We are trying to figure out how to use rados bench to reproduce. Ceph itself doesn't seem to think there is any corruption, but when you do a verify inside the RBD, there is. Can rados bench verify the objects after they are written? It also seems to be primarily the filesystem metadata that is corrupted. If we fsck the volume, there is missing data (put into lost+found), but if it is there it is primarily OK. There only seems to be a few cases where a file's contents are corrupted. I would suspect on an object boundary. We would have to look at blockinfo to map that out and see if that is what is happening.
We stopped all the IO and did put the tier in writeback mode with recency 1, set the recency to 2 and started the test and there was corruption, so it doesn't seem to be limited to changing the mode. I don't know how that patch could cause the issue either. Unless there is a bug that reads from the back tier, but writes to cache tier, then the object gets promoted wiping that last write, but then it seems like it should not be as much corruption since the metadata should be in the cache pretty quick. We usually evited the cache before each try so we should not be evicting on writeback.
Sent from a mobile device, please excuse any typos.
On Mar 17, 2016 6:26 AM, "Sage Weil" <sweil@xxxxxxxxxx> wrote:
On Thu, 17 Mar 2016, Nick Fisk wrote:
> There is got to be something else going on here. All that PR does is to
> potentially delay the promotion to hit_set_period*recency instead of
> just doing it on the 2nd read regardless, it's got to be uncovering
> another bug.
>
> Do you see the same problem if the cache is in writeback mode before you
> start the unpacking. Ie is it the switching mid operation which causes
> the problem? If it only happens mid operation, does it still occur if
> you pause IO when you make the switch?
>
> Do you also see this if you perform on a RBD mount, to rule out any
> librbd/qemu weirdness?
>
> Do you know if it’s the actual data that is getting corrupted or if it's
> the FS metadata? I'm only wondering as unpacking should really only be
> writing to each object a couple of times, whereas FS metadata could
> potentially be being updated+read back lots of times for the same group
> of objects and ordering is very important.
>
> Thinking through it logically the only difference is that with recency=1
> the object will be copied up to the cache tier, where recency=6 it will
> be proxy read for a long time. If I had to guess I would say the issue
> would lie somewhere in the proxy read + writeback<->forward logic.
That seems reasonable. Was switching from writeback -> forward always
part of the sequence that resulted in corruption? Not that there is a
known ordering issue when switching to forward mode. I wouldn't really
expect it to bite real users but it's possible..
http://tracker.ceph.com/issues/12814
I've opened a ticket to track this:
http://tracker.ceph.com/issues/15171
What would be *really* great is if you could reproduce this with a
ceph_test_rados workload (from ceph-tests). I.e., get ceph_test_rados
running, and then find the sequence of operations that are sufficient to
trigger a failure.
sage
>
>
>
> > -----Original Message-----
> > From: ceph-users [mailto:ceph-users-bounces@xxxxxxxxxxxxxx] On Behalf Of
> > Mike Lovell
> > Sent: 16 March 2016 23:23
> > To: ceph-users <ceph-users@xxxxxxxxxxxxxx>; sweil@xxxxxxxxxx
> > Cc: Robert LeBlanc <robert.leblanc@xxxxxxxxxxxxx>; William Perkins
> > <william.perkins@xxxxxxxxxxxxx>
> > Subject: Re: data corruption with hammer
> >
> > just got done with a test against a build of 0.94.6 minus the two commits that
> > were backported in PR 7207. everything worked as it should with the cache-
> > mode set to writeback and the min_read_recency_for_promote set to 2.
> > assuming it works properly on master, there must be a commit that we're
> > missing on the backport to support this properly.
> >
> > sage,
> > i'm adding you to the recipients on this so hopefully you see it. the tl;dr
> > version is that the backport of the cache recency fix to hammer doesn't work
> > right and potentially corrupts data when
> > the min_read_recency_for_promote is set to greater than 1.
> >
> > mike
> >
> > On Wed, Mar 16, 2016 at 4:41 PM, Mike Lovell
> > <mike.lovell@xxxxxxxxxxxxx> wrote:
> > robert and i have done some further investigation the past couple days on
> > this. we have a test environment with a hard drive tier and an ssd tier as a
> > cache. several vms were created with volumes from the ceph cluster. i did a
> > test in each guest where i un-tarred the linux kernel source multiple times
> > and then did a md5sum check against all of the files in the resulting source
> > tree. i started off with the monitors and osds running 0.94.5 and never saw
> > any problems.
> >
> > a single node was then upgraded to 0.94.6 which has osds in both the ssd and
> > hard drive tier. i then proceeded to run the same test and, while the untar
> > and md5sum operations were running, i changed the ssd tier cache-mode
> > from forward to writeback. almost immediately the vms started reporting io
> > errors and odd data corruption. the remainder of the cluster was updated to
> > 0.94.6, including the monitors, and the same thing happened.
> >
> > things were cleaned up and reset and then a test was run
> > where min_read_recency_for_promote for the ssd cache pool was set to 1.
> > we previously had it set to 6. there was never an error with the recency
> > setting set to 1. i then tested with it set to 2 and it immediately caused
> > failures. we are currently thinking that it is related to the backport of the fix
> > for the recency promotion and are in progress of making a .6 build without
> > that backport to see if we can cause corruption. is anyone using a version
> > from after the original recency fix (PR 6702) with a cache tier in writeback
> > mode? anyone have a similar problem?
> >
> > mike
> >
> > On Mon, Mar 14, 2016 at 8:51 PM, Mike Lovell
> > <mike.lovell@xxxxxxxxxxxxx> wrote:
> > something weird happened on one of the ceph clusters that i administer
> > tonight which resulted in virtual machines using rbd volumes seeing
> > corruption in multiple forms.
> >
> > when everything was fine earlier in the day, the cluster was a number of
> > storage nodes spread across 3 different roots in the crush map. the first
> > bunch of storage nodes have both hard drives and ssds in them with the hard
> > drives in one root and the ssds in another. there is a pool for each and the
> > pool for the ssds is a cache tier for the hard drives. the last set of storage
> > nodes were in a separate root with their own pool that is being used for burn
> > in testing.
> >
> > these nodes had run for a while with test traffic and we decided to move
> > them to the main root and pools. the main cluster is running 0.94.5 and the
> > new nodes got 0.94.6 due to them getting configured after that was
> > released. i removed the test pool and did a ceph osd crush move to move
> > the first node into the main cluster, the hard drives into the root for that tier
> > of storage and the ssds into the root and pool for the cache tier. each set was
> > done about 45 minutes apart and they ran for a couple hours while
> > performing backfill without any issue other than high load on the cluster.
> >
> > we normally run the ssd tier in the forward cache-mode due to the ssds we
> > have not being able to keep up with the io of writeback. this results in io on
> > the hard drives slowing going up and performance of the cluster starting to
> > suffer. about once a week, i change the cache-mode between writeback and
> > forward for short periods of time to promote actively used data to the cache
> > tier. this moves io load from the hard drive tier to the ssd tier and has been
> > done multiple times without issue. i normally don't do this while there are
> > backfills or recoveries happening on the cluster but decided to go ahead
> > while backfill was happening due to the high load.
> >
> > i tried this procedure to change the ssd cache-tier between writeback and
> > forward cache-mode and things seemed okay from the ceph cluster. about
> > 10 minutes after the first attempt a changing the mode, vms using the ceph
> > cluster for their storage started seeing corruption in multiple forms. the
> > mode was flipped back and forth multiple times in that time frame and its
> > unknown if the corruption was noticed with the first change or subsequent
> > changes. the vms were having issues of filesystems having errors and getting
> > remounted RO and mysql databases seeing corruption (both myisam and
> > innodb). some of this was recoverable but on some filesystems there was
> > corruption that lead to things like lots of data ending up in the lost+found and
> > some of the databases were un-recoverable (backups are helping there).
> >
> > i'm not sure what would have happened to cause this corruption. the libvirt
> > logs for the qemu processes for the vms did not provide any output of
> > problems from the ceph client code. it doesn't look like any of the qemu
> > processes had crashed. also, it has now been several hours since this
> > happened with no additional corruption noticed by the vms. it doesn't
> > appear that we had any corruption happen before i attempted the flipping of
> > the ssd tier cache-mode.
> >
> > the only think i can think of that is different between this time doing this
> > procedure vs previous attempts was that there was the one storage node
> > running 0.94.6 where the remainder were running 0.94.5. is is possible that
> > something changed between these two releases that would have caused
> > problems with data consistency related to the cache tier? or otherwise? any
> > other thoughts or suggestions?
> >
> > thanks in advance for any help you can provide.
> >
> > mike
> >
>
>
>
>
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
