[PATCH v8 00/41] libify apply and use lib in am, part 2

Goal ~~~~

This is a patch series about libifying `git apply` functionality, and using this libified functionality in `git am`, so that no 'git apply' process is spawn anymore. This makes `git am` significantly faster, so `git rebase`, when it uses the am backend, is also significantly faster.

Previous discussions and patches series ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This has initially been discussed in the following thread:

  http://thread.gmane.org/gmane.comp.version-control.git/287236/

Then the following patch series were sent:

RFC: http://thread.gmane.org/gmane.comp.version-control.git/288489/
v1: http://thread.gmane.org/gmane.comp.version-control.git/292324/
v2: http://thread.gmane.org/gmane.comp.version-control.git/294248/
v3: http://thread.gmane.org/gmane.comp.version-control.git/295429/
v4: http://thread.gmane.org/gmane.comp.version-control.git/296350/
v5: http://thread.gmane.org/gmane.comp.version-control.git/296490/
v6: http://thread.gmane.org/gmane.comp.version-control.git/297024/
v7: http://thread.gmane.org/gmane.comp.version-control.git/297193/

Highlevel view of the patches in the series ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This new patch series is built on top of the above previous work.

More precisely, this is "part 2" of the full patch series which is built on top of the "part 1" of the full patch series. And as the "part 1" is now in "next", this "part 2" is built on top of "next".

  - Patches 01/41 is new in v8.

It renames some structs and constants that will be moved into apply.h to give them a more specific name as suggested by Junio.

  - Patches 02/41 to 32/41 were in v1, v2, v6 and v7.

They finish libifying the apply functionality that was in builtin/apply.c and move it into apply.{c,h}, but the libified functionality is not yet used in `git am`.

There are a number of changes in these patches compared to v7, so that apply_all_patches() will return 1 in case some patches don't apply and 128 in case a previously fatal error happened. This makes it possible for `git apply` to return the same exit code as it used to return before libification.

To do that a number of functions like find_header(), parse_chunk(), apply_patch(), etc have been made to return -128 instead of -1 when it was necessary.

  - Patches 33/41 to 39/41 were in v2, v6 and v7.

They implement a way to make the libified apply silent by adding a new 'be_silent' flag into 'struct apply_state'. It is a new feature in the libified apply functionality.

This could be in a separate series, but unfortunately using the libified apply in "git am" unmasks the fact that "git am", since it was a shell script, has been silencing the apply functionality by redirecting file descriptors to /dev/null and it looks like this is not acceptable in C.

There are no changes in these patches compared to v7.

  - Patch 40/41 was in v1, v2, v6 and v7.

This patch makes `git am` use the libified functionality. It is the same as in v7.

  - Patch 41/41 was in v6 and v7.

It replaces some calls to error() with calls to error_errno(). It is the nearly the same as in v7. The only change is that one call to error() is now replaced by error_errno() earlier in the series.

General comments ~~~~~~~~~~~~~~~~

Sorry if this patch series is still long. I can split it into two or more series if it is prefered.

I will send a diff between this version and the previous one, as a reply to this email.

The benefits are not just related to not creating new processes. When `git am` launched a `git apply` process, this new process had to read the index from disk. Then after the `git apply`process had terminated, `git am` dropped its index and read the index from disk to get the index that had been modified by the `git apply`process. This was inefficient and also prevented the split-index mechanism to provide many performance benefits.

Using this series as rebase material, Duy explains it like this:

 > Without the series, the picture is not so surprising. We run git-apply
 > 80+ times, each consists of this sequence
 >
 > read index
 > write index (cache tree updates only)
 > read index again
 > optionally initialize name hash (when new entries are added, I guess)
 > read packed-refs
 > write index
 >
 > With this series, we run a single git-apply which does
 >
 > read index (and sharedindex too if in split-index mode)
 > initialize name hash
 > write index 80+ times

(See: http://thread.gmane.org/gmane.comp.version-control.git/292324/focus=292460)

Links ~~~~~

This patch series is available here:

https://github.com/chriscool/git/commits/libify-apply-use-in-am

The previous versions are available there:

v1: https://github.com/chriscool/git/commits/libify-apply-use-in-am25 v2: https://github.com/chriscool/git/commits/libify-apply-use-in-am54 v6: https://github.com/chriscool/git/commits/libify-apply-use-in-am65 v7: https://github.com/chriscool/git/commits/libify-apply-use-in-am75

Performance numbers ~~~~~~~~~~~~~~~~~~~

Only tests on Linux have been performed using a very early version of this series. It could be interesting to test on other platforms especially Windows and perhaps OSX too.

  - Around mid April Ævar did a huge many-hundred commit rebase on the
    kernel with untracked cache.

command: git rebase --onto 1993b17 52bef0c 29dde7c

Vanilla "next" without split index: 1m54.953s Vanilla "next" with split index: 1m22.476s This series on top of "next" without split index: 1m12.034s This series on top of "next" with split index: 0m15.678s

Ævar used his Debian laptop with SSD.

  - Around mid April I tested rebasing 13 commits in Booking.com's
    monorepo on a Red Hat 6.5 server with split-index and
    GIT_TRACE_PERFORMANCE=1.

With Git v2.8.0, the rebase took 6.375888383 s, with the git am command launched by the rebase command taking 3.705677431 s.

With this series on top of next, the rebase took 3.044529494 s, with the git am command launched by the rebase command taking 0.583521168 s.

Christian Couder (41):
  apply: make some names more specific
  apply: move 'struct apply_state' to apply.h
  builtin/apply: make apply_patch() return -1 or -128 instead of
    die()ing
  builtin/apply: read_patch_file() return -1 instead of die()ing
  builtin/apply: make find_header() return -128 instead of die()ing
  builtin/apply: make parse_chunk() return a negative integer on error
  builtin/apply: make parse_single_patch() return -1 on error
  builtin/apply: make parse_whitespace_option() return -1 instead of
    die()ing
  builtin/apply: make parse_ignorewhitespace_option() return -1 instead
    of die()ing
  builtin/apply: move init_apply_state() to apply.c
  apply: make init_apply_state() return -1 instead of exit()ing
  builtin/apply: make check_apply_state() return -1 instead of die()ing
  builtin/apply: move check_apply_state() to apply.c
  builtin/apply: make apply_all_patches() return 128 or 1 on error
  builtin/apply: make parse_traditional_patch() return -1 on error
  builtin/apply: make gitdiff_*() return 1 at end of header
  builtin/apply: make gitdiff_*() return -1 on error
  builtin/apply: change die_on_unsafe_path() to check_unsafe_path()
  builtin/apply: make build_fake_ancestor() return -1 on error
  builtin/apply: make remove_file() return -1 on error
  builtin/apply: make add_conflicted_stages_file() return -1 on error
  builtin/apply: make add_index_file() return -1 on error
  builtin/apply: make create_file() return -1 on error
  builtin/apply: make write_out_one_result() return -1 on error
  builtin/apply: make write_out_results() return -1 on error
  builtin/apply: make try_create_file() return -1 on error
  builtin/apply: make create_one_file() return -1 on error
  builtin/apply: rename option parsing functions
  apply: rename and move opt constants to apply.h
  Move libified code from builtin/apply.c to apply.{c,h}
  apply: make some parsing functions static again
  environment: add set_index_file()
  write_or_die: use warning() instead of fprintf(stderr, ...)
  apply: add 'be_silent' variable to 'struct apply_state'
  apply: make 'be_silent' incompatible with 'apply_verbosely'
  apply: don't print on stdout when be_silent is set
  usage: add set_warn_routine()
  usage: add get_error_routine() and get_warn_routine()
  apply: change error_routine when be_silent is set
  builtin/am: use apply api in run_apply()
  apply: use error_errno() where possible

 Makefile               |    1 +
 apply.c                | 4891 ++++++++++++++++++++++++++++++++++++++++++++++++
 apply.h                |  133 ++
 builtin/am.c           |   91 +-
 builtin/apply.c        | 4803 +----------------------------------------------
 cache.h                |    1 +
 environment.c          |   10 +
 git-compat-util.h      |    3 +
 t/t4012-diff-binary.sh |    4 +-
 t/t4254-am-corrupt.sh  |    2 +-
 usage.c                |   15 +
 write_or_die.c         |    6 +-
 12 files changed, 5147 insertions(+), 4813 deletions(-)
 create mode 100644 apply.c
 create mode 100644 apply.h