Re: workqueue lockup debug

[John Garry <john.g.garry@oracle.com>]

On 29/10/2024 04:11, Dave Chinner wrote:
>> Can you advise on a robust method to get some debug from this system?
>>
>> Maybe this is a scheduler issue, as Dave mentioned in that same thread.
> I just got a new hit on the scheduler issue on 6.12-rc5:
> 
> [  172.477662] ------------[ cut here ]------------
> [  172.480660] se->on_rq
> [  172.480682] WARNING: CPU: 3 PID: 728318 at kernel/sched/fair.c:5629 pick_task_fair+0xb6/0x1b0
> [  172.487172] Modules linked in:
> [  172.488911] CPU: 3 UID: 0 PID: 728318 Comm: 291 Not tainted 6.12.0-rc5-dgc+ #273
> [  172.492022] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> [  172.496029] RIP: 0010:pick_task_fair+0xb6/0x1b0
> [  172.497978] Code: 40 f3 f3 03 01 74 2a 41 80 7f 50 00 74 96 f6 05 31 f3 f3 03 01 75 8d c6 05 28 f3 f3 03 01 48 c7 c7 41 26 e4 82 e8 7a d7 fa ff <0f> 0b e9 73 ff ff ff c6 05 0d f3 f3 03 01 48 c7 c7 2f 26 e4 82 e8
> [  172.505868] RSP: 0018:ffffc9002ec83d60 EFLAGS: 00010046
> [  172.508111] RAX: bdcdd05ee831a400 RBX: 0000000000000000 RCX: 0000000000000027
> [  172.511139] RDX: 0000000000000000 RSI: 00000000ffdfffff RDI: ffff88881fadc9c8
> [  172.514326] RBP: ffffc9002ec83d88 R08: 00000000001fffff R09: ffff88a018a00000
> [  172.517804] R10: 00000000005ffffd R11: 0000000000000004 R12: ffff888122b58c00
> [  172.520788] R13: ffff88881faf0140 R14: ffff88881faf00c0 R15: ffff88819e8cd180
> [  172.523748] FS:  0000000000000000(0000) GS:ffff88881fac0000(0000) knlGS:0000000000000000
> [  172.527167] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [  172.529571] CR2: 00007f378e8866fc CR3: 000000104daae000 CR4: 0000000000350ef0
> [  172.532601] Call Trace:
> [  172.533666]  <TASK>
> [  172.534557]  ? show_regs+0x5e/0x70
> [  172.535993]  ? __warn+0xd0/0x1d0
> [  172.537430]  ? pick_task_fair+0xb6/0x1b0
> [  172.539062]  ? report_bug+0x145/0x1f0
> [  172.540619]  ? handle_bug+0x67/0x90
> [  172.542060]  ? exc_invalid_op+0x1b/0x50
> [  172.543854]  ? asm_exc_invalid_op+0x1b/0x20
> [  172.545688]  ? pick_task_fair+0xb6/0x1b0
> [  172.547353]  pick_next_task_fair+0x27/0x330
> [  172.549090]  __schedule+0x2ad/0xb10
> [  172.550460]  do_task_dead+0x43/0x50
> [  172.551743]  do_exit+0x836/0xab0
> [  172.552950]  do_group_exit+0x8f/0x90
> [  172.554274]  __x64_sys_exit_group+0x17/0x20
> [  172.555808]  x64_sys_call+0x2ed3/0x2ee0
> [  172.557210]  do_syscall_64+0x68/0x130
> [  172.558531]  ? exc_page_fault+0x62/0xc0
> [  172.559907]  entry_SYSCALL_64_after_hwframe+0x76/0x7e
> [  172.561739] RIP: 0033:0x7f378e74b3c5
> [  172.563036] Code: Unable to access opcode bytes at 0x7f378e74b39b.
> [  172.565278] RSP: 002b:00007ffd21861488 EFLAGS: 00000206 ORIG_RAX: 00000000000000e7
> [  172.568001] RAX: ffffffffffffffda RBX: 00007f378e857fa8 RCX: 00007f378e74b3c5
> [  172.570589] RDX: 00000000000000e7 RSI: ffffffffffffff88 RDI: 0000000000000000
> [  172.573182] RBP: 0000000000000000 R08: 00007ffd21861428 R09: 00007ffd218613af
> [  172.575758] R10: 00007ffd21861320 R11: 0000000000000206 R12: 00007f378e856680
> [  172.578356] R13: 00007f378e8b9f40 R14: 0000000000000001 R15: 00007f378e857fc0
> [  172.580941]  </TASK>
> 
> There was nothing particularly obvious that caused it. It's a 64p
> VM, running a heavy concurrent fstests workload that takes about 20
> minutes to complete. There are a bit over 8000 tasks reported, most
> of them kernel threads. The load is consuming about 50 CPUs, 40GB of
> RAM, sustaining ~3GB/s of IO to/from disk across about 130 block
> devices and the scheduler is context switching around 800,000 times
> a second.
> 
> I have no idea how to reproduce this on demand - it randomly shows
> up in about 1 in every 10 test runs and all the tests running at
> that point in time report failure because they all detect this
> warning in dmesg.

I have not been able to test v6.12-rc5 yet.

Can you share your config? And/Or compare to mine at 
https://pastebin.com/Y7DXnMG2

Maybe we can see what was merged for the sched in 6.12, and I can try 
before and after. I don't like the prospect of a full bisect.

Thanks,
John