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 (1UL << 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 int chunk_size = 512*1024;
86
87 unsigned long htotal;
88
prefetch(const void * x)89 static inline void prefetch(const void *x)
90 {
91 #if defined(__x86_64__)
92 asm volatile("prefetcht0 %P0" : : "m" (*(const char *)x));
93 #endif
94 }
95
hash_zone(void * zone,unsigned int length)96 void hash_zone(void *zone, unsigned int length)
97 {
98 unsigned long temp = htotal;
99
100 while (length >= 8*sizeof(long)) {
101 prefetch(zone + 384);
102 temp ^= *(unsigned long *)zone;
103 temp ^= *(unsigned long *)(zone + sizeof(long));
104 temp ^= *(unsigned long *)(zone + 2*sizeof(long));
105 temp ^= *(unsigned long *)(zone + 3*sizeof(long));
106 temp ^= *(unsigned long *)(zone + 4*sizeof(long));
107 temp ^= *(unsigned long *)(zone + 5*sizeof(long));
108 temp ^= *(unsigned long *)(zone + 6*sizeof(long));
109 temp ^= *(unsigned long *)(zone + 7*sizeof(long));
110 zone += 8*sizeof(long);
111 length -= 8*sizeof(long);
112 }
113 while (length >= 1) {
114 temp ^= *(unsigned char *)zone;
115 zone += 1;
116 length--;
117 }
118 htotal = temp;
119 }
120
child_thread(void * arg)121 void *child_thread(void *arg)
122 {
123 unsigned long total_mmap = 0, total = 0;
124 struct tcp_zerocopy_receive zc;
125 unsigned long delta_usec;
126 int flags = MAP_SHARED;
127 struct timeval t0, t1;
128 char *buffer = NULL;
129 void *addr = NULL;
130 double throughput;
131 struct rusage ru;
132 int lu, fd;
133
134 fd = (int)(unsigned long)arg;
135
136 gettimeofday(&t0, NULL);
137
138 fcntl(fd, F_SETFL, O_NDELAY);
139 buffer = malloc(chunk_size);
140 if (!buffer) {
141 perror("malloc");
142 goto error;
143 }
144 if (zflg) {
145 addr = mmap(NULL, chunk_size, PROT_READ, flags, fd, 0);
146 if (addr == (void *)-1)
147 zflg = 0;
148 }
149 while (1) {
150 struct pollfd pfd = { .fd = fd, .events = POLLIN, };
151 int sub;
152
153 poll(&pfd, 1, 10000);
154 if (zflg) {
155 socklen_t zc_len = sizeof(zc);
156 int res;
157
158 zc.address = (__u64)addr;
159 zc.length = chunk_size;
160 zc.recv_skip_hint = 0;
161 res = getsockopt(fd, IPPROTO_TCP, TCP_ZEROCOPY_RECEIVE,
162 &zc, &zc_len);
163 if (res == -1)
164 break;
165
166 if (zc.length) {
167 assert(zc.length <= chunk_size);
168 total_mmap += zc.length;
169 if (xflg)
170 hash_zone(addr, zc.length);
171 total += zc.length;
172 }
173 if (zc.recv_skip_hint) {
174 assert(zc.recv_skip_hint <= chunk_size);
175 lu = read(fd, buffer, zc.recv_skip_hint);
176 if (lu > 0) {
177 if (xflg)
178 hash_zone(buffer, lu);
179 total += lu;
180 }
181 }
182 continue;
183 }
184 sub = 0;
185 while (sub < chunk_size) {
186 lu = read(fd, buffer + sub, chunk_size - sub);
187 if (lu == 0)
188 goto end;
189 if (lu < 0)
190 break;
191 if (xflg)
192 hash_zone(buffer + sub, lu);
193 total += lu;
194 sub += lu;
195 }
196 }
197 end:
198 gettimeofday(&t1, NULL);
199 delta_usec = (t1.tv_sec - t0.tv_sec) * 1000000 + t1.tv_usec - t0.tv_usec;
200
201 throughput = 0;
202 if (delta_usec)
203 throughput = total * 8.0 / (double)delta_usec / 1000.0;
204 getrusage(RUSAGE_THREAD, &ru);
205 if (total > 1024*1024) {
206 unsigned long total_usec;
207 unsigned long mb = total >> 20;
208 total_usec = 1000000*ru.ru_utime.tv_sec + ru.ru_utime.tv_usec +
209 1000000*ru.ru_stime.tv_sec + ru.ru_stime.tv_usec;
210 printf("received %lg MB (%lg %% mmap'ed) in %lg s, %lg Gbit\n"
211 " cpu usage user:%lg sys:%lg, %lg usec per MB, %lu c-switches\n",
212 total / (1024.0 * 1024.0),
213 100.0*total_mmap/total,
214 (double)delta_usec / 1000000.0,
215 throughput,
216 (double)ru.ru_utime.tv_sec + (double)ru.ru_utime.tv_usec / 1000000.0,
217 (double)ru.ru_stime.tv_sec + (double)ru.ru_stime.tv_usec / 1000000.0,
218 (double)total_usec/mb,
219 ru.ru_nvcsw);
220 }
221 error:
222 free(buffer);
223 close(fd);
224 if (zflg)
225 munmap(addr, chunk_size);
226 pthread_exit(0);
227 }
228
apply_rcvsnd_buf(int fd)229 static void apply_rcvsnd_buf(int fd)
230 {
231 if (rcvbuf && setsockopt(fd, SOL_SOCKET,
232 SO_RCVBUF, &rcvbuf, sizeof(rcvbuf)) == -1) {
233 perror("setsockopt SO_RCVBUF");
234 }
235
236 if (sndbuf && setsockopt(fd, SOL_SOCKET,
237 SO_SNDBUF, &sndbuf, sizeof(sndbuf)) == -1) {
238 perror("setsockopt SO_SNDBUF");
239 }
240 }
241
242
setup_sockaddr(int domain,const char * str_addr,struct sockaddr_storage * sockaddr)243 static void setup_sockaddr(int domain, const char *str_addr,
244 struct sockaddr_storage *sockaddr)
245 {
246 struct sockaddr_in6 *addr6 = (void *) sockaddr;
247 struct sockaddr_in *addr4 = (void *) sockaddr;
248
249 switch (domain) {
250 case PF_INET:
251 memset(addr4, 0, sizeof(*addr4));
252 addr4->sin_family = AF_INET;
253 addr4->sin_port = htons(cfg_port);
254 if (str_addr &&
255 inet_pton(AF_INET, str_addr, &(addr4->sin_addr)) != 1)
256 error(1, 0, "ipv4 parse error: %s", str_addr);
257 break;
258 case PF_INET6:
259 memset(addr6, 0, sizeof(*addr6));
260 addr6->sin6_family = AF_INET6;
261 addr6->sin6_port = htons(cfg_port);
262 if (str_addr &&
263 inet_pton(AF_INET6, str_addr, &(addr6->sin6_addr)) != 1)
264 error(1, 0, "ipv6 parse error: %s", str_addr);
265 break;
266 default:
267 error(1, 0, "illegal domain");
268 }
269 }
270
do_accept(int fdlisten)271 static void do_accept(int fdlisten)
272 {
273 if (setsockopt(fdlisten, SOL_SOCKET, SO_RCVLOWAT,
274 &chunk_size, sizeof(chunk_size)) == -1) {
275 perror("setsockopt SO_RCVLOWAT");
276 }
277
278 apply_rcvsnd_buf(fdlisten);
279
280 while (1) {
281 struct sockaddr_in addr;
282 socklen_t addrlen = sizeof(addr);
283 pthread_t th;
284 int fd, res;
285
286 fd = accept(fdlisten, (struct sockaddr *)&addr, &addrlen);
287 if (fd == -1) {
288 perror("accept");
289 continue;
290 }
291 res = pthread_create(&th, NULL, child_thread,
292 (void *)(unsigned long)fd);
293 if (res) {
294 errno = res;
295 perror("pthread_create");
296 close(fd);
297 }
298 }
299 }
300
main(int argc,char * argv[])301 int main(int argc, char *argv[])
302 {
303 struct sockaddr_storage listenaddr, addr;
304 unsigned int max_pacing_rate = 0;
305 unsigned long total = 0;
306 char *host = NULL;
307 int fd, c, on = 1;
308 char *buffer;
309 int sflg = 0;
310 int mss = 0;
311
312 while ((c = getopt(argc, argv, "46p:svr:w:H:zxkP:M:")) != -1) {
313 switch (c) {
314 case '4':
315 cfg_family = PF_INET;
316 cfg_alen = sizeof(struct sockaddr_in);
317 break;
318 case '6':
319 cfg_family = PF_INET6;
320 cfg_alen = sizeof(struct sockaddr_in6);
321 break;
322 case 'p':
323 cfg_port = atoi(optarg);
324 break;
325 case 'H':
326 host = optarg;
327 break;
328 case 's': /* server : listen for incoming connections */
329 sflg++;
330 break;
331 case 'r':
332 rcvbuf = atoi(optarg);
333 break;
334 case 'w':
335 sndbuf = atoi(optarg);
336 break;
337 case 'z':
338 zflg = 1;
339 break;
340 case 'M':
341 mss = atoi(optarg);
342 break;
343 case 'x':
344 xflg = 1;
345 break;
346 case 'k':
347 keepflag = 1;
348 break;
349 case 'P':
350 max_pacing_rate = atoi(optarg) ;
351 break;
352 default:
353 exit(1);
354 }
355 }
356 if (sflg) {
357 int fdlisten = socket(cfg_family, SOCK_STREAM, 0);
358
359 if (fdlisten == -1) {
360 perror("socket");
361 exit(1);
362 }
363 apply_rcvsnd_buf(fdlisten);
364 setsockopt(fdlisten, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on));
365
366 setup_sockaddr(cfg_family, host, &listenaddr);
367
368 if (mss &&
369 setsockopt(fdlisten, IPPROTO_TCP, TCP_MAXSEG,
370 &mss, sizeof(mss)) == -1) {
371 perror("setsockopt TCP_MAXSEG");
372 exit(1);
373 }
374 if (bind(fdlisten, (const struct sockaddr *)&listenaddr, cfg_alen) == -1) {
375 perror("bind");
376 exit(1);
377 }
378 if (listen(fdlisten, 128) == -1) {
379 perror("listen");
380 exit(1);
381 }
382 do_accept(fdlisten);
383 }
384 buffer = mmap(NULL, chunk_size, PROT_READ | PROT_WRITE,
385 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
386 if (buffer == (char *)-1) {
387 perror("mmap");
388 exit(1);
389 }
390
391 fd = socket(cfg_family, SOCK_STREAM, 0);
392 if (fd == -1) {
393 perror("socket");
394 exit(1);
395 }
396 apply_rcvsnd_buf(fd);
397
398 setup_sockaddr(cfg_family, host, &addr);
399
400 if (mss &&
401 setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, &mss, sizeof(mss)) == -1) {
402 perror("setsockopt TCP_MAXSEG");
403 exit(1);
404 }
405 if (connect(fd, (const struct sockaddr *)&addr, cfg_alen) == -1) {
406 perror("connect");
407 exit(1);
408 }
409 if (max_pacing_rate &&
410 setsockopt(fd, SOL_SOCKET, SO_MAX_PACING_RATE,
411 &max_pacing_rate, sizeof(max_pacing_rate)) == -1)
412 perror("setsockopt SO_MAX_PACING_RATE");
413
414 if (zflg && setsockopt(fd, SOL_SOCKET, SO_ZEROCOPY,
415 &on, sizeof(on)) == -1) {
416 perror("setsockopt SO_ZEROCOPY, (-z option disabled)");
417 zflg = 0;
418 }
419 while (total < FILE_SZ) {
420 long wr = FILE_SZ - total;
421
422 if (wr > chunk_size)
423 wr = chunk_size;
424 /* Note : we just want to fill the pipe with 0 bytes */
425 wr = send(fd, buffer, wr, zflg ? MSG_ZEROCOPY : 0);
426 if (wr <= 0)
427 break;
428 total += wr;
429 }
430 close(fd);
431 munmap(buffer, chunk_size);
432 return 0;
433 }
434
