[webkit-dev] Making WTF::StringImpl and WTF::AtomString thread-safe

[Chris Lord]

Hi all,

As part of the work for making FontCache thread-safe, it's necessary for
there to be a thread-safe AtomString. After discussion, it seems that a
thread-safe StringImpl is generally desirable and GPUProcess also has a
need of it. I've filed a bug to track this work:
https://bugs.webkit.org/show_bug.cgi?id=219285

Google have already done this for Blink and there's a nice plan and lots
of discussion to read. Their plan document is linked in the bug. I think
we'd be well-served by taking broadly the same approach, which is to
make ref-counting on StringImpl atomic and to guard AtomStringTable
access with a lock.

Making String thread-safe has implications, of course, and I'd like to
open discussion on this - Making ref-counting atomic on StringImpl has a
significant, negative impact on the expense of ref and deref operations.
I'm interested in discussing how we should approach this in terms of
tracking the work in Bugzilla and how to go about landing it. Perhaps
people also have alternative ideas?

On the bug is a first-run at implementing the above approach, currently
minus the follow-up of everywhere taking into consideration that
String/AtomString are now thread-safe. The impact on StringImpl
ref/deref performance has it running on my Xeon desktop machine at about
30-50% of non-atomic ref/deref performance. Speedometer 2.0 takes a 1-8%
hit considering error margins, but I'm fairly certain it's mostly on the
higher end of that and I've not run enough iterations just yet.
Jetstream 1.1 seems practically unaffected, I can't run 2.0 with or
without the patch, it appears to hang the browser on the bomb-workers
test (at least if it completes, it's not in a reasonable time-frame). I
would guess that results may vary wildly depending on platform and
available atomic access primitives. As one might expect, the impact is
far less on a debug build.

I think the initial patch of making it thread-safe vs. the follow-up of
altering various areas to take it into account could/should be split,
but I assume we'd want to land them at the same time. This is cumbersome
with how WebKit Bugzilla currently works and I'd like to hear what
people think and how similar changes have been made in the past.

Thoughts?

Regards,

Chris