[PATCH] splice: fix wrong __splice_from_pipe() usage

[PATCH] splice: fix wrong __splice_from_pipe() usage

[OGAWA Hirofumi]

Hi,

I've noticed the nfsd read corruption by recent change. And this patch
fixes the problem for me, is this right fix?
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>


__splice_from_pipe() is updating the sd->pos for the actor, but those
functions are passing the sd of reader side directory. So, splice
updates sd->pos twice.

This fixes usage of __splice_from_pipe().

Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
---

 drivers/block/loop.c |    5 +++--
 fs/nfsd/vfs.c        |    5 +++--
 2 files changed, 6 insertions(+), 4 deletions(-)

diff -puN fs/nfsd/vfs.c~splice-fix-__splice_from_pipe-usage fs/nfsd/vfs.c
--- linux-2.6/fs/nfsd/vfs.c~splice-fix-__splice_from_pipe-usage	2007-07-16 09:30:10.000000000 +0900
+++ linux-2.6-hirofumi/fs/nfsd/vfs.c	2007-07-16 09:30:10.000000000 +0900
@@ -842,9 +842,10 @@ nfsd_splice_actor(struct pipe_inode_info
 }
 
 static int nfsd_direct_splice_actor(struct pipe_inode_info *pipe,
-				    struct splice_desc *sd)
+				    struct splice_desc *__sd)
 {
-	return __splice_from_pipe(pipe, sd, nfsd_splice_actor);
+	struct splice_desc sd = *__sd;
+	return __splice_from_pipe(pipe, &sd, nfsd_splice_actor);
 }
 
 static __be32
diff -puN drivers/block/loop.c~splice-fix-__splice_from_pipe-usage drivers/block/loop.c
--- linux-2.6/drivers/block/loop.c~splice-fix-__splice_from_pipe-usage	2007-07-16 09:30:10.000000000 +0900
+++ linux-2.6-hirofumi/drivers/block/loop.c	2007-07-16 09:30:10.000000000 +0900
@@ -437,9 +437,10 @@ lo_splice_actor(struct pipe_inode_info *
 }
 
 static int
-lo_direct_splice_actor(struct pipe_inode_info *pipe, struct splice_desc *sd)
+lo_direct_splice_actor(struct pipe_inode_info *pipe, struct splice_desc *__sd)
 {
-	return __splice_from_pipe(pipe, sd, lo_splice_actor);
+	struct splice_desc sd = *__sd;
+	return __splice_from_pipe(pipe, &sd, lo_splice_actor);
 }
 
 static int
_

Re: [PATCH] splice: fix wrong __splice_from_pipe() usage

[Jens Axboe]

On Mon, Jul 16 2007, OGAWA Hirofumi wrote:
> Hi,
> 
> I've noticed the nfsd read corruption by recent change. And this patch
> fixes the problem for me, is this right fix?
> -- 
> OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
> 
> 
> __splice_from_pipe() is updating the sd->pos for the actor, but those
> functions are passing the sd of reader side directory. So, splice
> updates sd->pos twice.
> 
> This fixes usage of __splice_from_pipe().

For sendfile() usage, or the nfsd path that uses splice to send?

-- 
Jens Axboe

Re: [PATCH] splice: fix wrong __splice_from_pipe() usage

[OGAWA Hirofumi]

Jens Axboe <jens.axboe@oracle.com> writes:

> On Mon, Jul 16 2007, OGAWA Hirofumi wrote:
>> Hi,
>> 
>> I've noticed the nfsd read corruption by recent change. And this patch
>> fixes the problem for me, is this right fix?
>> -- 
>> OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
>> 
>> 
>> __splice_from_pipe() is updating the sd->pos for the actor, but those
>> functions are passing the sd of reader side directory. So, splice
>> updates sd->pos twice.
>> 
>> This fixes usage of __splice_from_pipe().
>
> For sendfile() usage, or the nfsd path that uses splice to send?

nfsd_vfs_read() path.

nfsd_vfs_read()
    splice_direct_to_actor()
        while(len) {
            do_splice_to()                     [update sd->pos]
                -> generic_file_splice_read()  [read from sd->pos]
            nfsd_direct_splice_actor()
                -> __splice_from_pipe()        [update sd->pos]
        }

do_splice_to() updates sd->pos for read, and pass updated sd to
__splice_from_pipe(), and __splice_from_pipe() updates sd->pos for
write. So, next do_splice_to() read from wrong position.
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: [PATCH] splice: fix wrong __splice_from_pipe() usage

[Jens Axboe]

On Mon, Jul 16 2007, OGAWA Hirofumi wrote:
> Jens Axboe <jens.axboe@oracle.com> writes:
> 
> > On Mon, Jul 16 2007, OGAWA Hirofumi wrote:
> >> Hi,
> >> 
> >> I've noticed the nfsd read corruption by recent change. And this patch
> >> fixes the problem for me, is this right fix?
> >> -- 
> >> OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
> >> 
> >> 
> >> __splice_from_pipe() is updating the sd->pos for the actor, but those
> >> functions are passing the sd of reader side directory. So, splice
> >> updates sd->pos twice.
> >> 
> >> This fixes usage of __splice_from_pipe().
> >
> > For sendfile() usage, or the nfsd path that uses splice to send?
> 
> nfsd_vfs_read() path.
> 
> nfsd_vfs_read()
>     splice_direct_to_actor()
>         while(len) {
>             do_splice_to()                     [update sd->pos]
>                 -> generic_file_splice_read()  [read from sd->pos]
>             nfsd_direct_splice_actor()
>                 -> __splice_from_pipe()        [update sd->pos]
>         }
> 
> do_splice_to() updates sd->pos for read, and pass updated sd to
> __splice_from_pipe(), and __splice_from_pipe() updates sd->pos for
> write. So, next do_splice_to() read from wrong position.

Ah, I see. Copying the structure over would work, but it seems like a
big work-around for something that is essentially bad behaviour in the
splice core for direct splicing.

Can you check if this works for you?

diff --git a/fs/splice.c b/fs/splice.c
index 6c98286..5e5071a 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -1061,8 +1061,9 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 
 	while (len) {
 		size_t read_len;
+		loff_t pos = sd->pos;
 
-		ret = do_splice_to(in, &sd->pos, pipe, len, flags);
+		ret = do_splice_to(in, &pos, pipe, len, flags);
 		if (unlikely(ret <= 0))
 			goto out_release;
 
@@ -1080,6 +1081,7 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
 
 		bytes += ret;
 		len -= ret;
+		sd->pos = pos;
 
 		if (ret < read_len)
 			goto out_release;

-- 
Jens Axboe

Re: [PATCH] splice: fix wrong __splice_from_pipe() usage

[OGAWA Hirofumi]

Jens Axboe <jens.axboe@oracle.com> writes:

>> nfsd_vfs_read() path.
>> 
>> nfsd_vfs_read()
>>     splice_direct_to_actor()
>>         while(len) {
>>             do_splice_to()                     [update sd->pos]
>>                 -> generic_file_splice_read()  [read from sd->pos]
>>             nfsd_direct_splice_actor()
>>                 -> __splice_from_pipe()        [update sd->pos]
>>         }
>> 
>> do_splice_to() updates sd->pos for read, and pass updated sd to
>> __splice_from_pipe(), and __splice_from_pipe() updates sd->pos for
>> write. So, next do_splice_to() read from wrong position.
>
> Ah, I see. Copying the structure over would work, but it seems like a
> big work-around for something that is essentially bad behaviour in the
> splice core for direct splicing.
>
> Can you check if this works for you?

Simple test was done. The patch works for me.

Thanks.
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: [PATCH] splice: fix wrong __splice_from_pipe() usage

[Jens Axboe]

On Mon, Jul 16 2007, OGAWA Hirofumi wrote:
> Jens Axboe <jens.axboe@oracle.com> writes:
> 
> >> nfsd_vfs_read() path.
> >> 
> >> nfsd_vfs_read()
> >>     splice_direct_to_actor()
> >>         while(len) {
> >>             do_splice_to()                     [update sd->pos]
> >>                 -> generic_file_splice_read()  [read from sd->pos]
> >>             nfsd_direct_splice_actor()
> >>                 -> __splice_from_pipe()        [update sd->pos]
> >>         }
> >> 
> >> do_splice_to() updates sd->pos for read, and pass updated sd to
> >> __splice_from_pipe(), and __splice_from_pipe() updates sd->pos for
> >> write. So, next do_splice_to() read from wrong position.
> >
> > Ah, I see. Copying the structure over would work, but it seems like a
> > big work-around for something that is essentially bad behaviour in the
> > splice core for direct splicing.
> >
> > Can you check if this works for you?
> 
> Simple test was done. The patch works for me.

Great, thanks a lot for retesting. I'll get the patch upstream asap.

-- 
Jens Axboe