691309 - Monitor all main thread IO

Status: NEW

(Core :: General, defect)

Description

[David Teller [:Yoric] - still alive but not very active]

We are not the only ones hunting down I/O in the main thread. Android offers a developer tool to monitor I/O happening in the main thread and, essentially, kill off applications that have any: http://www.talkandroid.com/23545-strict-mode-in-android-2-3-gingerbread-helps-developers-make-better-leaner-apps/

I do not think we have anything quite equivalent for Mozilla, although perhaps Instruments or XPerf can do something related. I am wondering if we could add a build mode that provides the same kind of feature.

Comment 1

[David Teller [:Yoric] - still alive but not very active]

Attachment: Monitoring PRLog on main thread

Attaching a small utility patch that makes any PRLog launched from the main thread easier to filter out.

Comment 2

[David Teller [:Yoric] - still alive but not very active]

With the profiler, we now have ways to do this by instrumenting low-level functions performing main thread IO.

So this bug is now about instrumenting all low-level IO functions with SAMPLE_LABEL("IO", "FunctionName"). This will let us find all main thread IO by just filtering upon "IO::" in Cleopatra.

More info about SAMPLE_LABEL: https://developer.mozilla.org/en-US/docs/Performance/Profiling_with_the_Built-in_Profiler#Pseudostack
And a few examples: http://dxr.mozilla.org/search.cgi?tree=mozilla-central&string=SAMPLE_LABEL

Comment 3

[David Teller [:Yoric] - still alive but not very active]

We can start by labelling as IO:
- the relevant methods of nsLocalFileUnix, nsLocalFileWin;
- the relevant functions of sqlite.

Or we can try and locate all Posix and Win32 IO functions.

Comment 4

[u415160]

I'll start working on this bug. Feel free to assign to me.

Comment 5

[David Teller [:Yoric] - still alive but not very active]

Actually, BenWa has advanced a lot on the more general question of monitoring I/O since the bug was filed. BenWa, can Richard help you at this stage?

Comment 6

[Benoit Girard (:BenWa)]

We talked about this during the snappy work week:
- Profile can highlight IO functions. This will find slow IO calls but the probability of it being missed decreases as the runtime for the IO is less then the sampling interval.
- We plan on intercepting certain Gecko IO calls that will cover most IO calls in practice. We will instrument the code to repo this to the profiler and telemetry.

Nathan what are the calls we are interested in watching?

Comment 7

[Nathan Froyd [:froydnj]]

(In reply to Benoit Girard (:BenWa) from comment #6)
> Nathan what are the calls we are interested in watching?

We were planning on watching:

- Wrapper functions in storage/src/TelemetryVFS.cpp;
- Calls made by OS.File;
- Calls made by NSPR (PR_Read/PR_Write/etc.).

Those three categories might not catch everything, but they ought to catch the majority of interesting cases.

Comment 8

[David Teller [:Yoric] - still alive but not very active]

Fwiw, barring any bug, OS.File should not be able to do main thread I/O, so they might not need watching – at least on the main thread.

Comment 9

[David Teller [:Yoric] - still alive but not very active]

Richard, if you are still interested in this bug, you should synchronize with BenWa and froydnj over irc.

Comment 10

[u415160]

Unfortunately i won't be able to look at this anymore.