Fix an off-by-one bug in the untracked cache code

Frankly, I am surprised it took me this long to discover this bug, as I am running with the untracked cache for ages now.

But as of recent, I am running more and more long-running commands, mostly rebases to keep what I call "ever-green" branches up to date, where Git for Windows' patch thicket is rebased onto core Git's four integration branches, and when there are no merge conflicts to stop the rebase early, I got this really obscure crash (no error message, no nothing to indicate what it was).

It cost me two full work days to chase this one down.

Johannes Schindelin (1):
untracked cache: fix off-by-one

 dir.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

base-commit: e35b8cb8e212e3557efc565157ceb5cbaaf0d87f
Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-178%2Fdscho%2Funtracked-cache-off-by-one-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-178/dscho/untracked-cache-off-by-one-v1
Pull-Request: https://github.com/gitgitgadget/git/pull/178

gitgitgadget

From: Johannes Schindelin <[email protected]>

In f9e6c649589e (untracked cache: load from UNTR index extension, 2015-03-08), code was added to read back the untracked cache from an index extension.

Probably in the endeavor to avoid the calloc() implied by FLEX_ALLOC_STR() (it is hard to know why exactly, the commit message of that commit is a bit parsimonious with information), it calls malloc() manually and then memcpy()s the bits and pieces into place.

It allocates the size of struct untracked_cache_dir plus the string length of the untracked file name, then copies the information in two
steps: first the fixed-size metadata, then the name. And here lies the
rub: it includes the trailing NUL byte in the name.

If FLEX_ARRAY is defined as 0, this results in a buffer overrun.

To fix this, let's just add 1, for the trailing NUL byte. Technically, this overallocates on platforms where FLEX_ARRAY is 1, but it should not matter much in reality, as malloc() usually overallocates anyway, unless the size to allocate aligns exactly with some internal chunk size (see below for more on that).

The real strange thing is that neither valgrind nor DrMemory catches this bug. In this developer's tests, a memcpy() (but not a memset()!) could write up to 4 bytes after the allocated memory range before valgrind would start reporting an issue.

However, when running Git built with nedmalloc as allocator, under rare conditions (and inconsistently at that), this bug triggered an abort() because nedmalloc rounds up the size to be malloc()ed to a multiple of a certain chunk size, then adds a few bytes to be used for storing some internal state. If there is no rounding up to do (because the size is already a multiple of that chunk size), and if the buffer is overrun as in the code patched in this commit, the internal state is corrupted.

The scenario that triggered this here bug fix entailed a git.git checkout with an extra copy of the source code in an untracked subdirectory, meaning that there was an untracked subdirectory called "thunderbird-patch-inline" whose name's length is exactly 24 bytes, which, added to the size of above-mentioned struct untracked_cache_dir that weighs in with 104 bytes on a 64-bit system, amounts to 128, aligning perfectly with nedmalloc's chunk size.

As there is no obvious way to trigger this bug reliably, on all platforms supported by Git, and as the bug is obvious enough, this patch comes without a regression test.

---
 dir.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index b2cabadf25..f5293a6536 100644
@@ -2760,7 +2760,7 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
 	next = data + len + 1;
 	if (next > rd->end)
 		return -1;
-	*untracked_ = untracked = xmalloc(st_add(sizeof(*untracked), len));
+	*untracked_ = untracked = xmalloc(st_add3(sizeof(*untracked), len, 1));
 	memcpy(untracked, &ud, sizeof(ud));
 	memcpy(untracked->name, data, len + 1);
 	data = next;

gitgitgadget

You have a talent for understatement. :)

Yeah, every struct on a platform where FLEX_ARRAY is 1 ends up over-allocated by 1 byte. We could account for that in all of our flex allocations, but I don't it affects enough platforms to be worth the bother.

I couldn't get it to trigger with ASan, either. I assume it's due to alignment (i.e., those tools are stuck poisoning at the end of a page, but the start of the struct needs to be page-aligned). But that would imply that we could trigger it with different path lengths, which I wasn't able to do. So I dunno.

Right, that makes sense that the length would impact it there.

Makes sense. This code path should be well-covered by the existing tests anyway, so even if we could get those tools to trigger, I don't think there would be much point in adding a new test.

  len = strlen((const char *)data);

	return error("malformed untracked cache extension");

It was an explicit design decision to declare that the overallocation was a non-problem back when we invented FLEX_ARRAY macro. We could probably have instead decided to pass number of bytes to be stored in the flex-array (including NUL if that is a character array) then subtract the value of FLEX_ARRAY if FLEX_ARRAY were limited to only 0 or 1, but that was not workable as it is the most natural to define FLEX_ARRAY to an empty string, i.e. making an array field decl to "type field[];", for the more recent compilers.

Yeah, it is already super that this obscure bug was spotted in the first place. It is true that this may regress without a test, but I do not think it is reasonable to expect that it is possible to write a reliable crash-detecting test for this one.

True, true. I wonder if folks intereseted in libFuzzer can chime in with something useful here, but that is totally independent.

[+cc Duy as the master of all things untracked-cache]

Just so we don't forget about it, I wrote this fix up as a patch. And in fact it led to a few other cleanups. I think the first one is definitely worth doing now, even if there are other similar cases lurking in the rest of the index code.

The other two are optional, though I think they are worth it (and not too hard to verify that they are doing the right thing).

These are on top of js/untracked-cache-allocfix (though they could easily be ported to a separate topic if we want).

[1/3]: untracked-cache: be defensive about missing NULs in index [2/3]: untracked-cache: simplify parsing by dropping "next" [3/3]: untracked-cache: simplify parsing by dropping "len"

 dir.c | 41 ++++++++++++++++++-----------------------
 1 file changed, 18 insertions(+), 23 deletions(-)

-Peff

The on-disk format for the untracked-cache extension contains NUL-terminated filenames. We parse these from the mmap'd file using string functions like strlen(). This works fine in the normal case, but if we see a malformed or corrupted index, we might read off the end of our mmap.

Instead, let's use memchr() to find the trailing NUL within the bytes we know are available, and return an error if it's missing.

Note that we can further simplify by folding another range check into our conditional. After we find the end of the string, we set "next" to the byte after the string and treat it as an error if there are no such bytes left. That saves us from having to do a range check at the beginning of each subsequent string (and works because there is always data after each string). We can do both range checks together by checking "!eos" (we didn't find a NUL) and "eos == end" (it was on the last available byte, meaning there's nothing after). This replaces the existing "next > end" checks.

Note also that the decode_varint() calls have a similar problem (we don't even pass them "end"; they just keep parsing). These are probably OK in practice since varints have a finite length (we stop parsing when we'd overflow a uintmax_t), so the worst case is that we'd overflow into reading the trailing bytes of the index.

---
 dir.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)

index f5293a6536..7b0513c476 100644
@@ -2733,6 +2733,7 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
 {
 	struct untracked_cache_dir ud, *untracked;
 	const unsigned char *next, *data = rd->data, *end = rd->end;
+	const unsigned char *eos;
 	unsigned int value;
 	int i, len;
 

@@ -2756,21 +2757,24 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
 	ALLOC_ARRAY(ud.dirs, ud.dirs_nr);
 	data = next;
 
-	len = strlen((const char *)data);
-	next = data + len + 1;
-	if (next > rd->end)
+	eos = memchr(data, '\0', end - data);
+	if (!eos || eos == end)
 		return -1;
+	len = eos - data;
+	next = eos + 1;
+
 	*untracked_ = untracked = xmalloc(st_add3(sizeof(*untracked), len, 1));
 	memcpy(untracked, &ud, sizeof(ud));
 	memcpy(untracked->name, data, len + 1);
 	data = next;
 
 	for (i = 0; i < untracked->untracked_nr; i++) {
-		len = strlen((const char *)data);
-		next = data + len + 1;
-		if (next > rd->end)
+		eos = memchr(data, '\0', end - data);
+		if (!eos || eos == end)
 			return -1;
-		untracked->untracked[i] = xstrdup((const char*)data);
+		len = eos - data;
+		next = eos + 1;
+		untracked->untracked[i] = xmemdupz(data, len);
 		data = next;
 	}
 

2.21.0.1092.g8b0302e9c4

When we parse an on-disk untracked cache, we have two pointers, "data" and "next". As we parse, we point "next" to the end of an element, and then later update "data" to match.

But we actually don't need two pointers. Each parsing step can just update "data" directly from other variables we hold (and we don't have to worry about bailing in an intermediate state, since any parsing failure causes us to immediately discard "data" and return).

---
 dir.c | 20 +++++++-------------
 1 file changed, 7 insertions(+), 13 deletions(-)

index 7b0513c476..17865f44df 100644
@@ -2732,50 +2732,44 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
 			struct read_data *rd)
 {
 	struct untracked_cache_dir ud, *untracked;
-	const unsigned char *next, *data = rd->data, *end = rd->end;
+	const unsigned char *data = rd->data, *end = rd->end;
 	const unsigned char *eos;
 	unsigned int value;
 	int i, len;
 
 	memset(&ud, 0, sizeof(ud));
 
-	next = data;
-	value = decode_varint(&next);
-	if (next > end)
+	value = decode_varint(&data);
+	if (data > end)
 		return -1;
 	ud.recurse	   = 1;
 	ud.untracked_alloc = value;
 	ud.untracked_nr	   = value;
 	if (ud.untracked_nr)
 		ALLOC_ARRAY(ud.untracked, ud.untracked_nr);
-	data = next;
 
-	next = data;
-	ud.dirs_alloc = ud.dirs_nr = decode_varint(&next);
-	if (next > end)
+	ud.dirs_alloc = ud.dirs_nr = decode_varint(&data);
+	if (data > end)
 		return -1;
 	ALLOC_ARRAY(ud.dirs, ud.dirs_nr);
-	data = next;
 
 	eos = memchr(data, '\0', end - data);
 	if (!eos || eos == end)
 		return -1;
 	len = eos - data;
-	next = eos + 1;
 
 	*untracked_ = untracked = xmalloc(st_add3(sizeof(*untracked), len, 1));
 	memcpy(untracked, &ud, sizeof(ud));
 	memcpy(untracked->name, data, len + 1);
-	data = next;
+	data = eos + 1;
 
 	for (i = 0; i < untracked->untracked_nr; i++) {
 		eos = memchr(data, '\0', end - data);
 		if (!eos || eos == end)
 			return -1;
 		len = eos - data;
-		next = eos + 1;
 		untracked->untracked[i] = xmemdupz(data, len);
-		data = next;
+		data = eos + 1;
 	}
 
 	rd->ucd[rd->index++] = untracked;

2.21.0.1092.g8b0302e9c4

The code which parses untracked-cache extensions from disk keeps a "len" variable, which is the size of the string we are parsing. But since we now have an "end of string" variable, we can just use that to get the length when we need it. This eliminates the need to keep "len" up to date (and removes the possibility of any errors where "len" and "eos" get out of sync).

As a bonus, it means we are not storing a string length in an "int", which is a potential source of overflows (though in this case it seems fairly unlikely for that to cause any memory problems).

---
 dir.c | 13 +++++--------
 1 file changed, 5 insertions(+), 8 deletions(-)

index 17865f44df..60438b2cdc 100644
@@ -2735,7 +2735,7 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
 	const unsigned char *data = rd->data, *end = rd->end;
 	const unsigned char *eos;
 	unsigned int value;
-	int i, len;
+	int i;
 
 	memset(&ud, 0, sizeof(ud));
 

@@ -2756,28 +2756,25 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
 	eos = memchr(data, '\0', end - data);
 	if (!eos || eos == end)
 		return -1;
-	len = eos - data;
 
-	*untracked_ = untracked = xmalloc(st_add3(sizeof(*untracked), len, 1));
+	*untracked_ = untracked = xmalloc(st_add3(sizeof(*untracked), eos - data, 1));
 	memcpy(untracked, &ud, sizeof(ud));
-	memcpy(untracked->name, data, len + 1);
+	memcpy(untracked->name, data, eos - data + 1);
 	data = eos + 1;
 
 	for (i = 0; i < untracked->untracked_nr; i++) {
 		eos = memchr(data, '\0', end - data);
 		if (!eos || eos == end)
 			return -1;
-		len = eos - data;
-		untracked->untracked[i] = xmemdupz(data, len);
+		untracked->untracked[i] = xmemdupz(data, eos - data);
 		data = eos + 1;
 	}
 
 	rd->ucd[rd->index++] = untracked;
 	rd->data = data;
 
 	for (i = 0; i < untracked->dirs_nr; i++) {
-		len = read_one_dir(untracked->dirs + i, rd);
-		if (len < 0)
+		if (read_one_dir(untracked->dirs + i, rd) < 0)
 			return -1;
 	}
 	return 0;

2.21.0.1092.g8b0302e9c4

I also wondered if we could just accept the cost of calloc() here and use FLEX_ALLOC to simplify things. That resulted in the patch below, but I didn't include it with the initial 3, because I think it's too subtle/gross for my tastes.

-- >8 --
Subject: untracked-cache: use FLEX_ALLOC to create internal structs

The untracked_cache_dir struct has a FLEX_ARRAY in it. Let's use FLEX_ALLOC_MEM to allocate it, which saves us having to compute the length ourselves.

In theory this could be slightly slower, since the FLEX_ALLOC macros use calloc (and we just memcpy over most of the contents anyway). But in practice this distinction is not generally measurable.

Note that because we then fill in the pre-flex elements of the struct using a memcpy, we need to take care to use the exact size of that space and _not_ "sizeof(ud)", since the latter may include padding (or even an extra byte on systems where FLEX_ARRAY is 1).

---
If we wanted to go this route, I think it would make sense to provide a FLEX_ALLOC macro that takes a "template" set of bytes as a ptr/len pair, and writes it before we fill in the flex portion.

Then we could do something like:

  FLEX_ALLOC_COPY(untracked, &ud, sizeof(ud), name, data, eos - data);

If this is the only such case, it's probably not worth it (I didn't really look around for more, though).

 dir.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

index 60438b2cdc..7cd4eec198 100644
@@ -2757,9 +2757,9 @@ static int read_one_dir(struct untracked_cache_dir **untracked_,
 	if (!eos || eos == end)
 		return -1;
 
-	*untracked_ = untracked = xmalloc(st_add3(sizeof(*untracked), eos - data, 1));
-	memcpy(untracked, &ud, sizeof(ud));
-	memcpy(untracked->name, data, eos - data + 1);
+	FLEX_ALLOC_MEM(untracked, name, data, eos - data);
+	memcpy(untracked, &ud, offsetof(struct untracked_cache_dir, name));
+	*untracked_ = untracked;
 	data = eos + 1;
 
 	for (i = 0; i < untracked->untracked_nr; i++) {

2.21.0.1092.g8b0302e9c4

It's probably ok. If I remember correctly, reading UNTR extension (on a huge repo) took the longest time after Ben added support for reading the index with multiple threads. So performance is a concern, but I don't think calloc() would be the problem. malloc() itself without the memory pool could probably slow down more when we have lots and lots of directories.

Duy

;-)

My first reaction was "you can do so now you have introduced eos--why didn't you do that in the previous step?", but losing 'next' from the varint parsing step would certainly have been possible even before that change. So I agree that it makes much more sense to do this step separately from the previous one.

The code after the patch certainly reads easier and simpler.

Thanks.

I think if we do the FLEX_ALLOC_COPY() thing I mentioned that it would probably _not_ use calloc() there, since we'd know we were copying in the content from elsewhere. So that concern would go away either way. :)

(But I'm still skeptical that FLEX_ALLOC_COPY() is worth it unless we can find at least one other caller).

-Peff