1273301 - Use DMD to record allocations during a specific segment of the code

Status: RESOLVED INCOMPLETE

(Core :: DMD, defect)

Description

[Andrew McCreight [:mccr8]]

I'm looking at temporary increases in memory usage during IPC serialization/deserialization. I'd like to be able to record all allocations during a certain part of the code, with a StartRecording() and StopRecording() kind of thing. It turns out this is reasonably easy to implement this by misusing dmd::ResetEverything().

The stacks I'm getting are bogus. I'm not sure if I'm just having some problem with DMD, or if the way I'm using ResetEverything is causing issues.

Comment 1

[Andrew McCreight [:mccr8]]

Attachment: Add a way to record memory allocations in one segment.

Comment 2

[Andrew McCreight [:mccr8]]

Attachment: example allocation log

I had to use GCC because Clang was giving me weird stacks.

I might hack up an "Analyze and Reset" method because with my current approach you get some noise from the actual logging process.

Comment 3

[Andrew McCreight [:mccr8]]

Anyways, in the log, you can see pretty easily the redundant allocation that is happening: there are 3 separate buffers of 8,192 bytes, inside CacheFileChunk::EnsureBufSize(), Pickle::Resize() and the nsCString that NS_ReadInputStreamToString() stores data into.

Comment 4

[Andrew McCreight [:mccr8]]

This isn't something that needs to be tracked for e10s, because it is just a local diagnostic tool, but I thought I'd make it block bug 1262918 to give it more visibility.

The patch here is a little problematic. The GCing that DMD does causes a UAF crash with this patch, so I just commented it out. I was also hitting a JS assertion from adding something to the hash table of allocated addresses that is already in there. I have no idea how that could happen.