Skip to contentSkip to conversation

Bug: Segfault when doing "git diff"

PATCH In review Mathias L. Baumann 4 participants 7 messages Jun 15, 2016

lore
Share & tools
lore
#0 lore raw

Hello dear git devs,

I just stumbled upon a segfault when doing just "git diff" in my repo.

I managed to create a minimal repo setup where the bug is reproducable.

The problem seems to be a mix of having an untracked submodule and having set an alternates file for one submodule.

Attached you'll find my setup that will reproduce the problem. Simply run 'git diff' in bugtest1.

In case the attachment is refused, I also uploaded it here:

http://supraverse.net/bugdemo.tar.gz

cheers,

--Marenz

#1 lore raw
Quoted 10 lines · Mathias L. Baumann

Hello dear git devs,

I just stumbled upon a segfault when doing just "git diff" in my repo.

I managed to create a minimal repo setup where the bug is reproducable.

The problem seems to be a mix of having an untracked submodule and having set an alternates file for one submodule.

Attached you'll find my setup that will reproduce the problem. Simply run 'git diff' in bugtest1.

Quoted 7 lines

In case the attachment is refused, I also uploaded it here:

http://supraverse.net/bugdemo.tar.gz

cheers,

--Marenz

Hello Marenz,

I have just tried to reproduce segfault with the provided archive:

[del@del-debian bugtest1 (master)]$ git diff

submodules/bugtest2 +1 -1
@@ -1 +1 @@
-Subproject commit cd0b9ee2946d2df3626943347332a4d86f93b126
+Subproject commit cd0b9ee2946d2df3626943347332a4d86f93b126-dirty

No segfault occured. I am using

git version 2.6.2.308.g3b8f10c

Could you please specify which version of git you are using and also try to reproduce it with latest 2.6.2?

--
Victor

#2 lore raw

I was using the latest git version 2.6.2 already.
I suspect it is due to a .gitconfig. This is what is probably required:

➜ ~ cat .gitconfig
[diff]
submodule = log

Quoted 43 lines · Victor Leschuk
Quoted 12 lines · Mathias L. Baumann

Hello dear git devs,

I just stumbled upon a segfault when doing just "git diff" in my repo.

I managed to create a minimal repo setup where the bug is reproducable.

The problem seems to be a mix of having an untracked submodule and having set an alternates file for one submodule.

Attached you'll find my setup that will reproduce the problem. Simply run 'git diff' in bugtest1.

Quoted 7 lines

In case the attachment is refused, I also uploaded it here:

http://supraverse.net/bugdemo.tar.gz

cheers,

--Marenz

Hello Marenz,

I have just tried to reproduce segfault with the provided archive:

[del@del-debian bugtest1 (master)]$ git diff

submodules/bugtest2 +1 -1
@@ -1 +1 @@
-Subproject commit cd0b9ee2946d2df3626943347332a4d86f93b126
+Subproject commit cd0b9ee2946d2df3626943347332a4d86f93b126-dirty

No segfault occured. I am using

git version 2.6.2.308.g3b8f10c

Could you please specify which version of git you are using and also try to reproduce it with latest 2.6.2?

--
Victor

#3 lore raw

PATCH 1/1 add_submodule_odb: initialize alt_odb list earlier

Quoted 6 lines · Mathias L. Baumann

I was using the latest git version 2.6.2 already.
I suspect it is due to a .gitconfig. This is what is probably required:

➜ ~ cat .gitconfig
[diff]
submodule = log

Yeah, I can reproduce it easily with that. Thanks for providing the repository. It takes a rather convoluted set of conditions to trigger the bug. :)

Here's the fix:

-- >8 --
Subject: add_submodule_odb: initialize alt_odb list earlier

The add_submodule_odb function tries to add a submodule's object store as an "alternate". It needs the existing list to be initialized (from the objects/info/alternates file) for two reasons:

1. We look for duplicates with the existing alternate stores, but obviously this doesn't work if we haven't loaded any yet.

2. We link our new entry into the list by prepending it to alt_odb_list. But we do _not_ modify alt_odb_tail. This variable starts as NULL, and is a signal to the alt_odb code that the list has not yet been initialized.

We then call read_info_alternates on the submodule (to recursively load its alternates), which will try to append to that tail, assuming it has been initialized. This causes us to segfault if it is NULL.

This rarely comes up in practice, because we will have initialized the alt_odb any time we do an object lookup. So you can trigger this only when:

- you try to access a submodule (e.g., a diff with diff.submodule=log)

- the access happens before any other object has been accessed (e.g., because the diff is between the working tree and the index)

- the submodule contains an alternates file (so we try to add an entry to the NULL alt_odb_tail)

To fix this, we just need to call prepare_alt_odb at the start of the function (and if we have already initialized, it is a noop).

Note that we can remove the prepare_alt_odb call from the end. It is guaranteed to be a noop, since we will have called it earlier.

---
 submodule.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
submodule.c +1 -1
index 5879cfb..88af54c 100644
@@ -130,6 +130,7 @@ static int add_submodule_odb(const char *path)
 		goto done;
 	}
 	/* avoid adding it twice */
+	prepare_alt_odb();
 	for (alt_odb = alt_odb_list; alt_odb; alt_odb = alt_odb->next)
 		if (alt_odb->name - alt_odb->base == objects_directory.len &&
 				!strncmp(alt_odb->base, objects_directory.buf,
@@ -148,7 +149,6 @@ static int add_submodule_odb(const char *path)
 
 	/* add possible alternates from the submodule */
 	read_info_alternates(objects_directory.buf, 0);
-	prepare_alt_odb();
 done:
 	strbuf_release(&objects_directory);
 	return ret;
Signature · 1 line
2.6.2.572.g6ed22dd
#4 lore raw

Re: [PATCH] add_submodule_odb: initialize alt_odb list earlier

Quoted 43 lines · Jeff King

Yeah, I can reproduce it easily with that. Thanks for providing the repository. It takes a rather convoluted set of conditions to trigger the bug. :)

Here's the fix:

-- >8 --
Subject: add_submodule_odb: initialize alt_odb list earlier

The add_submodule_odb function tries to add a submodule's object store as an "alternate". It needs the existing list to be initialized (from the objects/info/alternates file) for two reasons:

1. We look for duplicates with the existing alternate stores, but obviously this doesn't work if we haven't loaded any yet.

2. We link our new entry into the list by prepending it to alt_odb_list. But we do _not_ modify alt_odb_tail. This variable starts as NULL, and is a signal to the alt_odb code that the list has not yet been initialized.

We then call read_info_alternates on the submodule (to recursively load its alternates), which will try to append to that tail, assuming it has been initialized. This causes us to segfault if it is NULL.

This rarely comes up in practice, because we will have initialized the alt_odb any time we do an object lookup. So you can trigger this only when:

- you try to access a submodule (e.g., a diff with diff.submodule=log)

- the access happens before any other object has been accessed (e.g., because the diff is between the working tree and the index)

- the submodule contains an alternates file (so we try to add an entry to the NULL alt_odb_tail)

Quoted 7 lines

To fix this, we just need to call prepare_alt_odb at the start of the function (and if we have already initialized, it is a noop).

Note that we can remove the prepare_alt_odb call from the end. It is guaranteed to be a noop, since we will have called it earlier.

Thanks for a quick and detailed diagnosis and a fix.

The removal is correct, but even without this fix, the order of calls in the original should have screamed "bug" loudly at us, I think. We shouldn't be reading data from alternates file without first preparing the place we read data into.

Quoted 25 lines
---
 submodule.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
submodule.c +1 -1
index 5879cfb..88af54c 100644
@@ -130,6 +130,7 @@ static int add_submodule_odb(const char *path)
 		goto done;
 	}
 	/* avoid adding it twice */
+	prepare_alt_odb();
 	for (alt_odb = alt_odb_list; alt_odb; alt_odb = alt_odb->next)
 		if (alt_odb->name - alt_odb->base == objects_directory.len &&
 				!strncmp(alt_odb->base, objects_directory.buf,
@@ -148,7 +149,6 @@ static int add_submodule_odb(const char *path)
 
 	/* add possible alternates from the submodule */
 	read_info_alternates(objects_directory.buf, 0);
-	prepare_alt_odb();
 done:
 	strbuf_release(&objects_directory);
 	return ret;
#5 lore raw

Re: [PATCH] add_submodule_odb: initialize alt_odb list earlier

Quoted 10 lines · Junio C Hamano
Quoted 3 lines

Note that we can remove the prepare_alt_odb call from the end. It is guaranteed to be a noop, since we will have called it earlier.

Thanks for a quick and detailed diagnosis and a fix.

The removal is correct, but even without this fix, the order of calls in the original should have screamed "bug" loudly at us, I think. We shouldn't be reading data from alternates file without first preparing the place we read data into.

Yeah, I agree. I spent a long time trying to figure out if that prepare_alt_odb was actually doing something useful (like if it was needed to somehow "cement" the new alt into place).

But I don't think it was.

In the majority of cases, it was a noop (we had already prepared when we looked up the first object). But for other cases...

- if read_info_alternates actually did something, we segfaulted (i.e., this bug)

- otherwise, we would prepare on _top_ of what we just added to the list, which was probably buggy (I didn't dig far enough to see if prepare_alt_odb() would overwrite what we just added to the list).

So some pretty dark corners of the code. :)

-Peff

#6 lore raw
In reply to Mathias L. Baumann
Quoted 6 lines · Mathias L. Baumann

I was using the latest git version 2.6.2 already.
I suspect it is due to a .gitconfig. This is what is probably required:

➜ ~ cat .gitconfig
[diff]
submodule = log

Yep, that did the trick.

The segfault is from

sha1_file.c:

/* add the alternate entry */
*alt_odb_tail = ent; /* <===== alt_obd_tail is NULL here */
alt_odb_tail = &(ent->next);
ent->next = NULL;

Will try to take a closer look at it.

--
Victor