104741 - NSS leaks on shutdown()

Status: RESOLVED FIXED

(NSS :: Libraries, defect, P2)

Description

[David Baron :dbaron: (⌚️UTC-4, no longer working on Mozilla)]

I attempted to merge the leak fixes provided by [e-mail] (Jeff
Hostetler) into a single patch.  It did reduce the shutdown leaks I observed
from NSS by a bit (I think around 200K), although it still didn't fix the
largest leak, which I'll try to file as a separate bug.  I'll attach the patch
and mark the bugs that the patch covers as depending on this one.

Comment 1

[David Baron :dbaron: (⌚️UTC-4, no longer working on Mozilla)]

Attachment: patch

Comment 2

[Wan-Teh Chang]

Assigned the bug to Bob.

Comment 3

[Robert Relyea]

I've expanded the meaning of this bug. The new NSS 3.4 code base leaks even more
on shutdown(). Many of the data structures have changed, but there are still
many that this patch still applies to.

bob

Comment 4

[David Baron :dbaron: (⌚️UTC-4, no longer working on Mozilla)]

cc:ing [e-mail], since these leaks might be a real issue if we want
to start up and shut down NSS multiple times due to profile changes in turbo mode.

Comment 5

[Robert Relyea]

Most of these are not 'real' leaks. Most are data elements that NSS needs to
run. They are created on the fly as needed, and one instance is ever created for
each of these objects. If this data is freed then reallocated on profile changes
and turbo mode it may affect performance in these cases.

One particular type of memory allocation is the free lists. NSS keeps a free
list of keys and objects that it allocates for reuse. This was to keep the
malloc contention down in server products (where were trying to do hundreds of
SSL connections a second and contending on malloc). A better fix for mozilla
would be to reduce the watermark for these free lists. That will reduce steady
state heap usage in PSM as well.

Anyway for 3.4 I'm proceeding under the assumption at the malloc's and frees
will not be a significant performance hit against turbo mode and will remove
noise from the leak detection tools. NOTE: one source of leakage: the Random
number generater, will specifically *NOT* be handled. The cost of keeping one
lock and a buffer around is worth keeping the entropy we collected to that point
-- particularly when Mozilla uses shutdown not to shutdown, but to restart with
a different profile.

bob

Comment 6

[Robert Relyea]

This should be fixed now with the NSS 3.4 landing.

Comment 7

[Wan-Teh Chang]

I am still seeing many one-time leaks in NSS_Init() reported
by BoundsChecker.  We need to examine these memory leaks to
verify that all the one-time leaks reported by [e-mail]
have been fixed in NSS 3.4.

I am reopening this bug, targeting NSS 3.4.1, priority P2.

Comment 8

[Wan-Teh Chang]

Changed the QA contact to Bishakha.

Comment 9

[Robert Relyea]

All the leaks associated with this bug has been fixed. There are separate bug
reports for any new leaks already.