1100744 - Massive amounts of readback from the X server when viewing a detailed PDF map

Status: RESOLVED WORKSFORME

(Core :: Graphics: Layers, defect)

Description

[Nicholas Nethercote [inactive]]

I did some cumulative heap profiling on Linux while viewing a very detailed map with pdf.js (http://www.vrr.de/imperia/md/content/fahrten/stadtlinienplaene/wuppertal_2012.pdf, from https://github.com/mozilla/pdf.js/issues/2541).

I will attach a file that show that 55% of the cumulative heap allocations are for Cairo and X server stuff under DrawTargetCairo::CopySurface(). roc says that this:

> shows a serious problem. We're doing massive amounts of readback from the X
> server. not only will it hog memory, but it will be incredibly slow."

I tried with e10s disabled and got a similar result; actually, the numbers were even worse.

I wanted to try with gfx.xrender.enabled set to false, but that just gave me a white empty window when I started Firefox.

Comment 1

[Nicholas Nethercote [inactive]]

Attachment: Child process data from the e10s build profile

Comment 2

[Nicholas Nethercote [inactive]]

Attachment: Data from the non-e10s build profile

All data is from a build from revision e9029cbd137c.

Comment 3

[Lee Salzman [:lsalzman]]

(In reply to Nicholas Nethercote [:njn] from comment #0)
> I did some cumulative heap profiling on Linux while viewing a very detailed
> map with pdf.js
> (http://www.vrr.de/imperia/md/content/fahrten/stadtlinienplaene/
> wuppertal_2012.pdf, from https://github.com/mozilla/pdf.js/issues/2541).
> 
> I will attach a file that show that 55% of the cumulative heap allocations
> are for Cairo and X server stuff under DrawTargetCairo::CopySurface(). roc
> says that this:
> 
> > shows a serious problem. We're doing massive amounts of readback from the X
> > server. not only will it hog memory, but it will be incredibly slow."
> 
> I tried with e10s disabled and got a similar result; actually, the numbers
> were even worse.
> 
> I wanted to try with gfx.xrender.enabled set to false, but that just gave me
> a white empty window when I started Firefox.

Do you still have a copy of the pdf? It is not accessible via that link anymore. Also, I can't seem to reproduce the particular traceback that you have attached via pdfs I have tried otherwise. Does the issue still occur for you with nightlies?

Comment 4

[Nicholas Nethercote [inactive]]

Attachment: wuppertal.pdf

Comment 5

[Nicholas Nethercote [inactive]]

Attachment: New DMD output

I attached the PDF in comment 4.

I ran again and got similar results. The stacks are a bit screwy in this profile, not sure why, but the basic shape of the results is unchanged.

Comment 6

[Lee Salzman [:lsalzman]]

Since we disabled xrender in bug 1180942, and switched to Skia canvas in 1244754, this bug is no longer applicable. We don't as yet plan on supporting xrender again by default for canvas/content rendering, maybe at best compositing, so I don't expect this should pop up again.