|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] Lots of connections led oxenstored stuck
Hi Joe, On 27/08/2014 02:59, Joe Jin wrote: No this patch does not intend to fix NR_OPEN(this is from system side, need to unlimited before start the daemon) and select limitations. We met the bug not because of the limitations, the original issue is when more connect(i.e 2000) request coming at the same time, accept() failed because of open fds > SYSCONF.OPEN_MAX, this is as expected. The thing is when client exited, oxenstored should close the sockets as well, but during our test, it did not, and oxenstored keeping reported accept failed, any new request hang as well. So my changes let oxenstored check and delete closed fds, then oxenstored able to ack new requests. During our testing, when issue happened also xenstored log enabled, xenstored.log full of below error and xenstored.log keeping be rotated: [20140827T15:48:25.399Z|error|xenstored] caught exception Unix.Unix_error(15, "accept", "") Thanks for the explanation, I now understand the intention of your patch better. Still, I think the change can only mitigate/delay the issue to certain extend. If I'm wrong on this, please help me to understand it better. Here is an example walking through: Suppose your xenstored process have the default max open fd setting of 1024 and it currently has 1024 fds open already. Now 2000 more connections are coming: * Without your change, all the 2000 connection failed with EMFILE one by one (by Unix.accept in process_special_fds), you'll see 2000 lines of errors of Unix_error(15) in xenstored.log. But since there are always finite number of connections coming, eventually xenstored should recover and continue the normal logic after all these go through. And the recovery can happen even before that, if there are small time gaps among all your incoming connections (which basically give the process_connection_fd a chance to run). * With your change, xenstored will handle existing fds first. So if there is any of the fds being closed already, it basically gives its slot away to the coming connections, so the issue won't happen at the very beginning. But IIUIC, this all depends on how many persistent connections currently exist. Suppose among the 1024 fds, 800 of them are from relatively persistent connections (e.g. qemu connections for long running VMs) , then only the first small batch in your 2000 incoming connections will succeed and the rest will face exactly the same situation. So, IMHO, the change might help a great deal in a artificial testcase where a lot of temporary connections come and go, but less so in real life system where a large number of the connections are persistent. Please let me know if my understanding is correct. I have no objection to the change itself, mitigation is positive change too. Hence, Acked-by: Zheng Li <dev@xxxxxxxx> A better way is probably wrapping try-with logic on each of process_special_fds and process_connections_fds and process_domains instead of the single top-level try-with, so that the failure on one aspect will not block the others. We can leave this as future work. * When select is given 1024 fds (which can still happen even with your patch), the behavior is _undefined_. IIRC, some bits in the bitmap might be reused (wrongly), so that the output (fds reported as ready for read/write) might be wrong for some fds, so that the following read/write might be blocked on them. The max open fd setting of a process normally defaults to 1024 for a reason. If, as you said, your tests raise the ulimit setting beforehand but still use select internally, then there are potential issues here. * Also, we generally prefer to handle special fds first, as the eventchn fd represents all the domain connections.Remove closed firstly may reduce system resource usage? My initial guess was the original authors of oxenstored might be handling special fds first for performance consideration. But after more thinking about it, I don't think that could make a great difference. We can also separate the eventchn fd from the other two special sockets and give it higher priority if it turns out to be a problem afterwards. I previously mentioned I've got patches for these. I'm currently testing with 1,000 Windows 7 VMs on a single host (each consume at least 2 persistent xenstored socket connections). Besides the two limits just mentioned, I've also fixed several bugs and bottlenecks along the way. I'm going to upstream these patches very soon, just a bit clean up and documentation are needed. However if you (or anyone) need them urgently or eager to have a test, please send me an private email separately. I'm happy to send you the patch in its current form --- a single non-disaggregated patch for multiple issues, not very well commented, but should just work.Can you please send a copy of your patch? I'd like to test when connections more than @nfds of poll, what happened. Some logic in the patch changes the xenstored process's fd limit to the NR_OPEN of system max (usually 1024x1024), though you obviously won't be able to reach that exact number as other processes will have open fds too. But I guess that's a big enough number for real life cases. I'll send out the patch to those on the threads (except xen-devel) late today. Thanks, Zheng On 26/08/2014 09:15, Joe Jin wrote: _______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxx http://lists.xen.org/xen-devel
|
 |
Lists.xenproject.org is hosted with RackSpace, monitoring our |