1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright 2018 Google Inc. 4 * Author: Eric Dumazet (edumazet@google.com) 5 * 6 * Reference program demonstrating tcp mmap() usage, 7 * and SO_RCVLOWAT hints for receiver. 8 * 9 * Note : NIC with header split is needed to use mmap() on TCP : 10 * Each incoming frame must be a multiple of PAGE_SIZE bytes of TCP payload. 11 * 12 * How to use on loopback interface : 13 * 14 * ifconfig lo mtu 61512 # 15*4096 + 40 (ipv6 header) + 32 (TCP with TS option header) 15 * tcp_mmap -s -z & 16 * tcp_mmap -H ::1 -z 17 * 18 * Or leave default lo mtu, but use -M option to set TCP_MAXSEG option to (4096 + 12) 19 * (4096 : page size on x86, 12: TCP TS option length) 20 * tcp_mmap -s -z -M $((4096+12)) & 21 * tcp_mmap -H ::1 -z -M $((4096+12)) 22 * 23 * Note: -z option on sender uses MSG_ZEROCOPY, which forces a copy when packets go through loopback interface. 24 * We might use sendfile() instead, but really this test program is about mmap(), for receivers ;) 25 * 26 * $ ./tcp_mmap -s & # Without mmap() 27 * $ for i in {1..4}; do ./tcp_mmap -H ::1 -z ; done 28 * received 32768 MB (0 % mmap'ed) in 14.1157 s, 19.4732 Gbit 29 * cpu usage user:0.057 sys:7.815, 240.234 usec per MB, 65531 c-switches 30 * received 32768 MB (0 % mmap'ed) in 14.6833 s, 18.7204 Gbit 31 * cpu usage user:0.043 sys:8.103, 248.596 usec per MB, 65524 c-switches 32 * received 32768 MB (0 % mmap'ed) in 11.143 s, 24.6682 Gbit 33 * cpu usage user:0.044 sys:6.576, 202.026 usec per MB, 65519 c-switches 34 * received 32768 MB (0 % mmap'ed) in 14.9056 s, 18.4413 Gbit 35 * cpu usage user:0.036 sys:8.193, 251.129 usec per MB, 65530 c-switches 36 * $ kill %1 # kill tcp_mmap server 37 * 38 * $ ./tcp_mmap -s -z & # With mmap() 39 * $ for i in {1..4}; do ./tcp_mmap -H ::1 -z ; done 40 * received 32768 MB (99.9939 % mmap'ed) in 6.73792 s, 40.7956 Gbit 41 * cpu usage user:0.045 sys:2.827, 87.6465 usec per MB, 65532 c-switches 42 * received 32768 MB (99.9939 % mmap'ed) in 7.26732 s, 37.8238 Gbit 43 * cpu usage user:0.037 sys:3.087, 95.3369 usec per MB, 65532 c-switches 44 * received 32768 MB (99.9939 % mmap'ed) in 7.61661 s, 36.0893 Gbit 45 * cpu usage user:0.046 sys:3.559, 110.016 usec per MB, 65529 c-switches 46 * received 32768 MB (99.9939 % mmap'ed) in 7.43764 s, 36.9577 Gbit 47 * cpu usage user:0.035 sys:3.467, 106.873 usec per MB, 65530 c-switches 48 */ 49 #define _GNU_SOURCE 50 #include <pthread.h> 51 #include <sys/types.h> 52 #include <fcntl.h> 53 #include <error.h> 54 #include <sys/socket.h> 55 #include <sys/mman.h> 56 #include <sys/resource.h> 57 #include <unistd.h> 58 #include <string.h> 59 #include <stdlib.h> 60 #include <stdio.h> 61 #include <errno.h> 62 #include <time.h> 63 #include <sys/time.h> 64 #include <netinet/in.h> 65 #include <arpa/inet.h> 66 #include <poll.h> 67 #include <linux/tcp.h> 68 #include <assert.h> 69 #include <openssl/pem.h> 70 71 #ifndef MSG_ZEROCOPY 72 #define MSG_ZEROCOPY 0x4000000 73 #endif 74 75 #ifndef min 76 #define min(a, b) ((a) < (b) ? (a) : (b)) 77 #endif 78 79 #define FILE_SZ (1ULL << 35) 80 static int cfg_family = AF_INET6; 81 static socklen_t cfg_alen = sizeof(struct sockaddr_in6); 82 static int cfg_port = 8787; 83 84 static int rcvbuf; /* Default: autotuning. Can be set with -r <integer> option */ 85 static int sndbuf; /* Default: autotuning. Can be set with -w <integer> option */ 86 static int zflg; /* zero copy option. (MSG_ZEROCOPY for sender, mmap() for receiver */ 87 static int xflg; /* hash received data (simple xor) (-h option) */ 88 static int keepflag; /* -k option: receiver shall keep all received file in memory (no munmap() calls) */ 89 static int integrity; /* -i option: sender and receiver compute sha256 over the data.*/ 90 91 static size_t chunk_size = 512*1024; 92 93 static size_t map_align; 94 95 unsigned long htotal; 96 unsigned int digest_len; 97 98 static inline void prefetch(const void *x) 99 { 100 #if defined(__x86_64__) 101 asm volatile("prefetcht0 %P0" : : "m" (*(const char *)x)); 102 #endif 103 } 104 105 void hash_zone(void *zone, unsigned int length) 106 { 107 unsigned long temp = htotal; 108 109 while (length >= 8*sizeof(long)) { 110 prefetch(zone + 384); 111 temp ^= *(unsigned long *)zone; 112 temp ^= *(unsigned long *)(zone + sizeof(long)); 113 temp ^= *(unsigned long *)(zone + 2*sizeof(long)); 114 temp ^= *(unsigned long *)(zone + 3*sizeof(long)); 115 temp ^= *(unsigned long *)(zone + 4*sizeof(long)); 116 temp ^= *(unsigned long *)(zone + 5*sizeof(long)); 117 temp ^= *(unsigned long *)(zone + 6*sizeof(long)); 118 temp ^= *(unsigned long *)(zone + 7*sizeof(long)); 119 zone += 8*sizeof(long); 120 length -= 8*sizeof(long); 121 } 122 while (length >= 1) { 123 temp ^= *(unsigned char *)zone; 124 zone += 1; 125 length--; 126 } 127 htotal = temp; 128 } 129 130 #define ALIGN_UP(x, align_to) (((x) + ((align_to)-1)) & ~((align_to)-1)) 131 #define ALIGN_PTR_UP(p, ptr_align_to) ((typeof(p))ALIGN_UP((unsigned long)(p), ptr_align_to)) 132 133 134 static void *mmap_large_buffer(size_t need, size_t *allocated) 135 { 136 void *buffer; 137 size_t sz; 138 139 /* Attempt to use huge pages if possible. */ 140 sz = ALIGN_UP(need, map_align); 141 buffer = mmap(NULL, sz, PROT_READ | PROT_WRITE, 142 MAP_PRIVATE | MAP_ANONYMOUS | MAP_HUGETLB, -1, 0); 143 144 if (buffer == (void *)-1) { 145 sz = need; 146 buffer = mmap(NULL, sz, PROT_READ | PROT_WRITE, 147 MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE, 148 -1, 0); 149 if (buffer != (void *)-1) 150 fprintf(stderr, "MAP_HUGETLB attempt failed, look at /sys/kernel/mm/hugepages for optimal performance\n"); 151 } 152 *allocated = sz; 153 return buffer; 154 } 155 156 void *child_thread(void *arg) 157 { 158 unsigned char digest[SHA256_DIGEST_LENGTH]; 159 unsigned long total_mmap = 0, total = 0; 160 struct tcp_zerocopy_receive zc; 161 unsigned char *buffer = NULL; 162 unsigned long delta_usec; 163 EVP_MD_CTX *ctx = NULL; 164 int flags = MAP_SHARED; 165 struct timeval t0, t1; 166 void *raddr = NULL; 167 void *addr = NULL; 168 double throughput; 169 struct rusage ru; 170 size_t buffer_sz; 171 int lu, fd; 172 173 fd = (int)(unsigned long)arg; 174 175 gettimeofday(&t0, NULL); 176 177 fcntl(fd, F_SETFL, O_NDELAY); 178 buffer = mmap_large_buffer(chunk_size, &buffer_sz); 179 if (buffer == (void *)-1) { 180 perror("mmap"); 181 goto error; 182 } 183 if (zflg) { 184 raddr = mmap(NULL, chunk_size + map_align, PROT_READ, flags, fd, 0); 185 if (raddr == (void *)-1) { 186 perror("mmap"); 187 zflg = 0; 188 } else { 189 addr = ALIGN_PTR_UP(raddr, map_align); 190 } 191 } 192 if (integrity) { 193 ctx = EVP_MD_CTX_new(); 194 if (!ctx) { 195 perror("cannot enable SHA computing"); 196 goto error; 197 } 198 EVP_DigestInit_ex(ctx, EVP_sha256(), NULL); 199 } 200 while (1) { 201 struct pollfd pfd = { .fd = fd, .events = POLLIN, }; 202 int sub; 203 204 poll(&pfd, 1, 10000); 205 if (zflg) { 206 socklen_t zc_len = sizeof(zc); 207 int res; 208 209 memset(&zc, 0, sizeof(zc)); 210 zc.address = (__u64)((unsigned long)addr); 211 zc.length = min(chunk_size, FILE_SZ - total); 212 213 res = getsockopt(fd, IPPROTO_TCP, TCP_ZEROCOPY_RECEIVE, 214 &zc, &zc_len); 215 if (res == -1) 216 break; 217 218 if (zc.length) { 219 assert(zc.length <= chunk_size); 220 if (integrity) 221 EVP_DigestUpdate(ctx, addr, zc.length); 222 total_mmap += zc.length; 223 if (xflg) 224 hash_zone(addr, zc.length); 225 /* It is more efficient to unmap the pages right now, 226 * instead of doing this in next TCP_ZEROCOPY_RECEIVE. 227 */ 228 madvise(addr, zc.length, MADV_DONTNEED); 229 total += zc.length; 230 } 231 if (zc.recv_skip_hint) { 232 assert(zc.recv_skip_hint <= chunk_size); 233 lu = read(fd, buffer, min(zc.recv_skip_hint, 234 FILE_SZ - total)); 235 if (lu > 0) { 236 if (integrity) 237 EVP_DigestUpdate(ctx, buffer, lu); 238 if (xflg) 239 hash_zone(buffer, lu); 240 total += lu; 241 } 242 if (lu == 0) 243 goto end; 244 } 245 continue; 246 } 247 sub = 0; 248 while (sub < chunk_size) { 249 lu = read(fd, buffer + sub, min(chunk_size - sub, 250 FILE_SZ - total)); 251 if (lu == 0) 252 goto end; 253 if (lu < 0) 254 break; 255 if (integrity) 256 EVP_DigestUpdate(ctx, buffer + sub, lu); 257 if (xflg) 258 hash_zone(buffer + sub, lu); 259 total += lu; 260 sub += lu; 261 } 262 } 263 end: 264 gettimeofday(&t1, NULL); 265 delta_usec = (t1.tv_sec - t0.tv_sec) * 1000000 + t1.tv_usec - t0.tv_usec; 266 267 if (integrity) { 268 fcntl(fd, F_SETFL, 0); 269 EVP_DigestFinal_ex(ctx, digest, &digest_len); 270 lu = read(fd, buffer, SHA256_DIGEST_LENGTH); 271 if (lu != SHA256_DIGEST_LENGTH) 272 perror("Error: Cannot read SHA256\n"); 273 274 if (memcmp(digest, buffer, 275 SHA256_DIGEST_LENGTH)) 276 fprintf(stderr, "Error: SHA256 of the data is not right\n"); 277 else 278 printf("\nSHA256 is correct\n"); 279 } 280 281 throughput = 0; 282 if (delta_usec) 283 throughput = total * 8.0 / (double)delta_usec / 1000.0; 284 getrusage(RUSAGE_THREAD, &ru); 285 if (total > 1024*1024) { 286 unsigned long total_usec; 287 unsigned long mb = total >> 20; 288 total_usec = 1000000*ru.ru_utime.tv_sec + ru.ru_utime.tv_usec + 289 1000000*ru.ru_stime.tv_sec + ru.ru_stime.tv_usec; 290 printf("received %lg MB (%lg %% mmap'ed) in %lg s, %lg Gbit\n" 291 " cpu usage user:%lg sys:%lg, %lg usec per MB, %lu c-switches\n", 292 total / (1024.0 * 1024.0), 293 100.0*total_mmap/total, 294 (double)delta_usec / 1000000.0, 295 throughput, 296 (double)ru.ru_utime.tv_sec + (double)ru.ru_utime.tv_usec / 1000000.0, 297 (double)ru.ru_stime.tv_sec + (double)ru.ru_stime.tv_usec / 1000000.0, 298 (double)total_usec/mb, 299 ru.ru_nvcsw); 300 } 301 error: 302 munmap(buffer, buffer_sz); 303 close(fd); 304 if (zflg) 305 munmap(raddr, chunk_size + map_align); 306 pthread_exit(0); 307 } 308 309 static void apply_rcvsnd_buf(int fd) 310 { 311 if (rcvbuf && setsockopt(fd, SOL_SOCKET, 312 SO_RCVBUF, &rcvbuf, sizeof(rcvbuf)) == -1) { 313 perror("setsockopt SO_RCVBUF"); 314 } 315 316 if (sndbuf && setsockopt(fd, SOL_SOCKET, 317 SO_SNDBUF, &sndbuf, sizeof(sndbuf)) == -1) { 318 perror("setsockopt SO_SNDBUF"); 319 } 320 } 321 322 323 static void setup_sockaddr(int domain, const char *str_addr, 324 struct sockaddr_storage *sockaddr) 325 { 326 struct sockaddr_in6 *addr6 = (void *) sockaddr; 327 struct sockaddr_in *addr4 = (void *) sockaddr; 328 329 switch (domain) { 330 case PF_INET: 331 memset(addr4, 0, sizeof(*addr4)); 332 addr4->sin_family = AF_INET; 333 addr4->sin_port = htons(cfg_port); 334 if (str_addr && 335 inet_pton(AF_INET, str_addr, &(addr4->sin_addr)) != 1) 336 error(1, 0, "ipv4 parse error: %s", str_addr); 337 break; 338 case PF_INET6: 339 memset(addr6, 0, sizeof(*addr6)); 340 addr6->sin6_family = AF_INET6; 341 addr6->sin6_port = htons(cfg_port); 342 if (str_addr && 343 inet_pton(AF_INET6, str_addr, &(addr6->sin6_addr)) != 1) 344 error(1, 0, "ipv6 parse error: %s", str_addr); 345 break; 346 default: 347 error(1, 0, "illegal domain"); 348 } 349 } 350 351 static void do_accept(int fdlisten) 352 { 353 pthread_attr_t attr; 354 int rcvlowat; 355 356 pthread_attr_init(&attr); 357 pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED); 358 359 rcvlowat = chunk_size; 360 if (setsockopt(fdlisten, SOL_SOCKET, SO_RCVLOWAT, 361 &rcvlowat, sizeof(rcvlowat)) == -1) { 362 perror("setsockopt SO_RCVLOWAT"); 363 } 364 365 apply_rcvsnd_buf(fdlisten); 366 367 while (1) { 368 struct sockaddr_in addr; 369 socklen_t addrlen = sizeof(addr); 370 pthread_t th; 371 int fd, res; 372 373 fd = accept(fdlisten, (struct sockaddr *)&addr, &addrlen); 374 if (fd == -1) { 375 perror("accept"); 376 continue; 377 } 378 res = pthread_create(&th, &attr, child_thread, 379 (void *)(unsigned long)fd); 380 if (res) { 381 errno = res; 382 perror("pthread_create"); 383 close(fd); 384 } 385 } 386 } 387 388 /* Each thread should reserve a big enough vma to avoid 389 * spinlock collisions in ptl locks. 390 * This size is 2MB on x86_64, and is exported in /proc/meminfo. 391 */ 392 static unsigned long default_huge_page_size(void) 393 { 394 FILE *f = fopen("/proc/meminfo", "r"); 395 unsigned long hps = 0; 396 size_t linelen = 0; 397 char *line = NULL; 398 399 if (!f) 400 return 0; 401 while (getline(&line, &linelen, f) > 0) { 402 if (sscanf(line, "Hugepagesize: %lu kB", &hps) == 1) { 403 hps <<= 10; 404 break; 405 } 406 } 407 free(line); 408 fclose(f); 409 return hps; 410 } 411 412 static void randomize(void *target, size_t count) 413 { 414 static int urandom = -1; 415 ssize_t got; 416 417 urandom = open("/dev/urandom", O_RDONLY); 418 if (urandom < 0) { 419 perror("open /dev/urandom"); 420 exit(1); 421 } 422 got = read(urandom, target, count); 423 if (got != count) { 424 perror("read /dev/urandom"); 425 exit(1); 426 } 427 } 428 429 int main(int argc, char *argv[]) 430 { 431 unsigned char digest[SHA256_DIGEST_LENGTH]; 432 struct sockaddr_storage listenaddr, addr; 433 unsigned int max_pacing_rate = 0; 434 EVP_MD_CTX *ctx = NULL; 435 unsigned char *buffer; 436 uint64_t total = 0; 437 char *host = NULL; 438 int fd, c, on = 1; 439 size_t buffer_sz; 440 int sflg = 0; 441 int mss = 0; 442 443 while ((c = getopt(argc, argv, "46p:svr:w:H:zxkP:M:C:a:i")) != -1) { 444 switch (c) { 445 case '4': 446 cfg_family = PF_INET; 447 cfg_alen = sizeof(struct sockaddr_in); 448 break; 449 case '6': 450 cfg_family = PF_INET6; 451 cfg_alen = sizeof(struct sockaddr_in6); 452 break; 453 case 'p': 454 cfg_port = atoi(optarg); 455 break; 456 case 'H': 457 host = optarg; 458 break; 459 case 's': /* server : listen for incoming connections */ 460 sflg++; 461 break; 462 case 'r': 463 rcvbuf = atoi(optarg); 464 break; 465 case 'w': 466 sndbuf = atoi(optarg); 467 break; 468 case 'z': 469 zflg = 1; 470 break; 471 case 'M': 472 mss = atoi(optarg); 473 break; 474 case 'x': 475 xflg = 1; 476 break; 477 case 'k': 478 keepflag = 1; 479 break; 480 case 'P': 481 max_pacing_rate = atoi(optarg) ; 482 break; 483 case 'C': 484 chunk_size = atol(optarg); 485 break; 486 case 'a': 487 map_align = atol(optarg); 488 break; 489 case 'i': 490 integrity = 1; 491 break; 492 default: 493 exit(1); 494 } 495 } 496 if (!map_align) { 497 map_align = default_huge_page_size(); 498 /* if really /proc/meminfo is not helping, 499 * we use the default x86_64 hugepagesize. 500 */ 501 if (!map_align) 502 map_align = 2*1024*1024; 503 } 504 if (sflg) { 505 int fdlisten = socket(cfg_family, SOCK_STREAM, 0); 506 507 if (fdlisten == -1) { 508 perror("socket"); 509 exit(1); 510 } 511 apply_rcvsnd_buf(fdlisten); 512 setsockopt(fdlisten, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on)); 513 514 setup_sockaddr(cfg_family, host, &listenaddr); 515 516 if (mss && 517 setsockopt(fdlisten, IPPROTO_TCP, TCP_MAXSEG, 518 &mss, sizeof(mss)) == -1) { 519 perror("setsockopt TCP_MAXSEG"); 520 exit(1); 521 } 522 if (bind(fdlisten, (const struct sockaddr *)&listenaddr, cfg_alen) == -1) { 523 perror("bind"); 524 exit(1); 525 } 526 if (listen(fdlisten, 128) == -1) { 527 perror("listen"); 528 exit(1); 529 } 530 do_accept(fdlisten); 531 } 532 533 buffer = mmap_large_buffer(chunk_size, &buffer_sz); 534 if (buffer == (unsigned char *)-1) { 535 perror("mmap"); 536 exit(1); 537 } 538 539 fd = socket(cfg_family, SOCK_STREAM, 0); 540 if (fd == -1) { 541 perror("socket"); 542 exit(1); 543 } 544 apply_rcvsnd_buf(fd); 545 546 setup_sockaddr(cfg_family, host, &addr); 547 548 if (mss && 549 setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, &mss, sizeof(mss)) == -1) { 550 perror("setsockopt TCP_MAXSEG"); 551 exit(1); 552 } 553 if (connect(fd, (const struct sockaddr *)&addr, cfg_alen) == -1) { 554 perror("connect"); 555 exit(1); 556 } 557 if (max_pacing_rate && 558 setsockopt(fd, SOL_SOCKET, SO_MAX_PACING_RATE, 559 &max_pacing_rate, sizeof(max_pacing_rate)) == -1) 560 perror("setsockopt SO_MAX_PACING_RATE"); 561 562 if (zflg && setsockopt(fd, SOL_SOCKET, SO_ZEROCOPY, 563 &on, sizeof(on)) == -1) { 564 perror("setsockopt SO_ZEROCOPY, (-z option disabled)"); 565 zflg = 0; 566 } 567 if (integrity) { 568 randomize(buffer, buffer_sz); 569 ctx = EVP_MD_CTX_new(); 570 if (!ctx) { 571 perror("cannot enable SHA computing"); 572 exit(1); 573 } 574 EVP_DigestInit_ex(ctx, EVP_sha256(), NULL); 575 } 576 while (total < FILE_SZ) { 577 size_t offset = total % chunk_size; 578 int64_t wr = FILE_SZ - total; 579 580 if (wr > chunk_size - offset) 581 wr = chunk_size - offset; 582 /* Note : we just want to fill the pipe with random bytes */ 583 wr = send(fd, buffer + offset, 584 (size_t)wr, zflg ? MSG_ZEROCOPY : 0); 585 if (wr <= 0) 586 break; 587 if (integrity) 588 EVP_DigestUpdate(ctx, buffer + offset, wr); 589 total += wr; 590 } 591 if (integrity && total == FILE_SZ) { 592 EVP_DigestFinal_ex(ctx, digest, &digest_len); 593 send(fd, digest, (size_t)SHA256_DIGEST_LENGTH, 0); 594 } 595 close(fd); 596 munmap(buffer, buffer_sz); 597 return 0; 598 } 599