Bug encountered while comitting
Hello list,
I don't know if this is the right place to report a bug, but I'll just try and see what comes back.
I am trying to build a Wiki [1] using PHP, a hacked version of Markdown, and git for content tracking. I use the git core plumbing to do the history work.
The PHP script created directories under .git/objects which were only writable by www-data. There were other directories which were owned by user mk and group www-data, and they were group writable.
So, I had write access to only a part of the .git directory.
When I tried to commit, I got a message saying "Unable to write sha1 filename".
The result was, that only part of the commit was recorded and that I experienced repository corruption. refs/heads/master pointed to a non-existant object.
The expected behavior would have been an error message telling me I had insufficient write privileges and surely no repository corruption.
Thanks,
Matthias
[1]: http://spinlock.ch/cgi-bin/gitweb.pl?p=swisdk2/bugs.git;a=tree
(See Wiki_ctrl.php for source)
Signature · 1 line
:wq
Quoted 10 lines · Matthias Kestenholz
Hello list,
I don't know if this is the right place to report a bug, but I'll just try and see what comes back.
I am trying to build a Wiki [1] using PHP, a hacked version of Markdown, and git for content tracking. I use the git core plumbing to do the history work.
The PHP script created directories under .git/objects
Ouch... You're not really supposed to do that. The proper thing to do is to do things in the working tree and commit them to git later.
Quoted 6 lines
which were
only writable by www-data. There were other directories which were owned by user mk and group www-data, and they were group writable.So, I had write access to only a part of the .git directory.
Unless you're using the git tools (or things hooking in to the git core C functions somehow), don't touch the .git directory.
(this merits an exclamation marks, so brace yourselves) !
Quoted 3 lines
When I tried to commit, I got a message saying "Unable to write sha1 filename".
What file were you trying to write?
Quoted 4 lines
The result was, that only part of the commit was recorded and that I experienced repository corruption. refs/heads/master pointed to a non-existant object.
Did you use git tools to update .git/refs/heads/master ?
Quoted 4 lines
The expected behavior would have been an error message telling me I had insufficient write privileges and surely no repository corruption.
Didn't you get the strerror(3) message from that? If so, I'd consider it a bug.
Signature · 3 lines
Andreas Ericsson [email protected] OP5 AB www.op5.se Tel: +46 8-230225 Fax: +46 8-230231
Quoted 8 lines · * Andreas Ericsson ([email protected]) wrote:
Quoted 2 lines · Matthias Kestenholz
The PHP script created directories under .git/objects
Ouch... You're not really supposed to do that. The proper thing to do is to do things in the working tree and commit them to git later.
I think I did not express myself clearly. The PHP Script executes git commands which in turn create or modify the mentioned directories. I do not create them myself by hand.
Quoted 6 lines
Quoted 3 lines
When I tried to commit, I got a message saying "Unable to write sha1 filename".
What file were you trying to write?
Some file I was updating (in this case, a file holding some wiki content)
Quoted 7 lines
Quoted 4 lines
The result was, that only part of the commit was recorded and that I experienced repository corruption. refs/heads/master pointed to a non-existant object.
Did you use git tools to update .git/refs/heads/master ?
Yes.
Quoted 8 lines
Quoted 5 lines
The expected behavior would have been an error message telling me I had insufficient write privileges and surely no repository corruption.
Didn't you get the strerror(3) message from that? If so, I'd consider it a bug.
As I (only now) wrote in the other email, I got the "Permission denied" message, and that was a sufficient hint what was wrong.
Signature · 1 line
:wq
Quoted 5 lines · Matthias Kestenholz
The PHP script created directories under .git/objects which were only writable by www-data. There were other directories which were owned by user mk and group www-data, and they were group writable.
So, I had write access to only a part of the .git directory.
core.sharedrepository perhaps?
While it probably is not a good idea to have you in www-data, it appears that is essentially you will end up doing, because PHP scripts that may _create_ new directories better not to have privilege to give newly created directories away to you (busting your quota), so they will be owned by www-data.www-data and for you to be able to write into it you either need to be www-data user or in www-data group, with core.sharedrepostiory set.
Quoted 10 lines · * Junio C Hamano ([email protected]) wrote:
Quoted 5 lines · Matthias Kestenholz
The PHP script created directories under .git/objects which were only writable by www-data. There were other directories which were owned by user mk and group www-data, and they were group writable.
So, I had write access to only a part of the .git directory.
core.sharedrepository perhaps?
Quoted 8 lines
While it probably is not a good idea to have you in www-data, it appears that is essentially you will end up doing, because PHP scripts that may _create_ new directories better not to have privilege to give newly created directories away to you (busting your quota), so they will be owned by www-data.www-data and for you to be able to write into it you either need to be www-data user or in www-data group, with core.sharedrepostiory set.
Thanks for your answer; I did not know about this option (I should probably re-read all the docs).
Anyway, I think git should never corrupt a repository, even if it does not have the write permissions it needs.
The following commands were sufficient to create a corrupt repository with git (v1.2.4-1, debian package):
$ git-init-db $ echo test > file $ git-update-index --add file ; git commit -m 'message'
repeat (f.e. 10 times):
$ echo test >> file $ git-update-index file ; git commit -m 'message'
$ sudo chown root .git/objects/*
repeat the modification and commit commands until you get a message similar to the following:
unable to write sha1 filename .git/objects/90/b33..: Permission denied
fatal: 90b33... is not a valid 'tree' object
unable to write sha1 filename .git/objects/ba/fe4..: Permission denied
error: file: failed to insert into database
fatal: Unable to process file file
etc...
The result of this all is: refs/heads/master might now point to a non-existant commit object. Every git command now errors out with:
fatal: bad tree object HEAD
and git-log shows no output (probably since it does not find a commit to begin with)
git-commit should abort as soon as it encounters an error and not update HEAD.
Thanks,
Matthias
(Note: To find the last valid commit object, I could just scan the objects directory for the recently modified files and write the sha1 value to refs/heads/master, so I had no data loss.)
Signature · 1 line
:wq
Quoted 12 lines · Matthias Kestenholz
$ sudo chown root .git/objects/*repeat the modification and commit commands until you get a message similar to the following:
unable to write sha1 filename .git/objects/90/b33..: Permission denied
fatal: 90b33... is not a valid 'tree' object
unable to write sha1 filename .git/objects/ba/fe4..: Permission denied
error: file: failed to insert into database
fatal: Unable to process file file
etc...
Quoted 7 lines
The result of this all is: refs/heads/master might now point to a non-existant commit object. Every git command now errors out with:
fatal: bad tree object HEAD
and git-log shows no output (probably since it does not find a commit to begin with)
You are right. commit-tree does not seem to check if it successfully wrote the commit object. How about this?
-- >8 --
commit-tree.c +6 -3
@@ -125,7 +125,10 @@ int main(int argc, char **argv)while (fgets (comment,sizeof (comment), stdin) !=NULL )add_buffer (&buffer, &size,"%s" , comment); -write_sha1_file (buffer, size,"commit" , commit_sha1); -printf ("%s\n" ,sha1_to_hex (commit_sha1)); -return 0 ; +if (!write_sha1_file (buffer, size,"commit" , commit_sha1)) { +printf ("%s\n" ,sha1_to_hex (commit_sha1)); +return 0 ; + } +else +return 1 ; }
Quoted 12 lines · * Junio C Hamano ([email protected]) wrote:
You are right. commit-tree does not seem to check if it successfully wrote the commit object. How about this?
-- >8 --
commit-tree.c +0
-0@@ -125,7 +125,10 @@ int main(int argc, char **argv)while (fgets (comment,sizeof (comment), stdin) !=NULL )add_buffer (&buffer, &size,"%s" , comment);
commit-tree.c int main(int argc, char **argv) · line 134
Quoted 10 lines
commit-tree.c +6
-3-write_sha1_file (buffer, size,"commit" , commit_sha1); -printf ("%s\n" ,sha1_to_hex (commit_sha1)); -return 0 ; +if (!write_sha1_file (buffer, size,"commit" , commit_sha1)) { +printf ("%s\n" ,sha1_to_hex (commit_sha1)); +return 0 ; + } +else +return 1 ; }
This patch fixes the problem I was encountering (git-update-ref is not executed anymore).
Thank you all for this really great tool!
Signature · 1 line
:wq