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 #define _LGPL_SOURCE
  22 #include "shm.h"
  23 #include <unistd.h>
  24 #include <fcntl.h>
  25 #include <sys/mman.h>
  26 #include <sys/types.h>
  27 #include <sys/stat.h>   /* For mode constants */
  28 #include <fcntl.h>      /* For O_* constants */
  29 #include <assert.h>
  30 #include <stdio.h>
  31 #include <signal.h>
  32 #include <dirent.h>
  33 #include <lttng/align.h>
  34 #include <limits.h>
  35 #include <helper.h>
  36 #include <ust-fd.h>
  37
  38 /*
  39  * Ensure we have the required amount of space available by writing 0
  40  * into the entire buffer. Not doing so can trigger SIGBUS when going
  41  * beyond the available shm space.
  42  */
  43 static
  44 int zero_file(int fd, size_t len)
  45 {
  46         ssize_t retlen;
  47         size_t written = 0;
  48         char *zeropage;
  49         long pagelen;
  50         int ret;
  51
  52         pagelen = sysconf(_SC_PAGESIZE);
  53         if (pagelen < 0)
  54                 return (int) pagelen;
  55         zeropage = calloc(pagelen, 1);
  56         if (!zeropage)
  57                 return -ENOMEM;
  58
  59         while (len > written) {
  60                 do {
  61                         retlen = write(fd, zeropage,
  62                                 min_t(size_t, pagelen, len - written));
  63                 } while (retlen == -1UL && errno == EINTR);
  64                 if (retlen < 0) {
  65                         ret = (int) retlen;
  66                         goto error;
  67                 }
  68                 written += retlen;
  69         }
  70         ret = 0;
  71 error:
  72         free(zeropage);
  73         return ret;
  74 }
  75
  76 struct shm_object_table *shm_object_table_create(size_t max_nb_obj)
  77 {
  78         struct shm_object_table *table;
  79
  80         table = zmalloc(sizeof(struct shm_object_table) +
  81                         max_nb_obj * sizeof(table->objects[0]));
  82         if (!table)
  83                 return NULL;
  84         table->size = max_nb_obj;
  85         return table;
  86 }
  87
  88 static
  89 struct shm_object *_shm_object_table_alloc_shm(struct shm_object_table *table,
  90                                            size_t memory_map_size,
  91                                            int stream_fd)
  92 {
  93         int shmfd, waitfd[2], ret, i;
  94         struct shm_object *obj;
  95         char *memory_map;
  96
  97         if (stream_fd < 0)
  98                 return NULL;
  99         if (table->allocated_len >= table->size)
 100                 return NULL;
 101         obj = &table->objects[table->allocated_len];
 102
 103         /* wait_fd: create pipe */
 104         ret = pipe(waitfd);
 105         if (ret < 0) {
 106                 PERROR("pipe");
 107                 goto error_pipe;
 108         }
 109         for (i = 0; i < 2; i++) {
 110                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 111                 if (ret < 0) {
 112                         PERROR("fcntl");
 113                         goto error_fcntl;
 114                 }
 115         }
 116         /* The write end of the pipe needs to be non-blocking */
 117         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 118         if (ret < 0) {
 119                 PERROR("fcntl");
 120                 goto error_fcntl;
 121         }
 122         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 123
 124         /* create shm */
 125
 126         shmfd = stream_fd;
 127         ret = zero_file(shmfd, memory_map_size);
 128         if (ret) {
 129                 PERROR("zero_file");
 130                 goto error_zero_file;
 131         }
 132         ret = ftruncate(shmfd, memory_map_size);
 133         if (ret) {
 134                 PERROR("ftruncate");
 135                 goto error_ftruncate;
 136         }
 137         /*
 138          * Also ensure the file metadata is synced with the storage by using
 139          * fsync(2).
 140          */
 141         ret = fsync(shmfd);
 142         if (ret) {
 143                 PERROR("fsync");
 144                 goto error_fsync;
 145         }
 146         obj->shm_fd_ownership = 0;
 147         obj->shm_fd = shmfd;
 148
 149         /* memory_map: mmap */
 150         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 151                           MAP_SHARED, shmfd, 0);
 152         if (memory_map == MAP_FAILED) {
 153                 PERROR("mmap");
 154                 goto error_mmap;
 155         }
 156         obj->type = SHM_OBJECT_SHM;
 157         obj->memory_map = memory_map;
 158         obj->memory_map_size = memory_map_size;
 159         obj->allocated_len = 0;
 160         obj->index = table->allocated_len++;
 161
 162         return obj;
 163
 164 error_mmap:
 165 error_fsync:
 166 error_ftruncate:
 167 error_zero_file:
 168 error_fcntl:
 169         for (i = 0; i < 2; i++) {
 170                 ret = close(waitfd[i]);
 171                 if (ret) {
 172                         PERROR("close");
 173                         assert(0);
 174                 }
 175         }
 176 error_pipe:
 177         return NULL;
 178 }
 179
 180 static
 181 struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table,
 182                                            size_t memory_map_size)
 183 {
 184         struct shm_object *obj;
 185         void *memory_map;
 186         int waitfd[2], i, ret;
 187
 188         if (table->allocated_len >= table->size)
 189                 return NULL;
 190         obj = &table->objects[table->allocated_len];
 191
 192         memory_map = zmalloc(memory_map_size);
 193         if (!memory_map)
 194                 goto alloc_error;
 195
 196         /* wait_fd: create pipe */
 197         ret = pipe(waitfd);
 198         if (ret < 0) {
 199                 PERROR("pipe");
 200                 goto error_pipe;
 201         }
 202         for (i = 0; i < 2; i++) {
 203                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 204                 if (ret < 0) {
 205                         PERROR("fcntl");
 206                         goto error_fcntl;
 207                 }
 208         }
 209         /* The write end of the pipe needs to be non-blocking */
 210         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 211         if (ret < 0) {
 212                 PERROR("fcntl");
 213                 goto error_fcntl;
 214         }
 215         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 216
 217         /* no shm_fd */
 218         obj->shm_fd = -1;
 219         obj->shm_fd_ownership = 0;
 220
 221         obj->type = SHM_OBJECT_MEM;
 222         obj->memory_map = memory_map;
 223         obj->memory_map_size = memory_map_size;
 224         obj->allocated_len = 0;
 225         obj->index = table->allocated_len++;
 226
 227         return obj;
 228
 229 error_fcntl:
 230         for (i = 0; i < 2; i++) {
 231                 ret = close(waitfd[i]);
 232                 if (ret) {
 233                         PERROR("close");
 234                         assert(0);
 235                 }
 236         }
 237 error_pipe:
 238         free(memory_map);
 239 alloc_error:
 240         return NULL;
 241 }
 242
 243 struct shm_object *shm_object_table_alloc(struct shm_object_table *table,
 244                         size_t memory_map_size,
 245                         enum shm_object_type type,
 246                         int stream_fd)
 247 {
 248         switch (type) {
 249         case SHM_OBJECT_SHM:
 250                 return _shm_object_table_alloc_shm(table, memory_map_size,
 251                                 stream_fd);
 252         case SHM_OBJECT_MEM:
 253                 return _shm_object_table_alloc_mem(table, memory_map_size);
 254         default:
 255                 assert(0);
 256         }
 257         return NULL;
 258 }
 259
 260 struct shm_object *shm_object_table_append_shm(struct shm_object_table *table,
 261                         int shm_fd, int wakeup_fd, uint32_t stream_nr,
 262                         size_t memory_map_size)
 263 {
 264         struct shm_object *obj;
 265         char *memory_map;
 266         int ret;
 267
 268         if (table->allocated_len >= table->size)
 269                 return NULL;
 270         /* streams _must_ be received in sequential order, else fail. */
 271         if (stream_nr + 1 != table->allocated_len)
 272                 return NULL;
 273
 274         obj = &table->objects[table->allocated_len];
 275
 276         /* wait_fd: set write end of the pipe. */
 277         obj->wait_fd[0] = -1;   /* read end is unset */
 278         obj->wait_fd[1] = wakeup_fd;
 279         obj->shm_fd = shm_fd;
 280         obj->shm_fd_ownership = 1;
 281
 282         /* The write end of the pipe needs to be non-blocking */
 283         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 284         if (ret < 0) {
 285                 PERROR("fcntl");
 286                 goto error_fcntl;
 287         }
 288
 289         /* memory_map: mmap */
 290         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 291                           MAP_SHARED, shm_fd, 0);
 292         if (memory_map == MAP_FAILED) {
 293                 PERROR("mmap");
 294                 goto error_mmap;
 295         }
 296         obj->type = SHM_OBJECT_SHM;
 297         obj->memory_map = memory_map;
 298         obj->memory_map_size = memory_map_size;
 299         obj->allocated_len = memory_map_size;
 300         obj->index = table->allocated_len++;
 301
 302         return obj;
 303
 304 error_fcntl:
 305 error_mmap:
 306         return NULL;
 307 }
 308
 309 /*
 310  * Passing ownership of mem to object.
 311  */
 312 struct shm_object *shm_object_table_append_mem(struct shm_object_table *table,
 313                         void *mem, size_t memory_map_size, int wakeup_fd)
 314 {
 315         struct shm_object *obj;
 316         int ret;
 317
 318         if (table->allocated_len >= table->size)
 319                 return NULL;
 320         obj = &table->objects[table->allocated_len];
 321
 322         obj->wait_fd[0] = -1;   /* read end is unset */
 323         obj->wait_fd[1] = wakeup_fd;
 324         obj->shm_fd = -1;
 325         obj->shm_fd_ownership = 0;
 326
 327         ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC);
 328         if (ret < 0) {
 329                 PERROR("fcntl");
 330                 goto error_fcntl;
 331         }
 332         /* The write end of the pipe needs to be non-blocking */
 333         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 334         if (ret < 0) {
 335                 PERROR("fcntl");
 336                 goto error_fcntl;
 337         }
 338
 339         obj->type = SHM_OBJECT_MEM;
 340         obj->memory_map = mem;
 341         obj->memory_map_size = memory_map_size;
 342         obj->allocated_len = memory_map_size;
 343         obj->index = table->allocated_len++;
 344
 345         return obj;
 346
 347 error_fcntl:
 348         return NULL;
 349 }
 350
 351 static
 352 void shmp_object_destroy(struct shm_object *obj, int consumer)
 353 {
 354         switch (obj->type) {
 355         case SHM_OBJECT_SHM:
 356         {
 357                 int ret, i;
 358
 359                 ret = munmap(obj->memory_map, obj->memory_map_size);
 360                 if (ret) {
 361                         PERROR("umnmap");
 362                         assert(0);
 363                 }
 364
 365                 if (obj->shm_fd_ownership) {
 366                         /* Delete FDs only if called from app (not consumer). */
 367                         if (!consumer) {
 368                                 lttng_ust_lock_fd_tracker();
 369                                 ret = close(obj->shm_fd);
 370                                 if (!ret) {
 371                                         lttng_ust_delete_fd_from_tracker(obj->shm_fd);
 372                                 } else {
 373                                         PERROR("close");
 374                                         assert(0);
 375                                 }
 376                                 lttng_ust_unlock_fd_tracker();
 377                         } else {
 378                                 ret = close(obj->shm_fd);
 379                                 if (ret) {
 380                                         PERROR("close");
 381                                         assert(0);
 382                                 }
 383                         }
 384                 }
 385                 for (i = 0; i < 2; i++) {
 386                         if (obj->wait_fd[i] < 0)
 387                                 continue;
 388                         if (!consumer) {
 389                                 lttng_ust_lock_fd_tracker();
 390                                 ret = close(obj->wait_fd[i]);
 391                                 if (!ret) {
 392                                         lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
 393                                 } else {
 394                                         PERROR("close");
 395                                         assert(0);
 396                                 }
 397                                 lttng_ust_unlock_fd_tracker();
 398                         } else {
 399                                 ret = close(obj->wait_fd[i]);
 400                                 if (ret) {
 401                                         PERROR("close");
 402                                         assert(0);
 403                                 }
 404                         }
 405                 }
 406                 break;
 407         }
 408         case SHM_OBJECT_MEM:
 409         {
 410                 int ret, i;
 411
 412                 for (i = 0; i < 2; i++) {
 413                         if (obj->wait_fd[i] < 0)
 414                                 continue;
 415                         if (!consumer) {
 416                                 lttng_ust_lock_fd_tracker();
 417                                 ret = close(obj->wait_fd[i]);
 418                                 if (!ret) {
 419                                         lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
 420                                 } else {
 421                                         PERROR("close");
 422                                         assert(0);
 423                                 }
 424                                 lttng_ust_unlock_fd_tracker();
 425                         } else {
 426                                 ret = close(obj->wait_fd[i]);
 427                                 if (ret) {
 428                                         PERROR("close");
 429                                         assert(0);
 430                                 }
 431                         }
 432                 }
 433                 free(obj->memory_map);
 434                 break;
 435         }
 436         default:
 437                 assert(0);
 438         }
 439 }
 440
 441 void shm_object_table_destroy(struct shm_object_table *table, int consumer)
 442 {
 443         int i;
 444
 445         for (i = 0; i < table->allocated_len; i++)
 446                 shmp_object_destroy(&table->objects[i], consumer);
 447         free(table);
 448 }
 449
 450 /*
 451  * zalloc_shm - allocate memory within a shm object.
 452  *
 453  * Shared memory is already zeroed by shmget.
 454  * *NOT* multithread-safe (should be protected by mutex).
 455  * Returns a -1, -1 tuple on error.
 456  */
 457 struct shm_ref zalloc_shm(struct shm_object *obj, size_t len)
 458 {
 459         struct shm_ref ref;
 460         struct shm_ref shm_ref_error = { -1, -1 };
 461
 462         if (obj->memory_map_size - obj->allocated_len < len)
 463                 return shm_ref_error;
 464         ref.index = obj->index;
 465         ref.offset =  obj->allocated_len;
 466         obj->allocated_len += len;
 467         return ref;
 468 }
 469
 470 void align_shm(struct shm_object *obj, size_t align)
 471 {
 472         size_t offset_len = offset_align(obj->allocated_len, align);
 473         obj->allocated_len += offset_len;
 474 }