On 24/11/2021 10.18, Ong Boon Leong wrote:
Tx cycle time is in micro-seconds unit. By combining the batch size (-b M)
and Tx cycle time (-T|--tx-cycle N), xdpsock now can transmit batch-size of
packets every N-us periodically.
Does this also work for --poll mode (which is a wakeup mode) ?
For example to transmit 1 packet each 1ms cycle time for total of 2000000
packets:
$ xdpsock -i eth0 -T -N -z -T 1000 -b 1 -C 2000000
sock0@enp0s29f1:2 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 1000 1996872
sock0@enp0s29f1:2 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 1000 1997872
sock0@enp0s29f1:2 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 1000 1998872
sock0@enp0s29f1:2 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 1000 1999872
sock0@enp0s29f1:2 txonly xdp-drv
pps pkts 1.00
rx 0 0
tx 128 2000000
sock0@enp0s29f1:2 txonly xdp-drv
pps pkts 0.00
rx 0 0
tx 0 2000000
Signed-off-by: Ong Boon Leong <boon.leong.ong@xxxxxxxxx>
---
samples/bpf/xdpsock_user.c | 36 +++++++++++++++++++++++++++++++-----
1 file changed, 31 insertions(+), 5 deletions(-)
diff --git a/samples/bpf/xdpsock_user.c b/samples/bpf/xdpsock_user.c
index 691f442bbb2..61d4063f11a 100644
--- a/samples/bpf/xdpsock_user.c
+++ b/samples/bpf/xdpsock_user.c
@@ -111,6 +111,7 @@ static u32 opt_num_xsks = 1;
static u32 prog_id;
static bool opt_busy_poll;
static bool opt_reduced_cap;
+static unsigned long opt_cycle_time;
struct vlan_ethhdr {
unsigned char h_dest[6];
@@ -173,6 +174,8 @@ struct xsk_socket_info {
struct xsk_app_stats app_stats;
struct xsk_driver_stats drv_stats;
u32 outstanding_tx;
+ unsigned long prev_tx_time;
+ unsigned long tx_cycle_time;
};
static int num_socks;
@@ -972,6 +975,7 @@ static struct option long_options[] = {
{"tx-vlan-pri", required_argument, 0, 'K'},
{"tx-dmac", required_argument, 0, 'G'},
{"tx-smac", required_argument, 0, 'H'},
+ {"tx-cycle", required_argument, 0, 'T'},
{"extra-stats", no_argument, 0, 'x'},
{"quiet", no_argument, 0, 'Q'},
{"app-stats", no_argument, 0, 'a'},
@@ -1017,6 +1021,7 @@ static void usage(const char *prog)
" -K, --tx-vlan-pri=n Tx VLAN Priority [0-7]. Default: %d (For -V|--tx-vlan)\n"
" -G, --tx-dmac=<MAC> Dest MAC addr of TX frame in aa:bb:cc:dd:ee:ff format (For -V|--tx-vlan)\n"
" -H, --tx-smac=<MAC> Src MAC addr of TX frame in aa:bb:cc:dd:ee:ff format (For -V|--tx-vlan)\n"
+ " -T, --tx-cycle=n Tx cycle time in micro-seconds (For -t|--txonly).\n"
" -x, --extra-stats Display extra statistics.\n"
" -Q, --quiet Do not display any stats.\n"
" -a, --app-stats Display application (syscall) statistics.\n"
@@ -1039,7 +1044,7 @@ static void parse_command_line(int argc, char **argv)
opterr = 0;
for (;;) {
- c = getopt_long(argc, argv, "Frtli:q:pSNn:czf:muMd:b:C:s:P:VJ:K:G:H:xQaI:BR",
+ c = getopt_long(argc, argv, "Frtli:q:pSNn:czf:muMd:b:C:s:P:VJ:K:G:H:T:xQaI:BR",
long_options, &option_index);
if (c == -1)
break;
@@ -1145,6 +1150,10 @@ static void parse_command_line(int argc, char **argv)
usage(basename(argv[0]));
}
break;
+ case 'T':
+ opt_cycle_time = atoi(optarg);
+ opt_cycle_time *= 1000;
Converting to nanosec, right(?).
+ break;
case 'x':
opt_extra_stats = 1;
break;
@@ -1350,16 +1359,25 @@ static void rx_drop_all(void)
}
}
-static void tx_only(struct xsk_socket_info *xsk, u32 *frame_nb, int batch_size)
+static int tx_only(struct xsk_socket_info *xsk, u32 *frame_nb, int batch_size)
{
u32 idx;
unsigned int i;
+ if (xsk->tx_cycle_time) {
+ unsigned long now = get_nsecs();
+
+ if ((now - xsk->prev_tx_time) < xsk->tx_cycle_time)
+ return 0;
So, this test is actively spinning until the time is reached, spending
100% CPU time on this. I guess we can have this as a test for most
accurate transmit (cyclic period) with AF_XDP.
Do you have a use-case for this?
I have a customer use-case, but my customer don't want to actively spin.
My plan is to use clock_nanosleep() and wakeup slightly before the
target time and then we can spin shortly for the Tx time slot.
I will need to code this up for the customer soon anyway... perhaps we
can extend your code with this idea?
I have coded the period cycle Tx with UDP packets, here[1], if you like
to see some code using clock_nanosleep(). Next step (for me) is doing
this for AF_XDP (likely in my example[2].
[1]
https://github.com/netoptimizer/network-testing/blob/master/src/udp_pacer.c
[2]
https://github.com/xdp-project/bpf-examples/tree/master/AF_XDP-interaction
+
+ xsk->prev_tx_time = now;
Would it be valuable to know how-much we shoot "over" the tx_cycle_time?
For my use-case, I will be monitoring the other-side receiving the
packets (and using HW RX-time) to evaluate how accurate my sender is. In
this case, I would like to know if my software "knew" if was not 100%
accurate.
+ }
+
while (xsk_ring_prod__reserve(&xsk->tx, batch_size, &idx) <
batch_size) {
complete_tx_only(xsk, batch_size);
if (benchmark_done)
- return;
+ return 0;
}
I wonder if this step can introduce jitter/delay before the actual Tx
happens?
I mean, the real transmit cannot happen before xsk_ring_prod__submit()
is called. If the cycles spend are exactly the same, it doesn't matter
if you tx_cycle_time timestamp is done above.
Here you have a potential call to complete_tx_only(), which can
introduce variance for your period.
I will suggest moving the TX completion handling, so it doesn't
interfere with accurate TX.
for (i = 0; i < batch_size; i++) {
@@ -1375,6 +1393,8 @@ static void tx_only(struct xsk_socket_info *xsk, u32 *frame_nb, int batch_size)
*frame_nb += batch_size;
*frame_nb %= NUM_FRAMES;
complete_tx_only(xsk, batch_size);
+
+ return batch_size;
}
static inline int get_batch_size(int pkt_cnt)
@@ -1407,6 +1427,7 @@ static void complete_tx_only_all(void)
static void tx_only_all(void)
{
struct pollfd fds[MAX_SOCKS] = {};
+ unsigned long now = get_nsecs();
u32 frame_nb[MAX_SOCKS] = {};
int pkt_cnt = 0;
int i, ret;
@@ -1414,10 +1435,15 @@ static void tx_only_all(void)
for (i = 0; i < num_socks; i++) {
fds[0].fd = xsk_socket__fd(xsks[i]->xsk);
fds[0].events = POLLOUT;
+ if (opt_cycle_time) {
+ xsks[i]->prev_tx_time = now;
+ xsks[i]->tx_cycle_time = opt_cycle_time;
+ }
}
while ((opt_pkt_count && pkt_cnt < opt_pkt_count) || !opt_pkt_count) {
int batch_size = get_batch_size(pkt_cnt);
+ int tx_cnt = 0;
if (opt_poll) {
for (i = 0; i < num_socks; i++)
@@ -1431,9 +1457,9 @@ static void tx_only_all(void)
}
for (i = 0; i < num_socks; i++)
- tx_only(xsks[i], &frame_nb[i], batch_size);
+ tx_cnt += tx_only(xsks[i], &frame_nb[i], batch_size);
- pkt_cnt += batch_size;
+ pkt_cnt += tx_cnt;
if (benchmark_done)
break;
