Panic in the udp_input() under heavy load

[Robert N. M. Watson]

On 30 Dec 2011, at 19:48, Maxim Sobolev wrote:

> On 11/24/2011 11:24 PM, Robert N. M. Watson wrote:
>>>> There was recently a commit to fix a race condition in 10-CURRENT which
>>>> >>  I think is not slated to be merged for 9.0. You might check the commit
>>>> >>  logs there and see if that fixes the problems you have -- if so, we
>>>> >>  might want to reconsider the plan not to merge for 9.0.
>>>> >>
>>>> >>  (It relates to a race condition on closing sockets..)
>>> >
>>> >  Thank you for the tip. I will give it a try and see what happens. So far, after installing that trap we have not seen any panics yet. I have not checked logs yet if my trap actually has catch anything or not.
>> Do we know if this fix has been merged to stable/9 and releng/9.0? Given multiple reports of instability without it, I think we would be well-served to merge it at this point.
> 
> Hey Robert, sorry to bother you again, but can you at least point me towards specific SVN revisions that I need to get merged? I tried to google it and also browsed my svn history for the last 4 months using keyword "sockets", but nothing came up. This fix is critical for any system that does lot of fault-ctitical networking, and FreeBSD has always been solid in this regard. It saved us at least 10-15 crashes across 5 machines in the last month.

Hi Maxim:

Looking back at a recent post from you, it appears that you are on 8.x and not 9.x, as I had assumed form your original e-mail. The patch I was referring to in 9-CURRENT has long since been merged for 9.0 and will appear in that release. However, it does not apply to 8.x, as the bug it fixed was introduced during the 9.x cycle. We'll need to do a from-scratch diagnosis here rather than assume it's the same problem. Could I ask you to follow up to this post with version information, stack traces from relevant threads, etc? I am not aware of any other reports of UDP-related crashes along the lines of what you've described in 8.x, so it may be being triggered by some unusual aspect of your workload (or just bad luck). Sorry that there's no instant "merge a patch" fix for this one.

Thanks,

Robert