[lkml]   [2014]   [Mar]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [v3.13][v3.14][Regression] kthread:makekthread_create()killable
On 03/19/2014 07:49 AM, Tetsuo Handa wrote:
> Oleg Nesterov wrote:
>>>> If we need the urgent hack to fix the regression, then I suggest to change
>>>> scsi_host_alloc() temporary until mptsas (or whatever) is fixed.
>>> Device initialization taking longer than 30 seconds is possible and is not a
>>> hang up. It is systemd which needs to be fixed.
>> Perhaps systemd needs the fix too, I do not know. But this is irrelevant,
>> I think. Or at least this should be discussed separately.
> I confirmed that this problem goes away if systemd-udevd supports longer
> timeout.
>> kthread_run() can fail anyway, mptsas_probe() should not crash the kernel.
> Right. But mptsas_probe() triggering an OOPS is irrelevant to kthread_run()
> ( comment #27 ).
>> And btw, it is not clear to me if in this case device initialization really
>> needs more than 30 seconds... My understanding is probably wrong, so please
>> correct me. But it seems that before your "make kthread_create() killable"
>> - probe hangs
>> - SIGKILL wakes it up
>> - so I assume that the probe was interrupted and didn't finish
>> correctly ???
>> - initialization continues, does scsi_host_alloc(), etc, and
>> everything works fine even if probe was interrupted?
> I confirmed that device initialization really took more than 30 seconds
> ( comments #51 and #52 ).
>> So perhaps that probe should not hang and this should be fixed too ?
>> Do you know where exactly it hangs? And where it is woken up by SIGKILL ?
>> Or I totally misunderstood ?
> The probe did not hang. SIGKILL affected only wait_for_completion_killable()
> in kthread_create_on_node() called by mptsas_probe() via scsi_host_alloc().
> Thus, the probe was interrupted because kthread_run() returned an error.
>>> I think we need a bit different version, in order to take TIF_MEMDIE flag into
>>> account at the caller of kthread_create(), for the purpose of commit 786235ee
>>> is "try to die as soon as possible if chosen by the OOM killer".
>>> for (;;) {
>>> shost->ehandler = kthread_run(scsi_error_handler, shost,
>>> "scsi_eh_%d", shost->host_no);
>>> if (PTR_ERR(shost->ehandler) != -EINTR ||
>>> test_thread_flag(TIF_MEMDIE))
>> Well, personally I do not care about TIF_MEMDIE/oom at all. We need the
>> temporary hack (unless we have the "right" fix right now) which should be
>> reverted later.
> I do seriously care about TIF_MEMDIE/oom. Last week I respond to a trouble
> which hit "kernel: request_module() OOM local DoS" (RHBZ #853474) without
> any malice.
>> Not sure I understand... Yes, wait_for_completion_killable() can return
>> immediately if TIF_SIGPENDING will be set again for any reason. Say, another
>> signal. But the next iteration will clear TIF_SIGPENDING ?
>>> As I think it is difficult to prove that kmalloc(GFP_KERNEL) never sets
>> Ah, I see, you mean that kmalloc() can do this every time. No, this should
>> not happen or we have another problem.
> Then, what happens if somebody does
> while (1)
> kill(pid, SIGKILL);
> where pid is the process calling kthread_run() from the "for (;;)" loop in
> scsi_host_alloc()? Theoretically, it will form an infinite retry loop.
> Clearing TIF_SIGPENDING does not guarantee that next
> wait_for_completion_killable() does not return immediately.
> Doing retry decision at scsi_host_alloc() will make things worse than
> doing it at kthread_create_on_node().
>> Anyway. I agree with any hack in scsi_host_alloc/etc, this is up to
>> maintainers. I still think that your change uncovered the problems in
>> drivers/message/fusion/, these problems should be fixed somehow.
>> Dear maintainers, we need your help.
> Right. We found that we can fix this problem by updating systemd-udevd to
> support longer timeout ( comment #53 ). Joseph, would you consult systemd
> maintainers?

Thanks everyone for reviewing this bug. Message sent to systemd mailing

 \ /
  Last update: 2014-03-19 22:41    [W:0.110 / U:3.300 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site