lkml.org 
[lkml]   [2023]   [Jan]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/2] vhost: check for pending livepatches from vhost worker kthreads
On Fri 2023-01-20 16:12:22, Seth Forshee (DigitalOcean) wrote:
> Livepatch relies on stack checking of sleeping tasks to switch kthreads,
> so a busy kthread can block a livepatch transition indefinitely. We've
> seen this happen fairly often with busy vhost kthreads.

To be precise, it would be "indefinitely" only when the kthread never
sleeps.

But yes. I believe that the problem is real. It might be almost
impossible to livepatch some busy kthreads, especially when they
have a dedicated CPU.


> Add a check to call klp_switch_current() from vhost_worker() when a
> livepatch is pending. In testing this allowed vhost kthreads to switch
> immediately when they had previously blocked livepatch transitions for
> long periods of time.
>
> Signed-off-by: Seth Forshee (DigitalOcean) <sforshee@kernel.org>
> ---
> drivers/vhost/vhost.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index cbe72bfd2f1f..d8624f1f2d64 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -366,6 +367,9 @@ static int vhost_worker(void *data)
> if (need_resched())
> schedule();
> }
> +
> + if (unlikely(klp_patch_pending(current)))
> + klp_switch_current();

I suggest to use the following intead:

if (unlikely(klp_patch_pending(current)))
klp_update_patch_state(current);

We already use this in do_idle(). The reason is basically the same.
It is almost impossible to livepatch the idle task when a CPU is
very idle.

klp_update_patch_state(current) does not check the stack.
It switches the task immediately.

It should be safe because the kthread never leaves vhost_worker().
It means that the same kthread could never re-enter this function
and use the new code.

Best Regards,
Petr

\
 
 \ /
  Last update: 2023-03-26 23:54    [W:0.115 / U:0.196 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site