On Fri, Sep 2, 2022 at 12:43 AM Peter Zijlstra [off-list ref] wrote:

On Thu, Sep 01, 2022 at 10:34:48AM -0700, Suren Baghdasaryan wrote:

quoted

This is a proof of concept for per-vma locks idea that was discussed
during SPF [1] discussion at LSF/MM this year [2], which concluded with
suggestion that “a reader/writer semaphore could be put into the VMA
itself; that would have the effect of using the VMA as a sort of range
lock. There would still be contention at the VMA level, but it would be an
improvement.” This patchset implements this suggested approach.

The whole reason I started the SPF thing waay back when was because one
of the primary reporters at the time had very large VMAs and a per-vma
lock wouldn't actually help anything at all.

IIRC it was either scientific code initializing a huge matrix or a
database with a giant table; I'm sure the archives have better memory
than me.

Regardless of the initial intent, SPF happens to be very useful for
cases when we have multiple threads establishing some mappings
concurrently with page faults (see details at [1]). Android vendors
independently from each other were backporting your and Laurent's
patchset for years. I found internal reports of similar mmap_lock
contention issues in Google Fibers [2] and I suspect there are more
places this happens if people looked closer.

[1] https://lore.kernel.org/all/CAJuCfpE10y78SNPQ+LRY5EonDFhOG=1XjZ9FUUDiyhfhjZ54NA@mail.gmail.com/ (local)
[2] https://www.phoronix.com/scan.php?page=news_item&px=Google-Fibers-Toward-Open

--
To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe@android.com.