On Wed, May 26, 2021 at 10:17 AM Mike Kravetz [off-list ref] wrote:

On 5/25/21 8:19 PM, Muchun Song wrote:

quoted

On Wed, May 26, 2021 at 7:31 AM Mike Kravetz [off-list ref] wrote:

quoted

Here is a modification to the reservation tracking for fixup on errors.
It is a more general change, but should work for the hugetlb_mcopy_pte_atomic
case as well.

Perhaps use this as a prerequisite for your fix(es)?  Pretty sure this
will eliminate the need for the call to hugetlb_unreserve_pages.

Hi Mike,

It seems like someone is fixing a bug, right? Maybe a link should be
placed in the cover letter so that someone can know what issue
we are facing.

Thanks Muchun,

I wanted to first see if these patches would work in the code Mina is
modifying.  If this works for Mina, then a more formal patch and request
for inclusion will be sent.

So a quick test: I apply my patche and yours on top of linus/master,
and I remove the hugetlb_unreserve_pages() call that triggered this
conversation, and run the userfaultfd test, resv_huge_pages underflows
again, so it seems on the surface this doesn't quite work as is.

Not quite sure what to do off the top of my head. I think I will try
to debug why the 3 patches don't work together and I will fix either
your patch or mine. I haven't taken a deep look yet; I just ran a
quick test.

I believe this issue has existed since the introduction of hugetlb
reservations in v2.6.18.  Since the bug only shows up when we take error
paths, the issue may not have been observed.  Mina found a similar issue
in an error path which could also expose this issue.
--
Mike Kravetz