Compact sk_stream_mem_schedule() code

PATCH Applied Pavel Emelyanov 3 participants 3 messages Nov 19, 2007 – Nov 20, 2007

Subsystems (2)

Networking Network drivers

lore

Share & tools

lore

Pavel Emelyanov

author

2007-11-19 12:13 UTC #0 lore raw

PATCH 1/1 Compact sk_stream_mem_schedule() code

This function references sk->sk_prot->xxx for many times. It turned out, that there's so many code in it, that gcc cannot always optimize access to sk->sk_prot's fields.

After saving the sk->sk_prot on the stack and comparing disassembled code, it turned out that the function became ~10 bytes shorter and made less dereferences (on i386 and x86_64). Stack consumption didn't grow.

Besides, this patch drives most of this function into the 80 columns limit.

Signed-off-by Pavel Emelyanov <[email protected]>

---

net/core/stream.c +14 ~~-13~~

index 755bacb..b2fb846 100644
@@ -210,35 +210,36 @@ EXPORT_SYMBOL(__sk_stream_mem_reclaim);
 int sk_stream_mem_schedule(struct sock *sk, int size, int kind)
 {
 	int amt = sk_stream_pages(size);
+	struct proto *prot = sk->sk_prot;
 
 	sk->sk_forward_alloc += amt * SK_STREAM_MEM_QUANTUM;
-	atomic_add(amt, sk->sk_prot->memory_allocated);
+	atomic_add(amt, prot->memory_allocated);
 
 	/* Under limit. */
-	if (atomic_read(sk->sk_prot->memory_allocated) < sk->sk_prot->sysctl_mem[0]) {
-		if (*sk->sk_prot->memory_pressure)
-			*sk->sk_prot->memory_pressure = 0;
+	if (atomic_read(prot->memory_allocated) < prot->sysctl_mem[0]) {
+		if (*prot->memory_pressure)
+			*prot->memory_pressure = 0;
 		return 1;
 	}
 
 	/* Over hard limit. */
-	if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[2]) {
-		sk->sk_prot->enter_memory_pressure();
+	if (atomic_read(prot->memory_allocated) > prot->sysctl_mem[2]) {
+		prot->enter_memory_pressure();
 		goto suppress_allocation;
 	}
 
 	/* Under pressure. */
-	if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[1])
-		sk->sk_prot->enter_memory_pressure();
+	if (atomic_read(prot->memory_allocated) > prot->sysctl_mem[1])
+		prot->enter_memory_pressure();
 
 	if (kind) {
-		if (atomic_read(&sk->sk_rmem_alloc) < sk->sk_prot->sysctl_rmem[0])
+		if (atomic_read(&sk->sk_rmem_alloc) < prot->sysctl_rmem[0])
 			return 1;
-	} else if (sk->sk_wmem_queued < sk->sk_prot->sysctl_wmem[0])
+	} else if (sk->sk_wmem_queued < prot->sysctl_wmem[0])
 		return 1;
 
-	if (!*sk->sk_prot->memory_pressure ||
-	    sk->sk_prot->sysctl_mem[2] > atomic_read(sk->sk_prot->sockets_allocated) *
+	if (!*prot->memory_pressure ||
+	    prot->sysctl_mem[2] > atomic_read(prot->sockets_allocated) *
 				sk_stream_pages(sk->sk_wmem_queued +
 						atomic_read(&sk->sk_rmem_alloc) +
 						sk->sk_forward_alloc))

@@ -258,7 +259,7 @@ suppress_allocation:
 
 	/* Alas. Undo changes. */
 	sk->sk_forward_alloc -= amt * SK_STREAM_MEM_QUANTUM;
-	atomic_sub(amt, sk->sk_prot->memory_allocated);
+	atomic_sub(amt, prot->memory_allocated);
 	return 0;
 }

Arnaldo Carvalho de Melo

Acked-by on v1 1/1

2007-11-19 19:31 UTC #1 lore raw

Em Mon, Nov 19, 2007 at 03:13:44PM +0300, Pavel Emelyanov escreveu:

Quoted 9 lines

This function references sk->sk_prot->xxx for many times. It turned out, that there's so many code in it, that gcc cannot always optimize access to sk->sk_prot's fields.
After saving the sk->sk_prot on the stack and comparing disassembled code, it turned out that the function became ~10 bytes shorter and made less dereferences (on i386 and x86_64). Stack consumption didn't grow.

Quoted 4 lines

Besides, this patch drives most of this function into the 80 columns limit.
Signed-off-by Pavel Emelyanov <[email protected]>

I wonder if making it 'const struct proto *prot = sk->sk_prot;'

would make any difference.

Acked-by Arnaldo Carvalho de Melo <[email protected]>

David Miller

maintainer Applied

2007-11-20 07:22 UTC #2 lore raw

From: Arnaldo Carvalho de Melo <[email protected]>
Date: Mon, 19 Nov 2007 17:30:59 -0200

Quoted 10 lines

Em Mon, Nov 19, 2007 at 03:13:44PM +0300, Pavel Emelyanov escreveu:
Quoted 9 lines
This function references sk->sk_prot->xxx for many times. It turned out, that there's so many code in it, that gcc cannot always optimize access to sk->sk_prot's fields.
After saving the sk->sk_prot on the stack and comparing disassembled code, it turned out that the function became ~10 bytes shorter and made less dereferences (on i386 and x86_64). Stack consumption didn't grow.

Quoted 8 lines

Quoted 4 lines
Besides, this patch drives most of this function into the 80 columns limit.
Signed-off-by Pavel Emelyanov <[email protected]>
I wonder if making it 'const struct proto *prot = sk->sk_prot;'
would make any difference.

Such experiments are always useful, but I doubt there will be substantial gains in this case.

Quoted 1 line

Acked-by Arnaldo Carvalho de Melo <[email protected]>

I've applied the patch, thanks Pavel.