sequencer.c: set GIT_WORK_TREE in run_git_checkout
PATCH 1/1 sequencer.c: set GIT_WORK_TREE in run_git_checkout
When rebasing in a subdirectory of a worktree, Git fails due to a dirty worktree:
error: The following untracked working tree files would be overwritten
by merge:
a/b/c
Please move or remove them before you merge.
This occurs because "git rebase" invokes a "git checkout" without propagating the GIT_WORK_TREE environment variable, causing the worktree to be assumed to be ".". This was not an issue until bc3ae46b42 (rebase:
do not attempt to remove startup_info->original_cwd, 2021-12-09), when the .dir of the "git checkout" child process was changed such that it no longer runs at the root of the worktree.
Propagate GIT_WORK_TREE to the "git checkout" child process and test that rebase in a subdirectory of a worktree succeeds.
---
Here is a fix for the bug I found in [1]. I didn't look through the rest of the "preserve cwd" thread for other possible bugs pertaining to worktrees, but I didn't find any during a cursory glance at sequencer.c.
I'm also not sure if this is the idiomatic way to do it since, I assume, we'd always want to propagate GIT_WORK_TREE, but this is identical to how do_exec() sets GIT_WORK_TREE in the same file (perhaps this is something that should be refactored).
[1] https://lore.kernel.org/git/[email protected]
sequencer.c | 2 ++ t/t3400-rebase.sh | 29 +++++++++++++++++++++++++++++ 2 files changed, 31 insertions(+)
sequencer.c +2 -0
@@ -4233,6 +4233,8 @@ static int run_git_checkout(struct repository *r, struct replay_opts *opts,strvec_push (&cmd.args,"checkout" );strvec_push (&cmd.args, commit);strvec_pushf (&cmd.env_array,GIT_REFLOG_ACTION "=%s" , action); +strvec_pushf (&cmd.env_array,"GIT_WORK_TREE=%s" , +absolute_path (get_git_work_tree ()));if (opts->verbose) ret =run_command (&cmd);
t/t3400-rebase.sh +29 -0
@@ -416,4 +416,33 @@ test_expect_success MINGW,SYMLINKS_WINDOWS 'rebase when .git/logs is a symlink' mv actual_logs .git/logs ' +test_expect_success 'rebase when inside worktree subdirectory' ' + git init main-wt && + ( + cd main-wt && + git commit --allow-empty -m "initial" && + # create commit with foo/bar/baz + mkdir -p foo/bar && + touch foo/bar/baz && + git add foo/bar/baz && + git commit -m "add foo/bar/baz" && + # create commit with a/b/c + mkdir -p a/b && + touch a/b/c && + git add a/b/c && + git commit -m "add a/b/c" && + # create another branch for our other worktree + git branch other && + git worktree add ../other-wt other && + ( + cd ../other-wt && + mkdir -p random/dir && + ( + cd random/dir && + git rebase --onto HEAD^^ HEAD^ # drops the HEAD^ commit + ) + ) + ) +' + test_done
base-commit: bc3ae46b420f58dfe2bfd87714dca096a043d554
Signature · 1 line
2.33.GIT
Quoted 24 lines · Glen Choo
When rebasing in a subdirectory of a worktree, Git fails due to a dirty worktree:
error: The following untracked working tree files would be overwritten
by merge:a/b/cPlease move or remove them before you merge.
This occurs because "git rebase" invokes a "git checkout" without propagating the
GIT_WORK_TREEenvironment variable, causing the worktree to be assumed to be ".". This was not an issue untilbc3ae46b42(rebase:
do not attempt to removestartup_info->original_cwd, 2021-12-09), when the .dir of the "git checkout" child process was changed such that it no longer runs at the root of the worktree.Propagate
GIT_WORK_TREEto the "git checkout" child process and test that rebase in a subdirectory of a worktree succeeds.Signed-off-by Glen Choo <[email protected]>---
Here is a fix for the bug I found in [1]. I didn't look through the rest of the "preserve cwd" thread for other possible bugs pertaining to worktrees, but I didn't find any during a cursory glance atsequencer.c.
Quoted 6 lines
I'm also not sure if this is the idiomatic way to do it since, I assume, we'd always want to propagate
GIT_WORK_TREE, but this is identical to howdo_exec()setsGIT_WORK_TREEin the same file (perhaps this is something that should be refactored).
I forgot to add this to the previous email:
cc-ing Junio because this bug is in master.
Quoted 1 line · Glen Choo
...
Quoted 3 lines
I forgot to add this to the previous email:
cc-ing Junio because this bug is in master.
...and is a new regression in 2.35. :-(
Sorry about that.
Looks like you beat me to sending out a patch.
Quoted 16 lines · Glen Choo
When rebasing in a subdirectory of a worktree, Git fails due to a dirty worktree:
error: The following untracked working tree files would be overwritten
by merge:a/b/cPlease move or remove them before you merge.
This occurs because "git rebase" invokes a "git checkout" without propagating the
GIT_WORK_TREEenvironment variable, causing the worktree to be assumed to be ".". This was not an issue untilbc3ae46b42(rebase:
do not attempt to removestartup_info->original_cwd, 2021-12-09), when the .dir of the "git checkout" child process was changed such that it no longer runs at the root of the worktree.
Quoted 2 lines
Propagate
GIT_WORK_TREEto the "git checkout" child process and test that rebase in a subdirectory of a worktree succeeds.
This is correct, but leaves the reader wondering why the failure to set GIT_WORK_TREE fails when in a non-primary worktree, but does not fail in the primary working tree. Here's how I described it in the patch I was about to send to the list:
"""
In commit bc3ae46 ("rebase: do not attempt to remove startup_info->original_cwd", 2021-12-09), we wanted to allow the checkout subprocess to know which directory the parent process was running from, so that the subprocess could protect it. However...
When run from a non-main worktree, setup_git_directory() will note that the discovered git directory (/PATH/TO/.git/worktree/non-main-worktree) does not match DEFAULT_GIT_DIR_ENVIRONMENT (see setup_discovered_git_dir()), and decide to set GIT_DIR in the environment. This matters because...
Whenever git is run with the GIT_DIR environment variable set, and GIT_WORK_TREE not set, it presumes that '.' is the working tree. So...
This combination results in rebase failing when run from a subdirectory of a non-main worktree. Fix it by also setting the GIT_WORK_TREE environment variable along with setting cmd.dir.
A possibly more involved fix we could consider for later would be to make setup.c set GIT_WORK_TREE whenever (a) it discovers both the git directory and the working tree and (b) it decides to set GIT_DIR in the environment. I did not attempt that here as such would be too big of a change for a 2.35.1 release.
"""
(Another fix would also be making sequencer stop forking a "git checkout" subprocess; it should be able to call a library function, and then quit dropping and re-reading the index. But that's also a much bigger change...)
Quoted 5 lines
Signed-off-by Glen Choo <[email protected]>---
Here is a fix for the bug I found in [1]. I didn't look through the rest of the "preserve cwd" thread for other possible bugs pertaining to worktrees, but I didn't find any during a cursory glance atsequencer.c.
We'll need the same fix for stash due to 0fce211ccc ("stash: do not attempt to remove startup_info->original_cwd", 2021-12-09) -- in that case, though, it's calling git-clean rather than git-checkout. But the same idea.
Quoted 4 lines
I'm also not sure if this is the idiomatic way to do it since, I assume, we'd always want to propagate
GIT_WORK_TREE, but this is identical to howdo_exec()setsGIT_WORK_TREEin the same file (perhaps this is something that should be refactored).
I'd rather only set GIT_WORK_TREE if something is also setting GIT_DIR, but I don't think it hurts to set it in all cases.
Quoted 11 lines
[1] https://lore.kernel.org/git/[email protected]
sequencer.c | 2 ++ t/t3400-rebase.sh | 29 +++++++++++++++++++++++++++++ 2 files changed, 31 insertions(+)sequencer.c +0
-0
sequencer.c static int run_git_checkout(struct repository *r, struct replay_opts *opts, · line 4237
Quoted 5 lines
sequencer.c +2
-0@@ -4233,6 +4233,8 @@ static int run_git_checkout(struct repository *r, struct replay_opts *opts,strvec_push (&cmd.args,"checkout" );strvec_push (&cmd.args, commit);strvec_pushf (&cmd.env_array,GIT_REFLOG_ACTION "=%s" , action); +strvec_pushf (&cmd.env_array,"GIT_WORK_TREE=%s" , +absolute_path (get_git_work_tree ()));
struct repository *r already contains the path to the worktree in r->worktree; I think it'd be better to use it.
Quoted 28 lines
sequencer.c +0
-0if (opts->verbose) ret =run_command (&cmd);t/t3400-rebase.sh +18
-0@@ -416,4 +416,33 @@ test_expect_success MINGW,SYMLINKS_WINDOWS 'rebase when .git/logs is a symlink' mv actual_logs .git/logs ' +test_expect_success 'rebase when inside worktree subdirectory' ' + git init main-wt && + ( + cd main-wt && + git commit --allow-empty -m "initial" && + # create commit with foo/bar/baz + mkdir -p foo/bar && + touch foo/bar/baz && + git add foo/bar/baz && + git commit -m "add foo/bar/baz" && + # create commit with a/b/c + mkdir -p a/b && + touch a/b/c && + git add a/b/c && + git commit -m "add a/b/c" && + # create another branch for our other worktree + git branch other && + git worktree add ../other-wt other &&
t/t3400-rebase.sh test_expect_success MINGW,SYMLINKS_WINDOWS 'rebase when .git/logs is a symlink' · line 446
Quoted 10 lines
t/t3400-rebase.sh +10
-0+ ( + cd ../other-wt && + mkdir -p random/dir && + ( + cd random/dir && + git rebase --onto HEAD^^ HEAD^ # drops the HEAD^ commit + ) + ) + ) +'
Nice testcase. You would have seen potentially even more confusing results had some of these commits merely modified files rather than introducing totally new files.
Quoted 17 lines · Elijah Newren
Quoted 16 lines · Glen Choo
When rebasing in a subdirectory of a worktree, Git fails due to a dirty worktree:
error: The following untracked working tree files would be overwritten
by merge:a/b/cPlease move or remove them before you merge.
This occurs because "git rebase" invokes a "git checkout" without propagating the
GIT_WORK_TREEenvironment variable, causing the worktree to be assumed to be ".". This was not an issue untilbc3ae46b42(rebase:
do not attempt to removestartup_info->original_cwd, 2021-12-09), when the .dir of the "git checkout" child process was changed such that it no longer runs at the root of the worktree.
Quoted 6 lines
Quoted 2 lines
Propagate
GIT_WORK_TREEto the "git checkout" child process and test that rebase in a subdirectory of a worktree succeeds.This is correct, but leaves the reader wondering why the failure to set
GIT_WORK_TREEfails when in a non-primary worktree, but does not fail in the primary working tree.
Ah, that is true. Thanks for digging into it deeper.
Quoted 39 lines
Here's how I described it in the patch I was about to send to the list:
"""
In commit bc3ae46 ("rebase: do not attempt to removestartup_info->original_cwd", 2021-12-09), we wanted to allow the checkout subprocess to know which directory the parent process was running from, so that the subprocess could protect it. However...When run from a non-main worktree,
setup_git_directory()will note that the discovered git directory (/PATH/TO/.git/worktree/non-main-worktree) does not matchDEFAULT_GIT_DIR_ENVIRONMENT(seesetup_discovered_git_dir()), and decide to setGIT_DIRin the environment. This matters because...Whenever git is run with the
GIT_DIRenvironment variable set, andGIT_WORK_TREEnot set, it presumes that '.' is the working tree. So...This combination results in rebase failing when run from a subdirectory of a non-main worktree. Fix it by also setting the
GIT_WORK_TREEenvironment variable along with setting cmd.dir.A possibly more involved fix we could consider for later would be to make
setup.csetGIT_WORK_TREEwhenever (a) it discovers both the git directory and the working tree and (b) it decides to setGIT_DIRin the environment. I did not attempt that here as such would be too big of a change for a 2.35.1 release.
"""(Another fix would also be making sequencer stop forking a "git checkout" subprocess; it should be able to call a library function, and then quit dropping and re-reading the index. But that's also a much bigger change...)
Quoted 5 lines
Signed-off-by Glen Choo <[email protected]>---
Here is a fix for the bug I found in [1]. I didn't look through the rest of the "preserve cwd" thread for other possible bugs pertaining to worktrees, but I didn't find any during a cursory glance atsequencer.c.
Quoted 4 lines
We'll need the same fix for stash due to
0fce211ccc("stash: do not attempt to removestartup_info->original_cwd", 2021-12-09) -- in that case, though, it's calling git-clean rather than git-checkout. But the same idea.
Let's use your patch [1] instead of mine :)