Commits c0d0381ade79 and 87bf91d39bb5 changed the way huegtlb locking was performed to address BUGs. One specific change was to always take the i_mmap_rwsem in read mode during fault processing. One result of this change was a 33% regression for anon non-shared page faults [1].
Technically, i_mmap_rwsem only needs to be taken during page faults if the pmd can potentially be shared. pmd sharing is not possible for anon non-shared mappings (as in the reported regression), therefore the code can be modified to not acquire the semaphore in this case. Unfortunately, commit 87bf91d39bb5 depends on i_mmap_rwsem always being taken in the fault path to prevent fault/truncation races. So, that approach is no longer appropriate. Rather, the code now detects races and backs out operations. This code "works" in that it only takes i_mmap_rwsem when necessary and addresses the original BUGs. However, I am sending as an RFC because: - I am unsure if the added complexity is worth performance benefit. - There needs to be a better way/location to make a decison about taking the semaphore. See FIXME's in the code. Comments and suggestions would be appreciated. [1] https://lore.kernel.org/lkml/20200622005551.GK5535@shao2-debian Mike Kravetz (3): Revert: "hugetlbfs: Use i_mmap_rwsem to address page fault/truncate race" hugetlbfs: Only take i_mmap_rwsem when sharing is possible huegtlbfs: handle page fault/truncate races fs/hugetlbfs/inode.c | 69 +++++++++----------- mm/hugetlb.c | 150 ++++++++++++++++++++++++++++++------------- 2 files changed, 137 insertions(+), 82 deletions(-) -- 2.25.4
