samples/bpf/xdp_redirect_cpu_user.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright(c) 2017 Jesper Dangaard Brouer, Red Hat, Inc.
   3  */
   4 static const char *__doc__ =
   5         " XDP redirect with a CPU-map type \"BPF_MAP_TYPE_CPUMAP\"";
   6
   7 #include <errno.h>
   8 #include <signal.h>
   9 #include <stdio.h>
  10 #include <stdlib.h>
  11 #include <stdbool.h>
  12 #include <string.h>
  13 #include <unistd.h>
  14 #include <locale.h>
  15 #include <sys/resource.h>
  16 #include <sys/sysinfo.h>
  17 #include <getopt.h>
  18 #include <net/if.h>
  19 #include <time.h>
  20 #include <linux/limits.h>
  21
  22 #include <arpa/inet.h>
  23 #include <linux/if_link.h>
  24
  25 /* How many xdp_progs are defined in _kern.c */
  26 #define MAX_PROG 6
  27
  28 #include <bpf/bpf.h>
  29 #include <bpf/libbpf.h>
  30
  31 #include "bpf_util.h"
  32
  33 static int ifindex = -1;
  34 static char ifname_buf[IF_NAMESIZE];
  35 static char *ifname;
  36 static __u32 prog_id;
  37
  38 static __u32 xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
  39 static int n_cpus;
  40 static int cpu_map_fd;
  41 static int rx_cnt_map_fd;
  42 static int redirect_err_cnt_map_fd;
  43 static int cpumap_enqueue_cnt_map_fd;
  44 static int cpumap_kthread_cnt_map_fd;
  45 static int cpus_available_map_fd;
  46 static int cpus_count_map_fd;
  47 static int cpus_iterator_map_fd;
  48 static int exception_cnt_map_fd;
  49
  50 #define NUM_TP 5
  51 struct bpf_link *tp_links[NUM_TP] = { 0 };
  52 static int tp_cnt = 0;
  53
  54 /* Exit return codes */
  55 #define EXIT_OK         0
  56 #define EXIT_FAIL               1
  57 #define EXIT_FAIL_OPTION        2
  58 #define EXIT_FAIL_XDP           3
  59 #define EXIT_FAIL_BPF           4
  60 #define EXIT_FAIL_MEM           5
  61
  62 static const struct option long_options[] = {
  63         {"help",        no_argument,            NULL, 'h' },
  64         {"dev",         required_argument,      NULL, 'd' },
  65         {"skb-mode",    no_argument,            NULL, 'S' },
  66         {"sec",         required_argument,      NULL, 's' },
  67         {"progname",    required_argument,      NULL, 'p' },
  68         {"qsize",       required_argument,      NULL, 'q' },
  69         {"cpu",         required_argument,      NULL, 'c' },
  70         {"stress-mode", no_argument,            NULL, 'x' },
  71         {"no-separators", no_argument,          NULL, 'z' },
  72         {"force",       no_argument,            NULL, 'F' },
  73         {0, 0, NULL,  0 }
  74 };
  75
  76 static void int_exit(int sig)
  77 {
  78         __u32 curr_prog_id = 0;
  79
  80         if (ifindex > -1) {
  81                 if (bpf_get_link_xdp_id(ifindex, &curr_prog_id, xdp_flags)) {
  82                         printf("bpf_get_link_xdp_id failed\n");
  83                         exit(EXIT_FAIL);
  84                 }
  85                 if (prog_id == curr_prog_id) {
  86                         fprintf(stderr,
  87                                 "Interrupted: Removing XDP program on ifindex:%d device:%s\n",
  88                                 ifindex, ifname);
  89                         bpf_set_link_xdp_fd(ifindex, -1, xdp_flags);
  90                 } else if (!curr_prog_id) {
  91                         printf("couldn't find a prog id on a given iface\n");
  92                 } else {
  93                         printf("program on interface changed, not removing\n");
  94                 }
  95         }
  96         /* Detach tracepoints */
  97         while (tp_cnt)
  98                 bpf_link__destroy(tp_links[--tp_cnt]);
  99
 100         exit(EXIT_OK);
 101 }
 102
 103 static void print_avail_progs(struct bpf_object *obj)
 104 {
 105         struct bpf_program *pos;
 106
 107         bpf_object__for_each_program(pos, obj) {
 108                 if (bpf_program__is_xdp(pos))
 109                         printf(" %s\n", bpf_program__title(pos, false));
 110         }
 111 }
 112
 113 static void usage(char *argv[], struct bpf_object *obj)
 114 {
 115         int i;
 116
 117         printf("\nDOCUMENTATION:\n%s\n", __doc__);
 118         printf("\n");
 119         printf(" Usage: %s (options-see-below)\n", argv[0]);
 120         printf(" Listing options:\n");
 121         for (i = 0; long_options[i].name != 0; i++) {
 122                 printf(" --%-12s", long_options[i].name);
 123                 if (long_options[i].flag != NULL)
 124                         printf(" flag (internal value:%d)",
 125                                 *long_options[i].flag);
 126                 else
 127                         printf(" short-option: -%c",
 128                                 long_options[i].val);
 129                 printf("\n");
 130         }
 131         printf("\n Programs to be used for --progname:\n");
 132         print_avail_progs(obj);
 133         printf("\n");
 134 }
 135
 136 /* gettime returns the current time of day in nanoseconds.
 137  * Cost: clock_gettime (ns) => 26ns (CLOCK_MONOTONIC)
 138  *       clock_gettime (ns) =>  9ns (CLOCK_MONOTONIC_COARSE)
 139  */
 140 #define NANOSEC_PER_SEC 1000000000 /* 10^9 */
 141 static __u64 gettime(void)
 142 {
 143         struct timespec t;
 144         int res;
 145
 146         res = clock_gettime(CLOCK_MONOTONIC, &t);
 147         if (res < 0) {
 148                 fprintf(stderr, "Error with gettimeofday! (%i)\n", res);
 149                 exit(EXIT_FAIL);
 150         }
 151         return (__u64) t.tv_sec * NANOSEC_PER_SEC + t.tv_nsec;
 152 }
 153
 154 /* Common stats data record shared with _kern.c */
 155 struct datarec {
 156         __u64 processed;
 157         __u64 dropped;
 158         __u64 issue;
 159 };
 160 struct record {
 161         __u64 timestamp;
 162         struct datarec total;
 163         struct datarec *cpu;
 164 };
 165 struct stats_record {
 166         struct record rx_cnt;
 167         struct record redir_err;
 168         struct record kthread;
 169         struct record exception;
 170         struct record enq[];
 171 };
 172
 173 static bool map_collect_percpu(int fd, __u32 key, struct record *rec)
 174 {
 175         /* For percpu maps, userspace gets a value per possible CPU */
 176         unsigned int nr_cpus = bpf_num_possible_cpus();
 177         struct datarec values[nr_cpus];
 178         __u64 sum_processed = 0;
 179         __u64 sum_dropped = 0;
 180         __u64 sum_issue = 0;
 181         int i;
 182
 183         if ((bpf_map_lookup_elem(fd, &key, values)) != 0) {
 184                 fprintf(stderr,
 185                         "ERR: bpf_map_lookup_elem failed key:0x%X\n", key);
 186                 return false;
 187         }
 188         /* Get time as close as possible to reading map contents */
 189         rec->timestamp = gettime();
 190
 191         /* Record and sum values from each CPU */
 192         for (i = 0; i < nr_cpus; i++) {
 193                 rec->cpu[i].processed = values[i].processed;
 194                 sum_processed        += values[i].processed;
 195                 rec->cpu[i].dropped = values[i].dropped;
 196                 sum_dropped        += values[i].dropped;
 197                 rec->cpu[i].issue = values[i].issue;
 198                 sum_issue        += values[i].issue;
 199         }
 200         rec->total.processed = sum_processed;
 201         rec->total.dropped   = sum_dropped;
 202         rec->total.issue     = sum_issue;
 203         return true;
 204 }
 205
 206 static struct datarec *alloc_record_per_cpu(void)
 207 {
 208         unsigned int nr_cpus = bpf_num_possible_cpus();
 209         struct datarec *array;
 210
 211         array = calloc(nr_cpus, sizeof(struct datarec));
 212         if (!array) {
 213                 fprintf(stderr, "Mem alloc error (nr_cpus:%u)\n", nr_cpus);
 214                 exit(EXIT_FAIL_MEM);
 215         }
 216         return array;
 217 }
 218
 219 static struct stats_record *alloc_stats_record(void)
 220 {
 221         struct stats_record *rec;
 222         int i, size;
 223
 224         size = sizeof(*rec) + n_cpus * sizeof(struct record);
 225         rec = malloc(size);
 226         if (!rec) {
 227                 fprintf(stderr, "Mem alloc error\n");
 228                 exit(EXIT_FAIL_MEM);
 229         }
 230         memset(rec, 0, size);
 231         rec->rx_cnt.cpu    = alloc_record_per_cpu();
 232         rec->redir_err.cpu = alloc_record_per_cpu();
 233         rec->kthread.cpu   = alloc_record_per_cpu();
 234         rec->exception.cpu = alloc_record_per_cpu();
 235         for (i = 0; i < n_cpus; i++)
 236                 rec->enq[i].cpu = alloc_record_per_cpu();
 237
 238         return rec;
 239 }
 240
 241 static void free_stats_record(struct stats_record *r)
 242 {
 243         int i;
 244
 245         for (i = 0; i < n_cpus; i++)
 246                 free(r->enq[i].cpu);
 247         free(r->exception.cpu);
 248         free(r->kthread.cpu);
 249         free(r->redir_err.cpu);
 250         free(r->rx_cnt.cpu);
 251         free(r);
 252 }
 253
 254 static double calc_period(struct record *r, struct record *p)
 255 {
 256         double period_ = 0;
 257         __u64 period = 0;
 258
 259         period = r->timestamp - p->timestamp;
 260         if (period > 0)
 261                 period_ = ((double) period / NANOSEC_PER_SEC);
 262
 263         return period_;
 264 }
 265
 266 static __u64 calc_pps(struct datarec *r, struct datarec *p, double period_)
 267 {
 268         __u64 packets = 0;
 269         __u64 pps = 0;
 270
 271         if (period_ > 0) {
 272                 packets = r->processed - p->processed;
 273                 pps = packets / period_;
 274         }
 275         return pps;
 276 }
 277
 278 static __u64 calc_drop_pps(struct datarec *r, struct datarec *p, double period_)
 279 {
 280         __u64 packets = 0;
 281         __u64 pps = 0;
 282
 283         if (period_ > 0) {
 284                 packets = r->dropped - p->dropped;
 285                 pps = packets / period_;
 286         }
 287         return pps;
 288 }
 289
 290 static __u64 calc_errs_pps(struct datarec *r,
 291                             struct datarec *p, double period_)
 292 {
 293         __u64 packets = 0;
 294         __u64 pps = 0;
 295
 296         if (period_ > 0) {
 297                 packets = r->issue - p->issue;
 298                 pps = packets / period_;
 299         }
 300         return pps;
 301 }
 302
 303 static void stats_print(struct stats_record *stats_rec,
 304                         struct stats_record *stats_prev,
 305                         char *prog_name)
 306 {
 307         unsigned int nr_cpus = bpf_num_possible_cpus();
 308         double pps = 0, drop = 0, err = 0;
 309         struct record *rec, *prev;
 310         int to_cpu;
 311         double t;
 312         int i;
 313
 314         /* Header */
 315         printf("Running XDP/eBPF prog_name:%s\n", prog_name);
 316         printf("%-15s %-7s %-14s %-11s %-9s\n",
 317                "XDP-cpumap", "CPU:to", "pps", "drop-pps", "extra-info");
 318
 319         /* XDP rx_cnt */
 320         {
 321                 char *fmt_rx = "%-15s %-7d %'-14.0f %'-11.0f %'-10.0f %s\n";
 322                 char *fm2_rx = "%-15s %-7s %'-14.0f %'-11.0f\n";
 323                 char *errstr = "";
 324
 325                 rec  = &stats_rec->rx_cnt;
 326                 prev = &stats_prev->rx_cnt;
 327                 t = calc_period(rec, prev);
 328                 for (i = 0; i < nr_cpus; i++) {
 329                         struct datarec *r = &rec->cpu[i];
 330                         struct datarec *p = &prev->cpu[i];
 331
 332                         pps = calc_pps(r, p, t);
 333                         drop = calc_drop_pps(r, p, t);
 334                         err  = calc_errs_pps(r, p, t);
 335                         if (err > 0)
 336                                 errstr = "cpu-dest/err";
 337                         if (pps > 0)
 338                                 printf(fmt_rx, "XDP-RX",
 339                                         i, pps, drop, err, errstr);
 340                 }
 341                 pps  = calc_pps(&rec->total, &prev->total, t);
 342                 drop = calc_drop_pps(&rec->total, &prev->total, t);
 343                 err  = calc_errs_pps(&rec->total, &prev->total, t);
 344                 printf(fm2_rx, "XDP-RX", "total", pps, drop);
 345         }
 346
 347         /* cpumap enqueue stats */
 348         for (to_cpu = 0; to_cpu < n_cpus; to_cpu++) {
 349                 char *fmt = "%-15s %3d:%-3d %'-14.0f %'-11.0f %'-10.2f %s\n";
 350                 char *fm2 = "%-15s %3s:%-3d %'-14.0f %'-11.0f %'-10.2f %s\n";
 351                 char *errstr = "";
 352
 353                 rec  =  &stats_rec->enq[to_cpu];
 354                 prev = &stats_prev->enq[to_cpu];
 355                 t = calc_period(rec, prev);
 356                 for (i = 0; i < nr_cpus; i++) {
 357                         struct datarec *r = &rec->cpu[i];
 358                         struct datarec *p = &prev->cpu[i];
 359
 360                         pps  = calc_pps(r, p, t);
 361                         drop = calc_drop_pps(r, p, t);
 362                         err  = calc_errs_pps(r, p, t);
 363                         if (err > 0) {
 364                                 errstr = "bulk-average";
 365                                 err = pps / err; /* calc average bulk size */
 366                         }
 367                         if (pps > 0)
 368                                 printf(fmt, "cpumap-enqueue",
 369                                        i, to_cpu, pps, drop, err, errstr);
 370                 }
 371                 pps = calc_pps(&rec->total, &prev->total, t);
 372                 if (pps > 0) {
 373                         drop = calc_drop_pps(&rec->total, &prev->total, t);
 374                         err  = calc_errs_pps(&rec->total, &prev->total, t);
 375                         if (err > 0) {
 376                                 errstr = "bulk-average";
 377                                 err = pps / err; /* calc average bulk size */
 378                         }
 379                         printf(fm2, "cpumap-enqueue",
 380                                "sum", to_cpu, pps, drop, err, errstr);
 381                 }
 382         }
 383
 384         /* cpumap kthread stats */
 385         {
 386                 char *fmt_k = "%-15s %-7d %'-14.0f %'-11.0f %'-10.0f %s\n";
 387                 char *fm2_k = "%-15s %-7s %'-14.0f %'-11.0f %'-10.0f %s\n";
 388                 char *e_str = "";
 389
 390                 rec  = &stats_rec->kthread;
 391                 prev = &stats_prev->kthread;
 392                 t = calc_period(rec, prev);
 393                 for (i = 0; i < nr_cpus; i++) {
 394                         struct datarec *r = &rec->cpu[i];
 395                         struct datarec *p = &prev->cpu[i];
 396
 397                         pps  = calc_pps(r, p, t);
 398                         drop = calc_drop_pps(r, p, t);
 399                         err  = calc_errs_pps(r, p, t);
 400                         if (err > 0)
 401                                 e_str = "sched";
 402                         if (pps > 0)
 403                                 printf(fmt_k, "cpumap_kthread",
 404                                        i, pps, drop, err, e_str);
 405                 }
 406                 pps = calc_pps(&rec->total, &prev->total, t);
 407                 drop = calc_drop_pps(&rec->total, &prev->total, t);
 408                 err  = calc_errs_pps(&rec->total, &prev->total, t);
 409                 if (err > 0)
 410                         e_str = "sched-sum";
 411                 printf(fm2_k, "cpumap_kthread", "total", pps, drop, err, e_str);
 412         }
 413
 414         /* XDP redirect err tracepoints (very unlikely) */
 415         {
 416                 char *fmt_err = "%-15s %-7d %'-14.0f %'-11.0f\n";
 417                 char *fm2_err = "%-15s %-7s %'-14.0f %'-11.0f\n";
 418
 419                 rec  = &stats_rec->redir_err;
 420                 prev = &stats_prev->redir_err;
 421                 t = calc_period(rec, prev);
 422                 for (i = 0; i < nr_cpus; i++) {
 423                         struct datarec *r = &rec->cpu[i];
 424                         struct datarec *p = &prev->cpu[i];
 425
 426                         pps  = calc_pps(r, p, t);
 427                         drop = calc_drop_pps(r, p, t);
 428                         if (pps > 0)
 429                                 printf(fmt_err, "redirect_err", i, pps, drop);
 430                 }
 431                 pps = calc_pps(&rec->total, &prev->total, t);
 432                 drop = calc_drop_pps(&rec->total, &prev->total, t);
 433                 printf(fm2_err, "redirect_err", "total", pps, drop);
 434         }
 435
 436         /* XDP general exception tracepoints */
 437         {
 438                 char *fmt_err = "%-15s %-7d %'-14.0f %'-11.0f\n";
 439                 char *fm2_err = "%-15s %-7s %'-14.0f %'-11.0f\n";
 440
 441                 rec  = &stats_rec->exception;
 442                 prev = &stats_prev->exception;
 443                 t = calc_period(rec, prev);
 444                 for (i = 0; i < nr_cpus; i++) {
 445                         struct datarec *r = &rec->cpu[i];
 446                         struct datarec *p = &prev->cpu[i];
 447
 448                         pps  = calc_pps(r, p, t);
 449                         drop = calc_drop_pps(r, p, t);
 450                         if (pps > 0)
 451                                 printf(fmt_err, "xdp_exception", i, pps, drop);
 452                 }
 453                 pps = calc_pps(&rec->total, &prev->total, t);
 454                 drop = calc_drop_pps(&rec->total, &prev->total, t);
 455                 printf(fm2_err, "xdp_exception", "total", pps, drop);
 456         }
 457
 458         printf("\n");
 459         fflush(stdout);
 460 }
 461
 462 static void stats_collect(struct stats_record *rec)
 463 {
 464         int fd, i;
 465
 466         fd = rx_cnt_map_fd;
 467         map_collect_percpu(fd, 0, &rec->rx_cnt);
 468
 469         fd = redirect_err_cnt_map_fd;
 470         map_collect_percpu(fd, 1, &rec->redir_err);
 471
 472         fd = cpumap_enqueue_cnt_map_fd;
 473         for (i = 0; i < n_cpus; i++)
 474                 map_collect_percpu(fd, i, &rec->enq[i]);
 475
 476         fd = cpumap_kthread_cnt_map_fd;
 477         map_collect_percpu(fd, 0, &rec->kthread);
 478
 479         fd = exception_cnt_map_fd;
 480         map_collect_percpu(fd, 0, &rec->exception);
 481 }
 482
 483
 484 /* Pointer swap trick */
 485 static inline void swap(struct stats_record **a, struct stats_record **b)
 486 {
 487         struct stats_record *tmp;
 488
 489         tmp = *a;
 490         *a = *b;
 491         *b = tmp;
 492 }
 493
 494 static int create_cpu_entry(__u32 cpu, __u32 queue_size,
 495                             __u32 avail_idx, bool new)
 496 {
 497         __u32 curr_cpus_count = 0;
 498         __u32 key = 0;
 499         int ret;
 500
 501         /* Add a CPU entry to cpumap, as this allocate a cpu entry in
 502          * the kernel for the cpu.
 503          */
 504         ret = bpf_map_update_elem(cpu_map_fd, &cpu, &queue_size, 0);
 505         if (ret) {
 506                 fprintf(stderr, "Create CPU entry failed (err:%d)\n", ret);
 507                 exit(EXIT_FAIL_BPF);
 508         }
 509
 510         /* Inform bpf_prog's that a new CPU is available to select
 511          * from via some control maps.
 512          */
 513         ret = bpf_map_update_elem(cpus_available_map_fd, &avail_idx, &cpu, 0);
 514         if (ret) {
 515                 fprintf(stderr, "Add to avail CPUs failed\n");
 516                 exit(EXIT_FAIL_BPF);
 517         }
 518
 519         /* When not replacing/updating existing entry, bump the count */
 520         ret = bpf_map_lookup_elem(cpus_count_map_fd, &key, &curr_cpus_count);
 521         if (ret) {
 522                 fprintf(stderr, "Failed reading curr cpus_count\n");
 523                 exit(EXIT_FAIL_BPF);
 524         }
 525         if (new) {
 526                 curr_cpus_count++;
 527                 ret = bpf_map_update_elem(cpus_count_map_fd, &key,
 528                                           &curr_cpus_count, 0);
 529                 if (ret) {
 530                         fprintf(stderr, "Failed write curr cpus_count\n");
 531                         exit(EXIT_FAIL_BPF);
 532                 }
 533         }
 534         /* map_fd[7] = cpus_iterator */
 535         printf("%s CPU:%u as idx:%u queue_size:%d (total cpus_count:%u)\n",
 536                new ? "Add-new":"Replace", cpu, avail_idx,
 537                queue_size, curr_cpus_count);
 538
 539         return 0;
 540 }
 541
 542 /* CPUs are zero-indexed. Thus, add a special sentinel default value
 543  * in map cpus_available to mark CPU index'es not configured
 544  */
 545 static void mark_cpus_unavailable(void)
 546 {
 547         __u32 invalid_cpu = n_cpus;
 548         int ret, i;
 549
 550         for (i = 0; i < n_cpus; i++) {
 551                 ret = bpf_map_update_elem(cpus_available_map_fd, &i,
 552                                           &invalid_cpu, 0);
 553                 if (ret) {
 554                         fprintf(stderr, "Failed marking CPU unavailable\n");
 555                         exit(EXIT_FAIL_BPF);
 556                 }
 557         }
 558 }
 559
 560 /* Stress cpumap management code by concurrently changing underlying cpumap */
 561 static void stress_cpumap(void)
 562 {
 563         /* Changing qsize will cause kernel to free and alloc a new
 564          * bpf_cpu_map_entry, with an associated/complicated tear-down
 565          * procedure.
 566          */
 567         create_cpu_entry(1,  1024, 0, false);
 568         create_cpu_entry(1,     8, 0, false);
 569         create_cpu_entry(1, 16000, 0, false);
 570 }
 571
 572 static void stats_poll(int interval, bool use_separators, char *prog_name,
 573                        bool stress_mode)
 574 {
 575         struct stats_record *record, *prev;
 576
 577         record = alloc_stats_record();
 578         prev   = alloc_stats_record();
 579         stats_collect(record);
 580
 581         /* Trick to pretty printf with thousands separators use %' */
 582         if (use_separators)
 583                 setlocale(LC_NUMERIC, "en_US");
 584
 585         while (1) {
 586                 swap(&prev, &record);
 587                 stats_collect(record);
 588                 stats_print(record, prev, prog_name);
 589                 sleep(interval);
 590                 if (stress_mode)
 591                         stress_cpumap();
 592         }
 593
 594         free_stats_record(record);
 595         free_stats_record(prev);
 596 }
 597
 598 static struct bpf_link * attach_tp(struct bpf_object *obj,
 599                                    const char *tp_category,
 600                                    const char* tp_name)
 601 {
 602         struct bpf_program *prog;
 603         struct bpf_link *link;
 604         char sec_name[PATH_MAX];
 605         int len;
 606
 607         len = snprintf(sec_name, PATH_MAX, "tracepoint/%s/%s",
 608                        tp_category, tp_name);
 609         if (len < 0)
 610                 exit(EXIT_FAIL);
 611
 612         prog = bpf_object__find_program_by_title(obj, sec_name);
 613         if (!prog) {
 614                 fprintf(stderr, "ERR: finding progsec: %s\n", sec_name);
 615                 exit(EXIT_FAIL_BPF);
 616         }
 617
 618         link = bpf_program__attach_tracepoint(prog, tp_category, tp_name);
 619         if (libbpf_get_error(link))
 620                 exit(EXIT_FAIL_BPF);
 621
 622         return link;
 623 }
 624
 625 static void init_tracepoints(struct bpf_object *obj) {
 626         tp_links[tp_cnt++] = attach_tp(obj, "xdp", "xdp_redirect_err");
 627         tp_links[tp_cnt++] = attach_tp(obj, "xdp", "xdp_redirect_map_err");
 628         tp_links[tp_cnt++] = attach_tp(obj, "xdp", "xdp_exception");
 629         tp_links[tp_cnt++] = attach_tp(obj, "xdp", "xdp_cpumap_enqueue");
 630         tp_links[tp_cnt++] = attach_tp(obj, "xdp", "xdp_cpumap_kthread");
 631 }
 632
 633 static int init_map_fds(struct bpf_object *obj)
 634 {
 635         /* Maps updated by tracepoints */
 636         redirect_err_cnt_map_fd =
 637                 bpf_object__find_map_fd_by_name(obj, "redirect_err_cnt");
 638         exception_cnt_map_fd =
 639                 bpf_object__find_map_fd_by_name(obj, "exception_cnt");
 640         cpumap_enqueue_cnt_map_fd =
 641                 bpf_object__find_map_fd_by_name(obj, "cpumap_enqueue_cnt");
 642         cpumap_kthread_cnt_map_fd =
 643                 bpf_object__find_map_fd_by_name(obj, "cpumap_kthread_cnt");
 644
 645         /* Maps used by XDP */
 646         rx_cnt_map_fd = bpf_object__find_map_fd_by_name(obj, "rx_cnt");
 647         cpu_map_fd = bpf_object__find_map_fd_by_name(obj, "cpu_map");
 648         cpus_available_map_fd =
 649                 bpf_object__find_map_fd_by_name(obj, "cpus_available");
 650         cpus_count_map_fd = bpf_object__find_map_fd_by_name(obj, "cpus_count");
 651         cpus_iterator_map_fd =
 652                 bpf_object__find_map_fd_by_name(obj, "cpus_iterator");
 653
 654         if (cpu_map_fd < 0 || rx_cnt_map_fd < 0 ||
 655             redirect_err_cnt_map_fd < 0 || cpumap_enqueue_cnt_map_fd < 0 ||
 656             cpumap_kthread_cnt_map_fd < 0 || cpus_available_map_fd < 0 ||
 657             cpus_count_map_fd < 0 || cpus_iterator_map_fd < 0 ||
 658             exception_cnt_map_fd < 0)
 659                 return -ENOENT;
 660
 661         return 0;
 662 }
 663
 664 int main(int argc, char **argv)
 665 {
 666         struct rlimit r = {10 * 1024 * 1024, RLIM_INFINITY};
 667         char *prog_name = "xdp_cpu_map5_lb_hash_ip_pairs";
 668         struct bpf_prog_load_attr prog_load_attr = {
 669                 .prog_type      = BPF_PROG_TYPE_UNSPEC,
 670         };
 671         struct bpf_prog_info info = {};
 672         __u32 info_len = sizeof(info);
 673         bool use_separators = true;
 674         bool stress_mode = false;
 675         struct bpf_program *prog;
 676         struct bpf_object *obj;
 677         char filename[256];
 678         int added_cpus = 0;
 679         int longindex = 0;
 680         int interval = 2;
 681         int add_cpu = -1;
 682         int opt, err;
 683         int prog_fd;
 684         __u32 qsize;
 685
 686         n_cpus = get_nprocs_conf();
 687
 688         /* Notice: choosing he queue size is very important with the
 689          * ixgbe driver, because it's driver page recycling trick is
 690          * dependend on pages being returned quickly.  The number of
 691          * out-standing packets in the system must be less-than 2x
 692          * RX-ring size.
 693          */
 694         qsize = 128+64;
 695
 696         snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
 697         prog_load_attr.file = filename;
 698
 699         if (setrlimit(RLIMIT_MEMLOCK, &r)) {
 700                 perror("setrlimit(RLIMIT_MEMLOCK)");
 701                 return 1;
 702         }
 703
 704         if (bpf_prog_load_xattr(&prog_load_attr, &obj, &prog_fd))
 705                 return EXIT_FAIL;
 706
 707         if (prog_fd < 0) {
 708                 fprintf(stderr, "ERR: bpf_prog_load_xattr: %s\n",
 709                         strerror(errno));
 710                 return EXIT_FAIL;
 711         }
 712         init_tracepoints(obj);
 713         if (init_map_fds(obj) < 0) {
 714                 fprintf(stderr, "bpf_object__find_map_fd_by_name failed\n");
 715                 return EXIT_FAIL;
 716         }
 717         mark_cpus_unavailable();
 718
 719         /* Parse commands line args */
 720         while ((opt = getopt_long(argc, argv, "hSd:s:p:q:c:xzF",
 721                                   long_options, &longindex)) != -1) {
 722                 switch (opt) {
 723                 case 'd':
 724                         if (strlen(optarg) >= IF_NAMESIZE) {
 725                                 fprintf(stderr, "ERR: --dev name too long\n");
 726                                 goto error;
 727                         }
 728                         ifname = (char *)&ifname_buf;
 729                         strncpy(ifname, optarg, IF_NAMESIZE);
 730                         ifindex = if_nametoindex(ifname);
 731                         if (ifindex == 0) {
 732                                 fprintf(stderr,
 733                                         "ERR: --dev name unknown err(%d):%s\n",
 734                                         errno, strerror(errno));
 735                                 goto error;
 736                         }
 737                         break;
 738                 case 's':
 739                         interval = atoi(optarg);
 740                         break;
 741                 case 'S':
 742                         xdp_flags |= XDP_FLAGS_SKB_MODE;
 743                         break;
 744                 case 'x':
 745                         stress_mode = true;
 746                         break;
 747                 case 'z':
 748                         use_separators = false;
 749                         break;
 750                 case 'p':
 751                         /* Selecting eBPF prog to load */
 752                         prog_name = optarg;
 753                         break;
 754                 case 'c':
 755                         /* Add multiple CPUs */
 756                         add_cpu = strtoul(optarg, NULL, 0);
 757                         if (add_cpu >= n_cpus) {
 758                                 fprintf(stderr,
 759                                 "--cpu nr too large for cpumap err(%d):%s\n",
 760                                         errno, strerror(errno));
 761                                 goto error;
 762                         }
 763                         create_cpu_entry(add_cpu, qsize, added_cpus, true);
 764                         added_cpus++;
 765                         break;
 766                 case 'q':
 767                         qsize = atoi(optarg);
 768                         break;
 769                 case 'F':
 770                         xdp_flags &= ~XDP_FLAGS_UPDATE_IF_NOEXIST;
 771                         break;
 772                 case 'h':
 773                 error:
 774                 default:
 775                         usage(argv, obj);
 776                         return EXIT_FAIL_OPTION;
 777                 }
 778         }
 779
 780         if (!(xdp_flags & XDP_FLAGS_SKB_MODE))
 781                 xdp_flags |= XDP_FLAGS_DRV_MODE;
 782
 783         /* Required option */
 784         if (ifindex == -1) {
 785                 fprintf(stderr, "ERR: required option --dev missing\n");
 786                 usage(argv, obj);
 787                 return EXIT_FAIL_OPTION;
 788         }
 789         /* Required option */
 790         if (add_cpu == -1) {
 791                 fprintf(stderr, "ERR: required option --cpu missing\n");
 792                 fprintf(stderr, " Specify multiple --cpu option to add more\n");
 793                 usage(argv, obj);
 794                 return EXIT_FAIL_OPTION;
 795         }
 796
 797         /* Remove XDP program when program is interrupted or killed */
 798         signal(SIGINT, int_exit);
 799         signal(SIGTERM, int_exit);
 800
 801         prog = bpf_object__find_program_by_title(obj, prog_name);
 802         if (!prog) {
 803                 fprintf(stderr, "bpf_object__find_program_by_title failed\n");
 804                 return EXIT_FAIL;
 805         }
 806
 807         prog_fd = bpf_program__fd(prog);
 808         if (prog_fd < 0) {
 809                 fprintf(stderr, "bpf_program__fd failed\n");
 810                 return EXIT_FAIL;
 811         }
 812
 813         if (bpf_set_link_xdp_fd(ifindex, prog_fd, xdp_flags) < 0) {
 814                 fprintf(stderr, "link set xdp fd failed\n");
 815                 return EXIT_FAIL_XDP;
 816         }
 817
 818         err = bpf_obj_get_info_by_fd(prog_fd, &info, &info_len);
 819         if (err) {
 820                 printf("can't get prog info - %s\n", strerror(errno));
 821                 return err;
 822         }
 823         prog_id = info.id;
 824
 825         stats_poll(interval, use_separators, prog_name, stress_mode);
 826         return EXIT_OK;
 827 }