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>
69 #include <openssl/pem.h>
72 #define MSG_ZEROCOPY 0x4000000
76 #define min(a, b) ((a) < (b) ? (a) : (b))
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;
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.*/
91 static size_t chunk_size
= 512*1024;
93 static size_t map_align
;
96 unsigned int digest_len
;
98 static inline void prefetch(const void *x
)
100 #if defined(__x86_64__)
101 asm volatile("prefetcht0 %P0" : : "m" (*(const char *)x
));
105 void hash_zone(void *zone
, unsigned int length
)
107 unsigned long temp
= htotal
;
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);
122 while (length
>= 1) {
123 temp
^= *(unsigned char *)zone
;
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))
134 static void *mmap_large_buffer(size_t need
, size_t *allocated
)
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);
144 if (buffer
== (void *)-1) {
146 buffer
= mmap(NULL
, sz
, PROT_READ
| PROT_WRITE
,
147 MAP_PRIVATE
| MAP_ANONYMOUS
| MAP_POPULATE
,
149 if (buffer
!= (void *)-1)
150 fprintf(stderr
, "MAP_HUGETLB attempt failed, look at /sys/kernel/mm/hugepages for optimal performance\n");
156 static uint32_t tcp_info_get_rcv_mss(int fd
)
158 socklen_t sz
= sizeof(struct tcp_info
);
159 struct tcp_info info
;
161 if (getsockopt(fd
, IPPROTO_TCP
, TCP_INFO
, &info
, &sz
)) {
162 fprintf(stderr
, "Error fetching TCP_INFO\n");
166 return info
.tcpi_rcv_mss
;
169 void *child_thread(void *arg
)
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
;
186 fd
= (int)(unsigned long)arg
;
188 gettimeofday(&t0
, NULL
);
190 fcntl(fd
, F_SETFL
, O_NDELAY
);
191 buffer
= mmap_large_buffer(chunk_size
, &buffer_sz
);
192 if (buffer
== (void *)-1) {
197 raddr
= mmap(NULL
, chunk_size
+ map_align
, PROT_READ
, flags
, fd
, 0);
198 if (raddr
== (void *)-1) {
202 addr
= ALIGN_PTR_UP(raddr
, map_align
);
206 ctx
= EVP_MD_CTX_new();
208 perror("cannot enable SHA computing");
211 EVP_DigestInit_ex(ctx
, EVP_sha256(), NULL
);
214 struct pollfd pfd
= { .fd
= fd
, .events
= POLLIN
, };
217 poll(&pfd
, 1, 10000);
219 socklen_t zc_len
= sizeof(zc
);
222 memset(&zc
, 0, sizeof(zc
));
223 zc
.address
= (__u64
)((unsigned long)addr
);
224 zc
.length
= min(chunk_size
, FILE_SZ
- total
);
226 res
= getsockopt(fd
, IPPROTO_TCP
, TCP_ZEROCOPY_RECEIVE
,
232 assert(zc
.length
<= chunk_size
);
234 EVP_DigestUpdate(ctx
, addr
, zc
.length
);
235 total_mmap
+= zc
.length
;
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.
241 madvise(addr
, zc
.length
, MADV_DONTNEED
);
244 if (zc
.recv_skip_hint
) {
245 assert(zc
.recv_skip_hint
<= chunk_size
);
246 lu
= read(fd
, buffer
, min(zc
.recv_skip_hint
,
250 EVP_DigestUpdate(ctx
, buffer
, lu
);
252 hash_zone(buffer
, lu
);
261 while (sub
< chunk_size
) {
262 lu
= read(fd
, buffer
+ sub
, min(chunk_size
- sub
,
269 EVP_DigestUpdate(ctx
, buffer
+ sub
, lu
);
271 hash_zone(buffer
+ sub
, lu
);
277 gettimeofday(&t1
, NULL
);
278 delta_usec
= (t1
.tv_sec
- t0
.tv_sec
) * 1000000 + t1
.tv_usec
- t0
.tv_usec
;
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");
287 if (memcmp(digest
, buffer
,
288 SHA256_DIGEST_LENGTH
))
289 fprintf(stderr
, "Error: SHA256 of the data is not right\n");
291 printf("\nSHA256 is correct\n");
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,
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
,
313 tcp_info_get_rcv_mss(fd
));
316 munmap(buffer
, buffer_sz
);
319 munmap(raddr
, chunk_size
+ map_align
);
323 static void apply_rcvsnd_buf(int fd
)
325 if (rcvbuf
&& setsockopt(fd
, SOL_SOCKET
,
326 SO_RCVBUF
, &rcvbuf
, sizeof(rcvbuf
)) == -1) {
327 perror("setsockopt SO_RCVBUF");
330 if (sndbuf
&& setsockopt(fd
, SOL_SOCKET
,
331 SO_SNDBUF
, &sndbuf
, sizeof(sndbuf
)) == -1) {
332 perror("setsockopt SO_SNDBUF");
337 static void setup_sockaddr(int domain
, const char *str_addr
,
338 struct sockaddr_storage
*sockaddr
)
340 struct sockaddr_in6
*addr6
= (void *) sockaddr
;
341 struct sockaddr_in
*addr4
= (void *) sockaddr
;
345 memset(addr4
, 0, sizeof(*addr4
));
346 addr4
->sin_family
= AF_INET
;
347 addr4
->sin_port
= htons(cfg_port
);
349 inet_pton(AF_INET
, str_addr
, &(addr4
->sin_addr
)) != 1)
350 error(1, 0, "ipv4 parse error: %s", str_addr
);
353 memset(addr6
, 0, sizeof(*addr6
));
354 addr6
->sin6_family
= AF_INET6
;
355 addr6
->sin6_port
= htons(cfg_port
);
357 inet_pton(AF_INET6
, str_addr
, &(addr6
->sin6_addr
)) != 1)
358 error(1, 0, "ipv6 parse error: %s", str_addr
);
361 error(1, 0, "illegal domain");
365 static void do_accept(int fdlisten
)
370 pthread_attr_init(&attr
);
371 pthread_attr_setdetachstate(&attr
, PTHREAD_CREATE_DETACHED
);
373 rcvlowat
= chunk_size
;
374 if (setsockopt(fdlisten
, SOL_SOCKET
, SO_RCVLOWAT
,
375 &rcvlowat
, sizeof(rcvlowat
)) == -1) {
376 perror("setsockopt SO_RCVLOWAT");
379 apply_rcvsnd_buf(fdlisten
);
382 struct sockaddr_in addr
;
383 socklen_t addrlen
= sizeof(addr
);
387 fd
= accept(fdlisten
, (struct sockaddr
*)&addr
, &addrlen
);
392 res
= pthread_create(&th
, &attr
, child_thread
,
393 (void *)(unsigned long)fd
);
396 perror("pthread_create");
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.
406 static unsigned long default_huge_page_size(void)
408 FILE *f
= fopen("/proc/meminfo", "r");
409 unsigned long hps
= 0;
415 while (getline(&line
, &linelen
, f
) > 0) {
416 if (sscanf(line
, "Hugepagesize: %lu kB", &hps
) == 1) {
426 static void randomize(void *target
, size_t count
)
428 static int urandom
= -1;
431 urandom
= open("/dev/urandom", O_RDONLY
);
433 perror("open /dev/urandom");
436 got
= read(urandom
, target
, count
);
438 perror("read /dev/urandom");
443 int main(int argc
, char *argv
[])
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
;
457 while ((c
= getopt(argc
, argv
, "46p:svr:w:H:zxkP:M:C:a:i")) != -1) {
460 cfg_family
= PF_INET
;
461 cfg_alen
= sizeof(struct sockaddr_in
);
464 cfg_family
= PF_INET6
;
465 cfg_alen
= sizeof(struct sockaddr_in6
);
468 cfg_port
= atoi(optarg
);
473 case 's': /* server : listen for incoming connections */
477 rcvbuf
= atoi(optarg
);
480 sndbuf
= atoi(optarg
);
495 max_pacing_rate
= atoi(optarg
) ;
498 chunk_size
= atol(optarg
);
501 map_align
= atol(optarg
);
511 map_align
= default_huge_page_size();
512 /* if really /proc/meminfo is not helping,
513 * we use the default x86_64 hugepagesize.
516 map_align
= 2*1024*1024;
519 int fdlisten
= socket(cfg_family
, SOCK_STREAM
, 0);
521 if (fdlisten
== -1) {
525 apply_rcvsnd_buf(fdlisten
);
526 setsockopt(fdlisten
, SOL_SOCKET
, SO_REUSEADDR
, &on
, sizeof(on
));
528 setup_sockaddr(cfg_family
, host
, &listenaddr
);
531 setsockopt(fdlisten
, IPPROTO_TCP
, TCP_MAXSEG
,
532 &mss
, sizeof(mss
)) == -1) {
533 perror("setsockopt TCP_MAXSEG");
536 if (bind(fdlisten
, (const struct sockaddr
*)&listenaddr
, cfg_alen
) == -1) {
540 if (listen(fdlisten
, 128) == -1) {
547 buffer
= mmap_large_buffer(chunk_size
, &buffer_sz
);
548 if (buffer
== (unsigned char *)-1) {
553 fd
= socket(cfg_family
, SOCK_STREAM
, 0);
558 apply_rcvsnd_buf(fd
);
560 setup_sockaddr(cfg_family
, host
, &addr
);
563 setsockopt(fd
, IPPROTO_TCP
, TCP_MAXSEG
, &mss
, sizeof(mss
)) == -1) {
564 perror("setsockopt TCP_MAXSEG");
567 if (connect(fd
, (const struct sockaddr
*)&addr
, cfg_alen
) == -1) {
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");
576 if (zflg
&& setsockopt(fd
, SOL_SOCKET
, SO_ZEROCOPY
,
577 &on
, sizeof(on
)) == -1) {
578 perror("setsockopt SO_ZEROCOPY, (-z option disabled)");
582 randomize(buffer
, buffer_sz
);
583 ctx
= EVP_MD_CTX_new();
585 perror("cannot enable SHA computing");
588 EVP_DigestInit_ex(ctx
, EVP_sha256(), NULL
);
590 while (total
< FILE_SZ
) {
591 size_t offset
= total
% chunk_size
;
592 int64_t wr
= FILE_SZ
- total
;
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);
602 EVP_DigestUpdate(ctx
, buffer
+ offset
, wr
);
605 if (integrity
&& total
== FILE_SZ
) {
606 EVP_DigestFinal_ex(ctx
, digest
, &digest_len
);
607 send(fd
, digest
, (size_t)SHA256_DIGEST_LENGTH
, 0);
610 munmap(buffer
, buffer_sz
);