Re: [PATCH v3] lockfile: add PID file for debugging stale locks

From: Patrick Steinhardt <ps@pks.im>
Date: Jan 5, 2026, 12:23 UTC
Message-ID: <aVutN543nceW0f8W@pks.im>
In-Reply-To: <20251227075025.GC2071715@coredump.intra.peff.net>

On Sat, Dec 27, 2025 at 02:50:25AM -0500, Jeff King wrote:

Show 38 quoted lines

> On Wed, Dec 24, 2025 at 12:24:13PM +0000, Paulo Casaretto via GitGitGadget wrote:
> 
> > For a lock file "foo.lock", the PID file is named "foo~pid.lock". The
> > tilde character is forbidden in refnames and allowed in Windows
> > filenames, which guarantees no collision with the refs namespace
> > (e.g., refs "foo" and "foo~pid" cannot both exist). The file uses a
> > simple key-value format ("pid <value>") following the same pattern as
> > Git object headers, making it extensible for future metadata.
> 
> FWIW, I like this approach, as the earlier collision possibilities in
> earlier iterations made me a bit uncomfortable.
> 
> But then I wondered...
> 
> > The feature is controlled via core.lockfilePid configuration, which
> > accepts per-component values similar to core.fsync:
> > 
> >   - none: Disable for all components (default)
> >   - all: Enable for all components
> >   - index, config, refs, commit-graph, midx, shallow, gc, other:
> >     Enable for specific components
> 
> Do we really need this complexity now? I can see reasons why a user
> might want to switch the feature on (because they want the extra pid
> debugging info) or off (because they do not want the extra filesystem
> operations or potential cleanup hassle of extra stale files).
> 
> But if the collision problems in the ref namespace are now solved, does
> anybody really care about turning it on just for particular subsystems?
> I'd think they would want all or nothing.
> 
> 
> Trying to devil's advocate myself: maybe somebody is concerned about
> filesystem overhead for frequently-written files like refs but not for
> others, like config? That feels unlikely to me, but at least possible.
> 
> But if we were to ditch the subsystem-granularity for config, then all
> of the extra LOCKFILE_* arguments here could just go away.

I tend to agree. Based on my own experience, references are by far the most likely subsystem where one may hit stale lockfiles, so the feature is most useful there. But at the same time, it's also the most expensive subsystem to enable this feature for because we write so many small files there.

Consequently, if anybody cares, they would most likely want to enable this feature for refs. And, if so, the impact on other subsystems would very likely be dwarfed by the refs overhead.

So my approach would be to enable this setting with only "true" and "false" as possible values initially. And if we ever get a use case where somebody would be helped by configuring this per subsystem we can extend the config to learn about subsystems.

But ultimately I don't care too strongly about this. If the author wants to keep the current approach I'm fine with that, too.

Patrick