[v7,0/4] per-vma locks in userfaultfd

Message

Lokesh Gidra Feb. 15, 2024, 6:27 p.m. UTC

Performing userfaultfd operations (like copy/move etc.) in critical
section of mmap_lock (read-mode) causes significant contention on the
lock when operations requiring the lock in write-mode are taking place
concurrently. We can use per-vma locks instead to significantly reduce
the contention issue.

Android runtime's Garbage Collector uses userfaultfd for concurrent
compaction. mmap-lock contention during compaction potentially causes
jittery experience for the user. During one such reproducible scenario,
we observed the following improvements with this patch-set:

- Wall clock time of compaction phase came down from ~3s to <500ms
- Uninterruptible sleep time (across all threads in the process) was
  ~10ms (none in mmap_lock) during compaction, instead of >20s

Changes since v6 [6]:
- Added a patch to add vma_assert_locked() for !CONFIG_PER_VMA_LOCK, per
  Suren Baghdasaryan
- Replaced mmap_assert_locked() with vma_assert_locked() in
  move_pages_huge_pmd(), per Suren Baghdasaryan

Changes since v5 [5]:
- Use abstract function names (like uffd_mfill_lock/uffd_mfill_unlock)
  to avoid using too many #ifdef's, per Suren Baghdasaryan and Liam
  Howlett
- Use 'unlikely' (as earlier) to anon_vma related checks, per Liam Howlett
- Eliminate redundant ptr->err->ptr conversion, per Liam Howlett
- Use 'int' instead of 'long' for error return type, per Liam Howlett

Changes since v4 [4]:
- Fix possible deadlock in find_and_lock_vmas() which may arise if
  lock_vma() is used for both src and dst vmas.
- Ensure we lock vma only once if src and dst vmas are same.
- Fix error handling in move_pages() after successfully locking vmas.
- Introduce helper function for finding dst vma and preparing its
  anon_vma when done in mmap_lock critical section, per Liam Howlett.
- Introduce helper function for finding dst and src vmas when done in
  mmap_lock critical section.

Changes since v3 [3]:
- Rename function names to clearly reflect which lock is being taken,
  per Liam Howlett.
- Have separate functions and abstractions in mm/userfaultfd.c to avoid
  confusion around which lock is being acquired/released, per Liam Howlett.
- Prepare anon_vma for all private vmas, anonymous or file-backed,
  per Jann Horn.

Changes since v2 [2]:
- Implement and use lock_vma() which uses mmap_lock critical section
  to lock the VMA using per-vma lock if lock_vma_under_rcu() fails,
  per Liam R. Howlett. This helps simplify the code and also avoids
  performing the entire userfaultfd operation under mmap_lock.

Changes since v1 [1]:
- rebase patches on 'mm-unstable' branch

[1] https://lore.kernel.org/all/[e-mail]/
[2] https://lore.kernel.org/all/[e-mail]/
[3] https://lore.kernel.org/all/[e-mail]/
[4] https://lore.kernel.org/all/[e-mail]/
[5] https://lore.kernel.org/all/[e-mail]/
[6] https://lore.kernel.org/all/[e-mail]/

Lokesh Gidra (4):
  userfaultfd: move userfaultfd_ctx struct to header file
  userfaultfd: protect mmap_changing with rw_sem in userfaulfd_ctx
  mm: add vma_assert_locked() for !CONFIG_PER_VMA_LOCK
  userfaultfd: use per-vma locks in userfaultfd operations

 fs/userfaultfd.c              |  86 ++-----
 include/linux/mm.h            |   5 +
 include/linux/userfaultfd_k.h |  75 ++++--
 mm/huge_memory.c              |   5 +-
 mm/userfaultfd.c              | 438 +++++++++++++++++++++++++---------
 5 files changed, 413 insertions(+), 196 deletions(-)