896703 - Siesta example dashboard loads very slowly

Status: RESOLVED WORKSFORME

(Core :: Graphics, defect)

Description

[Mats]

Attachment: ACMQyd9x.dms

User Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_8_4) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/28.0.1500.71 Safari/537.36

Steps to reproduce:

Open our web app online (A JS testing tool called Siesta)


Actual results:

Firefox freezes on Mac (Macbook Pro Retina). Tried on FF22 and Nightly. Works fine on Win FF, Chrome and all others. Suspect something to do with HTML5 web fonts.


Expected results:

No freeze

Comment 1

[Mats]

URL to reproduce it: http://www.bryntum.com/examples/siesta-latest/examples/browse-all-desktop.html

Comment 2

[Kevin Brosnan [Ex-Mozilla]]

According to the profile 

46% of time is at http://mxr.mozilla.org/mozilla-central/source/gfx/thebes/gfxContext.cpp#303
41% if time is at http://mxr.mozilla.org/mozilla-central/source/widget/cocoa/nsAppShell.mm#815

Comment 3

[Milan Sreckovic [:milan] (needinfo for best results)]

I see this on Aurora (24), but not on the latest nightly (2013-07-23).

Comment 4

[Milan Sreckovic [:milan] (needinfo for best results)]

(In reply to Milan Sreckovic [:milan] from comment #3)
> I see this on Aurora (24), but not on the latest nightly (2013-07-23).

I do see it on the nightly one day before (2013-07-22), though it is not as bad as Aurora. Not convinced this is in graphics.

Comment 5

[Milan Sreckovic [:milan] (needinfo for best results)]

(In reply to Milan Sreckovic [:milan] from comment #4)
> (In reply to Milan Sreckovic [:milan] from comment #3)
> > I see this on Aurora (24), but not on the latest nightly (2013-07-23).
> 
> I do see it on the nightly one day before (2013-07-22), though it is not as
> bad as Aurora. Not convinced this is in graphics.

Though, perhaps the improvement is related to bug 896129.

Comment 6

[Mats]

Any idea of a workaround? If I delete 'siesta-all.css' from the page includes, all works fine...

Comment 7

[Mats]

This is the "offending" line of CSS.

background-image    : linear-gradient(rgba(221, 221, 221, .3) 1px, transparent 1px),
                          linear-gradient(90deg, rgba(221, 221, 221, .3) 1px, transparent 1px);


Removing it, all is good!

Comment 8

[Steven Michaud [:smichaud] (Retired)]

How do you reproduce these hangs?

Just loading http://www.bryntum.com/examples/siesta-latest/examples/browse-all-desktop.html isn't enough to do it.  Neither (at least in my tests) is clicking on the "Run all" button.

I tested with FF 22 on a Retina MacBook Pro.

Comment 9

[Mats]

Hehe, we released a new version without those lines of CSS so it won't be reproducible online anymore :)

Old version where this is reproducible can be found here (until we do another Siesta release, typically in 2w): 

http://www.bryntum.com/examples/siesta-latest-old/examples/browse-all-desktop.html

Comment 10

[Steven Michaud [:smichaud] (Retired)]

OK, thanks.

Even that page doesn't "hang", but it does load very slowly.  The window also redraws very slowly when you resize it.  I'll try to get some representative stack traces (which are often much more informative than a profile).

Even with the old version, running the tests (e.g. clicking on the "Run all" button) doesn't cause bad behavior.  Does that match your results?

By the way, if at all possible please give us a "permanent" link to your "old version" -- one that will work for at least (say) six months.  I suspect this isn't going to be an easy bug to fix :-(

Comment 11

[Steven Michaud [:smichaud] (Retired)]

The "old version" loads slowly even on a non-Retina Mac.

Comment 12

[Steven Michaud [:smichaud] (Retired)]

The "old version" loads very snappily in FF 15.0.1.  It's much slower in FF 16, but not nearly as bad as it is now.  Things seem to get as bad as now only with FF 18.

I'll see if I can find one or both of these regression ranges in mozilla-central nightlies.

Comment 13

[Steven Michaud [:smichaud] (Retired)]

The time required to load http://www.bryntum.com/examples/siesta-latest-old/examples/browse-all-desktop.html doubled with the firefox-2012-09-30-03-06-10-mozilla-central nightly -- which accounts for the FF 16/18 regression.

This implies the following regression range:
http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=c09a0c022b2e&tochange=a680fd777c3b

Comment 14

[Steven Michaud [:smichaud] (Retired)]

Just a guess, but I'd suspect one of the following:
http://hg.mozilla.org/mozilla-central/rev/a680fd777c3b
http://hg.mozilla.org/mozilla-central/rev/87346ce19cfc

Comment 15

[Steven Michaud [:smichaud] (Retired)]

Oops, it's actually the patch for bug 674373 (or part of it) that caused this particular regression.  Turning off HiDPI support (by setting gfx.hidpi.enabled to -1) makes it go away on a Retina MacBook Pro.

Comment 16

[Steven Michaud [:smichaud] (Retired)]

Here's the range for the FF 15/16 regression.  Unfortunately the 2012-07-07 mozilla-central nightly is missing, so this range is more than one day long:

firefox-2012-07-06-07-51-26-mozilla-central
firefox-2012-07-08-03-05-43-mozilla-central

http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=a8f682801a6d&tochange=1751d97cc9e4

Comment 17

[Steven Michaud [:smichaud] (Retired)]

Once again it's just a guess, but I suspect one of the patches for bug 752187, which has to do with gradients.

Comment 18

[Steven Michaud [:smichaud] (Retired)]

Tomorrow I'll hg bisect the FF 15/16 regression range.  I'll make this bug block the "guilty" patch.

Comment 19

[Milan Sreckovic [:milan] (needinfo for best results)]

(In reply to Milan Sreckovic [:milan] from comment #4)
> (In reply to Milan Sreckovic [:milan] from comment #3)
> > I see this on Aurora (24), but not on the latest nightly (2013-07-23).
> 
> I do see it on the nightly one day before (2013-07-22), though it is not as
> bad as Aurora.
Don't see the problem in nightly 2013-07-29. If it is gradients, seeing the problem in 2013-07-22, but not in 2013-07-23 would match three gradient bugs fixed on the 22nd: bug 895135, bug 895646 and bug 893977.

Comment 20

[Steven Michaud [:smichaud] (Retired)]

(Following up comment #10)

> By the way, if at all possible please give us a "permanent" link to
> your "old version" -- one that will work for at least (say) six
> months.  I suspect this isn't going to be an easy bug to fix :-(

Never mind.  I've created a mirror of it here:
http://people.mozilla.com/~stmichaud/bmo/bugzilla896703/www.bryntum.com/examples/siesta-latest-old/examples/browse-all-desktop.html

Comment 21

[Steven Michaud [:smichaud] (Retired)]

The FF 15/16 regression was triggered by the following patch:

http://hg.mozilla.org/mozilla-central/rev/1f3a325d58a4

 Bug 752187 - Part 7: Add aIsLegacy parameter to parsing functions. r=dbaron
author	Masatoshi Kimura <[e-mail]>
	Sat Jul 07 10:27:08 2012 -0400 (at Sat Jul 07 10:27:08 2012 -0400)

But (as best I can tell) all this does is introduce support for the "linear-gradient" keyword (as opposed to "-moz-linear-gradient").  So this patch didn't *cause* this bug -- it was there all along.

Comment 22

[Steven Michaud [:smichaud] (Retired)]

I also see a marked improvement between the following two nightlies:

firefox-2013-07-22-03-02-26-mozilla-central
firefox-2013-07-23-03-02-05-mozilla-central

http://hg.mozilla.org/mozilla-central/pushloghtml?fromchange=2268ff80683a&tochange=5ceea82a79c7

Next I'll try to find which patch or patches are responsible for it.

Comment 23

[Steven Michaud [:smichaud] (Retired)]

The following patch seems to be responsible for all the improvement:

http://hg.mozilla.org/mozilla-central/rev/9497f8c1a115

 Bug 895135. Avoid normalizing gradients to a smaller interval than 0..1. r=roc
author	Jeff Muizelaar <[e-mail]>
	Sat Jul 20 08:58:10 2013 -0400 (at Sat Jul 20 08:58:10 2013 -0400)

Comment 24

[Mats]

BUMP: This is still reproducible, in latest FF (Mac + PC) with Firebug or native FF dev tools open.

Comment 25

[Mats]

Attachment: New profile dump

Comment 26

[Steven Michaud [:smichaud] (Retired)]

> (Mac + PC)

Comment 27

[Milan Sreckovic [:milan] (needinfo for best results)]

Benoit, can you look at the two profiles - the same bug that came back, or did we regress linear gradients?

Comment 28

[Benoit Girard (:BenWa)]

I can't reproduce the problem on my end. I see that there's siesta-all.js that takes over 2 seconds to receive from the network. There's very brief period of page load that pause the browser' main loop (~200ms).

In your profile I see some time spent in Firebug. Can you test again with Firebug and let me know exactly what you think is bad. Is the browser locked up or are you spending too much time in connecting? For me I spend a total of ~4 seconds waiting on the network. Particularly the resources not served from a CDN seem to be very slow.

Comment 29

[Benoit Girard (:BenWa)]

without Firebug*, preferably with a clean profile.

Comment 30

[Mats]

I see this running my local copy of Siesta too, with all scripts served from localhost. Browser frozen for about 15-20 seconds. Would it help if I gave you a screenshare demo over Skype? My id is mats.bryntse.

Comment 31

[Benoit Girard (:BenWa)]

Perhaps, but first can you grab another profile but first set the following preferences with the Gecko Profiler add-on:
profiler.interval: 50
profiler.entries: 250000

Stop/Start the profiler after tweaking these preferences, load the page and hit Analyze within 2-3 seconds after the page load.

This will give a profile that will last the full 20 seconds of page load you experience.

Comment 32

[Mats]

Attachment: saWI0Oyq.dms

profile with Jank only set

Comment 33

[Mats]

Like this?

http://people.mozilla.org/~bgirard/cleopatra/#report=17b38085c223353887de9f40632644174e282fa0

Comment 34

[Benoit Girard (:BenWa)]

Unfortunatly the main thread was not sampled at all =\. I've never seen this before. Try unselecting 'Multi-thread' and profiling again.

Comment 35

[Mats]

Multi-thread was never enabled when I did these profiles. One more try, please ping me on Skype if this log doesn't help. http://people.mozilla.org/~bgirard/cleopatra/#report=c67a64bffe8c1d96d20bb4fa99994091bb289dd0

Comment 36

[Benoit Girard (:BenWa)]

We had a quick conversation on Skype. The remaining performance issues are only with Firebug open. My understanding is Firebug is using old APIs that aren't maintained anymore and that the new Firebug extension will move to complement the built-in devtools:

https://getfirebug.com/wiki/index.php/Firebug.next

Resolving as WORKSFORME since the original gradient issues was resolved but the patch landed in another bug.