Re: [PATCH v3 1/3] factor out refresh_and_write_cache function

From: Thomas Gummerer <t.gummerer@gmail.com>
Date: Sep 6, 2019, 14:18 UTC
Message-ID: <20190906141812.GA128436@cat>
In-Reply-To: <xmqqwoemo131.fsf@gitster-ct.c.googlers.com>

On 09/05, Junio C Hamano wrote:

Show 23 quoted lines

> Thomas Gummerer <t.gummerer@gmail.com> writes:
> 
> > Getting the lock for the index, refreshing it and then writing it is a
> > pattern that happens more than once throughout the codebase, and isn't
> > trivial to get right.  Factor out the refresh_and_write_cache function
> > from builtin/am.c to read-cache.c, so it can be re-used in other
> > places in a subsequent commit.
> >
> > Note that we return different error codes for failing to refresh the
> > cache, and failing to write the index.  The current caller only cares
> > about failing to write the index.  However for other callers we're
> > going to convert in subsequent patches we will need this distinction.
> >
> > Signed-off-by: Thomas Gummerer <t.gummerer@gmail.com>
> > ---
> >  builtin/am.c | 16 ++--------------
> >  cache.h      | 16 ++++++++++++++++
> >  read-cache.c | 19 +++++++++++++++++++
> >  3 files changed, 37 insertions(+), 14 deletions(-)
> 
> I think this goes in the right direction, but obviously conflicts
> with what Dscho wants to do in the builtin-add-i series, and needs
> to be reconciled by working better together.

Oops, I didn't realize there was another series in flight that also introduces 'repo_refresh_and_write_index'. Probably should have done a test merge of this with pu.

Show 7 quoted lines

> For now, I'll eject builtin-add-i and queue this for a few days to
> give it a bit more exposure, but after that requeue builtin-add-i
> and discard these three patches.  By that time, hopefully you two
> would have a rerolled version of this one and builtin-add-i that
> agree what kind of refresh-and-write-index behaviour they both want.
>
> The differences I see that need reconciling are:

Thanks for writing these down.

>  - builtin-add-i seems to allow 'gentle' and allow returning an
>    error when we cannot open the index for writing by passing false
>    to 'gentle'; this feature is not used yet, though.

Right, and if gentle is set to false, it avoids writing the index, which seems fine from my perspective.

Show 5 quoted lines

>  - This version allows to pass pathspec, seen and header_msg, while
>    the one in builtin-add-i cannot limit the part of the index
>    getting refreshed with pathspec.  It wouldn't be a brain surgery
>    to use this version and adjust the caller (there only is one) in
>    the builtin-add-i topic.

'pathspec', 'seen' and 'header_msg' are not used in my version either, I just implemented it for completeness and compatibility. So I'd be fine to do without them.

Show 10 quoted lines

>  - This version does not write the index back when refresh_index()
>    returns non-zero, but the one in builtin-add-i ignores the
>    returned value.  I think, as a performance measure, it probably
>    is a better idea to write it back, even when the function returns
>    non-zero (the local variable's name is has_errors, but having an
>    entry in the index that does not get refreshed is *not* an error;
>    e.g. an unmerged entry is a normal thing in the index, and as
>    long as we refreshed other entries while having an unmerged and
>    unrefreshable entry, we are making progress that is worth writing
>    out).

I'm happy with writing the index back even if there are errors. However I think we still need the option to get the return code from 'refresh_index()', as some callers where I'm using 'refresh_and_write_index()' in this series behave differently depending on its return code.

There's two more differences between the versions:

 - The version in my series allows passing in write_flags to be passed
   to write_locked_index, which is required to convert the callers in
   builtin/merge.c.

 - Dscho's version also calls 'repo_read_index_preload()', which I
   don't do in mine.  Some callers don't need to do that, so I think it
   would be nice to keep that outside of the
   'repo_refresh_and_write_index()' function.

I can think of a few ways forward here:

 - I incorporate features that are needed for the builtin-add-i series
   here, and that is rebased on top of this series.

 - We drop the first two patches of this series, so we only fix the
   problems in 'git stash' for now.  Later we can have a refactoring
   series that uses repo_refresh_and_write_index in the places we
   converted here, once the dust of the builtin-add-i series settled.

 - I rebase this on top of builtin-add-i.

I'm happy with either of the first two, but less so with the last option. I was hoping this series could potentially go to maint as it was a bugfix, which we obviously can't do with that option.

Dscho, what do you think? :)

Show 24 quoted lines

> Thanks.
> 
> > +int repo_refresh_and_write_index(struct  repository *repo,
> > +				 unsigned int refresh_flags,
> > +				 unsigned int write_flags,
> > +				 const struct pathspec *pathspec,
> > +				 char *seen, const char *header_msg)
> > +{
> > +	struct lock_file lock_file = LOCK_INIT;
> > +
> > +	repo_hold_locked_index(repo, &lock_file, LOCK_DIE_ON_ERROR);
> > +	if (refresh_index(repo->index, refresh_flags, pathspec, seen, header_msg)) {
> > +		rollback_lock_file(&lock_file);
> > +		return 1;
> > +	}
> > +	if (write_locked_index(repo->index, &lock_file, COMMIT_LOCK | write_flags))
> > +		return -1;
> > +	return 0;
> > +}
> > +
> > +
> >  int refresh_index(struct index_state *istate, unsigned int flags,
> >  		  const struct pathspec *pathspec,
> >  		  char *seen, const char *header_msg)