Virtualization - May 2023 - [PATCH 3/3] fork, vhost: Use CLONE_THREAD to fix freezer/ps regression

Oleg Nesterov <oleg at redhat.com> writes:
> On 05/23, Eric W. Biederman wrote:
>>
>> I want to point out that we need to consider not just SIGKILL, but
>> SIGABRT that causes a coredump, as well as the process peforming
>> an ordinary exit(2).  All of which will cause get_signal to return
>> SIGKILL in this context.
>
> Yes, but probably SIGABRT/exit doesn't really differ from SIGKILL wrt
> vhost_worker().
Actually I think it reveals that exiting with SIGABRT will cause
a deadlock.

coredump_wait will wait for all of the threads to reach
coredump_task_exit.  Meanwhile vhost_worker is waiting for
all of the other threads to reach exit_files to close their
file descriptors.


So it looks like the final pieces of work will actually need to be moved
into to vhost_xxx_flush or vhost_xxx_release to avoid the exiting
threads from waiting on each other, instead of depending upon the
vhost_worker to do the work.

Which gets back to most of your other questions.
>> It is probably not the worst thing in the world, but what this means
>> is now if you pass a copy of the vhost file descriptor to another
>> process the vhost_worker will persist, and thus the process will
persist
>> until that copy of the file descriptor is closed.
>
> Hadn't thought about it.
>
> I am fighting with internal bugzillas today, will try to write another
> email tomorrow.
>
> But before that, I would like to have an answer to my "main"
question in
> my previois email. Otherwise I am still not sure I understand what exactly
> we need to fix.
Let me repeat your "main" question just for clarity here.

If a signal comes in after the signal_pending check but before the
"work->fn(work)" call is "work->fn(work)" expected to
run correctly
with signal_pending() or fatal_signal_pending returning true?


Eric

On 05/24, Eric W. Biederman wrote:
>
> Oleg Nesterov <oleg at redhat.com> writes:
>
> > Yes, but probably SIGABRT/exit doesn't really differ from SIGKILL
wrt
> > vhost_worker().
>
> Actually I think it reveals that exiting with SIGABRT will cause
> a deadlock.
>
> coredump_wait will wait for all of the threads to reach
> coredump_task_exit.  Meanwhile vhost_worker is waiting for
> all of the other threads to reach exit_files to close their
> file descriptors.
Indeed, I didn't think about this.


So why do we actually need CLONE_THREAD ? Can't vhost_worker() be a kernel
thread?

kthread_create() won't be convenient, but how about kernel_thread() ? it
inherits
mm/cgroups/rlimits/etc, kthread_stop() should work just fine.

Oleg.