782866 - Pymake: commands that don't use shells don't honour exported PATH on Win32

Status: RESOLVED FIXED

(Firefox Build System :: General, defect)

Description

[Rain [:rain1] (inactive)]

See http://bugs.python.org/issue15451

Comment 1

[Rain [:rain1] (inactive)]

I think the right thing to do here will be to special-case PATH. Gah...

Comment 2

[Rain [:rain1] (inactive)]

See also: http://bugs.python.org/issue8557

Comment 3

[Mike Hommey [:glandium] (high latency)]

all non native commands should be running with sh -c on msys, cf. http://mxr.mozilla.org/mozilla-central/source/build/pymake/pymake/process.py#86

Comment 4

[Rain [:rain1] (inactive)]

For a simple command, e.g. |cl|, shellreason will be None.

Comment 5

[(not currently active) Ted Mielczarek]

Right, that was one of the big Pymake wins (not spawning extraneous shells). How did our build ever work if this is broken? Things like "rm foo" would presumably not work.

Comment 6

[Rain [:rain1] (inactive)]

Sorry, the title was a bit unclear: they do honour the parent's path normally. They just don't honour the PATH variable if it's exported from within the makefile, as we have in our Win32 mozconfigs.

I have a working patch -- struggling to write a test right now.

Comment 7

[Rain [:rain1] (inactive)]

Attachment: patch v1

I believe this should fix things up. I'm running a test build on my build slave to make sure it does.

Comment 8

[Rain [:rain1] (inactive)]

Yes, the build works now.

Comment 9

[Gregory Szorc [:gps]]

(In reply to Ted Mielczarek [:ted] from comment #5)
> Right, that was one of the big Pymake wins (not spawning extraneous shells).
> How did our build ever work if this is broken? Things like "rm foo" would
> presumably not work.

I would throw out that I believe a proper build system deduces absolute paths during a "configure" phase and all paths in the build phase (even for things that are not binaries) are absolute paths.

I threw in *all* paths because we have a lot of code today that works on relative paths. For example, the include paths are relative. In the case of includes, those relative paths get written in the auto-generated makefile dependency files, which means that that .mk file can only be evaluated from a specific directory. This prevents non-recursive evaluation. Doh.

Comment 10

[Rain [:rain1] (inactive)]

Attachment: patch with fixed executable bit

With this fix the test works on Linux as well.

Comment 11

[Gregory Szorc [:gps]]

Comment on attachment 653937 [details] [diff] [review]
patch with fixed executable bit

Review of attachment 653937 [details] [diff] [review]:
-----------------------------------------------------------------

The hack, it hurts.

This is effectively r+. I just want answers to my questions first.

::: pymake/process.py
@@ +172,3 @@
>          try:
> +            if self.env is not None and self.env.has_key('PATH'):
> +                os.environ['PATH'] = self.env['PATH']

I understand modifying os.environ is safe given the current execution model. But, it makes me uneasy.

To help me rest better, how do you feel about setting a module-level flag when running and asserting in run() that it isn't set? i.e. if we ever change the execution model of pymake, I want this to die loudly.

::: tests/subprocess-path.mk
@@ +1,4 @@
> +#T gmake skip
> +#T grep-for: "2f7cdd0b-7277-48c1-beaf-56cb0dbacb24"
> +
> +ifdef __WIN32__

Where does this come from?

@@ +22,5 @@
> +#
> +# Q. Why use an exe and not a batch file?
> +# A. The use cases here were all exe files without the extension. Batch file
> +#    lookup has broken semantics if the .bat extension isn't passed.
> +#  

Whitespace.

Comment 12

[Rain [:rain1] (inactive)]

Attachment: remove threadpool, and make sure we're in a different process

Good catch! This patch means that now all commands are run through child processes.

__WIN32__ is defined in runtests.py.

Comment 13

[(not currently active) Ted Mielczarek]

So um, you mean you regressed bug 602506?

Comment 14

[Rain [:rain1] (inactive)]

Hmm, I guess so. However I did do a full build on a build slave and it took about the same time, give or take a few seconds.

I don't see how this bug can be reconciled with bug 602506 :(

Comment 15

[Rain [:rain1] (inactive)]

On a build slave, a -j4 build with same-process Popen took 32:59, while a -j4 build with different-process Popen took 33:14. I don't think it's a big concern any more.

Comment 16

[Gregory Szorc [:gps]]

Comment on attachment 654042 [details] [diff] [review]
remove threadpool, and make sure we're in a different process

Review of attachment 654042 [details] [diff] [review]:
-----------------------------------------------------------------

The removal of the multiprocessing.dummy foo scared me. But, apparently it doesn't make a difference and Ted's OK with it. So, I'm OK with it.

Comment 17

[Rain [:rain1] (inactive)]

http://hg.mozilla.org/users/bsmedberg_mozilla.com/pymake/rev/5d2775d6a2d5
http://hg.mozilla.org/mozilla-central/rev/21a815b9f4c7