[LU-4722] IO Errors during the failover - SLES 11 SP2 - Lustre 2.4.2

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Affects Version/s: Lustre 2.4.2
• Labels: mn4
• Environment: SLES 11 SP2 ; Lustre 2.4.2
• Severity: 3
• Assignee: Hongchao Zhang
• Reporter: Rajeshwaran Ganesan

Description

We have applied the patch provided in teh LU-3645. And still the customer complains that the issue is can be reproduced.

Attaching the latest set of logs.

The issue re-occured on 18th Feb.

Comments

[Peter Jones]

Any update Rajesh?

[Rajeshwaran Ganesan]

We are in the process of applying the patch. I will get back to you with the results.

[Hongchao Zhang]

Hi Rajesh,

What is the result of the test?

Thanks.

[Hongchao Zhang]

there is a bug in obd_str2uuid,

 static inline void obd_str2uuid(struct obd_uuid *uuid, const char *tmp)
 {
        strncpy((char *)uuid->uuid, tmp, sizeof(*uuid));
        uuid->uuid[sizeof(*uuid) - 1] = '\0';
 }

it take "tmp" also as a implicit "obd_uuid" type, but it isn't in all cases, such as in "class_add_uuid", the "tmp" is
"lustre_cfg_string(lcfg, 1)", and obd_str2uuid will copy some undefined data beyond the "tmp" to "uuid" and could cause two same
"uuid" in config were thought to be different.

the patch against b2_4 is tracked at http://review.whamcloud.com/#/c/10269/

Hi Rajesh,
Could you please try the patch in your site?
Thanks!

[Rajeshwaran Ganesan]

Hello Hongchao,

I have uploaded the requested log files into ftp.whamcloud.com:/uploads/LU-4722

2014-05-08-SR30502_pfs2n17.llog.gz

Thanks,
Rajesh

[Hongchao Zhang]

there is no error in these configs.

Could you please collect the debug logs(lctl dk >XXX.log) at the problematic node just after mounting the client (make sure the "ha" is contained in "/proc/sys/lnet/debug")?
Thanks very much!

[Rajeshwaran Ganesan]

There seems to be some confusion about our systems. The system with
prefix pfsc and IP addresses 172.26.4.x is our test system and the
system with prefix pfs2 and IP addresses 172.26.17.x is our production
system. We had seen the issue on both systems and therefore provided
logs from both systems. The configuration of both systems should be
very similar and also the config was newly generated on both systems.

After the config was newly generated the remaining issue is that
duplicate IP addresses appear as failover_nids on the servers only.
This appears for /proc/fs/lustre/osp/*/import on the MDS but
astonishingly only on pfsc and not on pfs2. It also appears for
/proc/fs/lustre/osc/*/import if we mount the file systems on servers
but astonishingly only for some OSTs. This appears if we mount the file
system on an OSS, on an MDS which is currently MDT for another file
system or on an currently unused (failover) MDS.

and also, uploading the logs into ftp site

[Hongchao Zhang]

/proc/fs/lustre/osc/pfs2dat2-OST0012-osc-ffff881033429400/import:
failover_nids: [172.26.8.15@o2ib, 172.26.8.15@o2ib, 172.26.8.14@o2ib]

Do you mount Lustre client at MDT? normally, the OSC(is OSP actually, and there is a symlink in /proc/fs/lustre/osc/
for each entry in /proc/fs/lustre/osp/) name for MDT is (fsname)-OSTXXXX-osc-MDT0000,
the name for client is (fsname)OSTXXXX-osc(address of superblock).

[Hongchao Zhang]

the configs seems fine.
Do you also regenerate the config of pfscdat2? it adds "172.26.17.4@o2ib" and "172.26.17.3@o2ib" as target node, and it was "172.26.4.4@o2ib" and "172.26.4.3@o2ib" previously.

cmd=cf010 marker=10(0x1)pfscdat2-OST0000 'add osc'
cmd=cf005 nid=172.26.17.4@o2ib(0x50000ac1a1104) 0:(null)  1:172.26.17.4@o2ib
cmd=cf001 0:pfscdat2-OST0000-osc  1:osc  2:pfscdat2-clilov_UUID
cmd=cf003 0:pfscdat2-OST0000-osc  1:pfscdat2-OST0000_UUID  2:172.26.17.4@o2ib
cmd=cf005 nid=172.26.17.4@o2ib(0x50000ac1a1104) 0:(null)  1:172.26.17.4@o2ib
cmd=cf00b 0:pfscdat2-OST0000-osc  1:172.26.17.4@o2ib
cmd=cf005 nid=172.26.17.3@o2ib(0x50000ac1a1103) 0:(null)  1:172.26.17.3@o2ib
cmd=cf00b 0:pfscdat2-OST0000-osc  1:172.26.17.3@o2ib 
cmd=cf00d 0:pfscdat2-clilov  1:pfscdat2-OST0000_UUID  2:0  3:1
cmd=cf010 marker=10(0x2)pfscdat2-OST0000 'add osc'

Does the issue occur again in the new generated config?

btw, since only the OSC(OSP) on the MDT is affected, could you please dump the config of MDT (the config-uuid-name is fsname-MDT0000, say, pfscdat2-MDT0000).
Thanks!

[Rajeshwaran Ganesan]

tunefs.lustre --erase-params --mgsnode=172.26.8.12@o2ib
--mgsnode=172.26.8.13@o2ib --servicenode=172.26.8.14@o2ib
--servicenode=172.26.8.15@o2ib /dev/mapper/ost_pfs2dat2_0
tunefs.lustre --erase-params --mgsnode=172.26.8.12@o2ib
--mgsnode=172.26.8.13@o2ib --servicenode=172.26.8.14@o2ib
--servicenode=172.26.8.15@o2ib /dev/mapper/ost_pfs2dat2_1

2. I have uploaded the files in the FTP server

3. only the OSC on the MDT is affected, having the duplicate entries. Anyway, we should find the reason for the wrong behaviour
of the servers.

[Hongchao Zhang]

what is the command line to create the failover nid?

Could you please dump the config file of your system,
at MGS node:
lctl>dl (will show the device list, it will show the device number at the first column)
lctl>device #MGS (MGS index, say, 1)
lctl>dump_cfg pfs2dat2-client (it will dump the config to syslog)

normally one nid (Node Id) will have only one UUID to represent it (one UUID could have more nid)
Having duplicated failover nid could cause some recovery problem, for it will use these nids one by one and it is possible to miss the recovery window.

Thanks