lkml.org 
[lkml]   [2013]   [Dec]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Found it! (was Re: [3.10] Oopses in kmem_cache_allocate() via prepare_creds())

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Sat, Nov 30, 2013 at 1:08 PM, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
> >
> > I still don't see what could be wrong with the pipe_inode_info thing,
> > but the fact that it's been so consistent in your traces does make me
> > suspect it really is *that* particular slab.
>
> I think I finally found it.
>
> I've spent waaayy too much time looking at and thinking about that
> code without seeing anything wrong, but this morning I woke up and
> thought to myself "What if.."
>
> And looking at the code again, I went "BINGO".
>
> All our reference counting etc seems right, but we have one very
> subtle bug: on the freeing path, we have a pattern like this:
>
> spin_lock(&inode->i_lock);
> if (!--pipe->files) {
> inode->i_pipe = NULL;
> kill = 1;
> }
> spin_unlock(&inode->i_lock);
> __pipe_unlock(pipe);
> if (kill)
> free_pipe_info(pipe);
>
> which on the face of it is trying to be very careful in not accessing
> the pipe-info after it is released by having that "kill" flag, and
> doing the release last.
>
> And it's complete garbage.
>
> Why?
>
> Because the thread that decrements "pipe->files" *without* releasing
> it, will very much access it after it has been dropped: that
> "__pipe_unlock(pipe)" happens *after* we've decremented the pipe
> reference count and dropped the inode lock. So another CPU can come
> in and free the structure concurrently with that
> __pipe_unlock(pipe).
>
> This happens in two places, and we don't actually need or want the
> pipe lock for the pipe->files accesses (since pipe->files is
> protected by inode->i_lock, not the pipe lock), so the solution is
> to just do the __pipe_unlock() before the whole dance about the
> pipe->files reference count.
>
> Patch appended. And no wonder nobody has ever seen it, because the
> race is unlikely as hell to ever happen. Simon, I assume it will be
> another few months before we can say "yeah, that fixed it", but I
> really think this is it. It explains all the symptoms, including
> "DEBUG_PAGEALLOC didn't catch it" (because the access happens just
> as it is released, and DEBUG_PAGEALLOC takes too long to actually
> free unmap the page etc).
>
> Linus
> fs/pipe.c | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/fs/pipe.c b/fs/pipe.c
> index d2c45e14e6d8..18f1a4b2dbbc 100644
> --- a/fs/pipe.c
> +++ b/fs/pipe.c
> @@ -743,13 +743,14 @@ pipe_release(struct inode *inode, struct file *file)
> kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
> kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
> }
> + __pipe_unlock(pipe);
> +
> spin_lock(&inode->i_lock);
> if (!--pipe->files) {
> inode->i_pipe = NULL;
> kill = 1;
> }
> spin_unlock(&inode->i_lock);
> - __pipe_unlock(pipe);

I'm wondering why pipe-mutex was introduced. It was done fairly
recently, with no justification given:

From 72b0d9aacb89f3759931ec440e1b535671145bb4 Mon Sep 17 00:00:00 2001
From: Al Viro <viro@zeniv.linux.org.uk>
Date: Thu, 21 Mar 2013 02:32:24 -0400
Subject: [PATCH] pipe: don't use ->i_mutex

now it can be done - put mutex into pipe_inode_info, use it instead
of ->i_mutex

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
---
fs/ocfs2/file.c | 6 ++----
fs/pipe.c | 5 +++--
include/linux/pipe_fs_i.h | 2 ++
3 files changed, 7 insertions(+), 6 deletions(-)

It's not like there should be many (any?) VFS operations where a pipe
is used via i_mutex and pipe->mutex in parallel, which would improve
scalability - so I don't see the scalability advantage. (But I might
be missing something)

Barring such kind of workload the extra mutex just adds extra
micro-costs because now two locks have to be taken on
creation/destruction, plus it adds extra complexity and races.

So unless I'm missing something obvious, another good fix would be to
just revert pipe->mutex and rely on i_mutex as before?

Thanks,

Ingo


\
 
 \ /
  Last update: 2013-12-02 17:41    [W:0.094 / U:2.376 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site