[lustre-discuss] LDLM locks not expiring/cancelling

[Moreno Diego (ID SIS)]

Hi Steve,

I was having a similar problem in the past months where the MDS servers would go OOM because of SlabUnreclaim. The root cause has not yet been found but we stopped seeing this the day we disabled the NRS TBF (QoS) for any LDLM service (just in case you have it enabled). Just in case you have it enabled. It would be good to check as well what’s being consumed in the slab cache. In our case it was mostly kernel objects and not ldlm.

Diego


From: lustre-discuss <lustre-discuss-bounces at lists.lustre.org> on behalf of Steve Crusan <stevec at dug.com>
Date: Thursday, 2 January 2020 at 20:25
To: "lustre-discuss at lists.lustre.org" <lustre-discuss at lists.lustre.org>
Subject: [lustre-discuss] LDLM locks not expiring/cancelling

Hi all,

We are running into a bizarre situation where we aren't having stale locks cancel themselves, and even worse, it seems as if ldlm.namespaces.*.lru_size is being ignored.

For instance, I unmount our Lustre file systems on a client machine, then remount. Next, I'll run "lctl set_param ldlm.namespaces.*.lru_max_age=60s, lctl set_param ldlm.namespaces.*.lru_size=1024". This (I believe) theoretically would only allow 1024 ldlm locks per osc, and then I'd see a lot of lock cancels (via ldlm.namespaces.${ost}.pool.stats). We also should see cancels if the grant time > lru_max_age.

We can trigger this simply by running 'find' on the root of our Lustre file system, and waiting for awhile. Eventually the clients SUnreclaim value bloats to 60-70GB (!!!), and each of our OSTs have 30-40k LRU locks (via lock_count). This is early in the process:

"""
ldlm.namespaces.h5-OST003f-osc-ffff8802d8559000.lock_count=2090
ldlm.namespaces.h5-OST0040-osc-ffff8802d8559000.lock_count=2127
ldlm.namespaces.h5-OST0047-osc-ffff8802d8559000.lock_count=52
ldlm.namespaces.h5-OST0048-osc-ffff8802d8559000.lock_count=1962
ldlm.namespaces.h5-OST0049-osc-ffff8802d8559000.lock_count=1247
ldlm.namespaces.h5-OST004a-osc-ffff8802d8559000.lock_count=1642
ldlm.namespaces.h5-OST004b-osc-ffff8802d8559000.lock_count=1340
ldlm.namespaces.h5-OST004c-osc-ffff8802d8559000.lock_count=1208
ldlm.namespaces.h5-OST004d-osc-ffff8802d8559000.lock_count=1422
ldlm.namespaces.h5-OST004e-osc-ffff8802d8559000.lock_count=1244
ldlm.namespaces.h5-OST004f-osc-ffff8802d8559000.lock_count=1117
ldlm.namespaces.h5-OST0050-osc-ffff8802d8559000.lock_count=1165
"""

But this will grow over time, and eventually this compute node gets evicted from the MDS (after 10 minutes of cancelling locks/hanging). The only way we have been able to reduce the slab usage is to drop caches and set LRU=clear...but the problem just comes back depending on the workload.

We are running 2.10.3 client side, 2.10.1 server side. Have there been any fixes added into the codebase for 2.10 that we need to apply? This seems to be the closest to what we are experiencing:

https://jira.whamcloud.com/browse/LU-11518


PS: I've checked other systems across our cluster, and some of them have as many as 50k locks per OST. I am kind of wondering if these locks are staying around much longer than the lru_max_age default (65 minutes), but I cannot prove that. Is there a good way to translate held locks to fids? I have been messing around with lctl set_param debug="XXX" and lctl set_param ldlm.namespaces.*.dump_namespace, but I don't feel like I'm getting *all* of the locks.

~Steve
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20200106/fedc80f2/attachment.html>