samples/bpf/hbm.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2019 Facebook
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of version 2 of the GNU General Public
   6  * License as published by the Free Software Foundation.
   7  *
   8  * Example program for Host Bandwidth Managment
   9  *
  10  * This program loads a cgroup skb BPF program to enforce cgroup output
  11  * (egress) or input (ingress) bandwidth limits.
  12  *
  13  * USAGE: hbm [-d] [-l] [-n <id>] [-r <rate>] [-s] [-t <secs>] [-w] [-h] [prog]
  14  *   Where:
  15  *    -d        Print BPF trace debug buffer
  16  *    -l        Also limit flows doing loopback
  17  *    -n <#>    To create cgroup \"/hbm#\" and attach prog
  18  *              Default is /hbm1
  19  *    --no_cn   Do not return cn notifications
  20  *    -r <rate> Rate limit in Mbps
  21  *    -s        Get HBM stats (marked, dropped, etc.)
  22  *    -t <time> Exit after specified seconds (default is 0)
  23  *    -w        Work conserving flag. cgroup can increase its bandwidth
  24  *              beyond the rate limit specified while there is available
  25  *              bandwidth. Current implementation assumes there is only
  26  *              NIC (eth0), but can be extended to support multiple NICs.
  27  *              Currrently only supported for egress.
  28  *    -h        Print this info
  29  *    prog      BPF program file name. Name defaults to hbm_out_kern.o
  30  */
  31
  32 #define _GNU_SOURCE
  33
  34 #include <stdio.h>
  35 #include <stdlib.h>
  36 #include <assert.h>
  37 #include <sys/resource.h>
  38 #include <sys/time.h>
  39 #include <unistd.h>
  40 #include <errno.h>
  41 #include <fcntl.h>
  42 #include <linux/unistd.h>
  43 #include <linux/compiler.h>
  44
  45 #include <linux/bpf.h>
  46 #include <bpf/bpf.h>
  47 #include <getopt.h>
  48
  49 #include "bpf_rlimit.h"
  50 #include "cgroup_helpers.h"
  51 #include "hbm.h"
  52 #include "bpf_util.h"
  53 #include <bpf/libbpf.h>
  54
  55 bool outFlag = true;
  56 int minRate = 1000;             /* cgroup rate limit in Mbps */
  57 int rate = 1000;                /* can grow if rate conserving is enabled */
  58 int dur = 1;
  59 bool stats_flag;
  60 bool loopback_flag;
  61 bool debugFlag;
  62 bool work_conserving_flag;
  63 bool no_cn_flag;
  64 bool edt_flag;
  65
  66 static void Usage(void);
  67 static void read_trace_pipe2(void);
  68 static void do_error(char *msg, bool errno_flag);
  69
  70 #define DEBUGFS "/sys/kernel/debug/tracing/"
  71
  72 static struct bpf_program *bpf_prog;
  73 static struct bpf_object *obj;
  74 static int queue_stats_fd;
  75
  76 static void read_trace_pipe2(void)
  77 {
  78         int trace_fd;
  79         FILE *outf;
  80         char *outFname = "hbm_out.log";
  81
  82         trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
  83         if (trace_fd < 0) {
  84                 printf("Error opening trace_pipe\n");
  85                 return;
  86         }
  87
  88 //      Future support of ingress
  89 //      if (!outFlag)
  90 //              outFname = "hbm_in.log";
  91         outf = fopen(outFname, "w");
  92
  93         if (outf == NULL)
  94                 printf("Error creating %s\n", outFname);
  95
  96         while (1) {
  97                 static char buf[4097];
  98                 ssize_t sz;
  99
 100                 sz = read(trace_fd, buf, sizeof(buf) - 1);
 101                 if (sz > 0) {
 102                         buf[sz] = 0;
 103                         puts(buf);
 104                         if (outf != NULL) {
 105                                 fprintf(outf, "%s\n", buf);
 106                                 fflush(outf);
 107                         }
 108                 }
 109         }
 110 }
 111
 112 static void do_error(char *msg, bool errno_flag)
 113 {
 114         if (errno_flag)
 115                 printf("ERROR: %s, errno: %d\n", msg, errno);
 116         else
 117                 printf("ERROR: %s\n", msg);
 118         exit(1);
 119 }
 120
 121 static int prog_load(char *prog)
 122 {
 123         struct bpf_program *pos;
 124         const char *sec_name;
 125
 126         obj = bpf_object__open_file(prog, NULL);
 127         if (libbpf_get_error(obj)) {
 128                 printf("ERROR: opening BPF object file failed\n");
 129                 return 1;
 130         }
 131
 132         /* load BPF program */
 133         if (bpf_object__load(obj)) {
 134                 printf("ERROR: loading BPF object file failed\n");
 135                 goto err;
 136         }
 137
 138         bpf_object__for_each_program(pos, obj) {
 139                 sec_name = bpf_program__section_name(pos);
 140                 if (sec_name && !strcmp(sec_name, "cgroup_skb/egress")) {
 141                         bpf_prog = pos;
 142                         break;
 143                 }
 144         }
 145         if (!bpf_prog) {
 146                 printf("ERROR: finding a prog in obj file failed\n");
 147                 goto err;
 148         }
 149
 150         queue_stats_fd = bpf_object__find_map_fd_by_name(obj, "queue_stats");
 151         if (queue_stats_fd < 0) {
 152                 printf("ERROR: finding a map in obj file failed\n");
 153                 goto err;
 154         }
 155
 156         return 0;
 157
 158 err:
 159         bpf_object__close(obj);
 160         return 1;
 161 }
 162
 163 static int run_bpf_prog(char *prog, int cg_id)
 164 {
 165         struct hbm_queue_stats qstats = {0};
 166         char cg_dir[100], cg_pin_path[100];
 167         struct bpf_link *link = NULL;
 168         int key = 0;
 169         int cg1 = 0;
 170         int rc = 0;
 171
 172         sprintf(cg_dir, "/hbm%d", cg_id);
 173         rc = prog_load(prog);
 174         if (rc != 0)
 175                 return rc;
 176
 177         if (setup_cgroup_environment()) {
 178                 printf("ERROR: setting cgroup environment\n");
 179                 goto err;
 180         }
 181         cg1 = create_and_get_cgroup(cg_dir);
 182         if (!cg1) {
 183                 printf("ERROR: create_and_get_cgroup\n");
 184                 goto err;
 185         }
 186         if (join_cgroup(cg_dir)) {
 187                 printf("ERROR: join_cgroup\n");
 188                 goto err;
 189         }
 190
 191         qstats.rate = rate;
 192         qstats.stats = stats_flag ? 1 : 0;
 193         qstats.loopback = loopback_flag ? 1 : 0;
 194         qstats.no_cn = no_cn_flag ? 1 : 0;
 195         if (bpf_map_update_elem(queue_stats_fd, &key, &qstats, BPF_ANY)) {
 196                 printf("ERROR: Could not update map element\n");
 197                 goto err;
 198         }
 199
 200         if (!outFlag)
 201                 bpf_program__set_expected_attach_type(bpf_prog, BPF_CGROUP_INET_INGRESS);
 202
 203         link = bpf_program__attach_cgroup(bpf_prog, cg1);
 204         if (libbpf_get_error(link)) {
 205                 fprintf(stderr, "ERROR: bpf_program__attach_cgroup failed\n");
 206                 goto err;
 207         }
 208
 209         sprintf(cg_pin_path, "/sys/fs/bpf/hbm%d", cg_id);
 210         rc = bpf_link__pin(link, cg_pin_path);
 211         if (rc < 0) {
 212                 printf("ERROR: bpf_link__pin failed: %d\n", rc);
 213                 goto err;
 214         }
 215
 216         if (work_conserving_flag) {
 217                 struct timeval t0, t_last, t_new;
 218                 FILE *fin;
 219                 unsigned long long last_eth_tx_bytes, new_eth_tx_bytes;
 220                 signed long long last_cg_tx_bytes, new_cg_tx_bytes;
 221                 signed long long delta_time, delta_bytes, delta_rate;
 222                 int delta_ms;
 223 #define DELTA_RATE_CHECK 10000          /* in us */
 224 #define RATE_THRESHOLD 9500000000       /* 9.5 Gbps */
 225
 226                 bpf_map_lookup_elem(queue_stats_fd, &key, &qstats);
 227                 if (gettimeofday(&t0, NULL) < 0)
 228                         do_error("gettimeofday failed", true);
 229                 t_last = t0;
 230                 fin = fopen("/sys/class/net/eth0/statistics/tx_bytes", "r");
 231                 if (fscanf(fin, "%llu", &last_eth_tx_bytes) != 1)
 232                         do_error("fscanf fails", false);
 233                 fclose(fin);
 234                 last_cg_tx_bytes = qstats.bytes_total;
 235                 while (true) {
 236                         usleep(DELTA_RATE_CHECK);
 237                         if (gettimeofday(&t_new, NULL) < 0)
 238                                 do_error("gettimeofday failed", true);
 239                         delta_ms = (t_new.tv_sec - t0.tv_sec) * 1000 +
 240                                 (t_new.tv_usec - t0.tv_usec)/1000;
 241                         if (delta_ms > dur * 1000)
 242                                 break;
 243                         delta_time = (t_new.tv_sec - t_last.tv_sec) * 1000000 +
 244                                 (t_new.tv_usec - t_last.tv_usec);
 245                         if (delta_time == 0)
 246                                 continue;
 247                         t_last = t_new;
 248                         fin = fopen("/sys/class/net/eth0/statistics/tx_bytes",
 249                                     "r");
 250                         if (fscanf(fin, "%llu", &new_eth_tx_bytes) != 1)
 251                                 do_error("fscanf fails", false);
 252                         fclose(fin);
 253                         printf("  new_eth_tx_bytes:%llu\n",
 254                                new_eth_tx_bytes);
 255                         bpf_map_lookup_elem(queue_stats_fd, &key, &qstats);
 256                         new_cg_tx_bytes = qstats.bytes_total;
 257                         delta_bytes = new_eth_tx_bytes - last_eth_tx_bytes;
 258                         last_eth_tx_bytes = new_eth_tx_bytes;
 259                         delta_rate = (delta_bytes * 8000000) / delta_time;
 260                         printf("%5d - eth_rate:%.1fGbps cg_rate:%.3fGbps",
 261                                delta_ms, delta_rate/1000000000.0,
 262                                rate/1000.0);
 263                         if (delta_rate < RATE_THRESHOLD) {
 264                                 /* can increase cgroup rate limit, but first
 265                                  * check if we are using the current limit.
 266                                  * Currently increasing by 6.25%, unknown
 267                                  * if that is the optimal rate.
 268                                  */
 269                                 int rate_diff100;
 270
 271                                 delta_bytes = new_cg_tx_bytes -
 272                                         last_cg_tx_bytes;
 273                                 last_cg_tx_bytes = new_cg_tx_bytes;
 274                                 delta_rate = (delta_bytes * 8000000) /
 275                                         delta_time;
 276                                 printf(" rate:%.3fGbps",
 277                                        delta_rate/1000000000.0);
 278                                 rate_diff100 = (((long long)rate)*1000000 -
 279                                                      delta_rate) * 100 /
 280                                         (((long long) rate) * 1000000);
 281                                 printf("  rdiff:%d", rate_diff100);
 282                                 if (rate_diff100  <= 3) {
 283                                         rate += (rate >> 4);
 284                                         if (rate > RATE_THRESHOLD / 1000000)
 285                                                 rate = RATE_THRESHOLD / 1000000;
 286                                         qstats.rate = rate;
 287                                         printf(" INC\n");
 288                                 } else {
 289                                         printf("\n");
 290                                 }
 291                         } else {
 292                                 /* Need to decrease cgroup rate limit.
 293                                  * Currently decreasing by 12.5%, unknown
 294                                  * if that is optimal
 295                                  */
 296                                 printf(" DEC\n");
 297                                 rate -= (rate >> 3);
 298                                 if (rate < minRate)
 299                                         rate = minRate;
 300                                 qstats.rate = rate;
 301                         }
 302                         if (bpf_map_update_elem(queue_stats_fd, &key, &qstats, BPF_ANY))
 303                                 do_error("update map element fails", false);
 304                 }
 305         } else {
 306                 sleep(dur);
 307         }
 308         // Get stats!
 309         if (stats_flag && bpf_map_lookup_elem(queue_stats_fd, &key, &qstats)) {
 310                 char fname[100];
 311                 FILE *fout;
 312
 313                 if (!outFlag)
 314                         sprintf(fname, "hbm.%d.in", cg_id);
 315                 else
 316                         sprintf(fname, "hbm.%d.out", cg_id);
 317                 fout = fopen(fname, "w");
 318                 fprintf(fout, "id:%d\n", cg_id);
 319                 fprintf(fout, "ERROR: Could not lookup queue_stats\n");
 320         } else if (stats_flag && qstats.lastPacketTime >
 321                    qstats.firstPacketTime) {
 322                 long long delta_us = (qstats.lastPacketTime -
 323                                       qstats.firstPacketTime)/1000;
 324                 unsigned int rate_mbps = ((qstats.bytes_total -
 325                                            qstats.bytes_dropped) * 8 /
 326                                           delta_us);
 327                 double percent_pkts, percent_bytes;
 328                 char fname[100];
 329                 FILE *fout;
 330                 int k;
 331                 static const char *returnValNames[] = {
 332                         "DROP_PKT",
 333                         "ALLOW_PKT",
 334                         "DROP_PKT_CWR",
 335                         "ALLOW_PKT_CWR"
 336                 };
 337 #define RET_VAL_COUNT 4
 338
 339 // Future support of ingress
 340 //              if (!outFlag)
 341 //                      sprintf(fname, "hbm.%d.in", cg_id);
 342 //              else
 343                 sprintf(fname, "hbm.%d.out", cg_id);
 344                 fout = fopen(fname, "w");
 345                 fprintf(fout, "id:%d\n", cg_id);
 346                 fprintf(fout, "rate_mbps:%d\n", rate_mbps);
 347                 fprintf(fout, "duration:%.1f secs\n",
 348                         (qstats.lastPacketTime - qstats.firstPacketTime) /
 349                         1000000000.0);
 350                 fprintf(fout, "packets:%d\n", (int)qstats.pkts_total);
 351                 fprintf(fout, "bytes_MB:%d\n", (int)(qstats.bytes_total /
 352                                                      1000000));
 353                 fprintf(fout, "pkts_dropped:%d\n", (int)qstats.pkts_dropped);
 354                 fprintf(fout, "bytes_dropped_MB:%d\n",
 355                         (int)(qstats.bytes_dropped /
 356                                                        1000000));
 357                 // Marked Pkts and Bytes
 358                 percent_pkts = (qstats.pkts_marked * 100.0) /
 359                         (qstats.pkts_total + 1);
 360                 percent_bytes = (qstats.bytes_marked * 100.0) /
 361                         (qstats.bytes_total + 1);
 362                 fprintf(fout, "pkts_marked_percent:%6.2f\n", percent_pkts);
 363                 fprintf(fout, "bytes_marked_percent:%6.2f\n", percent_bytes);
 364
 365                 // Dropped Pkts and Bytes
 366                 percent_pkts = (qstats.pkts_dropped * 100.0) /
 367                         (qstats.pkts_total + 1);
 368                 percent_bytes = (qstats.bytes_dropped * 100.0) /
 369                         (qstats.bytes_total + 1);
 370                 fprintf(fout, "pkts_dropped_percent:%6.2f\n", percent_pkts);
 371                 fprintf(fout, "bytes_dropped_percent:%6.2f\n", percent_bytes);
 372
 373                 // ECN CE markings
 374                 percent_pkts = (qstats.pkts_ecn_ce * 100.0) /
 375                         (qstats.pkts_total + 1);
 376                 fprintf(fout, "pkts_ecn_ce:%6.2f (%d)\n", percent_pkts,
 377                         (int)qstats.pkts_ecn_ce);
 378
 379                 // Average cwnd
 380                 fprintf(fout, "avg cwnd:%d\n",
 381                         (int)(qstats.sum_cwnd / (qstats.sum_cwnd_cnt + 1)));
 382                 // Average rtt
 383                 fprintf(fout, "avg rtt:%d\n",
 384                         (int)(qstats.sum_rtt / (qstats.pkts_total + 1)));
 385                 // Average credit
 386                 if (edt_flag)
 387                         fprintf(fout, "avg credit_ms:%.03f\n",
 388                                 (qstats.sum_credit /
 389                                  (qstats.pkts_total + 1.0)) / 1000000.0);
 390                 else
 391                         fprintf(fout, "avg credit:%d\n",
 392                                 (int)(qstats.sum_credit /
 393                                       (1500 * ((int)qstats.pkts_total ) + 1)));
 394
 395                 // Return values stats
 396                 for (k = 0; k < RET_VAL_COUNT; k++) {
 397                         percent_pkts = (qstats.returnValCount[k] * 100.0) /
 398                                 (qstats.pkts_total + 1);
 399                         fprintf(fout, "%s:%6.2f (%d)\n", returnValNames[k],
 400                                 percent_pkts, (int)qstats.returnValCount[k]);
 401                 }
 402                 fclose(fout);
 403         }
 404
 405         if (debugFlag)
 406                 read_trace_pipe2();
 407         goto cleanup;
 408
 409 err:
 410         rc = 1;
 411
 412 cleanup:
 413         bpf_link__destroy(link);
 414         bpf_object__close(obj);
 415
 416         if (cg1 != -1)
 417                 close(cg1);
 418
 419         if (rc != 0)
 420                 cleanup_cgroup_environment();
 421         return rc;
 422 }
 423
 424 static void Usage(void)
 425 {
 426         printf("This program loads a cgroup skb BPF program to enforce\n"
 427                "cgroup output (egress) bandwidth limits.\n\n"
 428                "USAGE: hbm [-o] [-d]  [-l] [-n <id>] [--no_cn] [-r <rate>]\n"
 429                "           [-s] [-t <secs>] [-w] [-h] [prog]\n"
 430                "  Where:\n"
 431                "    -o         indicates egress direction (default)\n"
 432                "    -d         print BPF trace debug buffer\n"
 433                "    --edt      use fq's Earliest Departure Time\n"
 434                "    -l         also limit flows using loopback\n"
 435                "    -n <#>     to create cgroup \"/hbm#\" and attach prog\n"
 436                "               Default is /hbm1\n"
 437                "    --no_cn    disable CN notifications\n"
 438                "    -r <rate>  Rate in Mbps\n"
 439                "    -s         Update HBM stats\n"
 440                "    -t <time>  Exit after specified seconds (default is 0)\n"
 441                "    -w         Work conserving flag. cgroup can increase\n"
 442                "               bandwidth beyond the rate limit specified\n"
 443                "               while there is available bandwidth. Current\n"
 444                "               implementation assumes there is only eth0\n"
 445                "               but can be extended to support multiple NICs\n"
 446                "    -h         print this info\n"
 447                "    prog       BPF program file name. Name defaults to\n"
 448                "                 hbm_out_kern.o\n");
 449 }
 450
 451 int main(int argc, char **argv)
 452 {
 453         char *prog = "hbm_out_kern.o";
 454         int  k;
 455         int cg_id = 1;
 456         char *optstring = "iodln:r:st:wh";
 457         struct option loptions[] = {
 458                 {"no_cn", 0, NULL, 1},
 459                 {"edt", 0, NULL, 2},
 460                 {NULL, 0, NULL, 0}
 461         };
 462
 463         while ((k = getopt_long(argc, argv, optstring, loptions, NULL)) != -1) {
 464                 switch (k) {
 465                 case 1:
 466                         no_cn_flag = true;
 467                         break;
 468                 case 2:
 469                         prog = "hbm_edt_kern.o";
 470                         edt_flag = true;
 471                         break;
 472                 case'o':
 473                         break;
 474                 case 'd':
 475                         debugFlag = true;
 476                         break;
 477                 case 'l':
 478                         loopback_flag = true;
 479                         break;
 480                 case 'n':
 481                         cg_id = atoi(optarg);
 482                         break;
 483                 case 'r':
 484                         minRate = atoi(optarg) * 1.024;
 485                         rate = minRate;
 486                         break;
 487                 case 's':
 488                         stats_flag = true;
 489                         break;
 490                 case 't':
 491                         dur = atoi(optarg);
 492                         break;
 493                 case 'w':
 494                         work_conserving_flag = true;
 495                         break;
 496                 case '?':
 497                         if (optopt == 'n' || optopt == 'r' || optopt == 't')
 498                                 fprintf(stderr,
 499                                         "Option -%c requires an argument.\n\n",
 500                                         optopt);
 501                 case 'h':
 502                         __fallthrough;
 503                 default:
 504                         Usage();
 505                         return 0;
 506                 }
 507         }
 508
 509         if (optind < argc)
 510                 prog = argv[optind];
 511         printf("HBM prog: %s\n", prog != NULL ? prog : "NULL");
 512
 513         return run_bpf_prog(prog, cg_id);
 514 }