Re: [Cluster-devel] [PATCH] DLM: increase socket backlog to avoid hangs with 16 nodes

From: Alexander Aring
Date: Thu Feb 02 2023 - 17:26:09 EST

Hi,

On Thu, Feb 2, 2023 at 10:37 AM Edwin Török <edvin.torok@xxxxxxxxxx> wrote:
>
> On a 16 node virtual cluster with e1000 NICs joining the 12th node prints
> SYN flood warnings for the DLM port:
> Dec 21 01:46:41 localhost kernel: [ 2146.516664] TCP: request_sock_TCP: Possible SYN flooding on port 21064. Sending cookies. Check SNMP counters.
>
> And then joining a DLM lockspace hangs:
> ```
> Dec 21 01:49:00 localhost kernel: [ 2285.780913] INFO: task xapi-clusterd:17638 blocked for more than 120 seconds. │
> Dec 21 01:49:00 localhost kernel: [ 2285.786476] Not tainted 4.4.0+10 #1 │
> Dec 21 01:49:00 localhost kernel: [ 2285.789043] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. │
> Dec 21 01:49:00 localhost kernel: [ 2285.794611] xapi-clusterd D ffff88001930bc58 0 17638 1 0x00000000 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794615] ffff88001930bc58 ffff880025593800 ffff880022433800 ffff88001930c000 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794617] ffff88000ef4a660 ffff88000ef4a658 ffff880022433800 ffff88000ef4a000 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794619] ffff88001930bc70 ffffffff8159f6b4 7fffffffffffffff ffff88001930bd10
> Dec 21 01:49:00 localhost kernel: [ 2285.794644] [<ffffffff811570fe>] ? printk+0x4d/0x4f │
> Dec 21 01:49:00 localhost kernel: [ 2285.794647] [<ffffffff810b1741>] ? __raw_callee_save___pv_queued_spin_unlock+0x11/0x20 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794649] [<ffffffff815a085d>] wait_for_completion+0x9d/0x110 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794653] [<ffffffff810979e0>] ? wake_up_q+0x80/0x80 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794661] [<ffffffffa03fa4b8>] dlm_new_lockspace+0x908/0xac0 [dlm] │
> Dec 21 01:49:00 localhost kernel: [ 2285.794665] [<ffffffff810aaa60>] ? prepare_to_wait_event+0x100/0x100 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794670] [<ffffffffa0402e37>] device_write+0x497/0x6b0 [dlm] │
> Dec 21 01:49:00 localhost kernel: [ 2285.794673] [<ffffffff811834f0>] ? handle_mm_fault+0x7f0/0x13b0 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794677] [<ffffffff811b4438>] __vfs_write+0x28/0xd0 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794679] [<ffffffff811b4b7f>] ? rw_verify_area+0x6f/0xd0 ┤
> Dec 21 01:49:00 localhost kernel: [ 2285.794681] [<ffffffff811b4dc1>] vfs_write+0xb1/0x190 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794686] [<ffffffff8105ffc2>] ? __do_page_fault+0x302/0x420 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794688] [<ffffffff811b5986>] SyS_write+0x46/0xa0 │
> Dec 21 01:49:00 localhost kernel: [ 2285.794690] [<ffffffff815a31ae>] entry_SYSCALL_64_fastpath+0x12/0x71
> ```
>
> The previous limit of 5 seems like an arbitrary number, that doesn't match any
> known DLM cluster size upper bound limit.
>
> Signed-off-by: Edwin Török <edvin.torok@xxxxxxxxxx>
> Cc: Christine Caulfield <ccaulfie@xxxxxxxxxx>
> Cc: David Teigland <teigland@xxxxxxxxxx>
> Cc: cluster-devel@xxxxxxxxxx
> ---
> Notes from 2023:
> This patch was initially developed on 21 Dec 2017, and in production use ever since.
> I expected to drop out of our patchqueue at the next kernel upgrade, however it
> hasn't, so I probably forgot to send it.
>
> I haven't noticed this bug again with the patch applied, and the previous value
> of '5' seems like an arbitrary limit not matching any supported upper bounds
> on DLM cluster sizes, so this patch has (unintentionally) had a 5 year test
> cycle.

Correct, I guess the 128 coming from dlm_controld 128 max node
boundary [0] and I think it's okay to change it to 128, especially if
you start a cluster every "mostly" node wants to connect() at the same
time and the backlog of 5 could be not enough to handle them.
>From my understanding the connect() will return -ECONNREFUSED in the
case of backlog is full. As I see the code of kernel v4.0 dlm at this
point will _not_ slow down a reattempt to run connect() again, see
[1].

>
> Although the join hanging forever like that may still be a bug, if the SYN cookies
> consistently trigger it lets try to avoid the bug by avoiding the SYN cookies.

right, it should work even with 1 as a backlog... after digging into
the code. My guess is that [2] ran into the MAX_CONNECT_RETRIES
condition and we do _not_ reset con->retries to 0 if connect() [3]
fails. If this happens we are kind of stuck and we would not try to
connect() again. Means you run into the hung task situation, because
e.g. ping_members() will not succeed.

The current upstream code does not have any limitations of trying
retries and will always wait some time to try a connect() again if it
fails. It should retry to connect so long the cluster manager tells
dlm kernel that a specific node is still part of a lockspace resource
and there is a transmission happening to this particular node. However
I think pumping this value to the maximum amount of possible nodes in
a cluster is not a bad idea.

There is still a question of "why seeing SYN flooding" warning, but as
I said there are a lot of connect() happening at the beginning at
mostly the same time starting a cluster and some of them retrying
connect() fast.

...

- Alex

[0] https://pagure.io/dlm/blob/c1e2ee574794e0dea68fd45783991ef7b1835901/f/dlm_controld/dlm_daemon.h#_177
[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/dlm/lowcomms.c?h=v4.0#n1226
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/dlm/lowcomms.c?h=v4.0#n1168
[3] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/dlm/lowcomms.c?h=v4.0#n1220