libringbuffer/shm.c

   1 /*
   2  * libringbuffer/shm.c
   3  *
   4  * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; only
   9  * version 2.1 of the License.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "shm.h"
  22 #include <unistd.h>
  23 #include <fcntl.h>
  24 #include <sys/mman.h>
  25 #include <sys/stat.h>   /* For mode constants */
  26 #include <fcntl.h>      /* For O_* constants */
  27 #include <assert.h>
  28 #include <stdio.h>
  29 #include <signal.h>
  30 #include <dirent.h>
  31 #include <lttng/align.h>
  32 #include <helper.h>
  33 #include <limits.h>
  34 #include <helper.h>
  35
  36 /*
  37  * Ensure we have the required amount of space available by writing 0
  38  * into the entire buffer. Not doing so can trigger SIGBUS when going
  39  * beyond the available shm space.
  40  */
  41 static
  42 int zero_file(int fd, size_t len)
  43 {
  44         ssize_t retlen;
  45         size_t written = 0;
  46         char *zeropage;
  47         long pagelen;
  48         int ret;
  49
  50         pagelen = sysconf(_SC_PAGESIZE);
  51         if (pagelen < 0)
  52                 return (int) pagelen;
  53         zeropage = calloc(pagelen, 1);
  54         if (!zeropage)
  55                 return -ENOMEM;
  56
  57         while (len > written) {
  58                 do {
  59                         retlen = write(fd, zeropage,
  60                                 min_t(size_t, pagelen, len - written));
  61                 } while (retlen == -1UL && errno == EINTR);
  62                 if (retlen < 0) {
  63                         ret = (int) retlen;
  64                         goto error;
  65                 }
  66                 written += retlen;
  67         }
  68         ret = 0;
  69 error:
  70         free(zeropage);
  71         return ret;
  72 }
  73
  74 struct shm_object_table *shm_object_table_create(size_t max_nb_obj)
  75 {
  76         struct shm_object_table *table;
  77
  78         table = zmalloc(sizeof(struct shm_object_table) +
  79                         max_nb_obj * sizeof(table->objects[0]));
  80         table->size = max_nb_obj;
  81         return table;
  82 }
  83
  84 static
  85 struct shm_object *_shm_object_table_alloc_shm(struct shm_object_table *table,
  86                                            size_t memory_map_size)
  87 {
  88         int shmfd, waitfd[2], ret, i, sigblocked = 0;
  89         struct shm_object *obj;
  90         char *memory_map;
  91         char tmp_name[NAME_MAX] = "/ust-shm-tmp-XXXXXX";
  92         sigset_t all_sigs, orig_sigs;
  93
  94         if (table->allocated_len >= table->size)
  95                 return NULL;
  96         obj = &table->objects[table->allocated_len];
  97
  98         /* wait_fd: create pipe */
  99         ret = pipe(waitfd);
 100         if (ret < 0) {
 101                 PERROR("pipe");
 102                 goto error_pipe;
 103         }
 104         for (i = 0; i < 2; i++) {
 105                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 106                 if (ret < 0) {
 107                         PERROR("fcntl");
 108                         goto error_fcntl;
 109                 }
 110         }
 111         /* The write end of the pipe needs to be non-blocking */
 112         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 113         if (ret < 0) {
 114                 PERROR("fcntl");
 115                 goto error_fcntl;
 116         }
 117         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 118
 119         /* shm_fd: create shm */
 120
 121         /*
 122          * Theoretically, we could leak a shm if the application crashes
 123          * between open and unlink. Disable signals on this thread for
 124          * increased safety against this scenario.
 125          */
 126         sigfillset(&all_sigs);
 127         ret = pthread_sigmask(SIG_BLOCK, &all_sigs, &orig_sigs);
 128         if (ret == -1) {
 129                 PERROR("pthread_sigmask");
 130                 goto error_pthread_sigmask;
 131         }
 132         sigblocked = 1;
 133
 134         /*
 135          * Allocate shm, and immediately unlink its shm oject, keeping
 136          * only the file descriptor as a reference to the object. If it
 137          * already exists (caused by short race window during which the
 138          * global object exists in a concurrent shm_open), simply retry.
 139          * We specifically do _not_ use the / at the beginning of the
 140          * pathname so that some OS implementations can keep it local to
 141          * the process (POSIX leaves this implementation-defined).
 142          */
 143         do {
 144                 /*
 145                  * Using mktemp filename with O_CREAT | O_EXCL open
 146                  * flags.
 147                  */
 148                 (void) mktemp(tmp_name);
 149                 if (tmp_name[0] == '\0') {
 150                         PERROR("mktemp");
 151                         goto error_shm_open;
 152                 }
 153                 shmfd = shm_open(tmp_name,
 154                                  O_CREAT | O_EXCL | O_RDWR, 0700);
 155         } while (shmfd < 0 && (errno == EEXIST || errno == EACCES));
 156         if (shmfd < 0) {
 157                 PERROR("shm_open");
 158                 goto error_shm_open;
 159         }
 160         ret = shm_unlink(tmp_name);
 161         if (ret < 0 && errno != ENOENT) {
 162                 PERROR("shm_unlink");
 163                 goto error_shm_release;
 164         }
 165         sigblocked = 0;
 166         ret = pthread_sigmask(SIG_SETMASK, &orig_sigs, NULL);
 167         if (ret == -1) {
 168                 PERROR("pthread_sigmask");
 169                 goto error_sigmask_release;
 170         }
 171         ret = zero_file(shmfd, memory_map_size);
 172         if (ret) {
 173                 PERROR("zero_file");
 174                 goto error_zero_file;
 175         }
 176         ret = ftruncate(shmfd, memory_map_size);
 177         if (ret) {
 178                 PERROR("ftruncate");
 179                 goto error_ftruncate;
 180         }
 181         obj->shm_fd = shmfd;
 182
 183         /* memory_map: mmap */
 184         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 185                           MAP_SHARED, shmfd, 0);
 186         if (memory_map == MAP_FAILED) {
 187                 PERROR("mmap");
 188                 goto error_mmap;
 189         }
 190         obj->type = SHM_OBJECT_SHM;
 191         obj->memory_map = memory_map;
 192         obj->memory_map_size = memory_map_size;
 193         obj->allocated_len = 0;
 194         obj->index = table->allocated_len++;
 195
 196         return obj;
 197
 198 error_mmap:
 199 error_ftruncate:
 200 error_shm_release:
 201 error_zero_file:
 202 error_sigmask_release:
 203         ret = close(shmfd);
 204         if (ret) {
 205                 PERROR("close");
 206                 assert(0);
 207         }
 208 error_shm_open:
 209         if (sigblocked) {
 210                 ret = pthread_sigmask(SIG_SETMASK, &orig_sigs, NULL);
 211                 if (ret == -1) {
 212                         PERROR("pthread_sigmask");
 213                 }
 214         }
 215 error_pthread_sigmask:
 216 error_fcntl:
 217         for (i = 0; i < 2; i++) {
 218                 ret = close(waitfd[i]);
 219                 if (ret) {
 220                         PERROR("close");
 221                         assert(0);
 222                 }
 223         }
 224 error_pipe:
 225         return NULL;
 226 }
 227
 228 static
 229 struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table,
 230                                            size_t memory_map_size)
 231 {
 232         struct shm_object *obj;
 233         void *memory_map;
 234         int waitfd[2], i, ret;
 235
 236         if (table->allocated_len >= table->size)
 237                 return NULL;
 238         obj = &table->objects[table->allocated_len];
 239
 240         memory_map = zmalloc(memory_map_size);
 241         if (!memory_map)
 242                 goto alloc_error;
 243
 244         /* wait_fd: create pipe */
 245         ret = pipe(waitfd);
 246         if (ret < 0) {
 247                 PERROR("pipe");
 248                 goto error_pipe;
 249         }
 250         for (i = 0; i < 2; i++) {
 251                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 252                 if (ret < 0) {
 253                         PERROR("fcntl");
 254                         goto error_fcntl;
 255                 }
 256         }
 257         /* The write end of the pipe needs to be non-blocking */
 258         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 259         if (ret < 0) {
 260                 PERROR("fcntl");
 261                 goto error_fcntl;
 262         }
 263         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 264
 265         /* no shm_fd */
 266         obj->shm_fd = -1;
 267
 268         obj->type = SHM_OBJECT_MEM;
 269         obj->memory_map = memory_map;
 270         obj->memory_map_size = memory_map_size;
 271         obj->allocated_len = 0;
 272         obj->index = table->allocated_len++;
 273
 274         return obj;
 275
 276 error_fcntl:
 277         for (i = 0; i < 2; i++) {
 278                 ret = close(waitfd[i]);
 279                 if (ret) {
 280                         PERROR("close");
 281                         assert(0);
 282                 }
 283         }
 284 error_pipe:
 285         free(memory_map);
 286 alloc_error:
 287         return NULL;
 288 }
 289
 290 struct shm_object *shm_object_table_alloc(struct shm_object_table *table,
 291                         size_t memory_map_size,
 292                         enum shm_object_type type)
 293 {
 294         switch (type) {
 295         case SHM_OBJECT_SHM:
 296                 return _shm_object_table_alloc_shm(table, memory_map_size);
 297         case SHM_OBJECT_MEM:
 298                 return _shm_object_table_alloc_mem(table, memory_map_size);
 299         default:
 300                 assert(0);
 301         }
 302         return NULL;
 303 }
 304
 305 struct shm_object *shm_object_table_append_shm(struct shm_object_table *table,
 306                         int shm_fd, int wakeup_fd, uint32_t stream_nr,
 307                         size_t memory_map_size)
 308 {
 309         struct shm_object *obj;
 310         char *memory_map;
 311         int ret;
 312
 313         if (table->allocated_len >= table->size)
 314                 return NULL;
 315         /* streams _must_ be received in sequential order, else fail. */
 316         if (stream_nr + 1 != table->allocated_len)
 317                 return NULL;
 318
 319         obj = &table->objects[table->allocated_len];
 320
 321         /* wait_fd: set write end of the pipe. */
 322         obj->wait_fd[0] = -1;   /* read end is unset */
 323         obj->wait_fd[1] = wakeup_fd;
 324         obj->shm_fd = shm_fd;
 325
 326         ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC);
 327         if (ret < 0) {
 328                 PERROR("fcntl");
 329                 goto error_fcntl;
 330         }
 331         /* The write end of the pipe needs to be non-blocking */
 332         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 333         if (ret < 0) {
 334                 PERROR("fcntl");
 335                 goto error_fcntl;
 336         }
 337
 338         /* memory_map: mmap */
 339         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 340                           MAP_SHARED, shm_fd, 0);
 341         if (memory_map == MAP_FAILED) {
 342                 PERROR("mmap");
 343                 goto error_mmap;
 344         }
 345         obj->type = SHM_OBJECT_SHM;
 346         obj->memory_map = memory_map;
 347         obj->memory_map_size = memory_map_size;
 348         obj->allocated_len = memory_map_size;
 349         obj->index = table->allocated_len++;
 350
 351         return obj;
 352
 353 error_fcntl:
 354 error_mmap:
 355         return NULL;
 356 }
 357
 358 /*
 359  * Passing ownership of mem to object.
 360  */
 361 struct shm_object *shm_object_table_append_mem(struct shm_object_table *table,
 362                         void *mem, size_t memory_map_size, int wakeup_fd)
 363 {
 364         struct shm_object *obj;
 365         int ret;
 366
 367         if (table->allocated_len >= table->size)
 368                 return NULL;
 369         obj = &table->objects[table->allocated_len];
 370
 371         obj->wait_fd[0] = -1;   /* read end is unset */
 372         obj->wait_fd[1] = wakeup_fd;
 373         obj->shm_fd = -1;
 374
 375         ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC);
 376         if (ret < 0) {
 377                 PERROR("fcntl");
 378                 goto error_fcntl;
 379         }
 380         /* The write end of the pipe needs to be non-blocking */
 381         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 382         if (ret < 0) {
 383                 PERROR("fcntl");
 384                 goto error_fcntl;
 385         }
 386
 387         obj->type = SHM_OBJECT_MEM;
 388         obj->memory_map = mem;
 389         obj->memory_map_size = memory_map_size;
 390         obj->allocated_len = memory_map_size;
 391         obj->index = table->allocated_len++;
 392
 393         return obj;
 394
 395 error_fcntl:
 396         return NULL;
 397 }
 398
 399 static
 400 void shmp_object_destroy(struct shm_object *obj)
 401 {
 402         switch (obj->type) {
 403         case SHM_OBJECT_SHM:
 404         {
 405                 int ret, i;
 406
 407                 ret = munmap(obj->memory_map, obj->memory_map_size);
 408                 if (ret) {
 409                         PERROR("umnmap");
 410                         assert(0);
 411                 }
 412                 ret = close(obj->shm_fd);
 413                 if (ret) {
 414                         PERROR("close");
 415                         assert(0);
 416                 }
 417                 for (i = 0; i < 2; i++) {
 418                         if (obj->wait_fd[i] < 0)
 419                                 continue;
 420                         ret = close(obj->wait_fd[i]);
 421                         if (ret) {
 422                                 PERROR("close");
 423                                 assert(0);
 424                         }
 425                 }
 426                 break;
 427         }
 428         case SHM_OBJECT_MEM:
 429         {
 430                 int ret, i;
 431
 432                 for (i = 0; i < 2; i++) {
 433                         if (obj->wait_fd[i] < 0)
 434                                 continue;
 435                         ret = close(obj->wait_fd[i]);
 436                         if (ret) {
 437                                 PERROR("close");
 438                                 assert(0);
 439                         }
 440                 }
 441                 free(obj->memory_map);
 442                 break;
 443         }
 444         default:
 445                 assert(0);
 446         }
 447 }
 448
 449 void shm_object_table_destroy(struct shm_object_table *table)
 450 {
 451         int i;
 452
 453         for (i = 0; i < table->allocated_len; i++)
 454                 shmp_object_destroy(&table->objects[i]);
 455         free(table);
 456 }
 457
 458 /*
 459  * zalloc_shm - allocate memory within a shm object.
 460  *
 461  * Shared memory is already zeroed by shmget.
 462  * *NOT* multithread-safe (should be protected by mutex).
 463  * Returns a -1, -1 tuple on error.
 464  */
 465 struct shm_ref zalloc_shm(struct shm_object *obj, size_t len)
 466 {
 467         struct shm_ref ref;
 468         struct shm_ref shm_ref_error = { -1, -1 };
 469
 470         if (obj->memory_map_size - obj->allocated_len < len)
 471                 return shm_ref_error;
 472         ref.index = obj->index;
 473         ref.offset =  obj->allocated_len;
 474         obj->allocated_len += len;
 475         return ref;
 476 }
 477
 478 void align_shm(struct shm_object *obj, size_t align)
 479 {
 480         size_t offset_len = offset_align(obj->allocated_len, align);
 481         obj->allocated_len += offset_len;
 482 }