Clean up write_in_full() users

Linus Torvalds· Jan 12, 2007, 04:23 UTC · lore

With the new-and-improved write_in_full() semantics, where a partial write simply always returns a real error (and always sets 'errno' when that happens, including for the disk full case), a lot of the callers of write_in_full() were just unnecessarily complex.

In particular, there's no reason to ever check for a zero length or return: if the length was zero, we'll return zero, otherwise, if a disk full resulted in the actual write() system call returning zero the write_in_full() logic would have correctly turned that into a negative return value, with 'errno' set to ENOSPC.

I really wish every "write_in_full()" user would just check against "<0" now, but this fixes the nasty and stupid ones.

Signed-off-by: Linus Torvalds <torvalds@osdl.org>
---

I actually think "read_in_full()" should get the same loving tender care too, for all the same reasons. I think "read_or_die()" is totally broken. Anybody who uses "read_or_die()" is buggy by definition, since it will do a partial read AND NOT RETURN ANY INDICATION THAT IT WAS PARTIAL!

Can I please ask people who do these idiotic cleanups to get their act together?

I'll send a patch for that next, but I looked at "write_in_full()" first, for historical reasons.

diff --git a/sha1_file.c b/sha1_file.c
index 18dd89b..2a5be53 100644
--- a/sha1_file.c
+++ b/sha1_file.c
@@ -1618,14 +1618,7 @@ int move_temp_to_file(const char *tmpfile, const char *filename)
 
 static int write_buffer(int fd, const void *buf, size_t len)
 {
-	ssize_t size;
-
-	if (!len)
-		return 0;
-	size = write_in_full(fd, buf, len);
-	if (!size)
-		return error("file write: disk full");
-	if (size < 0)
+	if (write_in_full(fd, buf, len) < 0)
 		return error("file write error (%s)", strerror(errno));
 	return 0;
 }
diff --git a/write_or_die.c b/write_or_die.c
index 488de72..1224cac 100644
--- a/write_or_die.c
+++ b/write_or_die.c
@@ -58,14 +58,7 @@ int write_in_full(int fd, const void *buf, size_t count)
 
 void write_or_die(int fd, const void *buf, size_t count)
 {
-	ssize_t written;
-
-	if (!count)
-		return;
-	written = write_in_full(fd, buf, count);
-	if (written == 0)
-		die("disk full?");
-	else if (written < 0) {
+	if (write_in_full(fd, buf, count) < 0) {
 		if (errno == EPIPE)
 			exit(0);
 		die("write error (%s)", strerror(errno));
@@ -74,16 +67,7 @@ void write_or_die(int fd, const void *buf, size_t count)
 
 int write_or_whine_pipe(int fd, const void *buf, size_t count, const char *msg)
 {
-	ssize_t written;
-
-	if (!count)
-		return 1;
-	written = write_in_full(fd, buf, count);
-	if (written == 0) {
-		fprintf(stderr, "%s: disk full?\n", msg);
-		return 0;
-	}
-	else if (written < 0) {
+	if (write_in_full(fd, buf, count) < 0) {
 		if (errno == EPIPE)
 			exit(0);
 		fprintf(stderr, "%s: write error (%s)\n",
@@ -96,16 +80,7 @@ int write_or_whine_pipe(int fd, const void *buf, size_t count, const char *msg)
 
 int write_or_whine(int fd, const void *buf, size_t count, const char *msg)
 {
-	ssize_t written;
-
-	if (!count)
-		return 1;
-	written = write_in_full(fd, buf, count);
-	if (written == 0) {
-		fprintf(stderr, "%s: disk full?\n", msg);
-		return 0;
-	}
-	else if (written < 0) {
+	if (write_in_full(fd, buf, count) < 0) {
 		fprintf(stderr, "%s: write error (%s)\n",
 			msg, strerror(errno));
 		return 0;

Shawn O. Pearce· Jan 12, 2007, 04:33 UTC · re: Linus Torvalds · lore

Re: Clean up write_in_full() users

Linus Torvalds <torvalds@osdl.org> wrote:

> I actually think "read_in_full()" should get the same loving tender care 
> too, for all the same reasons. I think "read_or_die()" is totally broken. 
> Anybody who uses "read_or_die()" is buggy by definition, since it will do 
> a partial read AND NOT RETURN ANY INDICATION THAT IT WAS PARTIAL!

AFAIK the only user of read_or_die is sha1_file.c when it reads in the 12 byte pack header and the 20 byte pack trailer to "quickly" verify the packfile is sane before using it. If I recall correctly it was correct when I created it, but the read_in_full refactoring changed it.

-- 
Shawn.

Linus Torvalds· Jan 12, 2007, 04:43 UTC · re: Shawn O. Pearce · lore

Re: Clean up write_in_full() users

On Thu, 11 Jan 2007, Shawn O. Pearce wrote:

Show 6 quoted lines

> 
> AFAIK the only user of read_or_die is sha1_file.c when it reads in
> the 12 byte pack header and the 20 byte pack trailer to "quickly"
> verify the packfile is sane before using it.  If I recall correctly
> it was correct when I created it, but the read_in_full refactoring
> changed it.

Well, I just sent a patch to hopefully fix it, but I also think that the whole function is really misdesigned.

The thing is, "write_or_die()" actually makes sense. If you get a write error, things are really seriously broken, and it makes a lot of sense to just say "things are totally broken".

HOWEVER. The same thing is not true of a partial read. If a read doesn't succeed fully, the most _common_ case is that a file is truncated, and it does generally NOT make sense to just die and say "partial file".

So for a write error, it's ok to say "ok, I can't write, I'm dead".

For a read error, at the very least you have to say WHICH FILE couldn't be read, because it's usually a matter of some file just being too short, not some system-wide problem.

Looking at the two call-sites of "read_or_die()", they really are better off not dying, or if they died, the caller should have passed in the _reason_ (ie the name of the pack-file), or should have just used a non-dying version and tested the return value and written a message of its own.

But at least the function now WORKS after the patch I just sent out. Which it didn't do before.

			Linus

Morten Welinder· Jan 12, 2007, 14:37 UTC · re: Linus Torvalds · lore

Re: Clean up write_in_full() users

> The thing is, "write_or_die()" actually makes sense.

Except for a library. Calling exit is very unappealing in that case and trying to change die to use longjmp is even less appealing.

M.