Re: [PATCH 0/9] vhost: Support SIGKILL by flushing and exiting

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Apr 09, 2024 at 12:16:36PM +0800, Jason Wang wrote:
> On Sat, Mar 16, 2024 at 8:47 AM Mike Christie
> <michael.christie@oracle.com> wrote:
> >
> > The following patches were made over Linus's tree and also apply over
> > mst's vhost branch. The patches add the ability for vhost_tasks to
> > handle SIGKILL by flushing queued works, stop new works from being
> > queued, and prepare the task for an early exit.
> >
> > This removes the need for the signal/coredump hacks added in:
> >
> > Commit f9010dbdce91 ("fork, vhost: Use CLONE_THREAD to fix freezer/ps regression")
> >
> > when the vhost_task patches were initially merged and fix the issue
> > in this thread:
> >
> > https://lore.kernel.org/all/000000000000a41b82060e875721@google.com/
> >
> > Long Background:
> >
> > The original vhost worker code didn't support any signals. If the
> > userspace application that owned the worker got a SIGKILL, the app/
> > process would exit dropping all references to the device and then the
> > file operation's release function would be called. From there we would
> > wait on running IO then cleanup the device's memory.
> >
> > When we switched to vhost_tasks being a thread in the owner's process we
> > added some hacks to the signal/coredump code so we could continue to
> > wait on running IO and process it from the vhost_task. The idea was that
> > we would eventually remove the hacks. We recently hit this bug:
> >
> > https://lore.kernel.org/all/000000000000a41b82060e875721@google.com/
> >
> > It turns out only vhost-scsi had an issue where it would send a command
> > to the block/LIO layer, wait for a response and then process in the vhost
> > task.
> 
> Vhost-net TX zerocopy code did the same:
> 
> It sends zerocopy packets to the under layer and waits for the
> underlayer. When the DMA is completed, vhost_zerocopy_callback will be
> called to schedule vq work for used ring updating.

Yea. It's still experimental though so I'm not sure how
stressed to be about it. I guess we can ignore it for now -
but yes it was one of the big issues with tx zerocopy
and this patchset opens the path to productizing it.

> > So patches 1-5 prepares vhost-scsi to handle when the vhost_task
> > is killed while we still have commands outstanding. The next patches then
> > prepare and convert the vhost and vhost_task layers to handle SIGKILL
> > by flushing running works, marking the vhost_task as dead so there's
> > no future uses, then exiting.
> 
> Thanks
> 
> >
> >
> >