[regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Dan Williams]

Hi Tom, Jens,

My build system started reporting these error messages recently.
Reverting commit c3270e577c18b3d0e984c3371493205a4807db9d on top of
2.6.26-rc1 gets things working for me again.

--
Dan

Re: [regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Jens Axboe]

On Wed, May 07 2008, Dan Williams wrote:
> Hi Tom, Jens,
> 
> My build system started reporting these error messages recently.
> Reverting commit c3270e577c18b3d0e984c3371493205a4807db9d on top of
> 2.6.26-rc1 gets things working for me again.

Irk, that patch did scare me a bit (hence I asked Tom to double check as
wel :-). I'll take a look in the morning, all test boxes are off at this
point in time.

-- 
Jens Axboe

Re: [regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Tom Zanussi]

On Wed, 2008-05-07 at 23:16 +0200, Jens Axboe wrote:
> On Wed, May 07 2008, Dan Williams wrote:
> > Hi Tom, Jens,
> > 
> > My build system started reporting these error messages recently.
> > Reverting commit c3270e577c18b3d0e984c3371493205a4807db9d on top of
> > 2.6.26-rc1 gets things working for me again.
> 
> Irk, that patch did scare me a bit (hence I asked Tom to double check as
> wel :-). I'll take a look in the morning, all test boxes are off at this
> point in time.
> 

I did, and it still looks ok to me, but obviously it's not, so I'll have
to do some more digging.

The only thing I can think of right now that might be a possible cause
would be in splice_direct_to_actor(), if we had an incomplete transfer,
the sd->pos returned and assigned would have the value set by the failed
actor().  Maybe something like the following would take care of that
case, but I haven't had a chance to test it yet - will do that tomorrow
night...

Tom

diff --git a/fs/splice.c b/fs/splice.c
index 633f58e..1bb3f34 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -986,7 +986,7 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 
 	while (len) {
 		size_t read_len;
-		loff_t pos = sd->pos;
+		loff_t pos = sd->pos, prev_pos = pos;
 
 		ret = do_splice_to(in, &pos, pipe, len, flags);
 		if (unlikely(ret <= 0))
@@ -1001,8 +1001,10 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 		 * could get stuck data in the internal pipe:
 		 */
 		ret = actor(pipe, sd);
-		if (unlikely(ret <= 0))
+		if (unlikely(ret <= 0)) {
+			sd->pos = prev_pos;
 			goto out_release;
+		}
 
 		bytes += ret;
 		len -= ret;

Re: [regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Tom Zanussi]

On Thu, 2008-05-08 at 02:21 -0500, Tom Zanussi wrote:
> On Wed, 2008-05-07 at 23:16 +0200, Jens Axboe wrote:
> > On Wed, May 07 2008, Dan Williams wrote:
> > > Hi Tom, Jens,
> > > 
> > > My build system started reporting these error messages recently.
> > > Reverting commit c3270e577c18b3d0e984c3371493205a4807db9d on top of
> > > 2.6.26-rc1 gets things working for me again.
> > 
> > Irk, that patch did scare me a bit (hence I asked Tom to double check as
> > wel :-). I'll take a look in the morning, all test boxes are off at this
> > point in time.
> > 
> 
> I did, and it still looks ok to me, but obviously it's not, so I'll have
> to do some more digging.
> 
> The only thing I can think of right now that might be a possible cause
> would be in splice_direct_to_actor(), if we had an incomplete transfer,
> the sd->pos returned and assigned would have the value set by the failed
> actor().  Maybe something like the following would take care of that
> case, but I haven't had a chance to test it yet - will do that tomorrow
> night...
> 

Looks like I was on the right track - can you try this patch out
instead?  It makes sure sd.pos is updated correctly if the transfer was
incomplete or failed.  I ran some kernel compiles using distcc while
running blktrace in sendfile mode and didn't see any problems with
either.

Tom

diff --git a/fs/splice.c b/fs/splice.c
index 633f58e..3bd95a7 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -986,7 +986,7 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 
 	while (len) {
 		size_t read_len;
-		loff_t pos = sd->pos;
+		loff_t pos = sd->pos, prev_pos = pos;
 
 		ret = do_splice_to(in, &pos, pipe, len, flags);
 		if (unlikely(ret <= 0))
@@ -1001,15 +1001,19 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 		 * could get stuck data in the internal pipe:
 		 */
 		ret = actor(pipe, sd);
-		if (unlikely(ret <= 0))
+		if (unlikely(ret <= 0)) {
+			sd->pos = prev_pos;
 			goto out_release;
+		}
 
 		bytes += ret;
 		len -= ret;
 		sd->pos = pos;
 
-		if (ret < read_len)
+		if (ret < read_len) {
+			sd->pos = prev_pos + ret;
 			goto out_release;
+		}
 	}
 
 done:

Re: [regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Jens Axboe]

On Thu, May 08 2008, Tom Zanussi wrote:
> 
> On Thu, 2008-05-08 at 02:21 -0500, Tom Zanussi wrote:
> > On Wed, 2008-05-07 at 23:16 +0200, Jens Axboe wrote:
> > > On Wed, May 07 2008, Dan Williams wrote:
> > > > Hi Tom, Jens,
> > > > 
> > > > My build system started reporting these error messages recently.
> > > > Reverting commit c3270e577c18b3d0e984c3371493205a4807db9d on top of
> > > > 2.6.26-rc1 gets things working for me again.
> > > 
> > > Irk, that patch did scare me a bit (hence I asked Tom to double check as
> > > wel :-). I'll take a look in the morning, all test boxes are off at this
> > > point in time.
> > > 
> > 
> > I did, and it still looks ok to me, but obviously it's not, so I'll have
> > to do some more digging.
> > 
> > The only thing I can think of right now that might be a possible cause
> > would be in splice_direct_to_actor(), if we had an incomplete transfer,
> > the sd->pos returned and assigned would have the value set by the failed
> > actor().  Maybe something like the following would take care of that
> > case, but I haven't had a chance to test it yet - will do that tomorrow
> > night...
> > 
> 
> Looks like I was on the right track - can you try this patch out
> instead?  It makes sure sd.pos is updated correctly if the transfer was
> incomplete or failed.  I ran some kernel compiles using distcc while
> running blktrace in sendfile mode and didn't see any problems with
> either.

Dan, can I talk you into re-trying current git with this patch applied?
It's basically the (now) reverted broken bits plus the fix from Tom from
this email.

diff --git a/fs/splice.c b/fs/splice.c
index 7815003..a048ad2 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -983,7 +983,7 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 
 	while (len) {
 		size_t read_len;
-		loff_t pos = sd->pos;
+		loff_t pos = sd->pos, prev_pos = pos;
 
 		ret = do_splice_to(in, &pos, pipe, len, flags);
 		if (unlikely(ret <= 0))
@@ -998,15 +998,19 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 		 * could get stuck data in the internal pipe:
 		 */
 		ret = actor(pipe, sd);
-		if (unlikely(ret <= 0))
+		if (unlikely(ret <= 0)) {
+			sd->pos = prev_pos;
 			goto out_release;
+		}
 
 		bytes += ret;
 		len -= ret;
 		sd->pos = pos;
 
-		if (ret < read_len)
+		if (ret < read_len) {
+			sd->pos = prev_pos + ret;
 			goto out_release;
+		}
 	}
 
 done:
@@ -1072,7 +1076,7 @@ long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 
 	ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
 	if (ret > 0)
-		*ppos += ret;
+		*ppos = sd.pos;
 
 	return ret;
 }
diff --git a/kernel/relay.c b/kernel/relay.c
index bc24dcd..7de644c 100644
--- a/kernel/relay.c
+++ b/kernel/relay.c
@@ -1191,7 +1191,7 @@ static ssize_t relay_file_splice_read(struct file *in,
 	ret = 0;
 	spliced = 0;
 
-	while (len) {
+	while (len && !spliced) {
 		ret = subbuf_splice_actor(in, ppos, pipe, len, flags, &nonpad_ret);
 		if (ret < 0)
 			break;

-- 
Jens Axboe

Re: [regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Dan Williams]

On Fri, May 9, 2008 at 4:26 AM, Jens Axboe <jens.axboe@oracle.com> wrote:
>  Dan, can I talk you into re-trying current git with this patch applied?
>  It's basically the (now) reverted broken bits plus the fix from Tom from
>  this email.
>

distcc appears to be  happy.

Tested-by: Dan Williams <dan.j.williams@intel.com>

Re: [regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Jens Axboe]

On Fri, May 09 2008, Dan Williams wrote:
> On Fri, May 9, 2008 at 4:26 AM, Jens Axboe <jens.axboe@oracle.com> wrote:
> >  Dan, can I talk you into re-trying current git with this patch applied?
> >  It's basically the (now) reverted broken bits plus the fix from Tom from
> >  this email.
> >
> 
> distcc appears to be  happy.
> 
> Tested-by: Dan Williams <dan.j.williams@intel.com>

Super, Tom I already committed this patch as coming from you. If you
want a SOB added, let me know.

-- 
Jens Axboe

Re: [regression?] distcc says: (dcc_pump_sendfile) ERROR: sendfile returned 0? can't cope (bisected)

[Tom Zanussi]

On Fri, 2008-05-09 at 21:17 +0200, Jens Axboe wrote:
> On Fri, May 09 2008, Dan Williams wrote:
> > On Fri, May 9, 2008 at 4:26 AM, Jens Axboe <jens.axboe@oracle.com> wrote:
> > >  Dan, can I talk you into re-trying current git with this patch applied?
> > >  It's basically the (now) reverted broken bits plus the fix from Tom from
> > >  this email.
> > >
> > 
> > distcc appears to be  happy.
> > 
> > Tested-by: Dan Williams <dan.j.williams@intel.com>
> 
> Super, Tom I already committed this patch as coming from you. If you
> want a SOB added, let me know.
> 

OK, great, thanks.

Signed-off-by: Tom Zanussi <zanussi@comcast.net>