Skip to contentSkip to conversation

AF_ALG buggy with sendfile

PATCH In review Shawn Landden 7 participants 17 messages Nov 24, 2013 – Nov 29, 2013

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

If I use sendfile() to send to a accept()ed AF_ALG socket set up for "hash", I get the wrong
answer, if I read() and then write() I get the right answer. None of the system calls return an error.

test case attached.

Signature · 3 lines
---
Shawn Landden
+1 360 389 3001 (SMS preferred)

#include <sys/sendfile.h>
#include <sys/socket.h>
#include <linux/if_alg.h>
#include <stdio.h>
#include <sys/stat.h>
#include <fcntl.h>

int main(void)

{
	int opfd;
	int tfmfd;
	struct sockaddr_alg sa = {
		.salg_family = AF_ALG,
		.salg_type = "hash",
		.salg_name = "sha1"
	};
	char buf2[10000000];
	char buf[20];
	int i;
	struct stat st;

	tfmfd = socket(AF_ALG, SOCK_SEQPACKET, 0);

	bind(tfmfd, (struct sockaddr *)&sa, sizeof(sa));

	opfd = accept(tfmfd, NULL, 0);

        int true = open("/bin/true", O_RDONLY);
        fstat(true, &st);

        sendfile(opfd, true, NULL, st.st_size);
	read(opfd, &buf, 20);

	for (i = 0; i < 20; i++) {
		printf("%02x", (unsigned char)buf[i]);
	}
	printf("\n");

        lseek(true, 0, SEEK_SET);
        read(true, &buf2, st.st_size);
        write(opfd, &buf2, st.st_size);
	read(opfd, &buf, 20);

	for (i = 0; i < 20; i++) {
		printf("%02x", (unsigned char)buf[i]);
	}
	printf("\n");

	close(opfd);
	close(tfmfd);

	return 0;
}
#1 lore raw

heres a version of the test case that builds.....

Sorry about that.

Quoted 12 lines · Shawn Landden

If I use sendfile() to send to a accept()ed AF_ALG socket set up for "hash", I get the wrong
answer, if I read() and then write() I get the right answer. None of the system calls return an error.

test case attached.

--

---
Shawn Landden
+1 360 389 3001 (SMS preferred)

Signature · 3 lines
---
Shawn Landden
+1 360 389 3001 (SMS preferred)

#include <sys/sendfile.h>
#include <sys/socket.h>
#include <linux/if_alg.h>
#include <stdio.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <stdlib.h>
#include <unistd.h>
int main(void)

{
	int opfd;
	int tfmfd;
	struct sockaddr_alg sa = {
		.salg_family = AF_ALG,
		.salg_type = "hash",
		.salg_name = "sha1"
	};
	char *buf2;
	char buf[20];
	int i;
	struct stat st;
	ssize_t size;

	tfmfd = socket(AF_ALG, SOCK_SEQPACKET, 0);

	bind(tfmfd, (struct sockaddr *)&sa, sizeof(sa));

	opfd = accept(tfmfd, NULL, 0);

        int t = open("/bin/true", O_RDONLY);
        fstat(t, &st);

        size = sendfile(opfd, t, NULL, st.st_size);
        if (size != st.st_size)
                exit(1);                
	read(opfd, &buf, 20);

	for (i = 0; i < 20; i++) {
		printf("%02x", (unsigned char)buf[i]);
	}
	printf("\n");

        lseek(t, 0, SEEK_SET);
        buf2 = malloc(st.st_size + 1);
        read(t, buf2, st.st_size);
        write(opfd, buf2, st.st_size);
	read(opfd, &buf, 20);

	for (i = 0; i < 20; i++) {
		printf("%02x", (unsigned char)buf[i]);
	}
	printf("\n");

	close(opfd);
	close(tfmfd);

	return 0;
}
#2 lore raw

If you build https://kernel.googlesource.com/pub/scm/network/connman/connman/+/0.80/tools/alg-test.c
from the connman source code and compare the output to coreutils
sha1sum you can see the problem.

shawn@debian-T61:~/git/test$ make connman_afalg
cc     connman_afalg.c   -o connman_afalg
shawn@debian-T61:~/git/test$ ./connman_afalg /bin/true
send 27080 bytes
recv 20 bytes
45384483cf9cd0d82eba164131795b4807c6d39d  /bin/true
shawn@debian-T61:~/git/test$ sha1sum /bin/true
82667ba2ec681d8e55b0ee3b6db2970f9911680d  /bin/true
Quoted 25 lines · Shawn Landden

heres a version of the test case that builds.....

Sorry about that.

Quoted 5 lines · Shawn Landden

If I use sendfile() to send to a accept()ed AF_ALG socket set up for "hash", I get the wrong
answer, if I read() and then write() I get the right answer. None of the system calls return an error.

Quoted 7 lines

test case attached.

--

---
Shawn Landden
+1 360 389 3001 (SMS preferred)

--

---
Shawn Landden
+1 360 389 3001 (SMS preferred)

Signature · 3 lines
---
Shawn Landden
+1 360 389 3001 (SMS preferred)
#3 lore raw

PATCH 1/1 pipe_to_sendpage: Ensure that MSG_MORE is set if we set MSG_SENDPAGE_NOTLAST

In reply to Shawn Landden

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST.
We have to ensure that MSG_MORE is also set if we set MSG_SENDPAGE_NOTLAST. Otherwise users that check against MSG_MORE will not see it.

This fixes sendfile() on AF_ALG.

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x
Reported-and-tested-by Shawn Landden <[email protected]>
---
 fs/splice.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/splice.c +1 -1
index 3b7ee65..b93f1b8 100644
@@ -701,7 +701,7 @@ static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 	more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 
 	if (sd->len < sd->total_len && pipe->nrbufs > 1)
-		more |= MSG_SENDPAGE_NOTLAST;
+		more |= MSG_SENDPAGE_NOTLAST | MSG_MORE;
 
 	return file->f_op->sendpage(file, buf->page, buf->offset,
 				    sd->len, &pos, more);
Signature · 1 line
1.8.1.4
#4 lore raw

Re: [PATCH] pipe_to_sendpage: Ensure that MSG_MORE is set if we set MSG_SENDPAGE_NOTLAST

Quoted 11 lines · Richard Weinberger

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST.
We have to ensure that MSG_MORE is also set if we set MSG_SENDPAGE_NOTLAST. Otherwise users that check against MSG_MORE will not see it.

This fixes sendfile() on AF_ALG.

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x

The offending commit also got backported to the 3.2 stable kernel, so we need this fix there as well.
---
Shawn Landden
+1 360 389 3001 (SMS preferred)

#5 lore raw

Re: [PATCH] pipe_to_sendpage: Ensure that MSG_MORE is set if we set MSG_SENDPAGE_NOTLAST

In reply to Richard Weinberger
Quoted 22 lines · Richard Weinberger

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST.
We have to ensure that MSG_MORE is also set if we set MSG_SENDPAGE_NOTLAST. Otherwise users that check against MSG_MORE will not see it.

This fixes sendfile() on AF_ALG.

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x
Reported-and-tested-by Shawn Landden <[email protected]>
---
 fs/splice.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/splice.c +0 -0
index 3b7ee65..b93f1b8 100644

fs/splice.c static int pipe_to_sendpage(struct pipe_inode_info *pipe, · line 707

Quoted 8 lines
fs/splice.c +1 -1
@@ -701,7 +701,7 @@ static int pipe_to_sendpage(struct pipe_inode_info *pipe,
 	more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 
 	if (sd->len < sd->total_len && pipe->nrbufs > 1)
-		more |= MSG_SENDPAGE_NOTLAST;
+		more |= MSG_SENDPAGE_NOTLAST | MSG_MORE;
 
 	return file->f_op->sendpage(file, buf->page, buf->offset,
 				    sd->len, &pos, more);

I do not think this patch is right. It looks like a revert of a useful patch for TCP zero copy. Given the time it took to discover this regression, I bet tcp zero copy has more users than AF_ALG, by 5 or 6 order of magnitude ;)

Here we want to make the difference between the two flags, not merge them.

If AF_ALG do not care of the difference, try instead :

crypto/algif_hash.c +3 -0
index ef5356cd280a..850246206b12 100644
@@ -114,6 +114,9 @@ static ssize_t hash_sendpage(struct socket *sock, struct page *page,
 	struct hash_ctx *ctx = ask->private;
 	int err;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	lock_sock(sk);
 	sg_init_table(ctx->sgl.sg, 1);
 	sg_set_page(ctx->sgl.sg, page, size, offset);
#6 lore raw

Re: [PATCH] pipe_to_sendpage: Ensure that MSG_MORE is set if we set MSG_SENDPAGE_NOTLAST

Quoted 59 lines · Eric Dumazet
Quoted 22 lines · Richard Weinberger

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST.
We have to ensure that MSG_MORE is also set if we set MSG_SENDPAGE_NOTLAST. Otherwise users that check against MSG_MORE will not see it.

This fixes sendfile() on AF_ALG.

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x
Reported-and-tested-by Shawn Landden <[email protected]>
---
 fs/splice.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/splice.c +0 -0
index 3b7ee65..b93f1b8 100644

fs/splice.c static int pipe_to_sendpage(struct pipe_inode_info *pipe, · line 707

Quoted 8 lines
fs/splice.c +1 -1
@@ -701,7 +701,7 @@ static int pipe_to_sendpage(struct pipe_inode_info *pipe,
      more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 
      if (sd->len < sd->total_len && pipe->nrbufs > 1)
-             more |= MSG_SENDPAGE_NOTLAST;
+             more |= MSG_SENDPAGE_NOTLAST | MSG_MORE;
 
      return file->f_op->sendpage(file, buf->page, buf->offset,
                                  sd->len, &pos, more);

I do not think this patch is right. It looks like a revert of a useful patch for TCP zero copy. Given the time it took to discover this regression, I bet tcp zero copy has more users than AF_ALG, by 5 or 6 order of magnitude ;)

Here we want to make the difference between the two flags, not merge them.

If AF_ALG do not care of the difference, try instead :

crypto/algif_hash.c +3 -0
index ef5356cd280a..850246206b12 100644
@@ -114,6 +114,9 @@ static ssize_t hash_sendpage(struct socket *sock, struct page *page,
        struct hash_ctx *ctx = ask->private;
        int err;
 
+       if (flags & MSG_SENDPAGE_NOTLAST)
+               flags |= MSG_MORE;
+
        lock_sock(sk);
        sg_init_table(ctx->sgl.sg, 1);
        sg_set_page(ctx->sgl.sg, page, size, offset);

From my testing this works.

Signature · 3 lines
---
Shawn Landden
+1 360 389 3001 (SMS preferred)
#7 lore raw

PATCH 1/1 Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

In reply to Eric Dumazet

algif_hash and algif_skcipher used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x + 3.2.x
Reported-and-tested-by Shawn Landden <[email protected]>

Original-patch: Richard Weinberger <[email protected]>

---
 crypto/algif_hash.c     | 3 +++
 crypto/algif_skcipher.c | 3 +++
 2 files changed, 6 insertions(+)
crypto/algif_hash.c +3 -0
index ef5356c..8502462 100644
@@ -114,6 +114,9 @@ static ssize_t hash_sendpage(struct socket *sock, struct page *page,
 	struct hash_ctx *ctx = ask->private;
 	int err;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	lock_sock(sk);
 	sg_init_table(ctx->sgl.sg, 1);
 	sg_set_page(ctx->sgl.sg, page, size, offset);
crypto/algif_skcipher.c +3 -0
index 6a6dfc0..a19c027 100644
@@ -378,6 +378,9 @@ static ssize_t skcipher_sendpage(struct socket *sock, struct page *page,
 	struct skcipher_sg_list *sgl;
 	int err = -EINVAL;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	lock_sock(sk);
 	if (!ctx->more && ctx->used)
 		goto unlock;
Signature · 1 line
1.8.4.4
#8 lore raw

PATCH 1/1 update consumers of MSG_MORE to recognize MSG_SENDPAGE_NOTLAST

In reply to Eric Dumazet

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

algif_hash and algif_skcipher used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x + 3.2.x
Reported-and-tested-by Shawn Landden <[email protected]>

Original-patch: Richard Weinberger <[email protected]>

---
 crypto/algif_hash.c     | 3 +++
 crypto/algif_skcipher.c | 3 +++
 2 files changed, 6 insertions(+)
crypto/algif_hash.c +3 -0
index ef5356c..8502462 100644
@@ -114,6 +114,9 @@ static ssize_t hash_sendpage(struct socket *sock, struct page *page,
 	struct hash_ctx *ctx = ask->private;
 	int err;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	lock_sock(sk);
 	sg_init_table(ctx->sgl.sg, 1);
 	sg_set_page(ctx->sgl.sg, page, size, offset);
crypto/algif_skcipher.c +3 -0
index 6a6dfc0..a19c027 100644
@@ -378,6 +378,9 @@ static ssize_t skcipher_sendpage(struct socket *sock, struct page *page,
 	struct skcipher_sg_list *sgl;
 	int err = -EINVAL;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	lock_sock(sk);
 	if (!ctx->more && ctx->used)
 		goto unlock;
Signature · 1 line
1.8.4.4
#9 lore raw

Re: [PATCH] update consumers of MSG_MORE to recognize MSG_SENDPAGE_NOTLAST

Quoted 8 lines · Shawn Landden

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

algif_hash and algif_skcipher used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

Don't we need a similar fix for udp_sendpage?

Greetings,

Hannes

#10 lore raw

PATCH 1/1 update consumers of MSG_MORE to recognize MSG_SENDPAGE_NOTLAST

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

algif_hash, algif_skcipher, and udp used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

v3: also fix udp

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x + 3.2.x
Reported-and-tested-by Shawn Landden <[email protected]>

Original-patch: Richard Weinberger <[email protected]>

---
 crypto/algif_hash.c     | 3 +++
 crypto/algif_skcipher.c | 3 +++
 net/ipv4/udp.c          | 3 +++
 3 files changed, 9 insertions(+)
crypto/algif_hash.c +3 -0
index ef5356c..8502462 100644
@@ -114,6 +114,9 @@ static ssize_t hash_sendpage(struct socket *sock, struct page *page,
 	struct hash_ctx *ctx = ask->private;
 	int err;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	lock_sock(sk);
 	sg_init_table(ctx->sgl.sg, 1);
 	sg_set_page(ctx->sgl.sg, page, size, offset);
crypto/algif_skcipher.c +3 -0
index 6a6dfc0..a19c027 100644
@@ -378,6 +378,9 @@ static ssize_t skcipher_sendpage(struct socket *sock, struct page *page,
 	struct skcipher_sg_list *sgl;
 	int err = -EINVAL;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	lock_sock(sk);
 	if (!ctx->more && ctx->used)
 		goto unlock;
net/ipv4/udp.c +3 -0
index 5944d7d..8bd04df 100644
@@ -1098,6 +1098,9 @@ int udp_sendpage(struct sock *sk, struct page *page, int offset,
 	struct udp_sock *up = udp_sk(sk);
 	int ret;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+
 	if (!up->pending) {
 		struct msghdr msg = {	.msg_flags = flags|MSG_MORE };
 
Signature · 1 line
1.8.4.4
#11 lore raw

Re: [PATCH] update consumers of MSG_MORE to recognize MSG_SENDPAGE_NOTLAST

Quoted 11 lines · Shawn Landden

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

algif_hash, algif_skcipher, and udp used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

v3: also fix udp

Quoted 7 lines
Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x + 3.2.x
Reported-and-tested-by Shawn Landden <[email protected]>

Original-patch: Richard Weinberger <[email protected]>

May I ask why you took over the my patch without even CC'in me nor replying to the original
thread "[PATCH] pipe_to_sendpage: Ensure that MSG_MORE is set if we set MSG_SENDPAGE_NOTLAST"?
You are acting very rude.

The discussion at the original thread is not done.
Does skcipher_sendpage() really also need fixing? or UDP?
I didn't send another patch because I'm waiting for Eric's answer first.

Thanks,
//richard

Quoted 55 lines
---
 crypto/algif_hash.c     | 3 +++
 crypto/algif_skcipher.c | 3 +++
 net/ipv4/udp.c          | 3 +++
 3 files changed, 9 insertions(+)
crypto/algif_hash.c +3 -0
index ef5356c..8502462 100644
@@ -114,6 +114,9 @@ static ssize_t hash_sendpage(struct socket *sock, struct page *page,
        struct hash_ctx *ctx = ask->private;
        int err;
 
+       if (flags & MSG_SENDPAGE_NOTLAST)
+               flags |= MSG_MORE;
+
        lock_sock(sk);
        sg_init_table(ctx->sgl.sg, 1);
        sg_set_page(ctx->sgl.sg, page, size, offset);
crypto/algif_skcipher.c +3 -0
index 6a6dfc0..a19c027 100644
@@ -378,6 +378,9 @@ static ssize_t skcipher_sendpage(struct socket *sock, struct page *page,
        struct skcipher_sg_list *sgl;
        int err = -EINVAL;
 
+       if (flags & MSG_SENDPAGE_NOTLAST)
+               flags |= MSG_MORE;
+
        lock_sock(sk);
        if (!ctx->more && ctx->used)
                goto unlock;
net/ipv4/udp.c +3 -1
index 5944d7d..8bd04df 100644
@@ -1098,6 +1098,9 @@ int udp_sendpage(struct sock *sk, struct page *page, int offset,
        struct udp_sock *up = udp_sk(sk);
        int ret;
 
+       if (flags & MSG_SENDPAGE_NOTLAST)
+               flags |= MSG_MORE;
+
        if (!up->pending) {
                struct msghdr msg = {   .msg_flags = flags|MSG_MORE };
 
--

1.8.4.4

--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Signature · 2 lines
Thanks,
//richard
#12 lore raw

Re: [PATCH] update consumers of MSG_MORE to recognize MSG_SENDPAGE_NOTLAST

In reply to Shawn Landden
Quoted 10 lines · Shawn Landden

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

algif_hash, algif_skcipher, and udp used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

v3: also fix udp

The UDP bits look fine to me.

Greetings,

Hannes

#13 lore raw

Re: [PATCH] update consumers of MSG_MORE to recognize MSG_SENDPAGE_NOTLAST

Shawn,

Am Freitag, 29. November 2013, 06:47:04 schrieb Hannes Frederic Sowa:

Quoted 5 lines · Shawn Landden
Quoted 4 lines

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

Quoted 8 lines
Quoted 6 lines

algif_hash, algif_skcipher, and udp used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

v3: also fix udp

The UDP bits look fine to me.

please don't forget to resend the patch (only the UDP fix). I sent the AF_ALG part already.

Thanks,
//richard

maintainer
#14 lore raw

Re: [PATCH] update consumers of MSG_MORE to recognize MSG_SENDPAGE_NOTLAST

In reply to Shawn Landden

From: Shawn Landden <[email protected]>
Date: Sun, 24 Nov 2013 22:36:28 -0800

Quoted 11 lines

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST, similar to MSG_MORE.

algif_hash, algif_skcipher, and udp used MSG_MORE from tcp_sendpages() and need to see the new flag as identical to MSG_MORE.

This fixes sendfile() on AF_ALG.

v3: also fix udp

Quoted 7 lines
Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x + 3.2.x
Reported-and-tested-by Shawn Landden <[email protected]>

Original-patch: Richard Weinberger <[email protected]>

Applied and queued up for -stable, thanks everyone.

#15 lore raw

Re: [PATCH] pipe_to_sendpage: Ensure that MSG_MORE is set if we set MSG_SENDPAGE_NOTLAST

In reply to Eric Dumazet

Am Sonntag, 24. November 2013, 17:25:06 schrieb Eric Dumazet:

Quoted 33 lines · Richard Weinberger
Quoted 32 lines

Commit 35f9c09fe (tcp: tcp_sendpages() should call tcp_push() once) added an internal flag MSG_SENDPAGE_NOTLAST.
We have to ensure that MSG_MORE is also set if we set MSG_SENDPAGE_NOTLAST.
Otherwise users that check against MSG_MORE will not see it.

This fixes sendfile() on AF_ALG.

Cc Tom Herbert <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc David S. Miller <[email protected]>
Cc <[email protected]> # 3.4.x
Reported-and-tested-by Shawn Landden <[email protected]>

---

 fs/splice.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
fs/splice.c +1 -1
index 3b7ee65..b93f1b8 100644
@@ -701,7 +701,7 @@ static int pipe_to_sendpage(struct pipe_inode_info
 *pipe,> 
 	more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
 
 	if (sd->len < sd->total_len && pipe->nrbufs > 1)
 
-		more |= MSG_SENDPAGE_NOTLAST;
+		more |= MSG_SENDPAGE_NOTLAST | MSG_MORE;
 
Quoted 8 lines

fs/splice.c

Quoted 3 lines
fs/splice.c +0 -0
 	return file->f_op->sendpage(file, buf->page, buf->offset,
 
 				    sd->len, &pos, more);

I do not think this patch is right. It looks like a revert of a useful patch for TCP zero copy. Given the time it took to discover this regression, I bet tcp zero copy has more users than AF_ALG, by 5 or 6 order of magnitude ;)

Yeah, but AF_ALG broke. That's why I did the patch.

Quoted 10 lines

Here we want to make the difference between the two flags, not merge them.

If AF_ALG do not care of the difference, try instead :

crypto/algif_hash.c +0 -0
index ef5356cd280a..850246206b12 100644

crypto/algif_hash.c static ssize_t hash_sendpage(struct socket *sock, struct · line 119

Quoted 6 lines
crypto/algif_hash.c +3 -0
@@ -114,6 +114,9 @@ static ssize_t hash_sendpage(struct socket *sock, struct
 page *page, struct hash_ctx *ctx = ask->private;
 	int err;
 
+	if (flags & MSG_SENDPAGE_NOTLAST)
+		flags |= MSG_MORE;
+

In the commit message of your patch you wrote "For all sendpage() providers, its a transparent change.". Why does AF_ALG need special handling? If users have to care about MSG_SENDPAGE_NOTLAST it is no longer really an internal flag.

Thanks,
//richard

#16 lore raw

Re: [PATCH] pipe_to_sendpage: Ensure that MSG_MORE is set if we set MSG_SENDPAGE_NOTLAST

Quoted 4 lines · Richard Weinberger

In the commit message of your patch you wrote "For all sendpage() providers, its a transparent change.". Why does AF_ALG need special handling? If users have to care about MSG_SENDPAGE_NOTLAST it is no longer really an internal flag.

MSG_SENDPAGE_NOTLAST is an internal (kernel) flag.

Fact that I missed some MSG_MORE 'users' in the kernel was an oversight.

I am not saying your patch is not needed, I am only saying it reverts a useful TCP optimization, and we can do better, dont we ?