Failure to allocate ArrayStorage in emit_op_new_array leads to poisonous JSArray

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Mark Hahnenberg]

When emit_op_new_array successfully allocates a new JSArray but fails to allocate the corresponding ArrayStorage for it, it falls back to the out-of-line stub call to constructArray, which constructs and entirely new JSArray/ArrayStorage pair. This leaves us with a JSArray hanging around on the stack or in a register that did not go through its own constructor, thus giving it uninitialized memory in the two fields that are checked in JSArray::visitChildren.

The fix for this is to detect when we have successfully allocated a JSArray but failed to allocate its backing store and to write zeros into the correct places in the object prior to abandoning it.

[Mark Hahnenberg]

Created attachment 138492 [details]
Patch

[Geoffrey Garen]

Comment on attachment 138492 [details]
Patch

(1) Regression test, please.

(2) A better fix is to allocate the backing store first. That way, no special branching, and no zombie JSArray.

(3) Please re-verify that this inlining is still a performance win. It's a bit odd to have an optimization only in the slow JIT.

[Geoffrey Garen]

<rdar://problem/11309000>

[Mark Hahnenberg]

Created attachment 138583 [details]
Bencher results

These are the benchmark results for ToT without the optimization (just a stub call like we used to do) compared to the optimization along with this particular fix. Looks like a small win. I think when we finish bug 79198 arrays will be in a much better state in general.

[Mark Hahnenberg]

Created attachment 138605 [details]
Patch

[Geoffrey Garen]

Comment on attachment 138605 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=138605&action=review

Boy, those perf numbers are whisper-thin. Still, the fix is good, so let's take it.

> Source/JavaScriptCore/jit/JITInlineMethods.h:490
> +    // Allocate the backing store for the array. We allocate the storage first
> +    // because it could fail and we have to take the slow path, which would leave 
> +    // behind a zombie JSArray with inconsistent state, potentially causing a GC crash.

Hard to understand a comment about a potentially bad alternate version of the code.

I would just say, "We allocate the backing store first to ensure that garbage collection doesn't happen during JSArray initialization".

[Mark Hahnenberg]

Fixed in http://trac.webkit.org/changeset/115096.

[Rafael Weinstein]

Reopening to attach new patch.

[Rafael Weinstein]

Created attachment 141188 [details]
Patch

[Mark Hahnenberg]

(In reply to comment #8)
> Reopening to attach new patch.

I don't think this is the bug you meant to reopen.

[Rafael Weinstein]

Sorry. No, it wasn't