[LU-3645] Interop 2.1.5 <--> 2.4 Write operations during failover errors out instead of stalling

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Affects Version/s: Lustre 2.4.0, Lustre 2.1.5
• Labels: mn1
• Severity: 3
• Assignee: Hongchao Zhang
• Reporter: Girish Shilamkar (Inactive)

Description

During acceptance testing for KIT, they tried OSS failover while running several applications. And applications got IO errors (can't create file and similar messages). This should not happen and IO should just stall till failover happens.
The clients were running 2.4 and servers were 2.1.5. We tried with 2.1.5 clients and did not see this issue. I have attached the client and server logs.

Comments

[Kit Westneat (Inactive)]

I tested adding the degraded flag to the oscc and it seemed to fix the problem on my test system:

        if(rc != 0 && rc != -ENOTCONN && rc != -ENODEV)
                GOTO(out_wake, rc);
           
+        if (rc == -ENOTCONN || rc ==  -ENODEV)
+                oscc->oscc_flags |= OSCC_FLAG_DEGRADED;

        /* Handle the critical type errors first.

I ran into one issue where the OST ran out of inodes and it caused the lov_create threads on the MDT to deadlock. I'm not sure if that's a new behavior or in stock 2.1.6 as well. I'll investigate.

[Kit Westneat (Inactive)]

Hi Hongchao,

Any updates?

Thanks,
Kit

[Kit Westneat (Inactive)]

Ah ok, that makes sense. Should handle_async_create add the RECOVERING flag to the oscc? It seems like osc_interpret_create would add it, but is it possible that it could get to oscc_internal_create without that flag?

[Hongchao Zhang]

Hi
the fake requests will be queued into osc_creater.oscc_wait_create_list in osc_create_async to wait the objects creation to complete, and it will be woken up
in osc_interpret_create with returned value from OST. in this case, -ENOTCONN or -ENODEV is returned and failed these fake requests, which should continue
waiting the recovery of OST to complete and the objects will be ready then.

I'll create a new debug patch with the http://review.whamcloud.com/#/c/7559/

Thanks!

[Kit Westneat (Inactive)]

Hi Hongchao,

I was trying to follow the create logic some and it's confused me some. It seems like the fake requests are handled by oscc_internal_create, but then if the oscc is recovering (due to the EAGAIN) all it does is return 0. Am I reading that right?

Thanks,
Kit

[Kit Westneat (Inactive)]

Hi Hongchao,

The patch fixed the data corruption, but we are still seeing the same errors on failover. Should we reproduce the issue with the original debugging patch + fix? Here are the kernel messages from the fix (patch 7559):

Sep 25 15:16:49 pfscn2 kernel: : LustreError: 11-0: an error occurred while communicating with 172.26.17.3@o2ib. The ost_setattr operation failed with -19
Sep 25 15:16:49 pfscn2 kernel: : Lustre: pfscdat2-OST0000-osc-MDT0000: Connection to pfscdat2-OST0000 (at 172.26.17.3@o2ib) was lost; in progress operations using this service will wait for recovery to complete
Sep 25 15:17:39 pfscn2 kernel: : LustreError: 9001:0:(lov_request.c:579:lov_update_create_set()) error creating fid 0x3288 sub-object on OST idx 0/1: rc = -11
Sep 25 15:17:39 pfscn2 kernel: : LustreError: 7308:0:(lov_request.c:579:lov_update_create_set()) error creating fid 0x3288 sub-object on OST idx 0/1: rc = -5
Sep 25 15:17:40 pfscn2 kernel: : Lustre: pfscdat2-OST0000-osc-MDT0000: Connection restored to pfscdat2-OST0000 (at 172.26.17.4@o2ib)
Sep 25 15:17:40 pfscn2 kernel: : Lustre: MDS mdd_obd-pfscdat2-MDT0000: pfscdat2-OST0000_UUID now active, resetting orphans

[Hongchao Zhang]

sorry, there is an error in the previous patch, and it has been updated (http://review.whamcloud.com/#/c/7559/)

[Kit Westneat (Inactive)]

I should add that we didn't see the problem any more after downgrading to stock 2.1.5. We did however see the issue at both 2.1.6+patch and 2.1.5+patch.

[Kit Westneat (Inactive)]

Hi Hongchao,

We've run into a problem with this patch. After upgrading the MDT, we are starting to get data corruption on the system. Here is a description of the issue:

Another admin has been working on the test file system this morning.
First, she could execute normal commands on the file system. Then
normal commands (like vi) added 112 MB rubbish to files. Even a touch
on a non-existing file created a file of 112 MB.

Another admin logged in and his .Xauthority was increased. He looked
at the binary data ant it seemed like the additional data came from a
software package which is installed on the same Lustre file system,
i.e. the rubbish seems to be no arbitrary data but seems to come from
another location.

root@iccn999:/software/all/tsm/sbin# touch gaga1
Wed Sep 11-14:38:21 (14/1012) - ACTIVE
root@iccn999:/software/all/tsm/sbin# ls -l gaga1
rw-rr- 1 root root 116430464 Sep 11 14:38 gaga1

On another client the behaviour is different:
root@iccn996:/software/all/tsm/sbin# touch gaga2
touch: setting times of `gaga2': No such file or directory
Wed Sep 11-14:39:15 (5/41)
root@iccn996:/software/all/tsm/sbin# ls -l gaga2
rw-rr- 1 root root 0 Sep 11 14:39 gaga2

I will upload the lctl dk logs (with vfstrace and rpctrace). Is there any other information we should get?

[Hongchao Zhang]

Hi Kit

the patch is for MDT only.

Thanks

[Kit Westneat (Inactive)]

Hi Hongchao,

Is this an MDT only patch, or is it for the clients too?

Thanks,
Kit