[xdp-hints] Re: [PATCH bpf-next v7 5/5] igc: Add launch time support to XDP ZC

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Tue, Feb 04, 2025 at 03:49:32PM +0100, Song, Yoong Siang wrote:
> On Tuesday, February 4, 2025 10:04 PM, Fijalkowski, Maciej <maciej.fijalkowski@intel.com> wrote:
> >On Tue, Feb 04, 2025 at 02:14:00PM +0100, Song, Yoong Siang wrote:
> >> On Tuesday, February 4, 2025 6:10 PM, Fijalkowski, Maciej<maciej.fijalkowski@intel.com> wrote:
> >> >On Tue, Feb 04, 2025 at 08:49:07AM +0800, Song Yoong Siang wrote:
> >> >
> >> >> Enable Launch Time Control (LTC) support for XDP zero copy via XDP Tx
> >> >> metadata framework.
> >> >>
> >> >> This patch has been tested with tools/testing/selftests/bpf/xdp_hw_metadata
> >> >> on Intel I225-LM Ethernet controller. Below are the test steps and result.
> >> >>
> >> >> Test 1: Send a single packet with the launch time set to 1 s in the future.
> >> >>
> >> >> Test steps:
> >> >> 1. On the DUT, start the xdp_hw_metadata selftest application:
> >> >>    $ sudo ./xdp_hw_metadata enp2s0 -l 1000000000 -L 1
> >> >>
> >> >> 2. On the Link Partner, send a UDP packet with VLAN priority 1 to port 9091
> >> >>    of the DUT.
> >> >>
> >> >> Result:
> >> >> When the launch time is set to 1 s in the future, the delta between the
> >> >> launch time and the transmit hardware timestamp is 0.016 us, as shown in
> >> >> printout of the xdp_hw_metadata application below.
> >> >>   0x562ff5dc8880: rx_desc[4]->addr=84110 addr=84110 comp_addr=84110 EoP
> >> >>   rx_hash: 0xE343384 with RSS type:0x1
> >> >>   HW RX-time:   1734578015467548904 (sec:1734578015.4675)
> >> >>                 delta to User RX-time sec:0.0002 (183.103 usec)
> >> >>   XDP RX-time:   1734578015467651698 (sec:1734578015.4677)
> >> >>                  delta to User RX-time sec:0.0001 (80.309 usec)
> >> >>   No rx_vlan_tci or rx_vlan_proto, err=-95
> >> >>   0x562ff5dc8880: ping-pong with csum=561c (want c7dd)
> >> >>                   csum_start=34 csum_offset=6
> >> >>   HW RX-time:   1734578015467548904 (sec:1734578015.4675)
> >> >>                 delta to HW Launch-time sec:1.0000 (1000000.000 usec)
> >> >>   0x562ff5dc8880: complete tx idx=4 addr=4018
> >> >>   HW Launch-time:   1734578016467548904 (sec:1734578016.4675)
> >> >>                     delta to HW TX-complete-time sec:0.0000 (0.016 usec)
> >> >>   HW TX-complete-time:   1734578016467548920 (sec:1734578016.4675)
> >> >>                          delta to User TX-complete-time sec:0.0000
> >> >>                          (32.546 usec)
> >> >>   XDP RX-time:   1734578015467651698 (sec:1734578015.4677)
> >> >>                  delta to User TX-complete-time sec:0.9999
> >> >>                  (999929.768 usec)
> >> >>   HW RX-time:   1734578015467548904 (sec:1734578015.4675)
> >> >>                 delta to HW TX-complete-time sec:1.0000 (1000000.016 usec)
> >> >>   0x562ff5dc8880: complete rx idx=132 addr=84110
> >> >>
> >> >> Test 2: Send 1000 packets with a 10 ms interval and the launch time set to
> >> >>         500 us in the future.
> >> >>
> >> >> Test steps:
> >> >> 1. On the DUT, start the xdp_hw_metadata selftest application:
> >> >>    $ sudo chrt -f 99 ./xdp_hw_metadata enp2s0 -l 500000 -L 1 > \
> >> >>      /dev/shm/result.log
> >> >>
> >> >> 2. On the Link Partner, send 1000 UDP packets with a 10 ms interval and
> >> >>    VLAN priority 1 to port 9091 of the DUT.
> >> >>
> >> >> Result:
> >> >> When the launch time is set to 500 us in the future, the average delta
> >> >> between the launch time and the transmit hardware timestamp is 0.016 us,
> >> >> as shown in the analysis of /dev/shm/result.log below. The XDP launch time
> >> >> works correctly in sending 1000 packets continuously.
> >> >>   Min delta: 0.005 us
> >> >>   Avr delta: 0.016 us
> >> >>   Max delta: 0.031 us
> >> >>   Total packets forwarded: 1000
> >> >>
> >> >> Signed-off-by: Song Yoong Siang <yoong.siang.song@intel.com>
> >> >> ---
> >> >>  drivers/net/ethernet/intel/igc/igc_main.c | 42 +++++++++++++++++++++--
> >> >>  1 file changed, 40 insertions(+), 2 deletions(-)
> >> >>
> >> >> diff --git a/drivers/net/ethernet/intel/igc/igc_main.c
> >> >b/drivers/net/ethernet/intel/igc/igc_main.c
> >> >> index c3edd8bcf633..535d340c71c9 100644
> >> >> --- a/drivers/net/ethernet/intel/igc/igc_main.c
> >> >> +++ b/drivers/net/ethernet/intel/igc/igc_main.c
> >> >> @@ -2951,9 +2951,33 @@ static u64 igc_xsk_fill_timestamp(void *_priv)
> >> >>  	return *(u64 *)_priv;
> >> >>  }
> >> >>
> >> >> +static void igc_xsk_request_launch_time(u64 launch_time, void *_priv)
> >> >> +{
> >> >> +	struct igc_metadata_request *meta_req = _priv;
> >> >> +	struct igc_ring *tx_ring = meta_req->tx_ring;
> >> >> +	__le32 launch_time_offset;
> >> >> +	bool insert_empty = false;
> >> >> +	bool first_flag = false;
> >> >> +
> >> >> +	if (!tx_ring->launchtime_enable)
> >> >> +		return;
> >> >> +
> >> >> +	launch_time_offset = igc_tx_launchtime(tx_ring,
> >> >> +					       ns_to_ktime(launch_time),
> >> >> +					       &first_flag, &insert_empty);
> >> >> +	if (insert_empty) {
> >> >> +		igc_insert_empty_packet(tx_ring);
> >> >> +		meta_req->tx_buffer =
> >> >> +			&tx_ring->tx_buffer_info[tx_ring->next_to_use];
> >> >
> >> >in this case I think you currently are leaking the skbs and dma mappings
> >> >that igc_init_empty_frame() did. you're going to mix
> >> >IGC_TX_BUFFER_TYPE_XSK with IGC_TX_BUFFER_TYPE_SKB and the latter is not
> >> >explicitly initialized. Even though IGC_TX_BUFFER_TYPE_SKB happens to be
> >> >equal to 0, igc_tx_buffer::type is never cleared in the tx clean desc
> >> >routine.
> >> >
> >>
> >> Hi Fijalkowski Maciej,
> >>
> >> Thanks for your inputs.
> >>
> >> Yes, you are right, IGC_TX_BUFFER_TYPE_SKB is mixed together with
> >> IGC_TX_BUFFER_TYPE_XSK. Regarding the skb and dma map,
> >> following code in igc_clean_tx_irq() will free the skb and unmap the dma,
> >> Do these answer your concern on leaking?
> >>
> >> igc_main.c:3133:                case IGC_TX_BUFFER_TYPE_SKB:
> >> igc_main.c-3134-                        napi_consume_skb(tx_buffer->skb, napi_budget);
> >> igc_main.c-3135-                        igc_unmap_tx_buffer(tx_ring->dev, tx_buffer);
> >> igc_main.c-3136-                        break;
> >>
> >> Regarding the igc_tx_buffer::type never cleared, I think the
> >> important thing is making the igc_tx_buffer::next_to_watch NULL
> >> to indicate no remaining packet. Since transmit function will
> >> always set the igc_tx_buffer::type to a proper type,
> >
> >igc_init_tx_empty_descriptor() does not set ::type, that was my point. So
> >these empty descs might be treated as IGC_TX_BUFFER_TYPE_XSK, which is
> >wrong and your qouted code above will never get called. You will then leak
> >skb and dma map plus you will confuse XSK code due to xsk_frames miscount.
> >
> 
> Now I understand what you mean. Thanks for the explanation. I will add the
> missing IGC_TX_BUFFER_TYPE_SKB initialization in igc_init_empty_frame(),
> as below.
> 
> --- a/drivers/net/ethernet/intel/igc/igc_main.c
> +++ b/drivers/net/ethernet/intel/igc/igc_main.c
> @@ -1096,6 +1096,7 @@ static int igc_init_empty_frame(struct igc_ring *ring,
>                 return -ENOMEM;
>         }
> 
> +       buffer->type = IGC_TX_BUFFER_TYPE_SKB;
>         buffer->skb = skb;
>         buffer->protocol = 0;
>         buffer->bytecount = skb->len;
> 
> With above, IMHO, we no need to clear igc_tx_buffer::type,
> Are you agree?

Yes, the contract should be that every routine that produces descriptors
should be setting the ::type explicitly. Sorry for not being clear from
beginning but I'm glad we're on the same page now.

I'm afraid this single line should be send as a fix, though :< even
without your patch set empty descs could be mixed IGC_TX_BUFFER_TYPE_XDP
type.

> 
> >> I think it is optional for us to clear it.
> >> Is that make sense to you?
> >>
> >> >> +	}
> >> >> +
> >> >> +	igc_tx_ctxtdesc(tx_ring, launch_time_offset, first_flag, 0, 0, 0);
> >> >> +}
> >> >> +
> >> >>  const struct xsk_tx_metadata_ops igc_xsk_tx_metadata_ops = {
> >> >>  	.tmo_request_timestamp		= igc_xsk_request_timestamp,
> >> >>  	.tmo_fill_timestamp		= igc_xsk_fill_timestamp,
> >> >> +	.tmo_request_launch_time	= igc_xsk_request_launch_time,
> >> >>  };
> >> >>
> >> >>  static void igc_xdp_xmit_zc(struct igc_ring *ring)
> >> >> @@ -2976,7 +3000,13 @@ static void igc_xdp_xmit_zc(struct igc_ring *ring)
> >> >>  	ntu = ring->next_to_use;
> >> >>  	budget = igc_desc_unused(ring);
> >> >>
> >> >> -	while (xsk_tx_peek_desc(pool, &xdp_desc) && budget--) {
> >> >> +	/* Packets with launch time require one data descriptor and one context
> >> >> +	 * descriptor. When the launch time falls into the next Qbv cycle, we
> >> >> +	 * may need to insert an empty packet, which requires two more
> >> >> +	 * descriptors. Therefore, to be safe, we always ensure we have at least
> >> >> +	 * 4 descriptors available.
> >> >> +	 */
> >> >> +	while (xsk_tx_peek_desc(pool, &xdp_desc) && budget >= 4) {
> >> >>  		struct igc_metadata_request meta_req;
> >> >>  		struct xsk_tx_metadata *meta = NULL;
> >> >>  		struct igc_tx_buffer *bi;
> >> >> @@ -3000,6 +3030,12 @@ static void igc_xdp_xmit_zc(struct igc_ring *ring)
> >> >>  		xsk_tx_metadata_request(meta, &igc_xsk_tx_metadata_ops,
> >> >>  					&meta_req);
> >> >>
> >> >> +		/* xsk_tx_metadata_request() may have updated next_to_use */
> >> >> +		ntu = ring->next_to_use;
> >> >> +
> >> >> +		/* xsk_tx_metadata_request() may have updated Tx buffer info */
> >> >> +		bi = meta_req.tx_buffer;
> >> >> +
> >> >>  		tx_desc = IGC_TX_DESC(ring, ntu);
> >> >>  		tx_desc->read.cmd_type_len = cpu_to_le32(meta_req.cmd_type);
> >> >>  		tx_desc->read.olinfo_status = cpu_to_le32(olinfo_status);
> >> >> @@ -3017,9 +3053,11 @@ static void igc_xdp_xmit_zc(struct igc_ring *ring)
> >> >>  		ntu++;
> >> >>  		if (ntu == ring->count)
> >> >>  			ntu = 0;
> >> >> +
> >> >> +		ring->next_to_use = ntu;
> >> >> +		budget = igc_desc_unused(ring);
> >> >
> >> >why count the remaining space in loop? couldn't you decrement it
> >> >accordingly to the count of descriptors you have produced? writing ntu
> >> >back and forth between local var and ring struct performance-wise does not
> >> >look fine.
> >> >
> >>
> >> Yes, I can check the number of used descriptor in xsk_tx_metadata_request()
> >> by introducing a new field named used_desc in struct igc_metadata_request,
> >> and then decreases the budget with it.
> >>
> >> Do this way looked good to you?
> >
> >Yes this makes sense to me, thanks!
> >
> 
> Thanks, I will rework and submit next version.
> 
> >>
> >> Thanks & Regards
> >> Siang
> >>
> >> >>  	}
> >> >>
> >> >> -	ring->next_to_use = ntu;
> >> >>  	if (tx_desc) {
> >> >>  		igc_flush_tx_descriptors(ring);
> >> >>  		xsk_tx_release(pool);
> >> >> --
> >> >> 2.34.1
> >> >>
> 
> Thanks & Regards
> Siang
>