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  *    -r <rate> Rate limit in Mbps
  20  *    -s        Get HBM stats (marked, dropped, etc.)
  21  *    -t <time> Exit after specified seconds (deault is 0)
  22  *    -w        Work conserving flag. cgroup can increase its bandwidth
  23  *              beyond the rate limit specified while there is available
  24  *              bandwidth. Current implementation assumes there is only
  25  *              NIC (eth0), but can be extended to support multiple NICs.
  26  *              Currrently only supported for egress.
  27  *    -h        Print this info
  28  *    prog      BPF program file name. Name defaults to hbm_out_kern.o
  29  */
  30
  31 #define _GNU_SOURCE
  32
  33 #include <stdio.h>
  34 #include <stdlib.h>
  35 #include <assert.h>
  36 #include <sys/resource.h>
  37 #include <sys/time.h>
  38 #include <unistd.h>
  39 #include <errno.h>
  40 #include <fcntl.h>
  41 #include <linux/unistd.h>
  42
  43 #include <linux/bpf.h>
  44 #include <bpf/bpf.h>
  45
  46 #include "bpf_load.h"
  47 #include "bpf_rlimit.h"
  48 #include "cgroup_helpers.h"
  49 #include "hbm.h"
  50 #include "bpf_util.h"
  51 #include "bpf/bpf.h"
  52 #include "bpf/libbpf.h"
  53
  54 bool outFlag = true;
  55 int minRate = 1000;             /* cgroup rate limit in Mbps */
  56 int rate = 1000;                /* can grow if rate conserving is enabled */
  57 int dur = 1;
  58 bool stats_flag;
  59 bool loopback_flag;
  60 bool debugFlag;
  61 bool work_conserving_flag;
  62
  63 static void Usage(void);
  64 static void read_trace_pipe2(void);
  65 static void do_error(char *msg, bool errno_flag);
  66
  67 #define DEBUGFS "/sys/kernel/debug/tracing/"
  68
  69 struct bpf_object *obj;
  70 int bpfprog_fd;
  71 int cgroup_storage_fd;
  72
  73 static void read_trace_pipe2(void)
  74 {
  75         int trace_fd;
  76         FILE *outf;
  77         char *outFname = "hbm_out.log";
  78
  79         trace_fd = open(DEBUGFS "trace_pipe", O_RDONLY, 0);
  80         if (trace_fd < 0) {
  81                 printf("Error opening trace_pipe\n");
  82                 return;
  83         }
  84
  85 //      Future support of ingress
  86 //      if (!outFlag)
  87 //              outFname = "hbm_in.log";
  88         outf = fopen(outFname, "w");
  89
  90         if (outf == NULL)
  91                 printf("Error creating %s\n", outFname);
  92
  93         while (1) {
  94                 static char buf[4097];
  95                 ssize_t sz;
  96
  97                 sz = read(trace_fd, buf, sizeof(buf) - 1);
  98                 if (sz > 0) {
  99                         buf[sz] = 0;
 100                         puts(buf);
 101                         if (outf != NULL) {
 102                                 fprintf(outf, "%s\n", buf);
 103                                 fflush(outf);
 104                         }
 105                 }
 106         }
 107 }
 108
 109 static void do_error(char *msg, bool errno_flag)
 110 {
 111         if (errno_flag)
 112                 printf("ERROR: %s, errno: %d\n", msg, errno);
 113         else
 114                 printf("ERROR: %s\n", msg);
 115         exit(1);
 116 }
 117
 118 static int prog_load(char *prog)
 119 {
 120         struct bpf_prog_load_attr prog_load_attr = {
 121                 .prog_type = BPF_PROG_TYPE_CGROUP_SKB,
 122                 .file = prog,
 123                 .expected_attach_type = BPF_CGROUP_INET_EGRESS,
 124         };
 125         int map_fd;
 126         struct bpf_map *map;
 127
 128         int ret = 0;
 129
 130         if (access(prog, O_RDONLY) < 0) {
 131                 printf("Error accessing file %s: %s\n", prog, strerror(errno));
 132                 return 1;
 133         }
 134         if (bpf_prog_load_xattr(&prog_load_attr, &obj, &bpfprog_fd))
 135                 ret = 1;
 136         if (!ret) {
 137                 map = bpf_object__find_map_by_name(obj, "queue_stats");
 138                 map_fd = bpf_map__fd(map);
 139                 if (map_fd < 0) {
 140                         printf("Map not found: %s\n", strerror(map_fd));
 141                         ret = 1;
 142                 }
 143         }
 144
 145         if (ret) {
 146                 printf("ERROR: load_bpf_file failed for: %s\n", prog);
 147                 printf("  Output from verifier:\n%s\n------\n", bpf_log_buf);
 148                 ret = -1;
 149         } else {
 150                 ret = map_fd;
 151         }
 152
 153         return ret;
 154 }
 155
 156 static int run_bpf_prog(char *prog, int cg_id)
 157 {
 158         int map_fd;
 159         int rc = 0;
 160         int key = 0;
 161         int cg1 = 0;
 162         int type = BPF_CGROUP_INET_EGRESS;
 163         char cg_dir[100];
 164         struct hbm_queue_stats qstats = {0};
 165
 166         sprintf(cg_dir, "/hbm%d", cg_id);
 167         map_fd = prog_load(prog);
 168         if (map_fd  == -1)
 169                 return 1;
 170
 171         if (setup_cgroup_environment()) {
 172                 printf("ERROR: setting cgroup environment\n");
 173                 goto err;
 174         }
 175         cg1 = create_and_get_cgroup(cg_dir);
 176         if (!cg1) {
 177                 printf("ERROR: create_and_get_cgroup\n");
 178                 goto err;
 179         }
 180         if (join_cgroup(cg_dir)) {
 181                 printf("ERROR: join_cgroup\n");
 182                 goto err;
 183         }
 184
 185         qstats.rate = rate;
 186         qstats.stats = stats_flag ? 1 : 0;
 187         qstats.loopback = loopback_flag ? 1 : 0;
 188         if (bpf_map_update_elem(map_fd, &key, &qstats, BPF_ANY)) {
 189                 printf("ERROR: Could not update map element\n");
 190                 goto err;
 191         }
 192
 193         if (!outFlag)
 194                 type = BPF_CGROUP_INET_INGRESS;
 195         if (bpf_prog_attach(bpfprog_fd, cg1, type, 0)) {
 196                 printf("ERROR: bpf_prog_attach fails!\n");
 197                 log_err("Attaching prog");
 198                 goto err;
 199         }
 200
 201         if (work_conserving_flag) {
 202                 struct timeval t0, t_last, t_new;
 203                 FILE *fin;
 204                 unsigned long long last_eth_tx_bytes, new_eth_tx_bytes;
 205                 signed long long last_cg_tx_bytes, new_cg_tx_bytes;
 206                 signed long long delta_time, delta_bytes, delta_rate;
 207                 int delta_ms;
 208 #define DELTA_RATE_CHECK 10000          /* in us */
 209 #define RATE_THRESHOLD 9500000000       /* 9.5 Gbps */
 210
 211                 bpf_map_lookup_elem(map_fd, &key, &qstats);
 212                 if (gettimeofday(&t0, NULL) < 0)
 213                         do_error("gettimeofday failed", true);
 214                 t_last = t0;
 215                 fin = fopen("/sys/class/net/eth0/statistics/tx_bytes", "r");
 216                 if (fscanf(fin, "%llu", &last_eth_tx_bytes) != 1)
 217                         do_error("fscanf fails", false);
 218                 fclose(fin);
 219                 last_cg_tx_bytes = qstats.bytes_total;
 220                 while (true) {
 221                         usleep(DELTA_RATE_CHECK);
 222                         if (gettimeofday(&t_new, NULL) < 0)
 223                                 do_error("gettimeofday failed", true);
 224                         delta_ms = (t_new.tv_sec - t0.tv_sec) * 1000 +
 225                                 (t_new.tv_usec - t0.tv_usec)/1000;
 226                         if (delta_ms > dur * 1000)
 227                                 break;
 228                         delta_time = (t_new.tv_sec - t_last.tv_sec) * 1000000 +
 229                                 (t_new.tv_usec - t_last.tv_usec);
 230                         if (delta_time == 0)
 231                                 continue;
 232                         t_last = t_new;
 233                         fin = fopen("/sys/class/net/eth0/statistics/tx_bytes",
 234                                     "r");
 235                         if (fscanf(fin, "%llu", &new_eth_tx_bytes) != 1)
 236                                 do_error("fscanf fails", false);
 237                         fclose(fin);
 238                         printf("  new_eth_tx_bytes:%llu\n",
 239                                new_eth_tx_bytes);
 240                         bpf_map_lookup_elem(map_fd, &key, &qstats);
 241                         new_cg_tx_bytes = qstats.bytes_total;
 242                         delta_bytes = new_eth_tx_bytes - last_eth_tx_bytes;
 243                         last_eth_tx_bytes = new_eth_tx_bytes;
 244                         delta_rate = (delta_bytes * 8000000) / delta_time;
 245                         printf("%5d - eth_rate:%.1fGbps cg_rate:%.3fGbps",
 246                                delta_ms, delta_rate/1000000000.0,
 247                                rate/1000.0);
 248                         if (delta_rate < RATE_THRESHOLD) {
 249                                 /* can increase cgroup rate limit, but first
 250                                  * check if we are using the current limit.
 251                                  * Currently increasing by 6.25%, unknown
 252                                  * if that is the optimal rate.
 253                                  */
 254                                 int rate_diff100;
 255
 256                                 delta_bytes = new_cg_tx_bytes -
 257                                         last_cg_tx_bytes;
 258                                 last_cg_tx_bytes = new_cg_tx_bytes;
 259                                 delta_rate = (delta_bytes * 8000000) /
 260                                         delta_time;
 261                                 printf(" rate:%.3fGbps",
 262                                        delta_rate/1000000000.0);
 263                                 rate_diff100 = (((long long)rate)*1000000 -
 264                                                      delta_rate) * 100 /
 265                                         (((long long) rate) * 1000000);
 266                                 printf("  rdiff:%d", rate_diff100);
 267                                 if (rate_diff100  <= 3) {
 268                                         rate += (rate >> 4);
 269                                         if (rate > RATE_THRESHOLD / 1000000)
 270                                                 rate = RATE_THRESHOLD / 1000000;
 271                                         qstats.rate = rate;
 272                                         printf(" INC\n");
 273                                 } else {
 274                                         printf("\n");
 275                                 }
 276                         } else {
 277                                 /* Need to decrease cgroup rate limit.
 278                                  * Currently decreasing by 12.5%, unknown
 279                                  * if that is optimal
 280                                  */
 281                                 printf(" DEC\n");
 282                                 rate -= (rate >> 3);
 283                                 if (rate < minRate)
 284                                         rate = minRate;
 285                                 qstats.rate = rate;
 286                         }
 287                         if (bpf_map_update_elem(map_fd, &key, &qstats, BPF_ANY))
 288                                 do_error("update map element fails", false);
 289                 }
 290         } else {
 291                 sleep(dur);
 292         }
 293         // Get stats!
 294         if (stats_flag && bpf_map_lookup_elem(map_fd, &key, &qstats)) {
 295                 char fname[100];
 296                 FILE *fout;
 297
 298                 if (!outFlag)
 299                         sprintf(fname, "hbm.%d.in", cg_id);
 300                 else
 301                         sprintf(fname, "hbm.%d.out", cg_id);
 302                 fout = fopen(fname, "w");
 303                 fprintf(fout, "id:%d\n", cg_id);
 304                 fprintf(fout, "ERROR: Could not lookup queue_stats\n");
 305         } else if (stats_flag && qstats.lastPacketTime >
 306                    qstats.firstPacketTime) {
 307                 long long delta_us = (qstats.lastPacketTime -
 308                                       qstats.firstPacketTime)/1000;
 309                 unsigned int rate_mbps = ((qstats.bytes_total -
 310                                            qstats.bytes_dropped) * 8 /
 311                                           delta_us);
 312                 double percent_pkts, percent_bytes;
 313                 char fname[100];
 314                 FILE *fout;
 315
 316 // Future support of ingress
 317 //              if (!outFlag)
 318 //                      sprintf(fname, "hbm.%d.in", cg_id);
 319 //              else
 320                 sprintf(fname, "hbm.%d.out", cg_id);
 321                 fout = fopen(fname, "w");
 322                 fprintf(fout, "id:%d\n", cg_id);
 323                 fprintf(fout, "rate_mbps:%d\n", rate_mbps);
 324                 fprintf(fout, "duration:%.1f secs\n",
 325                         (qstats.lastPacketTime - qstats.firstPacketTime) /
 326                         1000000000.0);
 327                 fprintf(fout, "packets:%d\n", (int)qstats.pkts_total);
 328                 fprintf(fout, "bytes_MB:%d\n", (int)(qstats.bytes_total /
 329                                                      1000000));
 330                 fprintf(fout, "pkts_dropped:%d\n", (int)qstats.pkts_dropped);
 331                 fprintf(fout, "bytes_dropped_MB:%d\n",
 332                         (int)(qstats.bytes_dropped /
 333                                                        1000000));
 334                 // Marked Pkts and Bytes
 335                 percent_pkts = (qstats.pkts_marked * 100.0) /
 336                         (qstats.pkts_total + 1);
 337                 percent_bytes = (qstats.bytes_marked * 100.0) /
 338                         (qstats.bytes_total + 1);
 339                 fprintf(fout, "pkts_marked_percent:%6.2f\n", percent_pkts);
 340                 fprintf(fout, "bytes_marked_percent:%6.2f\n", percent_bytes);
 341
 342                 // Dropped Pkts and Bytes
 343                 percent_pkts = (qstats.pkts_dropped * 100.0) /
 344                         (qstats.pkts_total + 1);
 345                 percent_bytes = (qstats.bytes_dropped * 100.0) /
 346                         (qstats.bytes_total + 1);
 347                 fprintf(fout, "pkts_dropped_percent:%6.2f\n", percent_pkts);
 348                 fprintf(fout, "bytes_dropped_percent:%6.2f\n", percent_bytes);
 349                 fclose(fout);
 350         }
 351
 352         if (debugFlag)
 353                 read_trace_pipe2();
 354         return rc;
 355 err:
 356         rc = 1;
 357
 358         if (cg1)
 359                 close(cg1);
 360         cleanup_cgroup_environment();
 361
 362         return rc;
 363 }
 364
 365 static void Usage(void)
 366 {
 367         printf("This program loads a cgroup skb BPF program to enforce\n"
 368                "cgroup output (egress) bandwidth limits.\n\n"
 369                "USAGE: hbm [-o] [-d]  [-l] [-n <id>] [-r <rate>] [-s]\n"
 370                "           [-t <secs>] [-w] [-h] [prog]\n"
 371                "  Where:\n"
 372                "    -o         indicates egress direction (default)\n"
 373                "    -d         print BPF trace debug buffer\n"
 374                "    -l         also limit flows using loopback\n"
 375                "    -n <#>     to create cgroup \"/hbm#\" and attach prog\n"
 376                "               Default is /hbm1\n"
 377                "    -r <rate>  Rate in Mbps\n"
 378                "    -s         Update HBM stats\n"
 379                "    -t <time>  Exit after specified seconds (deault is 0)\n"
 380                "    -w         Work conserving flag. cgroup can increase\n"
 381                "               bandwidth beyond the rate limit specified\n"
 382                "               while there is available bandwidth. Current\n"
 383                "               implementation assumes there is only eth0\n"
 384                "               but can be extended to support multiple NICs\n"
 385                "    -h         print this info\n"
 386                "    prog       BPF program file name. Name defaults to\n"
 387                "                 hbm_out_kern.o\n");
 388 }
 389
 390 int main(int argc, char **argv)
 391 {
 392         char *prog = "hbm_out_kern.o";
 393         int  k;
 394         int cg_id = 1;
 395         char *optstring = "iodln:r:st:wh";
 396
 397         while ((k = getopt(argc, argv, optstring)) != -1) {
 398                 switch (k) {
 399                 case'o':
 400                         break;
 401                 case 'd':
 402                         debugFlag = true;
 403                         break;
 404                 case 'l':
 405                         loopback_flag = true;
 406                         break;
 407                 case 'n':
 408                         cg_id = atoi(optarg);
 409                         break;
 410                 case 'r':
 411                         minRate = atoi(optarg) * 1.024;
 412                         rate = minRate;
 413                         break;
 414                 case 's':
 415                         stats_flag = true;
 416                         break;
 417                 case 't':
 418                         dur = atoi(optarg);
 419                         break;
 420                 case 'w':
 421                         work_conserving_flag = true;
 422                         break;
 423                 case '?':
 424                         if (optopt == 'n' || optopt == 'r' || optopt == 't')
 425                                 fprintf(stderr,
 426                                         "Option -%c requires an argument.\n\n",
 427                                         optopt);
 428                 case 'h':
 429                         // fallthrough
 430                 default:
 431                         Usage();
 432                         return 0;
 433                 }
 434         }
 435
 436         if (optind < argc)
 437                 prog = argv[optind];
 438         printf("HBM prog: %s\n", prog != NULL ? prog : "NULL");
 439
 440         return run_bpf_prog(prog, cg_id);
 441 }