src/libringbuffer/shm.c

   1 /*
   2  * SPDX-License-Identifier: LGPL-2.1-only
   3  *
   4  * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   5  */
   6
   7 #define _LGPL_SOURCE
   8 #include "shm.h"
   9 #include <unistd.h>
  10 #include <fcntl.h>
  11 #include <sys/mman.h>
  12 #include <sys/types.h>
  13 #include <sys/stat.h>   /* For mode constants */
  14 #include <fcntl.h>      /* For O_* constants */
  15 #include <assert.h>
  16 #include <stdio.h>
  17 #include <signal.h>
  18 #include <dirent.h>
  19 #include <limits.h>
  20 #include <stdbool.h>
  21 #include <stdint.h>
  22
  23 #ifdef HAVE_LIBNUMA
  24 #include <numa.h>
  25 #include <numaif.h>
  26 #endif
  27
  28 #include <lttng/ust-utils.h>
  29
  30 #include <ust-helper.h>
  31 #include <ust-fd.h>
  32 #include "mmap.h"
  33
  34 /*
  35  * Ensure we have the required amount of space available by writing 0
  36  * into the entire buffer. Not doing so can trigger SIGBUS when going
  37  * beyond the available shm space.
  38  */
  39 static
  40 int zero_file(int fd, size_t len)
  41 {
  42         ssize_t retlen;
  43         size_t written = 0;
  44         char *zeropage;
  45         long pagelen;
  46         int ret;
  47
  48         pagelen = sysconf(_SC_PAGESIZE);
  49         if (pagelen < 0)
  50                 return (int) pagelen;
  51         zeropage = calloc(pagelen, 1);
  52         if (!zeropage)
  53                 return -ENOMEM;
  54
  55         while (len > written) {
  56                 do {
  57                         retlen = write(fd, zeropage,
  58                                 min_t(size_t, pagelen, len - written));
  59                 } while (retlen == -1UL && errno == EINTR);
  60                 if (retlen < 0) {
  61                         ret = (int) retlen;
  62                         goto error;
  63                 }
  64                 written += retlen;
  65         }
  66         ret = 0;
  67 error:
  68         free(zeropage);
  69         return ret;
  70 }
  71
  72 struct shm_object_table *shm_object_table_create(size_t max_nb_obj)
  73 {
  74         struct shm_object_table *table;
  75
  76         table = zmalloc(sizeof(struct shm_object_table) +
  77                         max_nb_obj * sizeof(table->objects[0]));
  78         if (!table)
  79                 return NULL;
  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                                            int stream_fd)
  88 {
  89         int shmfd, waitfd[2], ret, i;
  90         struct shm_object *obj;
  91         char *memory_map;
  92
  93         if (stream_fd < 0)
  94                 return NULL;
  95         if (table->allocated_len >= table->size)
  96                 return NULL;
  97         obj = &table->objects[table->allocated_len];
  98
  99         /* wait_fd: create pipe */
 100         ret = pipe(waitfd);
 101         if (ret < 0) {
 102                 PERROR("pipe");
 103                 goto error_pipe;
 104         }
 105         for (i = 0; i < 2; i++) {
 106                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 107                 if (ret < 0) {
 108                         PERROR("fcntl");
 109                         goto error_fcntl;
 110                 }
 111         }
 112         /* The write end of the pipe needs to be non-blocking */
 113         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 114         if (ret < 0) {
 115                 PERROR("fcntl");
 116                 goto error_fcntl;
 117         }
 118         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 119
 120         /*
 121          * Set POSIX shared memory object size
 122          *
 123          * First, use ftruncate() to set its size, some implementations won't
 124          * allow writes past the size set by ftruncate.
 125          * Then, use write() to fill it with zeros, this allows us to fully
 126          * allocate it and detect a shortage of shm space without dealing with
 127          * a SIGBUS.
 128          */
 129
 130         shmfd = stream_fd;
 131         ret = ftruncate(shmfd, memory_map_size);
 132         if (ret) {
 133                 PERROR("ftruncate");
 134                 goto error_ftruncate;
 135         }
 136         ret = zero_file(shmfd, memory_map_size);
 137         if (ret) {
 138                 PERROR("zero_file");
 139                 goto error_zero_file;
 140         }
 141
 142         /*
 143          * Also ensure the file metadata is synced with the storage by using
 144          * fsync(2). Some platforms don't allow fsync on POSIX shm fds, ignore
 145          * EINVAL accordingly.
 146          */
 147         ret = fsync(shmfd);
 148         if (ret && errno != EINVAL) {
 149                 PERROR("fsync");
 150                 goto error_fsync;
 151         }
 152         obj->shm_fd_ownership = 0;
 153         obj->shm_fd = shmfd;
 154
 155         /* memory_map: mmap */
 156         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 157                           MAP_SHARED | LTTNG_MAP_POPULATE, shmfd, 0);
 158         if (memory_map == MAP_FAILED) {
 159                 PERROR("mmap");
 160                 goto error_mmap;
 161         }
 162         obj->type = SHM_OBJECT_SHM;
 163         obj->memory_map = memory_map;
 164         obj->memory_map_size = memory_map_size;
 165         obj->allocated_len = 0;
 166         obj->index = table->allocated_len++;
 167
 168         return obj;
 169
 170 error_mmap:
 171 error_fsync:
 172 error_ftruncate:
 173 error_zero_file:
 174 error_fcntl:
 175         for (i = 0; i < 2; i++) {
 176                 ret = close(waitfd[i]);
 177                 if (ret) {
 178                         PERROR("close");
 179                         assert(0);
 180                 }
 181         }
 182 error_pipe:
 183         return NULL;
 184 }
 185
 186 static
 187 struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table,
 188                                            size_t memory_map_size)
 189 {
 190         struct shm_object *obj;
 191         void *memory_map;
 192         int waitfd[2], i, ret;
 193
 194         if (table->allocated_len >= table->size)
 195                 return NULL;
 196         obj = &table->objects[table->allocated_len];
 197
 198         memory_map = zmalloc(memory_map_size);
 199         if (!memory_map)
 200                 goto alloc_error;
 201
 202         /* wait_fd: create pipe */
 203         ret = pipe(waitfd);
 204         if (ret < 0) {
 205                 PERROR("pipe");
 206                 goto error_pipe;
 207         }
 208         for (i = 0; i < 2; i++) {
 209                 ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
 210                 if (ret < 0) {
 211                         PERROR("fcntl");
 212                         goto error_fcntl;
 213                 }
 214         }
 215         /* The write end of the pipe needs to be non-blocking */
 216         ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
 217         if (ret < 0) {
 218                 PERROR("fcntl");
 219                 goto error_fcntl;
 220         }
 221         memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
 222
 223         /* no shm_fd */
 224         obj->shm_fd = -1;
 225         obj->shm_fd_ownership = 0;
 226
 227         obj->type = SHM_OBJECT_MEM;
 228         obj->memory_map = memory_map;
 229         obj->memory_map_size = memory_map_size;
 230         obj->allocated_len = 0;
 231         obj->index = table->allocated_len++;
 232
 233         return obj;
 234
 235 error_fcntl:
 236         for (i = 0; i < 2; i++) {
 237                 ret = close(waitfd[i]);
 238                 if (ret) {
 239                         PERROR("close");
 240                         assert(0);
 241                 }
 242         }
 243 error_pipe:
 244         free(memory_map);
 245 alloc_error:
 246         return NULL;
 247 }
 248
 249 /*
 250  * libnuma prints errors on the console even for numa_available().
 251  * Work-around this limitation by using get_mempolicy() directly to
 252  * check whether the kernel supports mempolicy.
 253  */
 254 #ifdef HAVE_LIBNUMA
 255 static bool lttng_is_numa_available(void)
 256 {
 257         int ret;
 258
 259         ret = get_mempolicy(NULL, NULL, 0, NULL, 0);
 260         if (ret && errno == ENOSYS) {
 261                 return false;
 262         }
 263         return numa_available() > 0;
 264 }
 265 #endif
 266
 267 struct shm_object *shm_object_table_alloc(struct shm_object_table *table,
 268                         size_t memory_map_size,
 269                         enum shm_object_type type,
 270                         int stream_fd,
 271                         int cpu)
 272 {
 273         struct shm_object *shm_object;
 274 #ifdef HAVE_LIBNUMA
 275         int oldnode = 0, node;
 276         bool numa_avail;
 277
 278         numa_avail = lttng_is_numa_available();
 279         if (numa_avail) {
 280                 oldnode = numa_preferred();
 281                 if (cpu >= 0) {
 282                         node = numa_node_of_cpu(cpu);
 283                         if (node >= 0)
 284                                 numa_set_preferred(node);
 285                 }
 286                 if (cpu < 0 || node < 0)
 287                         numa_set_localalloc();
 288         }
 289 #endif /* HAVE_LIBNUMA */
 290         switch (type) {
 291         case SHM_OBJECT_SHM:
 292                 shm_object = _shm_object_table_alloc_shm(table, memory_map_size,
 293                                 stream_fd);
 294                 break;
 295         case SHM_OBJECT_MEM:
 296                 shm_object = _shm_object_table_alloc_mem(table, memory_map_size);
 297                 break;
 298         default:
 299                 assert(0);
 300         }
 301 #ifdef HAVE_LIBNUMA
 302         if (numa_avail)
 303                 numa_set_preferred(oldnode);
 304 #endif /* HAVE_LIBNUMA */
 305         return shm_object;
 306 }
 307
 308 struct shm_object *shm_object_table_append_shm(struct shm_object_table *table,
 309                         int shm_fd, int wakeup_fd, uint32_t stream_nr,
 310                         size_t memory_map_size)
 311 {
 312         struct shm_object *obj;
 313         char *memory_map;
 314         int ret;
 315
 316         if (table->allocated_len >= table->size)
 317                 return NULL;
 318         /* streams _must_ be received in sequential order, else fail. */
 319         if (stream_nr + 1 != table->allocated_len)
 320                 return NULL;
 321
 322         obj = &table->objects[table->allocated_len];
 323
 324         /* wait_fd: set write end of the pipe. */
 325         obj->wait_fd[0] = -1;   /* read end is unset */
 326         obj->wait_fd[1] = wakeup_fd;
 327         obj->shm_fd = shm_fd;
 328         obj->shm_fd_ownership = 1;
 329
 330         /* The write end of the pipe needs to be non-blocking */
 331         ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
 332         if (ret < 0) {
 333                 PERROR("fcntl");
 334                 goto error_fcntl;
 335         }
 336
 337         /* memory_map: mmap */
 338         memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
 339                           MAP_SHARED | LTTNG_MAP_POPULATE, shm_fd, 0);
 340         if (memory_map == MAP_FAILED) {
 341                 PERROR("mmap");
 342                 goto error_mmap;
 343         }
 344         obj->type = SHM_OBJECT_SHM;
 345         obj->memory_map = memory_map;
 346         obj->memory_map_size = memory_map_size;
 347         obj->allocated_len = memory_map_size;
 348         obj->index = table->allocated_len++;
 349
 350         return obj;
 351
 352 error_fcntl:
 353 error_mmap:
 354         return NULL;
 355 }
 356
 357 /*
 358  * Passing ownership of mem to object.
 359  */
 360 struct shm_object *shm_object_table_append_mem(struct shm_object_table *table,
 361                         void *mem, size_t memory_map_size, int wakeup_fd)
 362 {
 363         struct shm_object *obj;
 364         int ret;
 365
 366         if (table->allocated_len >= table->size)
 367                 return NULL;
 368         obj = &table->objects[table->allocated_len];
 369
 370         obj->wait_fd[0] = -1;   /* read end is unset */
 371         obj->wait_fd[1] = wakeup_fd;
 372         obj->shm_fd = -1;
 373         obj->shm_fd_ownership = 0;
 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, int consumer)
 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
 413                 if (obj->shm_fd_ownership) {
 414                         /* Delete FDs only if called from app (not consumer). */
 415                         if (!consumer) {
 416                                 lttng_ust_lock_fd_tracker();
 417                                 ret = close(obj->shm_fd);
 418                                 if (!ret) {
 419                                         lttng_ust_delete_fd_from_tracker(obj->shm_fd);
 420                                 } else {
 421                                         PERROR("close");
 422                                         assert(0);
 423                                 }
 424                                 lttng_ust_unlock_fd_tracker();
 425                         } else {
 426                                 ret = close(obj->shm_fd);
 427                                 if (ret) {
 428                                         PERROR("close");
 429                                         assert(0);
 430                                 }
 431                         }
 432                 }
 433                 for (i = 0; i < 2; i++) {
 434                         if (obj->wait_fd[i] < 0)
 435                                 continue;
 436                         if (!consumer) {
 437                                 lttng_ust_lock_fd_tracker();
 438                                 ret = close(obj->wait_fd[i]);
 439                                 if (!ret) {
 440                                         lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
 441                                 } else {
 442                                         PERROR("close");
 443                                         assert(0);
 444                                 }
 445                                 lttng_ust_unlock_fd_tracker();
 446                         } else {
 447                                 ret = close(obj->wait_fd[i]);
 448                                 if (ret) {
 449                                         PERROR("close");
 450                                         assert(0);
 451                                 }
 452                         }
 453                 }
 454                 break;
 455         }
 456         case SHM_OBJECT_MEM:
 457         {
 458                 int ret, i;
 459
 460                 for (i = 0; i < 2; i++) {
 461                         if (obj->wait_fd[i] < 0)
 462                                 continue;
 463                         if (!consumer) {
 464                                 lttng_ust_lock_fd_tracker();
 465                                 ret = close(obj->wait_fd[i]);
 466                                 if (!ret) {
 467                                         lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
 468                                 } else {
 469                                         PERROR("close");
 470                                         assert(0);
 471                                 }
 472                                 lttng_ust_unlock_fd_tracker();
 473                         } else {
 474                                 ret = close(obj->wait_fd[i]);
 475                                 if (ret) {
 476                                         PERROR("close");
 477                                         assert(0);
 478                                 }
 479                         }
 480                 }
 481                 free(obj->memory_map);
 482                 break;
 483         }
 484         default:
 485                 assert(0);
 486         }
 487 }
 488
 489 void shm_object_table_destroy(struct shm_object_table *table, int consumer)
 490 {
 491         int i;
 492
 493         for (i = 0; i < table->allocated_len; i++)
 494                 shmp_object_destroy(&table->objects[i], consumer);
 495         free(table);
 496 }
 497
 498 /*
 499  * zalloc_shm - allocate memory within a shm object.
 500  *
 501  * Shared memory is already zeroed by shmget.
 502  * *NOT* multithread-safe (should be protected by mutex).
 503  * Returns a -1, -1 tuple on error.
 504  */
 505 struct shm_ref zalloc_shm(struct shm_object *obj, size_t len)
 506 {
 507         struct shm_ref ref;
 508         struct shm_ref shm_ref_error = { -1, -1 };
 509
 510         if (obj->memory_map_size - obj->allocated_len < len)
 511                 return shm_ref_error;
 512         ref.index = obj->index;
 513         ref.offset =  obj->allocated_len;
 514         obj->allocated_len += len;
 515         return ref;
 516 }
 517
 518 void align_shm(struct shm_object *obj, size_t align)
 519 {
 520         size_t offset_len = lttng_ust_offset_align(obj->allocated_len, align);
 521         obj->allocated_len += offset_len;
 522 }