Virtualization - Jun 2023 - [CFT][PATCH v3] fork, vhost: Use CLONE_THREAD to fix freezer/ps regression

On 6/5/23 10:10 AM, Oleg Nesterov wrote:
> On 06/03, michael.christie at oracle.com wrote:
>>
>> On 6/2/23 11:15 PM, Eric W. Biederman wrote:
>> The problem is that as part of the flush the drivers/vhost/scsi.c code
>> will wait for outstanding commands, because we can't free the
device and
>> it's resources before the commands complete or we will hit the
accessing
>> freed memory bug.
> 
> ignoring send-fd/clone issues, can we assume that the final fput/release
> should always come from vhost_worker's sub-thread (which shares mm/etc)
?
I think I'm misunderstanding the sub-thread term.

- Is it the task_struct's context that we did the
kernel/vhost_taskc.c:vhost_task_create() from? Below it would be the
thread we did VHOST_SET_OWNER from.

If so, then yes.

- Is it the task_struct that gets created by
kernel/vhost_taskc.c:vhost_task_create()?

If so, then the answer is no. vhost_task_create has set the no_files
arg on kernel_clone_args, so copy_files() sets task_struct->files to NULL
and we don't clone or dup the files.

So it works like if we were using a kthread still:

1. Userapce thread0 opens /dev/vhost-$something.
2. thread0 does VHOST_SET_OWNER ioctl. This calls vhost_task_create() to
create the task_struct which runs the vhost_worker() function which handles
the work->fns.
3. If userspace now does a SIGKILL or just exits without doing a close() on
/dev/vhost-$something, then when thread0 does exit_files() that will do the
fput that does vhost-$something's file_operations->release.

On 06/05, Mike Christie wrote:
>
> On 6/5/23 10:10 AM, Oleg Nesterov wrote:
> > On 06/03, michael.christie at oracle.com wrote:
> >>
> >> On 6/2/23 11:15 PM, Eric W. Biederman wrote:
> >> The problem is that as part of the flush the drivers/vhost/scsi.c
code
> >> will wait for outstanding commands, because we can't free the
device and
> >> it's resources before the commands complete or we will hit the
accessing
> >> freed memory bug.
> >
> > ignoring send-fd/clone issues, can we assume that the final
fput/release
> > should always come from vhost_worker's sub-thread (which shares
mm/etc) ?
>
> I think I'm misunderstanding the sub-thread term.
>
> - Is it the task_struct's context that we did the
> kernel/vhost_taskc.c:vhost_task_create() from? Below it would be the
> thread we did VHOST_SET_OWNER from.
Yes,
> So it works like if we were using a kthread still:
>
> 1. Userapce thread0 opens /dev/vhost-$something.
> 2. thread0 does VHOST_SET_OWNER ioctl. This calls vhost_task_create() to
> create the task_struct which runs the vhost_worker() function which handles
> the work->fns.
> 3. If userspace now does a SIGKILL or just exits without doing a close() on
> /dev/vhost-$something, then when thread0 does exit_files() that will do the
> fput that does vhost-$something's file_operations->release.
So, at least in this simple case vhost_worker() can just exit after SIGKILL,
and thread0 can flush the outstanding commands when it calls vhost_dev_flush()
rather than wait for vhost_worker().

Right?

not that I think this can help in the general case ...

Oleg.