On Tue, May 11, 2021 at 9:38 AM Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote: > > On 5/10/21 11:45 PM, Axel Rasmussen wrote: > > Thanks for the investigation, Mike! > > > > Mina, since hugetlb_mcopy_atomic_pte is specific to userfaultfd, I'm > > happy to take a deeper look at it this week as well. > > > > For context, we have seen the WARN_ON Mina described trigger in > > production before, but were never able to reproduce it. The > > userfaultfd self test turns out to reproduce it reliably, so the > > thinking up to this point was that it just happened to reproduce some > > non-userfaultfd-specific issue. But from Mike's description, it seems > > this bug is very specific to userfaultfd after all. :) > > Certainly, this case is userfaultfd specific. > > However, I too recall seeing 'transient' underflows in the past. Pretty > sure this was not userfaultfd specific. Specifically, when working on > commit 22146c3ce989 "hugetlbfs: dirty pages as they are added to pagecache" > I recall seeing transient underflow. After fixing the issue in 22146c3ce989, > I could not reproduce transient underflows and stopped looking for the > cause. We added code to a production kernel in an attmempt to catch > the issue: > https://github.com/oracle/linux-uek/commit/bd697676290d91762ef2bf79832f653b44e6f83b#diff-fb6066ca63d9afdc2e3660c85e2dcc04cf31b37900ca9df2a1019ee8fa80dce0 > We also have local changes that detect resv_huge_pages underflowing, and they trigger in production every few months and we don't use userfaultfd in production. We haven't been able to reproduce these issues though, and we run an older version of the kernel so I haven't had anything useful to share until now. I'm (quite tentatively to be honest) considering proposing adding some of these WARN_ONCE_ON() to the kernel after I fix this issue and soliciting reproducers from the community to fix these underflows. I'm hesitant because I'm not confident I can fix the ensuing bug reports in a timely manner and the volume of reports may be large. But we can cross that bridge when this issue is figured out :) > I'll try running some non-userfaultfd specific tests to see if I can > reproduce. > -- > Mike Kravetz
