On 16.10.20 01:14, Mike Kravetz wrote: > On 10/14/20 11:31 AM, Mike Kravetz wrote: >> On 10/14/20 11:18 AM, David Hildenbrand wrote: >> >> FWIW - I ran libhugetlbfs tests which do a bunch of hole punching >> with (and without) hugetlb controller enabled and did not see this issue. >> > > I took a closer look after running just the fallocate_stress test > in libhugetlbfs. Here are the cgroup counter values: > > hugetlb.2MB.failcnt 0 > hugetlb.2MB.limit_in_bytes 9223372036854771712 > hugetlb.2MB.max_usage_in_bytes 209715200 > hugetlb.2MB.rsvd.failcnt 0 > hugetlb.2MB.rsvd.limit_in_bytes 9223372036854771712 > hugetlb.2MB.rsvd.max_usage_in_bytes 601882624 > hugetlb.2MB.rsvd.usage_in_bytes 392167424 > hugetlb.2MB.usage_in_bytes 0 > > We did not hit the WARN_ON_ONCE(), but the 'rsvd.usage_in_bytes' value > is not correct in that it should be zero. No huge page reservations > remain after the test. > > HugePages_Total: 1024 > HugePages_Free: 1024 > HugePages_Rsvd: 0 > HugePages_Surp: 0 > Hugepagesize: 2048 kB > Hugetlb: 2097152 kB > > To try and better understand the reservation cgroup controller, I addded > a few printks to the code. While running fallocate_stress with the > printks, I can consistently hit the WARN_ON_ONCE() due to the counter > going negative. Here are the cgroup counter values after such a run: > > hugetlb.2MB.failcnt 0 > hugetlb.2MB.limit_in_bytes 9223372036854771712 > hugetlb.2MB.max_usage_in_bytes 209715200 > hugetlb.2MB.rsvd.failcnt 3 > hugetlb.2MB.rsvd.limit_in_bytes 9223372036854771712 > hugetlb.2MB.rsvd.max_usage_in_bytes 251658240 > hugetlb.2MB.rsvd.usage_in_bytes 18446744073487253504 > hugetlb.2MB.usage_in_bytes 0 > > Again, no reserved pages after the test. > > HugePages_Total: 1024 > HugePages_Free: 1024 > HugePages_Rsvd: 0 > HugePages_Surp: 0 > Hugepagesize: 2048 kB > Hugetlb: 2097152 kB > > I have some basic hugetlb hole punch functionality tests. Running > these on the kernel with added printk's does not cause any issues. > In order to reproduce, I need to run fallocate_stress test which > will cause hole punch to race with page fault. Best guess at this > time is that some of the error/race detection reservation back out > code is not properly dealing with cgroup accounting. > > I'll take a look at this as well. > I'm bisecting the warning right now. Looks like it was introduced in v5.7. -- Thanks, David / dhildenb
