Corner case bug caused by shell dependent behavior
Hi
When your system shell (/bin/sh) is a dash control sequences in strings get interpreted by the echo command. A commit message which ends with the string '\n' may result in a garbage line in the todo list of an interactive rebase which causes the rebase to fail.
To reproduce the behavior (with dash as /bin/sh):
mkdir test && cd test && git init echo 1 >foo && git add foo git commit -m"this commit message ends with '\n'" echo 2 >foo && git commit -a --fixup HEAD git rebase -i --autosquash --root
Now the editor opens with garbage in line 3 which has to be removed or the rebase fails.
The attached one-line patch fixes the bug. Be free to edit the commit message when it's too long.
Maybe there are more places where it would be more robust to use printf instead of echo.
Uwe
Quoted 1 line
From
53262bc8a7a3ec9d9a6b0e8ecaaea598257b87feMon Sep 17 00:00:00 2001
From: Uwe Storbeck <[email protected]>
Date: Fri, 14 Mar 2014 00:28:33 +0100
Subject: [PATCH] git-rebase--interactive: replace echo by printf
to avoid shell dependent behavior.
When your system shell (/bin/sh) is a dash control sequences in strings get interpreted by the echo command. A commit message which ends with the string '\n' may result in a garbage line in the todo list of an interactive rebase which causes the rebase to fail.
To reproduce the behavior (with dash as /bin/sh):
mkdir test && cd test && git init echo 1 >foo && git add foo git commit -m"this commit message ends with '\n'" echo 2 >foo && git commit -a --fixup HEAD git rebase -i --autosquash --root
Now the editor opens with garbage in line 3 which has to be removed or the rebase fails.
--- git-rebase--interactive.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
git-rebase--interactive.sh +1 -1
@@ -749,7 +749,7 @@ rearrange_squash () { ;; esac done - echo "$sha1 $action $prefix $rest" + printf "%s %s %s %s\n" "$sha1" "$action" "$prefix" "$rest" # if it's a single word, try to resolve to a full sha1 and # emit a second copy. This allows us to match on both message # and on sha1 prefix
Signature · 1 line
1.9.0
Jonathan Nieder
Reviewed-byHi,
Quoted 10 lines · Uwe Storbeck
To reproduce the behavior (with dash as
/bin/sh):mkdir test && cd test && git init echo 1 >foo && git add foo git commit -m"this commit message ends with '\n'" echo 2 >foo && git commit -a --fixup HEAD git rebase -i --autosquash --rootNow the editor opens with garbage in line 3 which has to be removed or the rebase fails.
Would it make sense to add this as a test to e.g.t/t3404-rebase-interactive.sh?
Quoted 1 line
The attached one-line patch fixes the bug.
May we have your sign-off? (See Documentation/SubmittingPatches section "Sign your work" for what this means.
Looks obviously correct, so for what it's worth,
Thanks,
Jonathan
Quoted 2 lines · Jonathan Nieder
May we have your sign-off? (See
Documentation/SubmittingPatchessection "Sign your work" for what this means.
I could have found that myself .. thanks! I'll try to follow it now. :)
I'll resend the patch. Hopefully I'll do it right.
Quoted 2 lines
Would it make sense to add this as a test to e.g.
t/t3404-rebase-interactive.sh?
It's a rather special case, so I'm not sure if it's worth it. I'll send a patch which adds a test for it. The test works for me, but as I don't understand the test mechanisms already good enough a few questions:
- Is it correct to call the fake editor with an empty variable FAKE_LINES when you want it to not change the todo list of a rebase -i and use it as it is (the work is already done by the autosquash option)? I can achieve the same with EDITOR=true. What's the preferred way? Is there an advantage to use the fake editor also in this case?
- The tests in t3404-rebase-interactive.sh use their variables a bit differently, some just set the variables, some export the variables and some use a subshell to encapsulate them. Also some of the tests reset their rebase state so that subsequent tests, which also use rebase, do not fail when the rebase fails. Other tests don't do that.
What's the expected resp. recommended behavior?
While trying to understand the test mechanisms I stumbled over two other problematic uses of echo. These although only affect the test output, not git itself.
Uwe
Quoted 6 lines · Uwe Storbeck
When your system shell (
/bin/sh) is a dash control sequences in strings get interpreted by the echo command. A commit message which ends with the string '\n' may result in a garbage line in the todo list of an interactive rebase which causes the rebase to fail.
Quoted 7 lines
To reproduce the behavior (with dash as
/bin/sh):mkdir test && cd test && git init echo 1 >foo && git add foo git commit -m"this commit message ends with '\n'" echo 2 >foo && git commit -a --fixup HEAD git rebase -i --autosquash --root
Hmph. We ran into this before and fixed all of the sites (e.g., d1c3b10 and 938791c). This one appears to have been added a few months later (by 68d5d03).
Quoted 2 lines
Maybe there are more places where it would be more robust to use printf instead of echo.
FWIW, I just looked through the other uses of "echo" in git-rebase*.sh, and I think this is the only problematic case.
git-rebase--interactive.sh rearrange_squash () { · line 752
Quoted 2 lines
git-rebase--interactive.sh +1
-1- echo "$sha1 $action $prefix $rest" + printf "%s %s %s %s\n" "$sha1" "$action" "$prefix" "$rest"
Looks obviously correct. The echo just below here does not need the same treatment, as "$rest" is the problematic bit ("$prefix" is always "fixup" or "squash").
-Peff
Quoted 10 lines · Jeff King
Hmph. We ran into this before and fixed all of the sites (e.g., d1c3b10 and 938791c). This one appears to have been added a few months later (by 68d5d03).
Quoted 2 lines
Maybe there are more places where it would be more robust to use printf instead of echo.
FWIW, I just looked through the other uses of "echo" in git-rebase*.sh, and I think this is the only problematic case.
Quoted 6 lines
git-rebase--interactive.sh rearrange_squash () { · line 752
Quoted 2 lines
git-rebase--interactive.sh +1
-1- echo "$sha1 $action $prefix $rest" + printf "%s %s %s %s\n" "$sha1" "$action" "$prefix" "$rest"Looks obviously correct. The echo just below here does not need the same treatment, as "$rest" is the problematic bit ("$prefix" is always "fixup" or "squash").
I'd not rationalize this away by deep analysis. Copy&paste is a thing, so to just use printf whenever _any_ seriously variable strings (source not immediately the shell script itself, perhaps even _any_ nonconstant strings) are involved keeps people from introducing bugs by following apparent practice.
Signature · 1 line
David Kastrup