Add kGCCallbackFlagCollectAllExternalMemory

Add kGCCallbackFlagCollectAllExternalMemory

BUG=570268

[keishi, 2016-08-12 06:00:09]

Description was changed from

==========
Add kGCCallbackFlagCollectAllExternalMemory

BUG=
==========

to

==========
Add kGCCallbackFlagCollectAllExternalMemory

BUG=570268
==========

[keishi, 2016-08-12 06:58:14]

keishi@chromium.org changed reviewers:
+ haraken@chromium.org

[keishi, 2016-08-12 06:58:14]

This CL is a fix for c)

Explanation:

After investigating the memory leak for the html below (crbug.com/570268) I saw
three problems.
https://dl.dropboxusercontent.com/u/15217362/bloburl/bitmaptest.html

a) AdjustAmountOfExternalMemory is not called for ImageResource.
b) external_memory_limit increases too much because it is adjusted after a v8
gc, not blink gc.
c) V8Followup blinkgc doesn't always run after a "external memory induced V8 GC"
because ImageResource data is allocated in malloc and blinkgc doesn't know about
malloc.

I confirmed that we can reduce peak memory usage by 500MB (after 20 seconds of
running the html) if these three problems are eliminated.

The fix for a) is https://codereview.chromium.org/2230133002/
To keep the change simple and avoid issues like double counting memory this only
reports the ImageResource data size to v8.

The fix for c) is https://codereview.chromium.org/2235653002/ and
https://codereview.chromium.org/2238173002/
This adds a new callback flag so Blink can be extra careful to free the external
allocations.

external_memory_limit is set to (external_memory + 192MB) after a V8 GC. When
external_memory becomes greater than external_memory_limit a V8 GC is triggered.
external_memory_limit increases too much because the external_memory hasn't
shrunk yet.
For example, for this html, i observed that the Image object requires two GC
cycles for it to be collected.
I'm not sure why but I believe the fact that the onload event handler references
the Image object makes the V8 Image object uncollectable until after a Blink GC.
The ImageResource data is not freed until a BlinkGC happens after that.
The significant delay from when the V8 GC is triggered by external memory alloc,
and when the external memory is a problem but I don't think we can fix that.
Instead I propose we reduce external_memory_limit when external_memory reduces
https://codereview.chromium.org/2233683002/
This will make it harder for the external_memory_limit to just keep on
increasing, and I'm hoping it will have minimal negative impact.

[haraken, 2016-08-12 07:11:00]

On 2016/08/12 06:58:14, keishi wrote:
> This CL is a fix for c)
> 
> Explanation:
> 
> After investigating the memory leak for the html below (crbug.com/570268) I
saw
> three problems.
> https://dl.dropboxusercontent.com/u/15217362/bloburl/bitmaptest.html
> 
> a) AdjustAmountOfExternalMemory is not called for ImageResource.
> b) external_memory_limit increases too much because it is adjusted after a v8
> gc, not blink gc.
> c) V8Followup blinkgc doesn't always run after a "external memory induced V8
GC"
> because ImageResource data is allocated in malloc and blinkgc doesn't know
about
> malloc.
> 
> I confirmed that we can reduce peak memory usage by 500MB (after 20 seconds of
> running the html) if these three problems are eliminated.
> 
> The fix for a) is https://codereview.chromium.org/2230133002/
> To keep the change simple and avoid issues like double counting memory this
only
> reports the ImageResource data size to v8.
> 
> The fix for c) is https://codereview.chromium.org/2235653002/ and
> https://codereview.chromium.org/2238173002/
> This adds a new callback flag so Blink can be extra careful to free the
external
> allocations.
> 
> external_memory_limit is set to (external_memory + 192MB) after a V8 GC. When
> external_memory becomes greater than external_memory_limit a V8 GC is
triggered.
> external_memory_limit increases too much because the external_memory hasn't
> shrunk yet.
> For example, for this html, i observed that the Image object requires two GC
> cycles for it to be collected.
> I'm not sure why but I believe the fact that the onload event handler
references
> the Image object makes the V8 Image object uncollectable until after a Blink
GC.
> The ImageResource data is not freed until a BlinkGC happens after that.
> The significant delay from when the V8 GC is triggered by external memory
alloc,
> and when the external memory is a problem but I don't think we can fix that.
> Instead I propose we reduce external_memory_limit when external_memory reduces
> https://codereview.chromium.org/2233683002/
> This will make it harder for the external_memory_limit to just keep on
> increasing, and I'm hoping it will have minimal negative impact.

(Let's address https://codereview.chromium.org/2238173002/ first before asking
V8 exports for review.)