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 44 #include <linux/bpf.h> 45 #include <bpf/bpf.h> 46 #include <getopt.h> 47 48 #include "bpf_load.h" 49 #include "bpf_rlimit.h" 50 #include "cgroup_helpers.h" 51 #include "hbm.h" 52 #include "bpf_util.h" 53 #include "bpf.h" 54 #include "libbpf.h" 55 56 bool outFlag = true; 57 int minRate = 1000; /* cgroup rate limit in Mbps */ 58 int rate = 1000; /* can grow if rate conserving is enabled */ 59 int dur = 1; 60 bool stats_flag; 61 bool loopback_flag; 62 bool debugFlag; 63 bool work_conserving_flag; 64 bool no_cn_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 struct bpf_object *obj; 73 int bpfprog_fd; 74 int cgroup_storage_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_prog_load_attr prog_load_attr = { 124 .prog_type = BPF_PROG_TYPE_CGROUP_SKB, 125 .file = prog, 126 .expected_attach_type = BPF_CGROUP_INET_EGRESS, 127 }; 128 int map_fd; 129 struct bpf_map *map; 130 131 int ret = 0; 132 133 if (access(prog, O_RDONLY) < 0) { 134 printf("Error accessing file %s: %s\n", prog, strerror(errno)); 135 return 1; 136 } 137 if (bpf_prog_load_xattr(&prog_load_attr, &obj, &bpfprog_fd)) 138 ret = 1; 139 if (!ret) { 140 map = bpf_object__find_map_by_name(obj, "queue_stats"); 141 map_fd = bpf_map__fd(map); 142 if (map_fd < 0) { 143 printf("Map not found: %s\n", strerror(map_fd)); 144 ret = 1; 145 } 146 } 147 148 if (ret) { 149 printf("ERROR: load_bpf_file failed for: %s\n", prog); 150 printf(" Output from verifier:\n%s\n------\n", bpf_log_buf); 151 ret = -1; 152 } else { 153 ret = map_fd; 154 } 155 156 return ret; 157 } 158 159 static int run_bpf_prog(char *prog, int cg_id) 160 { 161 int map_fd; 162 int rc = 0; 163 int key = 0; 164 int cg1 = 0; 165 int type = BPF_CGROUP_INET_EGRESS; 166 char cg_dir[100]; 167 struct hbm_queue_stats qstats = {0}; 168 169 sprintf(cg_dir, "/hbm%d", cg_id); 170 map_fd = prog_load(prog); 171 if (map_fd == -1) 172 return 1; 173 174 if (setup_cgroup_environment()) { 175 printf("ERROR: setting cgroup environment\n"); 176 goto err; 177 } 178 cg1 = create_and_get_cgroup(cg_dir); 179 if (!cg1) { 180 printf("ERROR: create_and_get_cgroup\n"); 181 goto err; 182 } 183 if (join_cgroup(cg_dir)) { 184 printf("ERROR: join_cgroup\n"); 185 goto err; 186 } 187 188 qstats.rate = rate; 189 qstats.stats = stats_flag ? 1 : 0; 190 qstats.loopback = loopback_flag ? 1 : 0; 191 qstats.no_cn = no_cn_flag ? 1 : 0; 192 if (bpf_map_update_elem(map_fd, &key, &qstats, BPF_ANY)) { 193 printf("ERROR: Could not update map element\n"); 194 goto err; 195 } 196 197 if (!outFlag) 198 type = BPF_CGROUP_INET_INGRESS; 199 if (bpf_prog_attach(bpfprog_fd, cg1, type, 0)) { 200 printf("ERROR: bpf_prog_attach fails!\n"); 201 log_err("Attaching prog"); 202 goto err; 203 } 204 205 if (work_conserving_flag) { 206 struct timeval t0, t_last, t_new; 207 FILE *fin; 208 unsigned long long last_eth_tx_bytes, new_eth_tx_bytes; 209 signed long long last_cg_tx_bytes, new_cg_tx_bytes; 210 signed long long delta_time, delta_bytes, delta_rate; 211 int delta_ms; 212 #define DELTA_RATE_CHECK 10000 /* in us */ 213 #define RATE_THRESHOLD 9500000000 /* 9.5 Gbps */ 214 215 bpf_map_lookup_elem(map_fd, &key, &qstats); 216 if (gettimeofday(&t0, NULL) < 0) 217 do_error("gettimeofday failed", true); 218 t_last = t0; 219 fin = fopen("/sys/class/net/eth0/statistics/tx_bytes", "r"); 220 if (fscanf(fin, "%llu", &last_eth_tx_bytes) != 1) 221 do_error("fscanf fails", false); 222 fclose(fin); 223 last_cg_tx_bytes = qstats.bytes_total; 224 while (true) { 225 usleep(DELTA_RATE_CHECK); 226 if (gettimeofday(&t_new, NULL) < 0) 227 do_error("gettimeofday failed", true); 228 delta_ms = (t_new.tv_sec - t0.tv_sec) * 1000 + 229 (t_new.tv_usec - t0.tv_usec)/1000; 230 if (delta_ms > dur * 1000) 231 break; 232 delta_time = (t_new.tv_sec - t_last.tv_sec) * 1000000 + 233 (t_new.tv_usec - t_last.tv_usec); 234 if (delta_time == 0) 235 continue; 236 t_last = t_new; 237 fin = fopen("/sys/class/net/eth0/statistics/tx_bytes", 238 "r"); 239 if (fscanf(fin, "%llu", &new_eth_tx_bytes) != 1) 240 do_error("fscanf fails", false); 241 fclose(fin); 242 printf(" new_eth_tx_bytes:%llu\n", 243 new_eth_tx_bytes); 244 bpf_map_lookup_elem(map_fd, &key, &qstats); 245 new_cg_tx_bytes = qstats.bytes_total; 246 delta_bytes = new_eth_tx_bytes - last_eth_tx_bytes; 247 last_eth_tx_bytes = new_eth_tx_bytes; 248 delta_rate = (delta_bytes * 8000000) / delta_time; 249 printf("%5d - eth_rate:%.1fGbps cg_rate:%.3fGbps", 250 delta_ms, delta_rate/1000000000.0, 251 rate/1000.0); 252 if (delta_rate < RATE_THRESHOLD) { 253 /* can increase cgroup rate limit, but first 254 * check if we are using the current limit. 255 * Currently increasing by 6.25%, unknown 256 * if that is the optimal rate. 257 */ 258 int rate_diff100; 259 260 delta_bytes = new_cg_tx_bytes - 261 last_cg_tx_bytes; 262 last_cg_tx_bytes = new_cg_tx_bytes; 263 delta_rate = (delta_bytes * 8000000) / 264 delta_time; 265 printf(" rate:%.3fGbps", 266 delta_rate/1000000000.0); 267 rate_diff100 = (((long long)rate)*1000000 - 268 delta_rate) * 100 / 269 (((long long) rate) * 1000000); 270 printf(" rdiff:%d", rate_diff100); 271 if (rate_diff100 <= 3) { 272 rate += (rate >> 4); 273 if (rate > RATE_THRESHOLD / 1000000) 274 rate = RATE_THRESHOLD / 1000000; 275 qstats.rate = rate; 276 printf(" INC\n"); 277 } else { 278 printf("\n"); 279 } 280 } else { 281 /* Need to decrease cgroup rate limit. 282 * Currently decreasing by 12.5%, unknown 283 * if that is optimal 284 */ 285 printf(" DEC\n"); 286 rate -= (rate >> 3); 287 if (rate < minRate) 288 rate = minRate; 289 qstats.rate = rate; 290 } 291 if (bpf_map_update_elem(map_fd, &key, &qstats, BPF_ANY)) 292 do_error("update map element fails", false); 293 } 294 } else { 295 sleep(dur); 296 } 297 // Get stats! 298 if (stats_flag && bpf_map_lookup_elem(map_fd, &key, &qstats)) { 299 char fname[100]; 300 FILE *fout; 301 302 if (!outFlag) 303 sprintf(fname, "hbm.%d.in", cg_id); 304 else 305 sprintf(fname, "hbm.%d.out", cg_id); 306 fout = fopen(fname, "w"); 307 fprintf(fout, "id:%d\n", cg_id); 308 fprintf(fout, "ERROR: Could not lookup queue_stats\n"); 309 } else if (stats_flag && qstats.lastPacketTime > 310 qstats.firstPacketTime) { 311 long long delta_us = (qstats.lastPacketTime - 312 qstats.firstPacketTime)/1000; 313 unsigned int rate_mbps = ((qstats.bytes_total - 314 qstats.bytes_dropped) * 8 / 315 delta_us); 316 double percent_pkts, percent_bytes; 317 char fname[100]; 318 FILE *fout; 319 int k; 320 static const char *returnValNames[] = { 321 "DROP_PKT", 322 "ALLOW_PKT", 323 "DROP_PKT_CWR", 324 "ALLOW_PKT_CWR" 325 }; 326 #define RET_VAL_COUNT 4 327 328 // Future support of ingress 329 // if (!outFlag) 330 // sprintf(fname, "hbm.%d.in", cg_id); 331 // else 332 sprintf(fname, "hbm.%d.out", cg_id); 333 fout = fopen(fname, "w"); 334 fprintf(fout, "id:%d\n", cg_id); 335 fprintf(fout, "rate_mbps:%d\n", rate_mbps); 336 fprintf(fout, "duration:%.1f secs\n", 337 (qstats.lastPacketTime - qstats.firstPacketTime) / 338 1000000000.0); 339 fprintf(fout, "packets:%d\n", (int)qstats.pkts_total); 340 fprintf(fout, "bytes_MB:%d\n", (int)(qstats.bytes_total / 341 1000000)); 342 fprintf(fout, "pkts_dropped:%d\n", (int)qstats.pkts_dropped); 343 fprintf(fout, "bytes_dropped_MB:%d\n", 344 (int)(qstats.bytes_dropped / 345 1000000)); 346 // Marked Pkts and Bytes 347 percent_pkts = (qstats.pkts_marked * 100.0) / 348 (qstats.pkts_total + 1); 349 percent_bytes = (qstats.bytes_marked * 100.0) / 350 (qstats.bytes_total + 1); 351 fprintf(fout, "pkts_marked_percent:%6.2f\n", percent_pkts); 352 fprintf(fout, "bytes_marked_percent:%6.2f\n", percent_bytes); 353 354 // Dropped Pkts and Bytes 355 percent_pkts = (qstats.pkts_dropped * 100.0) / 356 (qstats.pkts_total + 1); 357 percent_bytes = (qstats.bytes_dropped * 100.0) / 358 (qstats.bytes_total + 1); 359 fprintf(fout, "pkts_dropped_percent:%6.2f\n", percent_pkts); 360 fprintf(fout, "bytes_dropped_percent:%6.2f\n", percent_bytes); 361 362 // ECN CE markings 363 percent_pkts = (qstats.pkts_ecn_ce * 100.0) / 364 (qstats.pkts_total + 1); 365 fprintf(fout, "pkts_ecn_ce:%6.2f (%d)\n", percent_pkts, 366 (int)qstats.pkts_ecn_ce); 367 368 // Average cwnd 369 fprintf(fout, "avg cwnd:%d\n", 370 (int)(qstats.sum_cwnd / (qstats.sum_cwnd_cnt + 1))); 371 // Average rtt 372 fprintf(fout, "avg rtt:%d\n", 373 (int)(qstats.sum_rtt / (qstats.pkts_total + 1))); 374 // Average credit 375 fprintf(fout, "avg credit:%d\n", 376 (int)(qstats.sum_credit / 377 (1500 * ((int)qstats.pkts_total) + 1))); 378 379 // Return values stats 380 for (k = 0; k < RET_VAL_COUNT; k++) { 381 percent_pkts = (qstats.returnValCount[k] * 100.0) / 382 (qstats.pkts_total + 1); 383 fprintf(fout, "%s:%6.2f (%d)\n", returnValNames[k], 384 percent_pkts, (int)qstats.returnValCount[k]); 385 } 386 fclose(fout); 387 } 388 389 if (debugFlag) 390 read_trace_pipe2(); 391 return rc; 392 err: 393 rc = 1; 394 395 if (cg1) 396 close(cg1); 397 cleanup_cgroup_environment(); 398 399 return rc; 400 } 401 402 static void Usage(void) 403 { 404 printf("This program loads a cgroup skb BPF program to enforce\n" 405 "cgroup output (egress) bandwidth limits.\n\n" 406 "USAGE: hbm [-o] [-d] [-l] [-n <id>] [--no_cn] [-r <rate>]\n" 407 " [-s] [-t <secs>] [-w] [-h] [prog]\n" 408 " Where:\n" 409 " -o indicates egress direction (default)\n" 410 " -d print BPF trace debug buffer\n" 411 " -l also limit flows using loopback\n" 412 " -n <#> to create cgroup \"/hbm#\" and attach prog\n" 413 " Default is /hbm1\n" 414 " --no_cn disable CN notifications\n" 415 " -r <rate> Rate in Mbps\n" 416 " -s Update HBM stats\n" 417 " -t <time> Exit after specified seconds (default is 0)\n" 418 " -w Work conserving flag. cgroup can increase\n" 419 " bandwidth beyond the rate limit specified\n" 420 " while there is available bandwidth. Current\n" 421 " implementation assumes there is only eth0\n" 422 " but can be extended to support multiple NICs\n" 423 " -h print this info\n" 424 " prog BPF program file name. Name defaults to\n" 425 " hbm_out_kern.o\n"); 426 } 427 428 int main(int argc, char **argv) 429 { 430 char *prog = "hbm_out_kern.o"; 431 int k; 432 int cg_id = 1; 433 char *optstring = "iodln:r:st:wh"; 434 struct option loptions[] = { 435 {"no_cn", 0, NULL, 1}, 436 {NULL, 0, NULL, 0} 437 }; 438 439 while ((k = getopt_long(argc, argv, optstring, loptions, NULL)) != -1) { 440 switch (k) { 441 case 1: 442 no_cn_flag = true; 443 break; 444 case'o': 445 break; 446 case 'd': 447 debugFlag = true; 448 break; 449 case 'l': 450 loopback_flag = true; 451 break; 452 case 'n': 453 cg_id = atoi(optarg); 454 break; 455 case 'r': 456 minRate = atoi(optarg) * 1.024; 457 rate = minRate; 458 break; 459 case 's': 460 stats_flag = true; 461 break; 462 case 't': 463 dur = atoi(optarg); 464 break; 465 case 'w': 466 work_conserving_flag = true; 467 break; 468 case '?': 469 if (optopt == 'n' || optopt == 'r' || optopt == 't') 470 fprintf(stderr, 471 "Option -%c requires an argument.\n\n", 472 optopt); 473 case 'h': 474 // fallthrough 475 default: 476 Usage(); 477 return 0; 478 } 479 } 480 481 if (optind < argc) 482 prog = argv[optind]; 483 printf("HBM prog: %s\n", prog != NULL ? prog : "NULL"); 484 485 return run_bpf_prog(prog, cg_id); 486 } 487