Bug 230124 – [libpas] Update to 38b9b0b92ccc9628627c742523de6200acc08211 and fully enable on AS

• Status: RESOLVED
• Resolution: FIXED
• Component: bmalloc

[Filip Pizlo]

...

[Filip Pizlo]

Created attachment 437783 [details]
patch for landing, maybe

[Filip Pizlo]

Created attachment 437786 [details]
better changelog

[Filip Pizlo]

Created attachment 437793 [details]
added RAMification number to changelog

[Yusuke Suzuki]

Comment on attachment 437793 [details]
added RAMification number to changelog

View in context: https://bugs.webkit.org/attachment.cgi?id=437793&action=review

r=me

> Source/bmalloc/ChangeLog:25
> +        - Replace biasing and magazines with per-thread view caches. A view cache is a bounded-size
> +          queue of exclusive_views (i.e. pages) that is local to a thread. Each thread has a view
> +          cache for every segregated_size_directory that it talks to and that has view caching
> +          enabled. The size directories can control view cache size somewhat dynamically (different
> +          directories can have different size view caches). Views get enqueued when a thread frees
> +          the first object in the page. Views get dequeued whenever a thread would have asked the
> +          directory for a view but the view cache was non-empty.
> +
> +          This change increases the efficiency of local_allocator_refill, since pages have a longer
> +          time to "cook" before anyone allocates from them (since the view cache is a queue). As
> +          well, view caches' thread-locality means that there is no locking or contention when
> +          accessing them. The scavenger has clever tricks for clearing out view caches, similarly to
> +          how it clears out local allocators (it uses thread_suspend).

Yeah, I tried using magazine in bmalloc, and the result was, thread-local one was better (lock is not necessary, cpu-number can change relatively easily etc.).

> Source/bmalloc/bmalloc/BPlatform.h:323
> +#if defined(BENABLE_LIBPAS) && BENABLE_LIBPAS

We can use `if BENABLE(LIBPAS)`.

[Filip Pizlo]

Created attachment 438314 [details]
patch for landing?

[Yusuke Suzuki]

Looks like mac-AS-debug-wk2 failures are due to test flakiness. Both tests are failing frequently these days.

https://results.webkit.org/?suite=layout-tests&test=imported%2Fw3c%2Fweb-platform-tests%2Fcss%2Fcss-transforms%2Fcrashtests%2Ftransform-marquee-resize-div-image-001.html
https://results.webkit.org/?suite=layout-tests&test=imported%2Fw3c%2Fweb-platform-tests%2Fwebrtc%2FRTCDataChannel-close.html

[Filip Pizlo]

Landed in https://trac.webkit.org/changeset/282556/webkit

[Radar WebKit Bug Importer]

<rdar://problem/83195529>

[Truitt Savell]

It looks like the changes in nhttps://trac.webkit.org/changeset/282556/webkit

broke multiple tests on arm64 JSC testers:

https://build.webkit.org/#/builders/166/builds/2069
https://build.webkit.org/#/builders/102/builds/3448

[Filip Pizlo]

(In reply to Truitt Savell from comment #9)
> It looks like the changes in nhttps://trac.webkit.org/changeset/282556/webkit
> 
> broke multiple tests on arm64 JSC testers:
> 
> https://build.webkit.org/#/builders/166/builds/2069
> https://build.webkit.org/#/builders/102/builds/3448

Fix: https://bugs.webkit.org/show_bug.cgi?id=230362