814312 - Avoid PL_ArenaPool recycling in the layout engine.

Status: RESOLVED FIXED

(Core :: Layout, defect)

Description

[Nicholas Nethercote [inactive]]

PLArenaPool has a global free list of chunks.  If you call PL_FreeArenaPool
(as opposed to PL_FinishArenaPool) any outstanding chunks in the arena pool
are put on the free list.  As a result, the free list can have chunks of any
size, and arbitrarily-sized chunks can be handed out to any other arena pool.
In other words, the sizing of the chunks obtained by any one arena pool
depends partially on what other arena pools do(!)

Furthermore, this recycling of chunks is really confusing when looking at heap
profiles (e.g. from DMD) because chunks that were allocated by one arena pool
may be (a) on the free list, or even worse (b) held by another arena pool.  (I
spent 20 minutes scratching my head over a (b) case before bz explained about
the free list.)

My experience is that if you have a decent allocator (such as jemalloc) it's
simpler and better to let it do this kind of chunk recycling for you.  (See
bug 715770 for a related example.)

Comment 1

[Nicholas Nethercote [inactive]]

Attachment: Avoid PL_ArenaPool recycling in the layout engine.

This patch changes two places in the layout engine that call
PL_FreeArenaPool() and then PL_FinishArenaPool(), so they instead just call
the latter.  This avoids the recycling behaviour.

This reduces the amount of memory held on the free list from typically a few
10s of KiBs to typically zero.  (There's a use in NSS that causes it to rise to
~300KiB at shut-down, but that's not worth fixing.)

Comment 2

[Nicholas Nethercote [inactive]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/0a2f6f173e52

Comment 3

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/0a2f6f173e52