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 <stdbool.h>
  36 #include <stdint.h>
  37 #ifdef HAVE_LIBNUMA
  38 #include <numa.h>
  39 #include <numaif.h>
  40 #endif
  41 #include <helper.h>
  42 #include <ust-fd.h>
  43 #include "mmap.h"
  44
  45 /*
  46  * Ensure we have the required amount of space available by writing 0
  47  * into the entire buffer. Not doing so can trigger SIGBUS when going
  48  * beyond the available shm space.
  49  */
  50 static
  51 int zero_file(int fd, size_t len)
  52 {
  53         ssize_t retlen;
  54         size_t written = 0;
  55         char *zeropage;
  56         long pagelen;
  57         int ret;
  58
  59         pagelen = sysconf(_SC_PAGESIZE);
  60         if (pagelen < 0)
  61                 return (int) pagelen;
  62         zeropage = calloc(pagelen, 1);
  63         if (!zeropage)
  64                 return -ENOMEM;
  65
  66         while (len > written) {
  67                 do {
  68                         retlen = write(fd, zeropage,
  69                                 min_t(size_t, pagelen, len - written));
  70                 } while (retlen == -1UL && errno == EINTR);
  71                 if (retlen < 0) {
  72                         ret = (int) retlen;
  73                         goto error;
  74                 }
  75                 written += retlen;
  76         }
  77         ret = 0;
  78 error:
  79         free(zeropage);
  80         return ret;
  81 }
  82
  83 struct shm_object_table *shm_object_table_create(size_t max_nb_obj)
  84 {
  85         struct shm_object_table *table;
  86
  87         table = zmalloc(sizeof(struct shm_object_table) +
  88                         max_nb_obj * sizeof(table->objects[0]));
  89         if (!table)
  90                 return NULL;
  91         table->size = max_nb_obj;
  92         return table;
  93 }
  94
  95 static
  96 struct shm_object *_shm_object_table_alloc_shm(struct shm_object_table *table,
  97                                            size_t memory_map_size,
  98                                            int stream_fd)
  99 {
 100         int shmfd, waitfd[2], ret, i;
 101         struct shm_object *obj;
 102         char *memory_map;
 103
 104         if (stream_fd < 0)
 105                 return NULL;
 106         if (table->allocated_len >= table->size)
 107                 return NULL;
 108         obj = &table->objects[table->allocated_len];
 109
 110         /* wait_fd: create pipe */
 111         ret = pipe(waitfd);
 112         if (ret < 0) {
 113                 PERROR("pipe");
 114                 goto error_pipe;
 115         }
 116         for (i = 0; i < 2; i++) {
 117                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 118                 if (ret < 0) {
 119                         PERROR("fcntl");
 120                         goto error_fcntl;
 121                 }
 122         }
 123         /* The write end of the pipe needs to be non-blocking */
 124         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 125         if (ret < 0) {
 126                 PERROR("fcntl");
 127                 goto error_fcntl;
 128         }
 129         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 130
 131         /*
 132          * Set POSIX shared memory object size
 133          *
 134          * First, use ftruncate() to set its size, some implementations won't
 135          * allow writes past the size set by ftruncate.
 136          * Then, use write() to fill it with zeros, this allows us to fully
 137          * allocate it and detect a shortage of shm space without dealing with
 138          * a SIGBUS.
 139          */
 140
 141         shmfd = stream_fd;
 142         ret = ftruncate(shmfd, memory_map_size);
 143         if (ret) {
 144                 PERROR("ftruncate");
 145                 goto error_ftruncate;
 146         }
 147         ret = zero_file(shmfd, memory_map_size);
 148         if (ret) {
 149                 PERROR("zero_file");
 150                 goto error_zero_file;
 151         }
 152         /*
 153          * Also ensure the file metadata is synced with the storage by using
 154          * fsync(2).
 155          */
 156         ret = fsync(shmfd);
 157         if (ret) {
 158                 PERROR("fsync");
 159                 goto error_fsync;
 160         }
 161         obj->shm_fd_ownership = 0;
 162         obj->shm_fd = shmfd;
 163
 164         /* memory_map: mmap */
 165         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 166                           MAP_SHARED | LTTNG_MAP_POPULATE, shmfd, 0);
 167         if (memory_map == MAP_FAILED) {
 168                 PERROR("mmap");
 169                 goto error_mmap;
 170         }
 171         obj->type = SHM_OBJECT_SHM;
 172         obj->memory_map = memory_map;
 173         obj->memory_map_size = memory_map_size;
 174         obj->allocated_len = 0;
 175         obj->index = table->allocated_len++;
 176
 177         return obj;
 178
 179 error_mmap:
 180 error_fsync:
 181 error_ftruncate:
 182 error_zero_file:
 183 error_fcntl:
 184         for (i = 0; i < 2; i++) {
 185                 ret = close(waitfd[i]);
 186                 if (ret) {
 187                         PERROR("close");
 188                         assert(0);
 189                 }
 190         }
 191 error_pipe:
 192         return NULL;
 193 }
 194
 195 static
 196 struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table,
 197                                            size_t memory_map_size)
 198 {
 199         struct shm_object *obj;
 200         void *memory_map;
 201         int waitfd[2], i, ret;
 202
 203         if (table->allocated_len >= table->size)
 204                 return NULL;
 205         obj = &table->objects[table->allocated_len];
 206
 207         memory_map = zmalloc(memory_map_size);
 208         if (!memory_map)
 209                 goto alloc_error;
 210
 211         /* wait_fd: create pipe */
 212         ret = pipe(waitfd);
 213         if (ret < 0) {
 214                 PERROR("pipe");
 215                 goto error_pipe;
 216         }
 217         for (i = 0; i < 2; i++) {
 218                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 219                 if (ret < 0) {
 220                         PERROR("fcntl");
 221                         goto error_fcntl;
 222                 }
 223         }
 224         /* The write end of the pipe needs to be non-blocking */
 225         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 226         if (ret < 0) {
 227                 PERROR("fcntl");
 228                 goto error_fcntl;
 229         }
 230         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 231
 232         /* no shm_fd */
 233         obj->shm_fd = -1;
 234         obj->shm_fd_ownership = 0;
 235
 236         obj->type = SHM_OBJECT_MEM;
 237         obj->memory_map = memory_map;
 238         obj->memory_map_size = memory_map_size;
 239         obj->allocated_len = 0;
 240         obj->index = table->allocated_len++;
 241
 242         return obj;
 243
 244 error_fcntl:
 245         for (i = 0; i < 2; i++) {
 246                 ret = close(waitfd[i]);
 247                 if (ret) {
 248                         PERROR("close");
 249                         assert(0);
 250                 }
 251         }
 252 error_pipe:
 253         free(memory_map);
 254 alloc_error:
 255         return NULL;
 256 }
 257
 258 /*
 259  * libnuma prints errors on the console even for numa_available().
 260  * Work-around this limitation by using get_mempolicy() directly to
 261  * check whether the kernel supports mempolicy.
 262  */
 263 #ifdef HAVE_LIBNUMA
 264 static bool lttng_is_numa_available(void)
 265 {
 266         int ret;
 267
 268         ret = get_mempolicy(NULL, NULL, 0, NULL, 0);
 269         if (ret && errno == ENOSYS) {
 270                 return false;
 271         }
 272         return numa_available() > 0;
 273 }
 274 #endif
 275
 276 struct shm_object *shm_object_table_alloc(struct shm_object_table *table,
 277                         size_t memory_map_size,
 278                         enum shm_object_type type,
 279                         int stream_fd,
 280                         int cpu)
 281 {
 282         struct shm_object *shm_object;
 283 #ifdef HAVE_LIBNUMA
 284         int oldnode = 0, node;
 285         bool numa_avail;
 286
 287         numa_avail = lttng_is_numa_available();
 288         if (numa_avail) {
 289                 oldnode = numa_preferred();
 290                 if (cpu >= 0) {
 291                         node = numa_node_of_cpu(cpu);
 292                         if (node >= 0)
 293                                 numa_set_preferred(node);
 294                 }
 295                 if (cpu < 0 || node < 0)
 296                         numa_set_localalloc();
 297         }
 298 #endif /* HAVE_LIBNUMA */
 299         switch (type) {
 300         case SHM_OBJECT_SHM:
 301                 shm_object = _shm_object_table_alloc_shm(table, memory_map_size,
 302                                 stream_fd);
 303                 break;
 304         case SHM_OBJECT_MEM:
 305                 shm_object = _shm_object_table_alloc_mem(table, memory_map_size);
 306                 break;
 307         default:
 308                 assert(0);
 309         }
 310 #ifdef HAVE_LIBNUMA
 311         if (numa_avail)
 312                 numa_set_preferred(oldnode);
 313 #endif /* HAVE_LIBNUMA */
 314         return shm_object;
 315 }
 316
 317 struct shm_object *shm_object_table_append_shm(struct shm_object_table *table,
 318                         int shm_fd, int wakeup_fd, uint32_t stream_nr,
 319                         size_t memory_map_size)
 320 {
 321         struct shm_object *obj;
 322         char *memory_map;
 323         int ret;
 324
 325         if (table->allocated_len >= table->size)
 326                 return NULL;
 327         /* streams _must_ be received in sequential order, else fail. */
 328         if (stream_nr + 1 != table->allocated_len)
 329                 return NULL;
 330
 331         obj = &table->objects[table->allocated_len];
 332
 333         /* wait_fd: set write end of the pipe. */
 334         obj->wait_fd[0] = -1;   /* read end is unset */
 335         obj->wait_fd[1] = wakeup_fd;
 336         obj->shm_fd = shm_fd;
 337         obj->shm_fd_ownership = 1;
 338
 339         /* The write end of the pipe needs to be non-blocking */
 340         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 341         if (ret < 0) {
 342                 PERROR("fcntl");
 343                 goto error_fcntl;
 344         }
 345
 346         /* memory_map: mmap */
 347         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 348                           MAP_SHARED | LTTNG_MAP_POPULATE, shm_fd, 0);
 349         if (memory_map == MAP_FAILED) {
 350                 PERROR("mmap");
 351                 goto error_mmap;
 352         }
 353         obj->type = SHM_OBJECT_SHM;
 354         obj->memory_map = memory_map;
 355         obj->memory_map_size = memory_map_size;
 356         obj->allocated_len = memory_map_size;
 357         obj->index = table->allocated_len++;
 358
 359         return obj;
 360
 361 error_fcntl:
 362 error_mmap:
 363         return NULL;
 364 }
 365
 366 /*
 367  * Passing ownership of mem to object.
 368  */
 369 struct shm_object *shm_object_table_append_mem(struct shm_object_table *table,
 370                         void *mem, size_t memory_map_size, int wakeup_fd)
 371 {
 372         struct shm_object *obj;
 373         int ret;
 374
 375         if (table->allocated_len >= table->size)
 376                 return NULL;
 377         obj = &table->objects[table->allocated_len];
 378
 379         obj->wait_fd[0] = -1;   /* read end is unset */
 380         obj->wait_fd[1] = wakeup_fd;
 381         obj->shm_fd = -1;
 382         obj->shm_fd_ownership = 0;
 383
 384         ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC);
 385         if (ret < 0) {
 386                 PERROR("fcntl");
 387                 goto error_fcntl;
 388         }
 389         /* The write end of the pipe needs to be non-blocking */
 390         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 391         if (ret < 0) {
 392                 PERROR("fcntl");
 393                 goto error_fcntl;
 394         }
 395
 396         obj->type = SHM_OBJECT_MEM;
 397         obj->memory_map = mem;
 398         obj->memory_map_size = memory_map_size;
 399         obj->allocated_len = memory_map_size;
 400         obj->index = table->allocated_len++;
 401
 402         return obj;
 403
 404 error_fcntl:
 405         return NULL;
 406 }
 407
 408 static
 409 void shmp_object_destroy(struct shm_object *obj, int consumer)
 410 {
 411         switch (obj->type) {
 412         case SHM_OBJECT_SHM:
 413         {
 414                 int ret, i;
 415
 416                 ret = munmap(obj->memory_map, obj->memory_map_size);
 417                 if (ret) {
 418                         PERROR("umnmap");
 419                         assert(0);
 420                 }
 421
 422                 if (obj->shm_fd_ownership) {
 423                         /* Delete FDs only if called from app (not consumer). */
 424                         if (!consumer) {
 425                                 lttng_ust_lock_fd_tracker();
 426                                 ret = close(obj->shm_fd);
 427                                 if (!ret) {
 428                                         lttng_ust_delete_fd_from_tracker(obj->shm_fd);
 429                                 } else {
 430                                         PERROR("close");
 431                                         assert(0);
 432                                 }
 433                                 lttng_ust_unlock_fd_tracker();
 434                         } else {
 435                                 ret = close(obj->shm_fd);
 436                                 if (ret) {
 437                                         PERROR("close");
 438                                         assert(0);
 439                                 }
 440                         }
 441                 }
 442                 for (i = 0; i < 2; i++) {
 443                         if (obj->wait_fd[i] < 0)
 444                                 continue;
 445                         if (!consumer) {
 446                                 lttng_ust_lock_fd_tracker();
 447                                 ret = close(obj->wait_fd[i]);
 448                                 if (!ret) {
 449                                         lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
 450                                 } else {
 451                                         PERROR("close");
 452                                         assert(0);
 453                                 }
 454                                 lttng_ust_unlock_fd_tracker();
 455                         } else {
 456                                 ret = close(obj->wait_fd[i]);
 457                                 if (ret) {
 458                                         PERROR("close");
 459                                         assert(0);
 460                                 }
 461                         }
 462                 }
 463                 break;
 464         }
 465         case SHM_OBJECT_MEM:
 466         {
 467                 int ret, i;
 468
 469                 for (i = 0; i < 2; i++) {
 470                         if (obj->wait_fd[i] < 0)
 471                                 continue;
 472                         if (!consumer) {
 473                                 lttng_ust_lock_fd_tracker();
 474                                 ret = close(obj->wait_fd[i]);
 475                                 if (!ret) {
 476                                         lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
 477                                 } else {
 478                                         PERROR("close");
 479                                         assert(0);
 480                                 }
 481                                 lttng_ust_unlock_fd_tracker();
 482                         } else {
 483                                 ret = close(obj->wait_fd[i]);
 484                                 if (ret) {
 485                                         PERROR("close");
 486                                         assert(0);
 487                                 }
 488                         }
 489                 }
 490                 free(obj->memory_map);
 491                 break;
 492         }
 493         default:
 494                 assert(0);
 495         }
 496 }
 497
 498 void shm_object_table_destroy(struct shm_object_table *table, int consumer)
 499 {
 500         int i;
 501
 502         for (i = 0; i < table->allocated_len; i++)
 503                 shmp_object_destroy(&table->objects[i], consumer);
 504         free(table);
 505 }
 506
 507 /*
 508  * zalloc_shm - allocate memory within a shm object.
 509  *
 510  * Shared memory is already zeroed by shmget.
 511  * *NOT* multithread-safe (should be protected by mutex).
 512  * Returns a -1, -1 tuple on error.
 513  */
 514 struct shm_ref zalloc_shm(struct shm_object *obj, size_t len)
 515 {
 516         struct shm_ref ref;
 517         struct shm_ref shm_ref_error = { -1, -1 };
 518
 519         if (obj->memory_map_size - obj->allocated_len < len)
 520                 return shm_ref_error;
 521         ref.index = obj->index;
 522         ref.offset =  obj->allocated_len;
 523         obj->allocated_len += len;
 524         return ref;
 525 }
 526
 527 void align_shm(struct shm_object *obj, size_t align)
 528 {
 529         size_t offset_len = lttng_ust_offset_align(obj->allocated_len, align);
 530         obj->allocated_len += offset_len;
 531 }