Date: Mon, 20 Jun 2016 13:48:19 +0300 From: Konstantin Belousov <kostikbel@gmail.com> To: Julien Charbon <jch@freebsd.org> Cc: Gleb Smirnoff <glebius@FreeBSD.org>, rrs@FreeBSD.org, current@FreeBSD.org, hselasky@FreeBSD.org, net@FreeBSD.org Subject: Re: panic with tcp timers Message-ID: <20160620104819.GV38613@kib.kiev.ua> In-Reply-To: <1d18d0e2-3e42-cb26-928c-2989d0751884@freebsd.org> References: <20160617045319.GE1076@FreeBSD.org> <1f28844b-b4ea-b544-3892-811f2be327b9@freebsd.org> <20160620073917.GI1076@FreeBSD.org> <1d18d0e2-3e42-cb26-928c-2989d0751884@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Jun 20, 2016 at 11:55:55AM +0200, Julien Charbon wrote: > > Hi, > > On 6/20/16 9:39 AM, Gleb Smirnoff wrote: > > On Fri, Jun 17, 2016 at 11:27:39AM +0200, Julien Charbon wrote: > > J> > Comparing stable/10 and head, I see two changes that could > > J> > affect that: > > J> > > > J> > - callout_async_drain > > J> > - switch to READ lock for inp info in tcp timers > > J> > > > J> > That's why you are in To, Julien and Hans :) > > J> > > > J> > We continue investigating, and I will keep you updated. > > J> > However, any help is welcome. I can share cores. > > > > Now, spending some time with cores and adding a bunch of > > extra CTRs, I have a sequence of events that lead to the > > panic. In short, the bug is in the callout system. It seems > > to be not relevant to the callout_async_drain, at least for > > now. The transition to READ lock unmasked the problem, that's > > why NetflixBSD 10 doesn't panic. > > > > The panic requires heavy contention on the TCP info lock. > > > > [CPU 1] the callout fires, tcp_timer_keep entered > > [CPU 1] blocks on INP_INFO_RLOCK(&V_tcbinfo); > > [CPU 2] schedules the callout > > [CPU 2] tcp_discardcb called > > [CPU 2] callout successfully canceled > > [CPU 2] tcpcb freed > > [CPU 1] unblocks... panic > > > > When the lock was WLOCK, all contenders were resumed in a > > sequence they came to the lock. Now, that they are readers, > > once the lock is released, readers are resumed in a "random" > > order, and this allows tcp_discardcb to go before the old > > running callout, and this unmasks the panic. > > Highly interesting. I should be able to reproduce that (will be useful > for testing the corresponding fix). > > Fix proposal: If callout_async_drain() returns 0 (fail) (instead of 1 > (success) here) when the callout cancellation is a success _but_ the > callout is current running, that should fix it. > > For the history: It comes back to my old callout question: > > Does _callout_stop_safe() is allowed to return 1 (success) even if the > callout is still currently running; a.k.a. it is not because you > successfully cancelled a callout that the callout is not currently running. > > We did propose a patch to make _callout_stop_safe() returns 0 (fail) > when the callout is currently running: > > callout_stop() should return 0 when the callout is currently being > serviced and indeed unstoppable > https://reviews.freebsd.org/differential/changeset/?ref=62513&whitespace=ignore-most > > But this change impacted too many old code paths and was interesting > only for TCP timers and thus was abandoned. Look at callout_stop CS_MIGRBLOCK flag and the fix in sleepq_check_timeout(). Or, at least, do not allow this use of callout_stop() to rot again, after previous dozen regressions and fixes there.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20160620104819.GV38613>