sts-sponsors team mailing list archive

Thread
Date

[Bug 1847924] Re: Introduce broken state parsing to mdadm

To: sts-sponsors@xxxxxxxxxxxxxxxxxxx
From: "Guilherme G. Piccoli" <1847924@xxxxxxxxxxxxxxxxxx>
Date: Wed, 06 Nov 2019 19:34:39 -0000
Reply-to: Bug 1847924 <1847924@xxxxxxxxxxxxxxxxxx>
Sender: bounces@xxxxxxxxxxxxx

** Description changed:

[Impact]

* Currently, mounted raid0/md-linear arrays have no indication/warning
when one or more members are removed or suffer from some non-recoverable
error condition. The mdadm tool shows "clean" state regardless if a
member was removed.

* The patch proposed in this SRU addresses this issue by introducing a
new state "broken", which is analog to "clean" but indicates that array
is not in a good/correct state. The commit, available upstream as
43ebc910 ("mdadm: Introduce new array state 'broken' for raid0/linear")
[0], was extensively discussed and received a good amount of
reviews/analysis by both the current mdadm maintainer as well as an old
maintainer.

* One important note here is that this patch requires a counter-part in the kernel to be fully functional, which was SRUed in LP: #1847773.
It works fine/transparently without this kernel counter-part though.
+
+ * We had reports of users testing a scenario of failed raid0 arrays, and
+ getting 'clean' in mdadm proved to cause confusion and doesn't help on
+ noticing something went wrong with the arrays.
+
+ * The potential situation this patch (with its kernel counter-part)
+ addresses is: an user has raid0/linear array, and it's mounted. If one
+ member fails and gets removed (either physically, like a power or
+ firmware issue, or in software, like a driver-induced removal due to
+ detected failure), _without_ this patch (and its kernel counter-part)
+ there's nothing to let user know it failed, except filesystem errors in
+ dmesg. Also, non-direct writes to the filesystem will succeed, due to
+ how page-cache/writeback work; even a 'sync' command run will succeed.
+
+ * The case described in above bullet was tested and the writes to failed
+ devices succeeded - after a reboot, the files written were present in
+ the array, but corrupted. An user wouldn't noticed that unless if the
+ writes were directed or some checksum was performed in the files. With
+ this patch (and its kernel counter-part), the writes to such failed
+ raid0/linear array are fast-failed and the filesystem goes read-only
+ quickly.

[Test case]

* To test this patch, create a raid0 or linear md array on Linux using
mdadm, like: "mdadm --create md0 --level=0 --raid-devices=2 /dev/nvme0n1
/dev/nvme1n1";

* Format the array using a FS of your choice (for example ext4) and
mount the array;

* Remove one member of the array, for example using sysfs interface (for
nvme: echo 1 > /sys/block/nvme0n1/device/device/remove, for scsi: echo 1
> /sys/block/sdX/device/delete);

* Without this patch, the array state shown by "mdadm --detail" is
"clean", regardless a member is missing/failed.

[Regression potential]

- * There's not much potential regression here; we just exhibit arrays'
- state as "broken" if they have one or more missing/failed members; we
- believe the most common "issue" that could be reported from this patch
- is if an userspace tool rely on the array status as being always "clean"
- even for broken devices, then such tool may behave differently with this
- patch.
+ * There are mainly two potential regressions here; the first is user-
+ visible changes introduced by this mdadm patch. The second is if the
+ patch itself has some unnoticed bug.
+
+ * For the first type of potential regression: this patch introduces a
+ change in how the array state is displayed in "mdadm --detail <array>"
+ output for raid0/linear arrays *only*. Currently, the tool shows just 2
+ states, "clean" or "active". In the patch being SRUed here, this changes
+ for raid0/linear arrays to read the sysfs array state instead. So for
+ example, we could read "readonly" state here for raid0/linear if the
+ user (or some tool) changes the array to such state. This only affects
+ raid0/linear, the output for other levels didn't change at all.
+
+ * Regarding potential unnoticed issues in the code, we changed mainly
+ structs and the "detail" command. Structs were incremented with the new
+ "broken" state and the detail output was changed for raid0/linear as
+ discussed in the previous bullet.

* Note that we *proactively* skipped Xenial SRU here, in order to
prevent potential regressions - Xenial mdadm tool lacks code
infrastructure used by this patch, so the decision was for
safety/stability, by only SRUing Bionic / Disco / Eoan mdadm versions.

[0]
https://git.kernel.org/pub/scm/utils/mdadm/mdadm.git/commit/?id=43ebc910

[other info]

- As mdadm for focal hasn't been merged yet, this will need to be added
- there during or after merge.
+ As mdadm for focal (20.04) hasn't been merged yet, this will need to be
+ added there during or after merge.

--
You received this bug notification because you are a member of STS
Sponsors, which is subscribed to the bug report.
https://bugs.launchpad.net/bugs/1847924

Title:
Introduce broken state parsing to mdadm

Status in mdadm package in Ubuntu:
In Progress
Status in mdadm source package in Bionic:
In Progress
Status in mdadm source package in Disco:
In Progress
Status in mdadm source package in Eoan:
In Progress
Status in mdadm source package in Focal:
In Progress
Status in mdadm package in Debian:
New

Bug description:
[Impact]

* Currently, mounted raid0/md-linear arrays have no indication/warning
when one or more members are removed or suffer from some non-
recoverable error condition. The mdadm tool shows "clean" state
regardless if a member was removed.

* The patch proposed in this SRU addresses this issue by introducing a
new state "broken", which is analog to "clean" but indicates that
array is not in a good/correct state. The commit, available upstream
as 43ebc910 ("mdadm: Introduce new array state 'broken' for
raid0/linear") [0], was extensively discussed and received a good
amount of reviews/analysis by both the current mdadm maintainer as
well as an old maintainer.

* We had reports of users testing a scenario of failed raid0 arrays,
and getting 'clean' in mdadm proved to cause confusion and doesn't
help on noticing something went wrong with the arrays.

* The potential situation this patch (with its kernel counter-part)
addresses is: an user has raid0/linear array, and it's mounted. If one
member fails and gets removed (either physically, like a power or
firmware issue, or in software, like a driver-induced removal due to
detected failure), _without_ this patch (and its kernel counter-part)
there's nothing to let user know it failed, except filesystem errors
in dmesg. Also, non-direct writes to the filesystem will succeed, due
to how page-cache/writeback work; even a 'sync' command run will
succeed.

* The case described in above bullet was tested and the writes to
failed devices succeeded - after a reboot, the files written were
present in the array, but corrupted. An user wouldn't noticed that
unless if the writes were directed or some checksum was performed in
the files. With this patch (and its kernel counter-part), the writes
to such failed raid0/linear array are fast-failed and the filesystem
goes read-only quickly.

[Test case]

* To test this patch, create a raid0 or linear md array on Linux using
mdadm, like: "mdadm --create md0 --level=0 --raid-devices=2
/dev/nvme0n1 /dev/nvme1n1";

* Format the array using a FS of your choice (for example ext4) and
mount the array;

* Remove one member of the array, for example using sysfs interface
(for nvme: echo 1 > /sys/block/nvme0n1/device/device/remove, for scsi:
echo 1 > /sys/block/sdX/device/delete);

* Without this patch, the array state shown by "mdadm --detail" is
"clean", regardless a member is missing/failed.

[Regression potential]

* There are mainly two potential regressions here; the first is user-
visible changes introduced by this mdadm patch. The second is if the
patch itself has some unnoticed bug.

* For the first type of potential regression: this patch introduces a
change in how the array state is displayed in "mdadm --detail <array>"
output for raid0/linear arrays *only*. Currently, the tool shows just
2 states, "clean" or "active". In the patch being SRUed here, this
changes for raid0/linear arrays to read the sysfs array state instead.
So for example, we could read "readonly" state here for raid0/linear
if the user (or some tool) changes the array to such state. This only
affects raid0/linear, the output for other levels didn't change at
all.

* Regarding potential unnoticed issues in the code, we changed mainly
structs and the "detail" command. Structs were incremented with the
new "broken" state and the detail output was changed for raid0/linear
as discussed in the previous bullet.

[0]
https://git.kernel.org/pub/scm/utils/mdadm/mdadm.git/commit/?id=43ebc910

[other info]

As mdadm for focal (20.04) hasn't been merged yet, this will need to
be added there during or after merge.

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/mdadm/+bug/1847924/+subscriptions