Re: patch-2.7.3 no longer applies relative symbolic link patches

[Josh Boyer <jwboyer@fedoraproject.org>]

On Mon, Jan 26, 2015 at 3:44 PM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
> On Mon, Jan 26, 2015 at 8:32 AM, Josh Boyer <jwboyer@fedoraproject.org> wrote:
>>
>> I went to do the Fedora 3.19-rc6 build this morning and it failed in
>> our buildsystem with:
>>
>> + '[' '!' -f /builddir/build/SOURCES/patch-3.19-rc6.xz ']'
>> + case "$patch" in
>> + unxz
>> + patch -p1 -F1 -s
>> symbolic link target '../../../../../include/dt-bindings' is invalid
>> error: Bad exit status from /var/tmp/rpm-tmp.mWE3ZL (%prep)
>
> Ugh. I don't see anything we can do about this on the git side, and I
> do kind of understand why 'patch' would be worried about '..' files.
> In a perfect world, patch would parse the filename and see that it
> stays within the directory structure of the project, but that is a
> rather harder thing to do than just say "no dot-dot files".
>
> The short-term fix is likely to just use "git apply" instead of "patch".

Well, that's one fix anyway.  I just removed the hunk from the local
copy of patch-3.19-rc6.xz and added the symlink manually.  See why
below.

> The long-term fix? I dunno. I don't see us not using symlinks, and a
> quick check says that every *single* symlink we have in the kernel
> source tree is one that points to a different directory using ".."
> format. And while I could imagine that "patch" ends up counting the
> dot-dot entries and checking that it's all inside the same tree it is
> patching, I could also easily see patch *not* doing that. So using
> "git apply" _might_ end up being the long-term fix too.

It could, but from a distro perspective that requires either doing
'untar linux-3.N.tar.xz; cd linux-3.N; git add .; git apply
patch-3.N+1-rcX' , or just using a git tree to begin with, which then
makes all of this unnecessary anyway.  Creating a git repo from a
tarball for each build is kind of silly.  Some might say not just
using a git tree to build from to begin with in 2015 is also kind of
silly.

Or did I miss a way that git-apply can take a git patch and apply it
to a tree that isn't a git repo?

> I suspect that if "patch" cannot apply even old-style kernel patches
> due to the symlinks we have in the tree, and people end up having to
> use "git apply" for them, I might end up starting to just use
> rename-patches (ie using "git diff -M") for the kernel.

I'm kind of wondering why we'd generate patches at all if you have to
apply them to a git repo, but maybe people like doing things the
old-fashioned way just for the hell of it.

> I've considered that for a while already, because "patch" _does_ kind
> of understand them these days, although I think it gets the
> cross-rename case wrong because it fundamentally works on a
> file-by-file basis. But if "patch" just ends up not working at all,
> the argument for trying to maintain backwards compatibility gets
> really weak.

Yeah.  I mostly wanted to give people a heads up on the issue.  I'm
sure it's going to impact more than just the kernel.  I think for us
it's mostly limited to the -rcX patches, because once the tarball for
the final release is out the symlink should be created by tar just
fine.

josh