(got no reply on this when I originally send it on 20061031, so resending
now that a bit of time has passed. The patch still applies cleanly to
Linus' git tree as of today.)
The Coverity checker spotted a potential problem in XFS.
The problem is that if, in xfs_mount(), this code triggers:
...
if (!mp->m_logdev_targp)
goto error0;
...
Then we'll end up calling xfs_unmountfs_close() with a NULL
'mp->m_logdev_targp'.
This in turn will result in a call to xfs_free_buftarg() with its 'btp'
argument == NULL. xfs_free_buftarg() dereferences 'btp' leading to
a NULL pointer dereference and crash.
I think this can happen, since the fatal call to xfs_free_buftarg()
happens when 'm_logdev_targp != m_ddev_targp' and due to a check of
'm_ddev_targp' against NULL in xfs_mount() (and subsequent return if it is
NULL) the two will never both be NULL when we hit the error0 label from
the two lines cited above.
Comments welcome (please keep me on Cc: on replies).
Here's a proposed patch to fix this by testing 'btp' against NULL in
xfs_free_buftarg().
Signed-off-by: Jesper Juhl <[email protected]>
---
fs/xfs/linux-2.6/xfs_buf.c | 3 +++
1 files changed, 3 insertions(+), 0 deletions(-)
diff --git a/fs/xfs/linux-2.6/xfs_buf.c b/fs/xfs/linux-2.6/xfs_buf.c
index db5f5a3..6ef1860 100644
--- a/fs/xfs/linux-2.6/xfs_buf.c
+++ b/fs/xfs/linux-2.6/xfs_buf.c
@@ -1450,6 +1450,9 @@ xfs_free_buftarg(
xfs_buftarg_t *btp,
int external)
{
+ if (unlikely(!btp))
+ return;
+
xfs_flush_buftarg(btp, 1);
if (external)
xfs_blkdev_put(btp->bt_bdev);
On 16/11/06, Shailendra Tripathi <[email protected]> wrote:
> Hey Jesper,
> Rather, it can be done as below. Nothing to say that
> your code wouldn't work. Just that catch it early, so that potential
> function call overhead to call xfs_free_buftarg can be avoided.
>
Hi Shailendra,
The reason I want to fix it in the freeing function is that many other
functions in the kernel that free resources are safe to call with NULL
pointers and this would make xfs_free_buftarg() follow that
convention. This would perhaps also allow for some cleanups in other
places that call the function since then there's no longer a need for
explicit NULL checks any more (haven't checked if there's anything to
gain there though).
I don't think the function call overhead matters much since this is in
a case of a failed mount, so it should happen very rarely.
> void
> xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
> {
> if (mp->m_logdev_targp && (mp->m_logdev_targp != mp->m_ddev_targp))
> xfs_free_buftarg(mp->m_logdev_targp, 1);
> if (mp->m_rtdev_targp)
> xfs_free_buftarg(mp->m_rtdev_targp, 1);
> xfs_free_buftarg(mp->m_ddev_targp, 0);
> }
>
--
Jesper Juhl <[email protected]>
Don't top-post http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please http://www.expita.com/nomime.html
Hey Jesper,
Rather, it can be done as below. Nothing to say that
your code wouldn't work. Just that catch it early, so that potential
function call overhead to call xfs_free_buftarg can be avoided.
void
xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
{
if (mp->m_logdev_targp && (mp->m_logdev_targp != mp->m_ddev_targp))
xfs_free_buftarg(mp->m_logdev_targp, 1);
if (mp->m_rtdev_targp)
xfs_free_buftarg(mp->m_rtdev_targp, 1);
xfs_free_buftarg(mp->m_ddev_targp, 0);
}
Jesper Juhl wrote:
> (got no reply on this when I originally send it on 20061031, so resending
> now that a bit of time has passed. The patch still applies cleanly to
> Linus' git tree as of today.)
>
>
> The Coverity checker spotted a potential problem in XFS.
>
> The problem is that if, in xfs_mount(), this code triggers:
>
> ...
> if (!mp->m_logdev_targp)
> goto error0;
> ...
>
> Then we'll end up calling xfs_unmountfs_close() with a NULL
> 'mp->m_logdev_targp'.
> This in turn will result in a call to xfs_free_buftarg() with its 'btp'
> argument == NULL. xfs_free_buftarg() dereferences 'btp' leading to
> a NULL pointer dereference and crash.
>
> I think this can happen, since the fatal call to xfs_free_buftarg()
> happens when 'm_logdev_targp != m_ddev_targp' and due to a check of
> 'm_ddev_targp' against NULL in xfs_mount() (and subsequent return if it is
> NULL) the two will never both be NULL when we hit the error0 label from
> the two lines cited above.
>
> Comments welcome (please keep me on Cc: on replies).
>
> Here's a proposed patch to fix this by testing 'btp' against NULL in
> xfs_free_buftarg().
>
>
> Signed-off-by: Jesper Juhl <[email protected]>
> ---
>
> fs/xfs/linux-2.6/xfs_buf.c | 3 +++
> 1 files changed, 3 insertions(+), 0 deletions(-)
>
> diff --git a/fs/xfs/linux-2.6/xfs_buf.c b/fs/xfs/linux-2.6/xfs_buf.c
> index db5f5a3..6ef1860 100644
> --- a/fs/xfs/linux-2.6/xfs_buf.c
> +++ b/fs/xfs/linux-2.6/xfs_buf.c
> @@ -1450,6 +1450,9 @@ xfs_free_buftarg(
> xfs_buftarg_t *btp,
> int external)
> {
> + if (unlikely(!btp))
> + return;
> +
> xfs_flush_buftarg(btp, 1);
> if (external)
> xfs_blkdev_put(btp->bt_bdev);
>
>
>
>
Jesper Juhl wrote:
> The reason I want to fix it in the freeing function is that many other
> functions in the kernel that free resources are safe to call with NULL
> pointers and this would make xfs_free_buftarg() follow that
> convention. This would perhaps also allow for some cleanups in other
> places that call the function since then there's no longer a need for
> explicit NULL checks any more (haven't checked if there's anything to
> gain there though).
> I don't think the function call overhead matters much since this is in
> a case of a failed mount, so it should happen very rarely.
>
I agree with you. However, cleanup functions should(/must?) check for
NULL etc and
in this case it is already doing so for other cases. So, perhaps not
required. Just a different viewpoint.
Your choice.
>> void
>> xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
>> {
>> if (mp->m_logdev_targp && (mp->m_logdev_targp !=
>> mp->m_ddev_targp))
>> xfs_free_buftarg(mp->m_logdev_targp, 1);
>> if (mp->m_rtdev_targp)
>> xfs_free_buftarg(mp->m_rtdev_targp, 1);
>> xfs_free_buftarg(mp->m_ddev_targp, 0);
>> }
>>
>
>
On Thu, Nov 16, 2006 at 10:18:26PM +0100, Jesper Juhl wrote:
> (got no reply on this when I originally send it on 20061031, so resending
> now that a bit of time has passed. The patch still applies cleanly to
> Linus' git tree as of today.)
>
>
> The Coverity checker spotted a potential problem in XFS.
>
> The problem is that if, in xfs_mount(), this code triggers:
>
> ...
> if (!mp->m_logdev_targp)
> goto error0;
> ...
>
> Then we'll end up calling xfs_unmountfs_close() with a NULL
> 'mp->m_logdev_targp'.
> This in turn will result in a call to xfs_free_buftarg() with its 'btp'
> argument == NULL. xfs_free_buftarg() dereferences 'btp' leading to
> a NULL pointer dereference and crash.
Interesting that coverity found that, but failed to find the other
leaks in that function from exactly the same code and error
case.....
> I think this can happen, since the fatal call to xfs_free_buftarg()
> happens when 'm_logdev_targp != m_ddev_targp' and due to a check of
> 'm_ddev_targp' against NULL in xfs_mount() (and subsequent return if it is
> NULL) the two will never both be NULL when we hit the error0 label from
> the two lines cited above.
>
> Comments welcome (please keep me on Cc: on replies).
>
> Here's a proposed patch to fix this by testing 'btp' against NULL in
> xfs_free_buftarg().
Not the right fix - we should only be trying to free valid
buftargs, which means xfs_unmountfs_close() is the correct
place to fix this....
e.g:
- if (mp->m_logdev_targp != mp->m_ddev_targp)
+ if (mp->m_logdev_targp && (mp->m_logdev_targp != mp->m_ddev_targp))
As to the afore-mentioned leaks, if we fail to allocate a realtime
buftarg, then we will leak a reference to both the rtdev and logdev,
and if we fail to allocate an external log buftarg we'll leak a
reference to the logdev. i.e., we fail to do one or both of:
xfs_blkdev_put(logdev);
xfs_blkdev_put(rtdev);
To remove the bdev references we may have gained earlier. Normally,
these references are released by xfs_free_buftarg(), but because we
failed to allocate the buftarg, we can't drop the references via
that method....
Cheers,
Dave.
--
Dave Chinner
Principal Engineer
SGI Australian Software Group
On 16/11/06, David Chinner <[email protected]> wrote:
> On Thu, Nov 16, 2006 at 10:18:26PM +0100, Jesper Juhl wrote:
> > (got no reply on this when I originally send it on 20061031, so resending
> > now that a bit of time has passed. The patch still applies cleanly to
> > Linus' git tree as of today.)
> >
> >
> > The Coverity checker spotted a potential problem in XFS.
> >
> > The problem is that if, in xfs_mount(), this code triggers:
> >
> > ...
> > if (!mp->m_logdev_targp)
> > goto error0;
> > ...
> >
> > Then we'll end up calling xfs_unmountfs_close() with a NULL
> > 'mp->m_logdev_targp'.
> > This in turn will result in a call to xfs_free_buftarg() with its 'btp'
> > argument == NULL. xfs_free_buftarg() dereferences 'btp' leading to
> > a NULL pointer dereference and crash.
>
> Interesting that coverity found that, but failed to find the other
> leaks in that function from exactly the same code and error
> case.....
>
> > I think this can happen, since the fatal call to xfs_free_buftarg()
> > happens when 'm_logdev_targp != m_ddev_targp' and due to a check of
> > 'm_ddev_targp' against NULL in xfs_mount() (and subsequent return if it is
> > NULL) the two will never both be NULL when we hit the error0 label from
> > the two lines cited above.
> >
> > Comments welcome (please keep me on Cc: on replies).
> >
> > Here's a proposed patch to fix this by testing 'btp' against NULL in
> > xfs_free_buftarg().
>
> Not the right fix - we should only be trying to free valid
> buftargs, which means xfs_unmountfs_close() is the correct
> place to fix this....
>
Ok.
> e.g:
>
> - if (mp->m_logdev_targp != mp->m_ddev_targp)
> + if (mp->m_logdev_targp && (mp->m_logdev_targp != mp->m_ddev_targp))
>
> As to the afore-mentioned leaks, if we fail to allocate a realtime
> buftarg, then we will leak a reference to both the rtdev and logdev,
> and if we fail to allocate an external log buftarg we'll leak a
> reference to the logdev. i.e., we fail to do one or both of:
>
> xfs_blkdev_put(logdev);
> xfs_blkdev_put(rtdev);
>
> To remove the bdev references we may have gained earlier. Normally,
> these references are released by xfs_free_buftarg(), but because we
> failed to allocate the buftarg, we can't drop the references via
> that method....
>
Thanks a lot for commenting Dave.
I don't have time to look more at this tonight or tomorrow, but I'll
try to cook up a proper fix covering all cases during the weekend.
--
Jesper Juhl <[email protected]>
Don't top-post http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please http://www.expita.com/nomime.html
On 16/11/06, Jesper Juhl <[email protected]> wrote:
> On 16/11/06, David Chinner <[email protected]> wrote:
> > On Thu, Nov 16, 2006 at 10:18:26PM +0100, Jesper Juhl wrote:
> > > (got no reply on this when I originally send it on 20061031, so resending
> > > now that a bit of time has passed. The patch still applies cleanly to
> > > Linus' git tree as of today.)
> > >
> > >
> > > The Coverity checker spotted a potential problem in XFS.
> > >
> > > The problem is that if, in xfs_mount(), this code triggers:
> > >
> > > ...
> > > if (!mp->m_logdev_targp)
> > > goto error0;
> > > ...
> > >
> > > Then we'll end up calling xfs_unmountfs_close() with a NULL
> > > 'mp->m_logdev_targp'.
> > > This in turn will result in a call to xfs_free_buftarg() with its 'btp'
> > > argument == NULL. xfs_free_buftarg() dereferences 'btp' leading to
> > > a NULL pointer dereference and crash.
> >
> > Interesting that coverity found that, but failed to find the other
> > leaks in that function from exactly the same code and error
> > case.....
> >
> > > I think this can happen, since the fatal call to xfs_free_buftarg()
> > > happens when 'm_logdev_targp != m_ddev_targp' and due to a check of
> > > 'm_ddev_targp' against NULL in xfs_mount() (and subsequent return if it is
> > > NULL) the two will never both be NULL when we hit the error0 label from
> > > the two lines cited above.
> > >
> > > Comments welcome (please keep me on Cc: on replies).
> > >
> > > Here's a proposed patch to fix this by testing 'btp' against NULL in
> > > xfs_free_buftarg().
> >
> > Not the right fix - we should only be trying to free valid
> > buftargs, which means xfs_unmountfs_close() is the correct
> > place to fix this....
> >
>
> Ok.
>
> > e.g:
> >
> > - if (mp->m_logdev_targp != mp->m_ddev_targp)
> > + if (mp->m_logdev_targp && (mp->m_logdev_targp != mp->m_ddev_targp))
> >
> > As to the afore-mentioned leaks, if we fail to allocate a realtime
> > buftarg, then we will leak a reference to both the rtdev and logdev,
> > and if we fail to allocate an external log buftarg we'll leak a
> > reference to the logdev. i.e., we fail to do one or both of:
> >
> > xfs_blkdev_put(logdev);
> > xfs_blkdev_put(rtdev);
> >
> > To remove the bdev references we may have gained earlier. Normally,
> > these references are released by xfs_free_buftarg(), but because we
> > failed to allocate the buftarg, we can't drop the references via
> > that method....
> >
>
How about something like the attached patch ?
(sorry about the attachment, but I can't inline the patch from my
current location - a whitespace damaged version is below though for
easy review)
Signed-off-by: Jesper Juhl <[email protected]>
---
fs/xfs/xfs_mount.c | 2 +-
fs/xfs/xfs_vfsops.c | 10 ++++++++--
2 files changed, 9 insertions(+), 3 deletions(-)
diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c
index 9dfae18..3497128 100644
--- a/fs/xfs/xfs_mount.c
+++ b/fs/xfs/xfs_mount.c
@@ -1135,7 +1135,7 @@ #endif
void
xfs_unmountfs_close(xfs_mount_t *mp, struct cred *cr)
{
- if (mp->m_logdev_targp != mp->m_ddev_targp)
+ if (mp->m_logdev_targp && mp->m_logdev_targp != mp->m_ddev_targp)
xfs_free_buftarg(mp->m_logdev_targp, 1);
if (mp->m_rtdev_targp)
xfs_free_buftarg(mp->m_rtdev_targp, 1);
diff --git a/fs/xfs/xfs_vfsops.c b/fs/xfs/xfs_vfsops.c
index 62336a4..6d7e8f1 100644
--- a/fs/xfs/xfs_vfsops.c
+++ b/fs/xfs/xfs_vfsops.c
@@ -473,13 +473,19 @@ xfs_mount(
}
if (rtdev) {
mp->m_rtdev_targp = xfs_alloc_buftarg(rtdev, 1);
- if (!mp->m_rtdev_targp)
+ if (!mp->m_rtdev_targp) {
+ xfs_blkdev_put(logdev);
+ xfs_blkdev_put(rtdev);
goto error0;
+ }
}
mp->m_logdev_targp = (logdev && logdev != ddev) ?
xfs_alloc_buftarg(logdev, 1) : mp->m_ddev_targp;
- if (!mp->m_logdev_targp)
+ if (!mp->m_logdev_targp) {
+ xfs_blkdev_put(logdev);
+ xfs_blkdev_put(rtdev);
goto error0;
+ }
/*
* Setup flags based on mount(2) options and then the superblock
--
Jesper Juhl <[email protected]>
Don't top-post http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please http://www.expita.com/nomime.html