Re: [PATCH di/fast-import-deltified-tree] Windows: define S_ISUID properly
- From
Dmitry Ivankov <divanorama@gmail.com>
- Date
- Sep 21, 2011, 13:44 UTC
- Message-ID
- <CA+gfSn8VaTLcrLqb3HGJpjL5WJMHZz4kPPUtyeHPCsdmO5iU8g@mail.gmail.com>
- In-Reply-To
- <7vfwjq2hoo.fsf@alter.siamese.dyndns.org>
On Wed, Sep 21, 2011 at 6:08 PM, Junio C Hamano <gitster@pobox.com> wrote:
Show 30 quoted lines
> Dmitry Ivankov <divanorama@gmail.com> writes: > >> Johannes Sixt <j.sixt <at> viscovery.net> writes: >>> >>> From: Johannes Sixt <j6t <at> kdbg.org> >>> >>> 8fb3ad76 (fast-import: prevent producing bad delta) introduced the first >>> use of S_ISUID. Since before this commit the value was irrelevant, we had >>> only a dummy definition in mingw.h. But beginning with this commit the >>> macro must expand to a reasonable value. Make it so. >>> #define S_ISVTX 0 >>> ... >> Ow, it's awkward that the issue was discussed in [1] but slipped and nobody >> noticed, especially me being a patch sender. >> >> If we choose patch from [1] I'd also change a comment to smth like >> /* >> * We abuse the 04000 bit on directories to mean "do not delta". >> * It is a S_ISUID bit on setuid platforms and an unused bit on >> * non-setuid platforms supported in git. In either case git ignores >> * the bit, so it's safe to abuse it locally. >> */ >> >> [1] http://thread.gmane.org/gmane.comp.version-control.git/179223/focus=179762 > > I think that the fix from Jonathan to stop abusing S_ISUID is much more > preferrable; the Windows platform shouldn't have to worry about this. > > And it would be even better to use a value that does not overlap with the > usual bits for do-not-delta bit if possible.
Depends on what is a usual bit. I'll use linux defines for mode bits.
There are S_ISVTX, S_ISUID completely unused in git, S_ISGID is used somehow. 9 lower rwx bits are used as well as S_IFREG and S_IFDIR. Remaining are S_IFIFO (used somehow), S_IFCHR (part of GITLNK).
S_ISUID in fast-import input stream isn't accepted, so the only danger is it may come from a tree object (but not yet, git-fsck doesn't allow it). Or if there appears a platform with different S_I{F,S}* definitions, which will break more git parts than just fast-import.
I remember there was a thread concerning platform vs git-core mode bits. I'd just use hard-coded 04000 bit in fast-import as a hot-fix and leave the rest for the bits topic. With S_ISUID in a comment near 04000 it'll be a grep-able hard-coded constant, so it should be ok.