13451 - Massive increase in footprint on Unix

Status: VERIFIED FIXED

(SeaMonkey :: General, defect, P1)

Description

[Mike Shaver (:shaver emeritus)]

We use 30M or more just loading the initial (test1) page in viewer.

shaver    3191 16.7  5.4 30488 13952 pts/3   S    13:17   0:03 viewer

This is new since last Friday, and doesn't appear to happen on Mac, dunno about
Windows.

Memory footprint checkpoints from various points during the run (first number is
before the call, second number is after):

NS_InitXPCOM:
 3880   3448

NS_SetupRegistry_1
 3880   4784

nsICmdLineService:
 4784   5988

nsIDOMAppCoresManager:
 5988   6196

nsComponentManager::CreateInstance(kAppShellCID
 6280   9248

mAppShell->Create(0, nsnull)
 9248  10048

nsServiceManager::GetService(kWindowMediatorCID
10048  11068

CreateHiddenWindow
11068  18236

There's still some 12 MB unaccounted for, but I wasn't able to easily step
through it.  Purify on NT doesn't show any incredible lossage, but bug #13030
keeps us from testing on Unix.

Comment 1

[Mike Shaver (:shaver emeritus)]

Adding 13030 as a dependency.

Comment 2

[Mike Shaver (:shaver emeritus)]

Looks like I'm leading the charge on this one.  Waterson, do you have tools for
tracking memory usage?

I'm going to spam the world in a few minutes to tell them that time itself will
wait for this bug.

Comment 3

[Suresh Duddi (gone)]

I am using apprunner. I see it in apprunner too. Here is what I saw:

10324 rv = nsServiceManager::GetService(kWindowMediatorCID,
28264 CreateHiddenWindow();

I tracked vm using:
	$ grep Vm /proc/5731/status

Comment 4

[Mike Shaver (:shaver emeritus)]

We've only got 2.5-3.0M on the heap and 14M (!) or so in read-only mappings, but
then there's this enormous (10M) /dev/zero mapping in necko.  Warren, what's the
reasoning behind such a large buffer?  Can we tune it down?

If we can get it to the point where the NTO guys (who don't have mmap, and are
running in a tight-memory world) are back in business, I'll downgrade this bug
and let the tree open tomorrow.

Comment 5

[Warren Harris]

Sure, we can eliminate it completely if we want. What platforms are the problem?
On Windows, it's just reserved, but not mapped. Is there some way to do that on
Unix?

There's a related bug on this: 11062 (maybe a dup).

Comment 6

[Warren Harris]

I've disabled the page manager until we have more time to make it work
properly.

Comment 7

[leger]

Checked with warren.  Ok to mark Verified.  Doing so now.