[3/4] flowcache: make struct flow_cache_percpu::hash_rnd_recalc bool

Submitted by Alexey Dobriyan on March 19, 2017, 10:27 p.m.

Details

Message ID 20170319222743.GC17015@avx2
State Changes Requested
Delegated to: David Miller
Headers show

Commit Message

Alexey Dobriyan March 19, 2017, 10:27 p.m.
->hash_rnd_recalc is only used in boolean context.

Space savings on x86_64 come from the fact that "MOV rm8, imm8" is
shorter than "MOV rm32, imm32" by at least 3 bytes.

	add/remove: 0/0 grow/shrink: 0/3 up/down: 0/-10 (-10)
	function                                     old     new   delta
	flow_cache_new_hashrnd                       166     163      -3
	flow_cache_cpu_up_prep                       171     168      -3
	flow_cache_lookup                           1148    1144      -4
	Total: Before=170822872, After=170822862, chg -0.00%

Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
---

 include/net/flowcache.h |    2 +-
 net/core/flow.c         |    6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

Comments

Eric Dumazet March 19, 2017, 11:11 p.m.
On Mon, 2017-03-20 at 01:27 +0300, Alexey Dobriyan wrote:
> ->hash_rnd_recalc is only used in boolean context.
> 
> Space savings on x86_64 come from the fact that "MOV rm8, imm8" is
> shorter than "MOV rm32, imm32" by at least 3 bytes.
> 
> 	add/remove: 0/0 grow/shrink: 0/3 up/down: 0/-10 (-10)
> 	function                                     old     new   delta
> 	flow_cache_new_hashrnd                       166     163      -3
> 	flow_cache_cpu_up_prep                       171     168      -3
> 	flow_cache_lookup                           1148    1144      -4
> 	Total: Before=170822872, After=170822862, chg -0.00%
> 
> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
> ---
> 
>  include/net/flowcache.h |    2 +-
>  net/core/flow.c         |    6 +++---
>  2 files changed, 4 insertions(+), 4 deletions(-)
> 
> --- a/include/net/flowcache.h
> +++ b/include/net/flowcache.h
> @@ -10,7 +10,7 @@ struct flow_cache_percpu {
>  	struct hlist_head		*hash_table;
>  	int				hash_count;
>  	u32				hash_rnd;
> -	int				hash_rnd_recalc;
> +	bool				hash_rnd_recalc;

hash_rnd_recalc can be written from flow_cache_new_hashrnd() without any
locking.

Some arches do not have the ability to store an u8 atomically.

So your patch adds a possibility that in the future, we might have a
bug, if another field is added there.

Basically the whole integer here should be reserved, or converted to a
real flag (that can be manipulated with clear_bit() and set_bit()),
but this would consume 8 bytes instead of 4 ;)
David Miller March 22, 2017, 2:09 a.m.
From: Alexey Dobriyan <adobriyan@gmail.com>
Date: Mon, 20 Mar 2017 01:27:43 +0300

> ->hash_rnd_recalc is only used in boolean context.
> 
> Space savings on x86_64 come from the fact that "MOV rm8, imm8" is
> shorter than "MOV rm32, imm32" by at least 3 bytes.
> 
> 	add/remove: 0/0 grow/shrink: 0/3 up/down: 0/-10 (-10)
> 	function                                     old     new   delta
> 	flow_cache_new_hashrnd                       166     163      -3
> 	flow_cache_cpu_up_prep                       171     168      -3
> 	flow_cache_lookup                           1148    1144      -4
> 	Total: Before=170822872, After=170822862, chg -0.00%
> 
> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>

I agree with Eric Dumazet that we might have atomicity issues in the
future because of this change.

Why don't you drop this and resubmit just the other 3 patches which
seem to be much less controversial?

Thanks.
Alexey Dobriyan March 22, 2017, 10:38 a.m.
On Wed, Mar 22, 2017 at 5:09 AM, David Miller <davem@davemloft.net> wrote:
> From: Alexey Dobriyan <adobriyan@gmail.com>
> Date: Mon, 20 Mar 2017 01:27:43 +0300
>
>> ->hash_rnd_recalc is only used in boolean context.
>>
>> Space savings on x86_64 come from the fact that "MOV rm8, imm8" is
>> shorter than "MOV rm32, imm32" by at least 3 bytes.
>>
>>       add/remove: 0/0 grow/shrink: 0/3 up/down: 0/-10 (-10)
>>       function                                     old     new   delta
>>       flow_cache_new_hashrnd                       166     163      -3
>>       flow_cache_cpu_up_prep                       171     168      -3
>>       flow_cache_lookup                           1148    1144      -4
>>       Total: Before=170822872, After=170822862, chg -0.00%
>>
>> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
>
> I agree with Eric Dumazet that we might have atomicity issues in the
> future because of this change.
>
> Why don't you drop this and resubmit just the other 3 patches which
> seem to be much less controversial?

Sure, bool patch was last minute change.

Patch hide | download patch | download mbox

--- a/include/net/flowcache.h
+++ b/include/net/flowcache.h
@@ -10,7 +10,7 @@  struct flow_cache_percpu {
 	struct hlist_head		*hash_table;
 	int				hash_count;
 	u32				hash_rnd;
-	int				hash_rnd_recalc;
+	bool				hash_rnd_recalc;
 	struct tasklet_struct		flush_tasklet;
 };
 
--- a/net/core/flow.c
+++ b/net/core/flow.c
@@ -56,7 +56,7 @@  static void flow_cache_new_hashrnd(unsigned long arg)
 	int i;
 
 	for_each_possible_cpu(i)
-		per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
+		per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = true;
 
 	fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 	add_timer(&fc->rnd_timer);
@@ -155,7 +155,7 @@  static void flow_new_hash_rnd(struct flow_cache *fc,
 			      struct flow_cache_percpu *fcp)
 {
 	get_random_bytes(&fcp->hash_rnd, sizeof(u32));
-	fcp->hash_rnd_recalc = 0;
+	fcp->hash_rnd_recalc = false;
 	__flow_cache_shrink(fc, fcp, 0);
 }
 
@@ -412,7 +412,7 @@  static int flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
 			pr_err("NET: failed to allocate flow cache sz %u\n", sz);
 			return -ENOMEM;
 		}
-		fcp->hash_rnd_recalc = 1;
+		fcp->hash_rnd_recalc = true;
 		fcp->hash_count = 0;
 		tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 	}