> On Oct 23, 2024, at 3:27 PM, Benoît Gschwind <benoit.gschwind@xxxxxxxxxxxxxxxxx> wrote: > > Hello, > > I have a nfs server using debian 11 (Linux hostname 6.1.0-25-amd64 #1 > SMP PREEMPT_DYNAMIC Debian 6.1.106-3 (2024-08-26) x86_64 GNU/Linux) > > In some heavy workload some nfsd goes in D state and seems to never > leave this state. I did a python script to monitor how long a process > stay in particular state and I use it to monitor nfsd state. I get the > following result : > > [...] > 178056 I (idle) 0:25:24.475 [nfsd] > 178057 I (idle) 0:25:24.475 [nfsd] > 178058 I (idle) 0:25:24.475 [nfsd] > 178059 I (idle) 0:25:24.475 [nfsd] > 178060 I (idle) 0:25:24.475 [nfsd] > 178061 I (idle) 0:25:24.475 [nfsd] > 178062 I (idle) 0:24:15.638 [nfsd] > 178063 I (idle) 0:24:13.488 [nfsd] > 178064 I (idle) 0:24:13.488 [nfsd] > 178065 I (idle) 0:00:00.000 [nfsd] > 178066 I (idle) 0:00:00.000 [nfsd] > 178067 I (idle) 0:00:00.000 [nfsd] > 178068 I (idle) 0:00:00.000 [nfsd] > 178069 S (sleeping) 0:00:02.147 [nfsd] > 178070 S (sleeping) 0:00:02.147 [nfsd] > 178071 S (sleeping) 0:00:02.147 [nfsd] > 178072 S (sleeping) 0:00:02.147 [nfsd] > 178073 S (sleeping) 0:00:02.147 [nfsd] > 178074 D (disk sleep) 1:29:25.809 [nfsd] > 178075 S (sleeping) 0:00:02.147 [nfsd] > 178076 S (sleeping) 0:00:02.147 [nfsd] > 178077 S (sleeping) 0:00:02.147 [nfsd] > 178078 S (sleeping) 0:00:02.147 [nfsd] > 178079 S (sleeping) 0:00:02.147 [nfsd] > 178080 D (disk sleep) 1:29:25.809 [nfsd] > 178081 D (disk sleep) 1:29:25.809 [nfsd] > 178082 D (disk sleep) 0:28:04.444 [nfsd] > > All process not shown are in idle state. Columns are the following: > PID, state, state name, amoung of time the state did not changed and > the process was not interrupted, and /proc/PID/status Name entry. > > As you can read some nfsd process are in disk sleep state since more > than 1 hour, but looking at the disk activity, there is almost no I/O. > > I tried to restart nfs-server but I get the following error from the > kernel: > > oct. 23 11:59:49 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:49 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:49 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:49 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:49 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:59 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:59 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:59 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:59 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 11:59:59 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 12:00:09 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 12:00:09 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 12:00:09 hostname kernel: rpc-srv/tcp: nfsd: got error -104 when sending 20 bytes - shutting down socket > oct. 23 12:00:10 hostname kernel: rpc-srv/tcp: nfsd: got error -32 when sending 20 bytes - shutting down socket > oct. 23 12:00:10 hostname kernel: rpc-srv/tcp: nfsd: got error -32 when sending 20 bytes - shutting down socket > > The only way to recover seems to reboot the kernel. I guess because the > kernel force the reboot after a given timeout. > > My setup involve in order : > - scsi driver > - mdraid on top of scsi (raid6) > - btrfs ontop of mdraid > - nfsd ontop of btrfs > > > The setup is not very fast as expected, but it seems that in some > situation nfsd never leave the disk sleep state. the exports options > are: gss/krb5i(rw,sync,no_wdelay,no_subtree_check,fsid=XXXXX). The > situation is not commun but it's always happen at some point. For > instance in the case I report here, my server booted the 2024-10-01 and > was stuck about the 2024-10-23. I did reduced by a large amount the > frequency of issue by using no_wdelay (I did thought that I did solved > the issue when I started to use this option). > > My guess is hadware bug, scsi bug, btrfs bug or nfsd bug ? > > Any clue on this topic or any advice is wellcome. Generate stack traces for each process on the system using "sudo echo t > /proc/sysrq-trigger" and then examine the output in the system journal. Note the stack contents for the processes that look stuck. -- Chuck Lever
