Due to a recent update on Javascript code a full page refresh on your browser might be needed.
Bug 1408101 - Fix potential socket_poller thread deadlock and resource leak
Summary: Fix potential socket_poller thread deadlock and resource leak
Status: CLOSED DUPLICATE of bug 1561332
Alias: None
Product: GlusterFS
Classification: Community
Component: rpc
Version: mainline
Hardware: Unspecified
OS: Unspecified
Target Milestone: ---
Assignee: Raghavendra G
QA Contact:
Depends On:
Blocks: 1408104
TreeView+ depends on / blocked
Reported: 2016-12-22 06:32 UTC by Kaushal
Modified: 2019-11-25 12:51 UTC (History)
3 users (show)

Fixed In Version:
Doc Type: If docs needed, set a value
Doc Text:
Clone Of:
: 1408104 (view as bug list)
Last Closed: 2019-11-25 12:50:52 UTC
Regression: ---
Mount Type: ---
Documentation: ---
Verified Versions:

Attachments (Terms of Use)

Description Kaushal 2016-12-22 06:32:00 UTC
The fix for bug #1404181 [1], has a potential deadlock and resource leak of the socket_poller thread.

A disconnect caused by a PARENT_DOWN event during a fuse graph switch, can lead to the socket_poller thread being deadlocked. The deadlock doesn't affect the fuse client as no new fops are sent on the old graph.

In addition to the above, the race in gfapi solved by [1] can also occur in other codepaths, and need to be solved.

Quoting Raghavendra G's comment from the review,
- The race addressed by this patch (race b/w socket_disconnect cleaning up resources in priv and socket_poller using the same and resulting in undefined behaviour - crash/corruption etc) can potentially happen irrespective of the codepaths socket_disconnect is invoked from (like glusterd, client_portmap_cbk, handling of PARENT_DOWN, changelog etc). Note the usage of word "potential" here and I am not saying that this race happens in existing code. However, I would like this issue gets fixed for these potential cases too.
- If there are fops in progress at the time of graph switch, sending PARENT_DOWN event on the currently active (soon to be old) graph is deferred till all the fops are complete (though new graph becomes active and new I/O is redirected to that graph). So, PARENT_DOWN event can be sent after processing last response (to fop). This means PARENT_DOWN can be sent in thread executing socket_poller itself. Since PARENT_DOWN triggers a disconnect and disconnect waits for socket_poller to complete, we've a deadlock. Specifically the deadlock is: socket_poller -> notify-msg-received -> fuse processes fop response -> fuse sends PARENT_DOWN -> rpc-clnt calls rpc_clnt_disable -> socket_disconnect -> wait till socket_poller to complete before returning from socket_disconnect. Luckily we've have a socket_poller thread for each transport and threads that deadlock are the threads belonging to transports from older graphs on which no I/O happening. So, at worst this will be a case of resource leakage (threads/sockets etc) of old graph.


[1] https://review.gluster.org/16141

Comment 1 Amar Tumballi 2019-05-09 20:31:56 UTC
Is this still relevant?

Comment 2 Xavi Hernandez 2019-11-20 10:23:21 UTC
Does this problem still exist in current versions ?

Otherwise I'll close the bug.

Comment 3 Raghavendra G 2019-11-25 12:50:52 UTC
[1] removes the socket_pollerr thread itself. So, these issues are no longer relevant.

[1] https://review.gluster.org/c/glusterfs/+/19308/

*** This bug has been marked as a duplicate of bug 1561332 ***

Note You need to log in before you can comment on or make changes to this bug.