linux/samples/bpf/xdp_rxq_info_user.c

582 lines
13 KiB
C
Raw Normal View History

/* SPDX-License-Identifier: GPL-2.0
* Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
*/
static const char *__doc__ = " XDP RX-queue info extract example\n\n"
"Monitor how many packets per sec (pps) are received\n"
"per NIC RX queue index and which CPU processed the packet\n"
;
#include <errno.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <stdbool.h>
#include <string.h>
#include <unistd.h>
#include <locale.h>
#include <sys/resource.h>
#include <getopt.h>
#include <net/if.h>
#include <time.h>
#include <arpa/inet.h>
#include <linux/if_link.h>
#include "bpf/bpf.h"
#include "bpf/libbpf.h"
#include "bpf_util.h"
static int ifindex = -1;
static char ifname_buf[IF_NAMESIZE];
static char *ifname;
static __u32 xdp_flags;
static struct bpf_map *stats_global_map;
static struct bpf_map *rx_queue_index_map;
/* Exit return codes */
#define EXIT_OK 0
#define EXIT_FAIL 1
#define EXIT_FAIL_OPTION 2
#define EXIT_FAIL_XDP 3
#define EXIT_FAIL_BPF 4
#define EXIT_FAIL_MEM 5
static const struct option long_options[] = {
{"help", no_argument, NULL, 'h' },
{"dev", required_argument, NULL, 'd' },
{"skb-mode", no_argument, NULL, 'S' },
{"sec", required_argument, NULL, 's' },
{"no-separators", no_argument, NULL, 'z' },
{"action", required_argument, NULL, 'a' },
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
{"readmem", no_argument, NULL, 'r' },
{"swapmac", no_argument, NULL, 'm' },
{0, 0, NULL, 0 }
};
static void int_exit(int sig)
{
fprintf(stderr,
"Interrupted: Removing XDP program on ifindex:%d device:%s\n",
ifindex, ifname);
if (ifindex > -1)
bpf_set_link_xdp_fd(ifindex, -1, xdp_flags);
exit(EXIT_OK);
}
struct config {
__u32 action;
int ifindex;
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
__u32 options;
};
enum cfg_options_flags {
NO_TOUCH = 0x0U,
READ_MEM = 0x1U,
SWAP_MAC = 0x2U,
};
#define XDP_ACTION_MAX (XDP_TX + 1)
#define XDP_ACTION_MAX_STRLEN 11
static const char *xdp_action_names[XDP_ACTION_MAX] = {
[XDP_ABORTED] = "XDP_ABORTED",
[XDP_DROP] = "XDP_DROP",
[XDP_PASS] = "XDP_PASS",
[XDP_TX] = "XDP_TX",
};
static const char *action2str(int action)
{
if (action < XDP_ACTION_MAX)
return xdp_action_names[action];
return NULL;
}
static int parse_xdp_action(char *action_str)
{
size_t maxlen;
__u64 action = -1;
int i;
for (i = 0; i < XDP_ACTION_MAX; i++) {
maxlen = XDP_ACTION_MAX_STRLEN;
if (strncmp(xdp_action_names[i], action_str, maxlen) == 0) {
action = i;
break;
}
}
return action;
}
static void list_xdp_actions(void)
{
int i;
printf("Available XDP --action <options>\n");
for (i = 0; i < XDP_ACTION_MAX; i++)
printf("\t%s\n", xdp_action_names[i]);
printf("\n");
}
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
static char* options2str(enum cfg_options_flags flag)
{
if (flag == NO_TOUCH)
return "no_touch";
if (flag & SWAP_MAC)
return "swapmac";
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
if (flag & READ_MEM)
return "read";
fprintf(stderr, "ERR: Unknown config option flags");
exit(EXIT_FAIL);
}
static void usage(char *argv[])
{
int i;
printf("\nDOCUMENTATION:\n%s\n", __doc__);
printf(" Usage: %s (options-see-below)\n", argv[0]);
printf(" Listing options:\n");
for (i = 0; long_options[i].name != 0; i++) {
printf(" --%-12s", long_options[i].name);
if (long_options[i].flag != NULL)
printf(" flag (internal value:%d)",
*long_options[i].flag);
else
printf(" short-option: -%c",
long_options[i].val);
printf("\n");
}
printf("\n");
list_xdp_actions();
}
#define NANOSEC_PER_SEC 1000000000 /* 10^9 */
static __u64 gettime(void)
{
struct timespec t;
int res;
res = clock_gettime(CLOCK_MONOTONIC, &t);
if (res < 0) {
fprintf(stderr, "Error with gettimeofday! (%i)\n", res);
exit(EXIT_FAIL);
}
return (__u64) t.tv_sec * NANOSEC_PER_SEC + t.tv_nsec;
}
/* Common stats data record shared with _kern.c */
struct datarec {
__u64 processed;
__u64 issue;
};
struct record {
__u64 timestamp;
struct datarec total;
struct datarec *cpu;
};
struct stats_record {
struct record stats;
struct record *rxq;
};
static struct datarec *alloc_record_per_cpu(void)
{
unsigned int nr_cpus = bpf_num_possible_cpus();
struct datarec *array;
size_t size;
size = sizeof(struct datarec) * nr_cpus;
array = malloc(size);
memset(array, 0, size);
if (!array) {
fprintf(stderr, "Mem alloc error (nr_cpus:%u)\n", nr_cpus);
exit(EXIT_FAIL_MEM);
}
return array;
}
static struct record *alloc_record_per_rxq(void)
{
unsigned int nr_rxqs = bpf_map__def(rx_queue_index_map)->max_entries;
struct record *array;
size_t size;
size = sizeof(struct record) * nr_rxqs;
array = malloc(size);
memset(array, 0, size);
if (!array) {
fprintf(stderr, "Mem alloc error (nr_rxqs:%u)\n", nr_rxqs);
exit(EXIT_FAIL_MEM);
}
return array;
}
static struct stats_record *alloc_stats_record(void)
{
unsigned int nr_rxqs = bpf_map__def(rx_queue_index_map)->max_entries;
struct stats_record *rec;
int i;
rec = malloc(sizeof(*rec));
memset(rec, 0, sizeof(*rec));
if (!rec) {
fprintf(stderr, "Mem alloc error\n");
exit(EXIT_FAIL_MEM);
}
rec->rxq = alloc_record_per_rxq();
for (i = 0; i < nr_rxqs; i++)
rec->rxq[i].cpu = alloc_record_per_cpu();
rec->stats.cpu = alloc_record_per_cpu();
return rec;
}
static void free_stats_record(struct stats_record *r)
{
unsigned int nr_rxqs = bpf_map__def(rx_queue_index_map)->max_entries;
int i;
for (i = 0; i < nr_rxqs; i++)
free(r->rxq[i].cpu);
free(r->rxq);
free(r->stats.cpu);
free(r);
}
static bool map_collect_percpu(int fd, __u32 key, struct record *rec)
{
/* For percpu maps, userspace gets a value per possible CPU */
unsigned int nr_cpus = bpf_num_possible_cpus();
struct datarec values[nr_cpus];
__u64 sum_processed = 0;
__u64 sum_issue = 0;
int i;
if ((bpf_map_lookup_elem(fd, &key, values)) != 0) {
fprintf(stderr,
"ERR: bpf_map_lookup_elem failed key:0x%X\n", key);
return false;
}
/* Get time as close as possible to reading map contents */
rec->timestamp = gettime();
/* Record and sum values from each CPU */
for (i = 0; i < nr_cpus; i++) {
rec->cpu[i].processed = values[i].processed;
sum_processed += values[i].processed;
rec->cpu[i].issue = values[i].issue;
sum_issue += values[i].issue;
}
rec->total.processed = sum_processed;
rec->total.issue = sum_issue;
return true;
}
static void stats_collect(struct stats_record *rec)
{
int fd, i, max_rxqs;
fd = bpf_map__fd(stats_global_map);
map_collect_percpu(fd, 0, &rec->stats);
fd = bpf_map__fd(rx_queue_index_map);
max_rxqs = bpf_map__def(rx_queue_index_map)->max_entries;
for (i = 0; i < max_rxqs; i++)
map_collect_percpu(fd, i, &rec->rxq[i]);
}
static double calc_period(struct record *r, struct record *p)
{
double period_ = 0;
__u64 period = 0;
period = r->timestamp - p->timestamp;
if (period > 0)
period_ = ((double) period / NANOSEC_PER_SEC);
return period_;
}
static __u64 calc_pps(struct datarec *r, struct datarec *p, double period_)
{
__u64 packets = 0;
__u64 pps = 0;
if (period_ > 0) {
packets = r->processed - p->processed;
pps = packets / period_;
}
return pps;
}
static __u64 calc_errs_pps(struct datarec *r,
struct datarec *p, double period_)
{
__u64 packets = 0;
__u64 pps = 0;
if (period_ > 0) {
packets = r->issue - p->issue;
pps = packets / period_;
}
return pps;
}
static void stats_print(struct stats_record *stats_rec,
struct stats_record *stats_prev,
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
int action, __u32 cfg_opt)
{
unsigned int nr_rxqs = bpf_map__def(rx_queue_index_map)->max_entries;
unsigned int nr_cpus = bpf_num_possible_cpus();
double pps = 0, err = 0;
struct record *rec, *prev;
double t;
int rxq;
int i;
/* Header */
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
printf("\nRunning XDP on dev:%s (ifindex:%d) action:%s options:%s\n",
ifname, ifindex, action2str(action), options2str(cfg_opt));
/* stats_global_map */
{
char *fmt_rx = "%-15s %-7d %'-11.0f %'-10.0f %s\n";
char *fm2_rx = "%-15s %-7s %'-11.0f\n";
char *errstr = "";
printf("%-15s %-7s %-11s %-11s\n",
"XDP stats", "CPU", "pps", "issue-pps");
rec = &stats_rec->stats;
prev = &stats_prev->stats;
t = calc_period(rec, prev);
for (i = 0; i < nr_cpus; i++) {
struct datarec *r = &rec->cpu[i];
struct datarec *p = &prev->cpu[i];
pps = calc_pps (r, p, t);
err = calc_errs_pps(r, p, t);
if (err > 0)
errstr = "invalid-ifindex";
if (pps > 0)
printf(fmt_rx, "XDP-RX CPU",
i, pps, err, errstr);
}
pps = calc_pps (&rec->total, &prev->total, t);
err = calc_errs_pps(&rec->total, &prev->total, t);
printf(fm2_rx, "XDP-RX CPU", "total", pps, err);
}
/* rx_queue_index_map */
printf("\n%-15s %-7s %-11s %-11s\n",
"RXQ stats", "RXQ:CPU", "pps", "issue-pps");
for (rxq = 0; rxq < nr_rxqs; rxq++) {
char *fmt_rx = "%-15s %3d:%-3d %'-11.0f %'-10.0f %s\n";
char *fm2_rx = "%-15s %3d:%-3s %'-11.0f\n";
char *errstr = "";
int rxq_ = rxq;
/* Last RXQ in map catch overflows */
if (rxq_ == nr_rxqs - 1)
rxq_ = -1;
rec = &stats_rec->rxq[rxq];
prev = &stats_prev->rxq[rxq];
t = calc_period(rec, prev);
for (i = 0; i < nr_cpus; i++) {
struct datarec *r = &rec->cpu[i];
struct datarec *p = &prev->cpu[i];
pps = calc_pps (r, p, t);
err = calc_errs_pps(r, p, t);
if (err > 0) {
if (rxq_ == -1)
errstr = "map-overflow-RXQ";
else
errstr = "err";
}
if (pps > 0)
printf(fmt_rx, "rx_queue_index",
rxq_, i, pps, err, errstr);
}
pps = calc_pps (&rec->total, &prev->total, t);
err = calc_errs_pps(&rec->total, &prev->total, t);
if (pps || err)
printf(fm2_rx, "rx_queue_index", rxq_, "sum", pps, err);
}
}
/* Pointer swap trick */
static inline void swap(struct stats_record **a, struct stats_record **b)
{
struct stats_record *tmp;
tmp = *a;
*a = *b;
*b = tmp;
}
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
static void stats_poll(int interval, int action, __u32 cfg_opt)
{
struct stats_record *record, *prev;
record = alloc_stats_record();
prev = alloc_stats_record();
stats_collect(record);
while (1) {
swap(&prev, &record);
stats_collect(record);
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
stats_print(record, prev, action, cfg_opt);
sleep(interval);
}
free_stats_record(record);
free_stats_record(prev);
}
int main(int argc, char **argv)
{
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
__u32 cfg_options= NO_TOUCH ; /* Default: Don't touch packet memory */
struct rlimit r = {10 * 1024 * 1024, RLIM_INFINITY};
struct bpf_prog_load_attr prog_load_attr = {
.prog_type = BPF_PROG_TYPE_XDP,
};
int prog_fd, map_fd, opt, err;
bool use_separators = true;
struct config cfg = { 0 };
struct bpf_object *obj;
struct bpf_map *map;
char filename[256];
int longindex = 0;
int interval = 2;
__u32 key = 0;
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
char action_str_buf[XDP_ACTION_MAX_STRLEN + 1 /* for \0 */] = { 0 };
int action = XDP_PASS; /* Default action */
char *action_str = NULL;
snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
prog_load_attr.file = filename;
if (setrlimit(RLIMIT_MEMLOCK, &r)) {
perror("setrlimit(RLIMIT_MEMLOCK)");
return 1;
}
if (bpf_prog_load_xattr(&prog_load_attr, &obj, &prog_fd))
return EXIT_FAIL;
map = bpf_map__next(NULL, obj);
stats_global_map = bpf_map__next(map, obj);
rx_queue_index_map = bpf_map__next(stats_global_map, obj);
if (!map || !stats_global_map || !rx_queue_index_map) {
printf("finding a map in obj file failed\n");
return EXIT_FAIL;
}
map_fd = bpf_map__fd(map);
if (!prog_fd) {
fprintf(stderr, "ERR: load_bpf_file: %s\n", strerror(errno));
return EXIT_FAIL;
}
/* Parse commands line args */
while ((opt = getopt_long(argc, argv, "hSd:",
long_options, &longindex)) != -1) {
switch (opt) {
case 'd':
if (strlen(optarg) >= IF_NAMESIZE) {
fprintf(stderr, "ERR: --dev name too long\n");
goto error;
}
ifname = (char *)&ifname_buf;
strncpy(ifname, optarg, IF_NAMESIZE);
ifindex = if_nametoindex(ifname);
if (ifindex == 0) {
fprintf(stderr,
"ERR: --dev name unknown err(%d):%s\n",
errno, strerror(errno));
goto error;
}
break;
case 's':
interval = atoi(optarg);
break;
case 'S':
xdp_flags |= XDP_FLAGS_SKB_MODE;
break;
case 'z':
use_separators = false;
break;
case 'a':
action_str = (char *)&action_str_buf;
strncpy(action_str, optarg, XDP_ACTION_MAX_STRLEN);
break;
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
case 'r':
cfg_options |= READ_MEM;
break;
case 'm':
cfg_options |= SWAP_MAC;
break;
case 'h':
error:
default:
usage(argv);
return EXIT_FAIL_OPTION;
}
}
/* Required option */
if (ifindex == -1) {
fprintf(stderr, "ERR: required option --dev missing\n");
usage(argv);
return EXIT_FAIL_OPTION;
}
cfg.ifindex = ifindex;
/* Parse action string */
if (action_str) {
action = parse_xdp_action(action_str);
if (action < 0) {
fprintf(stderr, "ERR: Invalid XDP --action: %s\n",
action_str);
list_xdp_actions();
return EXIT_FAIL_OPTION;
}
}
cfg.action = action;
/* XDP_TX requires changing MAC-addrs, else HW may drop */
if (action == XDP_TX)
cfg_options |= SWAP_MAC;
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
cfg.options = cfg_options;
/* Trick to pretty printf with thousands separators use %' */
if (use_separators)
setlocale(LC_NUMERIC, "en_US");
/* User-side setup ifindex in config_map */
err = bpf_map_update_elem(map_fd, &key, &cfg, 0);
if (err) {
fprintf(stderr, "Store config failed (err:%d)\n", err);
exit(EXIT_FAIL_BPF);
}
/* Remove XDP program when program is interrupted or killed */
signal(SIGINT, int_exit);
signal(SIGTERM, int_exit);
if (bpf_set_link_xdp_fd(ifindex, prog_fd, xdp_flags) < 0) {
fprintf(stderr, "link set xdp fd failed\n");
return EXIT_FAIL_XDP;
}
samples/bpf: extend xdp_rxq_info to read packet payload There is a cost associated with reading the packet data payload that this test ignored. Add option --read to allow enabling reading part of the payload. This sample/tool helps us analyse an issue observed with a NIC mlx5 (ConnectX-5 Ex) and an Intel(R) Xeon(R) CPU E5-1650 v4. With no_touch of data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:no_touch XDP stats CPU pps issue-pps XDP-RX CPU 0 14,465,157 0 XDP-RX CPU 1 14,464,728 0 XDP-RX CPU 2 14,465,283 0 XDP-RX CPU 3 14,465,282 0 XDP-RX CPU 4 14,464,159 0 XDP-RX CPU 5 14,465,379 0 XDP-RX CPU total 86,789,992 When not touching data, we observe that the CPUs have idle cycles. When reading data the CPUs are 100% busy in softirq. With reading data: Running XDP on dev:mlx5p1 (ifindex:8) action:XDP_DROP options:read XDP stats CPU pps issue-pps XDP-RX CPU 0 9,620,639 0 XDP-RX CPU 1 9,489,843 0 XDP-RX CPU 2 9,407,854 0 XDP-RX CPU 3 9,422,289 0 XDP-RX CPU 4 9,321,959 0 XDP-RX CPU 5 9,395,242 0 XDP-RX CPU total 56,657,828 The effect seen above is a result of cache-misses occuring when more RXQs are being used. Based on perf-event observations, our conclusion is that the CPUs DDIO (Direct Data I/O) choose to deliver packet into main memory, instead of L3-cache. We also found, that this can be mitigated by either using less RXQs or by reducing NICs the RX-ring size. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk> Acked-by: Song Liu <songliubraving@fb.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-06-25 17:27:43 +03:00
stats_poll(interval, action, cfg_options);
return EXIT_OK;
}