event/octeontx2: remove WFE from dualslot dequeue
diff mbox series

Message ID 20200214064525.1895-1-pbhagavatula@marvell.com
State Accepted
Delegated to: Jerin Jacob
Headers show
Series
  • event/octeontx2: remove WFE from dualslot dequeue
Related show

Checks

Context Check Description
ci/Intel-compilation success Compilation OK
ci/travis-robot success Travis build: passed
ci/iol-mellanox-Performance success Performance Testing PASS
ci/iol-testing success Testing PASS
ci/checkpatch warning coding style issues

Commit Message

Pavan Nikhilesh Bhagavatula Feb. 14, 2020, 6:45 a.m. UTC
From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Each workslot is always bound to a specific lcore there is no multi-core
contention to cause cache trashing as a result it is safe to remove the
WFE. Also, in dual workslot dequeue work will mostlikely be available on
the pair workslot making WFE impractical.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---

Also, this in-turn reduces the branch misses

Before:
	0 arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,branch_filter=1,jitter=1,min_latency=0/
	0 dummy:u
	0 llc-miss
	0 tlb-miss
	853 branch-miss
	0 remote-access
	0 l1d-miss

After:
	0 arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,branch_filter=1,jitter=1,min_latency=0/
	0 dummy:u
	0 llc-miss
	0 tlb-miss
	250 branch-miss
	0 remote-access
	0 l1d-miss

WFE Data:

0x4C40 - WFI_WFE_WAIT_CYCLES - Number of cycles waiting at a WFI or WFE instruction.

- WFE Cycles before the patch for Dual workslot
#perf stat -C 20 -e r4C40 sleep 1
Performance counter stats for 'CPU(s) 20':

               264      r4C40
       1.002494168 seconds time elapsed

- WFE Cycles for single workslot
#perf stat -C 20 -e r4C40 sleep 1
Performance counter stats for 'CPU(s) 20':

       908,778,351      r4C40
       1.002598253 seconds time elapsed

 drivers/event/octeontx2/otx2_worker_dual.h | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

--
2.17.1

Comments

Gavin Hu Feb. 15, 2020, 5:56 a.m. UTC | #1
Hi Pavan,

> -----Original Message-----
> From: pbhagavatula@marvell.com <pbhagavatula@marvell.com>
> Sent: Friday, February 14, 2020 2:45 PM
> To: jerinj@marvell.com; Pavan Nikhilesh <pbhagavatula@marvell.com>
> Cc: Gavin Hu <Gavin.Hu@arm.com>; dev@dpdk.org
> Subject: [dpdk-dev] [PATCH] event/octeontx2: remove WFE from dualslot
> dequeue
> 
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> 
> Each workslot is always bound to a specific lcore there is no multi-core
> contention to cause cache trashing as a result it is safe to remove the
> WFE. Also, in dual workslot dequeue work will mostlikely be available on
> the pair workslot making WFE impractical.

Does SSO still signal EVENTI to exit from WFE?  Then the core ignore it?
Can this be disabled as WFE is removed?  

> 
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> ---
> 
> Also, this in-turn reduces the branch misses
> 
> Before:
> 	0
> arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,branch_filter=1,jitter=1,
> min_latency=0/
> 	0 dummy:u
> 	0 llc-miss
> 	0 tlb-miss
> 	853 branch-miss
> 	0 remote-access
> 	0 l1d-miss
> 
> After:
> 	0
> arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,branch_filter=1,jitter=1,
> min_latency=0/
> 	0 dummy:u
> 	0 llc-miss
> 	0 tlb-miss
> 	250 branch-miss
> 	0 remote-access
> 	0 l1d-miss
> 
> WFE Data:
> 
> 0x4C40 - WFI_WFE_WAIT_CYCLES - Number of cycles waiting at a WFI or
> WFE instruction.
> 
> - WFE Cycles before the patch for Dual workslot
> #perf stat -C 20 -e r4C40 sleep 1
> Performance counter stats for 'CPU(s) 20':
> 
>                264      r4C40
>        1.002494168 seconds time elapsed
> 
> - WFE Cycles for single workslot
> #perf stat -C 20 -e r4C40 sleep 1
> Performance counter stats for 'CPU(s) 20':
> 
>        908,778,351      r4C40
>        1.002598253 seconds time elapsed
> 
>  drivers/event/octeontx2/otx2_worker_dual.h | 6 +-----
>  1 file changed, 1 insertion(+), 5 deletions(-)
> 
> diff --git a/drivers/event/octeontx2/otx2_worker_dual.h
> b/drivers/event/octeontx2/otx2_worker_dual.h
> index 5134e3d52..c88420eb4 100644
> --- a/drivers/event/octeontx2/otx2_worker_dual.h
> +++ b/drivers/event/octeontx2/otx2_worker_dual.h
> @@ -29,11 +29,7 @@ otx2_ssogws_dual_get_work(struct
> otx2_ssogws_state *ws,
>  		rte_prefetch_non_temporal(lookup_mem);
>  #ifdef RTE_ARCH_ARM64
>  	asm volatile(
> -			"        ldr %[tag], [%[tag_loc]]    \n"
> -			"        ldr %[wqp], [%[wqp_loc]]    \n"
> -			"        tbz %[tag], 63, done%=      \n"
> -			"        sevl                        \n"
> -			"rty%=:  wfe                         \n"
> +			"rty%=:	                             \n"
>  			"        ldr %[tag], [%[tag_loc]]    \n"
>  			"        ldr %[wqp], [%[wqp_loc]]    \n"
>  			"        tbnz %[tag], 63, rty%=      \n"
> --
> 2.17.1
Pavan Nikhilesh Bhagavatula Feb. 15, 2020, 9:43 a.m. UTC | #2
Hi Gavin,

>Hi Pavan,
>
>> -----Original Message-----
>> From: pbhagavatula@marvell.com <pbhagavatula@marvell.com>
>> Sent: Friday, February 14, 2020 2:45 PM
>> To: jerinj@marvell.com; Pavan Nikhilesh
><pbhagavatula@marvell.com>
>> Cc: Gavin Hu <Gavin.Hu@arm.com>; dev@dpdk.org
>> Subject: [dpdk-dev] [PATCH] event/octeontx2: remove WFE from
>dualslot
>> dequeue
>>
>> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>>
>> Each workslot is always bound to a specific lcore there is no multi-core
>> contention to cause cache trashing as a result it is safe to remove the
>> WFE. Also, in dual workslot dequeue work will mostlikely be available
>on
>> the pair workslot making WFE impractical.
>
>Does SSO still signal EVENTI to exit from WFE?  Then the core ignore it?

All transactions on SSO bus take the core out of WFE.

>Can this be disabled as WFE is removed?

This can't be disabled.

>
>>
>> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
>> ---
>>
>> Also, this in-turn reduces the branch misses
>>
>> Before:
>> 	0
>>
>arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,branch_filter=1,jit
>ter=1,
>> min_latency=0/
>> 	0 dummy:u
>> 	0 llc-miss
>> 	0 tlb-miss
>> 	853 branch-miss
>> 	0 remote-access
>> 	0 l1d-miss
>>
>> After:
>> 	0
>>
>arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,branch_filter=1,jit
>ter=1,
>> min_latency=0/
>> 	0 dummy:u
>> 	0 llc-miss
>> 	0 tlb-miss
>> 	250 branch-miss
>> 	0 remote-access
>> 	0 l1d-miss
>>
>> WFE Data:
>>
>> 0x4C40 - WFI_WFE_WAIT_CYCLES - Number of cycles waiting at a WFI
>or
>> WFE instruction.
>>
>> - WFE Cycles before the patch for Dual workslot
>> #perf stat -C 20 -e r4C40 sleep 1
>> Performance counter stats for 'CPU(s) 20':
>>
>>                264      r4C40
>>        1.002494168 seconds time elapsed
>>
>> - WFE Cycles for single workslot
>> #perf stat -C 20 -e r4C40 sleep 1
>> Performance counter stats for 'CPU(s) 20':
>>
>>        908,778,351      r4C40
>>        1.002598253 seconds time elapsed
>>
>>  drivers/event/octeontx2/otx2_worker_dual.h | 6 +-----
>>  1 file changed, 1 insertion(+), 5 deletions(-)
>>
>> diff --git a/drivers/event/octeontx2/otx2_worker_dual.h
>> b/drivers/event/octeontx2/otx2_worker_dual.h
>> index 5134e3d52..c88420eb4 100644
>> --- a/drivers/event/octeontx2/otx2_worker_dual.h
>> +++ b/drivers/event/octeontx2/otx2_worker_dual.h
>> @@ -29,11 +29,7 @@ otx2_ssogws_dual_get_work(struct
>> otx2_ssogws_state *ws,
>>  		rte_prefetch_non_temporal(lookup_mem);
>>  #ifdef RTE_ARCH_ARM64
>>  	asm volatile(
>> -			"        ldr %[tag], [%[tag_loc]]    \n"
>> -			"        ldr %[wqp], [%[wqp_loc]]    \n"
>> -			"        tbz %[tag], 63, done%=      \n"
>> -			"        sevl                        \n"
>> -			"rty%=:  wfe                         \n"
>> +			"rty%=:	                             \n"
>>  			"        ldr %[tag], [%[tag_loc]]    \n"
>>  			"        ldr %[wqp], [%[wqp_loc]]    \n"
>>  			"        tbnz %[tag], 63, rty%=      \n"
>> --
>> 2.17.1
Gavin Hu Feb. 15, 2020, 12:23 p.m. UTC | #3
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
Jerin Jacob Feb. 15, 2020, 1:51 p.m. UTC | #4
On Sat, Feb 15, 2020 at 5:53 PM Gavin Hu <Gavin.Hu@arm.com> wrote:
>
> Reviewed-by: Gavin Hu <gavin.hu@arm.com>

Applied to dpdk-next-eventdev/master. Thanks.

Patch
diff mbox series

diff --git a/drivers/event/octeontx2/otx2_worker_dual.h b/drivers/event/octeontx2/otx2_worker_dual.h
index 5134e3d52..c88420eb4 100644
--- a/drivers/event/octeontx2/otx2_worker_dual.h
+++ b/drivers/event/octeontx2/otx2_worker_dual.h
@@ -29,11 +29,7 @@  otx2_ssogws_dual_get_work(struct otx2_ssogws_state *ws,
 		rte_prefetch_non_temporal(lookup_mem);
 #ifdef RTE_ARCH_ARM64
 	asm volatile(
-			"        ldr %[tag], [%[tag_loc]]    \n"
-			"        ldr %[wqp], [%[wqp_loc]]    \n"
-			"        tbz %[tag], 63, done%=      \n"
-			"        sevl                        \n"
-			"rty%=:  wfe                         \n"
+			"rty%=:	                             \n"
 			"        ldr %[tag], [%[tag_loc]]    \n"
 			"        ldr %[wqp], [%[wqp_loc]]    \n"
 			"        tbnz %[tag], 63, rty%=      \n"