Re: t3010 broken by 2eac2a4

Eric Sunshine [off-list ref] writes:

On Fri, Aug 23, 2013 at 1:36 AM, Junio C Hamano [off-list ref] wrote:

quoted

Eric Sunshine [off-list ref] writes:

quoted

I sent a patch [1] which resolves the problem, although the solution
is not especially pretty (due to some ugliness in the existing
implementation).

Yeah, thanks.

I tend to agree with you that fixing the "icase" callee not to rely
on having the trailing slash (which is looking past the end of the
given string), instead of working that breakage around on the
caller's side like your patch did, would be a better alternative,
though.

My concern with fixing directory_exists_in_index_icase() to add the
'/' itself was that it would have to copy the string to make space for
the '/', which could be expensive. However, I reworked the code so
that the existing strbufs now get passed to
directory_exists_in_index_icase(), which allows it to add its needed
'/' without duplicating the string. So, the trailing '/' requirement
of directory_exists_in_index_icase() is now a private implementation
detail, placing no burden on the caller.

When 5102c617 (Add case insensitivity support for directories when
using git status, 2010-10-03) added the directories to the name-hash
with trailing slash, there was only a single name hash table to
which both real cache entries and leading directory prefixes are
registered, so it made some sense to register them with trailing
slashes so that we can tell what kind of entry is being returned.

But since 2092678c (name-hash.c: fix endless loop with
core.ignorecase=true, 2013-02-28), these directory entries that are
not the cache entries are kept track of in a separate hashtable,
which makes me wonder if it still makes sense to register
directories with trailing slashes.

And if we stop doing that (and instead if we shrunk the namelen when
an unconverted caller asks for a name with a trailing slash to see
if a directory exists in the index), wouldn't it automatically fix
the directory_exists_in_index_icase()?  It does not need to assume
that dirname[len] has '/'; after all, it may not even be a valid
memory location in the first place.

On Fri, Aug 23, 2013 at 10:15:55AM -0700, Junio C Hamano wrote:

When 5102c617 (Add case insensitivity support for directories when
using git status, 2010-10-03) added the directories to the name-hash
with trailing slash, there was only a single name hash table to
which both real cache entries and leading directory prefixes are
registered, so it made some sense to register them with trailing
slashes so that we can tell what kind of entry is being returned.

But since 2092678c (name-hash.c: fix endless loop with
core.ignorecase=true, 2013-02-28), these directory entries that are
not the cache entries are kept track of in a separate hashtable,
which makes me wonder if it still makes sense to register
directories with trailing slashes.

And if we stop doing that (and instead if we shrunk the namelen when
an unconverted caller asks for a name with a trailing slash to see
if a directory exists in the index), wouldn't it automatically fix
the directory_exists_in_index_icase()?  It does not need to assume
that dirname[len] has '/'; after all, it may not even be a valid
memory location in the first place.

Yeah, I think that is sane overall direction. When I did the sketch that
eventually turned into Karsten's 2092678c, that was one of the goals.
But I did not keep up with his response and the final patch, and I'm not
sure if the slashes still serve some function. So it would definitely
need somebody looking carefully at the current logic.

More details are in this sub-thread:

  http://thread.gmane.org/gmane.comp.version-control.git/215820/focus=216284

-Peff