mbox series

[0/8] hugetlb: Use new vma mutex for huge pmd sharing synchronization

Message ID 20220824175757.20590-1-mike.kravetz@oracle.com (mailing list archive)
Headers show
Series hugetlb: Use new vma mutex for huge pmd sharing synchronization | expand

Message

Mike Kravetz Aug. 24, 2022, 5:57 p.m. UTC
hugetlb fault scalability regressions have recently been reported [1].
This is not the first such report, as regressions were also noted when
commit c0d0381ade79 ("hugetlbfs: use i_mmap_rwsem for more pmd sharing
synchronization") was added [2] in v5.7.  At that time, a proposal to
address the regression was suggested [3] but went nowhere.

The regression and benefit of this patch series is not evident when
using the vm_scalability benchmark reported in [2] on a recent kernel.
Results from running,
"./usemem -n 48 --prealloc --prefault -O -U 3448054972"

			48 sample Avg
next-20220822		next-20220822			next-20220822
unmodified	revert i_mmap_sema locking	vma sema locking, this series
-----------------------------------------------------------------------------
494229 KB/s		495375 KB/s			495573 KB/s

The recent regression report [1] notes page fault and fork latency of
shared hugetlb mappings.  To measure this, I created two simple programs:
1) map a shared hugetlb area, write fault all pages, unmap area
   Do this in a continuous loop to measure faults per second
2) map a shared hugetlb area, write fault a few pages, fork and exit
   Do this in a continuous loop to measure forks per second
These programs were run on a 48 CPU VM with 320GB memory.  The shared
mapping size was 250GB.  For comparison, a single instance of the program
was run.  Then, multiple instances were run in parallel to introduce
lock contention.  Changing the locking scheme results in a significant
performance benefit.

test		instances	unmodified	revert		vma
--------------------------------------------------------------------------
faults per sec	1		397068		403411		394935
faults per sec  24		 68322		 83023		 82436
forks per sec   1		  2717		  2862		  2816
forks per sec   24		   404		   465		   499
Combined faults 24		  1528		 69090		 59544
Combined forks  24		   337		    66		   140

Combined test is when running both faulting program and forking program
simultaneously.

Patches 1 and 2 of this series revert c0d0381ade79 and 87bf91d39bb5 which
depends on c0d0381ade79.  Acquisition of i_mmap_rwsem is still required in
the fault path to establish pmd sharing, so this is moved back to
huge_pmd_share.  With c0d0381ade79 reverted, this race is exposed:

Faulting thread                                 Unsharing thread
...                                                  ...
ptep = huge_pte_offset()
      or
ptep = huge_pte_alloc()
...
                                                i_mmap_lock_write
                                                lock page table
ptep invalid   <------------------------        huge_pmd_unshare()
Could be in a previously                        unlock_page_table
sharing process or worse                        i_mmap_unlock_write
...
ptl = huge_pte_lock(ptep)
get/update pte
set_pte_at(pte, ptep)

Reverting 87bf91d39bb5 exposes races in page fault/file truncation.
Patches 3 and 4 of this series address those races.  This requires
using the hugetlb fault mutexes for more coordination between the fault
code and file page removal.

Patches 5 - 7 add infrastructure for a new vma based rw semaphore that
will be used for pmd sharing synchronization.  The idea is that this
semaphore will be held in read mode for the duration of fault processing,
and held in write mode for unmap operations which may call huge_pmd_unshare.
Acquiring i_mmap_rwsem is also still required to synchronize huge pmd
sharing.  However it is only required in the fault path when setting up
sharing, and will be acquired in huge_pmd_share().

Patch 8 makes use of this new vma lock.  Unfortunately, the fault code
and truncate/hole punch code would naturally take locks in the opposite
order which could lead to deadlock.  Since the performance of page faults
is more important, the truncation/hole punch code is modified to back
out and take locks in the correct order if necessary.

[1] https://lore.kernel.org/linux-mm/43faf292-245b-5db5-cce9-369d8fb6bd21@infradead.org/
[2] https://lore.kernel.org/lkml/20200622005551.GK5535@shao2-debian/
[3] https://lore.kernel.org/linux-mm/20200706202615.32111-1-mike.kravetz@oracle.com/

RFC -> v1
- Addressed many issues pointed out by Miaohe Lin.  Thank you!  Most
  significant was not attempting to backout pages in fault code due to
  races with truncation (patch 4).
- Rebased and retested on next-20220822

Mike Kravetz (8):
  hugetlbfs: revert use i_mmap_rwsem to address page fault/truncate race
  hugetlbfs: revert use i_mmap_rwsem for more pmd sharing
    synchronization
  hugetlb: rename remove_huge_page to hugetlb_delete_from_page_cache
  hugetlb: handle truncate racing with page faults
  hugetlb: rename vma_shareable() and refactor code
  hugetlb: add vma based lock for pmd sharing
  hugetlb: create hugetlb_unmap_file_folio to unmap single file folio
  hugetlb: use new vma_lock for pmd sharing synchronization

 fs/hugetlbfs/inode.c    | 364 ++++++++++++++++++++++++++++++----------
 include/linux/hugetlb.h |  38 ++++-
 kernel/fork.c           |   6 +-
 mm/hugetlb.c            | 354 ++++++++++++++++++++++++++++----------
 mm/memory.c             |   2 +
 mm/rmap.c               | 114 ++++++++-----
 mm/userfaultfd.c        |  14 +-
 7 files changed, 653 insertions(+), 239 deletions(-)


base-commit: cc2986f4dc67df7e6209e0cd74145fffbd30d693