Re: [PATCH v2] net/sctp/ulpevent.c: Deinline sctp_ulpevent_set_owner, save 1616 bytes

[Marcelo Ricardo Leitner]

On Wed, Jun 21, 2017 at 07:03:27PM +0200, Denys Vlasenko wrote:
> This function compiles to 147 bytes of machine code. 13 callsites.
> 
> I'm no expert on SCTP events, but quick reading of SCTP docs tells me that
> SCTP events are not happening on every packet.
> They are ASSOC_CHANGE, PEER_ADDR_CHANGE, REMOTE_ERROR and such.
> Does not look performance critical.
> 
> Signed-off-by: Denys Vlasenko 
> CC: Vlad Yasevich 
> CC: Neil Horman 
> CC: David Miller 
> CC: linux-s...@vger.kernel.org
> CC: netdev@vger.kernel.org
> CC: linux-ker...@vger.kernel.org
> ---
> Changed since v1:
> * reindented function argument list

Dave, this patch is marked as Changes Requested on patchwork, but the v2
here looks good. There was no change request on it, only on the v1, and
they were satified on v2. The only info I asked on v2 was to know he had
a bigger plan for this, not sure if that's what caused confusion or not.

Thanks,
Marcelo

Re: [PATCH v2] net/sctp/ulpevent.c: Deinline sctp_ulpevent_set_owner, save 1616 bytes

[Denys Vlasenko]

On 06/21/2017 09:24 PM, Marcelo Ricardo Leitner wrote:

On Wed, Jun 21, 2017 at 07:03:27PM +0200, Denys Vlasenko wrote:

This function compiles to 147 bytes of machine code. 13 callsites.

I'm no expert on SCTP events, but quick reading of SCTP docs tells me that
SCTP events are not happening on every packet.
They are ASSOC_CHANGE, PEER_ADDR_CHANGE, REMOTE_ERROR and such.
Does not look performance critical.

Signed-off-by: Denys Vlasenko 
CC: Vlad Yasevich 
CC: Neil Horman 
CC: David Miller 
CC: linux-s...@vger.kernel.org
CC: netdev@vger.kernel.org
CC: linux-ker...@vger.kernel.org


Acked-by: Marcelo Ricardo Leitner 

Just wondering, are you conducting further research on this topic?
Because we probably could use such review on SCTP stack.


Here is the list of sctp inlines which has any (however small)
benefit when deinlined.

filename:lineno:inline_name:lines_of_source_code:saved_bytes_by_deinlining:size_of_code_of_deinlined_function

include/net/sctp/command.h:228:sctp_add_cmd_sf:7:8306:38
net/sctp/ulpevent.c:91:sctp_ulpevent_set_owner:13:1616:147
include/net/sctp/sctp.h:414:sctp_skb_set_owner_r:10:934:75
net/sctp/input.c:840:sctp_hash_key:13:896:359
net/sctp/input.c:823:sctp_hash_obj:13:704:409
include/net/sctp/checksum.h:60:sctp_compute_cksum:13:595:85
net/sctp/input.c:800:sctp_hash_cmp:18:320:124
net/sctp/socket.c:117:sctp_wspace:19:256:76
include/net/sctp/sctp.h:272:sctp_max_rto:7:204:68
net/sctp/socket.c:173:sctp_verify_addr:15:192:72
include/net/sctp/checksum.h:46:sctp_csum_update:4:147:21
include/net/sctp/sctp.h:519:param_type2af:8:134:35
include/net/sctp/sctp.h:399:sctp_list_dequeue:7:123:59
include/net/sctp/sctp.h:596:sctp_transport_dst_check:4:120:60
include/net/sctp/sctp.h:435:sctp_frag_point:12:65:65
net/sctp/outqueue.c:82:sctp_outq_dequeue_data:10:64:87
include/net/sctp/command.h:243:sctp_next_cmd:4:64:37
include/net/sctp/sm.h:347:sctp_data_size:6:19:19

For .c files, the patches are trivial and I have them auto-generated,
I'll send them to you privately to save you the mechanical work.

Unfortunately, for inlines in .h files this does not work
(a human is needed to decide where to more the function).

Of course, not every deinlining makes sense.

Re: [PATCH v2] net/sctp/ulpevent.c: Deinline sctp_ulpevent_set_owner, save 1616 bytes

[Marcelo Ricardo Leitner]

On Wed, Jun 21, 2017 at 07:03:27PM +0200, Denys Vlasenko wrote:
> This function compiles to 147 bytes of machine code. 13 callsites.
> 
> I'm no expert on SCTP events, but quick reading of SCTP docs tells me that
> SCTP events are not happening on every packet.
> They are ASSOC_CHANGE, PEER_ADDR_CHANGE, REMOTE_ERROR and such.
> Does not look performance critical.
> 
> Signed-off-by: Denys Vlasenko 
> CC: Vlad Yasevich 
> CC: Neil Horman 
> CC: David Miller 
> CC: linux-s...@vger.kernel.org
> CC: netdev@vger.kernel.org
> CC: linux-ker...@vger.kernel.org

Acked-by: Marcelo Ricardo Leitner 

Just wondering, are you conducting further research on this topic?
Because we probably could use such review on SCTP stack.

> ---
> Changed since v1:
> * reindented function argument list
> 
>  net/sctp/ulpevent.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/net/sctp/ulpevent.c b/net/sctp/ulpevent.c
> index ec2b3e0..bc3f495 100644
> --- a/net/sctp/ulpevent.c
> +++ b/net/sctp/ulpevent.c
> @@ -88,8 +88,8 @@ int sctp_ulpevent_is_notification(const struct 
> sctp_ulpevent *event)
>  /* Hold the association in case the msg_name needs read out of
>   * the association.
>   */
> -static inline void sctp_ulpevent_set_owner(struct sctp_ulpevent *event,
> -const struct sctp_association *asoc)
> +static void sctp_ulpevent_set_owner(struct sctp_ulpevent *event,
> + const struct sctp_association *asoc)
>  {
>   struct sctp_chunk *chunk = event->chunk;
>   struct sk_buff *skb;
> -- 
> 1.8.3.1
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-sctp" in
> the body of a message to majord...@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 

[PATCH v2] net/sctp/ulpevent.c: Deinline sctp_ulpevent_set_owner, save 1616 bytes

[Denys Vlasenko]

This function compiles to 147 bytes of machine code. 13 callsites.

I'm no expert on SCTP events, but quick reading of SCTP docs tells me that
SCTP events are not happening on every packet.
They are ASSOC_CHANGE, PEER_ADDR_CHANGE, REMOTE_ERROR and such.
Does not look performance critical.

Signed-off-by: Denys Vlasenko 
CC: Vlad Yasevich 
CC: Neil Horman 
CC: David Miller 
CC: linux-s...@vger.kernel.org
CC: netdev@vger.kernel.org
CC: linux-ker...@vger.kernel.org
---
Changed since v1:
* reindented function argument list

 net/sctp/ulpevent.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/sctp/ulpevent.c b/net/sctp/ulpevent.c
index ec2b3e0..bc3f495 100644
--- a/net/sctp/ulpevent.c
+++ b/net/sctp/ulpevent.c
@@ -88,8 +88,8 @@ int sctp_ulpevent_is_notification(const struct sctp_ulpevent 
*event)
 /* Hold the association in case the msg_name needs read out of
  * the association.
  */
-static inline void sctp_ulpevent_set_owner(struct sctp_ulpevent *event,
-  const struct sctp_association *asoc)
+static void sctp_ulpevent_set_owner(struct sctp_ulpevent *event,
+   const struct sctp_association *asoc)
 {
struct sctp_chunk *chunk = event->chunk;
struct sk_buff *skb;
-- 
1.8.3.1