diff mbox series

[RFT,net-next] net: ti: add pp skb recycling support

Message ID ef808c4d8447ee8cf832821a985ba68939455461.1623239847.git.lorenzo@kernel.org (mailing list archive)
State Superseded
Delegated to: Netdev Maintainers
Headers show
Series [RFT,net-next] net: ti: add pp skb recycling support | expand

Checks

Context Check Description
netdev/cover_letter success Link
netdev/fixes_present success Link
netdev/patch_count success Link
netdev/tree_selection success Clearly marked for net-next
netdev/subject_prefix success Link
netdev/cc_maintainers warning 9 maintainers not CCed: zhangchangzhong@huawei.com michael@walle.cc daniel@iogearbox.net camelia.groza@nxp.com shayagr@amazon.com m-karicheri2@ti.com yanaijie@huawei.com linux-omap@vger.kernel.org jesse.brandeburg@intel.com
netdev/source_inline success Was 0 now: 0
netdev/verify_signedoff success Link
netdev/module_param success Was 0 now: 0
netdev/build_32bit success Errors and warnings before: 0 this patch: 0
netdev/kdoc success Errors and warnings before: 0 this patch: 0
netdev/verify_fixes success Link
netdev/checkpatch success total: 0 errors, 0 warnings, 0 checks, 20 lines checked
netdev/build_allmodconfig_warn success Errors and warnings before: 0 this patch: 0
netdev/header_inline success Link

Commit Message

Lorenzo Bianconi June 9, 2021, 12:01 p.m. UTC
As already done for mvneta and mvpp2, enable skb recycling for ti
ethernet drivers

Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
This patch has just compile-tested
---
 drivers/net/ethernet/ti/cpsw.c     | 4 ++--
 drivers/net/ethernet/ti/cpsw_new.c | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

Comments

Matteo Croce June 9, 2021, 12:20 p.m. UTC | #1
On Wed, Jun 9, 2021 at 2:01 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:
>
> As already done for mvneta and mvpp2, enable skb recycling for ti
> ethernet drivers
>
> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>

Looks good! If someone with the HW could provide a with and without
the patch, that would be nice!
Grygorii Strashko June 9, 2021, 3:02 p.m. UTC | #2
hi

On 09/06/2021 15:20, Matteo Croce wrote:
> On Wed, Jun 9, 2021 at 2:01 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:
>>
>> As already done for mvneta and mvpp2, enable skb recycling for ti
>> ethernet drivers
>>
>> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> 
> Looks good! If someone with the HW could provide a with and without
> the patch, that would be nice!
> 

What test would you recommend to run?
Lorenzo Bianconi June 9, 2021, 3:12 p.m. UTC | #3
> hi
> 
> On 09/06/2021 15:20, Matteo Croce wrote:
> > On Wed, Jun 9, 2021 at 2:01 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> > > 
> > > As already done for mvneta and mvpp2, enable skb recycling for ti
> > > ethernet drivers
> > > 
> > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > 
> > Looks good! If someone with the HW could provide a with and without
> > the patch, that would be nice!
> > 
> 
> What test would you recommend to run?

Hi Grygorii,

Just pass the skb to the stack, so I think regular XDP_PASS use case is enough.

Regards,
Lorenzo

> 
> -- 
> Best regards,
> grygorii
>
Matteo Croce June 9, 2021, 3:43 p.m. UTC | #4
On Wed, Jun 9, 2021 at 5:03 PM Grygorii Strashko
<grygorii.strashko@ti.com> wrote:
>
> hi
>
> On 09/06/2021 15:20, Matteo Croce wrote:
> > On Wed, Jun 9, 2021 at 2:01 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> >>
> >> As already done for mvneta and mvpp2, enable skb recycling for ti
> >> ethernet drivers
> >>
> >> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> >
> > Looks good! If someone with the HW could provide a with and without
> > the patch, that would be nice!
> >
>
> What test would you recommend to run?
>
> --
> Best regards,
> grygorii

Hi Grygorii,

A test which benefits most from this kind of change is one in which
the frames are freed early.
One option would be to use mausezahn which by default sends frames
with an invalid ethertype, that are dropped very early from the stack
(I think in __netif_receive_skb_core() or near there).
Then, on the device I just watch the device statistics to count
packets per second:

mausezahn eth0 -c 0 -b $board_mac_address

This test should be precise enough for a gigabit link, on faster links
I usually use a DPDK or AF_XDP based one.

Regards,
Jesper Dangaard Brouer June 9, 2021, 3:55 p.m. UTC | #5
On Wed, 9 Jun 2021 17:43:57 +0200
Matteo Croce <mcroce@linux.microsoft.com> wrote:

> On Wed, Jun 9, 2021 at 5:03 PM Grygorii Strashko
> <grygorii.strashko@ti.com> wrote:
> >
> > On 09/06/2021 15:20, Matteo Croce wrote:  
> > > On Wed, Jun 9, 2021 at 2:01 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:  
> > >>
> > >> As already done for mvneta and mvpp2, enable skb recycling for ti
> > >> ethernet drivers
> > >>
> > >> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>  
> > >
> > > Looks good! If someone with the HW could provide a with and without
> > > the patch, that would be nice!
> > >  
> >
> > What test would you recommend to run?
> >
[...]
> 
> A test which benefits most from this kind of change is one in which
> the frames are freed early.

I would also recommend running an XDP_PASS program, and then running
something that let the packets travel as deep as possible into netstack.
Not to test performance, but to make sure we didn't break something!

I've hacked up bnxt driver (it's not as straight forward as this driver
to convert) and is running some TCP tests.  Not problems so-far :-0

I wanted to ask if someone knows howto setup the zero-copy TCP stuff
that google did? (but is that TX only?)
Eric Dumazet June 9, 2021, 4:58 p.m. UTC | #6
On 6/9/21 5:55 PM, Jesper Dangaard Brouer wrote:
> On Wed, 9 Jun 2021 17:43:57 +0200
> Matteo Croce <mcroce@linux.microsoft.com> wrote:
> 
>> On Wed, Jun 9, 2021 at 5:03 PM Grygorii Strashko
>> <grygorii.strashko@ti.com> wrote:
>>>
>>> On 09/06/2021 15:20, Matteo Croce wrote:  
>>>> On Wed, Jun 9, 2021 at 2:01 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:  
>>>>>
>>>>> As already done for mvneta and mvpp2, enable skb recycling for ti
>>>>> ethernet drivers
>>>>>
>>>>> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>  
>>>>
>>>> Looks good! If someone with the HW could provide a with and without
>>>> the patch, that would be nice!
>>>>  
>>>
>>> What test would you recommend to run?
>>>
> [...]
>>
>> A test which benefits most from this kind of change is one in which
>> the frames are freed early.
> 
> I would also recommend running an XDP_PASS program, and then running
> something that let the packets travel as deep as possible into netstack.
> Not to test performance, but to make sure we didn't break something!
> 
> I've hacked up bnxt driver (it's not as straight forward as this driver
> to convert) and is running some TCP tests.  Not problems so-far :-0
> 
> I wanted to ask if someone knows howto setup the zero-copy TCP stuff
> that google did? (but is that TX only?)
> 

TX does not need anything fancy really.

For RX part, you can look at https://netdevconf.info/0x14/pub/slides/62/Implementing%20TCP%20RX%20zero%20copy.pdf
Grygorii Strashko June 9, 2021, 5:03 p.m. UTC | #7
On 09/06/2021 15:20, Matteo Croce wrote:
> On Wed, Jun 9, 2021 at 2:01 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:
>>
>> As already done for mvneta and mvpp2, enable skb recycling for ti
>> ethernet drivers
>>
>> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> 
> Looks good! If someone with the HW could provide a with and without
> the patch, that would be nice!
> 

Not sure to which mail to answer, so answering here - thanks all

1) I've simulated packet drop using iperf
Host:
- arp -s <some IP> <unknown MAC>
- iperf -c <some IP> -u -l60 -b700M -t60 -i1

DUP:
- place interface in promisc mode
- check rx_packets stats

I see big improvement ~47Kpps vs ~64Kpps

2) I've run iperf3 tests - see no regressions, but also not too much improvements

3) I've applied 2 patches
- this one
- and [1]

Results are below, Thank you

Tested-by: Grygorii Strashko <grygorii.strashko@ti.com>
Reviewed-by: Grygorii Strashko <grygorii.strashko@ti.com>

[1] https://patchwork.kernel.org/project/netdevbpf/patch/20210609103326.278782-18-toke@redhat.com/

=========== Before:
[perf top]
  47.15%  [kernel]                   [k] _raw_spin_unlock_irqrestore
   11.77%  [kernel]                   [k] __cpdma_chan_free
    3.16%  [kernel]                   [k] ___bpf_prog_run
    2.52%  [kernel]                   [k] cpsw_rx_vlan_encap
    2.34%  [kernel]                   [k] __netif_receive_skb_core
    2.27%  [kernel]                   [k] free_unref_page
    2.26%  [kernel]                   [k] kmem_cache_free
    2.24%  [kernel]                   [k] kmem_cache_alloc
    1.69%  [kernel]                   [k] __softirqentry_text_start
    1.61%  [kernel]                   [k] cpsw_rx_handler
    1.19%  [kernel]                   [k] page_pool_release_page
    1.19%  [kernel]                   [k] clear_bits_ll
    1.15%  [kernel]                   [k] page_frag_free
    1.06%  [kernel]                   [k] __dma_page_dev_to_cpu
    0.99%  [kernel]                   [k] memset
    0.94%  [kernel]                   [k] __alloc_pages_bulk
    0.92%  [kernel]                   [k] kfree_skb
    0.85%  [kernel]                   [k] packet_rcv
    0.78%  [kernel]                   [k] page_address
    0.75%  [kernel]                   [k] v7_dma_inv_range
    0.71%  [kernel]                   [k] __lock_text_start

[rx packets - with packet drop]
rxdiff 48004
rxdiff 47630
rxdiff 47538Š
…
[iperf3 TCP]
iperf3 -c 192.168.1.1 -i1
[  5]   0.00-10.00  sec   873 MBytes   732 Mbits/sec    0             sender
[  5]   0.00-10.01  sec   866 MBytes   726 Mbits/sec                  receiver

=========== After:
[perf top - with packet drop]
  40.58%  [kernel]                   [k] _raw_spin_unlock_irqrestore
   16.18%  [kernel]                   [k] __softirqentry_text_start
   10.33%  [kernel]                   [k] __cpdma_chan_free
    2.62%  [kernel]                   [k] ___bpf_prog_run
    2.05%  [kernel]                   [k] cpsw_rx_vlan_encap
    2.00%  [kernel]                   [k] kmem_cache_alloc
    1.86%  [kernel]                   [k] __netif_receive_skb_core
    1.80%  [kernel]                   [k] kmem_cache_free
    1.63%  [kernel]                   [k] cpsw_rx_handler
    1.12%  [kernel]                   [k] cpsw_rx_mq_poll
    1.11%  [kernel]                   [k] page_pool_put_page
    1.04%  [kernel]                   [k] _raw_spin_unlock
    0.97%  [kernel]                   [k] clear_bits_ll
    0.90%  [kernel]                   [k] packet_rcv
    0.88%  [kernel]                   [k] __dma_page_dev_to_cpu
    0.85%  [kernel]                   [k] kfree_skb
    0.80%  [kernel]                   [k] memset
    0.71%  [kernel]                   [k] __lock_text_start
    0.66%  [kernel]                   [k] v7_dma_inv_range
    0.64%  [kernel]                   [k] gen_pool_free_owner
    0.58%  [kernel]                   [k] __rcu_read_unlock

[rx packets - with packet drop]
rxdiff 65843
rxdiff 66722
rxdiff 65264

[iperf3 TCP]
iperf3 -c 192.168.1.1 -i1
[  5]   0.00-10.00  sec   884 MBytes   742 Mbits/sec    0             sender
[  5]   0.00-10.01  sec   878 MBytes   735 Mbits/sec                  receiver
diff mbox series

Patch

diff --git a/drivers/net/ethernet/ti/cpsw.c b/drivers/net/ethernet/ti/cpsw.c
index c0cd7de88316..049508667a6d 100644
--- a/drivers/net/ethernet/ti/cpsw.c
+++ b/drivers/net/ethernet/ti/cpsw.c
@@ -430,8 +430,8 @@  static void cpsw_rx_handler(void *token, int len, int status)
 		cpts_rx_timestamp(cpsw->cpts, skb);
 	skb->protocol = eth_type_trans(skb, ndev);
 
-	/* unmap page as no netstack skb page recycling */
-	page_pool_release_page(pool, page);
+	/* mark skb for recycling */
+	skb_mark_for_recycle(skb, page, pool);
 	netif_receive_skb(skb);
 
 	ndev->stats.rx_bytes += len;
diff --git a/drivers/net/ethernet/ti/cpsw_new.c b/drivers/net/ethernet/ti/cpsw_new.c
index 69b7a4e0220a..11f536138495 100644
--- a/drivers/net/ethernet/ti/cpsw_new.c
+++ b/drivers/net/ethernet/ti/cpsw_new.c
@@ -373,8 +373,8 @@  static void cpsw_rx_handler(void *token, int len, int status)
 		cpts_rx_timestamp(cpsw->cpts, skb);
 	skb->protocol = eth_type_trans(skb, ndev);
 
-	/* unmap page as no netstack skb page recycling */
-	page_pool_release_page(pool, page);
+	/* mark skb for recycling */
+	skb_mark_for_recycle(skb, page, pool);
 	netif_receive_skb(skb);
 
 	ndev->stats.rx_bytes += len;