threads / discuss / 46120

Git v2.13.1 SHA1 very broken

Subject: Git v2.13.1 SHA1 very broken

## tl;dr

26 messages between Jun 5, 2017 and Jul 3, 2017.

replies: 25people: 11as markdown or json

Adam Dinwoodie· Jun 5, 2017, 20:34 UTC · lore

I'm trying to compile Git v2.13.1 to release for Cygwin, but it appears a010391 ("sha1dc: update from upstream", 2017-05-20) is breaking a very significant number of test cases in both 32-bit and 64-bit Cygwin builds.

The first failure is t0000.46 "validate object ID of a known tree"; output with -x and -v is below, although it's not very interesting:

    expecting success:
	    test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
    ++ test ceb282701536fe61bea01075664405caa7d6343f = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
    + test_eval_ret_=1
    + want_trace
    + test t = t
    + test t = t
    + set +x
    error: last command exited with $?=1
    not ok 46 - validate object ID of a known tree
    #
    #               test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
    #

I have no idea where to even begin debugging this, but I'm happy to take pointers / try things out on my box.

Cheers,
Adam
Ævar Arnfjörð Bjarmason· Jun 5, 2017, 21:05 UTC · re: Adam Dinwoodie · lore

Re: Git v2.13.1 SHA1 very broken

On Mon, Jun 5, 2017 at 10:34 PM, Adam Dinwoodie <adam@dinwoodie.org> wrote:
Show 25 quoted lines
> I'm trying to compile Git v2.13.1 to release for Cygwin, but it appears
> a010391 ("sha1dc: update from upstream", 2017-05-20) is breaking a very
> significant number of test cases in both 32-bit and 64-bit Cygwin
> builds.
>
> The first failure is t0000.46 "validate object ID of a known tree"; output with
> -x and -v is below, although it's not very interesting:
>
>     expecting success:
>             test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>
>     ++ test ceb282701536fe61bea01075664405caa7d6343f = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>     + test_eval_ret_=1
>     + want_trace
>     + test t = t
>     + test t = t
>     + set +x
>     error: last command exited with $?=1
>     not ok 46 - validate object ID of a known tree
>     #
>     #               test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>     #
>
> I have no idea where to even begin debugging this, but I'm happy to take
> pointers / try things out on my box.
That looks scary, can you please comment out this:
    #define SHA1DC_ALLOW_UNALIGNED_ACCESS

In sha1dc/sha1.c and see if that helps, alternatively comment out the ifdefs guarded by "#ifdef _MSC_VER" calls in sha1dc/sha1.c

The functional differences between 2.13.0 and 2.13.1 on that platform should be none aside from possibly those changes, unless I've missed something.

Ramsay Jones· Jun 5, 2017, 23:20 UTC · re: Ævar Arnfjörð Bjarmason · lore

Re: Git v2.13.1 SHA1 very broken

On 05/06/17 22:05, Ævar Arnfjörð Bjarmason wrote:
Show 30 quoted lines
> On Mon, Jun 5, 2017 at 10:34 PM, Adam Dinwoodie <adam@dinwoodie.org> wrote:
>> I'm trying to compile Git v2.13.1 to release for Cygwin, but it appears
>> a010391 ("sha1dc: update from upstream", 2017-05-20) is breaking a very
>> significant number of test cases in both 32-bit and 64-bit Cygwin
>> builds.
>>
>> The first failure is t0000.46 "validate object ID of a known tree"; output with
>> -x and -v is below, although it's not very interesting:
>>
>>     expecting success:
>>             test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>>
>>     ++ test ceb282701536fe61bea01075664405caa7d6343f = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>>     + test_eval_ret_=1
>>     + want_trace
>>     + test t = t
>>     + test t = t
>>     + set +x
>>     error: last command exited with $?=1
>>     not ok 46 - validate object ID of a known tree
>>     #
>>     #               test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>>     #
>>
>> I have no idea where to even begin debugging this, but I'm happy to take
>> pointers / try things out on my box.
> 
> That looks scary, can you please comment out this:
> 
>     #define SHA1DC_ALLOW_UNALIGNED_ACCESS
No, that doesn't fix it.
> 
> In sha1dc/sha1.c and see if that helps, alternatively comment out the
> ifdefs guarded by "#ifdef _MSC_VER" calls in sha1dc/sha1.c
This can't possibly make a difference! ;-)
However, rebuilding with:
    $ make OPENSSL_SHA1=YesPlease >out2 2>&1
... make t0000-basic.sh pass just fine, so ...

ATB, Ramsay Jones

Ramsay Jones· Jun 6, 2017, 00:11 UTC · re: Ramsay Jones · lore

Re: Git v2.13.1 SHA1 very broken

On 06/06/17 00:20, Ramsay Jones wrote:
Show 47 quoted lines
> 
> 
> On 05/06/17 22:05, Ævar Arnfjörð Bjarmason wrote:
>> On Mon, Jun 5, 2017 at 10:34 PM, Adam Dinwoodie <adam@dinwoodie.org> wrote:
>>> I'm trying to compile Git v2.13.1 to release for Cygwin, but it appears
>>> a010391 ("sha1dc: update from upstream", 2017-05-20) is breaking a very
>>> significant number of test cases in both 32-bit and 64-bit Cygwin
>>> builds.
>>>
>>> The first failure is t0000.46 "validate object ID of a known tree"; output with
>>> -x and -v is below, although it's not very interesting:
>>>
>>>     expecting success:
>>>             test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>>>
>>>     ++ test ceb282701536fe61bea01075664405caa7d6343f = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>>>     + test_eval_ret_=1
>>>     + want_trace
>>>     + test t = t
>>>     + test t = t
>>>     + set +x
>>>     error: last command exited with $?=1
>>>     not ok 46 - validate object ID of a known tree
>>>     #
>>>     #               test "$tree" = 7bb943559a305bdd6bdee2cef6e5df2413c3d30a
>>>     #
>>>
>>> I have no idea where to even begin debugging this, but I'm happy to take
>>> pointers / try things out on my box.
>>
>> That looks scary, can you please comment out this:
>>
>>     #define SHA1DC_ALLOW_UNALIGNED_ACCESS
> 
> No, that doesn't fix it.
> 
>>
>> In sha1dc/sha1.c and see if that helps, alternatively comment out the
>> ifdefs guarded by "#ifdef _MSC_VER" calls in sha1dc/sha1.c
> 
> This can't possibly make a difference! ;-)
> 
> However, rebuilding with:
> 
>     $ make OPENSSL_SHA1=YesPlease >out2 2>&1
> 
> ... make t0000-basic.sh pass just fine, so ...

commit 7e71542e8b ("sha1dc: avoid CPP macro collisions", 25-03-2017) runs t0000-basic.sh just fine.

commit a0103914c2 ("sha1dc: update from upstream", 20-05-2017) fails when running t0000-basic.sh.

I will look into this more tomorrow (it is late, I need sleep), unless someone finds the solution overnight, of course.

Adam, you can build using the OPENSSL_SHA1 build variable for now (if you want to release v2.13.1), or wait for another maint release I suppose, ... I'll leave that to you! ;-)

ATB, Ramsay Jones

Junio C Hamano· Jun 6, 2017, 01:20 UTC · re: Ævar Arnfjörð Bjarmason · lore

Re: Git v2.13.1 SHA1 very broken

Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:
Show 6 quoted lines
> That looks scary, can you please comment out this:
>
>     #define SHA1DC_ALLOW_UNALIGNED_ACCESS
>
> In sha1dc/sha1.c and see if that helps, alternatively comment out the
> ifdefs guarded by "#ifdef _MSC_VER" calls in sha1dc/sha1.c

That is merely a performance (and theoretical correctness) thing, no?

> The functional differences between 2.13.0 and 2.13.1 on that platform
> should be none aside from possibly those changes, unless I've missed
> something.

If it does not hash correctly, the cause is more likely that the endianness detection is going haywire.

    make CFLAGS="-DSHA1DC_FORCE_LITTLEENDIAN -g -O2 -Wall"
or something like that, perhaps?
Adam Dinwoodie· Jun 6, 2017, 10:03 UTC · re: Junio C Hamano · lore

Re: Git v2.13.1 SHA1 very broken

On Tue, Jun 06, 2017 at 10:20:45AM +0900, Junio C Hamano wrote:
Show 11 quoted lines
> Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:
> 
> > That looks scary, can you please comment out this:
> >
> >     #define SHA1DC_ALLOW_UNALIGNED_ACCESS
> >
> > In sha1dc/sha1.c and see if that helps, alternatively comment out the
> > ifdefs guarded by "#ifdef _MSC_VER" calls in sha1dc/sha1.c
> 
> That is merely a performance (and theoretical correctness) thing,
> no?

Confirmed rebuilding with either of these suggested changes has t0000.46 still failing.

Show 8 quoted lines
> > The functional differences between 2.13.0 and 2.13.1 on that platform
> > should be none aside from possibly those changes, unless I've missed
> > something.
> 
> If it does not hash correctly, the cause is more likely that the
> endianness detection is going haywire.
> 
>     make CFLAGS="-DSHA1DC_FORCE_LITTLEENDIAN -g -O2 -Wall"

Confirmed rebuilding with this option has t0000 passing. I can also get the test to pass with Ramsay Jones' suggestion of using OpenSSL's SHA1.

Digging briefly into the endianness detection, it appears Cygwin has both _LITTLE_ENDIAN and _BIG_ENDIAN defined. Git's detection works by assuming it's in a little endian environment and switching to big endian if it detects any of the defines that indicate such, and a010391 adds _BIG_ENDIAN to the set of defines that indicate big endianness.

The obvious quick fix would be one of the two below patches. I'll also take the discussion to the Cygwin mailing list in the hope that someone can explain why Cygwin defines both _LITTLE_ENDIAN and _BIG_ENDIAN (and indeed _PDP_ENDIAN).

Patch 1 (probably safer?):
diff --git a/sha1dc/sha1.c b/sha1dc/sha1.c
index 3dff80ac7..e47d004b1 100644
--- a/sha1dc/sha1.c
+++ b/sha1dc/sha1.c
@@ -36,6 +36,7 @@
 #undef SHA1DC_BIGENDIAN
 #endif
 #if (!defined SHA1DC_FORCE_LITTLEENDIAN) && \
+    (!defined _LITTLE_ENDIAN) && \
     ((defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || \
     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __BIG_ENDIAN__)) || \
     defined(_BIG_ENDIAN) || defined(__BIG_ENDIAN__) || defined(__ARMEB__) || defined(__THUMBEB__) ||  defined(__AARCH64EB__) || \

Patch 2:

diff --git a/sha1dc/sha1.c b/sha1dc/sha1.c
index 3dff80ac7..8d7b1ee7d 100644
--- a/sha1dc/sha1.c
+++ b/sha1dc/sha1.c
@@ -38,7 +38,7 @@
 #if (!defined SHA1DC_FORCE_LITTLEENDIAN) && \
     ((defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || \
     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __BIG_ENDIAN__)) || \
-    defined(_BIG_ENDIAN) || defined(__BIG_ENDIAN__) || defined(__ARMEB__) || defined(__THUMBEB__) ||  defined(__AARCH64EB__) || \
+    defined(__BIG_ENDIAN__) || defined(__ARMEB__) || defined(__THUMBEB__) ||  defined(__AARCH64EB__) || \
     defined(_MIPSEB) || defined(__MIPSEB) || defined(__MIPSEB__) || defined(SHA1DC_FORCE_BIGENDIAN))

 #define SHA1DC_BIGENDIAN
Junio C Hamano· Jun 6, 2017, 11:55 UTC · re: Adam Dinwoodie · lore

Re: Git v2.13.1 SHA1 very broken

Adam Dinwoodie <adam@dinwoodie.org> writes:
Show 5 quoted lines
> Digging briefly into the endianness detection, it appears Cygwin has
> both _LITTLE_ENDIAN and _BIG_ENDIAN defined.  Git's detection works by
> assuming it's in a little endian environment and switching to big endian
> if it detects any of the defines that indicate such, and a010391 adds
> _BIG_ENDIAN to the set of defines that indicate big endianness.

I suspect that the upstream has already fixed this one to cope with FreeBSD. My preference is that we do another import on top of the ab/sha1dc-maint topic, below the commit on ab/sha1dc that adds the upstream as a submodule.

Adam Dinwoodie· Jun 6, 2017, 12:43 UTC · re: Junio C Hamano · lore

Re: Git v2.13.1 SHA1 very broken

On Tue, Jun 06, 2017 at 08:55:04PM +0900, Junio C Hamano wrote:
Show 12 quoted lines
> Adam Dinwoodie <adam@dinwoodie.org> writes:
> 
> > Digging briefly into the endianness detection, it appears Cygwin has
> > both _LITTLE_ENDIAN and _BIG_ENDIAN defined.  Git's detection works by
> > assuming it's in a little endian environment and switching to big endian
> > if it detects any of the defines that indicate such, and a010391 adds
> > _BIG_ENDIAN to the set of defines that indicate big endianness.
> 
> I suspect that the upstream has already fixed this one to cope with
> FreeBSD.  My preference is that we do another import on top of the
> ab/sha1dc-maint topic, below the commit on ab/sha1dc that adds the
> upstream as a submodule.

Apparently so! a010391 brings Git up to the upstream's cc46554 ("Skip temporary variable for SHA1DC_ALLOW_UNSIGNED_ACCESS", 2017-05-18); the problem has been fixed in upstream's a24eef5 ("rewrote Endianness selection", 2017-05-29).

In the interim, I'll use the CFLAGS route to try to get a v2.13.1 build ready to release for Cygwin.

Jason Pyeron· Jun 6, 2017, 14:47 UTC · re: Adam Dinwoodie · lore

Continous Integration (was: RE: Git v2.13.1 SHA1 very broken)

Do we have Jenkins (or something else) setup for Git?
We would be happy to donate (slave) VMs for cygwin builds og Git.  
-Jason Pyeron
Lars Schneider· Jun 6, 2017, 15:04 UTC · re: Jason Pyeron · lore

Re: Continous Integration (was: RE: Git v2.13.1 SHA1 very broken)

Show 8 quoted lines
> On 06 Jun 2017, at 16:47, Jason Pyeron <jpyeron@pdinc.us> wrote:
> 
> Do we have Jenkins (or something else) setup for Git?
> 
> We would be happy to donate (slave) VMs for cygwin builds og Git.  
> 
> -Jason Pyeron
>

We use TravisCI for Linux, Mac, and (in a special way) Windows: https://travis-ci.org/git/git

Windows is not supported by TravisCI. We just use TravisCI to trigger a build on MS Azure and read the results: https://github.com/git/git/commit/029aeeed55f6bfe8014e8ffe5fc7a6f2e5b110fc

Maybe we could trigger a Cgywin build on your slaves in the same way?
- Lars
Adam Dinwoodie· Jul 2, 2017, 20:35 UTC · re: Lars Schneider · lore

Re: Continous Integration (was: RE: Git v2.13.1 SHA1 very broken)

On Tue, Jun 06, 2017 at 05:04:32PM +0200, Lars Schneider wrote:
Show 17 quoted lines
> > On 06 Jun 2017, at 16:47, Jason Pyeron <jpyeron@pdinc.us> wrote:
> > 
> > Do we have Jenkins (or something else) setup for Git?
> > 
> > We would be happy to donate (slave) VMs for cygwin builds og Git.  
> > 
> > -Jason Pyeron
> > 
> 
> We use TravisCI for Linux, Mac, and (in a special way) Windows: 
> https://travis-ci.org/git/git
> 
> Windows is not supported by TravisCI. We just use TravisCI to
> trigger a build on MS Azure and read the results:
> https://github.com/git/git/commit/029aeeed55f6bfe8014e8ffe5fc7a6f2e5b110fc
> 
> Maybe we could trigger a Cgywin build on your slaves in the same way?

I tried setting up a Cygwin Git build on AppVeyor a little while ago, but the problem I found is that Cygwin builds are slow, particularly if you want to also run the test suite. I do the builds for the Cygwin distribution on my normal PC (so reasonably powerful but definitely not devoted to the purpose), and doing the build and running the default tests takes in the region of 8 hours for the 64-bit build and 12 hours for the 32-bit build.

At the moment, I'm trying to set up automated regular builds on my PC using BuildBot. I think that, short of someone donating some fairly significant resources for Cygwin builds, that's going to be the closest I'll be able to find for spotting problems early. It's a project in my currently limited spare time, though, and not something I've done before, so it's taking a little while to get going.

Adam
Johannes Schindelin· Jul 3, 2017, 12:34 UTC · re: Adam Dinwoodie · lore

Re: Continous Integration (was: RE: Git v2.13.1 SHA1 very broken)

Hi Adam,
On Sun, 2 Jul 2017, Adam Dinwoodie wrote:
> I do the builds for the Cygwin distribution on my normal PC (so
> reasonably powerful but definitely not devoted to the purpose), and
> doing the build and running the default tests takes in the region of 8
> hours for the 64-bit build and 12 hours for the 32-bit build.

Wow. 8 hours. I take it that you are bitten by the same issue as Git for Windows, where Git's test suite uses Unix shell scripting rather heavily, and Unix shell simply requiring tons of expensive emulation to run decently on Windows.

Show 6 quoted lines
> At the moment, I'm trying to set up automated regular builds on my PC
> using BuildBot.  I think that, short of someone donating some fairly
> significant resources for Cygwin builds, that's going to be the closest
> I'll be able to find for spotting problems early.  It's a project in my
> currently limited spare time, though, and not something I've done
> before, so it's taking a little while to get going.

How automated is your process? (Including, most importantly, updating the Cygwin build setup...)

I ask because I could possibly set up something using the existing infrastructure for Git for Windows' testing, as long as

1) everything is scripted, and
2) you do not need interactive access to any box (I mainly use Docker
   containers to run the builds)
If you are interested, feel free to ping me via private mail.

Ciao, Johannes

Ævar Arnfjörð Bjarmason· Jun 6, 2017, 15:12 UTC · re: Adam Dinwoodie · lore

[PATCH 0/3] update sha1dc

This updates sha1dc fixing the issue on Cygwin introduced in 2.13.1, and hopefully not regressing elsewhere. Liam, it would be much appreciated if you could test this on SPARC.

As before the "sha1dc: update from upstream" patch is what should fast-track to master/maint and be in 2.13.2, the other two are the cooking submodule use, that's all unchanged aside from of course the submodule pointing to the same upstream commit as the code import itself does.

Junio: There's a whitespace change to sha1.h that am warns about, but
which it applies anyway that you didn't apply from my previous
patch. I think it probably makes sense to just take upstream's
whitespace shenanigans as-is instead of seeing that diff every time we
update. I guess we could also send them a pull request...
Junio C Hamano (1):
  sha1collisiondetection: automatically enable when submodule is
    populated
Ævar Arnfjörð Bjarmason (2):
  sha1dc: update from upstream
  sha1dc: optionally use sha1collisiondetection as a submodule
 .gitmodules            |  4 ++++
 Makefile               | 16 ++++++++++++++++
 hash.h                 |  4 ++++
 sha1collisiondetection |  1 +
 sha1dc/sha1.c          | 30 ++++++++++++++++++++++++------
 sha1dc/sha1.h          |  6 +++---
 6 files changed, 52 insertions(+), 9 deletions(-)
 create mode 100644 .gitmodules
 create mode 160000 sha1collisiondetection
-- 
2.13.0.506.g27d5fe0cd
Ævar Arnfjörð Bjarmason· Jun 6, 2017, 15:12 UTC · re: Ævar Arnfjörð Bjarmason · lore

[PATCH 1/3] sha1dc: update from upstream

Update sha1dc from the latest version by the upstream maintainer[1].

See commit a0103914c2 ("sha1dc: update from upstream", 2017-05-20) for the latest update. That update was done sans some whitespace changes by upstream, which is why the diff here isn't the same as the upstream cc46554..e139984.

It also brings in a change[2] upstream made which should hopefully address the breakage in 2.13.1 on Cygwin, see [3]. Cygwin defines both _BIG_ENDIAN and _LITTLE_ENDIAN.

Adam Dinwoodie reports on the mailing list that that upstream commit fixes the issue on Cygwin[4].

1. https://github.com/cr-marcstevens/sha1collisiondetection/commit/e1399840b501a68ac6c8d7ed9a5cb1455480200e
2. https://github.com/cr-marcstevens/sha1collisiondetection/commit/a24eef58c0684078405f8c7a89f9b78271432005
3. <20170606100355.GC25777@dinwoodie.org> (https://public-inbox.org/git/20170606100355.GC25777@dinwoodie.org/)
4. <20170606124323.GD25777@dinwoodie.org> (https://public-inbox.org/git/20170606124323.GD25777@dinwoodie.org/)
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
---
 sha1dc/sha1.c | 30 ++++++++++++++++++++++++------
 sha1dc/sha1.h |  6 +++---
 2 files changed, 27 insertions(+), 9 deletions(-)
diff --git a/sha1dc/sha1.c b/sha1dc/sha1.c
index 3dff80ac72..facea1bb56 100644
--- a/sha1dc/sha1.c
+++ b/sha1dc/sha1.c
@@ -35,15 +35,33 @@
 #ifdef SHA1DC_BIGENDIAN
 #undef SHA1DC_BIGENDIAN
 #endif
-#if (!defined SHA1DC_FORCE_LITTLEENDIAN) && \
-    ((defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || \
-    (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __BIG_ENDIAN__)) || \
-    defined(_BIG_ENDIAN) || defined(__BIG_ENDIAN__) || defined(__ARMEB__) || defined(__THUMBEB__) ||  defined(__AARCH64EB__) || \
-    defined(_MIPSEB) || defined(__MIPSEB) || defined(__MIPSEB__) || defined(SHA1DC_FORCE_BIGENDIAN))
 
+#if (defined(_BYTE_ORDER) || defined(__BYTE_ORDER) || defined(__BYTE_ORDER__))
+
+#if ((defined(_BYTE_ORDER) && (_BYTE_ORDER == _BIG_ENDIAN)) || \
+     (defined(__BYTE_ORDER) && (__BYTE_ORDER == __BIG_ENDIAN)) || \
+     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __BIG_ENDIAN__)) )
 #define SHA1DC_BIGENDIAN
+#endif
+
+#else
+
+#if (defined(_BIG_ENDIAN) || defined(__BIG_ENDIAN) || defined(__BIG_ENDIAN__) || \
+     defined(__ARMEB__) || defined(__THUMBEB__) || defined(__AARCH64EB__) || \
+     defined(__MIPSEB__) || defined(__MIPSEB) || defined(_MIPSEB) || \
+     defined(__sparc))
+#define SHA1DC_BIGENDIAN
+#endif
 
-#endif /*ENDIANNESS SELECTION*/
+#endif
+
+#if (defined(SHA1DC_FORCE_LITTLEENDIAN) && defined(SHA1DC_BIGENDIAN))
+#undef SHA1DC_BIGENDIAN
+#endif
+#if (defined(SHA1DC_FORCE_BIGENDIAN) && !defined(SHA1DC_BIGENDIAN))
+#define SHA1DC_BIGENDIAN
+#endif
+/*ENDIANNESS SELECTION*/
 
 #if (defined SHA1DC_FORCE_UNALIGNED_ACCESS || \
      defined(__amd64__) || defined(__amd64) || defined(__x86_64__) || defined(__x86_64) || \
diff --git a/sha1dc/sha1.h b/sha1dc/sha1.h
index a0ff5d1305..1e4e94be54 100644
--- a/sha1dc/sha1.h
+++ b/sha1dc/sha1.h
@@ -61,9 +61,9 @@ void SHA1DCInit(SHA1_CTX*);
     Function to enable safe SHA-1 hashing:
     Collision attacks are thwarted by hashing a detected near-collision block 3 times.
     Think of it as extending SHA-1 from 80-steps to 240-steps for such blocks:
-	The best collision attacks against SHA-1 have complexity about 2^60,
-	thus for 240-steps an immediate lower-bound for the best cryptanalytic attacks would be 2^180.
-	An attacker would be better off using a generic birthday search of complexity 2^80.
+        The best collision attacks against SHA-1 have complexity about 2^60,
+        thus for 240-steps an immediate lower-bound for the best cryptanalytic attacks would be 2^180.
+        An attacker would be better off using a generic birthday search of complexity 2^80.
 
    Enabling safe SHA-1 hashing will result in the correct SHA-1 hash for messages where no collision attack was detected,
    but it will result in a different SHA-1 hash for messages where a collision attack was detected.

-- 
2.13.0.506.g27d5fe0cd
Ævar Arnfjörð Bjarmason· Jun 6, 2017, 15:12 UTC · re: Ævar Arnfjörð Bjarmason · lore

[PATCH 2/3] sha1dc: optionally use sha1collisiondetection as a submodule

Add an option to use the sha1collisiondetection library from the submodule in sha1collisiondetection/ instead of in the copy in the sha1dc/ directory.

This allows us to try out the submodule in sha1collisiondetection without breaking the build for anyone who's not expecting them as we work out any kinks.

Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
---
 .gitmodules            |  4 ++++
 Makefile               | 12 ++++++++++++
 hash.h                 |  4 ++++
 sha1collisiondetection |  1 +
 4 files changed, 21 insertions(+)
 create mode 100644 .gitmodules
 create mode 160000 sha1collisiondetection
diff --git a/.gitmodules b/.gitmodules
new file mode 100644
index 0000000000..cbeebdab7a
--- /dev/null
+++ b/.gitmodules
@@ -0,0 +1,4 @@
+[submodule "sha1collisiondetection"]
+	path = sha1collisiondetection
+	url = https://github.com/cr-marcstevens/sha1collisiondetection.git
+	branch = master
diff --git a/Makefile b/Makefile
index 7c621f7f76..adeff26d57 100644
--- a/Makefile
+++ b/Makefile
@@ -146,6 +146,12 @@ all::
 # algorithm. This is slower, but may detect attempted collision attacks.
 # Takes priority over other *_SHA1 knobs.
 #
+# Define DC_SHA1_SUBMODULE in addition to DC_SHA1 to use the
+# sha1collisiondetection shipped as a submodule instead of the
+# non-submodule copy in sha1dc/. This is an experimental option used
+# by the git project to migrate to using sha1collisiondetection as a
+# submodule.
+#
 # Define OPENSSL_SHA1 environment variable when running make to link
 # with the SHA1 routine from openssl library.
 #
@@ -1416,8 +1422,14 @@ ifdef APPLE_COMMON_CRYPTO
 	BASIC_CFLAGS += -DSHA1_APPLE
 else
 	DC_SHA1 := YesPlease
+ifdef DC_SHA1_SUBMODULE
+	LIB_OBJS += sha1collisiondetection/lib/sha1.o
+	LIB_OBJS += sha1collisiondetection/lib/ubc_check.o
+	BASIC_CFLAGS += -DDC_SHA1_SUBMODULE
+else
 	LIB_OBJS += sha1dc/sha1.o
 	LIB_OBJS += sha1dc/ubc_check.o
+endif
 	BASIC_CFLAGS += \
 		-DSHA1_DC \
 		-DSHA1DC_NO_STANDARD_INCLUDES \
diff --git a/hash.h b/hash.h
index a11fc9233f..bef3e630a0 100644
--- a/hash.h
+++ b/hash.h
@@ -8,7 +8,11 @@
 #elif defined(SHA1_OPENSSL)
 #include <openssl/sha.h>
 #elif defined(SHA1_DC)
+#ifdef DC_SHA1_SUBMODULE
+#include "sha1collisiondetection/lib/sha1.h"
+#else
 #include "sha1dc/sha1.h"
+#endif
 #else /* SHA1_BLK */
 #include "block-sha1/sha1.h"
 #endif
diff --git a/sha1collisiondetection b/sha1collisiondetection
new file mode 160000
index 0000000000..e1399840b5
--- /dev/null
+++ b/sha1collisiondetection
@@ -0,0 +1 @@
+Subproject commit e1399840b501a68ac6c8d7ed9a5cb1455480200e

-- 
2.13.0.506.g27d5fe0cd
Stefan Beller· Jun 6, 2017, 18:48 UTC · re: Ævar Arnfjörð Bjarmason · lore

Re: [PATCH 2/3] sha1dc: optionally use sha1collisiondetection as a submodule

On Tue, Jun 6, 2017 at 8:12 AM, Ævar Arnfjörð Bjarmason <avarab@gmail.com> wrote:

Show 9 quoted lines
> Add an option to use the sha1collisiondetection library from the
> submodule in sha1collisiondetection/ instead of in the copy in the
> sha1dc/ directory.
>
> This allows us to try out the submodule in sha1collisiondetection
> without breaking the build for anyone who's not expecting them as we
> work out any kinks.
>
> Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>

Other projects using submodules sometimes have a .gitattributes entry to have .gitmodules not exported via git-archive. Do we want a similar thing?

Speaking of attributes, I wonder if we want to specify
the .gitmodules file to be text with unixy file endings:
Having an entry
    .gitattributes eol=crlf
to simulate a Windows environment doesn't harm
submodule operation, which is good. I'll check if we
have a test for that.
Ævar Arnfjörð Bjarmason· Jun 6, 2017, 19:03 UTC · re: Stefan Beller · lore

Re: [PATCH 2/3] sha1dc: optionally use sha1collisiondetection as a submodule

On Tue, Jun 6, 2017 at 8:48 PM, Stefan Beller <sbeller@google.com> wrote:
Show 15 quoted lines
> On Tue, Jun 6, 2017 at 8:12 AM, Ævar Arnfjörð Bjarmason
> <avarab@gmail.com> wrote:
>> Add an option to use the sha1collisiondetection library from the
>> submodule in sha1collisiondetection/ instead of in the copy in the
>> sha1dc/ directory.
>>
>> This allows us to try out the submodule in sha1collisiondetection
>> without breaking the build for anyone who's not expecting them as we
>> work out any kinks.
>>
>> Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
>
> Other projects using submodules sometimes have
> a .gitattributes entry to have .gitmodules not exported
> via git-archive. Do we want a similar thing?

Right now we end up with an empty directory due to the issue you noted in https://public-inbox.org/git/CAGZ79kZC98CxA69QjmX2s_SU6z1CSgKgwZeqvwiMRAQc6+S3xg@mail.gmail.com/

It's probably best to have the .gitmodules file as some hint that something should be there. We also ship the other .git* files.

Show 7 quoted lines
> Speaking of attributes, I wonder if we want to specify
> the .gitmodules file to be text with unixy file endings:
> Having an entry
>     .gitattributes eol=crlf
> to simulate a Windows environment doesn't harm
> submodule operation, which is good. I'll check if we
> have a test for that.

I have no idea what that would do or why we'd have it, but I'm going to understand this as you looking into it :)

Stefan Beller· Jun 6, 2017, 19:09 UTC · re: Ævar Arnfjörð Bjarmason · lore

Re: [PATCH 2/3] sha1dc: optionally use sha1collisiondetection as a submodule

On Tue, Jun 6, 2017 at 12:03 PM, Ævar Arnfjörð Bjarmason <avarab@gmail.com> wrote:

Show 22 quoted lines
> On Tue, Jun 6, 2017 at 8:48 PM, Stefan Beller <sbeller@google.com> wrote:
>> On Tue, Jun 6, 2017 at 8:12 AM, Ævar Arnfjörð Bjarmason
>> <avarab@gmail.com> wrote:
>>> Add an option to use the sha1collisiondetection library from the
>>> submodule in sha1collisiondetection/ instead of in the copy in the
>>> sha1dc/ directory.
>>>
>>> This allows us to try out the submodule in sha1collisiondetection
>>> without breaking the build for anyone who's not expecting them as we
>>> work out any kinks.
>>>
>>> Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
>>
>> Other projects using submodules sometimes have
>> a .gitattributes entry to have .gitmodules not exported
>> via git-archive. Do we want a similar thing?
>
> Right now we end up with an empty directory due to the issue you noted
> in https://public-inbox.org/git/CAGZ79kZC98CxA69QjmX2s_SU6z1CSgKgwZeqvwiMRAQc6+S3xg@mail.gmail.com/
>
> It's probably best to have the .gitmodules file as some hint that
> something should be there. We also ship the other .git* files.

Ok, but then let's talk about the other .git* files, would we want to distribute these via tarballs? (I guess it is a minor thing if at all and nobody downloading a git tarball would be surprised by these metadata files or annoyed by them, so all is good?)

Show 11 quoted lines
>
>> Speaking of attributes, I wonder if we want to specify
>> the .gitmodules file to be text with unixy file endings:
>> Having an entry
>>     .gitattributes eol=crlf
>> to simulate a Windows environment doesn't harm
>> submodule operation, which is good. I'll check if we
>> have a test for that.
>
> I have no idea what that would do or why we'd have it, but I'm going
> to understand this as you looking into it :)

I looked briefly into it and it seems to be no problem just as config files on Windows are no problem. I just spoke up too quickly.

Ævar Arnfjörð Bjarmason· Jun 6, 2017, 15:12 UTC · re: Ævar Arnfjörð Bjarmason · lore

[PATCH 3/3] sha1collisiondetection: automatically enable when submodule is populated

From: Junio C Hamano <gitster@pobox.com>

If a user wants to experiment with the version of collision detecting sha1 from the submodule, the user needed to not just populate the submodule but also needed to turn the knob.

A Makefile trick is easy enough to do so, so let's do this. When somebody with a copy of the submodule populated wants not to use it, that can be done by overriding it in config.mak or from the command line.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
---
 Makefile | 4 ++++
 1 file changed, 4 insertions(+)
diff --git a/Makefile b/Makefile
index adeff26d57..eeccbff1cd 100644
--- a/Makefile
+++ b/Makefile
@@ -993,6 +993,10 @@ EXTLIBS =
 
 GIT_USER_AGENT = git/$(GIT_VERSION)
 
+ifeq ($(wildcard sha1collisiondetection/lib/sha1.h),sha1collisiondetection/lib/sha1.h)
+DC_SHA1_SUBMODULE = auto
+endif
+
 include config.mak.uname
 -include config.mak.autogen
 -include config.mak

-- 
2.13.0.506.g27d5fe0cd
Stefan Beller· Jun 6, 2017, 18:23 UTC · re: Ævar Arnfjörð Bjarmason · lore

Re: [PATCH 0/3] update sha1dc

On Tue, Jun 6, 2017 at 8:12 AM, Ævar Arnfjörð Bjarmason <avarab@gmail.com> wrote:

Show 15 quoted lines
> This updates sha1dc fixing the issue on Cygwin introduced in 2.13.1,
> and hopefully not regressing elsewhere. Liam, it would be much
> appreciated if you could test this on SPARC.
>
> As before the "sha1dc: update from upstream" patch is what should
> fast-track to master/maint and be in 2.13.2, the other two are the
> cooking submodule use, that's all unchanged aside from of course the
> submodule pointing to the same upstream commit as the code import
> itself does.
>
> Junio: There's a whitespace change to sha1.h that am warns about, but
> which it applies anyway that you didn't apply from my previous
> patch. I think it probably makes sense to just take upstream's
> whitespace shenanigans as-is instead of seeing that diff every time we
> update. I guess we could also send them a pull request...
I would suggest the pull request.

Also as to not make the mistake from before that I jump on the submodule bandwagon here: Patch 1 ought to go in its on series/patch, so with that out the way we have more time to consider the pros and cons of the rest of the series?

Thanks, Stefan

Ævar Arnfjörð Bjarmason· Jun 6, 2017, 18:51 UTC · re: Stefan Beller · lore

Re: [PATCH 0/3] update sha1dc

On Tue, Jun 6, 2017 at 8:23 PM, Stefan Beller <sbeller@google.com> wrote:
Show 19 quoted lines
> On Tue, Jun 6, 2017 at 8:12 AM, Ævar Arnfjörð Bjarmason
> <avarab@gmail.com> wrote:
>> This updates sha1dc fixing the issue on Cygwin introduced in 2.13.1,
>> and hopefully not regressing elsewhere. Liam, it would be much
>> appreciated if you could test this on SPARC.
>>
>> As before the "sha1dc: update from upstream" patch is what should
>> fast-track to master/maint and be in 2.13.2, the other two are the
>> cooking submodule use, that's all unchanged aside from of course the
>> submodule pointing to the same upstream commit as the code import
>> itself does.
>>
>> Junio: There's a whitespace change to sha1.h that am warns about, but
>> which it applies anyway that you didn't apply from my previous
>> patch. I think it probably makes sense to just take upstream's
>> whitespace shenanigans as-is instead of seeing that diff every time we
>> update. I guess we could also send them a pull request...
>
> I would suggest the pull request.

Looking at this again it's not a bug, just upstream choosing to indent a comment with spaces, not a bug.

So it makes sense to just apply as-is so we don't have that diff with them / different sha1s on the files etc.

Show 5 quoted lines
> Also as to not make the mistake from before that I jump on the
> submodule bandwagon here:
> Patch 1 ought to go in its on series/patch, so with that out the way
> we have more time to consider the pros and cons of the rest of
> the series?

Yes it makes perfect sense to just take the 1st patch here and make the submodule changes cook. This is just how I submitted it the last time and Junio took the 1st patch into a maint topic, so I figured I'd send it like this again.

Jeff King· Jun 6, 2017, 19:01 UTC · re: Ævar Arnfjörð Bjarmason · lore

[PATCH] sha1dc: ignore indent-with-non-tab whitespace violations

On Tue, Jun 06, 2017 at 08:51:35PM +0200, Ævar Arnfjörð Bjarmason wrote:
Show 26 quoted lines
> On Tue, Jun 6, 2017 at 8:23 PM, Stefan Beller <sbeller@google.com> wrote:
> > On Tue, Jun 6, 2017 at 8:12 AM, Ævar Arnfjörð Bjarmason
> > <avarab@gmail.com> wrote:
> >> This updates sha1dc fixing the issue on Cygwin introduced in 2.13.1,
> >> and hopefully not regressing elsewhere. Liam, it would be much
> >> appreciated if you could test this on SPARC.
> >>
> >> As before the "sha1dc: update from upstream" patch is what should
> >> fast-track to master/maint and be in 2.13.2, the other two are the
> >> cooking submodule use, that's all unchanged aside from of course the
> >> submodule pointing to the same upstream commit as the code import
> >> itself does.
> >>
> >> Junio: There's a whitespace change to sha1.h that am warns about, but
> >> which it applies anyway that you didn't apply from my previous
> >> patch. I think it probably makes sense to just take upstream's
> >> whitespace shenanigans as-is instead of seeing that diff every time we
> >> update. I guess we could also send them a pull request...
> >
> > I would suggest the pull request.
> 
> Looking at this again it's not a bug, just upstream choosing to indent
> a comment with spaces, not a bug.
> 
> So it makes sense to just apply as-is so we don't have that diff with
> them / different sha1s on the files etc.
Agreed. Maybe we'd also want this patch:
-- >8 --
Subject: sha1dc: ignore indent-with-non-tab whitespace violations

The upstream sha1dc code indents some lines with spaces. While this doesn't match Git's coding guidelines, it's better to leave this imported code untouched than to try to make it match our style. However, we can use .gitattributes to tell "diff --check" and "git am" not to bother us about it.

Signed-off-by: Jeff King <peff@peff.net>
---
 sha1dc/.gitattributes | 1 +
 1 file changed, 1 insertion(+)
 create mode 100644 sha1dc/.gitattributes
diff --git a/sha1dc/.gitattributes b/sha1dc/.gitattributes
new file mode 100644
index 000000000..da53f4054
--- /dev/null
+++ b/sha1dc/.gitattributes
@@ -0,0 +1 @@
+* whitespace=-indent-with-non-tab

-- 
2.13.1.664.g1b5a21ec3
Ævar Arnfjörð Bjarmason· Jun 6, 2017, 19:04 UTC · re: Jeff King · lore

Re: [PATCH] sha1dc: ignore indent-with-non-tab whitespace violations

On Tue, Jun 6, 2017 at 9:01 PM, Jeff King <peff@peff.net> wrote:
Show 30 quoted lines
> On Tue, Jun 06, 2017 at 08:51:35PM +0200, Ævar Arnfjörð Bjarmason wrote:
>
>> On Tue, Jun 6, 2017 at 8:23 PM, Stefan Beller <sbeller@google.com> wrote:
>> > On Tue, Jun 6, 2017 at 8:12 AM, Ævar Arnfjörð Bjarmason
>> > <avarab@gmail.com> wrote:
>> >> This updates sha1dc fixing the issue on Cygwin introduced in 2.13.1,
>> >> and hopefully not regressing elsewhere. Liam, it would be much
>> >> appreciated if you could test this on SPARC.
>> >>
>> >> As before the "sha1dc: update from upstream" patch is what should
>> >> fast-track to master/maint and be in 2.13.2, the other two are the
>> >> cooking submodule use, that's all unchanged aside from of course the
>> >> submodule pointing to the same upstream commit as the code import
>> >> itself does.
>> >>
>> >> Junio: There's a whitespace change to sha1.h that am warns about, but
>> >> which it applies anyway that you didn't apply from my previous
>> >> patch. I think it probably makes sense to just take upstream's
>> >> whitespace shenanigans as-is instead of seeing that diff every time we
>> >> update. I guess we could also send them a pull request...
>> >
>> > I would suggest the pull request.
>>
>> Looking at this again it's not a bug, just upstream choosing to indent
>> a comment with spaces, not a bug.
>>
>> So it makes sense to just apply as-is so we don't have that diff with
>> them / different sha1s on the files etc.
>
> Agreed. Maybe we'd also want this patch:
Great, that makes perfect sense for prepending to the series.
Show 25 quoted lines
> -- >8 --
> Subject: sha1dc: ignore indent-with-non-tab whitespace violations
>
> The upstream sha1dc code indents some lines with spaces.
> While this doesn't match Git's coding guidelines, it's better
> to leave this imported code untouched than to try to make it
> match our style. However, we can use .gitattributes to tell
> "diff --check" and "git am" not to bother us about it.
>
> Signed-off-by: Jeff King <peff@peff.net>
> ---
>  sha1dc/.gitattributes | 1 +
>  1 file changed, 1 insertion(+)
>  create mode 100644 sha1dc/.gitattributes
>
> diff --git a/sha1dc/.gitattributes b/sha1dc/.gitattributes
> new file mode 100644
> index 000000000..da53f4054
> --- /dev/null
> +++ b/sha1dc/.gitattributes
> @@ -0,0 +1 @@
> +* whitespace=-indent-with-non-tab
> --
> 2.13.1.664.g1b5a21ec3
>
Stefan Beller· Jun 6, 2017, 19:05 UTC · re: Jeff King · lore

Re: [PATCH] sha1dc: ignore indent-with-non-tab whitespace violations

Show 9 quoted lines
> Subject: sha1dc: ignore indent-with-non-tab whitespace violations
>
> The upstream sha1dc code indents some lines with spaces.
> While this doesn't match Git's coding guidelines, it's better
> to leave this imported code untouched than to try to make it
> match our style. However, we can use .gitattributes to tell
> "diff --check" and "git am" not to bother us about it.
>
> Signed-off-by: Jeff King <peff@peff.net>
Reviewed-by: Stefan Beller <sbeller@google.com>
Show 15 quoted lines
> ---
>  sha1dc/.gitattributes | 1 +
>  1 file changed, 1 insertion(+)
>  create mode 100644 sha1dc/.gitattributes
>
> diff --git a/sha1dc/.gitattributes b/sha1dc/.gitattributes
> new file mode 100644
> index 000000000..da53f4054
> --- /dev/null
> +++ b/sha1dc/.gitattributes
> @@ -0,0 +1 @@
> +* whitespace=-indent-with-non-tab
> --
> 2.13.1.664.g1b5a21ec3
>
Liam R. Howlett· Jun 13, 2017, 02:09 UTC · re: Ævar Arnfjörð Bjarmason · lore

Re: [PATCH 0/3] update sha1dc

* ?var Arnfj?r? Bjarmason <avarab@gmail.com> [170606 11:12]:
> This updates sha1dc fixing the issue on Cygwin introduced in 2.13.1,
> and hopefully not regressing elsewhere. Liam, it would be much
> appreciated if you could test this on SPARC.
Tested and confirmed this works for SPARC with no build flags.
Sorry for the delay, I was away last week.

Thanks, Liam

Morten Welinder· Jun 6, 2017, 12:49 UTC · re: Junio C Hamano · lore

Re: Git v2.13.1 SHA1 very broken

One could have configure ask some existing dependency that has already determined the byte order. For example:

# perl -e 'use Config; $o=$Config{byteorder}; print(($o=~/^1234/ ? "little" : ($o=~/4321$/ ? "big" : "weird")), "\n");' little

Good: less #ifdef soup; bad: not so great for cross-compiling.

(That's the integer byte order. The floating-point byte order can be different; hopefully git doesn't care.)

Morten
On Tue, Jun 6, 2017 at 7:55 AM, Junio C Hamano <gitster@pobox.com> wrote:
Show 13 quoted lines
> Adam Dinwoodie <adam@dinwoodie.org> writes:
>
>> Digging briefly into the endianness detection, it appears Cygwin has
>> both _LITTLE_ENDIAN and _BIG_ENDIAN defined.  Git's detection works by
>> assuming it's in a little endian environment and switching to big endian
>> if it detects any of the defines that indicate such, and a010391 adds
>> _BIG_ENDIAN to the set of defines that indicate big endianness.
>
> I suspect that the upstream has already fixed this one to cope with
> FreeBSD.  My preference is that we do another import on top of the
> ab/sha1dc-maint topic, below the commit on ab/sha1dc that adds the
> upstream as a submodule.
>

← back to recent threads