diff mbox series

git-apply: fix --3way with binary patch

Message ID 20210728024434.20230-1-jerry@skydio.com (mailing list archive)
State New
Headers show
Series git-apply: fix --3way with binary patch | expand

Commit Message

Jerry Zhang July 28, 2021, 2:44 a.m. UTC
Binary patches applied with "--3way" will
always return a conflict even if the patch
should cleanly apply because the low level
merge function considers all binary merges
without a variant to be conflicting.

Fix by falling back to normal patch application
for all binary patches.

Add tests for --3way and normal applications
of binary patches.

Fixes: 923cd87ac8 ("git-apply: try threeway first when "--3way" is used")
Signed-off-by: Jerry Zhang <jerry@skydio.com>
---
 apply.c                   |  3 ++-
 t/t4108-apply-threeway.sh | 45 +++++++++++++++++++++++++++++++++++++++
 2 files changed, 47 insertions(+), 1 deletion(-)

Comments

Junio C Hamano July 28, 2021, 4:29 a.m. UTC | #1
Jerry Zhang <jerry@skydio.com> writes:

> diff --git a/t/t4108-apply-threeway.sh b/t/t4108-apply-threeway.sh
> index 65147efdea..d32748f899 100755
> --- a/t/t4108-apply-threeway.sh
> +++ b/t/t4108-apply-threeway.sh
> @@ -230,4 +230,49 @@ test_expect_success 'apply with --3way --cached and conflicts' '
>  	test_cmp expect.diff actual.diff
>  '
>  
> +test_expect_success 'apply binary file patch' '
> +	git reset --hard main &&
> +	cp $TEST_DIRECTORY/test-binary-1.png bin.png &&

Is it safe to use $TEST_DIRECTORY without quoting?  I doubt it, as
it is $(pwd) of whereever the testing user extracted our source
tarball.  

In other words, you'd need this.

diff --git w/t/t4108-apply-threeway.sh c/t/t4108-apply-threeway.sh
index d32748f899..cc3aa3314a 100755
--- w/t/t4108-apply-threeway.sh
+++ c/t/t4108-apply-threeway.sh
@@ -232,11 +232,11 @@ test_expect_success 'apply with --3way --cached and conflicts' '
 
 test_expect_success 'apply binary file patch' '
 	git reset --hard main &&
-	cp $TEST_DIRECTORY/test-binary-1.png bin.png &&
+	cp "$TEST_DIRECTORY/test-binary-1.png" bin.png &&
 	git add bin.png &&
 	git commit -m "add binary file" &&
 
-	cp $TEST_DIRECTORY/test-binary-2.png bin.png &&
+	cp "$TEST_DIRECTORY/test-binary-2.png" bin.png &&
 
 	git diff --binary >bin.diff &&
 	git reset --hard &&
@@ -247,11 +247,11 @@ test_expect_success 'apply binary file patch' '
 
 test_expect_success 'apply binary file patch with 3way' '
 	git reset --hard main &&
-	cp $TEST_DIRECTORY/test-binary-1.png bin.png &&
+	cp "$TEST_DIRECTORY/test-binary-1.png" bin.png &&
 	git add bin.png &&
 	git commit -m "add binary file" &&
 
-	cp $TEST_DIRECTORY/test-binary-2.png bin.png &&
+	cp "$TEST_DIRECTORY/test-binary-2.png" bin.png &&
 
 	git diff --binary >bin.diff &&
 	git reset --hard &&
@@ -262,11 +262,11 @@ test_expect_success 'apply binary file patch with 3way' '
 
 test_expect_success 'apply full-index patch with 3way' '
 	git reset --hard main &&
-	cp $TEST_DIRECTORY/test-binary-1.png bin.png &&
+	cp "$TEST_DIRECTORY/test-binary-1.png" bin.png &&
 	git add bin.png &&
 	git commit -m "add binary file" &&
 
-	cp $TEST_DIRECTORY/test-binary-2.png bin.png &&
+	cp "$TEST_DIRECTORY/test-binary-2.png" bin.png &&
 
 	git diff --full-index >bin.diff &&
 	git reset --hard &&
Junio C Hamano July 28, 2021, 4:30 a.m. UTC | #2
Jerry Zhang <jerry@skydio.com> writes:

> Binary patches applied with "--3way" will
> always return a conflict even if the patch
> should cleanly apply because the low level
> merge function considers all binary merges
> without a variant to be conflicting.
>
> Fix by falling back to normal patch application
> for all binary patches.
>
> Add tests for --3way and normal applications
> of binary patches.
>
> Fixes: 923cd87ac8 ("git-apply: try threeway first when "--3way" is used")
> Signed-off-by: Jerry Zhang <jerry@skydio.com>
> ---
>  apply.c                   |  3 ++-
>  t/t4108-apply-threeway.sh | 45 +++++++++++++++++++++++++++++++++++++++
>  2 files changed, 47 insertions(+), 1 deletion(-)
>
> diff --git a/apply.c b/apply.c
> index 1d2d7e124e..78e52f0dc1 100644
> --- a/apply.c
> +++ b/apply.c
> @@ -3638,7 +3638,8 @@ static int apply_data(struct apply_state *state, struct patch *patch,
>  	if (load_preimage(state, &image, patch, st, ce) < 0)
>  		return -1;
>  
> -	if (!state->threeway || try_threeway(state, &image, patch, st, ce) < 0) {
> +	if (!state->threeway || patch->is_binary ||
> +		try_threeway(state, &image, patch, st, ce) < 0) {

Thanks for a quick turnaround.  However.

Because apply.c::three_way_merge() calls into ll_merge() that lets
the low-level custom merge drivers to take over the actual merge, I
do not think your "if binary, bypass and never call try_threway() at
all" is the right solution.  The custom merge driver user uses for
the path may successfully perform such a "trivial" three-way merge
and return success.

Why does the current code that lets threeway tried first fails to
fall back to direct application?  The code before your change, if
fed a binary patch that does not apply, would have failed the direct
application first *and* then fell back to the threeway (if only to
fail because we do not let binary files be merged), no?

Is it that try_threeway()'s way to express failure slightly
different from how direct application reports failure, but your
change used the same "only if it is negative, we fail and fallback"
logic?  IIRC, apply_fragments() which is the meat of the direct
application logic reports failures by negative, but try_threeway()
can return positive non-zero to signal a "recoverable" failure (aka
"conflicted merge").  Which should lead us to explore a different
approach, which is ...

    Would it be possible for a patch to leave conflicts when
    try_threeway() was attempted, but will cleanly apply if direct
    application is done?

If so, perhaps

 - we first run try_threeway() and see if it cleanly resolves; if
   so, we are done.

 - then we try direct application and see if it cleanly applies; if
   so, we are done.

 - finally we run try_threeway() again and let it fail with
   conflict.

might be the right sequence?  We theoretically could omit the first
of these three steps, but that would mean we'd write 923cd87a
(git-apply: try threeway first when "--3way" is used, 2021-04-06)
off as a failed experiment and revert it, which would not be ideal.


Also, independent from this "if we claim we try threeway first and
fall back to direct application, we really should do so" fix we are
discussing, I think our default binary merge can be a bit more
lenient and resolve this particular case of applying the binary
patch taken from itself (i.e. a patch that takes A to B gets applied
using --3way option to A).  I wonder if it can be as simple as the
attached patch.  FWIW, this change is sufficient (without the change
to apply.c we are reviewing here) to make your new tests in t4108
pass.

---- >8 ------- >8 ------- >8 ------- >8 ------- >8 ------- >8 ----
Subject: ll-merge: teach ll_binary_merge() a trivial three-way merge

The low-level binary merge code assumed that the caller will not
feed trivial merges that would have been resolved at the tree level;
because of this, ll_binary_merge() assumes the ancestor is different
from either side, always failing the merge in conflict unless -Xours
or -Xtheirs is in effect.

But "git apply --3way" codepath could ask us to perform three-way
merge between two binaries A and B using A as the ancestor version.
The current code always fails such an application, but when given a
binary patch that turns A into B and asked to apply it to A, there
is no reason to fail such a request---we can trivially tell that the
result must be B.

Arguably, this fix may belong to one level higher at ll_merge()
function, which dispatches to lower-level merge drivers, possibly
even before it renormalizes the three input buffers.  But let's
first see how this goes.

Signed-off-by: Junio C Hamano <gitster@pobox.com>
---
 ll-merge.c | 56 +++++++++++++++++++++++++++++++++++++++-----------------
 1 file changed, 39 insertions(+), 17 deletions(-)

diff --git c/ll-merge.c w/ll-merge.c
index 261657578c..bc8038d404 100644
--- c/ll-merge.c
+++ w/ll-merge.c
@@ -46,6 +46,13 @@ void reset_merge_attributes(void)
 	merge_attributes = NULL;
 }
 
+static int same_mmfile(mmfile_t *a, mmfile_t *b)
+{
+	if (a->size != b->size)
+		return 0;
+	return !memcmp(a->ptr, b->ptr, a->size);
+}
+
 /*
  * Built-in low-levels
  */
@@ -58,9 +65,18 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
 			   const struct ll_merge_options *opts,
 			   int marker_size)
 {
+	int status;
 	mmfile_t *stolen;
 	assert(opts);
 
+	/*
+	 * With -Xtheirs or -Xours, we have cleanly merged;
+	 * otherwise we got a conflict, unless 3way trivially
+	 * resolves.
+	 */
+	status = (opts->variant == XDL_MERGE_FAVOR_OURS ||
+		  opts->variant == XDL_MERGE_FAVOR_THEIRS) ? 0 : 1;
+
 	/*
 	 * The tentative merge result is the common ancestor for an
 	 * internal merge.  For the final merge, it is "ours" by
@@ -68,18 +84,30 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
 	 */
 	if (opts->virtual_ancestor) {
 		stolen = orig;
+		status = 0;
 	} else {
-		switch (opts->variant) {
-		default:
-			warning("Cannot merge binary files: %s (%s vs. %s)",
-				path, name1, name2);
-			/* fallthru */
-		case XDL_MERGE_FAVOR_OURS:
-			stolen = src1;
-			break;
-		case XDL_MERGE_FAVOR_THEIRS:
+		if (same_mmfile(orig, src1)) {
 			stolen = src2;
-			break;
+			status = 0;
+		} else if (same_mmfile(orig, src2)) { 
+			stolen = src1;
+			status = 0;
+		} else if (same_mmfile(src1, src2)) {
+			stolen = src1;
+			status = 0;
+		} else {
+			switch (opts->variant) {
+			default:
+				warning("Cannot merge binary files: %s (%s vs. %s)",
+					path, name1, name2);
+				/* fallthru */
+			case XDL_MERGE_FAVOR_OURS:
+				stolen = src1;
+				break;
+			case XDL_MERGE_FAVOR_THEIRS:
+				stolen = src2;
+				break;
+			}
 		}
 	}
 
@@ -87,13 +115,7 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
 	result->size = stolen->size;
 	stolen->ptr = NULL;
 
-	/*
-	 * With -Xtheirs or -Xours, we have cleanly merged;
-	 * otherwise we got a conflict.
-	 */
-	return opts->variant == XDL_MERGE_FAVOR_OURS ||
-	       opts->variant == XDL_MERGE_FAVOR_THEIRS ?
-	       0 : 1;
+	return status;
 }
 
 static int ll_xdl_merge(const struct ll_merge_driver *drv_unused,
Jerry Zhang July 28, 2021, 7:38 p.m. UTC | #3
On Tue, Jul 27, 2021 at 9:30 PM Junio C Hamano <gitster@pobox.com> wrote:
>
> Jerry Zhang <jerry@skydio.com> writes:
>
> > Binary patches applied with "--3way" will
> > always return a conflict even if the patch
> > should cleanly apply because the low level
> > merge function considers all binary merges
> > without a variant to be conflicting.
> >
> > Fix by falling back to normal patch application
> > for all binary patches.
> >
> > Add tests for --3way and normal applications
> > of binary patches.
> >
> > Fixes: 923cd87ac8 ("git-apply: try threeway first when "--3way" is used")
> > Signed-off-by: Jerry Zhang <jerry@skydio.com>
> > ---
> >  apply.c                   |  3 ++-
> >  t/t4108-apply-threeway.sh | 45 +++++++++++++++++++++++++++++++++++++++
> >  2 files changed, 47 insertions(+), 1 deletion(-)
> >
> > diff --git a/apply.c b/apply.c
> > index 1d2d7e124e..78e52f0dc1 100644
> > --- a/apply.c
> > +++ b/apply.c
> > @@ -3638,7 +3638,8 @@ static int apply_data(struct apply_state *state, struct patch *patch,
> >       if (load_preimage(state, &image, patch, st, ce) < 0)
> >               return -1;
> >
> > -     if (!state->threeway || try_threeway(state, &image, patch, st, ce) < 0) {
> > +     if (!state->threeway || patch->is_binary ||
> > +             try_threeway(state, &image, patch, st, ce) < 0) {
>
> Thanks for a quick turnaround.  However.
>
> Because apply.c::three_way_merge() calls into ll_merge() that lets
> the low-level custom merge drivers to take over the actual merge, I
> do not think your "if binary, bypass and never call try_threway() at
> all" is the right solution.  The custom merge driver user uses for
> the path may successfully perform such a "trivial" three-way merge
> and return success.
I understand now, thanks for the explanation
>
> Why does the current code that lets threeway tried first fails to
> fall back to direct application?  The code before your change, if
> fed a binary patch that does not apply, would have failed the direct
> application first *and* then fell back to the threeway (if only to
> fail because we do not let binary files be merged), no?
>
> Is it that try_threeway()'s way to express failure slightly
> different from how direct application reports failure, but your
> change used the same "only if it is negative, we fail and fallback"
> logic?  IIRC, apply_fragments() which is the meat of the direct
> application logic reports failures by negative, but try_threeway()
> can return positive non-zero to signal a "recoverable" failure (aka
> "conflicted merge").  Which should lead us to explore a different
> approach, which is ...
>
>     Would it be possible for a patch to leave conflicts when
>     try_threeway() was attempted, but will cleanly apply if direct
>     application is done?
>
> If so, perhaps
>
>  - we first run try_threeway() and see if it cleanly resolves; if
>    so, we are done.
>
>  - then we try direct application and see if it cleanly applies; if
>    so, we are done.
>
>  - finally we run try_threeway() again and let it fail with
>    conflict.
>
> might be the right sequence?  We theoretically could omit the first
> of these three steps, but that would mean we'd write 923cd87a
> (git-apply: try threeway first when "--3way" is used, 2021-04-06)
> off as a failed experiment and revert it, which would not be ideal.
>
>
> Also, independent from this "if we claim we try threeway first and
> fall back to direct application, we really should do so" fix we are
> discussing, I think our default binary merge can be a bit more
> lenient and resolve this particular case of applying the binary
> patch taken from itself (i.e. a patch that takes A to B gets applied
> using --3way option to A).  I wonder if it can be as simple as the
> attached patch.  FWIW, this change is sufficient (without the change
> to apply.c we are reviewing here) to make your new tests in t4108
> pass.
So basically, another way of stating the problem would be that binary
patches can apply cleanly with direct application in some cases where
merge application is not clean. If i understand correctly this is unique
to binary files, although it would be possible for a user to supply a custom
merge driver for text files that is worse than direct application, that is
most likely heavy user error that we shouldn't have to cater to. However
the issue with binary is that the *default* merge driver is actually worse
than direct application (in some cases). Therefore our options are

1. do as you suggest and run 3way -> direct -> 3way. I would modify
this and say we should only attempt this for binary patches, since a text
file that fails 3way would most likely also fail direct, so it would be a waste
of time to try it. furthermore if we cache results from the first 3way and
return them after attempting direct, it can save us from having to compute
the 3way twice, so would be no worse than our current performance.

2. improve the default binary merge driver to be at least as good as direct
application. this would allow us to say overall that "merge drivers should
be at least as intelligent as direct patch application" and would greatly
simplify logic in apply.c. Your change is a good first step in allowing it
to handle more cases. A trivial way to make the binary merge driver
at least as good as patch application is to generate a patch and apply
it as part of the merge. I imagine this would have other consequences
though as many parts of git use the binary merge driver.

Separately I think it would be a worthwhile follow-up patch to also handle
trivial three-way merges in try_threeway(). This would:
1. Allow us to compare oid instead of the entire file buffer, which would be
faster.
2. Handle trivial merges of all file types, which would save time.

>
> ---- >8 ------- >8 ------- >8 ------- >8 ------- >8 ------- >8 ----
> Subject: ll-merge: teach ll_binary_merge() a trivial three-way merge
>
> The low-level binary merge code assumed that the caller will not
> feed trivial merges that would have been resolved at the tree level;
> because of this, ll_binary_merge() assumes the ancestor is different
> from either side, always failing the merge in conflict unless -Xours
> or -Xtheirs is in effect.
>
> But "git apply --3way" codepath could ask us to perform three-way
> merge between two binaries A and B using A as the ancestor version.
> The current code always fails such an application, but when given a
> binary patch that turns A into B and asked to apply it to A, there
> is no reason to fail such a request---we can trivially tell that the
> result must be B.
>
> Arguably, this fix may belong to one level higher at ll_merge()
> function, which dispatches to lower-level merge drivers, possibly
> even before it renormalizes the three input buffers.  But let's
> first see how this goes.
>
> Signed-off-by: Junio C Hamano <gitster@pobox.com>
> ---
>  ll-merge.c | 56 +++++++++++++++++++++++++++++++++++++++-----------------
>  1 file changed, 39 insertions(+), 17 deletions(-)
>
> diff --git c/ll-merge.c w/ll-merge.c
> index 261657578c..bc8038d404 100644
> --- c/ll-merge.c
> +++ w/ll-merge.c
> @@ -46,6 +46,13 @@ void reset_merge_attributes(void)
>         merge_attributes = NULL;
>  }
>
> +static int same_mmfile(mmfile_t *a, mmfile_t *b)
> +{
> +       if (a->size != b->size)
> +               return 0;
> +       return !memcmp(a->ptr, b->ptr, a->size);
> +}
> +
>  /*
>   * Built-in low-levels
>   */
> @@ -58,9 +65,18 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
>                            const struct ll_merge_options *opts,
>                            int marker_size)
>  {
> +       int status;
>         mmfile_t *stolen;
>         assert(opts);
>
> +       /*
> +        * With -Xtheirs or -Xours, we have cleanly merged;
> +        * otherwise we got a conflict, unless 3way trivially
> +        * resolves.
> +        */
> +       status = (opts->variant == XDL_MERGE_FAVOR_OURS ||
> +                 opts->variant == XDL_MERGE_FAVOR_THEIRS) ? 0 : 1;
> +
>         /*
>          * The tentative merge result is the common ancestor for an
>          * internal merge.  For the final merge, it is "ours" by
> @@ -68,18 +84,30 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
>          */
>         if (opts->virtual_ancestor) {
>                 stolen = orig;
> +               status = 0;
>         } else {
> -               switch (opts->variant) {
> -               default:
> -                       warning("Cannot merge binary files: %s (%s vs. %s)",
> -                               path, name1, name2);
> -                       /* fallthru */
> -               case XDL_MERGE_FAVOR_OURS:
> -                       stolen = src1;
> -                       break;
> -               case XDL_MERGE_FAVOR_THEIRS:
> +               if (same_mmfile(orig, src1)) {
>                         stolen = src2;
> -                       break;
> +                       status = 0;
> +               } else if (same_mmfile(orig, src2)) {
> +                       stolen = src1;
> +                       status = 0;
> +               } else if (same_mmfile(src1, src2)) {
> +                       stolen = src1;
> +                       status = 0;
> +               } else {
> +                       switch (opts->variant) {
> +                       default:
> +                               warning("Cannot merge binary files: %s (%s vs. %s)",
> +                                       path, name1, name2);
> +                               /* fallthru */
> +                       case XDL_MERGE_FAVOR_OURS:
> +                               stolen = src1;
> +                               break;
> +                       case XDL_MERGE_FAVOR_THEIRS:
> +                               stolen = src2;
> +                               break;
> +                       }
>                 }
>         }
>
> @@ -87,13 +115,7 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
>         result->size = stolen->size;
>         stolen->ptr = NULL;
>
> -       /*
> -        * With -Xtheirs or -Xours, we have cleanly merged;
> -        * otherwise we got a conflict.
> -        */
> -       return opts->variant == XDL_MERGE_FAVOR_OURS ||
> -              opts->variant == XDL_MERGE_FAVOR_THEIRS ?
> -              0 : 1;
> +       return status;
>  }
>
>  static int ll_xdl_merge(const struct ll_merge_driver *drv_unused,
Jerry Zhang July 28, 2021, 8:04 p.m. UTC | #4
On Wed, Jul 28, 2021 at 12:38 PM Jerry Zhang <jerry@skydio.com> wrote:
>
> On Tue, Jul 27, 2021 at 9:30 PM Junio C Hamano <gitster@pobox.com> wrote:
> >
> > Jerry Zhang <jerry@skydio.com> writes:
> >
> > > Binary patches applied with "--3way" will
> > > always return a conflict even if the patch
> > > should cleanly apply because the low level
> > > merge function considers all binary merges
> > > without a variant to be conflicting.
> > >
> > > Fix by falling back to normal patch application
> > > for all binary patches.
> > >
> > > Add tests for --3way and normal applications
> > > of binary patches.
> > >
> > > Fixes: 923cd87ac8 ("git-apply: try threeway first when "--3way" is used")
> > > Signed-off-by: Jerry Zhang <jerry@skydio.com>
> > > ---
> > >  apply.c                   |  3 ++-
> > >  t/t4108-apply-threeway.sh | 45 +++++++++++++++++++++++++++++++++++++++
> > >  2 files changed, 47 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/apply.c b/apply.c
> > > index 1d2d7e124e..78e52f0dc1 100644
> > > --- a/apply.c
> > > +++ b/apply.c
> > > @@ -3638,7 +3638,8 @@ static int apply_data(struct apply_state *state, struct patch *patch,
> > >       if (load_preimage(state, &image, patch, st, ce) < 0)
> > >               return -1;
> > >
> > > -     if (!state->threeway || try_threeway(state, &image, patch, st, ce) < 0) {
> > > +     if (!state->threeway || patch->is_binary ||
> > > +             try_threeway(state, &image, patch, st, ce) < 0) {
> >
> > Thanks for a quick turnaround.  However.
> >
> > Because apply.c::three_way_merge() calls into ll_merge() that lets
> > the low-level custom merge drivers to take over the actual merge, I
> > do not think your "if binary, bypass and never call try_threway() at
> > all" is the right solution.  The custom merge driver user uses for
> > the path may successfully perform such a "trivial" three-way merge
> > and return success.
> I understand now, thanks for the explanation
> >
> > Why does the current code that lets threeway tried first fails to
> > fall back to direct application?  The code before your change, if
> > fed a binary patch that does not apply, would have failed the direct
> > application first *and* then fell back to the threeway (if only to
> > fail because we do not let binary files be merged), no?
> >
> > Is it that try_threeway()'s way to express failure slightly
> > different from how direct application reports failure, but your
> > change used the same "only if it is negative, we fail and fallback"
> > logic?  IIRC, apply_fragments() which is the meat of the direct
> > application logic reports failures by negative, but try_threeway()
> > can return positive non-zero to signal a "recoverable" failure (aka
> > "conflicted merge").  Which should lead us to explore a different
> > approach, which is ...
> >
> >     Would it be possible for a patch to leave conflicts when
> >     try_threeway() was attempted, but will cleanly apply if direct
> >     application is done?
> >
> > If so, perhaps
> >
> >  - we first run try_threeway() and see if it cleanly resolves; if
> >    so, we are done.
> >
> >  - then we try direct application and see if it cleanly applies; if
> >    so, we are done.
> >
> >  - finally we run try_threeway() again and let it fail with
> >    conflict.
> >
> > might be the right sequence?  We theoretically could omit the first
> > of these three steps, but that would mean we'd write 923cd87a
> > (git-apply: try threeway first when "--3way" is used, 2021-04-06)
> > off as a failed experiment and revert it, which would not be ideal.
> >
> >
> > Also, independent from this "if we claim we try threeway first and
> > fall back to direct application, we really should do so" fix we are
> > discussing, I think our default binary merge can be a bit more
> > lenient and resolve this particular case of applying the binary
> > patch taken from itself (i.e. a patch that takes A to B gets applied
> > using --3way option to A).  I wonder if it can be as simple as the
> > attached patch.  FWIW, this change is sufficient (without the change
> > to apply.c we are reviewing here) to make your new tests in t4108
> > pass.
> So basically, another way of stating the problem would be that binary
> patches can apply cleanly with direct application in some cases where
> merge application is not clean. If i understand correctly this is unique
> to binary files, although it would be possible for a user to supply a custom
> merge driver for text files that is worse than direct application, that is
> most likely heavy user error that we shouldn't have to cater to. However
> the issue with binary is that the *default* merge driver is actually worse
> than direct application (in some cases). Therefore our options are
>
> 1. do as you suggest and run 3way -> direct -> 3way. I would modify
> this and say we should only attempt this for binary patches, since a text
> file that fails 3way would most likely also fail direct, so it would be a waste
> of time to try it. furthermore if we cache results from the first 3way and
> return them after attempting direct, it can save us from having to compute
> the 3way twice, so would be no worse than our current performance.
>
> 2. improve the default binary merge driver to be at least as good as direct
> application. this would allow us to say overall that "merge drivers should
> be at least as intelligent as direct patch application" and would greatly
> simplify logic in apply.c. Your change is a good first step in allowing it
> to handle more cases. A trivial way to make the binary merge driver
> at least as good as patch application is to generate a patch and apply
> it as part of the merge. I imagine this would have other consequences
> though as many parts of git use the binary merge driver.

Hmm I would have thought that binary patches allow context, similar to this
test snippet
"
 test_expect_success 'apply complex binary file patch' '
     git reset --hard main &&
     cp $TEST_DIRECTORY/test-binary-1.png bin.png &&
     git add bin.png &&
     git commit -m "add binary file" &&

     echo 1 >>bin.png &&
     git diff --binary >bin.diff &&
     git reset --hard &&

     cat $TEST_DIRECTORY/test-binary-2.png
$TEST_DIRECTORY/test-binary-1.png >bin.png &&
     git add bin.png &&
     git commit -m "change binary file" &&

     # Apply must succeed.
     git apply bin.diff
 '
"
but upon running it I see that normal patch application still requires the
preimage to match exactly.
"
error: the patch applies to 'bin.png'
(836481bd1b9b6bd7a1bb8939cf4ea01e05946850), which does not match the
current contents.
error: bin.png: patch does not apply
"

So at least in regards to making the default binary merge driver "at
least as intelligent" as
direct patch application, your patch ought to do it.

>
> Separately I think it would be a worthwhile follow-up patch to also handle
> trivial three-way merges in try_threeway(). This would:
> 1. Allow us to compare oid instead of the entire file buffer, which would be
> faster.
> 2. Handle trivial merges of all file types, which would save time.
>
> >
> > ---- >8 ------- >8 ------- >8 ------- >8 ------- >8 ------- >8 ----
> > Subject: ll-merge: teach ll_binary_merge() a trivial three-way merge
> >
> > The low-level binary merge code assumed that the caller will not
> > feed trivial merges that would have been resolved at the tree level;
> > because of this, ll_binary_merge() assumes the ancestor is different
> > from either side, always failing the merge in conflict unless -Xours
> > or -Xtheirs is in effect.
> >
> > But "git apply --3way" codepath could ask us to perform three-way
> > merge between two binaries A and B using A as the ancestor version.
> > The current code always fails such an application, but when given a
> > binary patch that turns A into B and asked to apply it to A, there
> > is no reason to fail such a request---we can trivially tell that the
> > result must be B.
> >
> > Arguably, this fix may belong to one level higher at ll_merge()
> > function, which dispatches to lower-level merge drivers, possibly
> > even before it renormalizes the three input buffers.  But let's
> > first see how this goes.
> >
> > Signed-off-by: Junio C Hamano <gitster@pobox.com>
> > ---
> >  ll-merge.c | 56 +++++++++++++++++++++++++++++++++++++++-----------------
> >  1 file changed, 39 insertions(+), 17 deletions(-)
> >
> > diff --git c/ll-merge.c w/ll-merge.c
> > index 261657578c..bc8038d404 100644
> > --- c/ll-merge.c
> > +++ w/ll-merge.c
> > @@ -46,6 +46,13 @@ void reset_merge_attributes(void)
> >         merge_attributes = NULL;
> >  }
> >
> > +static int same_mmfile(mmfile_t *a, mmfile_t *b)
> > +{
> > +       if (a->size != b->size)
> > +               return 0;
> > +       return !memcmp(a->ptr, b->ptr, a->size);
> > +}
> > +
> >  /*
> >   * Built-in low-levels
> >   */
> > @@ -58,9 +65,18 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
> >                            const struct ll_merge_options *opts,
> >                            int marker_size)
> >  {
> > +       int status;
> >         mmfile_t *stolen;
> >         assert(opts);
> >
> > +       /*
> > +        * With -Xtheirs or -Xours, we have cleanly merged;
> > +        * otherwise we got a conflict, unless 3way trivially
> > +        * resolves.
> > +        */
> > +       status = (opts->variant == XDL_MERGE_FAVOR_OURS ||
> > +                 opts->variant == XDL_MERGE_FAVOR_THEIRS) ? 0 : 1;
> > +
> >         /*
> >          * The tentative merge result is the common ancestor for an
> >          * internal merge.  For the final merge, it is "ours" by
> > @@ -68,18 +84,30 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
> >          */
> >         if (opts->virtual_ancestor) {
> >                 stolen = orig;
> > +               status = 0;
> >         } else {
> > -               switch (opts->variant) {
> > -               default:
> > -                       warning("Cannot merge binary files: %s (%s vs. %s)",
> > -                               path, name1, name2);
> > -                       /* fallthru */
> > -               case XDL_MERGE_FAVOR_OURS:
> > -                       stolen = src1;
> > -                       break;
> > -               case XDL_MERGE_FAVOR_THEIRS:
> > +               if (same_mmfile(orig, src1)) {
> >                         stolen = src2;
> > -                       break;
> > +                       status = 0;
> > +               } else if (same_mmfile(orig, src2)) {
> > +                       stolen = src1;
> > +                       status = 0;
> > +               } else if (same_mmfile(src1, src2)) {
> > +                       stolen = src1;
> > +                       status = 0;
> > +               } else {
> > +                       switch (opts->variant) {
> > +                       default:
> > +                               warning("Cannot merge binary files: %s (%s vs. %s)",
> > +                                       path, name1, name2);
> > +                               /* fallthru */
> > +                       case XDL_MERGE_FAVOR_OURS:
> > +                               stolen = src1;
> > +                               break;
> > +                       case XDL_MERGE_FAVOR_THEIRS:
> > +                               stolen = src2;
> > +                               break;
> > +                       }
> >                 }
> >         }
> >
> > @@ -87,13 +115,7 @@ static int ll_binary_merge(const struct ll_merge_driver *drv_unused,
> >         result->size = stolen->size;
> >         stolen->ptr = NULL;
> >
> > -       /*
> > -        * With -Xtheirs or -Xours, we have cleanly merged;
> > -        * otherwise we got a conflict.
> > -        */
> > -       return opts->variant == XDL_MERGE_FAVOR_OURS ||
> > -              opts->variant == XDL_MERGE_FAVOR_THEIRS ?
> > -              0 : 1;
> > +       return status;
> >  }
> >
> >  static int ll_xdl_merge(const struct ll_merge_driver *drv_unused,
Junio C Hamano July 28, 2021, 8:08 p.m. UTC | #5
Jerry Zhang <jerry@skydio.com> writes:

> So basically, another way of stating the problem would be that binary
> patches can apply cleanly with direct application in some cases where
> merge application is not clean. If i understand correctly this is unique
> to binary files, although it would be possible for a user to supply a custom
> merge driver for text files that is worse than direct application, that is
> most likely heavy user error that we shouldn't have to cater to.

Not really.  The built-in binary merge driver luckily had such
characteristics to allow us to catch this regression, but I see no
reason to believe that it is unique to binary.  Funky merge backends
like union merges can turn an otherwise conflicting merge into a
clean merge even for non-binary files.  And no, it is not an error
for a merge driver to fail "apply --3way" merge on incoming data
that "apply --no-3way" would apply cleanly.

> However
> the issue with binary is that the *default* merge driver is actually worse
> than direct application (in some cases).

> 1. do as you suggest and run 3way -> direct -> 3way. I would modify
> this and say we should only attempt this for binary patches, since a text
> file that fails 3way would most likely also fail direct,...

No, I do not trust our (myself and your) unsubstantiated belief that
it is limited to binary.  We saw a problem with binary, and I would
think it is a tip of iceberg for any non-straight-text-merge backend
(and I do not have any sound reason to believe that straight
text-merge backend will not have this issue).  I'd rather treat this
as coalmine canary.

I think the real problem, even without the "try threeway, fall back
to direct application, and then try threeway again", is that after
swapping the fallback order, a failed threeway does *not* fall back
to direct application in this case.  Regardless of what ll_merge()
and its backend does, if they fail, shouldn't the caller of
try_threeway() notice the failure and fall back to direct
application, just like the earlier code tried direct application
first and then _always_ fell back to threeway if it failed?  I do
not know exactly why today's code fails to do so, but I suspect that
fixing that is the real solution, no?

Independent from that, I suspect that it may be a good thing to do
to (at least optionally) allow ll_merge() to notice trivial merges
that proper merge frontends would never ask it to do and resolve
them trivially.  The patch you saw from me to ll_merge_binary() may
do so at a wrong layer (doing it in ll_merge() before it dispatches
to ll_merge_binary() and other backends might be a better approach)
but would be a good starting point for that independent effort, but
"apply --3way" should work correctly even with user-configured merge
drivers (after all, the "direct application first and then fall back
to 3way" code would have worked perfectly fine even with broken
custom merge drivers in the case we are discussing right now).

Thanks.
Jerry Zhang July 28, 2021, 8:37 p.m. UTC | #6
On Wed, Jul 28, 2021 at 1:08 PM Junio C Hamano <gitster@pobox.com> wrote:
>
> Jerry Zhang <jerry@skydio.com> writes:
>
> > So basically, another way of stating the problem would be that binary
> > patches can apply cleanly with direct application in some cases where
> > merge application is not clean. If i understand correctly this is unique
> > to binary files, although it would be possible for a user to supply a custom
> > merge driver for text files that is worse than direct application, that is
> > most likely heavy user error that we shouldn't have to cater to.
>
> Not really.  The built-in binary merge driver luckily had such
> characteristics to allow us to catch this regression, but I see no
> reason to believe that it is unique to binary.  Funky merge backends
> like union merges can turn an otherwise conflicting merge into a
> clean merge even for non-binary files.  And no, it is not an error
> for a merge driver to fail "apply --3way" merge on incoming data
> that "apply --no-3way" would apply cleanly.
>
> > However
> > the issue with binary is that the *default* merge driver is actually worse
> > than direct application (in some cases).
>
> > 1. do as you suggest and run 3way -> direct -> 3way. I would modify
> > this and say we should only attempt this for binary patches, since a text
> > file that fails 3way would most likely also fail direct,...
>
> No, I do not trust our (myself and your) unsubstantiated belief that
> it is limited to binary.  We saw a problem with binary, and I would
> think it is a tip of iceberg for any non-straight-text-merge backend
> (and I do not have any sound reason to believe that straight
> text-merge backend will not have this issue).  I'd rather treat this
> as coalmine canary.

I see, this could be true. But then we would also have the case where
a merge driver results in conflict and direct patch application applies cleanly,
*but* the direct patch application is actually incorrect (for reasons relating
to the original purpose of the patch to switch the order, that direct patch
application can be wrong for files of repeating content). In this case
the user might actually want to see the conflict as it is more correct, but
it is being hidden by our fallback.

If the backwards compatibility story is going to get messy like this, perhaps
the best solution is to make a new flag similar to "--actually-3way" that
will attempt 3way and nothing else, and users who know what they want
can use that to get what they want.
>
> I think the real problem, even without the "try threeway, fall back
> to direct application, and then try threeway again", is that after
> swapping the fallback order, a failed threeway does *not* fall back
> to direct application in this case.  Regardless of what ll_merge()
> and its backend does, if they fail, shouldn't the caller of
> try_threeway() notice the failure and fall back to direct
> application, just like the earlier code tried direct application
> first and then _always_ fell back to threeway if it failed?  I do
> not know exactly why today's code fails to do so, but I suspect that
> fixing that is the real solution, no?

Well it isn't really failing right? Failing 3way would be not finding
the object ids in the database, which would indicate a failure to
even attempt 3way. This would result in fallback to direct application.
What we're seeing is that 3way application results in conflicts where
direct application would not result in conflicts. Having a conflict is
currently not a reason for the code to fall back to direct application,
here is the relevant line:
"
         try_threeway(state, &image, patch, st, ce) < 0) {
"
try_threeway returns 1 in case of conflict, 0 for success, and -1
for true errors.

>
> Independent from that, I suspect that it may be a good thing to do
> to (at least optionally) allow ll_merge() to notice trivial merges
> that proper merge frontends would never ask it to do and resolve
> them trivially.  The patch you saw from me to ll_merge_binary() may
> do so at a wrong layer (doing it in ll_merge() before it dispatches
> to ll_merge_binary() and other backends might be a better approach)
> but would be a good starting point for that independent effort, but
> "apply --3way" should work correctly even with user-configured merge
> drivers (after all, the "direct application first and then fall back
> to 3way" code would have worked perfectly fine even with broken
> custom merge drivers in the case we are discussing right now).
>
> Thanks.
>
>
Junio C Hamano July 28, 2021, 9:01 p.m. UTC | #7
Jerry Zhang <jerry@skydio.com> writes:

> Well it isn't really failing right? Failing 3way would be not finding
> the object ids in the database, which would indicate a failure to
> even attempt 3way. This would result in fallback to direct application.
> What we're seeing is that 3way application results in conflicts where
> direct application would not result in conflicts. Having a conflict is
> currently not a reason for the code to fall back to direct application,
> here is the relevant line:
> "
>          try_threeway(state, &image, patch, st, ce) < 0) {
> "
> try_threeway returns 1 in case of conflict, 0 for success, and -1
> for true errors.

Yup, I know.  That is why I questioned if this "< 0" is a bug in my
earlier message in this exchange.

Thanks.
diff mbox series

Patch

diff --git a/apply.c b/apply.c
index 1d2d7e124e..78e52f0dc1 100644
--- a/apply.c
+++ b/apply.c
@@ -3638,7 +3638,8 @@  static int apply_data(struct apply_state *state, struct patch *patch,
 	if (load_preimage(state, &image, patch, st, ce) < 0)
 		return -1;
 
-	if (!state->threeway || try_threeway(state, &image, patch, st, ce) < 0) {
+	if (!state->threeway || patch->is_binary ||
+		try_threeway(state, &image, patch, st, ce) < 0) {
 		if (state->apply_verbosity > verbosity_silent &&
 		    state->threeway && !patch->direct_to_threeway)
 			fprintf(stderr, _("Falling back to direct application...\n"));
diff --git a/t/t4108-apply-threeway.sh b/t/t4108-apply-threeway.sh
index 65147efdea..d32748f899 100755
--- a/t/t4108-apply-threeway.sh
+++ b/t/t4108-apply-threeway.sh
@@ -230,4 +230,49 @@  test_expect_success 'apply with --3way --cached and conflicts' '
 	test_cmp expect.diff actual.diff
 '
 
+test_expect_success 'apply binary file patch' '
+	git reset --hard main &&
+	cp $TEST_DIRECTORY/test-binary-1.png bin.png &&
+	git add bin.png &&
+	git commit -m "add binary file" &&
+
+	cp $TEST_DIRECTORY/test-binary-2.png bin.png &&
+
+	git diff --binary >bin.diff &&
+	git reset --hard &&
+
+	# Apply must succeed.
+	git apply bin.diff
+'
+
+test_expect_success 'apply binary file patch with 3way' '
+	git reset --hard main &&
+	cp $TEST_DIRECTORY/test-binary-1.png bin.png &&
+	git add bin.png &&
+	git commit -m "add binary file" &&
+
+	cp $TEST_DIRECTORY/test-binary-2.png bin.png &&
+
+	git diff --binary >bin.diff &&
+	git reset --hard &&
+
+	# Apply must succeed.
+	git apply --3way --index bin.diff
+'
+
+test_expect_success 'apply full-index patch with 3way' '
+	git reset --hard main &&
+	cp $TEST_DIRECTORY/test-binary-1.png bin.png &&
+	git add bin.png &&
+	git commit -m "add binary file" &&
+
+	cp $TEST_DIRECTORY/test-binary-2.png bin.png &&
+
+	git diff --full-index >bin.diff &&
+	git reset --hard &&
+
+	# Apply must succeed.
+	git apply --3way --index bin.diff
+'
+
 test_done