[PATCH] dir: support platforms that require aligned reads

[PATCH] dir: support platforms that require aligned reads

[René Scharfe]

The untracked cache is stored on disk by concatenating its memory
structures without any padding.  Consequently some of the structs are
not aligned at a particular boundary when the whole extension is read
back in one go.  That's only OK on platforms without strict alignment
requirements, or for byte-aligned data like strings or hash values.

Decode struct ondisk_untracked_cache carefully from the extension
blob by using explicit pointer arithmetic with offsets, avoiding
alignment issues.  Use char pointers for passing stat_data objects to
stat_data_from_disk(), and use memcpy(3) in that function to  get the
contents into a properly aligned struct, then perform the byte-order
adjustment in place there.

Found with Clang's UBSan.

Signed-off-by: Rene Scharfe <l.s.r@web.de>
---
Side note: The OS name is not enough for determining the layout of
struct ondisk_untracked_cache.  Different platforms can have
different int sizes and padding.  Adding the machine type could
help, but that would be a breaking change.  At that point we would
be better off defining a machine-independent format, no?

 dir.c | 50 +++++++++++++++++++++++++++-----------------------
 1 file changed, 27 insertions(+), 23 deletions(-)

diff --git a/dir.c b/dir.c
index ae6f5c9636..1c55dc3e36 100644
--- a/dir.c
+++ b/dir.c
@@ -2398,7 +2398,8 @@ struct ondisk_untracked_cache {
 	char exclude_per_dir[FLEX_ARRAY];
 };
 
-#define ouc_size(len) (offsetof(struct ondisk_untracked_cache, exclude_per_dir) + len + 1)
+#define ouc_offset(x) offsetof(struct ondisk_untracked_cache, x)
+#define ouc_size(len) (ouc_offset(exclude_per_dir) + len + 1)
 
 struct write_data {
 	int index;	   /* number of written untracked_cache_dir */
@@ -2560,17 +2561,18 @@ struct read_data {
 	const unsigned char *end;
 };
 
-static void stat_data_from_disk(struct stat_data *to, const struct stat_data *from)
+static void stat_data_from_disk(struct stat_data *to, const unsigned char *data)
 {
-	to->sd_ctime.sec  = get_be32(&from->sd_ctime.sec);
-	to->sd_ctime.nsec = get_be32(&from->sd_ctime.nsec);
-	to->sd_mtime.sec  = get_be32(&from->sd_mtime.sec);
-	to->sd_mtime.nsec = get_be32(&from->sd_mtime.nsec);
-	to->sd_dev	  = get_be32(&from->sd_dev);
-	to->sd_ino	  = get_be32(&from->sd_ino);
-	to->sd_uid	  = get_be32(&from->sd_uid);
-	to->sd_gid	  = get_be32(&from->sd_gid);
-	to->sd_size	  = get_be32(&from->sd_size);
+	memcpy(to, data, sizeof(*to));
+	to->sd_ctime.sec  = ntohl(to->sd_ctime.sec);
+	to->sd_ctime.nsec = ntohl(to->sd_ctime.nsec);
+	to->sd_mtime.sec  = ntohl(to->sd_mtime.sec);
+	to->sd_mtime.nsec = ntohl(to->sd_mtime.nsec);
+	to->sd_dev	  = ntohl(to->sd_dev);
+	to->sd_ino	  = ntohl(to->sd_ino);
+	to->sd_uid	  = ntohl(to->sd_uid);
+	to->sd_gid	  = ntohl(to->sd_gid);
+	to->sd_size	  = ntohl(to->sd_size);
 }
 
 static int read_one_dir(struct untracked_cache_dir **untracked_,
@@ -2645,7 +2647,7 @@ static void read_stat(size_t pos, void *cb)
 		rd->data = rd->end + 1;
 		return;
 	}
-	stat_data_from_disk(&ud->stat_data, (struct stat_data *)rd->data);
+	stat_data_from_disk(&ud->stat_data, rd->data);
 	rd->data += sizeof(struct stat_data);
 	ud->valid = 1;
 }
@@ -2663,22 +2665,22 @@ static void read_sha1(size_t pos, void *cb)
 }
 
 static void load_sha1_stat(struct sha1_stat *sha1_stat,
-			   const struct stat_data *stat,
+			   const unsigned char *data,
 			   const unsigned char *sha1)
 {
-	stat_data_from_disk(&sha1_stat->stat, stat);
+	stat_data_from_disk(&sha1_stat->stat, data);
 	hashcpy(sha1_stat->sha1, sha1);
 	sha1_stat->valid = 1;
 }
 
 struct untracked_cache *read_untracked_extension(const void *data, unsigned long sz)
 {
-	const struct ondisk_untracked_cache *ouc;
 	struct untracked_cache *uc;
 	struct read_data rd;
 	const unsigned char *next = data, *end = (const unsigned char *)data + sz;
 	const char *ident;
 	int ident_len, len;
+	const char *exclude_per_dir;
 
 	if (sz <= 1 || end[-1] != '\0')
 		return NULL;
@@ -2690,21 +2692,23 @@ struct untracked_cache *read_untracked_extension(const void *data, unsigned long
 	ident = (const char *)next;
 	next += ident_len;
 
-	ouc = (const struct ondisk_untracked_cache *)next;
 	if (next + ouc_size(0) > end)
 		return NULL;
 
 	uc = xcalloc(1, sizeof(*uc));
 	strbuf_init(&uc->ident, ident_len);
 	strbuf_add(&uc->ident, ident, ident_len);
-	load_sha1_stat(&uc->ss_info_exclude, &ouc->info_exclude_stat,
-		       ouc->info_exclude_sha1);
-	load_sha1_stat(&uc->ss_excludes_file, &ouc->excludes_file_stat,
-		       ouc->excludes_file_sha1);
-	uc->dir_flags = get_be32(&ouc->dir_flags);
-	uc->exclude_per_dir = xstrdup(ouc->exclude_per_dir);
+	load_sha1_stat(&uc->ss_info_exclude,
+		       next + ouc_offset(info_exclude_stat),
+		       next + ouc_offset(info_exclude_sha1));
+	load_sha1_stat(&uc->ss_excludes_file,
+		       next + ouc_offset(excludes_file_stat),
+		       next + ouc_offset(excludes_file_sha1));
+	uc->dir_flags = get_be32(next + ouc_offset(dir_flags));
+	exclude_per_dir = (const char *)next + ouc_offset(exclude_per_dir);
+	uc->exclude_per_dir = xstrdup(exclude_per_dir);
 	/* NUL after exclude_per_dir is covered by sizeof(*ouc) */
-	next += ouc_size(strlen(ouc->exclude_per_dir));
+	next += ouc_size(strlen(exclude_per_dir));
 	if (next >= end)
 		goto done2;
 
-- 
2.13.3

Re: [PATCH] dir: support platforms that require aligned reads

[Jeff King]

On Sun, Jul 16, 2017 at 02:17:37PM +0200, René Scharfe wrote:

> -static void stat_data_from_disk(struct stat_data *to, const struct stat_data *from)
> +static void stat_data_from_disk(struct stat_data *to, const unsigned char *data)
>  {
> -	to->sd_ctime.sec  = get_be32(&from->sd_ctime.sec);
> -	to->sd_ctime.nsec = get_be32(&from->sd_ctime.nsec);
> -	to->sd_mtime.sec  = get_be32(&from->sd_mtime.sec);
> -	to->sd_mtime.nsec = get_be32(&from->sd_mtime.nsec);
> -	to->sd_dev	  = get_be32(&from->sd_dev);
> -	to->sd_ino	  = get_be32(&from->sd_ino);
> -	to->sd_uid	  = get_be32(&from->sd_uid);
> -	to->sd_gid	  = get_be32(&from->sd_gid);
> -	to->sd_size	  = get_be32(&from->sd_size);
> +	memcpy(to, data, sizeof(*to));
> +	to->sd_ctime.sec  = ntohl(to->sd_ctime.sec);
> +	to->sd_ctime.nsec = ntohl(to->sd_ctime.nsec);
> +	to->sd_mtime.sec  = ntohl(to->sd_mtime.sec);
> +	to->sd_mtime.nsec = ntohl(to->sd_mtime.nsec);
> +	to->sd_dev	  = ntohl(to->sd_dev);
> +	to->sd_ino	  = ntohl(to->sd_ino);
> +	to->sd_uid	  = ntohl(to->sd_uid);
> +	to->sd_gid	  = ntohl(to->sd_gid);
> +	to->sd_size	  = ntohl(to->sd_size);
>  }

Hmm. I would have written this to pull the bytes directly out of the
array, like:

  to->sd_ctime.sec  = get_be32(data); data += 4;
  to->sd_ctime.nsec = get_be32(data); data += 4;

etc. Or even a helper to do the advancing like:

  to->sd_ctime.sec = parse_be32(&data);

That reduces assumptions about padding in "struct stat_data". But
looking more at this code, and reading your comment:

> Side note: The OS name is not enough for determining the layout of
> struct ondisk_untracked_cache.  Different platforms can have different
> int sizes and padding.  Adding the machine type could help, but that
> would be a breaking change.  At that point we would be better off
> defining a machine-independent format, no?

it looks like assumptions about struct layout are pervasive and part of
the on-disk format. Yuck. :(

-Peff

Re: [PATCH] dir: support platforms that require aligned reads

[René Scharfe]

Am 16.07.2017 um 16:04 schrieb Jeff King:
> On Sun, Jul 16, 2017 at 02:17:37PM +0200, René Scharfe wrote:
> 
>> -static void stat_data_from_disk(struct stat_data *to, const struct stat_data *from)
>> +static void stat_data_from_disk(struct stat_data *to, const unsigned char *data)
>>   {
>> -	to->sd_ctime.sec  = get_be32(&from->sd_ctime.sec);
>> -	to->sd_ctime.nsec = get_be32(&from->sd_ctime.nsec);
>> -	to->sd_mtime.sec  = get_be32(&from->sd_mtime.sec);
>> -	to->sd_mtime.nsec = get_be32(&from->sd_mtime.nsec);
>> -	to->sd_dev	  = get_be32(&from->sd_dev);
>> -	to->sd_ino	  = get_be32(&from->sd_ino);
>> -	to->sd_uid	  = get_be32(&from->sd_uid);
>> -	to->sd_gid	  = get_be32(&from->sd_gid);
>> -	to->sd_size	  = get_be32(&from->sd_size);
>> +	memcpy(to, data, sizeof(*to));
>> +	to->sd_ctime.sec  = ntohl(to->sd_ctime.sec);
>> +	to->sd_ctime.nsec = ntohl(to->sd_ctime.nsec);
>> +	to->sd_mtime.sec  = ntohl(to->sd_mtime.sec);
>> +	to->sd_mtime.nsec = ntohl(to->sd_mtime.nsec);
>> +	to->sd_dev	  = ntohl(to->sd_dev);
>> +	to->sd_ino	  = ntohl(to->sd_ino);
>> +	to->sd_uid	  = ntohl(to->sd_uid);
>> +	to->sd_gid	  = ntohl(to->sd_gid);
>> +	to->sd_size	  = ntohl(to->sd_size);
>>   }
> 
> Hmm. I would have written this to pull the bytes directly out of the
> array, like:
> 
>    to->sd_ctime.sec  = get_be32(data); data += 4;
>    to->sd_ctime.nsec = get_be32(data); data += 4;
> 
> etc. Or even a helper to do the advancing like:
> 
>    to->sd_ctime.sec = parse_be32(&data);
> 
> That reduces assumptions about padding in "struct stat_data". But
> looking more at this code, and reading your comment:
> 
>> Side note: The OS name is not enough for determining the layout of
>> struct ondisk_untracked_cache.  Different platforms can have different
>> int sizes and padding.  Adding the machine type could help, but that
>> would be a breaking change.  At that point we would be better off
>> defining a machine-independent format, no?
> 
> it looks like assumptions about struct layout are pervasive and part of
> the on-disk format. Yuck. :(

Assuming that there is no padding probably even works for the platforms
the code currently supports (basically x86), but I don't know about
others.  We'd need to change the writing side as well to match, though.
Which is probably a good idea, but I tried to keep the patch small and
its impact low.  Cross-machine usability is currently explicitly not
supported -- not sure why, though.

René

Re: [PATCH] dir: support platforms that require aligned reads

[Jeff King]

On Sun, Jul 16, 2017 at 05:18:27PM +0200, René Scharfe wrote:

> > it looks like assumptions about struct layout are pervasive and part of
> > the on-disk format. Yuck. :(
> 
> Assuming that there is no padding probably even works for the platforms
> the code currently supports (basically x86), but I don't know about
> others.  We'd need to change the writing side as well to match, though.
> Which is probably a good idea, but I tried to keep the patch small and
> its impact low.  Cross-machine usability is currently explicitly not
> supported -- not sure why, though.

Yeah, I think your patch is a good first step, and we don't have to go
further for now.

-Peff