diff mbox series

GIT-VERSION-GEN: allow it to be run in parallel

Message ID pull.1850.git.1736432663587.gitgitgadget@gmail.com (mailing list archive)
State Superseded
Headers show
Series GIT-VERSION-GEN: allow it to be run in parallel | expand

Commit Message

Johannes Schindelin Jan. 9, 2025, 2:24 p.m. UTC
From: Johannes Schindelin <johannes.schindelin@gmx.de>

"Why would one want to run it in parallel?" I hear you ask. I am glad
you are curious, because a curious story is what it is, indeed.

The `GIT-VERSION-GEN` script is quite a pillar of Git's source code,
with most lines being unchanged for the past 15 years. Until the v2.48.0
release candidate cycle.

Its original purpose was to generate the version string and store it in
the `GIT-VERSION-FILE`.

This paradigm changed quite dramatically when support for building with
Meson was introduced. Most crucially, a38edab7c88b (Makefile: generate
doc versions via GIT-VERSION-GEN, 2024-12-06) changed the way the
documentation is built by using the `GIT-VERSION-GEN` file to write out
the `asciidocor-extensions.rb` and `asciidoc.conf` files with now
hard-coded version strings.

Crucially, the Makefile rule to generate those files needs to be run in
every build because `GIT_VERSION` could have been specified in the
`make` command-line, which would require these files to be modified.

This introduced a surprising race condition!

And this is how that race surfaces: When calling `make -j2 html man`
from the top-level directory (a variant of which is invoked in Git for
Windows' release process), two sub-processes are spawned, a `make -C
Documentation html` one and a `make -C Documentation man` one. Both run
the rule to (re-)generate `asciidoctor-extensions.rb` or
`asciidoc.conf`, invoking `GIT-VERSION-GEN` to do so. That script first
generates a temporary file (appending the `+` character to the
filename), then looks whether it contains something different than the
already existing file (if it exists, that is), and either replaces it if
needed, or removes the temporary file. If one of the two parallel
invocations removes that temporary file before the other can compare it,
or even worse: if one tries to replace the target file just after the
other _started_ writing the temporary file (but did not finish writing
it yet), that race condition now causes bad builds.

This may sound highly theoretical, but due to the design of Git's build
process, Git for Windows is forced to use a (slow) POSIX emulation layer
to run that script and in the blink of an eye it becomes very much not
theoretical at all. See Exhibit A: These GitHub workflow runs failed
because one of the two competing `make` processes tried to remove the
temporary file when the other process had already done so:

https://github.com/git-for-windows/git-sdk-32/actions/runs/12663456654
https://github.com/git-for-windows/git-sdk-32/actions/runs/12683174970
https://github.com/git-for-windows/git-sdk-64/actions/runs/12649348496

While it is undesirable to run this script over and over again,
certainly when this involves above-mentioned slow POSIX emulation layer,
the stage of the release cycle in which we are presently finding
ourselves does not lend itself to a re-design where this script could be
run once, and once only, but instead dictates that a quick and reliable
work-around be implemented that prevents the race condition without
changing the overall architecture of the build process.

This patch does that: By using a filename suffix for the temporary file
which is based on the currently-executing script's process ID, We
guarantee that the two competing invocations cannot overwrite or remove
each others' temporary files.

Incidentally, this also fixes something else: The `+` character is
not even a valid filename character on Windows. The only reason why Git
for Windows did not need this is that above-mentioned POSIX emulation
layer also plays a couple of tricks with filenames (tricks that are not
interoperable with regular Windows programs, though), and previous
attempts to remedy this in git/git were unsuccessful, see e.g.
https://lore.kernel.org/git/pull.216.git.gitgitgadget@gmail.com/

This commit fixes one of the issues that are currently delaying Git for
Windows v2.48.0-rc2.

Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
---
    GIT-VERSION-GEN: allow it to be run in parallel

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1850%2Fdscho%2Fasciidoctor-extensions-gen-race-work-around-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1850/dscho/asciidoctor-extensions-gen-race-work-around-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/1850

 GIT-VERSION-GEN | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)


base-commit: a60673e9252b08d4eca90543b3729f4798b9aafd

Comments

Martin Ågren Jan. 9, 2025, 6:33 p.m. UTC | #1
On Thu, 9 Jan 2025 at 15:24, Johannes Schindelin via GitGitGadget
<gitgitgadget@gmail.com> wrote:
> And this is how that race surfaces: When calling `make -j2 html man`
> from the top-level directory (a variant of which is invoked in Git for
> Windows' release process), two sub-processes are spawned, a `make -C
> Documentation html` one and a `make -C Documentation man` one. Both run
> the rule to (re-)generate `asciidoctor-extensions.rb` or
> `asciidoc.conf`, invoking `GIT-VERSION-GEN` to do so.

Nicely described. Indeed, there's a reason recursive make is considered
harmful. This is of course not the time or place for addressing that.

> Incidentally, this also fixes something else: The `+` character is
> not even a valid filename character on Windows. The only reason why Git
> for Windows did not need this is that above-mentioned POSIX emulation
> layer also plays a couple of tricks with filenames (tricks that are not
> interoperable with regular Windows programs, though), and previous
> attempts to remedy this in git/git were unsuccessful, see e.g.
> https://lore.kernel.org/git/pull.216.git.gitgitgadget@gmail.com/

> -       "$INPUT" >"$OUTPUT"+
> +       "$INPUT" >"$OUTPUT".$$
>
> -if ! test -f "$OUTPUT" || ! cmp "$OUTPUT"+ "$OUTPUT" >/dev/null
> +if ! test -f "$OUTPUT" || ! cmp "$OUTPUT".$$ "$OUTPUT" >/dev/null
>  then
> -       mv "$OUTPUT"+ "$OUTPUT"
> +       mv "$OUTPUT".$$ "$OUTPUT"
>  else
> -       rm "$OUTPUT"+
> +       rm "$OUTPUT".$$
>  fi

Our `.gitignore` contains an entry "*+" to ignore this sort of temporary
files. Yes, they're supposed to disappear within a second or so, but
according to f9bbaa384e (Add intermediate build products to .gitignore,
2009-11-08), they can linger after interrupted builds. Maybe separate
tooling built around git could pick up these as untracked files for a
second, causing them to come and go in whatever GUI.

You could use "$OUTPUT"."$$"+ to restore this. That of course
invalidates your remark about "Incidentally, ..." above, but might give
this fix a tiny bit less chance of regressing something somewhere?

Martin
Junio C Hamano Jan. 9, 2025, 8 p.m. UTC | #2
Martin Ågren <martin.agren@gmail.com> writes:

>> ... attempts to remedy this in git/git were unsuccessful, see e.g.
>> https://lore.kernel.org/git/pull.216.git.gitgitgadget@gmail.com/
> ...
> You could use "$OUTPUT"."$$"+ to restore this. That of course
> invalidates your remark about "Incidentally, ..." above, but might give
> this fix a tiny bit less chance of regressing something somewhere?

Thanks for being careful.

My reading of that old thread cited there tells me that the reason
that previous one failed was mostly because it wasn't being self
consistent and only touched the use of "+" in the Documentation
directory but not what the top-level Makefile did, and also because
it did not adjust .gitignore patterns, so it is good that somebody
actually read the cited thread and made sure this time we do better.

Again, I was not opposed to moving from "+" to something else that
is equally short-and-sweet, and I still am not ("~" is a fine suffix
for this kind of thing, for example).  But if we are aiming for a
short-term fix, I think your ".$$+" may make the most sense.

Thanks.
diff mbox series

Patch

diff --git a/GIT-VERSION-GEN b/GIT-VERSION-GEN
index 6d1cb66d69a..5b49e2d72fb 100755
--- a/GIT-VERSION-GEN
+++ b/GIT-VERSION-GEN
@@ -86,11 +86,11 @@  sed -e "s|@GIT_VERSION@|$GIT_VERSION|" \
 	-e "s|@GIT_BUILT_FROM_COMMIT@|$GIT_BUILT_FROM_COMMIT|" \
 	-e "s|@GIT_USER_AGENT@|$GIT_USER_AGENT|" \
 	-e "s|@GIT_DATE@|$GIT_DATE|" \
-	"$INPUT" >"$OUTPUT"+
+	"$INPUT" >"$OUTPUT".$$
 
-if ! test -f "$OUTPUT" || ! cmp "$OUTPUT"+ "$OUTPUT" >/dev/null
+if ! test -f "$OUTPUT" || ! cmp "$OUTPUT".$$ "$OUTPUT" >/dev/null
 then
-	mv "$OUTPUT"+ "$OUTPUT"
+	mv "$OUTPUT".$$ "$OUTPUT"
 else
-	rm "$OUTPUT"+
+	rm "$OUTPUT".$$
 fi