Hi Linus,
Please consider pulling the following additional patch for the GFS2 file system.
Regards,
Bob Peterson
----------------------------------------------------------------
The following changes since commit ae50dfd61665086e617cc9e554a1285d52765670:
Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net (2017-03-14 21:31:23 -0700)
are available in the git repository at:
git://git.kernel.org/pub/scm/linux/kernel/git/gfs2/linux-gfs2.git tags/gfs2-4.11-rc3.fixes
for you to fetch changes up to 28ea06c46fbcab63fd9a55531387b7928a18a590:
gfs2: Avoid alignment hole in struct lm_lockname (2017-03-15 10:06:07 -0400)
----------------------------------------------------------------
This is an emergency patch for 4.11-rc3
The GFS2 developers uncovered a really nasty problem that
can lead to random corruption and kernel panic, much like
the last one. Andreas Gruenbacher wrote this simple one-line
patch to fix the problem:
28ea06c gfs2: Avoid alignment hole in struct lm_lockname
----------------------------------------------------------------
Andreas Gruenbacher (1):
gfs2: Avoid alignment hole in struct lm_lockname
fs/gfs2/incore.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
On Wed, Mar 15, 2017 at 7:32 AM, Bob Peterson <[email protected]> wrote:
>
> Andreas Gruenbacher (1):
> gfs2: Avoid alignment hole in struct lm_lockname
So I've pulled this because I think it fixes a real bug, but honestly
I think it's the wrong fix.
Marking that lm_lockname structure "packed, aligned(4)" means that the
compiler will now think that the 64-bit fields in it may be unaligned
- including on architectures where that can be very expensive and the
compiler now might generate stupid unaligned instruction sequences to
load those values.
So the *correct* fix, I think, would have been:
- add a comment about not having holes in the struct due to the hashing
- sort the fields by size (so "ln_number" first, then "ln_sbd", then "ln_type")
- use offsetofend(struct lm_lockname, ln_type) instead of sizeof() when hashing
which avoids the "possibly generate garbage code" issue due to the
quick-and-dirty one-liner approach.
Hmm?
Linus
----- Original Message -----
| On Wed, Mar 15, 2017 at 7:32 AM, Bob Peterson <[email protected]> wrote:
| >
| > Andreas Gruenbacher (1):
| > gfs2: Avoid alignment hole in struct lm_lockname
|
| So I've pulled this because I think it fixes a real bug, but honestly
| I think it's the wrong fix.
|
| Marking that lm_lockname structure "packed, aligned(4)" means that the
| compiler will now think that the 64-bit fields in it may be unaligned
| - including on architectures where that can be very expensive and the
| compiler now might generate stupid unaligned instruction sequences to
| load those values.
|
| So the *correct* fix, I think, would have been:
|
| - add a comment about not having holes in the struct due to the hashing
|
| - sort the fields by size (so "ln_number" first, then "ln_sbd", then
| "ln_type")
|
| - use offsetofend(struct lm_lockname, ln_type) instead of sizeof() when
| hashing
|
| which avoids the "possibly generate garbage code" issue due to the
| quick-and-dirty one-liner approach.
|
| Hmm?
|
| Linus
Hi Linus,
Thanks. Yes, good ideas.
I see your point and I'll see if we can get that fixed up for the next merge window.
Bob Peterson