[lustre-discuss] Connectivity issues after client crash

Wed Dec 11 01:10:48 PST 2024

Jesse,
what Lustre version are you using?  I'm wondering if this is an issue that has already been fixed, or if there is a bug with the peer discovery in some cases that needs to be fixed?

On Dec 10, 2024, at 07:40, Jesse Stroik <jesse.stroik at ssec.wisc.edu<mailto:jesse.stroik at ssec.wisc.edu>> wrote:

[You don't often get email from jesse.stroik at ssec.wisc.edu<mailto:jesse.stroik at ssec.wisc.edu>. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]

I spent a good deal of time tracking down and cleaning up infiniband errors (we had some bad cables) but that didn't matter. The solution was to disable peer discovery on all of the ko2iblnd hosts because they were mistakenly showing up as multi-rail in some cases. I suspect it's related to multi-port IB HCAs.

For this to work consistently we had to disable discovery, clear up all peers on the hosts, then let the lustre file systems mount.

Thanks for lending me a bit of your experience, Shane.

Jesse

________________________________________
From: Nehring, Shane R [ITS] <snehring at iastate.edu<mailto:snehring at iastate.edu>>
Sent: Wednesday, December 4, 2024 11:08 AM
To: Jesse Stroik; lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
Subject: Re: Connectivity issues after client crash

No, we're using o2iblnd primarily now, we just used socklnd until we were able to migrate the bulk of nodes to infiniband. There are a small number of nodes that are still using socklnd either via ethernet or remaining opa nodes.

Our cluster is relatively small and everything is single rail, so discovery isn't necessary in our case.

________________________________
From: Jesse Stroik <jesse.stroik at ssec.wisc.edu<mailto:jesse.stroik at ssec.wisc.edu>>
Sent: Tuesday, December 3, 2024 18:18
To: Nehring, Shane R [ITS] <snehring at iastate.edu<mailto:snehring at iastate.edu>>; lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org> <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>>
Subject: Re: Connectivity issues after client crash

Thanks for your reply.

I did consider disabling peer discovery on the servers for one of the file systems as a test, but did not yet test it. i'll keep that idea in my back pocket in case the issue recurs with our workaround or we have the issue happen on any of our other 2.15 file systems.

Did you use your ib equipment just using ksocklnd and forgo rdma?

For us, this would be a fairly major change as our clusters and lustre systems are all on infiniband directly, but we have a lot of ethernet clients as well. And we have seven lustre file systems most of which are routed to many ethernet clients.

Jesse

________________________________________
From: Nehring, Shane R [ITS]
Sent: Tuesday, December 3, 2024 2:50 PM
To: lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>; Jesse Stroik
Subject: Re: Connectivity issues after client crash

Hello Jesse,

What I think we may have been hitting in this particular case was
https://jira.whamcloud.com/browse/LU-16349<https://urldefense.com/v3/__https://jira.whamcloud.com/browse/LU-16349__;!!Mak6IKo!KjIS-mlygGvZXA84UL7NpPd6BbzEHT4hP-NTUTL4RmVJ2GjVu_zPrP0oO1MAkx3CWy4hebKZ4MYH7r4O-XxnkKucCgg$> which we in the short term worked
around by switching to ksocklnd from o2iblnd, and then ultimately mostly moved
away from omnipath to ib.

I've seen some weird behavior in our environment around peer discovery (we have
tcp and o2ib nids on the oss and mds nodes to provide access over both methods)
that've caused problems in the past, such that we just outright disable peer
discovery with modprobe (eg options lnet lnet_peer_discovery_disabled=1 in
/etc/modprobe.d/lnet.conf). I dunno if you've tried that yet, but if not it
might be a place to start (assuming you don't need automatic multi-rail)

Shane

On Tue, 2024-12-03 at 20:29 +0000, Jesse Stroik wrote:
Hi Shane,

I realize this is quite an old post but I think it is worth responding for
posterity and because I suspect others who upgrade may run into this issue.

I'm observing some similar issues to what you describe. They started this
weekend for us on two of our servers which were upgraded to rocky 8 and lustre
2.15.5 a couple of months ago. Like you, we run infiniband. We also route to
ethernet clients and our lustre routers are now running rocky 8 and either
lustre 2.15.5 or 2.15.3.

The behavior that we observe is widely similar to yours though we have a few
observations that may be relevant. I suspect that if we can limit the
infiniband communication to certain segments or LIDs it seems to behave fine.

Today i was able to reproduce the problem when routing to an ethernet client
that ran our robinhood policy engine. With the file systems behaving normally
to our infiniband network clients, I enabled routing on them again and tested
with all lrouters enabled or just individual lrouters enabled. In all cases
that i tested, mounting one of those file systems resulted in the
communication problems on the MDS and OSS servers for that file system. In one
case an MDS panicked as soon as that client mounted its file system. I put an
IB card into that policy engine client, removed all lnet routing for those
file systems and lnet behavior was stable with that client mounting those file
systems and ingesting the changelogs.

When the servers have failed lnet communication, they can sometimes be fixed
by stopping the lustre service / unmounting the OSTs and MDTs and restarting /
reconfiguring lnet. If i'm unable to do that, rebooting the affected systems
seems to be the only solution and they sometimes hang indefinitely trying to
shut down lnet necessitating a power cycle.

When the issue first started I did suspect the infiniband fabric and looked
for problems with that, though, like you, I didn't find anything conclusive. A
recent change we've had is our OpenSM subnet manager definitely restarted and
moved when we updated the kernel and lustre-client versions on two of our
lrouters. However, even when lnet is unable to communicate, the nodes can all
still ping each other with ipv4 ping IPoIB which suggests the subnet manager
and infiniband are working.

I did also attempt manual lnet peer discovery on individual systems while they
were affected to see if i could recover them that way. Discovery failed on
those systems, but only when we were already observing the communication
problem with lnet.

Jesse

________________________________________
From: lustre-discuss <lustre-discuss-bounces at lists.lustre.org<mailto:lustre-discuss-bounces at lists.lustre.org>> on behalf of
Nehring, Shane R [LAS] via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>>
Sent: Wednesday, December 21, 2022 11:49 AM
To: lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
Subject: [lustre-discuss] Connectivity issues after client crash

Hello all,

I'm hoping that someone might be able to help me with an issue I've been
seeing
periodically since updating to 2.15.x.

Back in July we remade our scratch storage volume with 2.15 after running 2.12
for a long time. As part of the upgrade we reinstalled our oss and mds nodes
with rhel 8 so we'd finally be able to take advantage of project quotas. For
background, our nodes are connected via omnipath and we have a mix of el7
clients, el8 clients, and el9 clients. We've been piecemeal updating our
clients
to el9 over the past few months with the goal of being 99% el9 clients by mid
January.

Since upgrading we've seen recurring connectivity issues arise in the cluster
from time to time which seem to be very strongly correlated with a client
crashing. The fabric itself seems fine. There's no evidence of error or any
packet loss, so I cannot confidently blame it.

On an oss that's having trouble communicating we see the following messages
for
various osts:

kernel: LNetError: 31282:0:(o2iblnd_cb.c:3358:kiblnd_check_txs_locked()) Timed
out tx: active_txs(WSQ:100), 19 seconds
kernel: LNetError: 31282:0:(o2iblnd_cb.c:3358:kiblnd_check_txs_locked())
Skipped
6 previous similar messages
kernel: LNetError: 31282:0:(o2iblnd_cb.c:3428:kiblnd_check_conns()) Timed out
RDMA with 172.16.100.19 at o2ib (4): c: 31, oc: 0, rc: 31
kernel: LNetError: 31282:0:(o2iblnd_cb.c:3428:kiblnd_check_conns()) Skipped 6
previous similar messages
kernel: LustreError: 109351:0:(ldlm_lib.c:3543:target_bulk_io()) @@@ network
error on bulk WRITE  req at 00000000f848e208 x1751146351599040/t0(0) o4-
10ccde33-01ef-47cf-873d-da9a1b6bb1ea at 172.16.100.19<mailto:10ccde33-01ef-47cf-873d-da9a1b6bb1ea at 172.16.100.19>@o2ib:19/0 lens 488/448 e
0
to 0 dl 1671634949 ref 1 fl Interpret:/2/0 rc 0/0 job:'873032'
kernel: Lustre: work-OST0007: Bulk IO write error with 10ccde33-01ef-47cf-
873d-
da9a1b6bb1ea (at 172.16.100.19 at o2ib), client will retry: rc = -110
kernel: Lustre: Skipped 5 previous similar messages
kernel: LustreError: 109351:0:(ldlm_lib.c:3543:target_bulk_io()) @@@ network
error on bulk WRITE  req at 000000008176189a x1751153771751616/t0(0) o4-
63e88958-d0b1-4c7f-8413-da96b181cd92 at 172.16.100.25<mailto:63e88958-d0b1-4c7f-8413-da96b181cd92 at 172.16.100.25>@o2ib:94/0 lens 488/448 e
0
to 0 dl 1671635024 ref 1 fl Interpret:/0/0 rc 0/0 job:'873641'
Lustre: work-OST000d: Bulk IO write error with 63e88958-d0b1-4c7f-8413-
da96b181cd92 (at 172.16.100.25 at o2ib), client will retry: rc = -110

An already connected client will see messages along the lines of:

kernel: Lustre: 3736:0:(client.c:2295:ptlrpc_expire_one_request()) @@@ Request
sent has timed out for slow reply: [sent 1671634931/real 1671634931]
req at ffff94a0ed490000 x1751139078233344/t0(0) o9-
work-OST0001-osc-ffff95552570a000 at 172.16.100.253<mailto:work-OST0001-osc-ffff95552570a000 at 172.16.100.253>@o2ib:28/4 lens 224/224 e 0
to
1

A stack trace on one of the clients of an ls against the volume that hangs:

[<ffffffffc1576d05>] cl_sync_io_wait+0x1c5/0x480 [obdclass]
[<ffffffffc1572943>] cl_lock_request+0x1d3/0x210 [obdclass]
[<ffffffffc17db1d9>] cl_glimpse_lock+0x329/0x380 [lustre]
[<ffffffffc17db5a5>] cl_glimpse_size0+0x255/0x280 [lustre]
[<ffffffffc1793cdc>] ll_getattr_dentry+0x50c/0x9c0 [lustre]
[<ffffffffc17941ae>] ll_getattr+0x1e/0x20 [lustre]
[<ffffffff99a53d49>] vfs_getattr+0x49/0x80
[<ffffffff99a53e55>] vfs_fstatat+0x75/0xc0
[<ffffffff99a54261>] SYSC_newlstat+0x31/0x60
[<ffffffff99a546ce>] SyS_newlstat+0xe/0x10
[<ffffffff99f99f92>] system_call_fastpath+0x25/0x2a
[<ffffffffffffffff>] 0xffffffffffffffff

If we try to reboot a client to clear the issue it won't be able to mount the
filesystem; the mount process hangs in D until it times out. A stack trace
from
the mount:
[<0>] llog_process_or_fork+0x2de/0x570 [obdclass]
[<0>] llog_process+0x10/0x20 [obdclass]
[<0>] class_config_parse_llog+0x1eb/0x3e0 [obdclass]
[<0>] mgc_process_cfg_log+0x659/0xc90 [mgc]
[<0>] mgc_process_log+0x667/0x800 [mgc]
[<0>] mgc_process_config+0x42b/0x6e0 [mgc]
[<0>] obd_process_config.constprop.0+0x76/0x1a0 [obdclass]
[<0>] lustre_process_log+0x562/0x8f0 [obdclass]
[<0>] ll_fill_super+0x6ec/0x1020 [lustre]
[<0>] lustre_fill_super+0xe4/0x470 [lustre]
[<0>] mount_nodev+0x41/0x90
[<0>] legacy_get_tree+0x24/0x40
[<0>] vfs_get_tree+0x22/0xb0
[<0>] do_new_mount+0x176/0x310
[<0>] __x64_sys_mount+0x103/0x140
[<0>] do_syscall_64+0x38/0x90
[<0>] entry_SYSCALL_64_after_hwframe+0x44/0xae

When this has happened in the past we've had to shut down the entire lustre
backend, and if we're lucky clients would start recovering once everything was
back up again. The last time this happened we actually had to shut down the
entire cluster.

In this particular case rebooting the osses and mds/mgs has stopped the errors
on the server side, but el9 and el8 clients are now seeing:

kernel: LNetError: 149017:0:(o2iblnd_cb.c:966:kiblnd_post_tx_locked()) Error -
22
posting transmit to 172.16.100.250 at o2ib
kernel: LNetError: 149017:0:(o2iblnd_cb.c:966:kiblnd_post_tx_locked()) Skipped
31 previous similar messages
and
LustreError: 3585788:0:(file.c:5096:ll_inode_revalidate_fini()) work:
revalidate
FID [0x200000007:0x1:0x0] error: rc = -4

el7 clients have recovered seemingly without issue.

Wondering if anyone might have any suggestions on where to look as to why this
is breaking, or pointers as to how to recover from this situation without
rebooting if possible.

_______________________________________________
lustre-discuss mailing list
lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org<https://urldefense.com/v3/__http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org__;!!Mak6IKo!KjIS-mlygGvZXA84UL7NpPd6BbzEHT4hP-NTUTL4RmVJ2GjVu_zPrP0oO1MAkx3CWy4hebKZ4MYH7r4O-Xxn04MGbMU$>

_______________________________________________
lustre-discuss mailing list
lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org

Cheers, Andreas
--
Andreas Dilger
Lustre Principal Architect
Whamcloud

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20241211/c4eef9df/attachment-0001.htm>