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 70 #ifndef MSG_ZEROCOPY 71 #define MSG_ZEROCOPY 0x4000000 72 #endif 73 74 #define FILE_SZ (1ULL << 35) 75 static int cfg_family = AF_INET6; 76 static socklen_t cfg_alen = sizeof(struct sockaddr_in6); 77 static int cfg_port = 8787; 78 79 static int rcvbuf; /* Default: autotuning. Can be set with -r <integer> option */ 80 static int sndbuf; /* Default: autotuning. Can be set with -w <integer> option */ 81 static int zflg; /* zero copy option. (MSG_ZEROCOPY for sender, mmap() for receiver */ 82 static int xflg; /* hash received data (simple xor) (-h option) */ 83 static int keepflag; /* -k option: receiver shall keep all received file in memory (no munmap() calls) */ 84 85 static size_t chunk_size = 512*1024; 86 87 static size_t map_align; 88 89 unsigned long htotal; 90 91 static inline void prefetch(const void *x) 92 { 93 #if defined(__x86_64__) 94 asm volatile("prefetcht0 %P0" : : "m" (*(const char *)x)); 95 #endif 96 } 97 98 void hash_zone(void *zone, unsigned int length) 99 { 100 unsigned long temp = htotal; 101 102 while (length >= 8*sizeof(long)) { 103 prefetch(zone + 384); 104 temp ^= *(unsigned long *)zone; 105 temp ^= *(unsigned long *)(zone + sizeof(long)); 106 temp ^= *(unsigned long *)(zone + 2*sizeof(long)); 107 temp ^= *(unsigned long *)(zone + 3*sizeof(long)); 108 temp ^= *(unsigned long *)(zone + 4*sizeof(long)); 109 temp ^= *(unsigned long *)(zone + 5*sizeof(long)); 110 temp ^= *(unsigned long *)(zone + 6*sizeof(long)); 111 temp ^= *(unsigned long *)(zone + 7*sizeof(long)); 112 zone += 8*sizeof(long); 113 length -= 8*sizeof(long); 114 } 115 while (length >= 1) { 116 temp ^= *(unsigned char *)zone; 117 zone += 1; 118 length--; 119 } 120 htotal = temp; 121 } 122 123 #define ALIGN_UP(x, align_to) (((x) + ((align_to)-1)) & ~((align_to)-1)) 124 #define ALIGN_PTR_UP(p, ptr_align_to) ((typeof(p))ALIGN_UP((unsigned long)(p), ptr_align_to)) 125 126 void *child_thread(void *arg) 127 { 128 unsigned long total_mmap = 0, total = 0; 129 struct tcp_zerocopy_receive zc; 130 unsigned long delta_usec; 131 int flags = MAP_SHARED; 132 struct timeval t0, t1; 133 char *buffer = NULL; 134 void *raddr = NULL; 135 void *addr = NULL; 136 double throughput; 137 struct rusage ru; 138 int lu, fd; 139 140 fd = (int)(unsigned long)arg; 141 142 gettimeofday(&t0, NULL); 143 144 fcntl(fd, F_SETFL, O_NDELAY); 145 buffer = malloc(chunk_size); 146 if (!buffer) { 147 perror("malloc"); 148 goto error; 149 } 150 if (zflg) { 151 raddr = mmap(NULL, chunk_size + map_align, PROT_READ, flags, fd, 0); 152 if (raddr == (void *)-1) { 153 perror("mmap"); 154 zflg = 0; 155 } else { 156 addr = ALIGN_PTR_UP(raddr, map_align); 157 } 158 } 159 while (1) { 160 struct pollfd pfd = { .fd = fd, .events = POLLIN, }; 161 int sub; 162 163 poll(&pfd, 1, 10000); 164 if (zflg) { 165 socklen_t zc_len = sizeof(zc); 166 int res; 167 168 zc.address = (__u64)((unsigned long)addr); 169 zc.length = chunk_size; 170 zc.recv_skip_hint = 0; 171 res = getsockopt(fd, IPPROTO_TCP, TCP_ZEROCOPY_RECEIVE, 172 &zc, &zc_len); 173 if (res == -1) 174 break; 175 176 if (zc.length) { 177 assert(zc.length <= chunk_size); 178 total_mmap += zc.length; 179 if (xflg) 180 hash_zone(addr, zc.length); 181 total += zc.length; 182 } 183 if (zc.recv_skip_hint) { 184 assert(zc.recv_skip_hint <= chunk_size); 185 lu = read(fd, buffer, zc.recv_skip_hint); 186 if (lu > 0) { 187 if (xflg) 188 hash_zone(buffer, lu); 189 total += lu; 190 } 191 } 192 continue; 193 } 194 sub = 0; 195 while (sub < chunk_size) { 196 lu = read(fd, buffer + sub, chunk_size - sub); 197 if (lu == 0) 198 goto end; 199 if (lu < 0) 200 break; 201 if (xflg) 202 hash_zone(buffer + sub, lu); 203 total += lu; 204 sub += lu; 205 } 206 } 207 end: 208 gettimeofday(&t1, NULL); 209 delta_usec = (t1.tv_sec - t0.tv_sec) * 1000000 + t1.tv_usec - t0.tv_usec; 210 211 throughput = 0; 212 if (delta_usec) 213 throughput = total * 8.0 / (double)delta_usec / 1000.0; 214 getrusage(RUSAGE_THREAD, &ru); 215 if (total > 1024*1024) { 216 unsigned long total_usec; 217 unsigned long mb = total >> 20; 218 total_usec = 1000000*ru.ru_utime.tv_sec + ru.ru_utime.tv_usec + 219 1000000*ru.ru_stime.tv_sec + ru.ru_stime.tv_usec; 220 printf("received %lg MB (%lg %% mmap'ed) in %lg s, %lg Gbit\n" 221 " cpu usage user:%lg sys:%lg, %lg usec per MB, %lu c-switches\n", 222 total / (1024.0 * 1024.0), 223 100.0*total_mmap/total, 224 (double)delta_usec / 1000000.0, 225 throughput, 226 (double)ru.ru_utime.tv_sec + (double)ru.ru_utime.tv_usec / 1000000.0, 227 (double)ru.ru_stime.tv_sec + (double)ru.ru_stime.tv_usec / 1000000.0, 228 (double)total_usec/mb, 229 ru.ru_nvcsw); 230 } 231 error: 232 free(buffer); 233 close(fd); 234 if (zflg) 235 munmap(raddr, chunk_size + map_align); 236 pthread_exit(0); 237 } 238 239 static void apply_rcvsnd_buf(int fd) 240 { 241 if (rcvbuf && setsockopt(fd, SOL_SOCKET, 242 SO_RCVBUF, &rcvbuf, sizeof(rcvbuf)) == -1) { 243 perror("setsockopt SO_RCVBUF"); 244 } 245 246 if (sndbuf && setsockopt(fd, SOL_SOCKET, 247 SO_SNDBUF, &sndbuf, sizeof(sndbuf)) == -1) { 248 perror("setsockopt SO_SNDBUF"); 249 } 250 } 251 252 253 static void setup_sockaddr(int domain, const char *str_addr, 254 struct sockaddr_storage *sockaddr) 255 { 256 struct sockaddr_in6 *addr6 = (void *) sockaddr; 257 struct sockaddr_in *addr4 = (void *) sockaddr; 258 259 switch (domain) { 260 case PF_INET: 261 memset(addr4, 0, sizeof(*addr4)); 262 addr4->sin_family = AF_INET; 263 addr4->sin_port = htons(cfg_port); 264 if (str_addr && 265 inet_pton(AF_INET, str_addr, &(addr4->sin_addr)) != 1) 266 error(1, 0, "ipv4 parse error: %s", str_addr); 267 break; 268 case PF_INET6: 269 memset(addr6, 0, sizeof(*addr6)); 270 addr6->sin6_family = AF_INET6; 271 addr6->sin6_port = htons(cfg_port); 272 if (str_addr && 273 inet_pton(AF_INET6, str_addr, &(addr6->sin6_addr)) != 1) 274 error(1, 0, "ipv6 parse error: %s", str_addr); 275 break; 276 default: 277 error(1, 0, "illegal domain"); 278 } 279 } 280 281 static void do_accept(int fdlisten) 282 { 283 pthread_attr_t attr; 284 285 pthread_attr_init(&attr); 286 pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED); 287 288 if (setsockopt(fdlisten, SOL_SOCKET, SO_RCVLOWAT, 289 &chunk_size, sizeof(chunk_size)) == -1) { 290 perror("setsockopt SO_RCVLOWAT"); 291 } 292 293 apply_rcvsnd_buf(fdlisten); 294 295 while (1) { 296 struct sockaddr_in addr; 297 socklen_t addrlen = sizeof(addr); 298 pthread_t th; 299 int fd, res; 300 301 fd = accept(fdlisten, (struct sockaddr *)&addr, &addrlen); 302 if (fd == -1) { 303 perror("accept"); 304 continue; 305 } 306 res = pthread_create(&th, &attr, child_thread, 307 (void *)(unsigned long)fd); 308 if (res) { 309 errno = res; 310 perror("pthread_create"); 311 close(fd); 312 } 313 } 314 } 315 316 /* Each thread should reserve a big enough vma to avoid 317 * spinlock collisions in ptl locks. 318 * This size is 2MB on x86_64, and is exported in /proc/meminfo. 319 */ 320 static unsigned long default_huge_page_size(void) 321 { 322 FILE *f = fopen("/proc/meminfo", "r"); 323 unsigned long hps = 0; 324 size_t linelen = 0; 325 char *line = NULL; 326 327 if (!f) 328 return 0; 329 while (getline(&line, &linelen, f) > 0) { 330 if (sscanf(line, "Hugepagesize: %lu kB", &hps) == 1) { 331 hps <<= 10; 332 break; 333 } 334 } 335 free(line); 336 fclose(f); 337 return hps; 338 } 339 340 int main(int argc, char *argv[]) 341 { 342 struct sockaddr_storage listenaddr, addr; 343 unsigned int max_pacing_rate = 0; 344 size_t total = 0; 345 char *host = NULL; 346 int fd, c, on = 1; 347 char *buffer; 348 int sflg = 0; 349 int mss = 0; 350 351 while ((c = getopt(argc, argv, "46p:svr:w:H:zxkP:M:C:a:")) != -1) { 352 switch (c) { 353 case '4': 354 cfg_family = PF_INET; 355 cfg_alen = sizeof(struct sockaddr_in); 356 break; 357 case '6': 358 cfg_family = PF_INET6; 359 cfg_alen = sizeof(struct sockaddr_in6); 360 break; 361 case 'p': 362 cfg_port = atoi(optarg); 363 break; 364 case 'H': 365 host = optarg; 366 break; 367 case 's': /* server : listen for incoming connections */ 368 sflg++; 369 break; 370 case 'r': 371 rcvbuf = atoi(optarg); 372 break; 373 case 'w': 374 sndbuf = atoi(optarg); 375 break; 376 case 'z': 377 zflg = 1; 378 break; 379 case 'M': 380 mss = atoi(optarg); 381 break; 382 case 'x': 383 xflg = 1; 384 break; 385 case 'k': 386 keepflag = 1; 387 break; 388 case 'P': 389 max_pacing_rate = atoi(optarg) ; 390 break; 391 case 'C': 392 chunk_size = atol(optarg); 393 break; 394 case 'a': 395 map_align = atol(optarg); 396 break; 397 default: 398 exit(1); 399 } 400 } 401 if (!map_align) { 402 map_align = default_huge_page_size(); 403 /* if really /proc/meminfo is not helping, 404 * we use the default x86_64 hugepagesize. 405 */ 406 if (!map_align) 407 map_align = 2*1024*1024; 408 } 409 if (sflg) { 410 int fdlisten = socket(cfg_family, SOCK_STREAM, 0); 411 412 if (fdlisten == -1) { 413 perror("socket"); 414 exit(1); 415 } 416 apply_rcvsnd_buf(fdlisten); 417 setsockopt(fdlisten, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on)); 418 419 setup_sockaddr(cfg_family, host, &listenaddr); 420 421 if (mss && 422 setsockopt(fdlisten, IPPROTO_TCP, TCP_MAXSEG, 423 &mss, sizeof(mss)) == -1) { 424 perror("setsockopt TCP_MAXSEG"); 425 exit(1); 426 } 427 if (bind(fdlisten, (const struct sockaddr *)&listenaddr, cfg_alen) == -1) { 428 perror("bind"); 429 exit(1); 430 } 431 if (listen(fdlisten, 128) == -1) { 432 perror("listen"); 433 exit(1); 434 } 435 do_accept(fdlisten); 436 } 437 buffer = mmap(NULL, chunk_size, PROT_READ | PROT_WRITE, 438 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); 439 if (buffer == (char *)-1) { 440 perror("mmap"); 441 exit(1); 442 } 443 444 fd = socket(cfg_family, SOCK_STREAM, 0); 445 if (fd == -1) { 446 perror("socket"); 447 exit(1); 448 } 449 apply_rcvsnd_buf(fd); 450 451 setup_sockaddr(cfg_family, host, &addr); 452 453 if (mss && 454 setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, &mss, sizeof(mss)) == -1) { 455 perror("setsockopt TCP_MAXSEG"); 456 exit(1); 457 } 458 if (connect(fd, (const struct sockaddr *)&addr, cfg_alen) == -1) { 459 perror("connect"); 460 exit(1); 461 } 462 if (max_pacing_rate && 463 setsockopt(fd, SOL_SOCKET, SO_MAX_PACING_RATE, 464 &max_pacing_rate, sizeof(max_pacing_rate)) == -1) 465 perror("setsockopt SO_MAX_PACING_RATE"); 466 467 if (zflg && setsockopt(fd, SOL_SOCKET, SO_ZEROCOPY, 468 &on, sizeof(on)) == -1) { 469 perror("setsockopt SO_ZEROCOPY, (-z option disabled)"); 470 zflg = 0; 471 } 472 while (total < FILE_SZ) { 473 ssize_t wr = FILE_SZ - total; 474 475 if (wr > chunk_size) 476 wr = chunk_size; 477 /* Note : we just want to fill the pipe with 0 bytes */ 478 wr = send(fd, buffer, wr, zflg ? MSG_ZEROCOPY : 0); 479 if (wr <= 0) 480 break; 481 total += wr; 482 } 483 close(fd); 484 munmap(buffer, chunk_size); 485 return 0; 486 } 487