[bpf-next,v3,3/3] selftests/bpf: add Launch Time request to xdp_hw_metadata
Commit Message
This patch adds Launch Time hw offload request to xdp_hw_metadata. User can
configure the delta of HW launch time to HW RX-time by using "-l" argument.
The default delta is set to 0.1 second.
This patch is tested with stmmac on Intel Tiger Lake platform. Refer to
result below, the delta between pre-determined launch time and actual HW
transmit complete time is around 24 us.
$ sudo ./xdp_hw_metadata eth0
...
xsk_ring_cons__peek: 1
0x55e577c3a7a8: rx_desc[0]->addr=80100 addr=80100 comp_addr=80100 EoP
No rx_hash err=-95
HW RX-time: 1677762523393813392 (sec:1677762523.3938) delta to User RX-time sec:0.0003 (259.290 usec)
XDP RX-time: 1677762523394050576 (sec:1677762523.3941) delta to User RX-time sec:0.0000 (22.106 usec)
0x55e577c3a7a8: ping-pong with csum=5619 (want 8626) csum_start=34 csum_offset=6
HW RX-time: 1677762523393813392 (sec:1677762523.3938) delta to HW Launch-time sec:0.1000 (100000.000 usec)
0x55e577c3a7a8: complete tx idx=0 addr=18
HW Launch-time: 1677762523493813392 (sec:1677762523.4938) delta to HW TX-complete-time sec:0.0000 (24.181 usec)
HW TX-complete-time: 1677762523493837573 (sec:1677762523.4938) delta to User TX-complete-time sec:0.0007 (737.636 usec)
XDP RX-time: 1677762523394050576 (sec:1677762523.3941) delta to User TX-complete-time sec:0.1005 (100524.633 usec)
HW RX-time: 1677762523393813392 (sec:1677762523.3938) delta to HW TX-complete-time sec:0.1000 (100024.181 usec)
0x55e577c3a7a8: complete rx idx=128 addr=80100
$ sudo ./xdp_hw_metadata eth0 -l 10000000
...
poll: 1 (0) skip=17 fail=0 redir=17
xsk_ring_cons__peek: 1
0x558336d397a8: rx_desc[15]->addr=9e100 addr=9e100 comp_addr=9e100 EoP
No rx_hash err=-95
HW RX-time: 1677762699254666655 (sec:1677762699.2547) delta to User RX-time sec:0.0003 (256.928 usec)
XDP RX-time: 1677762699254901232 (sec:1677762699.2549) delta to User RX-time sec:0.0000 (22.351 usec)
0x558336d397a8: ping-pong with csum=5619 (want 8626) csum_start=34 csum_offset=6
HW RX-time: 1677762699254666655 (sec:1677762699.2547) delta to HW Launch-time sec:0.0100 (10000.000 usec)
0x558336d397a8: complete tx idx=15 addr=f018
HW Launch-time: 1677762699264666655 (sec:1677762699.2647) delta to HW TX-complete-time sec:0.0000 (24.307 usec)
HW TX-complete-time: 1677762699264690962 (sec:1677762699.2647) delta to User TX-complete-time sec:0.0003 (309.901 usec)
XDP RX-time: 1677762699254901232 (sec:1677762699.2549) delta to User TX-complete-time sec:0.0101 (10099.631 usec)
HW RX-time: 1677762699254666655 (sec:1677762699.2547) delta to HW TX-complete-time sec:0.0100 (10024.307 usec)
0x558336d397a8: complete rx idx=143 addr=9e100
Signed-off-by: Song Yoong Siang <yoong.siang.song@intel.com>
---
tools/testing/selftests/bpf/xdp_hw_metadata.c | 19 ++++++++++++++++++-
1 file changed, 18 insertions(+), 1 deletion(-)
Comments
On 12/3/23 17:51, Song Yoong Siang wrote:
> This patch is tested with stmmac on Intel Tiger Lake platform. Refer to
> result below, the delta between pre-determined launch time and actual HW
> transmit complete time is around 24 us.
Is there any NIC setup (e.g. ethtool/qdisc) requirements to enable HW
for this feature?
--Jesper
Jesper Dangaard Brouer wrote:
>
>
> On 12/3/23 17:51, Song Yoong Siang wrote:
> > This patch is tested with stmmac on Intel Tiger Lake platform. Refer to
> > result below, the delta between pre-determined launch time and actual HW
> > transmit complete time is around 24 us.
>
> Is there any NIC setup (e.g. ethtool/qdisc) requirements to enable HW
> for this feature?
Judging from how we currently use this with FQ and ETF, no.
See for instance tools/testing/selftests/net/so_txtime.sh
On Monday, December 4, 2023 11:00 PM, Willem de Bruijn wrote:
>Jesper Dangaard Brouer wrote:
>>
>>
>> On 12/3/23 17:51, Song Yoong Siang wrote:
>> > This patch is tested with stmmac on Intel Tiger Lake platform. Refer to
>> > result below, the delta between pre-determined launch time and actual HW
>> > transmit complete time is around 24 us.
>>
>> Is there any NIC setup (e.g. ethtool/qdisc) requirements to enable HW
>> for this feature?
>
>Judging from how we currently use this with FQ and ETF, no.
>
>See for instance tools/testing/selftests/net/so_txtime.sh
For stmmac, we need to enable per queue launch time support
by using tc qdisc etf offload command.
I believe igc also using the same way.
I will add into documentation to make it clear that
XDP_TXMD_FLAGS_LAUNCH_TIME only used to pass the
launch time value, not enable launch time feature.
@@ -13,6 +13,7 @@
* - UDP 9091 packets trigger TX reply
* - TX HW timestamp is requested and reported back upon completion
* - TX checksum is requested
+ * - HW launch time is set for transmission
*/
#include <test_progs.h>
@@ -61,6 +62,8 @@ int rxq;
bool skip_tx;
__u64 last_hw_rx_timestamp;
__u64 last_xdp_rx_timestamp;
+__u64 last_launch_time;
+__u64 launch_time_delta_to_hw_rx_timestamp = 100000000; /* 0.1 second */
void test__fail(void) { /* for network_helpers.c */ }
@@ -274,6 +277,8 @@ static bool complete_tx(struct xsk *xsk, clockid_t clock_id)
if (meta->completion.tx_timestamp) {
__u64 ref_tstamp = gettime(clock_id);
+ print_tstamp_delta("HW Launch-time", "HW TX-complete-time",
+ last_launch_time, meta->completion.tx_timestamp);
print_tstamp_delta("HW TX-complete-time", "User TX-complete-time",
meta->completion.tx_timestamp, ref_tstamp);
print_tstamp_delta("XDP RX-time", "User TX-complete-time",
@@ -371,6 +376,14 @@ static void ping_pong(struct xsk *xsk, void *rx_packet, clockid_t clock_id)
xsk, ntohs(udph->check), ntohs(want_csum),
meta->request.csum_start, meta->request.csum_offset);
+ /* Set the value of launch time */
+ meta->flags |= XDP_TXMD_FLAGS_LAUNCH_TIME;
+ meta->request.launch_time = last_hw_rx_timestamp +
+ launch_time_delta_to_hw_rx_timestamp;
+ last_launch_time = meta->request.launch_time;
+ print_tstamp_delta("HW RX-time", "HW Launch-time", last_hw_rx_timestamp,
+ meta->request.launch_time);
+
memcpy(data, rx_packet, len); /* don't share umem chunk for simplicity */
tx_desc->options |= XDP_TX_METADATA;
tx_desc->len = len;
@@ -595,6 +608,7 @@ static void print_usage(void)
" -h Display this help and exit\n\n"
" -m Enable multi-buffer XDP for larger MTU\n"
" -r Don't generate AF_XDP reply (rx metadata only)\n"
+ " -l Delta of HW Launch-time to HW RX-time in ns (default: 0.1s)\n"
"Generate test packets on the other machine with:\n"
" echo -n xdp | nc -u -q1 <dst_ip> 9091\n";
@@ -605,7 +619,7 @@ static void read_args(int argc, char *argv[])
{
int opt;
- while ((opt = getopt(argc, argv, "chmr")) != -1) {
+ while ((opt = getopt(argc, argv, "chmrl:")) != -1) {
switch (opt) {
case 'c':
bind_flags &= ~XDP_USE_NEED_WAKEUP;
@@ -621,6 +635,9 @@ static void read_args(int argc, char *argv[])
case 'r':
skip_tx = true;
break;
+ case 'l':
+ launch_time_delta_to_hw_rx_timestamp = atoll(optarg);
+ break;
case '?':
if (isprint(optopt))
fprintf(stderr, "Unknown option: -%c\n", optopt);