+error_pipe:
+ free(memory_map);
+alloc_error:
+ return NULL;
+}
+
+/*
+ * libnuma prints errors on the console even for numa_available().
+ * Work-around this limitation by using get_mempolicy() directly to
+ * check whether the kernel supports mempolicy.
+ */
+#ifdef HAVE_LIBNUMA
+static bool lttng_is_numa_available(void)
+{
+ int ret;
+
+ ret = get_mempolicy(NULL, NULL, 0, NULL, 0);
+ if (ret && errno == ENOSYS) {
+ return false;
+ }
+ return numa_available() > 0;
+}
+#endif
+
+struct shm_object *shm_object_table_alloc(struct shm_object_table *table,
+ size_t memory_map_size,
+ enum shm_object_type type,
+ int stream_fd,
+ int cpu)
+{
+ struct shm_object *shm_object;
+#ifdef HAVE_LIBNUMA
+ int oldnode = 0, node;
+ bool numa_avail;
+
+ numa_avail = lttng_is_numa_available();
+ if (numa_avail) {
+ oldnode = numa_preferred();
+ if (cpu >= 0) {
+ node = numa_node_of_cpu(cpu);
+ if (node >= 0)
+ numa_set_preferred(node);
+ }
+ if (cpu < 0 || node < 0)
+ numa_set_localalloc();
+ }
+#endif /* HAVE_LIBNUMA */
+ switch (type) {
+ case SHM_OBJECT_SHM:
+ shm_object = _shm_object_table_alloc_shm(table, memory_map_size,
+ stream_fd);
+ break;
+ case SHM_OBJECT_MEM:
+ shm_object = _shm_object_table_alloc_mem(table, memory_map_size);
+ break;
+ default:
+ assert(0);
+ }
+#ifdef HAVE_LIBNUMA
+ if (numa_avail)
+ numa_set_preferred(oldnode);
+#endif /* HAVE_LIBNUMA */
+ return shm_object;
+}
+
+struct shm_object *shm_object_table_append_shm(struct shm_object_table *table,
+ int shm_fd, int wakeup_fd, uint32_t stream_nr,
+ size_t memory_map_size)
+{
+ struct shm_object *obj;
+ char *memory_map;
+ int ret;
+
+ if (table->allocated_len >= table->size)
+ return NULL;
+ /* streams _must_ be received in sequential order, else fail. */
+ if (stream_nr + 1 != table->allocated_len)
+ return NULL;
+
+ obj = &table->objects[table->allocated_len];
+
+ /* wait_fd: set write end of the pipe. */
+ obj->wait_fd[0] = -1; /* read end is unset */
+ obj->wait_fd[1] = wakeup_fd;
+ obj->shm_fd = shm_fd;
+ obj->shm_fd_ownership = 1;
+
+ ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC);
+ if (ret < 0) {
+ PERROR("fcntl");
+ goto error_fcntl;
+ }
+ /* The write end of the pipe needs to be non-blocking */
+ ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
+ if (ret < 0) {
+ PERROR("fcntl");
+ goto error_fcntl;
+ }
+
+ /* memory_map: mmap */
+ memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
+ MAP_SHARED, shm_fd, 0);
+ if (memory_map == MAP_FAILED) {
+ PERROR("mmap");
+ goto error_mmap;
+ }
+ obj->type = SHM_OBJECT_SHM;
+ obj->memory_map = memory_map;
+ obj->memory_map_size = memory_map_size;
+ obj->allocated_len = memory_map_size;
+ obj->index = table->allocated_len++;
+
+ return obj;
+
+error_fcntl:
+error_mmap:
+ return NULL;
+}
+
+/*
+ * Passing ownership of mem to object.
+ */
+struct shm_object *shm_object_table_append_mem(struct shm_object_table *table,
+ void *mem, size_t memory_map_size, int wakeup_fd)
+{
+ struct shm_object *obj;
+ int ret;
+
+ if (table->allocated_len >= table->size)
+ return NULL;
+ obj = &table->objects[table->allocated_len];
+
+ obj->wait_fd[0] = -1; /* read end is unset */
+ obj->wait_fd[1] = wakeup_fd;
+ obj->shm_fd = -1;
+ obj->shm_fd_ownership = 0;
+
+ ret = fcntl(obj->wait_fd[1], F_SETFD, FD_CLOEXEC);
+ if (ret < 0) {
+ PERROR("fcntl");
+ goto error_fcntl;
+ }
+ /* The write end of the pipe needs to be non-blocking */
+ ret = fcntl(obj->wait_fd[1], F_SETFL, O_NONBLOCK);
+ if (ret < 0) {
+ PERROR("fcntl");
+ goto error_fcntl;
+ }
+
+ obj->type = SHM_OBJECT_MEM;
+ obj->memory_map = mem;
+ obj->memory_map_size = memory_map_size;
+ obj->allocated_len = memory_map_size;
+ obj->index = table->allocated_len++;
+
+ return obj;
+
+error_fcntl:
+ return NULL;
+}
+
+static
+void shmp_object_destroy(struct shm_object *obj, int consumer)
+{
+ switch (obj->type) {
+ case SHM_OBJECT_SHM:
+ {
+ int ret, i;
+
+ ret = munmap(obj->memory_map, obj->memory_map_size);
+ if (ret) {
+ PERROR("umnmap");
+ assert(0);
+ }
+
+ if (obj->shm_fd_ownership) {
+ /* Delete FDs only if called from app (not consumer). */
+ if (!consumer) {
+ lttng_ust_lock_fd_tracker();
+ ret = close(obj->shm_fd);
+ if (!ret) {
+ lttng_ust_delete_fd_from_tracker(obj->shm_fd);
+ } else {
+ PERROR("close");
+ assert(0);
+ }
+ lttng_ust_unlock_fd_tracker();
+ } else {
+ ret = close(obj->shm_fd);
+ if (ret) {
+ PERROR("close");
+ assert(0);
+ }
+ }
+ }
+ for (i = 0; i < 2; i++) {
+ if (obj->wait_fd[i] < 0)
+ continue;
+ if (!consumer) {
+ lttng_ust_lock_fd_tracker();
+ ret = close(obj->wait_fd[i]);
+ if (!ret) {
+ lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
+ } else {
+ PERROR("close");
+ assert(0);
+ }
+ lttng_ust_unlock_fd_tracker();
+ } else {
+ ret = close(obj->wait_fd[i]);
+ if (ret) {
+ PERROR("close");
+ assert(0);
+ }
+ }
+ }
+ break;
+ }
+ case SHM_OBJECT_MEM:
+ {
+ int ret, i;
+
+ for (i = 0; i < 2; i++) {
+ if (obj->wait_fd[i] < 0)
+ continue;
+ if (!consumer) {
+ lttng_ust_lock_fd_tracker();
+ ret = close(obj->wait_fd[i]);
+ if (!ret) {
+ lttng_ust_delete_fd_from_tracker(obj->wait_fd[i]);
+ } else {
+ PERROR("close");
+ assert(0);
+ }
+ lttng_ust_unlock_fd_tracker();
+ } else {
+ ret = close(obj->wait_fd[i]);
+ if (ret) {
+ PERROR("close");
+ assert(0);
+ }
+ }
+ }
+ free(obj->memory_map);
+ break;
+ }
+ default:
+ assert(0);
+ }