675216 - Update about:memory's description of heap-committed

Status: RESOLVED FIXED

(Toolkit :: about:memory, defect)

Description

[Justin Lebar (not reading bugmail)]

I keep telling people that it's not committed, non-shared, non-copy-on-write pages. At least, it appears that it can't be, because heap-committed is often many times larger than RSS, and I don't think we've swapped out 4/5 of Firefox in these examples [1, 2].

So what *does* heap-committed measure?  We should figure this out.

[1] Bug 671702 comment 34
[2] Bug 671702 comment 49

Comment 1

[Justin Lebar (not reading bugmail)]

Ah, I think I may have a guess here.  Without DECOMMIT (so, on Linux) jemalloc madvise DONT_NEED's blocks instead of decommitting them.  But those blocks aren't taking up any space in memory.

Comment 2

[Justin Lebar (not reading bugmail)]

Attachment: Patch v1

Comment 3

[Nicholas Nethercote [inactive]]

Comment on attachment 550785 [details] [diff] [review]
Patch v1

Review of attachment 550785 [details] [diff] [review]:
-----------------------------------------------------------------

rs=me.  At this point you understand this stuff much better than I do :)

Comment 4

[Dão Gottwald [:dao]]

backed out from mozilla-inbound for busting builds

Comment 5

[Justin Lebar (not reading bugmail)]

Maemo is picky about a semicolon.

Comment 6

[Justin Lebar (not reading bugmail)]

This push is healthier.  Thanks for the quick backout, Dao!

http://hg.mozilla.org/integration/mozilla-inbound/rev/b0d6e197c741

Comment 7

[Kyle Huey (Exited; not receiving bugmail, old account, do not use)]

http://hg.mozilla.org/mozilla-central/rev/b0d6e197c741

Comment 8

[Christian Ascheberg]

The description says that "heap-committed should equal heap-allocated + heap-unallocated". I am using current nightly on Windows Vista, and this is an example of my about:memory after some browsing:

 78,900,180 B -- heap-allocated
 86,630,400 B -- heap-committed
  3,530,752 B -- heap-dirty
 20,713,834 B -- heap-unallocated

-> heap-committed != heap-allocated + heap-unallocated

Why is that?

Comment 9

[Justin Lebar (not reading bugmail)]

> Why is that?

I'm not sure.  The end of jemalloc_stats() says:

(jemalloc.c)
> #ifndef MALLOC_DECOMMIT
>   stats->committed = stats->mapped;
> #endif

so committed = mapped on Windows, where we define MALLOC_DECOMMIT.

heap-allocated is stats.allocated, and heap-unallocated is stats.mapped - stats.allocated (nsMemoryReporterManager.cpp).

So allocated + unallocated = mapped = committed.

Weird...

Comment 10

[Christian Ascheberg]

(In reply to Justin Lebar [:jlebar] from comment #9)
> Weird...

Can you reproduce the problem? If so, do you want to file a new bug? You obviously know more about the details than I do.

Comment 11

[Justin Lebar (not reading bugmail)]

(In reply to Christian Ascheberg from comment #10)
> Can you reproduce the problem? If so, do you want to file a new bug? You
> obviously know more about the details than I do.

All fixed now in bug 684592.  My mistake was in comment 9; I read #ifndef MALLOC_DECOMMIT as #ifdef!