On Wed, 2015-04-08 at 15:08 +0200, Johannes Berg wrote:
> On Wed, 2015-04-08 at 13:03 +0100, David Woodhouse wrote:
> 
> > I'm not sure if this is entirely fixed. In Fedora 22 (4.0.0-rc5-git4)
> > I'm occasionally seeing glibc deadlock in __check_pf() on a netlink
> > recvmsg(), here:
> > https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/unix/sysv/linux/check_pf.c;h=162606d7;hb=glibc-2.21#l166
> > 
> > As I understand it, this shouldn't happen. Even if messages are
> > dropped (which surely shouldn't happen as often as I'm seeing this),
> > glibc should get ENOBUFS from the recvmsg() call.
> > 
> > https://bugzilla.redhat.com/show_bug.cgi?id=1209433
> > 
> > I haven't bisected and proved that it *was* this commit which
> > introduced the problem, as it only happens after a day or two of
> > running Evolution and I haven't managed to trigger it more reliably.
> 
> I don't see the connection to this change.
> 
> The issue with my patch was that some code for NLM_F_DUMP would have
> this pattern:
> 
>  int fill_function(...)
>  {
>     ...
>     return nlmsg_end(...);
>  }
> 
>  loop (...) {
>    if (fill_function() <= 0)
>      break; /* continue in next dump */
>  }
> 
> and that all had to be converted to be just "< 0" now.
> 
> Additionally, the failure mode of this was the process running out of
> memory due to receiving the same results over and over again - does that
> happen for you? It seems it was stuck in recvmsg(), but that may just be
> a side effect of happening to interrupt at that point?
> 

I don't think the problem was introduced by your change. At 
https://github.com/nahi/httpclient/issues/232 it seems to have been
observed even in November of last year.

I've added some debugging, and it seems that when it deadlocks, glibc
doesn't get *any* response to its RTM_GETADDR request. I know we'd get
ENOBUFS is a *response* was dropped... but what about when the request
itself is dropped? Does userspace get any hint of that? Is this purely
a glibc bug, for assuming its request got delivered and unconditionally
waiting for a response?

I don't know why it suddenly started happening to me in the 4.0 kernel
when I'd never seen it before, but it's still happening. I've put a
poll() in the glibc code (referenced above), and made it fail after a 5
-second timeout. That will at least prevent me from throwing my
computer out the window for the time being...

-- 
dwmw2