On Thu, Jun 25, 2020 at 12:13 PM Kent Gibson [off-list ref] wrote:

On Thu, Jun 25, 2020 at 11:44:21AM +0300, Andy Shevchenko wrote:

quoted

On Thu, Jun 25, 2020 at 1:58 AM Kent Gibson [off-list ref] wrote:

quoted

On Wed, Jun 24, 2020 at 11:57:14PM +0800, Kent Gibson wrote:

...

quoted

quoted

Perhaps you are referring to the case where the copy_to_user fails?

Yes.

quoted

To be honest I considered that to be so unlikely that I ignored it.
Is there a relevant failure mode that I'm missing?

The traditional question for such cases is "what can possibly go wrong?"
I wouldn't underestimate the probability of failure.

The worst case is the watch is enabled and the userspace gets an
EFAULT so it thinks it failed.  If userspace retries then they get
EBUSY, so userspace accounting gets muddled.

We can clear the watch bit if the copy_to_user fails - before
returning the EFAULT. Would that be satisfactory?

Perhaps. I didn't check that scenario.

Back to the failure, is it possible for the copy_to_user fail here,
given that the corresponding copy_from_user has succeeded?

Of course. The general rule is  if on SMP system you have not strongly
serialized sequence of calls (means no preemption, no interrupts, etc)
anything can happen in between.

If so, can that be manually triggered for test purposes?

Unfortunately not an expert in mm, no idea, sorry.

-- 
With Best Regards,
Andy Shevchenko