Fix: consumerd: fd still open after `lttng snapshot record` returns

[lttng-tools.git] / src / common / trace-chunk.c

diff --git a/src/common/trace-chunk.c b/src/common/trace-chunk.c
index f909a682f6b376769b9cdbd27662e7d36f5e64b7..5518051960e654b654c118429e00e81b211e3cdc 100644 (file)
--- a/src/common/trace-chunk.c
+++ b/src/common/trace-chunk.c
@@ -97,8 +97,14 @@ struct lttng_trace_chunk {
        char *path;
        /* An unset id means the chunk is anonymous. */
        LTTNG_OPTIONAL(uint64_t) id;
+
+       /*
+        * The creation and close timestamps are NOT monotonic.
+        * They must not be used in context were monotonicity is required.
+        */
        LTTNG_OPTIONAL(time_t) timestamp_creation;
        LTTNG_OPTIONAL(time_t) timestamp_close;
+
        LTTNG_OPTIONAL(struct chunk_credentials) credentials;
        struct lttng_directory_handle *session_output_directory;
        struct lttng_directory_handle *chunk_directory;
@@ -599,11 +605,20 @@ enum lttng_trace_chunk_status lttng_trace_chunk_set_close_timestamp(
                status = LTTNG_TRACE_CHUNK_STATUS_INVALID_OPERATION;
                goto end;
        }
+
+       /*
+        * Note: we do not enforce that the closing timestamp be greater or
+        * equal to the begin timestamp. These timestamps are used for
+        * generating the chunk name and should only be used in context where
+        * the monotonicity of time is not important. The source of those
+        * timestamps is NOT monotonic and represent the system calendar time,
+        * also know as the wall time.
+        */
        if (chunk->timestamp_creation.value > close_ts) {
-               ERR("Failed to set trace chunk close timestamp: close timestamp is before creation timestamp");
-               status = LTTNG_TRACE_CHUNK_STATUS_INVALID_ARGUMENT;
-               goto end;
+               WARN("Set trace chunk close timestamp: close timestamp is before creation timestamp, begin : %ld, close : %ld",
+                               chunk->timestamp_creation.value, close_ts);
        }
+
        LTTNG_OPTIONAL_SET(&chunk->timestamp_close, close_ts);
        if (!chunk->name_overridden) {
                free(chunk->name);
@@ -859,7 +874,7 @@ enum lttng_trace_chunk_status lttng_trace_chunk_rename_path_no_lock(
                 */
                chunk->chunk_directory = rename_directory;
                rename_directory = NULL;
-       } else {
+       } else if (old_path) {
                size_t i, count = lttng_dynamic_pointer_array_get_count(
                                &chunk->top_level_directories);
                const bool reference_acquired = lttng_directory_handle_get(
@@ -905,9 +920,12 @@ enum lttng_trace_chunk_status lttng_trace_chunk_rename_path_no_lock(
                if (status != LTTNG_TRACE_CHUNK_STATUS_OK) {
                        ERR("Error removing subdirectory '%s' file when deleting chunk",
                                old_path);
-                       ret = -1;
                        goto end;
                }
+       } else {
+               /* Unexpected !old_path && !path. */
+               status = LTTNG_TRACE_CHUNK_STATUS_INVALID_ARGUMENT;
+               goto end;
        }
 
 skip_move:
@@ -1027,7 +1045,7 @@ enum lttng_trace_chunk_status lttng_trace_chunk_set_as_owner(
                status = LTTNG_TRACE_CHUNK_STATUS_ERROR;
                goto end;
        }
-       if (chunk->path[0] != '\0') {
+       if (chunk->path && chunk->path[0] != '\0') {
                ret = lttng_directory_handle_create_subdirectory_as_user(
                                session_output_directory,
                                chunk->path,
@@ -1491,7 +1509,7 @@ end:
        return status;
 }
 
-LTTNG_HIDDEN
+static
 int lttng_trace_chunk_remove_subdirectory_recursive(struct lttng_trace_chunk *chunk,
                const char *path)
 {
@@ -1810,6 +1828,37 @@ const char *lttng_trace_chunk_command_type_get_name(
        }
 }
 
+LTTNG_HIDDEN
+bool lttng_trace_chunk_ids_equal(const struct lttng_trace_chunk *chunk_a,
+               const struct lttng_trace_chunk *chunk_b)
+{
+       bool equal = false;
+
+       if (chunk_a == chunk_b) {
+               equal = true;
+               goto end;
+       }
+
+       if (!!chunk_a ^ !!chunk_b) {
+               goto end;
+       }
+
+       if (chunk_a->id.is_set ^ chunk_a->id.is_set) {
+               /* One id is set and not the other, thus they are not equal. */
+               goto end;
+       }
+
+       if (!chunk_a->id.is_set) {
+               /* Both ids are unset. */
+               equal = true;
+       } else {
+               equal = chunk_a->id.value == chunk_b->id.value;
+       }
+
+end:
+       return equal;
+}
+
 LTTNG_HIDDEN
 bool lttng_trace_chunk_get(struct lttng_trace_chunk *chunk)
 {
@@ -1822,7 +1871,6 @@ void free_lttng_trace_chunk_registry_element(struct rcu_head *node)
        struct lttng_trace_chunk_registry_element *element =
                        container_of(node, typeof(*element), rcu_node);
 
-       lttng_trace_chunk_fini(&element->chunk);
        free(element);
 }
 
@@ -1843,6 +1891,24 @@ void lttng_trace_chunk_release(struct urcu_ref *ref)
        if (chunk->in_registry_element) {
                struct lttng_trace_chunk_registry_element *element;
 
+               /*
+                * Release internal chunk attributes immediately and
+                * only use the deferred `call_rcu` work to reclaim the
+                * storage.
+                *
+                * This ensures that file handles are released as soon as
+                * possible which works around a problem we encounter with PRAM fs
+                * mounts (and possibly other non-POSIX compliant file systems):
+                * directories that contain files which are open can't be
+                * rmdir().
+                *
+                * This means that the recording of a snapshot could be
+                * completed, but that it would be impossible for the user to
+                * delete it until the deferred clean-up released the file
+                * handles to its contents.
+                */
+               lttng_trace_chunk_fini(chunk);
+
                element = container_of(chunk, typeof(*element), chunk);
                if (element->registry) {
                        rcu_read_lock();
@@ -1952,7 +2018,20 @@ LTTNG_HIDDEN
 struct lttng_trace_chunk *
 lttng_trace_chunk_registry_publish_chunk(
                struct lttng_trace_chunk_registry *registry,
-               uint64_t session_id, struct lttng_trace_chunk *chunk)
+               uint64_t session_id,
+               struct lttng_trace_chunk *chunk)
+{
+       bool unused;
+
+       return lttng_trace_chunk_registry_publish_chunk_published(
+                       registry, session_id, chunk, &unused);
+}
+
+struct lttng_trace_chunk *
+lttng_trace_chunk_registry_publish_chunk_published(
+               struct lttng_trace_chunk_registry *registry,
+               uint64_t session_id, struct lttng_trace_chunk *chunk,
+               bool *previously_published)
 {
        struct lttng_trace_chunk_registry_element *element;
        unsigned long element_hash;
@@ -1987,6 +2066,7 @@ lttng_trace_chunk_registry_publish_chunk(
                        element->registry = registry;
                        /* Acquire a reference for the caller. */
                        if (lttng_trace_chunk_get(&element->chunk)) {
+                               *previously_published = false;
                                break;
                        } else {
                                /*
@@ -2013,6 +2093,7 @@ lttng_trace_chunk_registry_publish_chunk(
                if (lttng_trace_chunk_get(published_chunk)) {
                        lttng_trace_chunk_put(&element->chunk);
                        element = published_element;
+                       *previously_published = true;
                        break;
                }
                /*