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                 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
 235         if (table->allocated_len >= table->size)
 236                 return NULL;
 237         obj = &table->objects[table->allocated_len];
 238
 239         memory_map = zmalloc(memory_map_size);
 240         if (!memory_map)
 241                 goto alloc_error;
 242
 243         obj->wait_fd[0] = -1;
 244         obj->wait_fd[1] = -1;
 245         obj->shm_fd = -1;
 246
 247         obj->type = SHM_OBJECT_MEM;
 248         obj->memory_map = memory_map;
 249         obj->memory_map_size = memory_map_size;
 250         obj->allocated_len = 0;
 251         obj->index = table->allocated_len++;
 252
 253         return obj;
 254
 255 alloc_error:
 256         return NULL;
 257 }
 258
 259 struct shm_object *shm_object_table_alloc(struct shm_object_table *table,
 260                         size_t memory_map_size,
 261                         enum shm_object_type type)
 262 {
 263         switch (type) {
 264         case SHM_OBJECT_SHM:
 265                 return _shm_object_table_alloc_shm(table, memory_map_size);
 266         case SHM_OBJECT_MEM:
 267                 return _shm_object_table_alloc_mem(table, memory_map_size);
 268         default:
 269                 assert(0);
 270         }
 271         return NULL;
 272 }
 273
 274 struct shm_object *shm_object_table_append_shm(struct shm_object_table *table,
 275                         int shm_fd, int wakeup_fd, uint32_t stream_nr,
 276                         size_t memory_map_size)
 277 {
 278         struct shm_object *obj;
 279         char *memory_map;
 280         int ret;
 281
 282         if (table->allocated_len >= table->size)
 283                 return NULL;
 284         /* streams _must_ be received in sequential order, else fail. */
 285         if (stream_nr + 1 != table->allocated_len)
 286                 return NULL;
 287
 288         obj = &table->objects[table->allocated_len];
 289
 290         /* wait_fd: set write end of the pipe. */
 291         obj->wait_fd[0] = -1;   /* read end is unset */
 292         obj->wait_fd[1] = wakeup_fd;
 293         obj->shm_fd = shm_fd;
 294
 295         ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC);
 296         if (ret < 0) {
 297                 PERROR("fcntl");
 298                 goto error_fcntl;
 299         }
 300         /* The write end of the pipe needs to be non-blocking */
 301         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 302         if (ret < 0) {
 303                 PERROR("fcntl");
 304                 goto error_fcntl;
 305         }
 306
 307         /* memory_map: mmap */
 308         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 309                           MAP_SHARED, shm_fd, 0);
 310         if (memory_map == MAP_FAILED) {
 311                 PERROR("mmap");
 312                 goto error_mmap;
 313         }
 314         obj->type = SHM_OBJECT_SHM;
 315         obj->memory_map = memory_map;
 316         obj->memory_map_size = memory_map_size;
 317         obj->allocated_len = memory_map_size;
 318         obj->index = table->allocated_len++;
 319
 320         return obj;
 321
 322 error_fcntl:
 323 error_mmap:
 324         return NULL;
 325 }
 326
 327 /*
 328  * Passing ownership of mem to object.
 329  */
 330 struct shm_object *shm_object_table_append_mem(struct shm_object_table *table,
 331                         void *mem, size_t memory_map_size)
 332 {
 333         struct shm_object *obj;
 334
 335         if (table->allocated_len >= table->size)
 336                 return NULL;
 337         obj = &table->objects[table->allocated_len];
 338
 339         obj->wait_fd[0] = -1;
 340         obj->wait_fd[1] = -1;
 341         obj->shm_fd = -1;
 342
 343         obj->type = SHM_OBJECT_MEM;
 344         obj->memory_map = mem;
 345         obj->memory_map_size = memory_map_size;
 346         obj->allocated_len = memory_map_size;
 347         obj->index = table->allocated_len++;
 348
 349         return obj;
 350 }
 351
 352 static
 353 void shmp_object_destroy(struct shm_object *obj)
 354 {
 355         switch (obj->type) {
 356         case SHM_OBJECT_SHM:
 357         {
 358                 int ret, i;
 359
 360                 ret = munmap(obj->memory_map, obj->memory_map_size);
 361                 if (ret) {
 362                         PERROR("umnmap");
 363                         assert(0);
 364                 }
 365                 ret = close(obj->shm_fd);
 366                 if (ret) {
 367                         PERROR("close");
 368                         assert(0);
 369                 }
 370                 for (i = 0; i < 2; i++) {
 371                         if (obj->wait_fd[i] < 0)
 372                                 continue;
 373                         ret = close(obj->wait_fd[i]);
 374                         if (ret) {
 375                                 PERROR("close");
 376                                 assert(0);
 377                         }
 378                 }
 379                 break;
 380         }
 381         case SHM_OBJECT_MEM:
 382                 free(obj->memory_map);
 383                 break;
 384         default:
 385                 assert(0);
 386         }
 387 }
 388
 389 void shm_object_table_destroy(struct shm_object_table *table)
 390 {
 391         int i;
 392
 393         for (i = 0; i < table->allocated_len; i++)
 394                 shmp_object_destroy(&table->objects[i]);
 395         free(table);
 396 }
 397
 398 /*
 399  * zalloc_shm - allocate memory within a shm object.
 400  *
 401  * Shared memory is already zeroed by shmget.
 402  * *NOT* multithread-safe (should be protected by mutex).
 403  * Returns a -1, -1 tuple on error.
 404  */
 405 struct shm_ref zalloc_shm(struct shm_object *obj, size_t len)
 406 {
 407         struct shm_ref ref;
 408         struct shm_ref shm_ref_error = { -1, -1 };
 409
 410         if (obj->memory_map_size - obj->allocated_len < len)
 411                 return shm_ref_error;
 412         ref.index = obj->index;
 413         ref.offset =  obj->allocated_len;
 414         obj->allocated_len += len;
 415         return ref;
 416 }
 417
 418 void align_shm(struct shm_object *obj, size_t align)
 419 {
 420         size_t offset_len = offset_align(obj->allocated_len, align);
 421         obj->allocated_len += offset_len;
 422 }