svn commit: r278729 - head/sys/sys

[Gleb Smirnoff]

On Fri, Feb 17, 2017 at 11:06:21AM -0800, hiren panchasara wrote:
h> On 03/19/15 at 11:08P, hiren panchasara wrote:
h> > On 03/16/15 at 06:06P, hiren panchasara wrote:
h> > > On 03/16/15 at 03:39P, Gleb Smirnoff wrote:
h> > > > On Wed, Mar 11, 2015 at 02:36:07PM -0700, hiren panchasara wrote:
h> > > > h> On 02/13/15 at 11:19P, Simon J. Gerraty wrote:
h> > > > h> > Author: sjg
h> > > > h> > Date: Fri Feb 13 23:19:35 2015
h> > > > h> > New Revision: 278729
h> > > > h> > URL: https://svnweb.freebsd.org/changeset/base/278729
h> > > > h> > 
h> > > > h> > Log:
h> > > > h> >   sbspace: size of bleft, mleft must match sockbuf fields to avoid
h> > > > h> >   overflow on amd64
h> > > > h> >   
h> > > > h> >   Submitted by:	anshukla at juniper.net
h> > > > h> >   Obtained from:	Juniper Networks
h> > > > h> 
h> > > > h> Talking to sjg on -arch to MFC this. If he cannot get around doing that,
h> > > > h> I'll do it tomorrow. 
h> > > > h> 
h> > > > h> Letting people know here to see if there are any objections.
h> > > > 
h> > > > Would that fix the bug we've been discussing?
h> > > 
h> > > Unsure as I am not sure what caused the issue I saw.
h> > > 
h> > > For those who do not know the details, we recently saw a userland
h> > > process stuck spinning at 100% around sbcut_internal(). Inside
h> > > sbflush_internal(), the sb_cc was grown to be about 4G. And before
h> > > passing it to sbcut_internal(), we cast it from uint to int which
h> > > would make that valud -ve.
h> > > 
h> > > Gleb pointed out to me that sbspace() is supposed to check/stop sb_cc
h> > > from growing that large.
h> > > 
h> > > Now, I am not sure if we'd ever run into this situation again but
h> > > current fix is a great catch anyways.
h> > > 
h> > > I still have 2 questions around what we saw. It'd be great if someone can
h> > > clarify them for my understanding:
h> > > 
h> > > 1) Even if we get into such a scenario that we were in, following would
h> > > help by not looping endlessly.
h> > > 
h> > > --- uipc_sockbuf.c.0    2015-03-11 15:49:52.000000000 -0700
h> > > +++ uipc_sockbuf.c      2015-03-11 15:51:48.000000000 -0700
h> > > @@ -877,6 +877,9 @@
h> > > {
h> > >          struct mbuf *m, *n, *next, *mfree;
h> > > 
h> > > +      if (len < 0)
h> > > +               panic("%s: len is %d and it is supposed to be +ve",
h> > > +		    __func__, len);
h> > > +
h> > >          next = (m = sb->sb_mb) ? m->m_nextpkt : 0;
h> > >          mfree = NULL
h> > > 
h> > > 2) We need 1) because we are casting a uint to int which _may_ rander a
h> > > value -ve. Is there a way we can avoid the casting?
h> > 
h> > It'd be useful if someone with knowledge in this area can weigh in.
h> 
h> Ran into this again today. While the real question of how sb_ccc grew
h> this large is still unsolved, any objection to adding this patch to
h> avoid a hang and panic instead?

I am all for adding KASSERT now and then fixing the bug.

-- 
Totus tuus, Glebius.