在 2025/8/31 2:10, Kenta Akagi 写道:
On 2025/08/30 17:48, Li Nan wrote:
在 2025/8/29 20:21, Kenta Akagi 写道:
On 2025/08/29 11:54, Li Nan wrote:
在 2025/8/29 0:32, Kenta Akagi 写道:
This commit ensures that an MD_FAILFAST IO failure does not put
the array into a broken state.
When failfast is enabled on rdev in RAID1 or RAID10,
the array may be flagged MD_BROKEN in the following cases.
- If MD_FAILFAST IOs to multiple rdevs fail simultaneously
- If an MD_FAILFAST metadata write to the 'last' rdev fails
[...]
diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c
index 408c26398321..8a61fd93b3ff 100644
--- a/drivers/md/raid1.c
+++ b/drivers/md/raid1.c
@@ -470,6 +470,7 @@ static void raid1_end_write_request(struct bio *bio)
(bio->bi_opf & MD_FAILFAST) &&
/* We never try FailFast to WriteMostly devices */
!test_bit(WriteMostly, &rdev->flags)) {
+ set_bit(FailfastIOFailure, &rdev->flags);
md_error(r1_bio->mddev, rdev);
}
@@ -1746,8 +1747,12 @@ static void raid1_status(struct seq_file *seq, struct mddev *mddev)
* - recovery is interrupted.
* - &mddev->degraded is bumped.
*
- * @rdev is marked as &Faulty excluding case when array is failed and
- * &mddev->fail_last_dev is off.
+ * If @rdev has &FailfastIOFailure and it is the 'last' rdev,
+ * then @mddev and @rdev will not be marked as failed.
+ *
+ * @rdev is marked as &Faulty excluding any cases:
+ * - when @mddev is failed and &mddev->fail_last_dev is off
+ * - when @rdev is last device and &FailfastIOFailure flag is set
*/
static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
{
@@ -1758,6 +1763,13 @@ static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
if (test_bit(In_sync, &rdev->flags) &&
(conf->raid_disks - mddev->degraded) == 1) {
+ if (test_and_clear_bit(FailfastIOFailure, &rdev->flags)) {
+ spin_unlock_irqrestore(&conf->device_lock, flags);
+ pr_warn_ratelimited("md/raid1:%s: Failfast IO failure on %pg, "
+ "last device but ignoring it\n",
+ mdname(mddev), rdev->bdev);
+ return;
+ }
set_bit(MD_BROKEN, &mddev->flags);
if (!mddev->fail_last_dev) {
@@ -2148,6 +2160,7 @@ static int fix_sync_read_error(struct r1bio *r1_bio)
if (test_bit(FailFast, &rdev->flags)) {
/* Don't try recovering from here - just fail it
* ... unless it is the last working device of course */
+ set_bit(FailfastIOFailure, &rdev->flags);
md_error(mddev, rdev);
if (test_bit(Faulty, &rdev->flags))
/* Don't try to read from here, but make sure
@@ -2652,6 +2665,7 @@ static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
fix_read_error(conf, r1_bio);
unfreeze_array(conf);
} else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
+ set_bit(FailfastIOFailure, &rdev->flags);
md_error(mddev, rdev);
} else {
r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
index b60c30bfb6c7..530ad6503189 100644
--- a/drivers/md/raid10.c
+++ b/drivers/md/raid10.c
@@ -488,6 +488,7 @@ static void raid10_end_write_request(struct bio *bio)
dec_rdev = 0;
if (test_bit(FailFast, &rdev->flags) &&
(bio->bi_opf & MD_FAILFAST)) {
+ set_bit(FailfastIOFailure, &rdev->flags);
md_error(rdev->mddev, rdev);
}
Thank you for the patch. There may be an issue with 'test_and_clear'.
If two write IO go to the same rdev, MD_BROKEN may be set as below:
IO1 IO2
set FailfastIOFailure
set FailfastIOFailure
md_error
raid1_error
test_and_clear FailfastIOFailur
md_error
raid1_error
//FailfastIOFailur is cleared
set MD_BROKEN
Maybe we should check whether FailfastIOFailure is already set before
setting it. It also needs to be considered in metadata writes.
Thank you for reviewing.
I agree, this seems to be as you described.
So, should it be implemented as follows?
bool old=false;
do{
spin_lock_irqsave(&conf->device_lock, flags);
old = test_and_set_bit(FailfastIOFailure, &rdev->flags);
spin_unlock_irqrestore(&conf->device_lock, flags);
}while(old);
However, since I am concerned about potential deadlocks,
so I am considering two alternative approaches:
* Add an atomic_t counter to md_rdev to track failfast IO failures.
This may set MD_BROKEN at a slightly incorrect timing, but mixing
error handling of Failfast and non-Failfast IOs appears to be rare.
In any case, the final outcome would be the same, i.e. the array
ends up with MD_BROKEN. Therefore, I think this should not cause
issues. I think the same applies to test_and_set_bit.
IO1 IO2 IO3
FailfastIOFailure Normal IOFailure FailfastIOFailure
atomic_inc
md_error atomic_inc
raid1_error
atomic_dec //2to1
md_error
raid1_error md_error
atomic_dec //1to0 raid1_error
atomic_dec //0
set MD_BROKEN
* Alternatively, create a separate error handler,
e.g. md_error_failfast(), that clearly does not fail the array.
This approach would require somewhat larger changes and may not
be very elegant, but it seems to be a reliable way to ensure
MD_BROKEN is never set at the wrong timing.
Which of these three approaches would you consider preferable?
I would appreciate your feedback.
For metadata writes, I plan to clear_bit MD_FAILFAST_SUPPORTED
when the array is degraded.
Thanks,
Akagi
I took a closer look at the FailFast code and found a few issues, using
RAID1 as an example:
For normal read/write IO, FailFast is only triggered when there is another
disk is available, as seen in read_balance() and raid1_write_request().
In raid1_error(), MD_BROKEN is set only when no other disks are available.
Hi,
Agree, I think so too.
So, the FailFast for normal read/write is not triggered in the scenario you
described in cover-letter.
This corresponds to the case described in the commit message of PATCH v3 1/3.
"Normally, MD_FAILFAST IOs are not issued to the 'last' rdev, so this is
an edge case; however, it can occur if rdevs in a non-degraded
array share the same path and that path is lost, or if a metadata
write is triggered in a degraded array and fails due to failfast."
To describe it in more detail, the flow is as follows:
Prerequisites:
- Both rdevs are in-sync
- Both rdevs have in-flight MD_FAILFAST bios
- Both rdevs depend on the same lower-level path
(e.g., nvme-tcp over a single Ethernet interface)
Sequence:
- A bios with REQ_FAILFAST_DEV fails (e.g., due to a temporary network outage),
in the case of nvme-tcp:
- The Ethernet connection is lost on the node where md is running over 5 seconds
- Then the connection is restored. Idk the details of nvme-tcp implementation,
but it seems that failfast IOs finish only after the connection is back.
- All failfast bios fail, raid1_end_write_request is called.
- md_error() marks one rdev Faulty; the other rdev becomes the 'last' rdev.
- md_error() on the last rdev sets MD_BROKEN on the array - fail_last_dev=1 is unlikely.
- The write is retried via handle_write_finished -> narrow_write_error, usually succeeding.
- MD_BROKEN remains set, leaving the array in a state where no further writes can occur.
Thanks for your patient explanation. I understand. Maybe we need a separate
error-handling path for failfast. How about adding an extra parameter to
md_error()?
Kuai, do you have any suggestions?
Normal writes may call md_error() in narrow_write_error. Normal reads do
not execute md_error() on the last disk.
Perhaps you should get more information to confirm how MD_BROKEN is set in
normal read/write IO.
Should I add the above sequence of events to the cover letter, or commit message?
I think we should mention this in the commit message.
Thanks,
Akagi
--
Thanks,
Nan
.
--
Thanks,
Nan
