tree c6dfc8ba5d04fda4c5cb15cdebd5425ab674f8df
parent 69ac59647e66c1b53fb98fe8b6d0f2099cffad60
author Jason Baron <jbaron@redhat.com> 1126296117 -0700
committer Linus Torvalds <torvalds@g5.osdl.org> 1126299451 -0700

[PATCH] pty_chars_in_buffer oops fix

The idea of this patch is to lock both sides of a ptmx/pty pair during line
discipline changing.  This is needed to ensure that say a poll on one side of
the pty doesn't occur while the line discipline is actively being changed.
This resulted in an oops reported on lkml, see:

	http://marc.theaimsgroup.com/?l=linux-kernel&m=111342171410005&w=2

A 'hacky' approach was previously implmemented which served to eliminate the
poll vs.  line discipline changing race.  However, this patch takes a more
general approach to the issue.  The patch only adds locking on a less often
used path, the line-discipline changing path, as opposed to locking the
ptmx/pty pair on read/write/poll paths.

The patch below, takes both ldisc locks in either order b/c the locks are both
taken under the same spinlock().  I thought about locking the ptmx/pty
separately, such as master always first but that introduces a 3 way deadlock.
For example, process 1 does a blocking read on the slave side.  Then, process
2 does an ldisc change on the slave side, which acquires the master ldisc lock
but not the slave's.  Finally, process 3 does a write which blocks on the
process 2's ldisc reference.

This patch does introduce some changes in semantics.  For example, a line
discipline change on side 'a' of a ptmx/pty pair, will now wait for a
read/write to complete on the other side, or side 'b'.  The current behavior
is to simply wait for any read/writes on only side 'a', not both sides 'a' and
'b'.  I think this behavior makes sense, but I wanted to point it out.

I've tested the patch with a bunch of read/write/poll while changing the line
discipline out from underneath.

This patch obviates the need for the above "hide the problem" patch.

Signed-off-by: Jason Baron <jbaron@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>