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)
80static int cfg_family = AF_INET6;
81static socklen_t cfg_alen = sizeof(struct sockaddr_in6);
82static int cfg_port = 8787;
83
84static int rcvbuf; /* Default: autotuning.  Can be set with -r <integer> option */
85static int sndbuf; /* Default: autotuning.  Can be set with -w <integer> option */
86static int zflg; /* zero copy option. (MSG_ZEROCOPY for sender, mmap() for receiver */
87static int xflg; /* hash received data (simple xor) (-h option) */
88static int keepflag; /* -k option: receiver shall keep all received file in memory (no munmap() calls) */
89static int integrity; /* -i option: sender and receiver compute sha256 over the data.*/
90
91static size_t chunk_size  = 512*1024;
92
93static size_t map_align;
94
95unsigned long htotal;
96unsigned int digest_len;
97
98static 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
105void 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
134static 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
156static uint32_t tcp_info_get_rcv_mss(int fd)
157{
158	socklen_t sz = sizeof(struct tcp_info);
159	struct tcp_info info;
160
161	if (getsockopt(fd, IPPROTO_TCP, TCP_INFO, &info, &sz)) {
162		fprintf(stderr, "Error fetching TCP_INFO\n");
163		return 0;
164	}
165
166	return info.tcpi_rcv_mss;
167}
168
169void *child_thread(void *arg)
170{
171	unsigned char digest[SHA256_DIGEST_LENGTH];
172	unsigned long total_mmap = 0, total = 0;
173	struct tcp_zerocopy_receive zc;
174	unsigned char *buffer = NULL;
175	unsigned long delta_usec;
176	EVP_MD_CTX *ctx = NULL;
177	int flags = MAP_SHARED;
178	struct timeval t0, t1;
179	void *raddr = NULL;
180	void *addr = NULL;
181	double throughput;
182	struct rusage ru;
183	size_t buffer_sz;
184	int lu, fd;
185
186	fd = (int)(unsigned long)arg;
187
188	gettimeofday(&t0, NULL);
189
190	fcntl(fd, F_SETFL, O_NDELAY);
191	buffer = mmap_large_buffer(chunk_size, &buffer_sz);
192	if (buffer == (void *)-1) {
193		perror("mmap");
194		goto error;
195	}
196	if (zflg) {
197		raddr = mmap(NULL, chunk_size + map_align, PROT_READ, flags, fd, 0);
198		if (raddr == (void *)-1) {
199			perror("mmap");
200			zflg = 0;
201		} else {
202			addr = ALIGN_PTR_UP(raddr, map_align);
203		}
204	}
205	if (integrity) {
206		ctx = EVP_MD_CTX_new();
207		if (!ctx) {
208			perror("cannot enable SHA computing");
209			goto error;
210		}
211		EVP_DigestInit_ex(ctx, EVP_sha256(), NULL);
212	}
213	while (1) {
214		struct pollfd pfd = { .fd = fd, .events = POLLIN, };
215		int sub;
216
217		poll(&pfd, 1, 10000);
218		if (zflg) {
219			socklen_t zc_len = sizeof(zc);
220			int res;
221
222			memset(&zc, 0, sizeof(zc));
223			zc.address = (__u64)((unsigned long)addr);
224			zc.length = min(chunk_size, FILE_SZ - total);
225
226			res = getsockopt(fd, IPPROTO_TCP, TCP_ZEROCOPY_RECEIVE,
227					 &zc, &zc_len);
228			if (res == -1)
229				break;
230
231			if (zc.length) {
232				assert(zc.length <= chunk_size);
233				if (integrity)
234					EVP_DigestUpdate(ctx, addr, zc.length);
235				total_mmap += zc.length;
236				if (xflg)
237					hash_zone(addr, zc.length);
238				/* It is more efficient to unmap the pages right now,
239				 * instead of doing this in next TCP_ZEROCOPY_RECEIVE.
240				 */
241				madvise(addr, zc.length, MADV_DONTNEED);
242				total += zc.length;
243			}
244			if (zc.recv_skip_hint) {
245				assert(zc.recv_skip_hint <= chunk_size);
246				lu = read(fd, buffer, min(zc.recv_skip_hint,
247							  FILE_SZ - total));
248				if (lu > 0) {
249					if (integrity)
250						EVP_DigestUpdate(ctx, buffer, lu);
251					if (xflg)
252						hash_zone(buffer, lu);
253					total += lu;
254				}
255				if (lu == 0)
256					goto end;
257			}
258			continue;
259		}
260		sub = 0;
261		while (sub < chunk_size) {
262			lu = read(fd, buffer + sub, min(chunk_size - sub,
263							FILE_SZ - total));
264			if (lu == 0)
265				goto end;
266			if (lu < 0)
267				break;
268			if (integrity)
269				EVP_DigestUpdate(ctx, buffer + sub, lu);
270			if (xflg)
271				hash_zone(buffer + sub, lu);
272			total += lu;
273			sub += lu;
274		}
275	}
276end:
277	gettimeofday(&t1, NULL);
278	delta_usec = (t1.tv_sec - t0.tv_sec) * 1000000 + t1.tv_usec - t0.tv_usec;
279
280	if (integrity) {
281		fcntl(fd, F_SETFL, 0);
282		EVP_DigestFinal_ex(ctx, digest, &digest_len);
283		lu = read(fd, buffer, SHA256_DIGEST_LENGTH);
284		if (lu != SHA256_DIGEST_LENGTH)
285			perror("Error: Cannot read SHA256\n");
286
287		if (memcmp(digest, buffer,
288			   SHA256_DIGEST_LENGTH))
289			fprintf(stderr, "Error: SHA256 of the data is not right\n");
290		else
291			printf("\nSHA256 is correct\n");
292	}
293
294	throughput = 0;
295	if (delta_usec)
296		throughput = total * 8.0 / (double)delta_usec / 1000.0;
297	getrusage(RUSAGE_THREAD, &ru);
298	if (total > 1024*1024) {
299		unsigned long total_usec;
300		unsigned long mb = total >> 20;
301		total_usec = 1000000*ru.ru_utime.tv_sec + ru.ru_utime.tv_usec +
302			     1000000*ru.ru_stime.tv_sec + ru.ru_stime.tv_usec;
303		printf("received %lg MB (%lg %% mmap'ed) in %lg s, %lg Gbit\n"
304		       "  cpu usage user:%lg sys:%lg, %lg usec per MB, %lu c-switches, rcv_mss %u\n",
305				total / (1024.0 * 1024.0),
306				100.0*total_mmap/total,
307				(double)delta_usec / 1000000.0,
308				throughput,
309				(double)ru.ru_utime.tv_sec + (double)ru.ru_utime.tv_usec / 1000000.0,
310				(double)ru.ru_stime.tv_sec + (double)ru.ru_stime.tv_usec / 1000000.0,
311				(double)total_usec/mb,
312				ru.ru_nvcsw,
313				tcp_info_get_rcv_mss(fd));
314	}
315error:
316	munmap(buffer, buffer_sz);
317	close(fd);
318	if (zflg)
319		munmap(raddr, chunk_size + map_align);
320	pthread_exit(0);
321}
322
323static void apply_rcvsnd_buf(int fd)
324{
325	if (rcvbuf && setsockopt(fd, SOL_SOCKET,
326				 SO_RCVBUF, &rcvbuf, sizeof(rcvbuf)) == -1) {
327		perror("setsockopt SO_RCVBUF");
328	}
329
330	if (sndbuf && setsockopt(fd, SOL_SOCKET,
331				 SO_SNDBUF, &sndbuf, sizeof(sndbuf)) == -1) {
332		perror("setsockopt SO_SNDBUF");
333	}
334}
335
336
337static void setup_sockaddr(int domain, const char *str_addr,
338			   struct sockaddr_storage *sockaddr)
339{
340	struct sockaddr_in6 *addr6 = (void *) sockaddr;
341	struct sockaddr_in *addr4 = (void *) sockaddr;
342
343	switch (domain) {
344	case PF_INET:
345		memset(addr4, 0, sizeof(*addr4));
346		addr4->sin_family = AF_INET;
347		addr4->sin_port = htons(cfg_port);
348		if (str_addr &&
349		    inet_pton(AF_INET, str_addr, &(addr4->sin_addr)) != 1)
350			error(1, 0, "ipv4 parse error: %s", str_addr);
351		break;
352	case PF_INET6:
353		memset(addr6, 0, sizeof(*addr6));
354		addr6->sin6_family = AF_INET6;
355		addr6->sin6_port = htons(cfg_port);
356		if (str_addr &&
357		    inet_pton(AF_INET6, str_addr, &(addr6->sin6_addr)) != 1)
358			error(1, 0, "ipv6 parse error: %s", str_addr);
359		break;
360	default:
361		error(1, 0, "illegal domain");
362	}
363}
364
365static void do_accept(int fdlisten)
366{
367	pthread_attr_t attr;
368	int rcvlowat;
369
370	pthread_attr_init(&attr);
371	pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
372
373	rcvlowat = chunk_size;
374	if (setsockopt(fdlisten, SOL_SOCKET, SO_RCVLOWAT,
375		       &rcvlowat, sizeof(rcvlowat)) == -1) {
376		perror("setsockopt SO_RCVLOWAT");
377	}
378
379	apply_rcvsnd_buf(fdlisten);
380
381	while (1) {
382		struct sockaddr_in addr;
383		socklen_t addrlen = sizeof(addr);
384		pthread_t th;
385		int fd, res;
386
387		fd = accept(fdlisten, (struct sockaddr *)&addr, &addrlen);
388		if (fd == -1) {
389			perror("accept");
390			continue;
391		}
392		res = pthread_create(&th, &attr, child_thread,
393				     (void *)(unsigned long)fd);
394		if (res) {
395			errno = res;
396			perror("pthread_create");
397			close(fd);
398		}
399	}
400}
401
402/* Each thread should reserve a big enough vma to avoid
403 * spinlock collisions in ptl locks.
404 * This size is 2MB on x86_64, and is exported in /proc/meminfo.
405 */
406static unsigned long default_huge_page_size(void)
407{
408	FILE *f = fopen("/proc/meminfo", "r");
409	unsigned long hps = 0;
410	size_t linelen = 0;
411	char *line = NULL;
412
413	if (!f)
414		return 0;
415	while (getline(&line, &linelen, f) > 0) {
416		if (sscanf(line, "Hugepagesize:       %lu kB", &hps) == 1) {
417			hps <<= 10;
418			break;
419		}
420	}
421	free(line);
422	fclose(f);
423	return hps;
424}
425
426static void randomize(void *target, size_t count)
427{
428	static int urandom = -1;
429	ssize_t got;
430
431	urandom = open("/dev/urandom", O_RDONLY);
432	if (urandom < 0) {
433		perror("open /dev/urandom");
434		exit(1);
435	}
436	got = read(urandom, target, count);
437	if (got != count) {
438		perror("read /dev/urandom");
439		exit(1);
440	}
441}
442
443int main(int argc, char *argv[])
444{
445	unsigned char digest[SHA256_DIGEST_LENGTH];
446	struct sockaddr_storage listenaddr, addr;
447	unsigned int max_pacing_rate = 0;
448	EVP_MD_CTX *ctx = NULL;
449	unsigned char *buffer;
450	uint64_t total = 0;
451	char *host = NULL;
452	int fd, c, on = 1;
453	size_t buffer_sz;
454	int sflg = 0;
455	int mss = 0;
456
457	while ((c = getopt(argc, argv, "46p:svr:w:H:zxkP:M:C:a:i")) != -1) {
458		switch (c) {
459		case '4':
460			cfg_family = PF_INET;
461			cfg_alen = sizeof(struct sockaddr_in);
462			break;
463		case '6':
464			cfg_family = PF_INET6;
465			cfg_alen = sizeof(struct sockaddr_in6);
466			break;
467		case 'p':
468			cfg_port = atoi(optarg);
469			break;
470		case 'H':
471			host = optarg;
472			break;
473		case 's': /* server : listen for incoming connections */
474			sflg++;
475			break;
476		case 'r':
477			rcvbuf = atoi(optarg);
478			break;
479		case 'w':
480			sndbuf = atoi(optarg);
481			break;
482		case 'z':
483			zflg = 1;
484			break;
485		case 'M':
486			mss = atoi(optarg);
487			break;
488		case 'x':
489			xflg = 1;
490			break;
491		case 'k':
492			keepflag = 1;
493			break;
494		case 'P':
495			max_pacing_rate = atoi(optarg) ;
496			break;
497		case 'C':
498			chunk_size = atol(optarg);
499			break;
500		case 'a':
501			map_align = atol(optarg);
502			break;
503		case 'i':
504			integrity = 1;
505			break;
506		default:
507			exit(1);
508		}
509	}
510	if (!map_align) {
511		map_align = default_huge_page_size();
512		/* if really /proc/meminfo is not helping,
513		 * we use the default x86_64 hugepagesize.
514		 */
515		if (!map_align)
516			map_align = 2*1024*1024;
517	}
518	if (sflg) {
519		int fdlisten = socket(cfg_family, SOCK_STREAM, 0);
520
521		if (fdlisten == -1) {
522			perror("socket");
523			exit(1);
524		}
525		apply_rcvsnd_buf(fdlisten);
526		setsockopt(fdlisten, SOL_SOCKET, SO_REUSEADDR, &on, sizeof(on));
527
528		setup_sockaddr(cfg_family, host, &listenaddr);
529
530		if (mss &&
531		    setsockopt(fdlisten, IPPROTO_TCP, TCP_MAXSEG,
532			       &mss, sizeof(mss)) == -1) {
533			perror("setsockopt TCP_MAXSEG");
534			exit(1);
535		}
536		if (bind(fdlisten, (const struct sockaddr *)&listenaddr, cfg_alen) == -1) {
537			perror("bind");
538			exit(1);
539		}
540		if (listen(fdlisten, 128) == -1) {
541			perror("listen");
542			exit(1);
543		}
544		do_accept(fdlisten);
545	}
546
547	buffer = mmap_large_buffer(chunk_size, &buffer_sz);
548	if (buffer == (unsigned char *)-1) {
549		perror("mmap");
550		exit(1);
551	}
552
553	fd = socket(cfg_family, SOCK_STREAM, 0);
554	if (fd == -1) {
555		perror("socket");
556		exit(1);
557	}
558	apply_rcvsnd_buf(fd);
559
560	setup_sockaddr(cfg_family, host, &addr);
561
562	if (mss &&
563	    setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, &mss, sizeof(mss)) == -1) {
564		perror("setsockopt TCP_MAXSEG");
565		exit(1);
566	}
567	if (connect(fd, (const struct sockaddr *)&addr, cfg_alen) == -1) {
568		perror("connect");
569		exit(1);
570	}
571	if (max_pacing_rate &&
572	    setsockopt(fd, SOL_SOCKET, SO_MAX_PACING_RATE,
573		       &max_pacing_rate, sizeof(max_pacing_rate)) == -1)
574		perror("setsockopt SO_MAX_PACING_RATE");
575
576	if (zflg && setsockopt(fd, SOL_SOCKET, SO_ZEROCOPY,
577			       &on, sizeof(on)) == -1) {
578		perror("setsockopt SO_ZEROCOPY, (-z option disabled)");
579		zflg = 0;
580	}
581	if (integrity) {
582		randomize(buffer, buffer_sz);
583		ctx = EVP_MD_CTX_new();
584		if (!ctx) {
585			perror("cannot enable SHA computing");
586			exit(1);
587		}
588		EVP_DigestInit_ex(ctx, EVP_sha256(), NULL);
589	}
590	while (total < FILE_SZ) {
591		size_t offset = total % chunk_size;
592		int64_t wr = FILE_SZ - total;
593
594		if (wr > chunk_size - offset)
595			wr = chunk_size - offset;
596		/* Note : we just want to fill the pipe with random bytes */
597		wr = send(fd, buffer + offset,
598			  (size_t)wr, zflg ? MSG_ZEROCOPY : 0);
599		if (wr <= 0)
600			break;
601		if (integrity)
602			EVP_DigestUpdate(ctx, buffer + offset, wr);
603		total += wr;
604	}
605	if (integrity && total == FILE_SZ) {
606		EVP_DigestFinal_ex(ctx, digest, &digest_len);
607		send(fd, digest, (size_t)SHA256_DIGEST_LENGTH, 0);
608	}
609	close(fd);
610	munmap(buffer, buffer_sz);
611	return 0;
612}
613