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 obj = bpf_object__open_file(prog, NULL); 124 if (libbpf_get_error(obj)) { 125 printf("ERROR: opening BPF object file failed\n"); 126 return 1; 127 } 128 129 /* load BPF program */ 130 if (bpf_object__load(obj)) { 131 printf("ERROR: loading BPF object file failed\n"); 132 goto err; 133 } 134 135 bpf_prog = bpf_object__find_program_by_title(obj, "cgroup_skb/egress"); 136 if (!bpf_prog) { 137 printf("ERROR: finding a prog in obj file failed\n"); 138 goto err; 139 } 140 141 queue_stats_fd = bpf_object__find_map_fd_by_name(obj, "queue_stats"); 142 if (queue_stats_fd < 0) { 143 printf("ERROR: finding a map in obj file failed\n"); 144 goto err; 145 } 146 147 return 0; 148 149 err: 150 bpf_object__close(obj); 151 return 1; 152 } 153 154 static int run_bpf_prog(char *prog, int cg_id) 155 { 156 struct hbm_queue_stats qstats = {0}; 157 char cg_dir[100], cg_pin_path[100]; 158 struct bpf_link *link = NULL; 159 int key = 0; 160 int cg1 = 0; 161 int rc = 0; 162 163 sprintf(cg_dir, "/hbm%d", cg_id); 164 rc = prog_load(prog); 165 if (rc != 0) 166 return rc; 167 168 if (setup_cgroup_environment()) { 169 printf("ERROR: setting cgroup environment\n"); 170 goto err; 171 } 172 cg1 = create_and_get_cgroup(cg_dir); 173 if (!cg1) { 174 printf("ERROR: create_and_get_cgroup\n"); 175 goto err; 176 } 177 if (join_cgroup(cg_dir)) { 178 printf("ERROR: join_cgroup\n"); 179 goto err; 180 } 181 182 qstats.rate = rate; 183 qstats.stats = stats_flag ? 1 : 0; 184 qstats.loopback = loopback_flag ? 1 : 0; 185 qstats.no_cn = no_cn_flag ? 1 : 0; 186 if (bpf_map_update_elem(queue_stats_fd, &key, &qstats, BPF_ANY)) { 187 printf("ERROR: Could not update map element\n"); 188 goto err; 189 } 190 191 if (!outFlag) 192 bpf_program__set_expected_attach_type(bpf_prog, BPF_CGROUP_INET_INGRESS); 193 194 link = bpf_program__attach_cgroup(bpf_prog, cg1); 195 if (libbpf_get_error(link)) { 196 fprintf(stderr, "ERROR: bpf_program__attach_cgroup failed\n"); 197 goto err; 198 } 199 200 sprintf(cg_pin_path, "/sys/fs/bpf/hbm%d", cg_id); 201 rc = bpf_link__pin(link, cg_pin_path); 202 if (rc < 0) { 203 printf("ERROR: bpf_link__pin failed: %d\n", rc); 204 goto err; 205 } 206 207 if (work_conserving_flag) { 208 struct timeval t0, t_last, t_new; 209 FILE *fin; 210 unsigned long long last_eth_tx_bytes, new_eth_tx_bytes; 211 signed long long last_cg_tx_bytes, new_cg_tx_bytes; 212 signed long long delta_time, delta_bytes, delta_rate; 213 int delta_ms; 214 #define DELTA_RATE_CHECK 10000 /* in us */ 215 #define RATE_THRESHOLD 9500000000 /* 9.5 Gbps */ 216 217 bpf_map_lookup_elem(queue_stats_fd, &key, &qstats); 218 if (gettimeofday(&t0, NULL) < 0) 219 do_error("gettimeofday failed", true); 220 t_last = t0; 221 fin = fopen("/sys/class/net/eth0/statistics/tx_bytes", "r"); 222 if (fscanf(fin, "%llu", &last_eth_tx_bytes) != 1) 223 do_error("fscanf fails", false); 224 fclose(fin); 225 last_cg_tx_bytes = qstats.bytes_total; 226 while (true) { 227 usleep(DELTA_RATE_CHECK); 228 if (gettimeofday(&t_new, NULL) < 0) 229 do_error("gettimeofday failed", true); 230 delta_ms = (t_new.tv_sec - t0.tv_sec) * 1000 + 231 (t_new.tv_usec - t0.tv_usec)/1000; 232 if (delta_ms > dur * 1000) 233 break; 234 delta_time = (t_new.tv_sec - t_last.tv_sec) * 1000000 + 235 (t_new.tv_usec - t_last.tv_usec); 236 if (delta_time == 0) 237 continue; 238 t_last = t_new; 239 fin = fopen("/sys/class/net/eth0/statistics/tx_bytes", 240 "r"); 241 if (fscanf(fin, "%llu", &new_eth_tx_bytes) != 1) 242 do_error("fscanf fails", false); 243 fclose(fin); 244 printf(" new_eth_tx_bytes:%llu\n", 245 new_eth_tx_bytes); 246 bpf_map_lookup_elem(queue_stats_fd, &key, &qstats); 247 new_cg_tx_bytes = qstats.bytes_total; 248 delta_bytes = new_eth_tx_bytes - last_eth_tx_bytes; 249 last_eth_tx_bytes = new_eth_tx_bytes; 250 delta_rate = (delta_bytes * 8000000) / delta_time; 251 printf("%5d - eth_rate:%.1fGbps cg_rate:%.3fGbps", 252 delta_ms, delta_rate/1000000000.0, 253 rate/1000.0); 254 if (delta_rate < RATE_THRESHOLD) { 255 /* can increase cgroup rate limit, but first 256 * check if we are using the current limit. 257 * Currently increasing by 6.25%, unknown 258 * if that is the optimal rate. 259 */ 260 int rate_diff100; 261 262 delta_bytes = new_cg_tx_bytes - 263 last_cg_tx_bytes; 264 last_cg_tx_bytes = new_cg_tx_bytes; 265 delta_rate = (delta_bytes * 8000000) / 266 delta_time; 267 printf(" rate:%.3fGbps", 268 delta_rate/1000000000.0); 269 rate_diff100 = (((long long)rate)*1000000 - 270 delta_rate) * 100 / 271 (((long long) rate) * 1000000); 272 printf(" rdiff:%d", rate_diff100); 273 if (rate_diff100 <= 3) { 274 rate += (rate >> 4); 275 if (rate > RATE_THRESHOLD / 1000000) 276 rate = RATE_THRESHOLD / 1000000; 277 qstats.rate = rate; 278 printf(" INC\n"); 279 } else { 280 printf("\n"); 281 } 282 } else { 283 /* Need to decrease cgroup rate limit. 284 * Currently decreasing by 12.5%, unknown 285 * if that is optimal 286 */ 287 printf(" DEC\n"); 288 rate -= (rate >> 3); 289 if (rate < minRate) 290 rate = minRate; 291 qstats.rate = rate; 292 } 293 if (bpf_map_update_elem(queue_stats_fd, &key, &qstats, BPF_ANY)) 294 do_error("update map element fails", false); 295 } 296 } else { 297 sleep(dur); 298 } 299 // Get stats! 300 if (stats_flag && bpf_map_lookup_elem(queue_stats_fd, &key, &qstats)) { 301 char fname[100]; 302 FILE *fout; 303 304 if (!outFlag) 305 sprintf(fname, "hbm.%d.in", cg_id); 306 else 307 sprintf(fname, "hbm.%d.out", cg_id); 308 fout = fopen(fname, "w"); 309 fprintf(fout, "id:%d\n", cg_id); 310 fprintf(fout, "ERROR: Could not lookup queue_stats\n"); 311 } else if (stats_flag && qstats.lastPacketTime > 312 qstats.firstPacketTime) { 313 long long delta_us = (qstats.lastPacketTime - 314 qstats.firstPacketTime)/1000; 315 unsigned int rate_mbps = ((qstats.bytes_total - 316 qstats.bytes_dropped) * 8 / 317 delta_us); 318 double percent_pkts, percent_bytes; 319 char fname[100]; 320 FILE *fout; 321 int k; 322 static const char *returnValNames[] = { 323 "DROP_PKT", 324 "ALLOW_PKT", 325 "DROP_PKT_CWR", 326 "ALLOW_PKT_CWR" 327 }; 328 #define RET_VAL_COUNT 4 329 330 // Future support of ingress 331 // if (!outFlag) 332 // sprintf(fname, "hbm.%d.in", cg_id); 333 // else 334 sprintf(fname, "hbm.%d.out", cg_id); 335 fout = fopen(fname, "w"); 336 fprintf(fout, "id:%d\n", cg_id); 337 fprintf(fout, "rate_mbps:%d\n", rate_mbps); 338 fprintf(fout, "duration:%.1f secs\n", 339 (qstats.lastPacketTime - qstats.firstPacketTime) / 340 1000000000.0); 341 fprintf(fout, "packets:%d\n", (int)qstats.pkts_total); 342 fprintf(fout, "bytes_MB:%d\n", (int)(qstats.bytes_total / 343 1000000)); 344 fprintf(fout, "pkts_dropped:%d\n", (int)qstats.pkts_dropped); 345 fprintf(fout, "bytes_dropped_MB:%d\n", 346 (int)(qstats.bytes_dropped / 347 1000000)); 348 // Marked Pkts and Bytes 349 percent_pkts = (qstats.pkts_marked * 100.0) / 350 (qstats.pkts_total + 1); 351 percent_bytes = (qstats.bytes_marked * 100.0) / 352 (qstats.bytes_total + 1); 353 fprintf(fout, "pkts_marked_percent:%6.2f\n", percent_pkts); 354 fprintf(fout, "bytes_marked_percent:%6.2f\n", percent_bytes); 355 356 // Dropped Pkts and Bytes 357 percent_pkts = (qstats.pkts_dropped * 100.0) / 358 (qstats.pkts_total + 1); 359 percent_bytes = (qstats.bytes_dropped * 100.0) / 360 (qstats.bytes_total + 1); 361 fprintf(fout, "pkts_dropped_percent:%6.2f\n", percent_pkts); 362 fprintf(fout, "bytes_dropped_percent:%6.2f\n", percent_bytes); 363 364 // ECN CE markings 365 percent_pkts = (qstats.pkts_ecn_ce * 100.0) / 366 (qstats.pkts_total + 1); 367 fprintf(fout, "pkts_ecn_ce:%6.2f (%d)\n", percent_pkts, 368 (int)qstats.pkts_ecn_ce); 369 370 // Average cwnd 371 fprintf(fout, "avg cwnd:%d\n", 372 (int)(qstats.sum_cwnd / (qstats.sum_cwnd_cnt + 1))); 373 // Average rtt 374 fprintf(fout, "avg rtt:%d\n", 375 (int)(qstats.sum_rtt / (qstats.pkts_total + 1))); 376 // Average credit 377 if (edt_flag) 378 fprintf(fout, "avg credit_ms:%.03f\n", 379 (qstats.sum_credit / 380 (qstats.pkts_total + 1.0)) / 1000000.0); 381 else 382 fprintf(fout, "avg credit:%d\n", 383 (int)(qstats.sum_credit / 384 (1500 * ((int)qstats.pkts_total ) + 1))); 385 386 // Return values stats 387 for (k = 0; k < RET_VAL_COUNT; k++) { 388 percent_pkts = (qstats.returnValCount[k] * 100.0) / 389 (qstats.pkts_total + 1); 390 fprintf(fout, "%s:%6.2f (%d)\n", returnValNames[k], 391 percent_pkts, (int)qstats.returnValCount[k]); 392 } 393 fclose(fout); 394 } 395 396 if (debugFlag) 397 read_trace_pipe2(); 398 goto cleanup; 399 400 err: 401 rc = 1; 402 403 cleanup: 404 bpf_link__destroy(link); 405 bpf_object__close(obj); 406 407 if (cg1 != -1) 408 close(cg1); 409 410 if (rc != 0) 411 cleanup_cgroup_environment(); 412 return rc; 413 } 414 415 static void Usage(void) 416 { 417 printf("This program loads a cgroup skb BPF program to enforce\n" 418 "cgroup output (egress) bandwidth limits.\n\n" 419 "USAGE: hbm [-o] [-d] [-l] [-n <id>] [--no_cn] [-r <rate>]\n" 420 " [-s] [-t <secs>] [-w] [-h] [prog]\n" 421 " Where:\n" 422 " -o indicates egress direction (default)\n" 423 " -d print BPF trace debug buffer\n" 424 " --edt use fq's Earliest Departure Time\n" 425 " -l also limit flows using loopback\n" 426 " -n <#> to create cgroup \"/hbm#\" and attach prog\n" 427 " Default is /hbm1\n" 428 " --no_cn disable CN notifications\n" 429 " -r <rate> Rate in Mbps\n" 430 " -s Update HBM stats\n" 431 " -t <time> Exit after specified seconds (default is 0)\n" 432 " -w Work conserving flag. cgroup can increase\n" 433 " bandwidth beyond the rate limit specified\n" 434 " while there is available bandwidth. Current\n" 435 " implementation assumes there is only eth0\n" 436 " but can be extended to support multiple NICs\n" 437 " -h print this info\n" 438 " prog BPF program file name. Name defaults to\n" 439 " hbm_out_kern.o\n"); 440 } 441 442 int main(int argc, char **argv) 443 { 444 char *prog = "hbm_out_kern.o"; 445 int k; 446 int cg_id = 1; 447 char *optstring = "iodln:r:st:wh"; 448 struct option loptions[] = { 449 {"no_cn", 0, NULL, 1}, 450 {"edt", 0, NULL, 2}, 451 {NULL, 0, NULL, 0} 452 }; 453 454 while ((k = getopt_long(argc, argv, optstring, loptions, NULL)) != -1) { 455 switch (k) { 456 case 1: 457 no_cn_flag = true; 458 break; 459 case 2: 460 prog = "hbm_edt_kern.o"; 461 edt_flag = true; 462 break; 463 case'o': 464 break; 465 case 'd': 466 debugFlag = true; 467 break; 468 case 'l': 469 loopback_flag = true; 470 break; 471 case 'n': 472 cg_id = atoi(optarg); 473 break; 474 case 'r': 475 minRate = atoi(optarg) * 1.024; 476 rate = minRate; 477 break; 478 case 's': 479 stats_flag = true; 480 break; 481 case 't': 482 dur = atoi(optarg); 483 break; 484 case 'w': 485 work_conserving_flag = true; 486 break; 487 case '?': 488 if (optopt == 'n' || optopt == 'r' || optopt == 't') 489 fprintf(stderr, 490 "Option -%c requires an argument.\n\n", 491 optopt); 492 case 'h': 493 __fallthrough; 494 default: 495 Usage(); 496 return 0; 497 } 498 } 499 500 if (optind < argc) 501 prog = argv[optind]; 502 printf("HBM prog: %s\n", prog != NULL ? prog : "NULL"); 503 504 return run_bpf_prog(prog, cg_id); 505 } 506