X-Git-Url: http://git.lttng.org/?p=lttng-ust.git;a=blobdiff_plain;f=libringbuffer%2Fshm.c;h=0153578c93a2cfddf3bdd69ae57324b516086651;hp=dbd3f34812f55b314eeedcb89b447d301b14d7a6;hb=bfcda6cea270952898ea122c375f2ed19105adef;hpb=3a81f31dfe668b3c56207745ca698c7db4dc0890

diff --git a/libringbuffer/shm.c b/libringbuffer/shm.c
index dbd3f348..0153578c 100644
--- a/libringbuffer/shm.c
+++ b/libringbuffer/shm.c
@@ -18,10 +18,13 @@
  * 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>
@@ -29,9 +32,12 @@
 #include <signal.h>
 #include <dirent.h>
 #include <lttng/align.h>
-#include <helper.h>
 #include <limits.h>
+#ifdef HAVE_LIBNUMA
+#include <numa.h>
+#endif
 #include <helper.h>
+#include <ust-fd.h>
 
 /*
  * Ensure we have the required amount of space available by writing 0
@@ -77,19 +83,23 @@ struct shm_object_table *shm_object_table_create(size_t max_nb_obj)
 
 	table = zmalloc(sizeof(struct shm_object_table) +
 			max_nb_obj * sizeof(table->objects[0]));
+	if (!table)
+		return NULL;
 	table->size = max_nb_obj;
 	return table;
 }
 
-struct shm_object *shm_object_table_append(struct shm_object_table *table,
-					   size_t memory_map_size)
+static
+struct shm_object *_shm_object_table_alloc_shm(struct shm_object_table *table,
+					   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];
@@ -115,58 +125,9 @@ struct shm_object *shm_object_table_append(struct shm_object_table *table,
 	}
 	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.
-		 */
-		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");
@@ -177,6 +138,16 @@ struct shm_object *shm_object_table_append(struct shm_object_table *table,
 		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 */
@@ -186,6 +157,7 @@ struct shm_object *shm_object_table_append(struct shm_object_table *table,
 		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 = 0;
@@ -194,23 +166,70 @@ struct shm_object *shm_object_table_append(struct shm_object_table *table,
 	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_fcntl:
+	for (i = 0; i < 2; i++) {
+		ret = close(waitfd[i]);
+		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_pipe:
+	return NULL;
+}
+
+static
+struct shm_object *_shm_object_table_alloc_mem(struct shm_object_table *table,
+					   size_t memory_map_size)
+{
+	struct shm_object *obj;
+	void *memory_map;
+	int waitfd[2], i, ret;
+
+	if (table->allocated_len >= table->size)
+		return NULL;
+	obj = &table->objects[table->allocated_len];
+
+	memory_map = zmalloc(memory_map_size);
+	if (!memory_map)
+		goto alloc_error;
+
+	/* wait_fd: create pipe */
+	ret = pipe(waitfd);
+	if (ret < 0) {
+		PERROR("pipe");
+		goto error_pipe;
+	}
+	for (i = 0; i < 2; i++) {
+		ret = fcntl(waitfd[i], F_SETFD, FD_CLOEXEC);
+		if (ret < 0) {
+			PERROR("fcntl");
+			goto error_fcntl;
 		}
 	}
-error_pthread_sigmask:
+	/* The write end of the pipe needs to be non-blocking */
+	ret = fcntl(waitfd[1], F_SETFL, O_NONBLOCK);
+	if (ret < 0) {
+		PERROR("fcntl");
+		goto error_fcntl;
+	}
+	memcpy(obj->wait_fd, waitfd, sizeof(waitfd));
+
+	/* no shm_fd */
+	obj->shm_fd = -1;
+	obj->shm_fd_ownership = 0;
+
+	obj->type = SHM_OBJECT_MEM;
+	obj->memory_map = memory_map;
+	obj->memory_map_size = memory_map_size;
+	obj->allocated_len = 0;
+	obj->index = table->allocated_len++;
+
+	return obj;
+
 error_fcntl:
 	for (i = 0; i < 2; i++) {
 		ret = close(waitfd[i]);
@@ -220,24 +239,80 @@ error_fcntl:
 		}
 	}
 error_pipe:
+	free(memory_map);
+alloc_error:
 	return NULL;
-	
 }
 
-struct shm_object *shm_object_table_append_shadow(struct shm_object_table *table,
-			int shm_fd, int wait_fd, size_t memory_map_size)
+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, node;
+
+	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
+	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 read end of the pipe. */
-	obj->wait_fd[0] = wait_fd;
-	obj->wait_fd[1] = -1;	/* write end is unset. */
+	/* 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,
@@ -246,6 +321,7 @@ struct shm_object *shm_object_table_append_shadow(struct shm_object_table *table
 		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;
@@ -253,46 +329,149 @@ struct shm_object *shm_object_table_append_shadow(struct shm_object_table *table
 
 	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)
+void shmp_object_destroy(struct shm_object *obj, int consumer)
 {
-	int ret, i;
+	switch (obj->type) {
+	case SHM_OBJECT_SHM:
+	{
+		int ret, i;
 
-	if (!obj->is_shadow) {
 		ret = munmap(obj->memory_map, obj->memory_map_size);
 		if (ret) {
 			PERROR("umnmap");
 			assert(0);
 		}
-	}
-	if (obj->shm_fd >= 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;
+			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);
+	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);
 	}
 }
 
-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);
 }