RSS Atom Add a new entry named:

I wanted to import an old GNU arch repository into Git, but only had HTTP access via ArchZoom. I spent quite some time to try teaching git archimport to use HTTP access to that repository, but this didn't work out. Too bad -- but at least, using ArchZoom, I was able to get the individual revisions' tarballs:

$ ls -1 *.tar.gz
bpf--devel--0.0--base-0.tar.gz
bpf--devel--0.0--patch-1.tar.gz
bpf--devel--0.0--patch-10.tar.gz
bpf--devel--0.0--patch-11.tar.gz
bpf--devel--0.0--patch-12.tar.gz
bpf--devel--0.0--patch-2.tar.gz
bpf--devel--0.0--patch-3.tar.gz
[...]
bpf--devel--0.0--patch-9.tar.gz
bpf--release--0.1--base-0.tar.gz
bpf--release--0.1--patch-1.tar.gz
bpf--release--0.1--patch-2.tar.gz
[...]
bpf--release--0.1--patch-8.tar.gz

I unpacked these:

$ for f in *.tar.gz; do tar -xz < "$f" || echo >&2 "$f" failed; done

The last revision's tree apparently contains all previous revisions' commit information (author, date, message), so use that:

$ cp -a ↩
    bpf--release--0.1--patch-8/{arch}/bpf/bpf--devel/bpf--devel--0.0/info@hurdfr.org--hurdfr/patch-log ↩
    d-patch-log
$ cp -a ↩
    bpf--release--0.1--patch-8/{arch}/bpf/bpf--release/bpf--release--0.1/info@hurdfr.org--hurdfr/patch-log ↩
    r-patch-log

... and extract the information that we need:

$ base=bpf--devel--0.0-- && ↩
  for f in d-patch-log/*; do ↩
    grep < "$f" ^Creator: | head -n 1 ↩
      | { read j c && ↩
          echo "$c" | sed s%' <.*'%% ↩
            > "$base""$(basename "$f")".author_name && ↩
          echo "$c" | sed -e 's%.*<%%' -e 's%>.*%%' ↩
            > "$base""$(basename "$f")".author_email; } && ↩
    grep < "$f" ^Standard-date: | head -n 1 | { read j d && echo "$d" ↩
      > "$base""$(basename "$f")".author_date; } && ↩
    { grep < "$f" ^Summary: | head -n 1 | { read j m && echo "$m"; } && ↩
      echo && sed < "$f" '1,/^$/d'; } ↩
      > "$base""$(basename "$f")".log ↩
    || echo >&2 "$f" failed; ↩
  done
$ base=bpf--release--0.1-- && ↩
  for f in r-patch-log/*; [...]

(Of course, I could have used something more elaborate than shell scripting...)

Remove the GNU arch stuff that we don't need anymore:

$ find bpf--*/ -type d \( -name {arch} -o -name .arch-ids \) -print0 ↩
    | xargs -0 rm -r

The base-0 revisions are actually either empty (the devel one) or equivalent to the previous revision (the release one), so remove these:

$ rm -rf bpf--devel--0.0--base-0 bpf--release--0.1--base-0

Finally, import all the other ones:

$ mkdir g && ( cd g/ && git init )
$ for d in bpf--d*-? bpf--d*-?? bpf--r*; do ↩
    test -d "$d"/ || continue && ↩
    ( cd g/ && ↩
      rsync -a --delete --exclude=/.git ../"$d"/ ./ && ↩
      git add . && ↩
      GIT_AUTHOR_NAME="$(cat ../"$d".author_name)" ↩
        GIT_AUTHOR_EMAIL="$(cat ../"$d".author_email)" ↩
        GIT_AUTHOR_DATE="$(cat ../"$d".author_date)" ↩
        git commit -F ../"$d".log -a ); ↩
  done

Voilà!

Update 2009-06-25:

Half a day later, ?HurdFr published a git archimport-converted repository -- which was identical to my hand-crafted one (apart from having git-archimport-id: tags in the commit messages, and the first (empty) commit not being stripped off). :-)

Posted 2009-06-24 13:52:34 CEST

I was revisiting the issue of getting the Hurd's code base compiled with recent versions of GCC. Specifically, there were a lot of duplicate symbols shown at linking time, and all these were related to inline functions. Originally, in 2007, we had solved this problem already (or rather, shifted it) by using GCC's -fgnu89-inline option, but as we saw now, that one obviously doesn't help anymore if third-party code is using the Hurd's unfixed header files.

So I was revisiting this issue. I was already prepared that this would take some hours, with lots of editing, compiling cycles, plus some analyzing of the binaries. So I made up a fresh repository for this work.

$ mkdir hurd-ei
$ cd hurd-ei/
$ git init
[...]
$ git remote add savannah git://git.savannah.gnu.org/hurd/hurd.git
$ git fetch
[...]

Switch to a new topic-branch.

$ git checkout -b master-ei savannah/master
Branch master-ei set up to track remote branch master from savannah.
Switched to a new branch 'master-ei'

(ei is short for extern inline.)

The first thing to do was to disable that -fgnu89-inline option, so I edited Makeconf where it was added to CFLAGS.

I started editing, compiling, editing, compiling, and so on.

Finally, the tree was in a shape where everything was building fine and the resulting libraries contained the symbols they should, etc.

I committed the whole junk as one big blob commit, to store it in a safe place (you never know with these Hurd machines...), and to continue working on it the next day.

$ git commit -a

For the commit message, I already mostly assembled a ChangeLog-style log. Then:

$ git format-patch savannah/master..
0001-Bla.patch

... and here is 0001-Bla.patch.bz2 (compressed).

The next day, a.k.a. today, in a different Git repository.

$ git checkout -b master-fix_inline savannah/master
Branch master-fix_inline set up to track remote branch master from savannah.
Switched to a new branch 'master-fix_inline'
$ bunzip2 < ../some/where/0001-Bla.patch.bz2 | git am
Applying: Bla.

The big blob is now on top of savannah/master (which was 2772f5c6a6a51cf946fd95bf6ffe254273157a21, by the way -- in case that you want to reproduce this tutorial later, simply substitute savannah/master with 2772...).

By then, I had come to the conclusion that the commit essentially was fine, but should be split into two, and the configure hunk shouldn't be in there. So be it.

So, the HEAD of the active branch is our big blob commit that we want to work on. Check with git show HEAD:

$ git show HEAD
commit 93e97f3351337c349e2926f4041e61bc487ef9e6
Author: Thomas Schwinge <tschwinge@gnu.org>
Date:   Tue Jun 23 00:27:28 2009 +0200

    Bla.

    * console-client/timer.h (fetch_jiffies): Use static inline instead of extern
    inline.
    * ext2fs/ext2fs.h (test_bit, set_bit, clear_bit, dino, global_block_modified)
    (record_global_poke, sync_global_ptr, record_indir_poke, sync_global)
    (alloc_sync): Likewise.
    * libftpconn/priv.h (unexpected_reply): Likewise.
    * term/term.h (qsize, qavail, clear_queue, dequeue_quote, dequeue)
    (enqueue_internal, enqueue, enqueue_quote, unquote_char, char_quoted_p)
    (queue_erase): Likewise.
    * ufs/ufs.h (dino, indir_block, cg_locate, sync_disk_blocks, sync_dinode)
    (swab_short, swab_long, swab_long_long): Likewise.
    * term/munge.c (poutput): Use static inline instead of inline.

    * libdiskfs/diskfs.h: Apply inline optimization only ifdef
    [__USE_EXTERN_INLINES].  Use __extern_inline instead of extern inline.
    * libftpconn/ftpconn.h: Likewise.
    * libpipe/pipe.h: Likewise.
    * libpipe/pq.h: Likewise.
    * libshouldbeinlibc/idvec.h: Likewise.
    * libshouldbeinlibc/maptime.h: Likewise.
    * libshouldbeinlibc/ugids.h: Likewise.
    * libstore/store.h: Likewise.
    * libthreads/rwlock.h: Likewise.
    * libdiskfs/extern-inline.c: Adapt to these changes.
    * libftpconn/xinl.c: Likewise.  And don't #include "priv.h".
    * libpipe/pipe-funcs.c: Likewise.
    * libpipe/pq-funcs.c: Likewise.
    * libshouldbeinlibc/maptime-funcs.c: Likewise.  And remove superfluous
    includes.
    * libstore/xinl.c: Likewise.
    * libthreads/rwlock.c: Likewise.

    * Makeconf (CFLAGS): Don't append $(gnu89-inline-CFLAGS).
    * pfinet/Makefile (CFLAGS): Append $(gnu89-inline-CFLAGS).

diff --git a/Makeconf b/Makeconf
index e9b2045..236f1ec 100644
--- a/Makeconf
+++ b/Makeconf
@@ -65,7 +65,7 @@ INCLUDES += -I$(..)include -I$(top_srcdir)/include
 CPPFLAGS += $(INCLUDES) \
             -D_GNU_SOURCE -D_IO_MTSAFE_IO -D_FILE_OFFSET_BITS=64 \
            $($*-CPPFLAGS)
-CFLAGS += -std=gnu99 $(gnu89-inline-CFLAGS) -Wall -g -O3 \
+CFLAGS += -std=gnu99 -Wall -g -O3 \
[...]

We want to undo this one commit, but preserve its changes in the working directory.

$ git reset HEAD^
Makeconf: locally modified
configure: locally modified
console-client/timer.h: locally modified
ext2fs/ext2fs.h: locally modified
libdiskfs/diskfs.h: locally modified
libdiskfs/extern-inline.c: locally modified
libftpconn/ftpconn.h: locally modified
libftpconn/priv.h: locally modified
libftpconn/xinl.c: locally modified
libpipe/pipe-funcs.c: locally modified
libpipe/pipe.h: locally modified
libpipe/pq-funcs.c: locally modified
libpipe/pq.h: locally modified
libshouldbeinlibc/idvec.h: locally modified
libshouldbeinlibc/maptime-funcs.c: locally modified
libshouldbeinlibc/maptime.h: locally modified
libshouldbeinlibc/ugids.h: locally modified
libstore/store.h: locally modified
libstore/xinl.c: locally modified
libthreads/rwlock.c: locally modified
libthreads/rwlock.h: locally modified
pfinet/Makefile: locally modified
term/munge.c: locally modified
term/term.h: locally modified
ufs/ufs.h: locally modified

Now, HEAD points to the commit before the previous HEAD, i.e. HEAD^. Again, check with git show HEAD:

$ git show HEAD
commit 2772f5c6a6a51cf946fd95bf6ffe254273157a21
Author: Samuel Thibault <samuel.thibault@ens-lyon.org>
Date:   Thu Apr 2 23:06:37 2009 +0000

    2009-04-03  Samuel Thibault  <samuel.thibault@ens-lyon.org>

        * exec.c (prepare): Call PREPARE_STREAM earlier to permit calling
        finish_mapping on E even after errors, as is already done in do_exec.

diff --git a/exec/ChangeLog b/exec/ChangeLog
index 5a0ad1d..a9300bf 100644
--- a/exec/ChangeLog
+++ b/exec/ChangeLog
@@ -1,3 +1,8 @@
+2009-04-03  Samuel Thibault  <samuel.thibault@ens-lyon.org>
+
+       * exec.c (prepare): Call PREPARE_STREAM earlier to permit calling
+       finish_mapping on E even after errors, as is already done in do_exec.
+
 2008-06-10  Samuel Thibault  <samuel.thibault@ens-lyon.org>

        * elfcore.c (TIME_VALUE_TO_TIMESPEC): Completely implement instead of
diff --git a/exec/exec.c b/exec/exec.c
index 05dc883..cb3d741 100644
--- a/exec/exec.c
+++ b/exec/exec.c
@@ -726,6 +726,9 @@ prepare (file_t file, struct execdata *e)

   e->interp.section = NULL;

+  /* Initialize E's stdio stream.  */
+  prepare_stream (e);
[...]

Luckily, Git saves the previous (i.e. before the git reset) HEAD reference as ORIG_HEAD. Have a look at it with git show ORIG_HEAD -- it contains the big blob commit, including the preliminary commit message -- just what HEAD was before:

$ git show ORIG_HEAD
commit 93e97f3351337c349e2926f4041e61bc487ef9e6
Author: Thomas Schwinge <tschwinge@gnu.org>
Date:   Tue Jun 23 00:27:28 2009 +0200

    Bla.

    * console-client/timer.h (fetch_jiffies): Use static inline instead of extern
    inline.
[...]

diff --git a/Makeconf b/Makeconf
index e9b2045..236f1ec 100644
--- a/Makeconf
+++ b/Makeconf
@@ -65,7 +65,7 @@ INCLUDES += -I$(..)include -I$(top_srcdir)/include
 CPPFLAGS += $(INCLUDES) \
             -D_GNU_SOURCE -D_IO_MTSAFE_IO -D_FILE_OFFSET_BITS=64 \
            $($*-CPPFLAGS)
-CFLAGS += -std=gnu99 $(gnu89-inline-CFLAGS) -Wall -g -O3 \
+CFLAGS += -std=gnu99 -Wall -g -O3 \
[...]

OK, now let's pick the files that we want to have in the first of the envisioned two commits: these are the static inline instead of extern inline and apply inline optimization only... sections.

$ git add console-client/timer.h ext2fs/ext2fs.h [...] libthreads/rwlock.c

Oh, we forgot something: now that we're preparing this stuff to go into the master repository, update the copyright years. Edit, edit, edit, and then, again:

$ git add console-client/timer.h ext2fs/ext2fs.h [...] libthreads/rwlock.c

Now Git's staging area contains the changes that we want to commit (and the working directory contains the rest of the big blob). Commit these added files, and use big blob's commit message as a template for the new one, as it already contains most of what we want (don't forget to chop off the unneeded parts).

$ git commit -c ORIG_HEAD
Waiting for Emacs...
[master-fix_inline 51c15bc] Use static inline where appropriate.
 6 files changed, 50 insertions(+), 51 deletions(-)
$ git show HEAD
commit c6c9d7a69dea26e04bba7010582e7bcd612e710c
Author: Thomas Schwinge <tschwinge@gnu.org>
Date:   Tue Jun 23 00:27:28 2009 +0200

    Use static inline where appropriate and use glibc's __extern_inline machinery.

    * console-client/timer.h (fetch_jiffies): Use static inline instead of extern
    inline.
    * ext2fs/ext2fs.h (test_bit, set_bit, clear_bit, dino, global_block_modified)
    (record_global_poke, sync_global_ptr, record_indir_poke, sync_global)
    (alloc_sync): Likewise.
    * libftpconn/priv.h (unexpected_reply): Likewise.
    * term/term.h (qsize, qavail, clear_queue, dequeue_quote, dequeue)
    (enqueue_internal, enqueue, enqueue_quote, unquote_char, char_quoted_p)
    (queue_erase): Likewise.
    * ufs/ufs.h (dino, indir_block, cg_locate, sync_disk_blocks, sync_dinode)
    (swab_short, swab_long, swab_long_long): Likewise.
    * term/munge.c (poutput): Use static inline instead of inline.

    * libdiskfs/diskfs.h: Apply inline optimization only ifdef
    [__USE_EXTERN_INLINES].  Use __extern_inline instead of extern inline.
    * libftpconn/ftpconn.h: Likewise.
    * libpipe/pipe.h: Likewise.
    * libpipe/pq.h: Likewise.
    * libshouldbeinlibc/idvec.h: Likewise.
    * libshouldbeinlibc/maptime.h: Likewise.
    * libshouldbeinlibc/ugids.h: Likewise.
    * libstore/store.h: Likewise.
    * libthreads/rwlock.h: Likewise.
    * libdiskfs/extern-inline.c: Adapt to these changes.
    * libftpconn/xinl.c: Likewise.  And don't #include "priv.h".
    * libpipe/pipe-funcs.c: Likewise.
    * libpipe/pq-funcs.c: Likewise.
    * libshouldbeinlibc/maptime-funcs.c: Likewise.  And remove superfluous
    includes.
    * libstore/xinl.c: Likewise.
    * libthreads/rwlock.c: Likewise.

diff --git a/console-client/timer.h b/console-client/timer.h
index 4204192..5e64e97 100644
--- a/console-client/timer.h
+++ b/console-client/timer.h
@@ -1,5 +1,7 @@
 /* timer.h - Interface to a timer module for Mach.
-   Copyright (C) 1995,96,2000,02 Free Software Foundation, Inc.
+
+   Copyright (C) 1995, 1996, 2000, 2002, 2009 Free Software Foundation, Inc.
+
    Written by Michael I. Bushnell, p/BSG and Marcus Brinkmann.

    This file is part of the GNU Hurd.
@@ -54,7 +56,7 @@ int timer_remove (struct timer_list *timer);
 /* Change the expiration time of the timer TIMER to EXPIRES.  */
 void timer_change (struct timer_list *timer, long long expires);

-extern inline long long
+static inline long long
[...]

As you can see, HEAD now points to the new commit on top of the current branch. (ORIG_HEAD doesn't change.)

On to the next, and last one, only two changes should be left: the Makeconf and pfinet/Makefile ones.

$ git status
# On branch master-fix_inline
# Your branch is ahead of 'savannah/master' by 1 commit.
#
# Changed but not updated:
#   (use "git add <file>..." to update what will be committed)
#   (use "git checkout -- <file>..." to discard changes in working directory)
#
#       modified:   Makeconf
#       modified:   configure
#       modified:   pfinet/Makefile
#
# Untracked files:
#   (use "git add <file>..." to include in what will be committed)
#
#       0001-Bla.patch
#       autom4te.cache/
#       hurd_extern_inline_fix.patch?file_id=18191
no changes added to commit (use "git add" and/or "git commit -a")

Alright, there is as well still the configure hunk that we want to get rid of. But first for the real second commit, after editing for again adding the copyright year update:

$ git add Makeconf pfinet/Makefile
$ git commit -c ORIG_HEAD
Waiting for Emacs...
[master-fix_inline 6a967d1] We're now C99 inline safe -- apart from the Linux code in pfinet.
 2 files changed, 6 insertions(+), 3 deletions(-)

Check that we're in a clean state now:

$ git status
# On branch master-fix_inline
# Your branch is ahead of 'savannah/master' by 2 commits.
#
# Changed but not updated:
#   (use "git add <file>..." to update what will be committed)
#   (use "git checkout -- <file>..." to discard changes in working directory)
#
#       modified:   configure
#
# Untracked files:
#   (use "git add <file>..." to include in what will be committed)
#
#       0001-Bla.patch
#       autom4te.cache/
#       hurd_extern_inline_fix.patch?file_id=18191
no changes added to commit (use "git add" and/or "git commit -a")

Oops, we forgot something...

$ git checkout -- configure

Now, our tree is clean again. (Check with git status.)

By now, we came to the conclusion that the first of the two commits should have been further split into two separate ones. Of course, essentially we would do the same splitting again that we've done just now -- but how to easily modify the first commit, now that we have another one on top of it?

Alright, git rebase --interactive to the rescue -- let's interactively rebase the last two commits, to modify them as wanted.

$ git rebase --interactive HEAD~2
Waiting for Emacs...

Emacs wants us to tell which commits we want to keep as they are (pick), which should be merged into others (squash), and which we want to edit. In our scenario, we want to edit the first one and pick the second one. Change the file thusly and close it.

Stopped at 5becbb5... Use static inline where appropriate and use...
You can amend the commit now, with

        git commit --amend

Once you are satisfied with your changes, run

        git rebase --continue

We want to undo this first commit to split it into two. Again, use git reset for that, while preserving the commit's changes in the working directory.

$ git reset HEAD^
console-client/timer.h: locally modified
[...]

Pick the set of files that we want to have in the first of the envisioned two commits: the static inline instead of extern inline section, and commit them, again using the previous commit message as a template for the new one:

$ git add console-client/timer.h ext2fs/ext2fs.h [...] term/munge.c
$ git commit -c ORIG_HEAD
Waiting for Emacs...
[detached HEAD 51c15bc] Use static inline where appropriate.
 6 files changed, 50 insertions(+), 51 deletions(-)

Next part: apply inline optimization only.... Again, git add those files that shall be part of the next commit, i.e. all remaining ones. As before, use the previous commit message as a template.

$ git add libdiskfs/diskfs.h [...] libthreads/rwlock.c
$ git commit -c ORIG_HEAD
Waiting for Emacs...
[detached HEAD 8ac30ea] [__USE_EXTERN_INLINES]: Use glibc's __extern_inline machinery.
 16 files changed, 508 insertions(+), 356 deletions(-)

Now we're done with splitting that commit into two. (Check with git status that we didn't forget anything.) What's missing is getting back the other commit on top of the two now-split ones:

$ git rebase --continue
Successfully rebased and updated refs/heads/master-fix_inline.

Here we go. The other commit has been applied on top of the two new ones.

Due to time-honored tradition, I always double-check what I have just committed, before distributing it to the world:

$ git log --reverse -p -C --cc savannah/master..

... and promptly, I recognize some changes that shouldn't be in there: when using it on some files, Emacs' copyright-fix-years, aside from indeed fixing the list of copyright years, and adding the current year, also changed GPL ... version 2 into version 3, which would be nice, but which we can't do for the moment. The error is present only in the first and second commit. If it were in only in the third (the last) one, simply editing the files, and then using git commit --amend would be the solution. But again there is the problem about how to modify the first (HEAD~2) and second (HEAD~1, or HEAD^) commit now that there is another one on top of it. By now, we know the solution:

$ git rebase --interactive HEAD~3
Waiting for Emacs...

This time, we need to edit the first and second commits, and pick the third one.

Stopped at ffd215b... Use static inline where appropriate.
You can amend the commit now, with

        git commit --amend

Once you are satisfied with your changes, run

        git rebase --continue

git show (which defaults to showing HEAD, by the way) can again be used to have a look at the current HEAD (which is the first of the three commits), and then we revert the unwanted changes in the editor, resulting with the following changed files:

$ git status
# Not currently on any branch.
# Changed but not updated:
#   (use "git add <file>..." to update what will be committed)
#   (use "git checkout -- <file>..." to discard changes in working directory)
#
#       modified:   ext2fs/ext2fs.h
#       modified:   libftpconn/priv.h
#       modified:   term/munge.c
#       modified:   term/term.h
#       modified:   ufs/ufs.h
#
# Untracked files:
#   (use "git add <file>..." to include in what will be committed)
#
#       0001-Bla.patch
#       autom4te.cache/
#       hurd_extern_inline_fix.patch?file_id=18191
no changes added to commit (use "git add" and/or "git commit -a")

Then, we can -- as git rebase suggested above -- amend the existing HEAD commit with these changes (--amend and --all), reusing HEAD's commit message without spawning an editor (-C HEAD):

$ git commit --amend -C HEAD --all
[detached HEAD c6c9d7a] Use static inline where appropriate.
 6 files changed, 45 insertions(+), 46 deletions(-)

Continue with the next commit:

$ git rebase --continue
Stopped at 8ac30ea... [__USE_EXTERN_INLINES]: Use glibc's __extern_inline machinery.
You can amend the commit now, with

        git commit --amend

Once you are satisfied with your changes, run

        git rebase --continue

Again, have a look at the commit (git show), revert the unwanted changes, amend HEAD, and continue to the next commit:

$ git commit --amend -C HEAD --all
[detached HEAD 9990dc6] [__USE_EXTERN_INLINES]: Use glibc's __extern_inline machinery.
 16 files changed, 500 insertions(+), 348 deletions(-)
$ git rebase --continue
Stopped at 6a967d1... We're now C99 inline safe -- apart from the Linux code in pfinet.
You can amend the commit now, with

        git commit --amend

Once you are satisfied with your changes, run

        git rebase --continue

Two files are left to be edited (git show, etc., again), and finally:

$ git commit --amend -C HEAD --all
[detached HEAD 241c605] We're now C99 inline safe -- apart from the Linux code in pfinet.
 2 files changed, 5 insertions(+), 2 deletions(-)
$ git rebase --continue
Successfully rebased and updated refs/heads/master-fix_inline.

That's it. git log --reverse -p -C --cc savannah/master.. now looks as nice as can be.

Of course, this is only a small insight of what is possible to do with git rebase and friends -- see the manual for further explanations.

Posted 2009-06-23 17:15:59 CEST