Re: [PATCH] read-cache: tighten checks for do_read_index

Thomas Gummerer [off-list ref] writes:

03f15a7 read-cache: fix reading of split index moved the checks for the
correct order of index entries out of do_read_index.  This loosens the
checks more than necessary.  Re-introduce the checks for the order, but
don't error out when we have multiple stage-0 entries in the index.
Return a flag for the caller instead, if we have multiple stage-0
entries and let the caller decide if we need to error out.

Signed-off-by: Thomas Gummerer <redacted>
---

This is a patch on top of my previous patch, as that one has already
been merged to next.

I am not convinced that this is a good change in the first place.

The original before your fix was wrong exactly because it was too
tightly tied to the implementation of the index file format where
there was only one file whose contents must be sorted, and that is
why it was a broken check in a new world with split-index.  And your
fix in 'next' is the right fix---it makes the verification happen
only on the result is given to the caller for its consumption.

It may be true that entries may have to be sorted in a certain order
when reading the original index file format and also reading some
steps in reading the split-index, but that merely happens to be an
imprementation detail of the two format currently supported, and as
we improve these formats (or even introduce yet another one) in the
longer term, this patch would re-introduce the same issue your
earlier fix corrected, wouldn't it?

On 03/24, Junio C Hamano wrote:

Thomas Gummerer [off-list ref] writes:

quoted

03f15a7 read-cache: fix reading of split index moved the checks for the
correct order of index entries out of do_read_index.  This loosens the
checks more than necessary.  Re-introduce the checks for the order, but
don't error out when we have multiple stage-0 entries in the index.
Return a flag for the caller instead, if we have multiple stage-0
entries and let the caller decide if we need to error out.

Signed-off-by: Thomas Gummerer <redacted>
---

This is a patch on top of my previous patch, as that one has already
been merged to next.

I am not convinced that this is a good change in the first place.

The original before your fix was wrong exactly because it was too
tightly tied to the implementation of the index file format where
there was only one file whose contents must be sorted, and that is
why it was a broken check in a new world with split-index.  And your
fix in 'next' is the right fix---it makes the verification happen
only on the result is given to the caller for its consumption.

It may be true that entries may have to be sorted in a certain order
when reading the original index file format and also reading some
steps in reading the split-index, but that merely happens to be an
imprementation detail of the two format currently supported, and as
we improve these formats (or even introduce yet another one) in the
longer term, this patch would re-introduce the same issue your
earlier fix corrected, wouldn't it?

Yes, after looking at it again I completely agree.  Sorry for the noise.