MD: Long delay for container drive removal

MD: Long delay for container drive removal

[Mateusz Kusiak]

Hi all,
we have an issue submitted for SLES15SP6 that is caused by huge delays when trying to remove drive 
from a container.

The scenario is as follows:
1. Create two drive imsm container
# mdadm --create --run /dev/md/imsm --metadata=imsm --raid-devices=2 /dev/nvme[0-1]n1
2. Remove single drive from container
# mdadm /dev/md127 --remove /dev/nvme0n1

The problem is that drive removal may take up to 7 seconds, which causes timeouts for other 
components that are mdadm dependent.

We narrowed it down to be MD related. We tested this with inbox mdadm-4.3 and mdadm-4.2 on SP6 and 
delay time is pretty much the same. SP5 is free of this issue.

I also tried RHEL 8.9 and drive removal is almost instant.

Is it default behavior now, or should we treat this as an issue?

Thanks,
Mateusz

Re: MD: Long delay for container drive removal

[Mateusz Kusiak]

On 18.06.2024 16:24, Mateusz Kusiak wrote:
> Hi all,
> we have an issue submitted for SLES15SP6 that is caused by huge delays when trying to remove drive 
> from a container.
> 
> The scenario is as follows:
> 1. Create two drive imsm container
> # mdadm --create --run /dev/md/imsm --metadata=imsm --raid-devices=2 /dev/nvme[0-1]n1
> 2. Remove single drive from container
> # mdadm /dev/md127 --remove /dev/nvme0n1
> 
> The problem is that drive removal may take up to 7 seconds, which causes timeouts for other 
> components that are mdadm dependent.
> 
> We narrowed it down to be MD related. We tested this with inbox mdadm-4.3 and mdadm-4.2 on SP6 and 
> delay time is pretty much the same. SP5 is free of this issue.
> 
> I also tried RHEL 8.9 and drive removal is almost instant.
> 
> Is it default behavior now, or should we treat this as an issue?
> 
> Thanks,
> Mateusz
> 

I dug into this more. I retested this on:
- Ubuntu 24.04 with inbox kernel 6.6.0: No reproduction
- RHEL 9.4 with usptream kernel: 6.9.5-1: Got reproduction
(Note that SLES15SP6 comes with 6.8.0-rc4 inbox)

I plugged into mdadm with gdb and found out that ioctl call in hot_remove_disk() fails and it's 
causing a delay. The function looks as follows:

int hot_remove_disk(int mdfd, unsigned long dev, int force)
{
	int cnt = force ? 500 : 5;
	int ret;

	/* HOT_REMOVE_DISK can fail with EBUSY if there are
	 * outstanding IO requests to the device.
	 * In this case, it can be helpful to wait a little while,
	 * up to 5 seconds if 'force' is set, or 50 msec if not.
	 */
	while ((ret = ioctl(mdfd, HOT_REMOVE_DISK, dev)) == -1 &&
	       errno == EBUSY &&
	       cnt-- > 0)
		sleep_for(0, MSEC_TO_NSEC(10), true);

	return ret;
}
... if it fails, then it defaults to removing drive via sysfs call.

Looks like a kernel ioctl issue...

Thanks,
Mateusz

Re: MD: Long delay for container drive removal

[Mariusz Tkaczyk]

On Thu, 20 Jun 2024 14:43:50 +0200
Mateusz Kusiak <mateusz.kusiak@linux.intel.com> wrote:

> On 18.06.2024 16:24, Mateusz Kusiak wrote:
> > Hi all,
> > we have an issue submitted for SLES15SP6 that is caused by huge delays when
> > trying to remove drive from a container.
> > 
> > The scenario is as follows:
> > 1. Create two drive imsm container
> > # mdadm --create --run /dev/md/imsm --metadata=imsm --raid-devices=2
> > /dev/nvme[0-1]n1 2. Remove single drive from container
> > # mdadm /dev/md127 --remove /dev/nvme0n1
> > 
> > The problem is that drive removal may take up to 7 seconds, which causes
> > timeouts for other components that are mdadm dependent.
> > 
> > We narrowed it down to be MD related. We tested this with inbox mdadm-4.3
> > and mdadm-4.2 on SP6 and delay time is pretty much the same. SP5 is free of
> > this issue.
> > 
> > I also tried RHEL 8.9 and drive removal is almost instant.
> > 
> > Is it default behavior now, or should we treat this as an issue?
> > 
> > Thanks,
> > Mateusz
> >   
> 
> I dug into this more. I retested this on:
> - Ubuntu 24.04 with inbox kernel 6.6.0: No reproduction
> - RHEL 9.4 with usptream kernel: 6.9.5-1: Got reproduction
> (Note that SLES15SP6 comes with 6.8.0-rc4 inbox)
> 
> I plugged into mdadm with gdb and found out that ioctl call in
> hot_remove_disk() fails and it's causing a delay. The function looks as
> follows:
> 
> int hot_remove_disk(int mdfd, unsigned long dev, int force)
> {
> 	int cnt = force ? 500 : 5;
> 	int ret;
> 
> 	/* HOT_REMOVE_DISK can fail with EBUSY if there are
> 	 * outstanding IO requests to the device.
> 	 * In this case, it can be helpful to wait a little while,
> 	 * up to 5 seconds if 'force' is set, or 50 msec if not.
> 	 */
> 	while ((ret = ioctl(mdfd, HOT_REMOVE_DISK, dev)) == -1 &&
> 	       errno == EBUSY &&
> 	       cnt-- > 0)
> 		sleep_for(0, MSEC_TO_NSEC(10), true);
> 
> 	return ret;
> }
> ... if it fails, then it defaults to removing drive via sysfs call.
> 
> Looks like a kernel ioctl issue...
> 

Hello,
I investigated this. Looks like HOT_REMOVE_DRIVE ioctl almost always failed for
raid with no raid personality. At some point it was allowed but it was blocked
6 years ago in c42a0e2675 (this id leads to merge commit, so giving title "md:
fix NULL dereference of mddev->pers in remove_and_add_spares()").

And that explains why we have outdated comment in mdadm:

		if (err && errno == ENODEV) {
			/* Old kernels rejected this if no personality
			 * is registered */

I'm working to make it fixed in mdadm (for kernels with this hang), I will
remove ioctl call for external containers:
https://github.com/md-raid-utilities/mdadm/pull/31

On HOT_REMOVE_DRIVE ioctl path, there is a wait for clearing MD_RECOVERY_NEEDED
flag with timeout set to 5 seconds. When I disabled this for arrays
with no personality- it fixes issue. However, I'm not sure if it is right fix. I
would expect to not set MD_RECOVERY_NEEDED for arrays with no MD personality.
Kuai and Song could you please advice?

diff --git a/drivers/md/md.c b/drivers/md/md.c
index c0426a6d2fd1..bd1cedeb105b 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c
@@ -7827,7 +7827,7 @@ static int md_ioctl(struct block_device *bdev, blk_mode_t
mode, return get_bitmap_file(mddev, argp);
        }

-       if (cmd == HOT_REMOVE_DISK)
+       if (cmd == HOT_REMOVE_DISK && mddev->pers)
                /* need to ensure recovery thread has run */
                wait_event_interruptible_timeout(mddev->sb_wait,
                                                 !test_bit(MD_RECOVERY_NEEDED,


Thanks,
Mariusz  

Re: MD: Long delay for container drive removal

[Yu Kuai]

Hi,

在 2024/06/24 15:14, Mariusz Tkaczyk 写道:
> On Thu, 20 Jun 2024 14:43:50 +0200
> Mateusz Kusiak <mateusz.kusiak@linux.intel.com> wrote:
> 
>> On 18.06.2024 16:24, Mateusz Kusiak wrote:
>>> Hi all,
>>> we have an issue submitted for SLES15SP6 that is caused by huge delays when
>>> trying to remove drive from a container.
>>>
>>> The scenario is as follows:
>>> 1. Create two drive imsm container
>>> # mdadm --create --run /dev/md/imsm --metadata=imsm --raid-devices=2
>>> /dev/nvme[0-1]n1 2. Remove single drive from container
>>> # mdadm /dev/md127 --remove /dev/nvme0n1
>>>
>>> The problem is that drive removal may take up to 7 seconds, which causes
>>> timeouts for other components that are mdadm dependent.
>>>
>>> We narrowed it down to be MD related. We tested this with inbox mdadm-4.3
>>> and mdadm-4.2 on SP6 and delay time is pretty much the same. SP5 is free of
>>> this issue.
>>>
>>> I also tried RHEL 8.9 and drive removal is almost instant.
>>>
>>> Is it default behavior now, or should we treat this as an issue?
>>>
>>> Thanks,
>>> Mateusz
>>>    
>>
>> I dug into this more. I retested this on:
>> - Ubuntu 24.04 with inbox kernel 6.6.0: No reproduction
>> - RHEL 9.4 with usptream kernel: 6.9.5-1: Got reproduction
>> (Note that SLES15SP6 comes with 6.8.0-rc4 inbox)
>>
>> I plugged into mdadm with gdb and found out that ioctl call in
>> hot_remove_disk() fails and it's causing a delay. The function looks as
>> follows:
>>
>> int hot_remove_disk(int mdfd, unsigned long dev, int force)
>> {
>> 	int cnt = force ? 500 : 5;
>> 	int ret;
>>
>> 	/* HOT_REMOVE_DISK can fail with EBUSY if there are
>> 	 * outstanding IO requests to the device.
>> 	 * In this case, it can be helpful to wait a little while,
>> 	 * up to 5 seconds if 'force' is set, or 50 msec if not.
>> 	 */
>> 	while ((ret = ioctl(mdfd, HOT_REMOVE_DISK, dev)) == -1 &&
>> 	       errno == EBUSY &&
>> 	       cnt-- > 0)
>> 		sleep_for(0, MSEC_TO_NSEC(10), true);
>>
>> 	return ret;
>> }
>> ... if it fails, then it defaults to removing drive via sysfs call.
>>
>> Looks like a kernel ioctl issue...
>>
> 
> Hello,
> I investigated this. Looks like HOT_REMOVE_DRIVE ioctl almost always failed for
> raid with no raid personality. At some point it was allowed but it was blocked
> 6 years ago in c42a0e2675 (this id leads to merge commit, so giving title "md:
> fix NULL dereference of mddev->pers in remove_and_add_spares()").
> 
> And that explains why we have outdated comment in mdadm:
> 
> 		if (err && errno == ENODEV) {
> 			/* Old kernels rejected this if no personality
> 			 * is registered */
> 
> I'm working to make it fixed in mdadm (for kernels with this hang), I will
> remove ioctl call for external containers:
> https://github.com/md-raid-utilities/mdadm/pull/31
> 
> On HOT_REMOVE_DRIVE ioctl path, there is a wait for clearing MD_RECOVERY_NEEDED
> flag with timeout set to 5 seconds. When I disabled this for arrays
> with no personality- it fixes issue. However, I'm not sure if it is right fix. I
> would expect to not set MD_RECOVERY_NEEDED for arrays with no MD personality.
> Kuai and Song could you please advice?
> 
> diff --git a/drivers/md/md.c b/drivers/md/md.c
> index c0426a6d2fd1..bd1cedeb105b 100644
> --- a/drivers/md/md.c
> +++ b/drivers/md/md.c
> @@ -7827,7 +7827,7 @@ static int md_ioctl(struct block_device *bdev, blk_mode_t
> mode, return get_bitmap_file(mddev, argp);
>          }
> 
> -       if (cmd == HOT_REMOVE_DISK)
> +       if (cmd == HOT_REMOVE_DISK && mddev->pers)

This patch will work, however, I'm afraid this can't fix the problem
thoroughly, because this is called without 'reconfig_mutex', and
mddev->pers can be set later before hot_remove_disk().

After taking a look at the commit 90f5f7ad4f38, which introducing the
waiting, was trying to wait for a failed device to be removed by
md_check_recovery().  However, this doen't make sense now, because
remove_and_add_spares() is called diretcly from hot_remove_disk(),
hence failed device can be removed directly from ioctl, without
md_check_recovery().

The only thining to prevent a failed device to be removed from array
wound be MD_RECOVERY_RUNNING, however, we can't wait for this falg to be
cleared, hence I'll suggest to revert this patch.

Thanks,
Kuai

>                  /* need to ensure recovery thread has run */
>                  wait_event_interruptible_timeout(mddev->sb_wait,
>                                                   !test_bit(MD_RECOVERY_NEEDED,
> 
> 
> Thanks,
> Mariusz
> 
> .
>