net: unix: Align send data_len up to PAGE_SIZE

net: unix: Align send data_len up to PAGE_SIZE

[Kirill Tkhai]

Using whole of allocated pages reduces requested skb->data size.
This is just a more thriftily allocation.
    
netperf does not show difference with the current performance.
    
Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
---
 net/unix/af_unix.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
index 749f80c..0830005 100644
--- a/net/unix/af_unix.c
+++ b/net/unix/af_unix.c
@@ -1492,10 +1492,14 @@ static int unix_dgram_sendmsg(struct kiocb *kiocb, struct socket *sock,
 	if (len > sk->sk_sndbuf - 32)
 		goto out;
 
-	if (len > SKB_MAX_ALLOC)
+	if (len > SKB_MAX_ALLOC) {
 		data_len = min_t(size_t,
 				 len - SKB_MAX_ALLOC,
 				 MAX_SKB_FRAGS * PAGE_SIZE);
+		data_len = min_t(size_t,
+				 len,
+				 PAGE_ALIGN(data_len));
+	}
 
 	skb = sock_alloc_send_pskb(sk, len - data_len, data_len,
 				   msg->msg_flags & MSG_DONTWAIT, &err,
@@ -1670,6 +1674,8 @@ static int unix_stream_sendmsg(struct kiocb *kiocb, struct socket *sock,
 
 		data_len = max_t(int, 0, size - SKB_MAX_HEAD(0));
 
+		data_len = min_t(size_t, size, PAGE_ALIGN(data_len));
+
 		skb = sock_alloc_send_pskb(sk, size - data_len, data_len,
 					   msg->msg_flags & MSG_DONTWAIT, &err,
 					   get_order(UNIX_SKB_FRAGS_SZ));

Re: net: unix: Align send data_len up to PAGE_SIZE

[David Miller]

From: Kirill Tkhai <ktkhai@parallels.com>
Date: Mon, 12 May 2014 18:52:12 +0400

>  		data_len = min_t(size_t,
>  				 len - SKB_MAX_ALLOC,
>  				 MAX_SKB_FRAGS * PAGE_SIZE);
> +		data_len = min_t(size_t,
> +				 len,
> +				 PAGE_ALIGN(data_len));
> +	}

When I see:

		x = min(y - N, z1);
		x = min(y, z2);

I'm a bit suspicious.  Why are you not subtracting the constant
factor out of the first variable in the second min() call?

Re: net: unix: Align send data_len up to PAGE_SIZE

[Kirill Tkhai]

В Ср, 14/05/2014 в 15:15 -0400, David Miller пишет:
> From: Kirill Tkhai <ktkhai@parallels.com>
> Date: Mon, 12 May 2014 18:52:12 +0400
> 
> >  		data_len = min_t(size_t,
> >  				 len - SKB_MAX_ALLOC,
> >  				 MAX_SKB_FRAGS * PAGE_SIZE);
> > +		data_len = min_t(size_t,
> > +				 len,
> > +				 PAGE_ALIGN(data_len));
> > +	}
> 
> When I see:
> 
> 		x = min(y - N, z1);
> 		x = min(y, z2);
> 
> I'm a bit suspicious.  Why are you not subtracting the constant
> factor out of the first variable in the second min() call?

Because, in the most cases (len - SKB_MAX_ALLOC) < PAGE_ALIGN(data_len),
and the only payload of the patch is it tries to fix that :)
We use unused space of allocated page. For mem cache it's easier
to allocate

(len - PAGE_ALIGN(x)) than (len - x).


Here really should be 

 		data_len = min_t(size_t,
 				 len - SKB_MAX_ALLOC,
 				 MAX_SKB_FRAGS * PAGE_SIZE);
+		data_len = PAGE_ALIGN(data_len));


I added the second min, because I was afraid somebody plays with
SKB_MAX_ALLOC size in debug purposes. Not sure now.

Ok, yes, simple PAGE_ALIGN is much better :)

+ data_len = PAGE_ALIGN(data_len));

And I assumed it's OK to allocate skbs with zero header size like this:

sock_alloc_send_pskb(sk, 0, data_len)

Please say, if it's wrong.

Thanks!
Kirill

Re: net: unix: Align send data_len up to PAGE_SIZE

[Eric Dumazet]

On Thu, 2014-05-15 at 12:54 +0400, Kirill Tkhai wrote:

> 
> And I assumed it's OK to allocate skbs with zero header size like this:
> 
> sock_alloc_send_pskb(sk, 0, data_len)
> 
> Please say, if it's wrong.
> 

Its OK

Re: net: unix: Align send data_len up to PAGE_SIZE

[Kirill Tkhai]

В Чт, 15/05/2014 в 06:27 -0700, Eric Dumazet пишет:
> On Thu, 2014-05-15 at 12:54 +0400, Kirill Tkhai wrote:
> 
> > 
> > And I assumed it's OK to allocate skbs with zero header size like this:
> > 
> > sock_alloc_send_pskb(sk, 0, data_len)
> > 
> > Please say, if it's wrong.
> > 
> 
> Its OK
> 

Thanks for the reply. So, new version is below.

[PATCH]net: unix: Align send data_len up to PAGE_SIZE
    
Using whole of allocated pages reduces requested skb->data size.
This is just a little more thriftily allocation.
    
netperf does not show difference with the current performance.
    
Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
---
 net/unix/af_unix.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
index 749f80c..65d67c2 100644
--- a/net/unix/af_unix.c
+++ b/net/unix/af_unix.c
@@ -1492,10 +1492,12 @@ static int unix_dgram_sendmsg(struct kiocb *kiocb, struct socket *sock,
 	if (len > sk->sk_sndbuf - 32)
 		goto out;
 
-	if (len > SKB_MAX_ALLOC)
+	if (len > SKB_MAX_ALLOC) {
 		data_len = min_t(size_t,
 				 len - SKB_MAX_ALLOC,
 				 MAX_SKB_FRAGS * PAGE_SIZE);
+		data_len = PAGE_ALIGN(data_len);
+	}
 
 	skb = sock_alloc_send_pskb(sk, len - data_len, data_len,
 				   msg->msg_flags & MSG_DONTWAIT, &err,
@@ -1670,6 +1672,8 @@ static int unix_stream_sendmsg(struct kiocb *kiocb, struct socket *sock,
 
 		data_len = max_t(int, 0, size - SKB_MAX_HEAD(0));
 
+		data_len = min_t(size_t, size, PAGE_ALIGN(data_len));
+
 		skb = sock_alloc_send_pskb(sk, size - data_len, data_len,
 					   msg->msg_flags & MSG_DONTWAIT, &err,
 					   get_order(UNIX_SKB_FRAGS_SZ));

Re: net: unix: Align send data_len up to PAGE_SIZE

[Eric Dumazet]

On Thu, 2014-05-15 at 18:27 +0400, Kirill Tkhai wrote:
> В Чт, 15/05/2014 в 06:27 -0700, Eric Dumazet пишет:
> > On Thu, 2014-05-15 at 12:54 +0400, Kirill Tkhai wrote:
> > 
> > > 
> > > And I assumed it's OK to allocate skbs with zero header size like this:
> > > 
> > > sock_alloc_send_pskb(sk, 0, data_len)
> > > 
> > > Please say, if it's wrong.
> > > 
> > 
> > Its OK
> > 
> 
> Thanks for the reply. So, new version is below.
> 
> [PATCH]net: unix: Align send data_len up to PAGE_SIZE
>     
> Using whole of allocated pages reduces requested skb->data size.
> This is just a little more thriftily allocation.
>     
> netperf does not show difference with the current performance.
>     
> Signed-off-by: Kirill Tkhai <ktkhai@parallels.com>
> ---
>  net/unix/af_unix.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
> index 749f80c..65d67c2 100644
> --- a/net/unix/af_unix.c
> +++ b/net/unix/af_unix.c
> @@ -1492,10 +1492,12 @@ static int unix_dgram_sendmsg(struct kiocb *kiocb, struct socket *sock,
>  	if (len > sk->sk_sndbuf - 32)
>  		goto out;
>  
> -	if (len > SKB_MAX_ALLOC)
> +	if (len > SKB_MAX_ALLOC) {
>  		data_len = min_t(size_t,
>  				 len - SKB_MAX_ALLOC,
>  				 MAX_SKB_FRAGS * PAGE_SIZE);
> +		data_len = PAGE_ALIGN(data_len);


I would add a BUILD_BUG_ON(SKB_MAX_ALLOC < PAGE_SIZE); just to make sure
nobody ever tries to 

#define SKB_MAX_ALLOC (SKB_MAX_ORDER(0, 0))

Also please send a fresh message, do not append a new patch like you
did : http://patchwork.ozlabs.org/patch/349243/

Title should be

[PATCH net-next v3] net: unix: ....