Re: [PATCH v3 18/23] Makefiles: add and use wildcard "mkdir -p" template

From: "Ævar Arnfjörð Bjarmason" <avarab@gmail.com>
To: Mike Hommey <mh@glandium.org>
Cc: git@vger.kernel.org, Junio C Hamano <gitster@pobox.com>,
	Phillip Wood <phillip.wood123@gmail.com>,
	Jeff King <peff@peff.net>, Dan Jacques <dnj@google.com>,
	Eric Wong <e@80x24.org>, Jonathan Nieder <jrnieder@gmail.com>
Subject: Re: [PATCH v3 18/23] Makefiles: add and use wildcard "mkdir -p" template
Date: Wed, 17 Nov 2021 10:26:27 +0100	[thread overview]
Message-ID: <211117.86wnl76sal.gmgdl@evledraar.gmail.com> (raw)
In-Reply-To: <20211117025148.awdha4udu5kmzwbe@glandium.org>

On Wed, Nov 17 2021, Mike Hommey wrote:

> On Tue, Nov 16, 2021 at 01:00:18PM +0100, Ævar Arnfjörð Bjarmason wrote:
>> Add a template to do the "mkdir -p" of $(@D) (the parent dir of $@)
>> for us, and use it for the "make lint-docs" targets I added in
>> 8650c6298c1 (doc lint: make "lint-docs" non-.PHONY, 2021-10-15).
>> 
>> As seen in 4c64fb5aad9 (Documentation/Makefile: fix lint-docs mkdir
>> dependency, 2021-10-26) maintaining these manual lists of parent
>> directory dependencies is fragile, in addition to being obviously
>> verbose.
>> 
>> I used this pattern at the time because I couldn't find another method
>> than "order-only" prerequisites to avoid doing a "mkdir -p $(@D)" for
>> every file being created, which as noted in [1] would be significantly
>> slower.
>> 
>> But as it turns out we can use this neat trick of only doing a "mkdir
>> -p" if the $(wildcard) macro tells us the path doesn't exist. A re-run
>> of a performance test similar to thatnoted downthread of [1] in [2]
>> shows that this is faster, in addition to being less verbose and more
>> reliable (this uses my "git-hyperfine" thin wrapper for "hyperfine"[3]):
>> 
>>     $ git hyperfine -L rev HEAD~0,HEAD~1 -b 'make -C Documentation lint-docs' -p 'rm -rf Documentation/.build' 'make -C Documentation lint-docs'
>>     Benchmark 1: make -C Documentation lint-docs' in 'HEAD~0
>>       Time (mean ± σ):      2.129 s ±  0.011 s    [User: 1.840 s, System: 0.321 s]
>>       Range (min … max):    2.121 s …  2.158 s    10 runs
>> 
>>     Benchmark 2: make -C Documentation lint-docs' in 'HEAD~1
>>       Time (mean ± σ):      2.659 s ±  0.002 s    [User: 2.306 s, System: 0.397 s]
>>       Range (min … max):    2.657 s …  2.662 s    10 runs
>> 
>>     Summary
>>       'make -C Documentation lint-docs' in 'HEAD~0' ran
>>         1.25 ± 0.01 times faster than 'make -C Documentation lint-docs' in 'HEAD~1'
>> 
>> So let's use that pattern both for the "lint-docs" target, and a few
>> miscellaneous other targets.
>> 
>> This method of creating parent directories is explicitly racy in that
>> we don't know if we're going to say always create a "foo" followed by
>> a "foo/bar" under parallelism, or skip the "foo" because we created
>> "foo/bar" first. In this case it doesn't matter for anything except
>> that we aren't guaranteed to get the same number of rules firing when
>> running make in parallel.
>
> Something else that is racy is that $(wildcard) might be saying the
> directory doesn't exist while there's another make subprocess that has
> already started spawning `mkdir -p` for that directory.
> That doesn't make a huge difference, but you can probably still end up
> with multiple `mkdir -p` runs for the same directory.
>
> I think something like the following could work while avoiding those
> races:
>
> define create_parent_dir_RULE
> $(1): | $(dir $(1)).
> ALL_DIRS += $(dir $(1))
> endef
>
> define create_parent_dir_TARGET
> $(1)/.: $(dir $(1)).
> 	echo mkdir $$(@D)
> endef
>
> $(eval $(call create_parent_dir_RULE, first/path/file))
> $(eval $(call create_parent_dir_RULE, second/path/file))
> # ...
>
> $(foreach dir,$(sort $(ALL_DIRS)),$(eval $(call create_parent_dir_TARGET,$(dir:%/=%))))

I think the "race" just isn't a problem, and makes managing this much
simpler.

I.e. we already rely on "mkdir -p" not failing on an existing directory,
so the case where we redundantly try to create a directory that just got
created by a concurrent process is OK, and as the quoted benchmark shows
is much faster than a similar (but not quite the same as) a
dependency-based implementaiton.

I haven't implemented your solution, but it seems to be inherently more
complex.

I.e. with the one I've got you just stick the "mkdir if needed"
one-liner in each rule, with yours you'll need to accumulate things in
ALL_DIRS, and have some foreach somewhere or dependency relationship to
create those beforehand if they're nested, no?