diff options
author | Ravishankar N <ravishankar@redhat.com> | 2014-01-24 17:27:38 +0530 |
---|---|---|
committer | Vijay Bellur <vbellur@redhat.com> | 2014-02-03 05:45:03 -0800 |
commit | 8cf2a36dad6c8bac7cd3a05455fd555544ebb457 (patch) | |
tree | 796f4489363839009b18140d8d83a603316ffba7 /xlators/mgmt/glusterd | |
parent | 3395b98a341228e89cf0a88e0d86c90117dbb285 (diff) |
glusterd: Fix race in pid file update
This patch only removes lines of code. For personal gratification, giving a
detailed explanation of what the problem was.
When glusterd spawns the local brick process, say when a reboot of the node
occurs,the glusterd_brick_start() and subsequently the
glusterd_volume_start_glusterfs() function gets called twice; from
glusterd_spawn_daemons() and glusterd_do_volume_quorum_action() respectively.
This causes a race, best described by a pseudo-code of current behaviour.
glusterd_volume_start_glusterfs()
{
if(!brick process running) {
step-a) reap pid file( i.e. unlink it)
step-b) fork a brick process which creates and locks pid file and
binds the process to a socket.
}
}
Time Event
---- -----
T1 Call-1 arrives, completes step-a, starts step-b
T2 Call-2 arrives, enters step-a as Call-1's forked child is not
yet running.
T3 Call-1's forked child is alive, creates pidfile and locks it,binds
its address to a socket.
T4 Call-2 performs step-a; i.e.unlinks the pid file created by Call-1 !!
(files can still be stil be unlinked despite a lockf on it)
T5 Call-2 does step-b, and the forked child process creates a *new* pid file
with it's pid and locks this file.
T6 But Call-2's brick process is not able to bind to socket as it
is already in use (courtesy T3) and hence exits (so no locks anymore on the pidfile).
Result:
- Pid file now contains PID of an extinct brick process.
- `gluster volume status` shows this PID value. It also notices that there is no
lock held on pid file by the currently running brick process (created by Call-1)
and hence shows N/A for the online status.
Also, as a result of events at T4, "ls -l /proc/<brick process PID>/fd/pidfile"
shows up as deleted.
Fix:
1.Do not unlink pid file. i.e. avoid step-a. Now at T5,Call-2's brick process
cannot obtain lock on pid file (Call-1's process still holds it) and exits.
2. Unrelated, but remove lock-unlock sequence in glusterfs_pidfile_setup()
which does not seem to be doing anything useful.
Change-Id: I18d3e577bb56f68d85d3e0d0bfd268e50ed4f038
BUG: 1035586
Signed-off-by: Ravishankar N <ravishankar@redhat.com>
Reviewed-on: http://review.gluster.org/6786
Reviewed-by: Krishnan Parthasarathi <kparthas@redhat.com>
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Vijay Bellur <vbellur@redhat.com>
Diffstat (limited to 'xlators/mgmt/glusterd')
-rw-r--r-- | xlators/mgmt/glusterd/src/glusterd-utils.c | 12 |
1 files changed, 0 insertions, 12 deletions
diff --git a/xlators/mgmt/glusterd/src/glusterd-utils.c b/xlators/mgmt/glusterd/src/glusterd-utils.c index 7c58e51addb..9fb2ab31e04 100644 --- a/xlators/mgmt/glusterd/src/glusterd-utils.c +++ b/xlators/mgmt/glusterd/src/glusterd-utils.c @@ -1307,16 +1307,6 @@ out: return ret; } -/* Caller should ensure that brick process is not running*/ -static void -_reap_brick_process (char *pidfile, char *brickpath) -{ - unlink (pidfile); - /* Brick process is not running and pmap may have an entry for it.*/ - pmap_registry_remove (THIS, 0, brickpath, - GF_PMAP_PORT_BRICKSERVER, NULL); -} - static int _mk_rundir_p (glusterd_volinfo_t *volinfo) { @@ -1371,8 +1361,6 @@ glusterd_volume_start_glusterfs (glusterd_volinfo_t *volinfo, if (gf_is_service_running (pidfile, NULL)) goto connect; - _reap_brick_process (pidfile, brickinfo->path); - port = brickinfo->port; if (!port) port = pmap_registry_alloc (THIS); |