Thomas Jackson created TS-5052:
----------------------------------
Summary: Segfault in HostDB sync if something fails while not
holding the parent continuation mutex
Key: TS-5052
URL: https://issues.apache.org/jira/browse/TS-5052
Project: Traffic Server
Issue Type: Bug
Reporter: Thomas Jackson
What we noticed was the following in traffic.out:
{code}
Server {0x2af761e0d700} WARNING: <P_RefCountCache.h:510 (initialize_storage)>
Unable to create temporary file /var/trafficserver/
host.db.syncing, unable to persist hostdb: -13 error:Permission denied
traffic_server: Segmentation fault (Address not mapped to object
[0x28])traffic_server - STACK TRACE:
{code}
Which lead me to dig into it-- and it turns out the issue is related to changes
after the HostDB rewrite to move syncing outside of the main NET threads.
Before all calls to this syncer where done in a single net thread wherever it
was initially scheduled. Now we bounce between ET_TASK threads and ET_NET
threads (to avoid switching, lock contention, etc.)-- but the error handlers
weren't updated to handle this situation.
So to fix this, I've created a "set_error" and "return_error" method to the
RefCountCacheSerializer which will take this into consideration-- specifically
that it will immediately return the error if scheduled in the calling thread--
otherwise it'll reschedule onto that thread *then* return the error.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
