633794 - Pass --num-callers=50 to Valgrind

Status: RESOLVED FIXED

(Release Engineering :: General, defect, P3)

Description

[Phil Ringnalda (:philor)]

My feeble flailings in bug 631950 were the result of totally not understanding anything at all. Because by default Valgrind shows no more than 12 frames, what I was thinking were leaks where we weren't involved at all were just things where there were more than 12 frames between the alloc and us. To be sure I know what parts of us are the ones winding up looking like gtk is leaking, I need to see one log on each of linux and linux64 with --num-callers=50 (or whatever the highest number we can use without melting the slave turns out to be, 50 is just the maximum possible according to the docs).

Comment 1

[John Ford [:jhford] CET/CEST Berlin Time]

Phil, is this something that you can do on your local machine?

Comment 2

[Chris AtLee [:catlee]]

Attachment: Add --num-callers=50

Comment 3

[Phil Ringnalda (:philor)]

Comment on attachment 512934 [details] [diff] [review]
Add --num-callers=50

Perfect.

Comment 4

[Chris AtLee [:catlee]]

Comment on attachment 512934 [details] [diff] [review]
Add --num-callers=50

http://hg.mozilla.org/build/tools/rev/04faf34d30f0

Comment 5

[Phil Ringnalda (:philor)]

The local machine problem is that I'd need a pair of VMs pretty much identical to the actual slaves, at least for the dozens of system libraries that the leaks pass through, and I'd bet that nobody really knows how outdated the public downloadable refplatform VMs are, or how outdated the directions on the wiki about how they used to be prepped are, or how to explain how to figure out what puppet does to them now. The hurdle where I'd have to download two VMs over fringe 3G we could surmount, I could cash in some karma and get someone in MV to burn them on a DVD and mail it to me, but getting me from there to an actual clone of a build slave would probably be more work than anyone wants to take on.

Comment 6

[Chris Cooper [:coop] (he/him)]

catlee/philor: Is there anything left to do here? Do we need to revert the patch once the tests are run, or are we just going to keep the new num-callers?

Comment 7

[Phil Ringnalda (:philor)]

We were talking about just keeping it (the results are massively more useful) as long as it doesn't cause the slaves to melt, which it didn't for a couple of runs yesterday. I don't know whether there haven't been any since then because slaves are melting or because there haven't been enough free slaves since.

Comment 8

[Phil Ringnalda (:philor)]

Seems to be working fine.