[Mesa-dev] [PATCH] Mesa: Fix a race in the build

Dan Nicholson dbn.lists at gmail.com
Wed Jun 13 08:03:41 PDT 2012


On 6/13/12, Mike Frysinger <vapier at chromium.org> wrote:
> On Wed, Jun 13, 2012 at 8:13 AM, Dan Nicholson <dbn.lists at gmail.com> wrote:
>
>> On Tue, Jun 12, 2012 at 3:52 PM, Mike Frysinger <vapier at chromium.org>
>> wrote:
>> > On Tue, Jun 12, 2012 at 6:47 PM, Dan Nicholson <dbn.lists at gmail.com>
>> wrote:
>> >>
>> >> On 6/12/12, Stéphane Marchesin <marcheu at chromium.org> wrote:
>> >> > From: Mike Frysinger <vapier at chromium.org>
>> >> >
>> >> > The intent of the message above it is right (we need to build those
>> >> > dependencies in that order) but the implementation is wrong, as it
>> >> > can be executed in any order. To enforce the order, invoke make
>> >> > multiple times.
>> >> > ---
>> >> >  src/mesa/Makefile |    8 ++++++--
>> >> >  1 files changed, 6 insertions(+), 2 deletions(-)
>> >> >
>> >> > diff --git a/src/mesa/Makefile b/src/mesa/Makefile
>> >> > index b0b461f..c436890 100644
>> >> > --- a/src/mesa/Makefile
>> >> > +++ b/src/mesa/Makefile
>> >> > @@ -33,8 +33,12 @@ MESA_CXXFLAGS := $(LLVM_CFLAGS) $(CXXFLAGS)
>> >> >       $(CC) -c -o $@ $< $(MESA_CPPFLAGS) $(MESA_CFLAGS)
>> >> >
>> >> >  # Default: build dependencies, then asm_subdirs, GLSL built-in lib,
>> >> > -# then convenience libs (.a) and finally the device drivers:
>> >> > -default: $(DEPENDS) asm_subdirs $(MESA_LIBS) driver_subdirs
>> >> > +# then convenience libs (.a) and finally the device drivers.
>> >> > +# To ensure ordering, we have to invoke make each time:
>> >> > +default: $(DEPENDS)
>> >> > +     $(MAKE) asm_subdirs
>> >> > +     $(MAKE) $(MESA_LIBS)
>> >> > +     $(MAKE) driver_subdirs
>> >> >
>> >> >  .PHONY: main/git_sha1.h.tmp
>> >> >  main/git_sha1.h.tmp:
>> >>
>> >> The alternative that's more correct is to encode the dependencies in
>> >> the right locations. I.e., make $(DEPENDS) a prereq for asm_subdirs,
>> >> asm_subdirs as prereq for $(MESA_LIBS), etc. That enforces the order
>> >> in the correct place and allows you to do something like "make
>> >> driver_dirs" and have it work correctly. Some of that is already
>> >> there, and it would be useful to know which part is racing. I suppose
>> >> if no one is going to do that work, though...
>> >
>> >
>> > the trouble is that w/out GNU make, you can't declare a dependency as
>> > order-only.  so by having $(MESA_LIBS) depend on asm_subdirs (which is
>> > a
>> > PHONY target), things end up getting regenerated all the time.
>>
>> Right, PHONY asm_subdirs does break that. Looking back into this now,
>> I'm surprised it's broken. This did work for a long time. The reason
>> is because MESA_LIBS will not be built until asm_subdirs is completed.
>> This is because MESA_OBJECTS and MESA_GALLIUM_OBJECTS contain
>> MESA_ASM_FILES, which are what's produced by descending through
>> asm_subdirs. So, make will have to wait for asm_subdirs to complete
>> before it can try to link libmesa.a or libgallium.a. I suspect what's
>> broken this now is prefixing all the sources in sources.mak with
>> $(SRCDIR). At least, I'd really like to see an actual error.
>>
>
> iirc, the debug+fix session happened on irc, so i didn't post any of the
> errors, and the log files we were reviewing no longer exist on the bots.
>
> the error is related to not building in subdirs anymore.  look at
> src/mesa/{x86,x86-64}/Makefile -- you'll see that there's a matypes.h
> header generated, and the objects declare dependencies on that.  but that
> isn't bubbled up, so if you build asm_subdirs and $(MESA_LIBS) in parallel,
> the objects that need matypes.h are built before matypes.h is generated.

Hmm, I recall that had worked at one point, but looking again I don't
see how. I think your patch is correct and essentially does the same
thing as automake in serializing the directory recursion but without
all the generic syntax. Essentially this is like "SUBDIRS = x86 x86-64
. drivers" on x86_64 with the extra nit that DEPENDS happens at the
beginning and not at the same time as the . directory.

Reviewed-by: Dan Nicholson <dbn.lists at gmail.com>


More information about the mesa-dev mailing list