 0) {
 314                         shmmap_s->reserved = 1;
 315                         break;
 316                 }
 317                 shmmap_s++;
 318         }
 319         if (i >= shminfo.shmseg) {
 320                 error = EMFILE;
 321                 goto done;
 322         }
 323         size = round_page(shmseg->shm_segsz);
 324 #ifdef VM_PROT_READ_IS_EXEC
 325         prot = VM_PROT_READ | VM_PROT_EXECUTE;
 326 #else
 327         prot = VM_PROT_READ;
 328 #endif
 329         if ((uap->shmflg & SHM_RDONLY) == 0)
 330                 prot |= VM_PROT_WRITE;
 331         flags = MAP_ANON | MAP_SHARED;
 332         if (uap->shmaddr) {
 333                 flags |= MAP_FIXED;
 334                 if (uap->shmflg & SHM_RND) {
 335                         attach_va =
 336                             rounddown2((vm_offset_t)uap->shmaddr, SHMLBA);
 337                 } else if (((vm_offset_t)uap->shmaddr & (SHMLBA-1)) == 0) {
 338                         attach_va = (vm_offset_t)uap->shmaddr;
 339                 } else {
 340                         error = EINVAL;
 341                         shmmap_s->reserved = 0;
 342                         goto done;
 343                 }
 344         } else {
 345                 /*
 346                  * This is just a hint to vm_map_find() about where to put it.
 347                  */
 348                 attach_va = round_page((vm_offset_t)p->p_vmspace->vm_taddr +
 349                                        maxtsiz + maxdsiz);
 350         }
 351
 352         /*
 353          * Handle alignment.  For large memory maps it is possible
 354          * that the MMU can optimize the page table so align anything
 355          * that is a multiple of SEG_SIZE to SEG_SIZE.
 356          */
 357         if ((flags & MAP_FIXED) == 0 && (size & SEG_MASK) == 0)
 358                 align = SEG_SIZE;
 359         else
 360                 align = PAGE_SIZE;
 361
 362         shm_handle = shmseg->shm_internal;
 363         vm_object_hold(shm_handle->shm_object);
 364         vm_object_reference_locked(shm_handle->shm_object);
 365         rv = vm_map_find(&p->p_vmspace->vm_map,
 366                          shm_handle->shm_object, NULL,
 367                          0, &attach_va, size,
 368                          align,
 369                          ((flags & MAP_FIXED) ? 0 : 1),
 370                          VM_MAPTYPE_NORMAL, VM_SUBSYS_SHMEM,
 371                          prot, prot, 0);
 372         vm_object_drop(shm_handle->shm_object);
 373         if (rv != KERN_SUCCESS) {
 374                 vm_object_deallocate(shm_handle->shm_object);
 375                 shmmap_s->reserved = 0;
 376                 error = ENOMEM;
 377                 goto done;
 378         }
 379         vm_map_inherit(&p->p_vmspace->vm_map,
 380                        attach_va, attach_va + size, VM_INHERIT_SHARE);
 381
 382         KKASSERT(shmmap_s->shmid == -1);
 383         shmmap_s->va = attach_va;
 384         shmmap_s->shmid = uap->shmid;
 385         shmmap_s->reserved = 0;
 386         shmseg->shm_lpid = p->p_pid;
 387         shmseg->shm_atime = time_second;
 388         shmseg->shm_nattch++;
 389         sysmsg->sysmsg_resultp = (void *)attach_va;
 390         error = 0;
 391 done:
 392         lwkt_reltoken(&shm_token);
 393
 394         return error;
 395 }
 396
 397 /*
 398  * MPALMOSTSAFE
 399  */
 400 int
 401 sys_shmctl(struct sysmsg *sysmsg, const struct shmctl_args *uap)
 402 {
 403         struct thread *td = curthread;
 404         struct proc *p = td->td_proc;
 405         struct prison *pr = p->p_ucred->cr_prison;
 406         int error;
 407         struct shmid_ds inbuf;
 408         struct shmid_ds *shmseg;
 409
 410         if (pr && !PRISON_CAP_ISSET(pr->pr_caps, PRISON_CAP_SYS_SYSVIPC))
 411                 return (ENOSYS);
 412
 413         lwkt_gettoken(&shm_token);
 414         shmseg = shm_find_segment_by_shmid(uap->shmid);
 415         if (shmseg == NULL) {
 416                 error = EINVAL;
 417                 goto done;
 418         }
 419
 420         switch (uap->cmd) {
 421         case IPC_STAT:
 422                 error = ipcperm(p, &shmseg->shm_perm, IPC_R);
 423                 if (error == 0)
 424                         error = copyout(shmseg, uap->buf, sizeof(inbuf));
 425                 break;
 426         case IPC_SET:
 427                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 428                 if (error == 0)
 429                         error = copyin(uap->buf, &inbuf, sizeof(inbuf));
 430                 if (error == 0) {
 431                         shmseg->shm_perm.uid = inbuf.shm_perm.uid;
 432                         shmseg->shm_perm.gid = inbuf.shm_perm.gid;
 433                         shmseg->shm_perm.mode =
 434                             (shmseg->shm_perm.mode & ~ACCESSPERMS) |
 435                             (inbuf.shm_perm.mode & ACCESSPERMS);
 436                         shmseg->shm_ctime = time_second;
 437                 }
 438                 break;
 439         case IPC_RMID:
 440                 error = ipcperm(p, &shmseg->shm_perm, IPC_M);
 441                 if (error == 0) {
 442                         shmseg->shm_perm.key = IPC_PRIVATE;
 443                         shmseg->shm_perm.mode |= SHMSEG_REMOVED;
 444                         if (shmseg->shm_nattch <= 0) {
 445                                 shm_deallocate_segment(shmseg);
 446                                 shm_last_free = IPCID_TO_IX(uap->shmid);
 447                         }
 448                 }
 449                 break;
 450 #if 0
 451         case SHM_LOCK:
 452         case SHM_UNLOCK:
 453 #endif
 454         default:
 455                 error = EINVAL;
 456                 break;
 457         }
 458 done:
 459         lwkt_reltoken(&shm_token);
 460
 461         return error;
 462 }
 463
 464 static int
 465 shmget_existing(struct proc *p, struct sysmsg *sysmsg,
 466                 const struct shmget_args *uap, int mode, int segnum)
 467 {
 468         struct shmid_ds *shmseg;
 469         int error;
 470
 471         shmseg = &shmsegs[segnum];
 472         if (shmseg->shm_perm.mode & SHMSEG_REMOVED) {
 473                 /*
 474                  * This segment is in the process of being allocated.  Wait
 475                  * until it's done, and look the key up again (in case the
 476                  * allocation failed or it was freed).
 477                  */
 478                 shmseg->shm_perm.mode |= SHMSEG_WANTED;
 479                 error = tsleep((caddr_t)shmseg, PCATCH, "shmget", 0);
 480                 if (error)
 481                         return error;
 482                 return EAGAIN;
 483         }
 484         if ((uap->shmflg & (IPC_CREAT | IPC_EXCL)) == (IPC_CREAT | IPC_EXCL))
 485                 return EEXIST;
 486         error = ipcperm(p, &shmseg->shm_perm, mode);
 487         if (error)
 488                 return error;
 489         if (uap->size && uap->size > shmseg->shm_segsz)
 490                 return EINVAL;
 491         sysmsg->sysmsg_result = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 492         return 0;
 493 }
 494
 495 static int
 496 shmget_allocate_segment(struct proc *p, struct sysmsg *sysmsg,
 497                         const struct shmget_args *uap, int mode)
 498 {
 499         int i, segnum, shmid;
 500         size_t size;
 501         struct ucred *cred = p->p_ucred;
 502         struct shmid_ds *shmseg;
 503         struct shm_handle *shm_handle;
 504
 505         if (uap->size < shminfo.shmmin || uap->size > shminfo.shmmax)
 506                 return EINVAL;
 507         if (shm_nused >= shminfo.shmmni) /* any shmids left? */
 508                 return ENOSPC;
 509         size = round_page(uap->size);
 510         if (shm_committed + btoc(size) > shminfo.shmall)
 511                 return ENOMEM;
 512         if (shm_last_free < 0) {
 513                 shmrealloc();   /* maybe expand the shmsegs[] array */
 514                 for (i = 0; i < shmalloced; i++) {
 515                         if (shmsegs[i].shm_perm.mode & SHMSEG_FREE)
 516                                 break;
 517                 }
 518                 if (i == shmalloced)
 519                         return ENOSPC;
 520                 segnum = i;
 521         } else  {
 522                 segnum = shm_last_free;
 523                 shm_last_free = -1;
 524         }
 525         shmseg = &shmsegs[segnum];
 526         /*
 527          * In case we sleep in malloc(), mark the segment present but deleted
 528          * so that noone else tries to create the same key.
 529          */
 530         shmseg->shm_perm.mode = SHMSEG_ALLOCATED | SHMSEG_REMOVED;
 531         shmseg->shm_perm.key = uap->key;
 532         shmseg->shm_perm.seq = (shmseg->shm_perm.seq + 1) & 0x7fff;
 533         shm_handle = kmalloc(sizeof(struct shm_handle), M_SHM, M_WAITOK);
 534         shmid = IXSEQ_TO_IPCID(segnum, shmseg->shm_perm);
 535
 536         /*
 537          * We make sure that we have allocated a pager before we need
 538          * to.
 539          */
 540         if (shm_use_phys) {
 541                 shm_handle->shm_object =
 542                    phys_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 543         } else {
 544                 shm_handle->shm_object =
 545                    swap_pager_alloc(NULL, size, VM_PROT_DEFAULT, 0);
 546         }
 547         vm_object_clear_flag(shm_handle->shm_object, OBJ_ONEMAPPING);
 548         vm_object_set_flag(shm_handle->shm_object, OBJ_NOSPLIT);
 549
 550         shmseg->shm_internal = shm_handle;
 551         shmseg->shm_perm.cuid = shmseg->shm_perm.uid = cred->cr_uid;
 552         shmseg->shm_perm.cgid = shmseg->shm_perm.gid = cred->cr_gid;
 553         shmseg->shm_perm.mode = (shmseg->shm_perm.mode & SHMSEG_WANTED) |
 554             (mode & ACCESSPERMS) | SHMSEG_ALLOCATED;
 555         shmseg->shm_segsz = uap->size;
 556         shmseg->shm_cpid = p->p_pid;
 557         shmseg->shm_lpid = shmseg->shm_nattch = 0;
 558         shmseg->shm_atime = shmseg->shm_dtime = 0;
 559         shmseg->shm_ctime = time_second;
 560         shm_committed += btoc(size);
 561         shm_nused++;
 562
 563         /*
 564          * If a physical mapping is desired and we have a ton of free pages
 565          * we pre-allocate the pages here in order to avoid on-the-fly
 566          * allocation later.  This has a big effect on database warm-up
 567          * times since DFly supports concurrent page faults coming from the
 568          * same VM object for pages which already exist.
 569          *
 570          * This can hang the kernel for a while so only do it if shm_use_phys
 571          * is set to 2 or higher.
 572          */
 573         if (shm_use_phys > 1) {
 574                 vm_pindex_t pi, pmax;
 575                 vm_page_t m;
 576
 577                 pmax = round_page(shmseg->shm_segsz) >> PAGE_SHIFT;
 578                 vm_object_hold(shm_handle->shm_object);
 579                 if (pmax > vmstats.v_free_count)
 580                         pmax = vmstats.v_free_count;
 581                 for (pi = 0; pi < pmax; ++pi) {
 582                         m = vm_page_grab(shm_handle->shm_object, pi,
 583                                          VM_ALLOC_SYSTEM | VM_ALLOC_NULL_OK |
 584                                          VM_ALLOC_ZERO);
 585                         if (m == NULL)
 586                                 break;
 587                         vm_pager_get_page(shm_handle->shm_object, pi, &m, 1);
 588                         vm_page_activate(m);
 589                         vm_page_wakeup(m);
 590                         lwkt_yield();
 591                 }
 592                 vm_object_drop(shm_handle->shm_object);
 593         }
 594
 595         if (shmseg->shm_perm.mode & SHMSEG_WANTED) {
 596                 /*
 597                  * Somebody else wanted this key while we were asleep.  Wake
 598                  * them up now.
 599                  */
 600                 shmseg->shm_perm.mode &= ~SHMSEG_WANTED;
 601                 wakeup((caddr_t)shmseg);
 602         }
 603         sysmsg->sysmsg_result = shmid;
 604         return 0;
 605 }
 606
 607 /*
 608  * MPALMOSTSAFE
 609  */
 610 int
 611 sys_shmget(struct sysmsg *sysmsg, const struct shmget_args *uap)
 612 {
 613         struct thread *td = curthread;
 614         struct proc *p = td->td_proc;
 615         struct prison *pr = p->p_ucred->cr_prison;
 616         int segnum, mode, error;
 617
 618         if (pr && !PRISON_CAP_ISSET(pr->pr_caps, PRISON_CAP_SYS_SYSVIPC))
 619                 return (ENOSYS);
 620
 621         mode = uap->shmflg & ACCESSPERMS;
 622
 623         lwkt_gettoken(&shm_token);
 624
 625         if (uap->key != IPC_PRIVATE) {
 626         again:
 627                 segnum = shm_find_segment_by_key(uap->key);
 628                 if (segnum >= 0) {
 629                         error = shmget_existing(p, sysmsg, uap, mode, segnum);
 630                         if (error == EAGAIN)
 631                                 goto again;
 632                         goto done;
 633                 }
 634                 if ((uap->shmflg & IPC_CREAT) == 0) {
 635                         error = ENOENT;
 636                         goto done;
 637                 }
 638         }
 639         error = shmget_allocate_segment(p, sysmsg, uap, mode);
 640 done:
 641         lwkt_reltoken(&shm_token);
 642
 643         return (error);
 644 }
 645
 646 void
 647 shmfork(struct proc *p1, struct proc *p2)
 648 {
 649         struct shmmap_state *shmmap_s;
 650         size_t size;
 651         int i;
 652
 653         lwkt_gettoken(&shm_token);
 654         size = shminfo.shmseg * sizeof(struct shmmap_state);
 655         shmmap_s = kmalloc(size, M_SHM, M_WAITOK);
 656         bcopy((caddr_t)p1->p_vmspace->vm_shm, (caddr_t)shmmap_s, size);
 657         p2->p_vmspace->vm_shm = (caddr_t)shmmap_s;
 658         for (i = 0; i < shminfo.shmseg; i++, shmmap_s++) {
 659                 if (shmmap_s->shmid != -1)
 660                         shmsegs[IPCID_TO_IX(shmmap_s->shmid)].shm_nattch++;
 661         }
 662         lwkt_reltoken(&shm_token);
 663 }
 664
 665 void
 666 shmexit(struct vmspace *vm)
 667 {
 668         struct shmmap_state *base, *shm;
 669         int i;
 670
 671         if ((base = (struct shmmap_state *)vm->vm_shm) != NULL) {
 672                 vm->vm_shm = NULL;
 673                 lwkt_gettoken(&shm_token);
 674                 for (i = 0, shm = base; i < shminfo.shmseg; i++, shm++) {
 675                         if (shm->shmid != -1)
 676                                 shm_delete_mapping(vm, shm);
 677                 }
 678                 kfree(base, M_SHM);
 679                 lwkt_reltoken(&shm_token);
 680         }
 681 }
 682
 683 static void
 684 shmrealloc(void)
 685 {
 686         int i;
 687         struct shmid_ds *newsegs;
 688
 689         if (shmalloced >= shminfo.shmmni)
 690                 return;
 691
 692         newsegs = kmalloc(shminfo.shmmni * sizeof(*newsegs), M_SHM, M_WAITOK);
 693         for (i = 0; i < shmalloced; i++)
 694                 bcopy(&shmsegs[i], &newsegs[i], sizeof(newsegs[0]));
 695         for (; i < shminfo.shmmni; i++) {
 696                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 697                 shmsegs[i].shm_perm.seq = 0;
 698         }
 699         kfree(shmsegs, M_SHM);
 700         shmsegs = newsegs;
 701         shmalloced = shminfo.shmmni;
 702 }
 703
 704 static void
 705 shminit(void *dummy)
 706 {
 707         int i;
 708
 709         /*
 710          * If not overridden by a tunable set the maximum shm to
 711          * 2/3 of main memory.
 712          */
 713         if (shminfo.shmall == 0)
 714                 shminfo.shmall = (size_t)vmstats.v_page_count * 2 / 3;
 715
 716         shminfo.shmmax = shminfo.shmall * PAGE_SIZE;
 717         shmalloced = shminfo.shmmni;
 718         shmsegs = kmalloc(shmalloced * sizeof(shmsegs[0]), M_SHM, M_WAITOK);
 719         for (i = 0; i < shmalloced; i++) {
 720                 shmsegs[i].shm_perm.mode = SHMSEG_FREE;
 721                 shmsegs[i].shm_perm.seq = 0;
 722         }
 723         shm_last_free = 0;
 724         shm_nused = 0;
 725         shm_committed = 0;
 726 }
 727 SYSINIT(sysv_shm, SI_SUB_SYSV_SHM, SI_ORDER_FIRST, shminit, NULL);