raid1 + 2.6.27.7 issues

raid1 + 2.6.27.7 issues

[Jon Nelson]

I found some time to get back to some raid1 issues I have been having.


Briefly:

I have a pair of machines each with an 80G hard drive. One machine exports
this hard drive over NBD (network block device) to the other. When both
devices are available, they are combined via MD into a raid1. The raid1
looks like this:

/dev/md11:
        Version : 1.00
  Creation Time : Mon Dec 15 07:06:13 2008
     Raid Level : raid1
     Array Size : 78123988 (74.50 GiB 80.00 GB)
  Used Dev Size : 156247976 (149.01 GiB 160.00 GB)
   Raid Devices : 2
  Total Devices : 2
    Persistence : Superblock is persistent

  Intent Bitmap : Internal

    Update Time : Mon Feb  9 09:53:13 2009
          State : active, degraded, recovering
 Active Devices : 1
Working Devices : 2
 Failed Devices : 0
  Spare Devices : 1

 Rebuild Status : 9% complete

           Name : turnip:11  (local to host turnip)
           UUID : cf24d099:9e174a79:2a2f6797:dcff1420
         Events : 90220

    Number   Major   Minor   RaidDevice State
       2      43        0        0      writemostly spare rebuilding
/dev/nbd0
       3       8        0        1      active sync   /dev/sda




The typical use case for me is this: I will run the array (/dev/md11) in
degraded mode (without /dev/nbd0) for a week or so.
At some point, I will try to synchronize the underlying devices. To do this
I use:

mdadm /dev/md11 --re-add /dev/nbd0

The issues I encounter are this: the array goes into *recovery* mode rather
than *resync*, despite the fact that /dev/nbd0 was at one point a full
member (in sync) of the array. Typically, less than 1/3 of the array needs
to be resynchronized, often much less than that. I base this off of the
--examine-bitmap output from /dev/sda. Today it says:

Bitmap : 19074 bits (chunks), 6001 dirty (31.5%)

which is a substantially higher percentage than usual.

Indications of a problem: --examine and --examine-bitmap have an Events
count which agrees for /dev/sda but does *not* agree for /dev/nbd0. From
today, the --examine and --examine-bitmap output from /dev/nbd0:

          Magic : a92b4efc
        Version : 1.0
    Feature Map : 0x1
     Array UUID : cf24d099:9e174a79:2a2f6797:dcff1420
           Name : turnip:11  (local to host turnip)
  Creation Time : Mon Dec 15 07:06:13 2008
     Raid Level : raid1
   Raid Devices : 2

 Avail Dev Size : 160086384 (76.34 GiB 81.96 GB)
     Array Size : 156247976 (74.50 GiB 80.00 GB)
  Used Dev Size : 156247976 (74.50 GiB 80.00 GB)
   Super Offset : 160086512 sectors
          State : clean
    Device UUID : 01524a75:c309869c:6da972c9:084115c6

Internal Bitmap : 2 sectors from superblock
      Flags : write-mostly
    Update Time : Mon Feb  9 09:52:58 2009
       Checksum : 64058b3b - correct
         Events : 90192

    Array Slot : 2 (failed, failed, empty, 1)
   Array State : _u 2 failed
        Filename : /dev/nbd0
           Magic : 6d746962
         Version : 4
            UUID : cf24d099:9e174a79:2a2f6797:dcff1420
          Events : 81596
  Events Cleared : 81570
           State : OK
       Chunksize : 4 MB
          Daemon : 5s flush period
      Write Mode : Allow write behind, max 256
       Sync Size : 78123988 (74.50 GiB 80.00 GB)
          Bitmap : 19074 bits (chunks), 0 dirty (0.0%)


As you can see, --examine says that there are 90192 events, and
--examine-bitmap says there are 81596 events.

So that seems to be a bug or some other issue.

What am I doing wrong that causes the array to go into "recovery" instead of
"resync" mode? It's clearly showing /dev/nbd0 as a *spare* - does this have
anything to do with it?



--
Jon

Re: raid1 + 2.6.27.7 issues

[Iain Rauch]

> The typical use case for me is this: I will run the array (/dev/md11) in
> degraded mode (without /dev/nbd0) for a week or so.
> At some point, I will try to synchronize the underlying devices.

Sounds like rsync is more suited to your application.
Why are you using RAID?

Iain

Re: raid1 + 2.6.27.7 issues

[Ray Van Dolson]

On Mon, Feb 09, 2009 at 08:47:33AM -0800, Iain Rauch wrote:
> > The typical use case for me is this: I will run the array (/dev/md11) in
> > degraded mode (without /dev/nbd0) for a week or so.
> > At some point, I will try to synchronize the underlying devices.
> 
> Sounds like rsync is more suited to your application.
> Why are you using RAID?
> 
> Iain

Seems academic to me.  Whatever the reasons, the above _should_ work
should it not?

Could this be NBD's fault somehow?

Ray

Re: raid1 + 2.6.27.7 issues

[Jon Nelson]

On Mon, Feb 9, 2009 at 10:47 AM, Iain Rauch
<groups@email.iain.rauch.co.uk> wrote:
>> The typical use case for me is this: I will run the array (/dev/md11) in
>> degraded mode (without /dev/nbd0) for a week or so.
>> At some point, I will try to synchronize the underlying devices.
>
> Sounds like rsync is more suited to your application.
> Why are you using RAID?

If I were transferring files then rsync would be fine. Files have
mutable state (applications may be writing to them at the time of
transfer, etc...) and a whole host of other issues.

> Why are you using RAID?

I want a *mirror* of the block device.


-- 
Jon

Re: raid1 + 2.6.27.7 issues

[Jon Nelson]

On Mon, Feb 9, 2009 at 10:59 AM, Ray Van Dolson <rvandolson@esri.com> wrote:
> On Mon, Feb 09, 2009 at 08:47:33AM -0800, Iain Rauch wrote:
>> > The typical use case for me is this: I will run the array (/dev/md11) in
>> > degraded mode (without /dev/nbd0) for a week or so.
>> > At some point, I will try to synchronize the underlying devices.
>>
>> Sounds like rsync is more suited to your application.
>> Why are you using RAID?
>>
>> Iain
>
> Seems academic to me.  Whatever the reasons, the above _should_ work
> should it not?
>
> Could this be NBD's fault somehow?

I don't think so. This is how I *remove* the nbd device:

mdadm /dev/md11 --fail /dev/nbd0
sleep 3
mdadm /dev/md11 --remove /dev/nbd0

and then finally nbd-client -d /dev/nbd0

If necessary, I can try to simulate the problem by using a local
logical volume or some such.

-- 
Jon

Re: raid1 + 2.6.27.7 issues

[Ray Van Dolson]

On Mon, Feb 09, 2009 at 09:49:21AM -0800, Jon Nelson wrote:
> On Mon, Feb 9, 2009 at 10:59 AM, Ray Van Dolson <rvandolson@esri.com> wrote:
> > On Mon, Feb 09, 2009 at 08:47:33AM -0800, Iain Rauch wrote:
> >> > The typical use case for me is this: I will run the array (/dev/md11) in
> >> > degraded mode (without /dev/nbd0) for a week or so.
> >> > At some point, I will try to synchronize the underlying devices.
> >>
> >> Sounds like rsync is more suited to your application.
> >> Why are you using RAID?
> >>
> >> Iain
> >
> > Seems academic to me.  Whatever the reasons, the above _should_ work
> > should it not?
> >
> > Could this be NBD's fault somehow?
> 
> I don't think so. This is how I *remove* the nbd device:
> 
> mdadm /dev/md11 --fail /dev/nbd0
> sleep 3
> mdadm /dev/md11 --remove /dev/nbd0
> 
> and then finally nbd-client -d /dev/nbd0
> 
> If necessary, I can try to simulate the problem by using a local
> logical volume or some such.

Don't want to send you on a wild goose chase, but I'd be interested to
see the results of that and maybe the results of the same with an iSCSI
backed blocked device vs nbd.

I just thought there were some subtle differences between NBD and other
"SAN" over network protocols.  No idea how they'd play into the
scenario you're describing however.

Ray

Re: raid1 + 2.6.27.7 issues

[Jon Nelson]

On Mon, Feb 9, 2009 at 11:53 AM, Ray Van Dolson <rvandolson@esri.com> wrote:
> Don't want to send you on a wild goose chase, but I'd be interested to
> see the results of that and maybe the results of the same with an iSCSI
> backed blocked device vs nbd.

I'm trying with AoE. I've never had any luck setting up a software
iSCSI device. Maybe I'll give that a try one of these days.

Once the dev is done /recovering/ (4.5% done, 3 hours to go) we'll see
if AoE behaves any differently. I doubt it.

-- 
Jon

Re: raid1 + 2.6.27.7 issues

[Jon Nelson]

On Mon, Feb 9, 2009 at 1:30 PM, Bill Davidsen <davidsen@tmr.com> wrote:
> Jon Nelson wrote:
>>
>> I don't think so. This is how I *remove* the nbd device:
>>
>> mdadm /dev/md11 --fail /dev/nbd0
>> sleep 3
>> mdadm /dev/md11 --remove /dev/nbd0
>>
>> and then finally nbd-client -d /dev/nbd0
>>
>> If necessary, I can try to simulate the problem by using a local
>> logical volume or some such.
>>
>
> I would try this with just the first step. I looked at the code briefly, and
> I think the write intent bitmap will not get built for a removed device but
> will for failed. In any case it's certainly something you can easily try.

Nope. :-(  --remove is necessary.

turnip:~ # mdadm --fail /dev/md11 /dev/nbd0
mdadm: set /dev/nbd0 faulty in /dev/md11
turnip:~ # mdadm --re-add /dev/md11 /dev/nbd0
mdadm: Cannot open /dev/nbd0: Device or resource busy
turnip:~ # mdadm --add /dev/md11 /dev/nbd0
mdadm: Cannot open /dev/nbd0: Device or resource busy
turnip:~ #


-- 
Jon

Re: raid1 + 2.6.27.7 issues

[Neil Brown]

On Monday February 9, jnelson-linux-raid@jamponi.net wrote:
> 
> 
> 
> The typical use case for me is this: I will run the array (/dev/md11) in
> degraded mode (without /dev/nbd0) for a week or so.
> At some point, I will try to synchronize the underlying devices. To do this
> I use:
> 
> mdadm /dev/md11 --re-add /dev/nbd0
> 
> The issues I encounter are this: the array goes into *recovery* mode rather
> than *resync*, despite the fact that /dev/nbd0 was at one point a full
> member (in sync) of the array. Typically, less than 1/3 of the array needs
> to be resynchronized, often much less than that.

I've managed to reproduce this.

If you fail the write-mostly device when the array is 'clean' (as
reported by --examine), it works as expected.
If you fail it when the array is 'active', you get the full recovery.

The array is 'active' if there have been any writes in the last 200
msecs, and clean otherwise.

I'll have to have a bit of a think about this and figure out where
what the correct fix is.   Nag me if you haven't heard anything by the
end of the week.

Thanks,
NeilBrown

Re: raid1 + 2.6.27.7 issues

[Jon Nelson]

On Mon, Feb 9, 2009 at 4:17 PM, Neil Brown <neilb@suse.de> wrote:
...

> I've managed to reproduce this.
>
> If you fail the write-mostly device when the array is 'clean' (as
> reported by --examine), it works as expected.
> If you fail it when the array is 'active', you get the full recovery.
>
> The array is 'active' if there have been any writes in the last 200
> msecs, and clean otherwise.
>
> I'll have to have a bit of a think about this and figure out where
> what the correct fix is.   Nag me if you haven't heard anything by the
> end of the week.


Can-do. Here are some more wrinkles:

Wrinkle "A". I can't un-do "write-mostly". I used the md.txt docs that
ship with the kernel which suggest that the following should work:

1. I let the array come up to sync
2. I echoed "-writemostly" into /sys/block/md11/md/dev-nbd0/state
3. A 'cat state' showed "in_sync,write_mostly" before, and "in_sync" after.
4. --fail and --remove /dev/nbd0
5. --re-add /dev/nbd0
6. 'cat state' shows "in_sync,write_mostly" again. D'oh!



Wrinkle "B": When I did the above, when I --re-add'ed /dev/nbd0, it
went into "recovery" mode, which completed instantly. My recollection
of "recovery" is that it does not update the bitmap until the entire
process is complete. Is this correct? If so, I'd like to try to
convince you (Neil Brown) that it's worthwhile to behave the same WRT
the bitmap and up-to-dateness regardless of whether it's recovery or
resync.

I'm including the --examine, --examine-bitmap from both /dev/nbd0 and /dev/sda:

/dev/nbd0:
          Magic : a92b4efc
        Version : 1.0
    Feature Map : 0x1
     Array UUID : cf24d099:9e174a79:2a2f6797:dcff1420
           Name : turnip:11  (local to host turnip)
  Creation Time : Mon Dec 15 07:06:13 2008
     Raid Level : raid1
   Raid Devices : 2

 Avail Dev Size : 160086384 (76.34 GiB 81.96 GB)
     Array Size : 156247976 (74.50 GiB 80.00 GB)
  Used Dev Size : 156247976 (74.50 GiB 80.00 GB)
   Super Offset : 160086512 sectors
          State : clean
    Device UUID : 01524a75:c309869c:6da972c9:084115c6

Internal Bitmap : 2 sectors from superblock
      Flags : write-mostly
    Update Time : Mon Feb  9 17:49:21 2009
       Checksum : 6404fbcd - correct
         Events : 90426


    Array Slot : 2 (failed, failed, 0, 1)
   Array State : Uu 2 failed
        Filename : /dev/nbd0
           Magic : 6d746962
         Version : 4
            UUID : cf24d099:9e174a79:2a2f6797:dcff1420
          Events : 90426
  Events Cleared : 90398
           State : OK
       Chunksize : 4 MB
          Daemon : 5s flush period
      Write Mode : Allow write behind, max 256
       Sync Size : 78123988 (74.50 GiB 80.00 GB)
          Bitmap : 19074 bits (chunks), 0 dirty (0.0%)

/dev/sda:
          Magic : a92b4efc
        Version : 1.0
    Feature Map : 0x1
     Array UUID : cf24d099:9e174a79:2a2f6797:dcff1420
           Name : turnip:11  (local to host turnip)
  Creation Time : Mon Dec 15 07:06:13 2008
     Raid Level : raid1
   Raid Devices : 2

 Avail Dev Size : 160086384 (76.34 GiB 81.96 GB)
     Array Size : 156247976 (74.50 GiB 80.00 GB)
  Used Dev Size : 156247976 (74.50 GiB 80.00 GB)
   Super Offset : 160086512 sectors
          State : clean
    Device UUID : 0059434c:ecef51a0:2974482d:ba38f944

Internal Bitmap : 2 sectors from superblock
    Update Time : Mon Feb  9 17:57:34 2009
       Checksum : 2184ad61 - correct
         Events : 90446


    Array Slot : 3 (failed, failed, failed, 1)
   Array State : _U 3 failed
        Filename : /dev/sda
           Magic : 6d746962
         Version : 4
            UUID : cf24d099:9e174a79:2a2f6797:dcff1420
          Events : 90446
  Events Cleared : 90398
           State : OK
       Chunksize : 4 MB
          Daemon : 5s flush period
      Write Mode : Allow write behind, max 256
       Sync Size : 78123988 (74.50 GiB 80.00 GB)
          Bitmap : 19074 bits (chunks), 0 dirty (0.0%)




-- 
Jon

Re: raid1 + 2.6.27.7 issues

[Neil Brown]

On Monday February 9, jnelson-linux-raid@jamponi.net wrote:
> On Mon, Feb 9, 2009 at 4:17 PM, Neil Brown <neilb@suse.de> wrote:
> ...
> 
> > I've managed to reproduce this.
> >
> > If you fail the write-mostly device when the array is 'clean' (as
> > reported by --examine), it works as expected.
> > If you fail it when the array is 'active', you get the full recovery.
> >
> > The array is 'active' if there have been any writes in the last 200
> > msecs, and clean otherwise.
> >
> > I'll have to have a bit of a think about this and figure out where
> > what the correct fix is.   Nag me if you haven't heard anything by the
> > end of the week.

See below...

> 
> 
> Can-do. Here are some more wrinkles:
> 
> Wrinkle "A". I can't un-do "write-mostly". I used the md.txt docs that
> ship with the kernel which suggest that the following should work:


You want mdadm 2.6.8
  mdadm /dev/md0 --re-add --readwrite /dev/whatever

... or you would if it actually worked... That's odd, I cannot have
tested that....  I'll have to think about that too.

> 
> 
> Wrinkle "B": When I did the above, when I --re-add'ed /dev/nbd0, it
> went into "recovery" mode, which completed instantly. My recollection
> of "recovery" is that it does not update the bitmap until the entire
> process is complete. Is this correct? If so, I'd like to try to
> convince you (Neil Brown) that it's worthwhile to behave the same WRT
> the bitmap and up-to-dateness regardless of whether it's recovery or
> resync.

If the recovery is completing instantly, I wonder why you care exactly
when in that instant the bitmap is updated.... but I suspect that is
missing the point.

No, the bitmap isn't updated during recovery... Maybe it could be...
More thinking.


Mean while, this patch should fix your original problem.

commit 67ad8eaf70c5ca2948b482138d3f88764b3e8ee5
Author: NeilBrown <neilb@suse.de>
Date:   Wed Feb 11 15:33:21 2009 +1100

    md: never clear bit from the write-intent bitmap when the array is degraded.
    
    
    It is safe to clear a bit from the write-intent bitmap for a raid1
    when if we know the data has been written to all devices, which is
    what the current test does.
    But it is not always safe to update the 'events_cleared' counter in
    that case.  This is one request could complete successfully after some
    other request has partially failed.
    
    So simply disable the clearing and updating of events_cleared whenever
    the array is degraded.  This might end up not clearing some bits that
    could safely be cleared, but it is safest approach.
    
    Signed-off-by: NeilBrown <neilb@suse.de>

diff --git a/drivers/md/raid1.c b/drivers/md/raid1.c
index 01e3cff..d875172 100644
--- a/drivers/md/raid1.c
+++ b/drivers/md/raid1.c
@@ -386,7 +386,8 @@ static void raid1_end_write_request(struct bio *bio, int error)
 			/* clear the bitmap if all writes complete successfully */
 			bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 					r1_bio->sectors,
-					!test_bit(R1BIO_Degraded, &r1_bio->state),
+					!test_bit(R1BIO_Degraded, &r1_bio->state)
+					&& !r1_bio->mddev->degraded,
 					behind);
 			md_write_end(r1_bio->mddev);
 			raid_end_bio_io(r1_bio);
diff --git a/drivers/md/raid10.c b/drivers/md/raid10.c
index 6736d6d..9797a85 100644
--- a/drivers/md/raid10.c
+++ b/drivers/md/raid10.c
@@ -332,7 +332,8 @@ static void raid10_end_write_request(struct bio *bio, int error)
 		/* clear the bitmap if all writes complete successfully */
 		bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 				r10_bio->sectors,
-				!test_bit(R10BIO_Degraded, &r10_bio->state),
+				!test_bit(R10BIO_Degraded, &r10_bio->state) &&
+				!r10_bio->mddev->degraded,
 				0);
 		md_write_end(r10_bio->mddev);
 		raid_end_bio_io(r10_bio);
diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
index a5ba080..4d71cce 100644
--- a/drivers/md/raid5.c
+++ b/drivers/md/raid5.c
@@ -2076,7 +2076,8 @@ static void handle_stripe_clean_event(raid5_conf_t *conf,
 					bitmap_endwrite(conf->mddev->bitmap,
 							sh->sector,
 							STRIPE_SECTORS,
-					 !test_bit(STRIPE_DEGRADED, &sh->state),
+					 !test_bit(STRIPE_DEGRADED, &sh->state) &&
+					 !conf->mddev->degraded,
 							0);
 			}
 		}