[PATCH] midx.c: clear auxiliary MIDX files first

[PATCH] midx.c: clear auxiliary MIDX files first

[Taylor Blau]

Since they were added in c528e17966 (pack-bitmap: write multi-pack
bitmaps, 2021-08-31), the routine to remove MIDXs removed the
multi-pack-index file itself before removing its associated .bitmap and
.rev file(s), if any.

This creates a window where a MIDX's .bitmap file exists without its
corresponding MIDX. If a reader tries to load a MIDX bitmap during that
time, they will get a warning, and the MIDX bitmap code will gracefully
degrade.

Remove this window entirely by removing the MIDX last, and removing its
auxiliary files first.

The order here is important, too. We remove the MIDX's .bitmap file
ahead of its .rev, since callers try and read the .bitmap first. The
.rev file is no longer generated by modern versions of Git, but cleaning
up old ones generated by previous versions of Git is still important to
do.

Signed-off-by: Taylor Blau <me@ttaylorr.com>
---
 midx.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/midx.c b/midx.c
index 3a8dcfe98e..994129aecd 100644
--- a/midx.c
+++ b/midx.c
@@ -1619,12 +1619,12 @@ void clear_midx_file(struct repository *r)
 		r->objects->multi_pack_index = NULL;
 	}
 
-	if (remove_path(midx.buf))
-		die(_("failed to clear multi-pack-index at %s"), midx.buf);
-
 	clear_midx_files_ext(r->objects->odb->path, ".bitmap", NULL);
 	clear_midx_files_ext(r->objects->odb->path, ".rev", NULL);
 
+	if (remove_path(midx.buf))
+		die(_("failed to clear multi-pack-index at %s"), midx.buf);
+
 	strbuf_release(&midx);
 }
 
-- 
2.38.0.16.g393fd4c6db

Re: [PATCH] midx.c: clear auxiliary MIDX files first

[Jeff King]

On Tue, Oct 25, 2022 at 02:25:20PM -0400, Taylor Blau wrote:

> Since they were added in c528e17966 (pack-bitmap: write multi-pack
> bitmaps, 2021-08-31), the routine to remove MIDXs removed the
> multi-pack-index file itself before removing its associated .bitmap and
> .rev file(s), if any.
> 
> This creates a window where a MIDX's .bitmap file exists without its
> corresponding MIDX. If a reader tries to load a MIDX bitmap during that
> time, they will get a warning, and the MIDX bitmap code will gracefully
> degrade.
> 
> Remove this window entirely by removing the MIDX last, and removing its
> auxiliary files first.

We remove that window, but don't we create a new one where a reader may
see the midx but not the bitmap? That won't generate a warning (it just
looks like a midx that never had a bitmap generated), but it will cause
the reader to follow the slow, non-bitmap path.

Ideally this would just be atomic, but short of stuffing the metadata
into the same file, we can't do that. But the replacement of the midx
file itself is atomic, and I'd think everything would (or should at
least) follow from there.

I.e., why are we reading the midx bitmap without having opened the midx
file? Ideally we'd be holding a descriptor for it.

-Peff

Re: [PATCH] midx.c: clear auxiliary MIDX files first

[Derrick Stolee]

On 10/26/22 1:41 AM, Jeff King wrote:
> On Tue, Oct 25, 2022 at 02:25:20PM -0400, Taylor Blau wrote:
> 
>> Since they were added in c528e17966 (pack-bitmap: write multi-pack
>> bitmaps, 2021-08-31), the routine to remove MIDXs removed the
>> multi-pack-index file itself before removing its associated .bitmap and
>> .rev file(s), if any.
>>
>> This creates a window where a MIDX's .bitmap file exists without its
>> corresponding MIDX. If a reader tries to load a MIDX bitmap during that
>> time, they will get a warning, and the MIDX bitmap code will gracefully
>> degrade.
>>
>> Remove this window entirely by removing the MIDX last, and removing its
>> auxiliary files first.
> 
> We remove that window, but don't we create a new one where a reader may
> see the midx but not the bitmap? That won't generate a warning (it just
> looks like a midx that never had a bitmap generated), but it will cause
> the reader to follow the slow, non-bitmap path.

Yes, this is the worrisome direction. The midx is read first, then that
points to the .bitmap file (based on its trailing hash). If the midx isn't
there, then the .bitmap will not be read.

> Ideally this would just be atomic, but short of stuffing the metadata
> into the same file, we can't do that. But the replacement of the midx
> file itself is atomic, and I'd think everything would (or should at
> least) follow from there.

The interesting case here is that this is in clear_midx_file(), which
is called when repacking to a single pack and no longer needing a midx
file. So it's not using the atomic rewrite from the midx writing code,
but instead the "atomic" deletion.

In this case, a reader will check for the midx first, before looking
for individual packs. Further, the new pack is written, but the old
packs have not been deleted (or the midx would be invalid). So the
new code introduces the window where a midx exists without a bitmap,
so some readers will act as if no bitmap exists on-disk.

This was always possible before, too: the midx could be read by a
reader process before the repack process deletes that file. However,
if the reader does not also gain a handle on the corresponding
.bitmap before the repack process deletes that file, too, then the
reader is also left thinking that no .bitmap exists.

I think the old code is more correct, here. The window is slightly
smaller, and the new code creates a window where the filesystem
doesn't need to change for readers to get an imperfect view of
things.

Aside: in these cases where a .bitmap file is not found for a midx,
do we fall back to trying to find a .bitmap file for a pack-file?
That would assuage most of the concerns here about what happens in
this window where the repack has a new .pack/.bitmap pair but the
old midx still exists (without a .bitmap, depending on timing).

Thanks,
-Stolee

Re: [PATCH] midx.c: clear auxiliary MIDX files first

[Taylor Blau]

On Wed, Oct 26, 2022 at 09:31:28AM -0400, Derrick Stolee wrote:
> On 10/26/22 1:41 AM, Jeff King wrote:
> > On Tue, Oct 25, 2022 at 02:25:20PM -0400, Taylor Blau wrote:
> >
> >> Since they were added in c528e17966 (pack-bitmap: write multi-pack
> >> bitmaps, 2021-08-31), the routine to remove MIDXs removed the
> >> multi-pack-index file itself before removing its associated .bitmap and
> >> .rev file(s), if any.
> >>
> >> This creates a window where a MIDX's .bitmap file exists without its
> >> corresponding MIDX. If a reader tries to load a MIDX bitmap during that
> >> time, they will get a warning, and the MIDX bitmap code will gracefully
> >> degrade.
> >>
> >> Remove this window entirely by removing the MIDX last, and removing its
> >> auxiliary files first.
> >
> > We remove that window, but don't we create a new one where a reader may
> > see the midx but not the bitmap? That won't generate a warning (it just
> > looks like a midx that never had a bitmap generated), but it will cause
> > the reader to follow the slow, non-bitmap path.
>
> Yes, this is the worrisome direction. The midx is read first, then that
> points to the .bitmap file (based on its trailing hash). If the midx isn't
> there, then the .bitmap will not be read.

Yes, thinking on it more I agree with this and (the elided) analysis
below.

Let's drop this one. Thanks, both!


Thanks,
Taylor

Re: [PATCH] midx.c: clear auxiliary MIDX files first

[Jeff King]

On Wed, Oct 26, 2022 at 09:31:28AM -0400, Derrick Stolee wrote:

> This was always possible before, too: the midx could be read by a
> reader process before the repack process deletes that file. However,
> if the reader does not also gain a handle on the corresponding
> .bitmap before the repack process deletes that file, too, then the
> reader is also left thinking that no .bitmap exists.

Good point. Neither the writer _nor_ the reader is atomic. :)

> I think the old code is more correct, here. The window is slightly
> smaller, and the new code creates a window where the filesystem
> doesn't need to change for readers to get an imperfect view of
> things.

Yeah, I agree that the old code is nicer in that the window is a little
smaller.

Do we want to do something about the warning, though? Falling back to a
slow path sucks, of course, but racily generating user-visible warnings
for something that is not actually a problem is also not great.

> Aside: in these cases where a .bitmap file is not found for a midx,
> do we fall back to trying to find a .bitmap file for a pack-file?
> That would assuage most of the concerns here about what happens in
> this window where the repack has a new .pack/.bitmap pair but the
> old midx still exists (without a .bitmap, depending on timing).

Yes, we should. Code paths generally go through open_bitmap(), which
tries the midx first, then looks for pack bitmaps.

And in that sense, the race after this patch is fairly harmless. We fail
to see the midx bitmap, but we'll see the pack one (which must have been
created before we deleted the midx one, assuming this is a "git repack
-adb" or equivalent).

Is that also true of the race before this patch? I'm not sure which
warning is being generated, but if it's in open_midx_bitmap_1(), then
the same logic would apply.

-Peff