1 // SPDX-License-Identifier: GPL-2.0 2 /* Toeplitz test 3 * 4 * 1. Read packets and their rx_hash using PF_PACKET/TPACKET_V3 5 * 2. Compute the rx_hash in software based on the packet contents 6 * 3. Compare the two 7 * 8 * Optionally, either '-C $rx_irq_cpu_list' or '-r $rps_bitmap' may be given. 9 * 10 * If '-C $rx_irq_cpu_list' is given, also 11 * 12 * 4. Identify the cpu on which the packet arrived with PACKET_FANOUT_CPU 13 * 5. Compute the rxqueue that RSS would select based on this rx_hash 14 * 6. Using the $rx_irq_cpu_list map, identify the arriving cpu based on rxq irq 15 * 7. Compare the cpus from 4 and 6 16 * 17 * Else if '-r $rps_bitmap' is given, also 18 * 19 * 4. Identify the cpu on which the packet arrived with PACKET_FANOUT_CPU 20 * 5. Compute the cpu that RPS should select based on rx_hash and $rps_bitmap 21 * 6. Compare the cpus from 4 and 5 22 */ 23 24 #define _GNU_SOURCE 25 26 #include <arpa/inet.h> 27 #include <errno.h> 28 #include <error.h> 29 #include <fcntl.h> 30 #include <getopt.h> 31 #include <linux/filter.h> 32 #include <linux/if_ether.h> 33 #include <linux/if_packet.h> 34 #include <net/if.h> 35 #include <netdb.h> 36 #include <netinet/ip.h> 37 #include <netinet/ip6.h> 38 #include <netinet/tcp.h> 39 #include <netinet/udp.h> 40 #include <poll.h> 41 #include <stdbool.h> 42 #include <stddef.h> 43 #include <stdint.h> 44 #include <stdio.h> 45 #include <stdlib.h> 46 #include <string.h> 47 #include <sys/mman.h> 48 #include <sys/socket.h> 49 #include <sys/stat.h> 50 #include <sys/sysinfo.h> 51 #include <sys/time.h> 52 #include <sys/types.h> 53 #include <unistd.h> 54 55 #include "../../../kselftest.h" 56 #include "../../../net/lib/ksft.h" 57 58 #define TOEPLITZ_KEY_MIN_LEN 40 59 #define TOEPLITZ_KEY_MAX_LEN 60 60 61 #define TOEPLITZ_STR_LEN(K) (((K) * 3) - 1) /* hex encoded: AA:BB:CC:...:ZZ */ 62 #define TOEPLITZ_STR_MIN_LEN TOEPLITZ_STR_LEN(TOEPLITZ_KEY_MIN_LEN) 63 #define TOEPLITZ_STR_MAX_LEN TOEPLITZ_STR_LEN(TOEPLITZ_KEY_MAX_LEN) 64 65 #define FOUR_TUPLE_MAX_LEN ((sizeof(struct in6_addr) * 2) + (sizeof(uint16_t) * 2)) 66 67 #define RSS_MAX_CPUS (1 << 16) /* real constraint is PACKET_FANOUT_MAX */ 68 69 #define RPS_MAX_CPUS 16UL /* must be a power of 2 */ 70 71 /* configuration options (cmdline arguments) */ 72 static uint16_t cfg_dport = 8000; 73 static int cfg_family = AF_INET6; 74 static char *cfg_ifname = "eth0"; 75 static int cfg_num_queues; 76 static int cfg_num_rps_cpus; 77 static bool cfg_sink; 78 static int cfg_type = SOCK_STREAM; 79 static int cfg_timeout_msec = 1000; 80 static bool cfg_verbose; 81 82 /* global vars */ 83 static int num_cpus; 84 static int ring_block_nr; 85 static int ring_block_sz; 86 87 /* stats */ 88 static int frames_received; 89 static int frames_nohash; 90 static int frames_error; 91 92 #define log_verbose(args...) do { if (cfg_verbose) fprintf(stderr, args); } while (0) 93 94 /* tpacket ring */ 95 struct ring_state { 96 int fd; 97 char *mmap; 98 int idx; 99 int cpu; 100 }; 101 102 static unsigned int rx_irq_cpus[RSS_MAX_CPUS]; /* map from rxq to cpu */ 103 static int rps_silo_to_cpu[RPS_MAX_CPUS]; 104 static unsigned char toeplitz_key[TOEPLITZ_KEY_MAX_LEN]; 105 static struct ring_state rings[RSS_MAX_CPUS]; 106 107 static inline uint32_t toeplitz(const unsigned char *four_tuple, 108 const unsigned char *key) 109 { 110 int i, bit, ret = 0; 111 uint32_t key32; 112 113 key32 = ntohl(*((uint32_t *)key)); 114 key += 4; 115 116 for (i = 0; i < FOUR_TUPLE_MAX_LEN; i++) { 117 for (bit = 7; bit >= 0; bit--) { 118 if (four_tuple[i] & (1 << bit)) 119 ret ^= key32; 120 121 key32 <<= 1; 122 key32 |= !!(key[0] & (1 << bit)); 123 } 124 key++; 125 } 126 127 return ret; 128 } 129 130 /* Compare computed cpu with arrival cpu from packet_fanout_cpu */ 131 static void verify_rss(uint32_t rx_hash, int cpu) 132 { 133 int queue = rx_hash % cfg_num_queues; 134 135 log_verbose(" rxq %d (cpu %d)", queue, rx_irq_cpus[queue]); 136 if (rx_irq_cpus[queue] != cpu) { 137 log_verbose(". error: rss cpu mismatch (%d)", cpu); 138 frames_error++; 139 } 140 } 141 142 static void verify_rps(uint64_t rx_hash, int cpu) 143 { 144 int silo = (rx_hash * cfg_num_rps_cpus) >> 32; 145 146 log_verbose(" silo %d (cpu %d)", silo, rps_silo_to_cpu[silo]); 147 if (rps_silo_to_cpu[silo] != cpu) { 148 log_verbose(". error: rps cpu mismatch (%d)", cpu); 149 frames_error++; 150 } 151 } 152 153 static void log_rxhash(int cpu, uint32_t rx_hash, 154 const char *addrs, int addr_len) 155 { 156 char saddr[INET6_ADDRSTRLEN], daddr[INET6_ADDRSTRLEN]; 157 uint16_t *ports; 158 159 if (!inet_ntop(cfg_family, addrs, saddr, sizeof(saddr)) || 160 !inet_ntop(cfg_family, addrs + addr_len, daddr, sizeof(daddr))) 161 error(1, 0, "address parse error"); 162 163 ports = (void *)addrs + (addr_len * 2); 164 log_verbose("cpu %d: rx_hash 0x%08x [saddr %s daddr %s sport %02hu dport %02hu]", 165 cpu, rx_hash, saddr, daddr, 166 ntohs(ports[0]), ntohs(ports[1])); 167 } 168 169 /* Compare computed rxhash with rxhash received from tpacket_v3 */ 170 static void verify_rxhash(const char *pkt, uint32_t rx_hash, int cpu) 171 { 172 unsigned char four_tuple[FOUR_TUPLE_MAX_LEN] = {0}; 173 uint32_t rx_hash_sw; 174 const char *addrs; 175 int addr_len; 176 177 if (cfg_family == AF_INET) { 178 addr_len = sizeof(struct in_addr); 179 addrs = pkt + offsetof(struct iphdr, saddr); 180 } else { 181 addr_len = sizeof(struct in6_addr); 182 addrs = pkt + offsetof(struct ip6_hdr, ip6_src); 183 } 184 185 memcpy(four_tuple, addrs, (addr_len * 2) + (sizeof(uint16_t) * 2)); 186 rx_hash_sw = toeplitz(four_tuple, toeplitz_key); 187 188 if (cfg_verbose) 189 log_rxhash(cpu, rx_hash, addrs, addr_len); 190 191 if (rx_hash != rx_hash_sw) { 192 log_verbose(" != expected 0x%x\n", rx_hash_sw); 193 frames_error++; 194 return; 195 } 196 197 log_verbose(" OK"); 198 if (cfg_num_queues) 199 verify_rss(rx_hash, cpu); 200 else if (cfg_num_rps_cpus) 201 verify_rps(rx_hash, cpu); 202 log_verbose("\n"); 203 } 204 205 static char *recv_frame(const struct ring_state *ring, char *frame) 206 { 207 struct tpacket3_hdr *hdr = (void *)frame; 208 209 if (hdr->hv1.tp_rxhash) 210 verify_rxhash(frame + hdr->tp_net, hdr->hv1.tp_rxhash, 211 ring->cpu); 212 else 213 frames_nohash++; 214 215 return frame + hdr->tp_next_offset; 216 } 217 218 /* A single TPACKET_V3 block can hold multiple frames */ 219 static bool recv_block(struct ring_state *ring) 220 { 221 struct tpacket_block_desc *block; 222 char *frame; 223 int i; 224 225 block = (void *)(ring->mmap + ring->idx * ring_block_sz); 226 if (!(block->hdr.bh1.block_status & TP_STATUS_USER)) 227 return false; 228 229 frame = (char *)block; 230 frame += block->hdr.bh1.offset_to_first_pkt; 231 232 for (i = 0; i < block->hdr.bh1.num_pkts; i++) { 233 frame = recv_frame(ring, frame); 234 frames_received++; 235 } 236 237 block->hdr.bh1.block_status = TP_STATUS_KERNEL; 238 ring->idx = (ring->idx + 1) % ring_block_nr; 239 240 return true; 241 } 242 243 /* simple test: sleep once unconditionally and then process all rings */ 244 static void process_rings(void) 245 { 246 int i; 247 248 usleep(1000 * cfg_timeout_msec); 249 250 for (i = 0; i < num_cpus; i++) 251 do {} while (recv_block(&rings[i])); 252 253 fprintf(stderr, "count: pass=%u nohash=%u fail=%u\n", 254 frames_received - frames_nohash - frames_error, 255 frames_nohash, frames_error); 256 } 257 258 static char *setup_ring(int fd) 259 { 260 struct tpacket_req3 req3 = {0}; 261 void *ring; 262 263 req3.tp_retire_blk_tov = cfg_timeout_msec / 8; 264 req3.tp_feature_req_word = TP_FT_REQ_FILL_RXHASH; 265 266 req3.tp_frame_size = 2048; 267 req3.tp_frame_nr = 1 << 10; 268 req3.tp_block_nr = 16; 269 270 req3.tp_block_size = req3.tp_frame_size * req3.tp_frame_nr; 271 req3.tp_block_size /= req3.tp_block_nr; 272 273 if (setsockopt(fd, SOL_PACKET, PACKET_RX_RING, &req3, sizeof(req3))) 274 error(1, errno, "setsockopt PACKET_RX_RING"); 275 276 ring_block_sz = req3.tp_block_size; 277 ring_block_nr = req3.tp_block_nr; 278 279 ring = mmap(0, req3.tp_block_size * req3.tp_block_nr, 280 PROT_READ | PROT_WRITE, 281 MAP_SHARED | MAP_LOCKED | MAP_POPULATE, fd, 0); 282 if (ring == MAP_FAILED) 283 error(1, 0, "mmap failed"); 284 285 return ring; 286 } 287 288 static void __set_filter(int fd, int off_proto, uint8_t proto, int off_dport) 289 { 290 struct sock_filter filter[] = { 291 BPF_STMT(BPF_LD + BPF_B + BPF_ABS, SKF_AD_OFF + SKF_AD_PKTTYPE), 292 BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, PACKET_HOST, 0, 4), 293 BPF_STMT(BPF_LD + BPF_B + BPF_ABS, off_proto), 294 BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, proto, 0, 2), 295 BPF_STMT(BPF_LD + BPF_H + BPF_ABS, off_dport), 296 BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, cfg_dport, 1, 0), 297 BPF_STMT(BPF_RET + BPF_K, 0), 298 BPF_STMT(BPF_RET + BPF_K, 0xFFFF), 299 }; 300 struct sock_fprog prog = {}; 301 302 prog.filter = filter; 303 prog.len = ARRAY_SIZE(filter); 304 if (setsockopt(fd, SOL_SOCKET, SO_ATTACH_FILTER, &prog, sizeof(prog))) 305 error(1, errno, "setsockopt filter"); 306 } 307 308 /* filter on transport protocol and destination port */ 309 static void set_filter(int fd) 310 { 311 const int off_dport = offsetof(struct tcphdr, dest); /* same for udp */ 312 uint8_t proto; 313 314 proto = cfg_type == SOCK_STREAM ? IPPROTO_TCP : IPPROTO_UDP; 315 if (cfg_family == AF_INET) 316 __set_filter(fd, offsetof(struct iphdr, protocol), proto, 317 sizeof(struct iphdr) + off_dport); 318 else 319 __set_filter(fd, offsetof(struct ip6_hdr, ip6_nxt), proto, 320 sizeof(struct ip6_hdr) + off_dport); 321 } 322 323 /* drop everything: used temporarily during setup */ 324 static void set_filter_null(int fd) 325 { 326 struct sock_filter filter[] = { 327 BPF_STMT(BPF_RET + BPF_K, 0), 328 }; 329 struct sock_fprog prog = {}; 330 331 prog.filter = filter; 332 prog.len = ARRAY_SIZE(filter); 333 if (setsockopt(fd, SOL_SOCKET, SO_ATTACH_FILTER, &prog, sizeof(prog))) 334 error(1, errno, "setsockopt filter"); 335 } 336 337 static int create_ring(char **ring) 338 { 339 struct fanout_args args = { 340 .id = 1, 341 .type_flags = PACKET_FANOUT_CPU, 342 .max_num_members = RSS_MAX_CPUS 343 }; 344 struct sockaddr_ll ll = { 0 }; 345 int fd, val; 346 347 fd = socket(PF_PACKET, SOCK_DGRAM, 0); 348 if (fd == -1) 349 error(1, errno, "socket creation failed"); 350 351 val = TPACKET_V3; 352 if (setsockopt(fd, SOL_PACKET, PACKET_VERSION, &val, sizeof(val))) 353 error(1, errno, "setsockopt PACKET_VERSION"); 354 *ring = setup_ring(fd); 355 356 /* block packets until all rings are added to the fanout group: 357 * else packets can arrive during setup and get misclassified 358 */ 359 set_filter_null(fd); 360 361 ll.sll_family = AF_PACKET; 362 ll.sll_ifindex = if_nametoindex(cfg_ifname); 363 ll.sll_protocol = cfg_family == AF_INET ? htons(ETH_P_IP) : 364 htons(ETH_P_IPV6); 365 if (bind(fd, (void *)&ll, sizeof(ll))) 366 error(1, errno, "bind"); 367 368 /* must come after bind: verifies all programs in group match */ 369 if (setsockopt(fd, SOL_PACKET, PACKET_FANOUT, &args, sizeof(args))) { 370 /* on failure, retry using old API if that is sufficient: 371 * it has a hard limit of 256 sockets, so only try if 372 * (a) only testing rxhash, not RSS or (b) <= 256 cpus. 373 * in this API, the third argument is left implicit. 374 */ 375 if (cfg_num_queues || num_cpus > 256 || 376 setsockopt(fd, SOL_PACKET, PACKET_FANOUT, 377 &args, sizeof(uint32_t))) 378 error(1, errno, "setsockopt PACKET_FANOUT cpu"); 379 } 380 381 return fd; 382 } 383 384 /* setup inet(6) socket to blackhole the test traffic, if arg '-s' */ 385 static int setup_sink(void) 386 { 387 int fd, val; 388 389 fd = socket(cfg_family, cfg_type, 0); 390 if (fd == -1) 391 error(1, errno, "socket %d.%d", cfg_family, cfg_type); 392 393 val = 1 << 20; 394 if (setsockopt(fd, SOL_SOCKET, SO_RCVBUFFORCE, &val, sizeof(val))) 395 error(1, errno, "setsockopt rcvbuf"); 396 397 return fd; 398 } 399 400 static void setup_rings(void) 401 { 402 int i; 403 404 for (i = 0; i < num_cpus; i++) { 405 rings[i].cpu = i; 406 rings[i].fd = create_ring(&rings[i].mmap); 407 } 408 409 /* accept packets once all rings in the fanout group are up */ 410 for (i = 0; i < num_cpus; i++) 411 set_filter(rings[i].fd); 412 } 413 414 static void cleanup_rings(void) 415 { 416 int i; 417 418 for (i = 0; i < num_cpus; i++) { 419 if (munmap(rings[i].mmap, ring_block_nr * ring_block_sz)) 420 error(1, errno, "munmap"); 421 if (close(rings[i].fd)) 422 error(1, errno, "close"); 423 } 424 } 425 426 static void parse_cpulist(const char *arg) 427 { 428 do { 429 rx_irq_cpus[cfg_num_queues++] = strtol(arg, NULL, 10); 430 431 arg = strchr(arg, ','); 432 if (!arg) 433 break; 434 arg++; // skip ',' 435 } while (1); 436 } 437 438 static void show_cpulist(void) 439 { 440 int i; 441 442 for (i = 0; i < cfg_num_queues; i++) 443 fprintf(stderr, "rxq %d: cpu %d\n", i, rx_irq_cpus[i]); 444 } 445 446 static void show_silos(void) 447 { 448 int i; 449 450 for (i = 0; i < cfg_num_rps_cpus; i++) 451 fprintf(stderr, "silo %d: cpu %d\n", i, rps_silo_to_cpu[i]); 452 } 453 454 static void parse_toeplitz_key(const char *str, int slen, unsigned char *key) 455 { 456 int i, ret, off; 457 458 if (slen < TOEPLITZ_STR_MIN_LEN || 459 slen > TOEPLITZ_STR_MAX_LEN + 1) 460 error(1, 0, "invalid toeplitz key"); 461 462 for (i = 0, off = 0; off < slen; i++, off += 3) { 463 ret = sscanf(str + off, "%hhx", &key[i]); 464 if (ret != 1) 465 error(1, 0, "key parse error at %d off %d len %d", 466 i, off, slen); 467 } 468 } 469 470 static void parse_rps_bitmap(const char *arg) 471 { 472 unsigned long bitmap; 473 int i; 474 475 bitmap = strtoul(arg, NULL, 0); 476 477 if (bitmap & ~(RPS_MAX_CPUS - 1)) 478 error(1, 0, "rps bitmap 0x%lx out of bounds 0..%lu", 479 bitmap, RPS_MAX_CPUS - 1); 480 481 for (i = 0; i < RPS_MAX_CPUS; i++) 482 if (bitmap & 1UL << i) 483 rps_silo_to_cpu[cfg_num_rps_cpus++] = i; 484 } 485 486 static void parse_opts(int argc, char **argv) 487 { 488 static struct option long_options[] = { 489 {"dport", required_argument, 0, 'd'}, 490 {"cpus", required_argument, 0, 'C'}, 491 {"key", required_argument, 0, 'k'}, 492 {"iface", required_argument, 0, 'i'}, 493 {"ipv4", no_argument, 0, '4'}, 494 {"ipv6", no_argument, 0, '6'}, 495 {"sink", no_argument, 0, 's'}, 496 {"tcp", no_argument, 0, 't'}, 497 {"timeout", required_argument, 0, 'T'}, 498 {"udp", no_argument, 0, 'u'}, 499 {"verbose", no_argument, 0, 'v'}, 500 {"rps", required_argument, 0, 'r'}, 501 {0, 0, 0, 0} 502 }; 503 bool have_toeplitz = false; 504 int index, c; 505 506 while ((c = getopt_long(argc, argv, "46C:d:i:k:r:stT:uv", long_options, &index)) != -1) { 507 switch (c) { 508 case '4': 509 cfg_family = AF_INET; 510 break; 511 case '6': 512 cfg_family = AF_INET6; 513 break; 514 case 'C': 515 parse_cpulist(optarg); 516 break; 517 case 'd': 518 cfg_dport = strtol(optarg, NULL, 0); 519 break; 520 case 'i': 521 cfg_ifname = optarg; 522 break; 523 case 'k': 524 parse_toeplitz_key(optarg, strlen(optarg), 525 toeplitz_key); 526 have_toeplitz = true; 527 break; 528 case 'r': 529 parse_rps_bitmap(optarg); 530 break; 531 case 's': 532 cfg_sink = true; 533 break; 534 case 't': 535 cfg_type = SOCK_STREAM; 536 break; 537 case 'T': 538 cfg_timeout_msec = strtol(optarg, NULL, 0); 539 break; 540 case 'u': 541 cfg_type = SOCK_DGRAM; 542 break; 543 case 'v': 544 cfg_verbose = true; 545 break; 546 547 default: 548 error(1, 0, "unknown option %c", optopt); 549 break; 550 } 551 } 552 553 if (!have_toeplitz) 554 error(1, 0, "Must supply rss key ('-k')"); 555 556 num_cpus = get_nprocs(); 557 if (num_cpus > RSS_MAX_CPUS) 558 error(1, 0, "increase RSS_MAX_CPUS"); 559 560 if (cfg_num_queues && cfg_num_rps_cpus) 561 error(1, 0, 562 "Can't supply both RSS cpus ('-C') and RPS map ('-r')"); 563 if (cfg_verbose) { 564 show_cpulist(); 565 show_silos(); 566 } 567 } 568 569 int main(int argc, char **argv) 570 { 571 const int min_tests = 10; 572 int fd_sink = -1; 573 574 parse_opts(argc, argv); 575 576 if (cfg_sink) 577 fd_sink = setup_sink(); 578 579 setup_rings(); 580 581 /* Signal to test framework that we're ready to receive */ 582 ksft_ready(); 583 584 process_rings(); 585 cleanup_rings(); 586 587 if (cfg_sink && close(fd_sink)) 588 error(1, errno, "close sink"); 589 590 if (frames_received - frames_nohash < min_tests) 591 error(1, 0, "too few frames for verification"); 592 593 return frames_error; 594 } 595