1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright 2018 Google Inc.
4 * Author: Eric Dumazet (edumazet@google.com)
6 * Reference program demonstrating tcp mmap() usage,
7 * and SO_RCVLOWAT hints for receiver.
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.
12 * How to use on loopback interface :
14 * ifconfig lo mtu 61512 # 15*4096 + 40 (ipv6 header) + 32 (TCP with TS option header)
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))
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 ;)
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
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
51 #include <sys/types.h>
54 #include <sys/socket.h>
56 #include <sys/resource.h>
64 #include <netinet/in.h>
65 #include <arpa/inet.h>
67 #include <linux/tcp.h>
71 #define MSG_ZEROCOPY 0x4000000
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;
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) */
85 static size_t chunk_size = 512*1024;
87 static size_t map_align;
91 static inline void prefetch(const void *x)
93 #if defined(__x86_64__)
94 asm volatile("prefetcht0 %P0" : : "m" (*(const char *)x));
98 void hash_zone(void *zone, unsigned int length)
100 unsigned long temp = htotal;
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);
115 while (length >= 1) {
116 temp ^= *(unsigned char *)zone;
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))
126 void *child_thread(void *arg)
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;
140 fd = (int)(unsigned long)arg;
142 gettimeofday(&t0, NULL);
144 fcntl(fd, F_SETFL, O_NDELAY);
145 buffer = malloc(chunk_size);
151 raddr = mmap(NULL, chunk_size + map_align, PROT_READ, flags, fd, 0);
152 if (raddr == (void *)-1) {
156 addr = ALIGN_PTR_UP(raddr, map_align);
160 struct pollfd pfd = { .fd = fd, .events = POLLIN, };
163 poll(&pfd, 1, 10000);
165 socklen_t zc_len = sizeof(zc);
168 memset(&zc, 0, sizeof(zc));
169 zc.address = (__u64)((unsigned long)addr);
170 zc.length = chunk_size;
172 res = getsockopt(fd, IPPROTO_TCP, TCP_ZEROCOPY_RECEIVE,
178 assert(zc.length <= chunk_size);
179 total_mmap += zc.length;
181 hash_zone(addr, zc.length);
184 if (zc.recv_skip_hint) {
185 assert(zc.recv_skip_hint <= chunk_size);
186 lu = read(fd, buffer, zc.recv_skip_hint);
189 hash_zone(buffer, lu);
196 while (sub < chunk_size) {
197 lu = read(fd, buffer + sub, chunk_size - sub);
203 hash_zone(buffer + sub, lu);
209 gettimeofday(&t1, NULL);
210 delta_usec = (t1.tv_sec - t0.tv_sec) * 1000000 + t1.tv_usec - t0.tv_usec;
214 throughput = total * 8.0 / (double)delta_usec / 1000.0;
215 getrusage(RUSAGE_THREAD, &ru);
216 if (total > 1024*1024) {
217 unsigned long total_usec;
218 unsigned long mb = total >> 20;
219 total_usec = 1000000*ru.ru_utime.tv_sec + ru.ru_utime.tv_usec +
220 1000000*ru.ru_stime.tv_sec + ru.ru_stime.tv_usec;
221 printf("received %lg MB (%lg %% mmap'ed) in %lg s, %lg Gbit\n"
222 " cpu usage user:%lg sys:%lg, %lg usec per MB, %lu c-switches\n",
223 total / (1024.0 * 1024.0),
224 100.0*total_mmap/total,
225 (double)delta_usec / 1000000.0,
227 (double)ru.ru_utime.tv_sec + (double)ru.ru_utime.tv_usec / 1000000.0,
228 (double)ru.ru_stime.tv_sec + (double)ru.ru_stime.tv_usec / 1000000.0,
229 (double)total_usec/mb,
236 munmap(raddr, chunk_size + map_align);
240 static void apply_rcvsnd_buf(int fd)
242 if (rcvbuf && setsockopt(fd, SOL_SOCKET,
243 SO_RCVBUF, &rcvbuf, sizeof(rcvbuf)) == -1) {
244 perror("setsockopt SO_RCVBUF");
247 if (sndbuf && setsockopt(fd, SOL_SOCKET,
248 SO_SNDBUF, &sndbuf, sizeof(sndbuf)) == -1) {
249 perror("setsockopt SO_SNDBUF");
254 static void setup_sockaddr(int domain, const char *str_addr,
255 struct sockaddr_storage *sockaddr)
257 struct sockaddr_in6 *addr6 = (void *) sockaddr;
258 struct sockaddr_in *addr4 = (void *) sockaddr;
262 memset(addr4, 0, sizeof(*addr4));
263 addr4->sin_family = AF_INET;
264 addr4->sin_port = htons(cfg_port);
266 inet_pton(AF_INET, str_addr, &(addr4->sin_addr)) != 1)
267 error(1, 0, "ipv4 parse error: %s", str_addr);
270 memset(addr6, 0, sizeof(*addr6));
271 addr6->sin6_family = AF_INET6;
272 addr6->sin6_port = htons(cfg_port);
274 inet_pton(AF_INET6, str_addr, &(addr6->sin6_addr)) != 1)
275 error(1, 0, "ipv6 parse error: %s", str_addr);
278 error(1, 0, "illegal domain");
282 static void do_accept(int fdlisten)
286 pthread_attr_init(&attr);
287 pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
289 if (setsockopt(fdlisten, SOL_SOCKET, SO_RCVLOWAT,
290 &chunk_size, sizeof(chunk_size)) == -1) {
291 perror("setsockopt SO_RCVLOWAT");
294 apply_rcvsnd_buf(fdlisten);
297 struct sockaddr_in addr;
298 socklen_t addrlen = sizeof(addr);
302 fd = accept(fdlisten, (struct sockaddr *)&addr, &addrlen);
307 res = pthread_create(&th, &attr, child_thread,
308 (void *)(unsigned long)fd);
311 perror("pthread_create");
317 /* Each thread should reserve a big enough vma to avoid
318 * spinlock collisions in ptl locks.
319 * This size is 2MB on x86_64, and is exported in /proc/meminfo.
321 static unsigned long default_huge_page_size(void)
323 FILE *f = fopen("/proc/meminfo", "r");
324 unsigned long hps = 0;
330 while (getline(&line, &linelen, f) > 0) {
331 if (sscanf(line, "Hugepagesize: %lu kB", &hps) == 1) {
341 int main(int argc, char *argv[])
343 struct sockaddr_storage listenaddr, addr;
344 unsigned int max_pacing_rate = 0;
352 while ((c = getopt(argc, argv, "46p:svr:w:H:zxkP:M:C:a:")) != -1) {
355 cfg_family = PF_INET;
356 cfg_alen = sizeof(struct sockaddr_in);
359 cfg_family = PF_INET6;
360 cfg_alen = sizeof(struct sockaddr_in6);
363 cfg_port = atoi(optarg);
368 case 's': /* server : listen for incoming connections */
372 rcvbuf = atoi(optarg);
375 sndbuf = atoi(optarg);
390 max_pacing_rate = atoi(optarg) ;
393 chunk_size = atol(optarg);
396 map_align = atol(optarg);
403 map_align = default_huge_page_size();
404 /* if really /proc/meminfo is not helping,
405 * we use the default x86_64 hugepagesize.
408 map_align = 2*1024*1024;
411 int fdlisten = socket(cfg_family, SOCK_STREAM, 0);
413 if (fdlisten == -1) {
417 apply_rcvsnd_buf(fdlisten);
418 setsockopt(fdlisten, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on));
420 setup_sockaddr(cfg_family, host, &listenaddr);
423 setsockopt(fdlisten, IPPROTO_TCP, TCP_MAXSEG,
424 &mss, sizeof(mss)) == -1) {
425 perror("setsockopt TCP_MAXSEG");
428 if (bind(fdlisten, (const struct sockaddr *)&listenaddr, cfg_alen) == -1) {
432 if (listen(fdlisten, 128) == -1) {
438 buffer = mmap(NULL, chunk_size, PROT_READ | PROT_WRITE,
439 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
440 if (buffer == (char *)-1) {
445 fd = socket(cfg_family, SOCK_STREAM, 0);
450 apply_rcvsnd_buf(fd);
452 setup_sockaddr(cfg_family, host, &addr);
455 setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, &mss, sizeof(mss)) == -1) {
456 perror("setsockopt TCP_MAXSEG");
459 if (connect(fd, (const struct sockaddr *)&addr, cfg_alen) == -1) {
463 if (max_pacing_rate &&
464 setsockopt(fd, SOL_SOCKET, SO_MAX_PACING_RATE,
465 &max_pacing_rate, sizeof(max_pacing_rate)) == -1)
466 perror("setsockopt SO_MAX_PACING_RATE");
468 if (zflg && setsockopt(fd, SOL_SOCKET, SO_ZEROCOPY,
469 &on, sizeof(on)) == -1) {
470 perror("setsockopt SO_ZEROCOPY, (-z option disabled)");
473 while (total < FILE_SZ) {
474 ssize_t wr = FILE_SZ - total;
478 /* Note : we just want to fill the pipe with 0 bytes */
479 wr = send(fd, buffer, wr, zflg ? MSG_ZEROCOPY : 0);
485 munmap(buffer, chunk_size);