[PATCH] Pass --retry to dmsetup when removing flakey and error devices

[PATCH] Pass --retry to dmsetup when removing flakey and error devices

[fdmanana]

From: Filipe Manana <fdmanana@suse.com>

When running multiple tests using dm's flakey target, often (very rarely)
I was getting an error when a test tried to setup a dm flakey target,
resulting in the following messages in the test's .out.bad file:

  device-mapper: reload ioctl on flakey-test failed: Device or resource busy
  Command failed
  failed to create flakey device

Upon further investigation it turns out that it was because the previous
test that ran had failed to remove the flakey device because the device
was being used at the time but this has not made that previous test fail
because we simply redirect the stderr (and stdout) to /dev/null and don't
fail if the dmsetup remove command exits with a non-zero status.
The device was in use, when the test attempted to remove it, by an udev
rule (btrfs udev rule).

Fix this by passing the option --retry to dmsetup remove, which serves to
deal with such cases. From dmsetup's man page:

  "If an attempt to remove a device fails, perhaps because a process run
   from a quick udev rule temporarily opened the device, the --retry
   option will cause the operation to be retried for a few seconds before
   failing."

Signed-off-by: Filipe Manana <fdmanana@suse.com>
---
 common/dmerror  | 4 ++--
 common/dmflakey | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/common/dmerror b/common/dmerror
index 5d2c1b6..c898df0 100644
--- a/common/dmerror
+++ b/common/dmerror
@@ -22,7 +22,7 @@ _dmerror_init()
 {
 	local dm_backing_dev=$SCRATCH_DEV
 
-	$DMSETUP_PROG remove error-test > /dev/null 2>&1
+	$DMSETUP_PROG remove --retry error-test > /dev/null 2>&1
 
 	local blk_dev_size=`blockdev --getsz $dm_backing_dev`
 
@@ -57,7 +57,7 @@ _dmerror_cleanup()
 	# wait for device to be fully settled so that 'dmsetup remove' doesn't
 	# fail due to EBUSY
 	$UDEV_SETTLE_PROG >/dev/null 2>&1
-	$DMSETUP_PROG remove error-test > /dev/null 2>&1
+	$DMSETUP_PROG remove --retry error-test > /dev/null 2>&1
 }
 
 _dmerror_load_error_table()
diff --git a/common/dmflakey b/common/dmflakey
index 4434307..cc458f0 100644
--- a/common/dmflakey
+++ b/common/dmflakey
@@ -57,7 +57,7 @@ _cleanup_flakey()
 	# wait for device to be fully settled so that 'dmsetup remove' doesn't
 	# fail due to EBUSY
 	$UDEV_SETTLE_PROG >/dev/null 2>&1
-	$DMSETUP_PROG remove flakey-test > /dev/null 2>&1
+	$DMSETUP_PROG remove --retry flakey-test > /dev/null 2>&1
 	$DMSETUP_PROG mknodes > /dev/null 2>&1
 }
 
-- 
2.7.0.rc3

Re: [PATCH] Pass --retry to dmsetup when removing flakey and error devices

[Dave Chinner]

On Thu, Apr 07, 2016 at 07:12:28AM +0100, fdmanana@kernel.org wrote:
> From: Filipe Manana <fdmanana@suse.com>
> 
> When running multiple tests using dm's flakey target, often (very rarely)
> I was getting an error when a test tried to setup a dm flakey target,
> resulting in the following messages in the test's .out.bad file:
> 
>   device-mapper: reload ioctl on flakey-test failed: Device or resource busy
>   Command failed
>   failed to create flakey device
> 
> Upon further investigation it turns out that it was because the previous
> test that ran had failed to remove the flakey device because the device
> was being used at the time but this has not made that previous test fail
> because we simply redirect the stderr (and stdout) to /dev/null and don't
> fail if the dmsetup remove command exits with a non-zero status.
> The device was in use, when the test attempted to remove it, by an udev
> rule (btrfs udev rule).

And that's why we have $UDEV_SETTLE_PROG calls before removing the
dm device, which it appears _dmerror_init() does not have.

If there are things other than udev holding the device open, then we
ned to find what they are first rather than trying to ignore them...

> Fix this by passing the option --retry to dmsetup remove, which serves to
> deal with such cases. From dmsetup's man page:
> 
>   "If an attempt to remove a device fails, perhaps because a process run
>    from a quick udev rule temporarily opened the device, the --retry
>    option will cause the operation to be retried for a few seconds before
>    failing."

It also talks about udev, so running $UDEV_SETTLE_PROG before remove
should prevent this from happening. If it is not sufficient, then we
need to understand why it is not sufficient, not blindly hack around
it.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com