* 2. Once the locks are obtained, signal parent process via fd.
* 3. Invoke registered callback routine with locking status.
* 4. If the child process cannot get locks within certain time,
- * diagnose using /proc/locks and log warning message
+ * execute an external script to debug.
*
* ctdb_lock_record() - get a lock on a record
* ctdb_lock_db() - get a lock on a DB
* auto_mark - whether to mark/unmark DBs in before/after callback
*/
-/* FIXME: Add a tunable max_lock_processes_per_db */
-#define MAX_LOCK_PROCESSES_PER_DB (100)
-
enum lock_type {
LOCK_RECORD,
LOCK_DB,
"lock_record",
"lock_db",
"lock_alldb_prio",
- "lock_db",
+ "lock_alldb",
};
struct lock_request;
TDB_DATA key;
uint32_t priority;
bool auto_mark;
- struct lock_request *req_queue;
+ struct lock_request *request;
pid_t child;
int fd[2];
struct tevent_fd *tfd;
struct tevent_timer *ttimer;
- pid_t block_child;
- int block_fd[2];
struct timeval start_time;
+ uint32_t key_hash;
+ bool can_schedule;
};
/* lock_request is the client specific part for a lock request */
struct lock_request {
- struct lock_request *next, *prev;
struct lock_context *lctx;
void (*callback)(void *, bool);
void *private_data;
int ctdb_lockall_unmark_prio(struct ctdb_context *ctdb, uint32_t priority)
{
/*
- * This function is only used by the main dameon during recovery.
+ * This function is only used by the main daemon during recovery.
* At this stage, the databases have already been locked, by a
* dedicated child process. The freeze_mode variable is used to track
* whether the actual locks are held by the child process or not.
{
uint32_t priority;
- for (priority=NUM_DB_PRIORITIES; priority>=0; priority--) {
+ for (priority=NUM_DB_PRIORITIES; priority>0; priority--) {
if (ctdb_db_iterator(ctdb, priority, db_lock_unmark_handler, NULL) != 0) {
return -1;
}
if (lock_ctx->child > 0) {
ctdb_kill(lock_ctx->ctdb, lock_ctx->child, SIGKILL);
DLIST_REMOVE(lock_ctx->ctdb->lock_current, lock_ctx);
- lock_ctx->ctdb->lock_num_current--;
+ if (lock_ctx->ctdb_db) {
+ lock_ctx->ctdb_db->lock_num_current--;
+ }
CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_current);
- if (lock_ctx->type == LOCK_RECORD || lock_ctx->type == LOCK_DB) {
+ if (lock_ctx->ctdb_db) {
CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
}
} else {
DLIST_REMOVE(lock_ctx->ctdb->lock_pending, lock_ctx);
- lock_ctx->ctdb->lock_num_pending--;
CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
- if (lock_ctx->type == LOCK_RECORD || lock_ctx->type == LOCK_DB) {
+ if (lock_ctx->ctdb_db) {
CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
}
}
*/
static int ctdb_lock_request_destructor(struct lock_request *lock_request)
{
- DLIST_REMOVE(lock_request->lctx->req_queue, lock_request);
+ lock_request->lctx->request = NULL;
return 0;
}
-
void ctdb_lock_free_request_context(struct lock_request *lock_req)
{
struct lock_context *lock_ctx;
*/
static void process_callbacks(struct lock_context *lock_ctx, bool locked)
{
- struct lock_request *request, *next;
+ struct lock_request *request;
if (lock_ctx->auto_mark && locked) {
switch (lock_ctx->type) {
}
}
- /* Iterate through all callbacks */
- request = lock_ctx->req_queue;
- while (request) {
- if (lock_ctx->auto_mark) {
- /* Reset the destructor, so request is not removed from the list */
- talloc_set_destructor(request, NULL);
- }
-
- /* In case, callback frees the request, store next */
- next = request->next;
- request->callback(request->private_data, locked);
- request = next;
+ request = lock_ctx->request;
+ if (lock_ctx->auto_mark) {
+ /* Reset the destructor, so request is not removed from the list */
+ talloc_set_destructor(request, NULL);
}
+ request->callback(request->private_data, locked);
if (lock_ctx->auto_mark && locked) {
switch (lock_ctx->type) {
}
/* Read the status from the child process */
- read(lock_ctx->fd[0], &c, 1);
- locked = (c == 0 ? true : false);
+ if (read(lock_ctx->fd[0], &c, 1) != 1) {
+ locked = false;
+ } else {
+ locked = (c == 0 ? true : false);
+ }
/* Update statistics */
- CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_calls);
if (lock_ctx->ctdb_db) {
- CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_calls);
}
if (locked) {
if (lock_ctx->ctdb_db) {
- CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.buckets[id]);
CTDB_UPDATE_LATENCY(lock_ctx->ctdb, lock_ctx->ctdb_db,
lock_type_str[lock_ctx->type], locks.latency,
lock_ctx->start_time);
- CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
CTDB_UPDATE_DB_LATENCY(lock_ctx->ctdb_db, lock_type_str[lock_ctx->type], locks.latency, t);
CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.buckets[id]);
}
}
-static void ctdb_lock_find_blocker(struct lock_context *lock_ctx);
-
/*
* Callback routine when required locks are not obtained within timeout
* Called from parent context
struct timeval current_time,
void *private_data)
{
+ static const char * debug_locks = NULL;
struct lock_context *lock_ctx;
struct ctdb_context *ctdb;
+ pid_t pid;
lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
ctdb = lock_ctx->ctdb;
- /* fire a child process to find the blocking process */
- if (lock_ctx->block_child == -1) {
- ctdb_lock_find_blocker(lock_ctx);
+ if (lock_ctx->ctdb_db) {
+ DEBUG(DEBUG_WARNING,
+ ("Unable to get %s lock on database %s for %.0lf seconds\n",
+ (lock_ctx->type == LOCK_RECORD ? "RECORD" : "DB"),
+ lock_ctx->ctdb_db->db_name,
+ timeval_elapsed(&lock_ctx->start_time)));
+ } else {
+ DEBUG(DEBUG_WARNING,
+ ("Unable to get ALLDB locks for %.0lf seconds\n",
+ timeval_elapsed(&lock_ctx->start_time)));
+ }
+
+ /* Fire a child process to find the blocking process. */
+ if (debug_locks == NULL) {
+ debug_locks = getenv("CTDB_DEBUG_LOCKS");
+ if (debug_locks == NULL) {
+ debug_locks = talloc_asprintf(ctdb,
+ "%s/debug_locks.sh",
+ getenv("CTDB_BASE"));
+ }
+ }
+ if (debug_locks != NULL) {
+ pid = vfork();
+ if (pid == 0) {
+ execl(debug_locks, debug_locks, NULL);
+ _exit(0);
+ }
+ ctdb_track_child(ctdb, pid);
+ } else {
+ DEBUG(DEBUG_WARNING,
+ (__location__
+ " Unable to setup lock debugging - no memory?\n"));
}
/* reset the timeout timer */
CTDB_NO_MEMORY_VOID(ctdb, prog);
}
- if (ctdb->lock_num_current >= MAX_LOCK_PROCESSES_PER_DB) {
- return;
- }
-
if (ctdb->lock_pending == NULL) {
return;
}
/* Find a lock context with requests */
lock_ctx = ctdb->lock_pending;
while (lock_ctx != NULL) {
- if (! lock_ctx->req_queue) {
- next_ctx = lock_ctx->next;
- DEBUG(DEBUG_INFO, ("Removing lock context without lock requests\n"));
+ next_ctx = lock_ctx->next;
+ if (! lock_ctx->request) {
+ DEBUG(DEBUG_INFO, ("Removing lock context without lock request\n"));
DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
- ctdb->lock_num_pending--;
CTDB_DECREMENT_STAT(ctdb, locks.num_pending);
if (lock_ctx->ctdb_db) {
CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
}
talloc_free(lock_ctx);
- lock_ctx = next_ctx;
- continue;
} else {
- /* Found a lock context with lock requests */
- break;
+ if (lock_ctx->ctdb_db == NULL ||
+ lock_ctx->ctdb_db->lock_num_current < ctdb->tunable.lock_processes_per_db) {
+ /* Found a lock context with lock requests */
+ break;
+ }
}
+ lock_ctx = next_ctx;
}
if (lock_ctx == NULL) {
return;
}
- lock_ctx->child = ctdb_fork(ctdb);
+ lock_ctx->child = vfork();
if (lock_ctx->child == (pid_t)-1) {
DEBUG(DEBUG_ERR, ("Failed to create a child in ctdb_lock_schedule\n"));
}
/* Parent process */
+ ctdb_track_child(ctdb, lock_ctx->child);
close(lock_ctx->fd[1]);
talloc_set_destructor(lock_ctx, ctdb_lock_context_destructor);
/* Move the context from pending to current */
DLIST_REMOVE(ctdb->lock_pending, lock_ctx);
- ctdb->lock_num_pending--;
DLIST_ADD_END(ctdb->lock_current, lock_ctx, NULL);
- ctdb->lock_num_current++;
-}
-
-
-/*
- * Find the lock context of a given type
- */
-static struct lock_context *find_lock_context(struct lock_context *lock_list,
- struct ctdb_db_context *ctdb_db,
- TDB_DATA key,
- uint32_t priority,
- enum lock_type type)
-{
- struct lock_context *lock_ctx;
-
- /* Search active locks */
- for (lock_ctx=lock_list; lock_ctx; lock_ctx=lock_ctx->next) {
- if (lock_ctx->type != type) {
- continue;
- }
-
- switch (lock_ctx->type) {
- case LOCK_RECORD:
- if (ctdb_db == lock_ctx->ctdb_db &&
- key.dsize == lock_ctx->key.dsize &&
- memcmp(key.dptr, lock_ctx->key.dptr, key.dsize) == 0) {
- goto done;
- }
- break;
-
- case LOCK_DB:
- if (ctdb_db == lock_ctx->ctdb_db) {
- goto done;
- }
- break;
-
- case LOCK_ALLDB_PRIO:
- if (priority == lock_ctx->priority) {
- goto done;
- }
- break;
-
- case LOCK_ALLDB:
- goto done;
- break;
- }
+ CTDB_DECREMENT_STAT(lock_ctx->ctdb, locks.num_pending);
+ CTDB_INCREMENT_STAT(lock_ctx->ctdb, locks.num_current);
+ if (lock_ctx->ctdb_db) {
+ lock_ctx->ctdb_db->lock_num_current++;
+ CTDB_DECREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_pending);
+ CTDB_INCREMENT_DB_STAT(lock_ctx->ctdb_db, locks.num_current);
}
-
- /* Did not find the lock context we are searching for */
- lock_ctx = NULL;
-
-done:
- return lock_ctx;
-
}
enum lock_type type,
bool auto_mark)
{
- struct lock_context *lock_ctx;
+ struct lock_context *lock_ctx = NULL;
struct lock_request *request;
if (callback == NULL) {
return NULL;
}
- /* get a context for this key - search only the pending contexts,
- * current contexts might in the middle of processing callbacks */
- lock_ctx = find_lock_context(ctdb->lock_pending, ctdb_db, key, priority, type);
-
- /* No existing context, create one */
+ lock_ctx = talloc_zero(ctdb, struct lock_context);
if (lock_ctx == NULL) {
- lock_ctx = talloc_zero(ctdb, struct lock_context);
- if (lock_ctx == NULL) {
- DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
- return NULL;
- }
+ DEBUG(DEBUG_ERR, ("Failed to create a new lock context\n"));
+ return NULL;
+ }
- lock_ctx->type = type;
- lock_ctx->ctdb = ctdb;
- lock_ctx->ctdb_db = ctdb_db;
- lock_ctx->key.dsize = key.dsize;
- if (key.dsize > 0) {
- lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
- } else {
- lock_ctx->key.dptr = NULL;
+ if ((request = talloc_zero(lock_ctx, struct lock_request)) == NULL) {
+ talloc_free(lock_ctx);
+ return NULL;
+ }
+
+ lock_ctx->type = type;
+ lock_ctx->ctdb = ctdb;
+ lock_ctx->ctdb_db = ctdb_db;
+ lock_ctx->key.dsize = key.dsize;
+ if (key.dsize > 0) {
+ lock_ctx->key.dptr = talloc_memdup(lock_ctx, key.dptr, key.dsize);
+ if (lock_ctx->key.dptr == NULL) {
+ DEBUG(DEBUG_ERR, (__location__ "Memory allocation error\n"));
+ talloc_free(lock_ctx);
+ return NULL;
}
- lock_ctx->priority = priority;
- lock_ctx->auto_mark = auto_mark;
+ lock_ctx->key_hash = ctdb_hash(&key);
+ } else {
+ lock_ctx->key.dptr = NULL;
+ }
+ lock_ctx->priority = priority;
+ lock_ctx->auto_mark = auto_mark;
- lock_ctx->child = -1;
- lock_ctx->block_child = -1;
+ lock_ctx->request = request;
+ lock_ctx->child = -1;
+ /* Non-record locks are required by recovery and should be scheduled
+ * immediately, so keep them at the head of the pending queue.
+ */
+ if (lock_ctx->type == LOCK_RECORD) {
DLIST_ADD_END(ctdb->lock_pending, lock_ctx, NULL);
- ctdb->lock_num_pending++;
- CTDB_INCREMENT_STAT(ctdb, locks.num_pending);
- if (ctdb_db) {
- CTDB_INCREMENT_DB_STAT(ctdb_db, locks.num_pending);
- }
-
- /* Start the timer when we activate the context */
- lock_ctx->start_time = timeval_current();
+ } else {
+ DLIST_ADD(ctdb->lock_pending, lock_ctx);
}
-
- if ((request = talloc_zero(lock_ctx, struct lock_request)) == NULL) {
- return NULL;
+ CTDB_INCREMENT_STAT(ctdb, locks.num_pending);
+ if (ctdb_db) {
+ CTDB_INCREMENT_DB_STAT(ctdb_db, locks.num_pending);
}
+ /* Start the timer when we activate the context */
+ lock_ctx->start_time = timeval_current();
+
request->lctx = lock_ctx;
request->callback = callback;
request->private_data = private_data;
talloc_set_destructor(request, ctdb_lock_request_destructor);
- DLIST_ADD_END(lock_ctx->req_queue, request, NULL);
ctdb_lock_schedule(ctdb);
void (*callback)(void *, bool),
void *private_data)
{
- if (priority < 0 || priority > NUM_DB_PRIORITIES) {
+ if (priority < 1 || priority > NUM_DB_PRIORITIES) {
DEBUG(DEBUG_ERR, ("Invalid db priority: %u\n", priority));
return NULL;
}
auto_mark);
}
-/*
- * Callback routine to read the PID of blocking process from the child and log
- *
- */
-void ctdb_lock_blocked_handler(struct tevent_context *ev,
- struct tevent_fd *tfd,
- uint16_t flags,
- void *private_data)
-{
- struct lock_context *lock_ctx;
- pid_t blocker_pid = -1;
- char *process_name = NULL;
- const char *db_name = NULL;
- ino_t inode;
- struct ctdb_db_context *ctdb_db;
- int fd;
- struct stat stat_buf;
-
- lock_ctx = talloc_get_type_abort(private_data, struct lock_context);
-
- if (read(lock_ctx->block_fd[0], &blocker_pid, sizeof(blocker_pid)) != sizeof(blocker_pid)) {
- DEBUG(DEBUG_ERR, ("Error reading blocker process pid from child\n"));
- goto failed;
- }
- if (read(lock_ctx->block_fd[0], &inode, sizeof(inode)) != sizeof(inode)) {
- DEBUG(DEBUG_ERR, ("Error reading blocked inode from child\n"));
- goto failed;
- }
-
- if (blocker_pid < 0) {
- goto failed;
- }
-
- process_name = ctdb_get_process_name(blocker_pid);
-
- if (lock_ctx->type == LOCK_RECORD || lock_ctx->type == LOCK_DB) {
- db_name = lock_ctx->ctdb_db->ltdb->name;
- } else {
- for (ctdb_db = lock_ctx->ctdb->db_list; ctdb_db; ctdb_db = ctdb_db->next) {
- fd = tdb_fd(ctdb_db->ltdb->tdb);
- if (fstat(fd, &stat_buf) == 0) {
- if (stat_buf.st_ino == inode) {
- db_name = ctdb_db->ltdb->name;
- break;
- }
- }
- }
- }
-
- if (db_name) {
- DEBUG(DEBUG_WARNING,
- ("Process (pid=%d) blocked in locking\n", lock_ctx->child));
- DEBUG(DEBUG_WARNING,
- ("Process %s (pid=%d) locked database %s (inode %lu) for %.0lf seconds\n",
- (process_name ? process_name : "unknown"),
- blocker_pid, db_name, (unsigned long)inode,
- timeval_elapsed(&lock_ctx->start_time)));
- } else {
- DEBUG(DEBUG_WARNING,
- ("Process %s (pid=%d) locked database (inode %lu) for %.0lf seconds\n",
- (process_name ? process_name : "unknown"),
- blocker_pid, (unsigned long)inode,
- timeval_elapsed(&lock_ctx->start_time)));
- }
-
- /*
- * If ctdb is blocked by smbd for deadlock_interval, detect it as a deadlock
- * and kill smbd process.
- */
- if (lock_ctx->ctdb->tunable.deadlock_timeout > 0 &&
- timeval_elapsed(&lock_ctx->start_time) > lock_ctx->ctdb->tunable.deadlock_timeout &&
- process_name && strstr(process_name, "smbd")) {
- DEBUG(DEBUG_WARNING,
- ("Deadlock detected. Killing smbd process (pid=%d)", blocker_pid));
- kill(blocker_pid, SIGKILL);
- }
-
- free(process_name);
-
-failed:
- if (lock_ctx->block_child > 0) {
- ctdb_kill(lock_ctx->ctdb, lock_ctx->block_child, SIGKILL);
- }
- lock_ctx->block_child = -1;
- talloc_free(tfd);
-}
-
-
-/*
- * Find processes that holds lock we are interested in
- */
-void ctdb_lock_find_blocker(struct lock_context *lock_ctx)
-{
- struct tevent_fd *tfd;
- pid_t parent;
-
- if (pipe(lock_ctx->block_fd) < 0) {
- return;
- }
-
- parent = getpid();
-
- lock_ctx->block_child = ctdb_fork(lock_ctx->ctdb);
- if (lock_ctx->block_child == -1) {
- close(lock_ctx->block_fd[0]);
- close(lock_ctx->block_fd[1]);
- return;
- }
-
- /* Child process */
- if (lock_ctx->block_child == 0) {
- struct ctdb_lock_info reqlock;
- pid_t blocker_pid = -1;
- bool status;
-
- close(lock_ctx->block_fd[0]);
- if (ctdb_get_lock_info(lock_ctx->child, &reqlock)) {
- status = ctdb_get_blocker_pid(&reqlock, &blocker_pid);
- if (!status) {
- /* Could not find blocker pid */
- blocker_pid = -2;
- }
- }
- write(lock_ctx->block_fd[1], &blocker_pid, sizeof(blocker_pid));
- write(lock_ctx->block_fd[1], &reqlock.inode, sizeof(reqlock.inode));
-
- /* Hang around till parent dies */
- while (kill(parent, 0) == 0 || errno != ESRCH) {
- sleep(5);
- }
- _exit(0);
- }
-
- /* Parent process */
- close(lock_ctx->block_fd[1]);
- set_close_on_exec(lock_ctx->block_fd[0]);
-
- tfd = tevent_add_fd(lock_ctx->ctdb->ev,
- lock_ctx,
- lock_ctx->block_fd[0],
- EVENT_FD_READ,
- ctdb_lock_blocked_handler,
- (void *)lock_ctx);
- if (tfd == NULL) {
- ctdb_kill(lock_ctx->ctdb, lock_ctx->block_child, SIGKILL);
- close(lock_ctx->block_fd[0]);
- }
-}