hi.
On Thu, Oct 19, 2017 at 10:06:56AM +1100, Neil Brown wrote:
On Wed, Oct 18 2017, NeilBrown wrote:
On Wed, Oct 18 2017, Coly Li wrote:
We have a bug report about NULL dereference in md raid10 code. The
operations are,
- Create a raid10 device
- add a spare disk
- disconnect one of the online disks from the raid10 device
- wait to the recovery happens on the spare disk
- remove the spare disk which is recovering
And sometimes a kernel oops of NULL dereference in md raid10 module can
be observed, and crash tool reports the following information:
(gdb) list *(raid10d+0xad6)
0x5de6 is in raid10d (../drivers/md/raid10.c:2300).
2295 * the latter is free to free wbio2.
2296 */
2297 if (wbio2 && !wbio2->bi_end_io)
2298 wbio2 = NULL;
2299 if (wbio->bi_end_io) {
2300 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2301 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2302 generic_make_request(wbio);
2303 }
2304 if (wbio2) {
At line 2300, conf->mirrors[d].rdev is NULL, this causes a NULL pointer
dereference to conf->mirrors[d].rdev->nr_pending. After reading previous
raid10 patches, I find conf->mirrors[d].rdev can be NULL at any point
unless,
- a counted reference is held
- ->reconfig_mutex is held, or
- rcu_read_lock() is held
We have the same bug report about NULL dereference in md raid10 code.
operations are:
- create raid10, including 4 disk and 2 spare disk
mdadm -C /dev/md1 -l 10 -n 4 /dev/sda /dev/sdb /dev/sdc
/dev/sdd -x 2 /dev/sde /dev/sdf -R
- block and offline a disk in raid10
- hot remove a disk by
mdadm --manage /dev/md1 -r sdX
- add disk
mdadm --manage /dev/md1 -r sdX
(gdb) l *(0xffffffff815daafd)
0xffffffff815daafd is in rdev_clear_badblocks (drivers/md/md.c:8688).
8683 int is_new)
8684 {
8685 if (is_new)
8686 s += rdev->new_data_offset;
8687 else
8688 s += rdev->data_offset; <====================
8689 return md_clear_badblocks(&rdev->badblocks,
8690 s, sectors);
8691 }
8692 EXPORT_SYMBOL_GPL(rdev_clear_badblocks);
Call Trace and crash tools show it is triggered as following:
raid10d() => handle_write_completed() => rdev_clear_badblocks()
There is one other condition: MD_RECOVERY_RUNNING is set (without
MD_RECOVERY_DONE).
mirrors[d].rdev is only set to NULL by ->hot_remove_disk() which is only
called from remove_and_add_spares(), and that is never called while
resync/recovery/reshape is happening.
So there is no need for RCU protection here.
Only ... remove_and_add_spares() *can* sometimes be called during
resync -
Commit: 8430e7e0af9a ("md: disconnect device from personality before trying to remove it.")
added a called to remove_and_add_spares() when "remove" is written to
the device/state attribute. That was wrong.
This:
} else if (cmd_match(buf, "remove")) {
- if (rdev->mddev->pers) {
+ if (rdev->mddev->pers &&
+ !test_bit(MD_RECOVERY_RECOVER, &mddev->recovery)) {
clear_bit(Blocked, &rdev->flags);
remove_and_add_spares(rdev->mddev, rdev);
should fix it.
Actually, that doesn't even compile :-(
I think this is a better fix.
Thanks,
NeilBrown
From: NeilBrown <neilb@xxxxxxxx>
Date: Thu, 19 Oct 2017 09:58:19 +1100
Subject: [PATCH] md: only allow remove_and_add_spares() when no sync_thread running.
The locking protocols in md assume that a device will
never be removed from an array during resync/recovery/reshape.
When that isn't happening, rcu or reconfig_mutex is needed
to protect an rdev pointer while taking a refcount. When
it is happening, that protection isn't needed.
Unfortuantely there is a case were remove_and_add_spares() is
called when recovery might be happening: when "remove" is
written to the device/state sysfs attribute.
This call is an optimization and not essential so it doesn't
matter if it fails.
So change remove_and_add_spares() to abort early if
resync/recovery/reshape is happening.
As this can result in a NULL dereference, the fix is suitable
for -stable.
Cc: Tomasz Majchrzak <tomasz.majchrzak@xxxxxxxxx>
Fixes: 8430e7e0af9a ("md: disconnect device from personality before trying to remove it.")
Cc: stable@xxxxxxxxxxxxxx (v4.8+)
Signed-off-by: NeilBrown <neilb@xxxxxxxx>
---
drivers/md/md.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/drivers/md/md.c b/drivers/md/md.c
index b3192943de7d..01eac0dbca98 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c
@@ -8563,6 +8563,10 @@ static int remove_and_add_spares(struct mddev *mddev,
int removed = 0;
bool remove_some = false;
Sorry, I missed this patch.
I'd really appreciate if you can add the locking protocol into comments here,
digging changelog is much painful.
> + if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
+ /* Mustn't remove devices when resync thread is running */
+ return 0;
+
This will bypass hotadd too, is it what we want?
After applying this patch, there is no more Oops, but the number of
disks will become less and less, which is not expected.
I think it's caused by bypassing of hotadd. Can we just add spare disk
without removing disk?
Any suggestions for this fix or any progress on this patch?
Thanks a lot,
Yufen Yu
Thanks,
Shaohua
rdev_for_each(rdev, mddev) {
if ((this == NULL || rdev == this) &&
rdev->raid_disk >= 0 &&
--
2.14.0.rc0.dirty
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
.
.
.
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
