* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
+#define _LGPL_SOURCE
+#include <config.h>
#include "shm.h"
#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>
+#include <sys/types.h>
#include <sys/stat.h> /* For mode constants */
#include <fcntl.h> /* For O_* constants */
#include <assert.h>
#include <signal.h>
#include <dirent.h>
#include <lttng/align.h>
-#include <helper.h>
#include <limits.h>
+#include <stdbool.h>
+#ifdef HAVE_LIBNUMA
+#include <numa.h>
+#include <numaif.h>
+#endif
#include <helper.h>
+#include <ust-fd.h>
+#include "mmap.h"
/*
* Ensure we have the required amount of space available by writing 0
static
struct shm_object *_shm_object_table_alloc_shm(struct shm_object_table *table,
- size_t memory_map_size)
+ size_t memory_map_size,
+ int stream_fd)
{
- int shmfd, waitfd[2], ret, i, sigblocked = 0;
+ int shmfd, waitfd[2], ret, i;
struct shm_object *obj;
char *memory_map;
- char tmp_name[NAME_MAX] = "/ust-shm-tmp-XXXXXX";
- sigset_t all_sigs, orig_sigs;
+ if (stream_fd < 0)
+ return NULL;
if (table->allocated_len >= table->size)
return NULL;
obj = &table->objects[table->allocated_len];
}
memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
- /* shm_fd: create shm */
-
- /*
- * Theoretically, we could leak a shm if the application crashes
- * between open and unlink. Disable signals on this thread for
- * increased safety against this scenario.
- */
- sigfillset(&all_sigs);
- ret = pthread_sigmask(SIG_BLOCK, &all_sigs, &orig_sigs);
- if (ret == -1) {
- PERROR("pthread_sigmask");
- goto error_pthread_sigmask;
- }
- sigblocked = 1;
+ /* create shm */
- /*
- * Allocate shm, and immediately unlink its shm oject, keeping
- * only the file descriptor as a reference to the object. If it
- * already exists (caused by short race window during which the
- * global object exists in a concurrent shm_open), simply retry.
- * We specifically do _not_ use the / at the beginning of the
- * pathname so that some OS implementations can keep it local to
- * the process (POSIX leaves this implementation-defined).
- */
- do {
- /*
- * Using mktemp filename with O_CREAT | O_EXCL open
- * flags.
- */
- (void) mktemp(tmp_name);
- if (tmp_name[0] == '\0') {
- PERROR("mktemp");
- goto error_shm_open;
- }
- shmfd = shm_open(tmp_name,
- O_CREAT | O_EXCL | O_RDWR, 0700);
- } while (shmfd < 0 && (errno == EEXIST || errno == EACCES));
- if (shmfd < 0) {
- PERROR("shm_open");
- goto error_shm_open;
- }
- ret = shm_unlink(tmp_name);
- if (ret < 0 && errno != ENOENT) {
- PERROR("shm_unlink");
- goto error_shm_release;
- }
- sigblocked = 0;
- ret = pthread_sigmask(SIG_SETMASK, &orig_sigs, NULL);
- if (ret == -1) {
- PERROR("pthread_sigmask");
- goto error_sigmask_release;
- }
+ shmfd = stream_fd;
ret = zero_file(shmfd, memory_map_size);
if (ret) {
PERROR("zero_file");
PERROR("ftruncate");
goto error_ftruncate;
}
+ /*
+ * Also ensure the file metadata is synced with the storage by using
+ * fsync(2).
+ */
+ ret = fsync(shmfd);
+ if (ret) {
+ PERROR("fsync");
+ goto error_fsync;
+ }
+ obj->shm_fd_ownership = 0;
obj->shm_fd = shmfd;
/* memory_map: mmap */
memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
- MAP_SHARED, shmfd, 0);
+ MAP_SHARED | LTTNG_MAP_POPULATE, shmfd, 0);
if (memory_map == MAP_FAILED) {
PERROR("mmap");
goto error_mmap;
return obj;
error_mmap:
+error_fsync:
error_ftruncate:
-error_shm_release:
error_zero_file:
-error_sigmask_release:
- ret = close(shmfd);
- if (ret) {
- PERROR("close");
- assert(0);
- }
-error_shm_open:
- if (sigblocked) {
- ret = pthread_sigmask(SIG_SETMASK, &orig_sigs, NULL);
- if (ret == -1) {
- PERROR("pthread_sigmask");
- }
- }
-error_pthread_sigmask:
error_fcntl:
for (i = 0; i < 2; i++) {
ret = close(waitfd[i]);
/* no shm_fd */
obj->shm_fd = -1;
+ obj->shm_fd_ownership = 0;
obj->type = SHM_OBJECT_MEM;
obj->memory_map = memory_map;
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)
+ 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:
- return _shm_object_table_alloc_shm(table, memory_map_size);
+ shm_object = _shm_object_table_alloc_shm(table, memory_map_size,
+ stream_fd);
+ break;
case SHM_OBJECT_MEM:
- return _shm_object_table_alloc_mem(table, memory_map_size);
+ shm_object = _shm_object_table_alloc_mem(table, memory_map_size);
+ break;
default:
assert(0);
}
- return NULL;
+#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,
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) {
/* memory_map: mmap */
memory_map = mmap(NULL, memory_map_size, PROT_READ | PROT_WRITE,
- MAP_SHARED, shm_fd, 0);
+ MAP_SHARED | LTTNG_MAP_POPULATE, shm_fd, 0);
if (memory_map == MAP_FAILED) {
PERROR("mmap");
goto error_mmap;
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) {
}
static
-void shmp_object_destroy(struct shm_object *obj)
+void shmp_object_destroy(struct shm_object *obj, int consumer)
{
switch (obj->type) {
case SHM_OBJECT_SHM:
PERROR("umnmap");
assert(0);
}
- ret = close(obj->shm_fd);
- if (ret) {
- PERROR("close");
- 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;
- ret = close(obj->wait_fd[i]);
- if (ret) {
- PERROR("close");
- assert(0);
+ 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;
for (i = 0; i < 2; i++) {
if (obj->wait_fd[i] < 0)
continue;
- ret = close(obj->wait_fd[i]);
- if (ret) {
- PERROR("close");
- assert(0);
+ 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);
}
}
-void shm_object_table_destroy(struct shm_object_table *table)
+void shm_object_table_destroy(struct shm_object_table *table, int consumer)
{
int i;
for (i = 0; i < table->allocated_len; i++)
- shmp_object_destroy(&table->objects[i]);
+ shmp_object_destroy(&table->objects[i], consumer);
free(table);
}