On Mon, Jan 29, 2024 at 12:39 PM Liam R. Howlett
<Liam.Howlett@oracle.com> wrote:
>
> * Lokesh Gidra <lokeshgidra@google.com> [240129 14:35]:
> > 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.
>
> Is this really an issue? I'm surprised so much userfaultfd work is
> happening to create contention. Can you share some numbers and how your
> patch set changes the performance?
>
In Android we are using userfaultfd for Android Runtime's GC
compaction. mmap-lock (write-mode) operations like mmap/munmap/mlock
happening simultaneously elsewhere in the process caused significant
contention. Of course, this doesn't happen during every compaction,
but whenever it does it leads to a 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 less than 500ms
- Uninterruptible sleep time (across all threads in the process) was
~10ms (none was in mmap_lock) during compaction, instead of >20s
I will add these numbers in the cover letter in the next version of
this patchset.
> >
> > Changes since v1 [1]:
> > - rebase patches on 'mm-unstable' branch
> >
> > [1] https://lore.kernel.org/all/20240126182647.2748949-1-lokeshgidra@google.com/
> >
> > Lokesh Gidra (3):
> > userfaultfd: move userfaultfd_ctx struct to header file
> > userfaultfd: protect mmap_changing with rw_sem in userfaulfd_ctx
> > userfaultfd: use per-vma locks in userfaultfd operations
> >
> > fs/userfaultfd.c | 86 ++++---------
> > include/linux/userfaultfd_k.h | 75 ++++++++---
> > mm/userfaultfd.c | 229 ++++++++++++++++++++++------------
> > 3 files changed, 229 insertions(+), 161 deletions(-)
> >
> > --
> > 2.43.0.429.g432eaa2c6b-goog
> >
> >