[
https://issues.apache.org/jira/browse/TS-440?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12910369#action_12910369
]
Bryan Call commented on TS-440:
-------------------------------
John moved the enable lists to be atomic, so there is no need to wrap locks
around them:
ASLLM(UnixNetVConnection, NetState, read, enable_link) read_enable_list;
ASLLM(UnixNetVConnection, NetState, write, enable_link) write_enable_list;
> Traffic Server threads will spin under heavy load
> -------------------------------------------------
>
> Key: TS-440
> URL: https://issues.apache.org/jira/browse/TS-440
> Project: Traffic Server
> Issue Type: Bug
> Components: Network
> Reporter: Bryan Call
> Assignee: Bryan Call
> Fix For: 2.1.3
>
>
> This has been a problem for a long time. I created a fix for the problem
> yesterday and tested it overnight.
> The problem is when we remove from the enabled lists in
> close_UnixNetVConnection(). We lock the lists when we add to them, but not
> when we remove the connection from them in close_UnixNetVConnection(). This
> can corrupt the data structure.
> Because the data structure is corrupted it can fail the empty() test in
> NetHandler::mainNetEvent() and poll_timeout (really epoll timeout) will set
> to 0 and the thread will spin.
> The read enabled list (Queue<>) will normally have a NULL head, but a
> non-NULL tail and the empty() test on the Queue<> checks for NULL head and
> NULL tail to be considered empty.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
