On 03/19/15 at 11:08P, hiren panchasara wrote: > On 03/16/15 at 06:06P, hiren panchasara wrote: > > On 03/16/15 at 03:39P, Gleb Smirnoff wrote: > > > On Wed, Mar 11, 2015 at 02:36:07PM -0700, hiren panchasara wrote: > > > h> On 02/13/15 at 11:19P, Simon J. Gerraty wrote: > > > h> > Author: sjg > > > h> > Date: Fri Feb 13 23:19:35 2015 > > > h> > New Revision: 278729 > > > h> > URL: https://svnweb.freebsd.org/changeset/base/278729 > > > h> > > > > h> > Log: > > > h> > sbspace: size of bleft, mleft must match sockbuf fields to avoid > > > h> > overflow on amd64 > > > h> > > > > h> > Submitted by: anshu...@juniper.net > > > h> > Obtained from: Juniper Networks > > > h> > > > h> Talking to sjg on -arch to MFC this. If he cannot get around doing > > > that, > > > h> I'll do it tomorrow. > > > h> > > > h> Letting people know here to see if there are any objections. > > > > > > Would that fix the bug we've been discussing? > > > > Unsure as I am not sure what caused the issue I saw. > > > > For those who do not know the details, we recently saw a userland > > process stuck spinning at 100% around sbcut_internal(). Inside > > sbflush_internal(), the sb_cc was grown to be about 4G. And before > > passing it to sbcut_internal(), we cast it from uint to int which > > would make that valud -ve. > > > > Gleb pointed out to me that sbspace() is supposed to check/stop sb_cc > > from growing that large. > > > > Now, I am not sure if we'd ever run into this situation again but > > current fix is a great catch anyways. > > > > I still have 2 questions around what we saw. It'd be great if someone can > > clarify them for my understanding: > > > > 1) Even if we get into such a scenario that we were in, following would > > help by not looping endlessly. > > > > --- uipc_sockbuf.c.0 2015-03-11 15:49:52.000000000 -0700 > > +++ uipc_sockbuf.c 2015-03-11 15:51:48.000000000 -0700 > > @@ -877,6 +877,9 @@ > > { > > struct mbuf *m, *n, *next, *mfree; > > > > + if (len < 0) > > + panic("%s: len is %d and it is supposed to be +ve", > > + __func__, len); > > + > > next = (m = sb->sb_mb) ? m->m_nextpkt : 0; > > mfree = NULL > > > > 2) We need 1) because we are casting a uint to int which _may_ rander a > > value -ve. Is there a way we can avoid the casting? > > It'd be useful if someone with knowledge in this area can weigh in.
Ran into this again today. While the real question of how sb_ccc grew this large is still unsolved, any objection to adding this patch to avoid a hang and panic instead? Cheers, Hiren
pgpxuhhAOgLkb.pgp
Description: PGP signature