projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Get rid if --header-file argument for lex
[fio.git] / engines / net.c
... / ...
CommitLineData
1/*
2 * net engine
3 *
4 * IO engine that reads/writes to/from sockets.
5 *
6 */
7#include <stdio.h>
8#include <stdlib.h>
9#include <unistd.h>
10#include <signal.h>
11#include <errno.h>
12#include <assert.h>
13#include <netinet/in.h>
14#include <netinet/tcp.h>
15#include <arpa/inet.h>
16#include <netdb.h>
17#include <sys/poll.h>
18#include <sys/types.h>
19#include <sys/stat.h>
20#include <sys/socket.h>
21#include <sys/un.h>
22
23#include "../fio.h"
24#include "../verify.h"
25
26struct netio_data {
27 int listenfd;
28 int use_splice;
29 int seq_off;
30 int pipes[2];
31 struct sockaddr_in addr;
32 struct sockaddr_in6 addr6;
33 struct sockaddr_un addr_un;
34 uint64_t udp_send_seq;
35 uint64_t udp_recv_seq;
36};
37
38struct netio_options {
39 struct thread_data *td;
40 unsigned int port;
41 unsigned int proto;
42 unsigned int listen;
43 unsigned int pingpong;
44 unsigned int nodelay;
45 unsigned int ttl;
46 unsigned int window_size;
47 unsigned int mss;
48 char *intfc;
49};
50
51struct udp_close_msg {
52 uint32_t magic;
53 uint32_t cmd;
54};
55
56struct udp_seq {
57 uint64_t magic;
58 uint64_t seq;
59 uint64_t bs;
60};
61
62enum {
63 FIO_LINK_CLOSE = 0x89,
64 FIO_LINK_OPEN_CLOSE_MAGIC = 0x6c696e6b,
65 FIO_LINK_OPEN = 0x98,
66 FIO_UDP_SEQ_MAGIC = 0x657375716e556563ULL,
67
68 FIO_TYPE_TCP = 1,
69 FIO_TYPE_UDP = 2,
70 FIO_TYPE_UNIX = 3,
71 FIO_TYPE_TCP_V6 = 4,
72 FIO_TYPE_UDP_V6 = 5,
73};
74
75static int str_hostname_cb(void *data, const char *input);
76static struct fio_option options[] = {
77 {
78 .name = "hostname",
79 .lname = "net engine hostname",
80 .type = FIO_OPT_STR_STORE,
81 .cb = str_hostname_cb,
82 .help = "Hostname for net IO engine",
83 .category = FIO_OPT_C_ENGINE,
84 .group = FIO_OPT_G_NETIO,
85 },
86 {
87 .name = "port",
88 .lname = "net engine port",
89 .type = FIO_OPT_INT,
90 .off1 = offsetof(struct netio_options, port),
91 .minval = 1,
92 .maxval = 65535,
93 .help = "Port to use for TCP or UDP net connections",
94 .category = FIO_OPT_C_ENGINE,
95 .group = FIO_OPT_G_NETIO,
96 },
97 {
98 .name = "protocol",
99 .lname = "net engine protocol",
100 .alias = "proto",
101 .type = FIO_OPT_STR,
102 .off1 = offsetof(struct netio_options, proto),
103 .help = "Network protocol to use",
104 .def = "tcp",
105 .posval = {
106 { .ival = "tcp",
107 .oval = FIO_TYPE_TCP,
108 .help = "Transmission Control Protocol",
109 },
110#ifdef CONFIG_IPV6
111 { .ival = "tcpv6",
112 .oval = FIO_TYPE_TCP_V6,
113 .help = "Transmission Control Protocol V6",
114 },
115#endif
116 { .ival = "udp",
117 .oval = FIO_TYPE_UDP,
118 .help = "User Datagram Protocol",
119 },
120#ifdef CONFIG_IPV6
121 { .ival = "udpv6",
122 .oval = FIO_TYPE_UDP_V6,
123 .help = "User Datagram Protocol V6",
124 },
125#endif
126 { .ival = "unix",
127 .oval = FIO_TYPE_UNIX,
128 .help = "UNIX domain socket",
129 },
130 },
131 .category = FIO_OPT_C_ENGINE,
132 .group = FIO_OPT_G_NETIO,
133 },
134#ifdef CONFIG_TCP_NODELAY
135 {
136 .name = "nodelay",
137 .type = FIO_OPT_BOOL,
138 .off1 = offsetof(struct netio_options, nodelay),
139 .help = "Use TCP_NODELAY on TCP connections",
140 .category = FIO_OPT_C_ENGINE,
141 .group = FIO_OPT_G_NETIO,
142 },
143#endif
144 {
145 .name = "listen",
146 .lname = "net engine listen",
147 .type = FIO_OPT_STR_SET,
148 .off1 = offsetof(struct netio_options, listen),
149 .help = "Listen for incoming TCP connections",
150 .category = FIO_OPT_C_ENGINE,
151 .group = FIO_OPT_G_NETIO,
152 },
153 {
154 .name = "pingpong",
155 .type = FIO_OPT_STR_SET,
156 .off1 = offsetof(struct netio_options, pingpong),
157 .help = "Ping-pong IO requests",
158 .category = FIO_OPT_C_ENGINE,
159 .group = FIO_OPT_G_NETIO,
160 },
161 {
162 .name = "interface",
163 .lname = "net engine interface",
164 .type = FIO_OPT_STR_STORE,
165 .off1 = offsetof(struct netio_options, intfc),
166 .help = "Network interface to use",
167 .category = FIO_OPT_C_ENGINE,
168 .group = FIO_OPT_G_NETIO,
169 },
170 {
171 .name = "ttl",
172 .lname = "net engine multicast ttl",
173 .type = FIO_OPT_INT,
174 .off1 = offsetof(struct netio_options, ttl),
175 .def = "1",
176 .minval = 0,
177 .help = "Time-to-live value for outgoing UDP multicast packets",
178 .category = FIO_OPT_C_ENGINE,
179 .group = FIO_OPT_G_NETIO,
180 },
181#ifdef CONFIG_NET_WINDOWSIZE
182 {
183 .name = "window_size",
184 .lname = "Window Size",
185 .type = FIO_OPT_INT,
186 .off1 = offsetof(struct netio_options, window_size),
187 .minval = 0,
188 .help = "Set socket buffer window size",
189 .category = FIO_OPT_C_ENGINE,
190 .group = FIO_OPT_G_NETIO,
191 },
192#endif
193#ifdef CONFIG_NET_MSS
194 {
195 .name = "mss",
196 .lname = "Maximum segment size",
197 .type = FIO_OPT_INT,
198 .off1 = offsetof(struct netio_options, mss),
199 .minval = 0,
200 .help = "Set TCP maximum segment size",
201 .category = FIO_OPT_C_ENGINE,
202 .group = FIO_OPT_G_NETIO,
203 },
204#endif
205 {
206 .name = NULL,
207 },
208};
209
210static inline int is_udp(struct netio_options *o)
211{
212 return o->proto == FIO_TYPE_UDP || o->proto == FIO_TYPE_UDP_V6;
213}
214
215static inline int is_tcp(struct netio_options *o)
216{
217 return o->proto == FIO_TYPE_TCP || o->proto == FIO_TYPE_TCP_V6;
218}
219
220static inline int is_ipv6(struct netio_options *o)
221{
222 return o->proto == FIO_TYPE_UDP_V6 || o->proto == FIO_TYPE_TCP_V6;
223}
224
225static int set_window_size(struct thread_data *td, int fd)
226{
227#ifdef CONFIG_NET_WINDOWSIZE
228 struct netio_options *o = td->eo;
229 unsigned int wss;
230 int snd, rcv, ret;
231
232 if (!o->window_size)
233 return 0;
234
235 rcv = o->listen || o->pingpong;
236 snd = !o->listen || o->pingpong;
237 wss = o->window_size;
238 ret = 0;
239
240 if (rcv) {
241 ret = setsockopt(fd, SOL_SOCKET, SO_RCVBUF, (void *) &wss,
242 sizeof(wss));
243 if (ret < 0)
244 td_verror(td, errno, "rcvbuf window size");
245 }
246 if (snd && !ret) {
247 ret = setsockopt(fd, SOL_SOCKET, SO_SNDBUF, (void *) &wss,
248 sizeof(wss));
249 if (ret < 0)
250 td_verror(td, errno, "sndbuf window size");
251 }
252
253 return ret;
254#else
255 td_verror(td, -EINVAL, "setsockopt window size");
256 return -1;
257#endif
258}
259
260static int set_mss(struct thread_data *td, int fd)
261{
262#ifdef CONFIG_NET_MSS
263 struct netio_options *o = td->eo;
264 unsigned int mss;
265 int ret;
266
267 if (!o->mss || !is_tcp(o))
268 return 0;
269
270 mss = o->mss;
271 ret = setsockopt(fd, IPPROTO_TCP, TCP_MAXSEG, (void *) &mss,
272 sizeof(mss));
273 if (ret < 0)
274 td_verror(td, errno, "setsockopt TCP_MAXSEG");
275
276 return ret;
277#else
278 td_verror(td, -EINVAL, "setsockopt TCP_MAXSEG");
279 return -1;
280#endif
281}
282
283
284/*
285 * Return -1 for error and 'nr events' for a positive number
286 * of events
287 */
288static int poll_wait(struct thread_data *td, int fd, short events)
289{
290 struct pollfd pfd;
291 int ret;
292
293 while (!td->terminate) {
294 pfd.fd = fd;
295 pfd.events = events;
296 ret = poll(&pfd, 1, -1);
297 if (ret < 0) {
298 if (errno == EINTR)
299 break;
300
301 td_verror(td, errno, "poll");
302 return -1;
303 } else if (!ret)
304 continue;
305
306 break;
307 }
308
309 if (pfd.revents & events)
310 return 1;
311
312 return -1;
313}
314
315static int fio_netio_is_multicast(const char *mcaddr)
316{
317 in_addr_t addr = inet_network(mcaddr);
318 if (addr == -1)
319 return 0;
320
321 if (inet_network("224.0.0.0") <= addr &&
322 inet_network("239.255.255.255") >= addr)
323 return 1;
324
325 return 0;
326}
327
328
329static int fio_netio_prep(struct thread_data *td, struct io_u *io_u)
330{
331 struct netio_options *o = td->eo;
332
333 /*
334 * Make sure we don't see spurious reads to a receiver, and vice versa
335 */
336 if (is_tcp(o))
337 return 0;
338
339 if ((o->listen && io_u->ddir == DDIR_WRITE) ||
340 (!o->listen && io_u->ddir == DDIR_READ)) {
341 td_verror(td, EINVAL, "bad direction");
342 return 1;
343 }
344
345 return 0;
346}
347
348#ifdef CONFIG_LINUX_SPLICE
349static int splice_io_u(int fdin, int fdout, unsigned int len)
350{
351 int bytes = 0;
352
353 while (len) {
354 int ret = splice(fdin, NULL, fdout, NULL, len, 0);
355
356 if (ret < 0) {
357 if (!bytes)
358 bytes = ret;
359
360 break;
361 } else if (!ret)
362 break;
363
364 bytes += ret;
365 len -= ret;
366 }
367
368 return bytes;
369}
370
371/*
372 * Receive bytes from a socket and fill them into the internal pipe
373 */
374static int splice_in(struct thread_data *td, struct io_u *io_u)
375{
376 struct netio_data *nd = td->io_ops->data;
377
378 return splice_io_u(io_u->file->fd, nd->pipes[1], io_u->xfer_buflen);
379}
380
381/*
382 * Transmit 'len' bytes from the internal pipe
383 */
384static int splice_out(struct thread_data *td, struct io_u *io_u,
385 unsigned int len)
386{
387 struct netio_data *nd = td->io_ops->data;
388
389 return splice_io_u(nd->pipes[0], io_u->file->fd, len);
390}
391
392static int vmsplice_io_u(struct io_u *io_u, int fd, unsigned int len)
393{
394 struct iovec iov = {
395 .iov_base = io_u->xfer_buf,
396 .iov_len = len,
397 };
398 int bytes = 0;
399
400 while (iov.iov_len) {
401 int ret = vmsplice(fd, &iov, 1, SPLICE_F_MOVE);
402
403 if (ret < 0) {
404 if (!bytes)
405 bytes = ret;
406 break;
407 } else if (!ret)
408 break;
409
410 iov.iov_len -= ret;
411 iov.iov_base += ret;
412 bytes += ret;
413 }
414
415 return bytes;
416
417}
418
419/*
420 * vmsplice() pipe to io_u buffer
421 */
422static int vmsplice_io_u_out(struct thread_data *td, struct io_u *io_u,
423 unsigned int len)
424{
425 struct netio_data *nd = td->io_ops->data;
426
427 return vmsplice_io_u(io_u, nd->pipes[0], len);
428}
429
430/*
431 * vmsplice() io_u to pipe
432 */
433static int vmsplice_io_u_in(struct thread_data *td, struct io_u *io_u)
434{
435 struct netio_data *nd = td->io_ops->data;
436
437 return vmsplice_io_u(io_u, nd->pipes[1], io_u->xfer_buflen);
438}
439
440/*
441 * splice receive - transfer socket data into a pipe using splice, then map
442 * that pipe data into the io_u using vmsplice.
443 */
444static int fio_netio_splice_in(struct thread_data *td, struct io_u *io_u)
445{
446 int ret;
447
448 ret = splice_in(td, io_u);
449 if (ret > 0)
450 return vmsplice_io_u_out(td, io_u, ret);
451
452 return ret;
453}
454
455/*
456 * splice transmit - map data from the io_u into a pipe by using vmsplice,
457 * then transfer that pipe to a socket using splice.
458 */
459static int fio_netio_splice_out(struct thread_data *td, struct io_u *io_u)
460{
461 int ret;
462
463 ret = vmsplice_io_u_in(td, io_u);
464 if (ret > 0)
465 return splice_out(td, io_u, ret);
466
467 return ret;
468}
469#else
470static int fio_netio_splice_in(struct thread_data *td, struct io_u *io_u)
471{
472 errno = EOPNOTSUPP;
473 return -1;
474}
475
476static int fio_netio_splice_out(struct thread_data *td, struct io_u *io_u)
477{
478 errno = EOPNOTSUPP;
479 return -1;
480}
481#endif
482
483static void store_udp_seq(struct netio_data *nd, struct io_u *io_u)
484{
485 struct udp_seq *us;
486
487 us = io_u->xfer_buf + io_u->xfer_buflen - sizeof(*us);
488 us->magic = cpu_to_le64(FIO_UDP_SEQ_MAGIC);
489 us->bs = cpu_to_le64((uint64_t) io_u->xfer_buflen);
490 us->seq = cpu_to_le64(nd->udp_send_seq++);
491}
492
493static void verify_udp_seq(struct thread_data *td, struct netio_data *nd,
494 struct io_u *io_u)
495{
496 struct udp_seq *us;
497 uint64_t seq;
498
499 if (nd->seq_off)
500 return;
501
502 us = io_u->xfer_buf + io_u->xfer_buflen - sizeof(*us);
503 if (le64_to_cpu(us->magic) != FIO_UDP_SEQ_MAGIC)
504 return;
505 if (le64_to_cpu(us->bs) != io_u->xfer_buflen) {
506 nd->seq_off = 1;
507 return;
508 }
509
510 seq = le64_to_cpu(us->seq);
511
512 if (seq != nd->udp_recv_seq)
513 td->ts.drop_io_u[io_u->ddir] += seq - nd->udp_recv_seq;
514
515 nd->udp_recv_seq = seq + 1;
516}
517
518static int fio_netio_send(struct thread_data *td, struct io_u *io_u)
519{
520 struct netio_data *nd = td->io_ops->data;
521 struct netio_options *o = td->eo;
522 int ret, flags = 0;
523
524 do {
525 if (is_udp(o)) {
526 const struct sockaddr *to;
527 socklen_t len;
528
529 if (is_ipv6(o)) {
530 to = (struct sockaddr *) &nd->addr6;
531 len = sizeof(nd->addr6);
532 } else {
533 to = (struct sockaddr *) &nd->addr;
534 len = sizeof(nd->addr);
535 }
536
537 if (td->o.verify == VERIFY_NONE)
538 store_udp_seq(nd, io_u);
539
540 ret = sendto(io_u->file->fd, io_u->xfer_buf,
541 io_u->xfer_buflen, flags, to, len);
542 } else {
543 /*
544 * if we are going to write more, set MSG_MORE
545 */
546#ifdef MSG_MORE
547 if ((td->this_io_bytes[DDIR_WRITE] + io_u->xfer_buflen <
548 td->o.size) && !o->pingpong)
549 flags |= MSG_MORE;
550#endif
551 ret = send(io_u->file->fd, io_u->xfer_buf,
552 io_u->xfer_buflen, flags);
553 }
554 if (ret > 0)
555 break;
556
557 ret = poll_wait(td, io_u->file->fd, POLLOUT);
558 if (ret <= 0)
559 break;
560 } while (1);
561
562 return ret;
563}
564
565static int is_close_msg(struct io_u *io_u, int len)
566{
567 struct udp_close_msg *msg;
568
569 if (len != sizeof(struct udp_close_msg))
570 return 0;
571
572 msg = io_u->xfer_buf;
573 if (le32_to_cpu(msg->magic) != FIO_LINK_OPEN_CLOSE_MAGIC)
574 return 0;
575 if (le32_to_cpu(msg->cmd) != FIO_LINK_CLOSE)
576 return 0;
577
578 return 1;
579}
580
581static int fio_netio_recv(struct thread_data *td, struct io_u *io_u)
582{
583 struct netio_data *nd = td->io_ops->data;
584 struct netio_options *o = td->eo;
585 int ret, flags = 0;
586
587 do {
588 if (is_udp(o)) {
589 struct sockaddr *from;
590 socklen_t l, *len = &l;
591
592 if (o->listen) {
593 if (!is_ipv6(o)) {
594 from = (struct sockaddr *) &nd->addr;
595 *len = sizeof(nd->addr);
596 } else {
597 from = (struct sockaddr *) &nd->addr6;
598 *len = sizeof(nd->addr6);
599 }
600 } else {
601 from = NULL;
602 len = NULL;
603 }
604
605 ret = recvfrom(io_u->file->fd, io_u->xfer_buf,
606 io_u->xfer_buflen, flags, from, len);
607
608 if (is_close_msg(io_u, ret)) {
609 td->done = 1;
610 return 0;
611 }
612 } else {
613 ret = recv(io_u->file->fd, io_u->xfer_buf,
614 io_u->xfer_buflen, flags);
615
616 if (is_close_msg(io_u, ret)) {
617 td->done = 1;
618 return 0;
619 }
620 }
621 if (ret > 0)
622 break;
623 else if (!ret && (flags & MSG_WAITALL))
624 break;
625
626 ret = poll_wait(td, io_u->file->fd, POLLIN);
627 if (ret <= 0)
628 break;
629 flags |= MSG_WAITALL;
630 } while (1);
631
632 if (is_udp(o) && td->o.verify == VERIFY_NONE)
633 verify_udp_seq(td, nd, io_u);
634
635 return ret;
636}
637
638static int __fio_netio_queue(struct thread_data *td, struct io_u *io_u,
639 enum fio_ddir ddir)
640{
641 struct netio_data *nd = td->io_ops->data;
642 struct netio_options *o = td->eo;
643 int ret;
644
645 if (ddir == DDIR_WRITE) {
646 if (!nd->use_splice || is_udp(o) ||
647 o->proto == FIO_TYPE_UNIX)
648 ret = fio_netio_send(td, io_u);
649 else
650 ret = fio_netio_splice_out(td, io_u);
651 } else if (ddir == DDIR_READ) {
652 if (!nd->use_splice || is_udp(o) ||
653 o->proto == FIO_TYPE_UNIX)
654 ret = fio_netio_recv(td, io_u);
655 else
656 ret = fio_netio_splice_in(td, io_u);
657 } else
658 ret = 0; /* must be a SYNC */
659
660 if (ret != (int) io_u->xfer_buflen) {
661 if (ret > 0) {
662 io_u->resid = io_u->xfer_buflen - ret;
663 io_u->error = 0;
664 return FIO_Q_COMPLETED;
665 } else if (!ret)
666 return FIO_Q_BUSY;
667 else {
668 int err = errno;
669
670 if (ddir == DDIR_WRITE && err == EMSGSIZE)
671 return FIO_Q_BUSY;
672
673 io_u->error = err;
674 }
675 }
676
677 if (io_u->error)
678 td_verror(td, io_u->error, "xfer");
679
680 return FIO_Q_COMPLETED;
681}
682
683static int fio_netio_queue(struct thread_data *td, struct io_u *io_u)
684{
685 struct netio_options *o = td->eo;
686 int ret;
687
688 fio_ro_check(td, io_u);
689
690 ret = __fio_netio_queue(td, io_u, io_u->ddir);
691 if (!o->pingpong || ret != FIO_Q_COMPLETED)
692 return ret;
693
694 /*
695 * For ping-pong mode, receive or send reply as needed
696 */
697 if (td_read(td) && io_u->ddir == DDIR_READ)
698 ret = __fio_netio_queue(td, io_u, DDIR_WRITE);
699 else if (td_write(td) && io_u->ddir == DDIR_WRITE)
700 ret = __fio_netio_queue(td, io_u, DDIR_READ);
701
702 return ret;
703}
704
705static int fio_netio_connect(struct thread_data *td, struct fio_file *f)
706{
707 struct netio_data *nd = td->io_ops->data;
708 struct netio_options *o = td->eo;
709 int type, domain;
710
711 if (o->proto == FIO_TYPE_TCP) {
712 domain = AF_INET;
713 type = SOCK_STREAM;
714 } else if (o->proto == FIO_TYPE_TCP_V6) {
715 domain = AF_INET6;
716 type = SOCK_STREAM;
717 } else if (o->proto == FIO_TYPE_UDP) {
718 domain = AF_INET;
719 type = SOCK_DGRAM;
720 } else if (o->proto == FIO_TYPE_UDP_V6) {
721 domain = AF_INET6;
722 type = SOCK_DGRAM;
723 } else if (o->proto == FIO_TYPE_UNIX) {
724 domain = AF_UNIX;
725 type = SOCK_STREAM;
726 } else {
727 log_err("fio: bad network type %d\n", o->proto);
728 f->fd = -1;
729 return 1;
730 }
731
732 f->fd = socket(domain, type, 0);
733 if (f->fd < 0) {
734 td_verror(td, errno, "socket");
735 return 1;
736 }
737
738#ifdef CONFIG_TCP_NODELAY
739 if (o->nodelay && is_tcp(o)) {
740 int optval = 1;
741
742 if (setsockopt(f->fd, IPPROTO_TCP, TCP_NODELAY, (void *) &optval, sizeof(int)) < 0) {
743 log_err("fio: cannot set TCP_NODELAY option on socket (%s), disable with 'nodelay=0'\n", strerror(errno));
744 return 1;
745 }
746 }
747#endif
748
749 if (set_window_size(td, f->fd)) {
750 close(f->fd);
751 return 1;
752 }
753 if (set_mss(td, f->fd)) {
754 close(f->fd);
755 return 1;
756 }
757
758 if (is_udp(o)) {
759 if (!fio_netio_is_multicast(td->o.filename))
760 return 0;
761 if (is_ipv6(o)) {
762 log_err("fio: multicast not supported on IPv6\n");
763 close(f->fd);
764 return 1;
765 }
766
767 if (o->intfc) {
768 struct in_addr interface_addr;
769
770 if (inet_aton(o->intfc, &interface_addr) == 0) {
771 log_err("fio: interface not valid interface IP\n");
772 close(f->fd);
773 return 1;
774 }
775 if (setsockopt(f->fd, IPPROTO_IP, IP_MULTICAST_IF, (const char*)&interface_addr, sizeof(interface_addr)) < 0) {
776 td_verror(td, errno, "setsockopt IP_MULTICAST_IF");
777 close(f->fd);
778 return 1;
779 }
780 }
781 if (setsockopt(f->fd, IPPROTO_IP, IP_MULTICAST_TTL, (const char*)&o->ttl, sizeof(o->ttl)) < 0) {
782 td_verror(td, errno, "setsockopt IP_MULTICAST_TTL");
783 close(f->fd);
784 return 1;
785 }
786 return 0;
787 } else if (o->proto == FIO_TYPE_TCP) {
788 socklen_t len = sizeof(nd->addr);
789
790 if (connect(f->fd, (struct sockaddr *) &nd->addr, len) < 0) {
791 td_verror(td, errno, "connect");
792 close(f->fd);
793 return 1;
794 }
795 } else if (o->proto == FIO_TYPE_TCP_V6) {
796 socklen_t len = sizeof(nd->addr6);
797
798 if (connect(f->fd, (struct sockaddr *) &nd->addr6, len) < 0) {
799 td_verror(td, errno, "connect");
800 close(f->fd);
801 return 1;
802 }
803
804 } else {
805 struct sockaddr_un *addr = &nd->addr_un;
806 socklen_t len;
807
808 len = sizeof(addr->sun_family) + strlen(addr->sun_path) + 1;
809
810 if (connect(f->fd, (struct sockaddr *) addr, len) < 0) {
811 td_verror(td, errno, "connect");
812 close(f->fd);
813 return 1;
814 }
815 }
816
817 return 0;
818}
819
820static int fio_netio_accept(struct thread_data *td, struct fio_file *f)
821{
822 struct netio_data *nd = td->io_ops->data;
823 struct netio_options *o = td->eo;
824 socklen_t socklen;
825 int state;
826
827 if (is_udp(o)) {
828 f->fd = nd->listenfd;
829 return 0;
830 }
831
832 state = td->runstate;
833 td_set_runstate(td, TD_SETTING_UP);
834
835 log_info("fio: waiting for connection\n");
836
837 if (poll_wait(td, nd->listenfd, POLLIN) < 0)
838 goto err;
839
840 if (o->proto == FIO_TYPE_TCP) {
841 socklen = sizeof(nd->addr);
842 f->fd = accept(nd->listenfd, (struct sockaddr *) &nd->addr, &socklen);
843 } else {
844 socklen = sizeof(nd->addr6);
845 f->fd = accept(nd->listenfd, (struct sockaddr *) &nd->addr6, &socklen);
846 }
847
848 if (f->fd < 0) {
849 td_verror(td, errno, "accept");
850 goto err;
851 }
852
853#ifdef CONFIG_TCP_NODELAY
854 if (o->nodelay && is_tcp(o)) {
855 int optval = 1;
856
857 if (setsockopt(f->fd, IPPROTO_TCP, TCP_NODELAY, (void *) &optval, sizeof(int)) < 0) {
858 log_err("fio: cannot set TCP_NODELAY option on socket (%s), disable with 'nodelay=0'\n", strerror(errno));
859 return 1;
860 }
861 }
862#endif
863
864 reset_all_stats(td);
865 td_set_runstate(td, state);
866 return 0;
867err:
868 td_set_runstate(td, state);
869 return 1;
870}
871
872static void fio_netio_send_close(struct thread_data *td, struct fio_file *f)
873{
874 struct netio_data *nd = td->io_ops->data;
875 struct netio_options *o = td->eo;
876 struct udp_close_msg msg;
877 struct sockaddr *to;
878 socklen_t len;
879 int ret;
880
881 if (is_ipv6(o)) {
882 to = (struct sockaddr *) &nd->addr6;
883 len = sizeof(nd->addr6);
884 } else {
885 to = (struct sockaddr *) &nd->addr;
886 len = sizeof(nd->addr);
887 }
888
889 msg.magic = cpu_to_le32((uint32_t) FIO_LINK_OPEN_CLOSE_MAGIC);
890 msg.cmd = cpu_to_le32((uint32_t) FIO_LINK_CLOSE);
891
892 ret = sendto(f->fd, (void *) &msg, sizeof(msg), MSG_WAITALL, to, len);
893 if (ret < 0)
894 td_verror(td, errno, "sendto udp link close");
895}
896
897static int fio_netio_close_file(struct thread_data *td, struct fio_file *f)
898{
899 /*
900 * Notify the receiver that we are closing down the link
901 */
902 fio_netio_send_close(td, f);
903
904 return generic_close_file(td, f);
905}
906
907static int fio_netio_udp_recv_open(struct thread_data *td, struct fio_file *f)
908{
909 struct netio_data *nd = td->io_ops->data;
910 struct netio_options *o = td->eo;
911 struct udp_close_msg msg;
912 struct sockaddr *to;
913 socklen_t len;
914 int ret;
915
916 if (is_ipv6(o)) {
917 len = sizeof(nd->addr6);
918 to = (struct sockaddr *) &nd->addr6;
919 } else {
920 len = sizeof(nd->addr);
921 to = (struct sockaddr *) &nd->addr;
922 }
923
924 ret = recvfrom(f->fd, (void *) &msg, sizeof(msg), MSG_WAITALL, to, &len);
925 if (ret < 0) {
926 td_verror(td, errno, "recvfrom udp link open");
927 return ret;
928 }
929
930 if (ntohl(msg.magic) != FIO_LINK_OPEN_CLOSE_MAGIC ||
931 ntohl(msg.cmd) != FIO_LINK_OPEN) {
932 log_err("fio: bad udp open magic %x/%x\n", ntohl(msg.magic),
933 ntohl(msg.cmd));
934 return -1;
935 }
936
937 fio_gettime(&td->start, NULL);
938 return 0;
939}
940
941static int fio_netio_send_open(struct thread_data *td, struct fio_file *f)
942{
943 struct netio_data *nd = td->io_ops->data;
944 struct netio_options *o = td->eo;
945 struct udp_close_msg msg;
946 struct sockaddr *to;
947 socklen_t len;
948 int ret;
949
950 if (is_ipv6(o)) {
951 len = sizeof(nd->addr6);
952 to = (struct sockaddr *) &nd->addr6;
953 } else {
954 len = sizeof(nd->addr);
955 to = (struct sockaddr *) &nd->addr;
956 }
957
958 msg.magic = htonl(FIO_LINK_OPEN_CLOSE_MAGIC);
959 msg.cmd = htonl(FIO_LINK_OPEN);
960
961 ret = sendto(f->fd, (void *) &msg, sizeof(msg), MSG_WAITALL, to, len);
962 if (ret < 0) {
963 td_verror(td, errno, "sendto udp link open");
964 return ret;
965 }
966
967 return 0;
968}
969
970static int fio_netio_open_file(struct thread_data *td, struct fio_file *f)
971{
972 int ret;
973 struct netio_options *o = td->eo;
974
975 if (o->listen)
976 ret = fio_netio_accept(td, f);
977 else
978 ret = fio_netio_connect(td, f);
979
980 if (ret) {
981 f->fd = -1;
982 return ret;
983 }
984
985 if (is_udp(o)) {
986 if (td_write(td))
987 ret = fio_netio_send_open(td, f);
988 else {
989 int state;
990
991 state = td->runstate;
992 td_set_runstate(td, TD_SETTING_UP);
993 ret = fio_netio_udp_recv_open(td, f);
994 td_set_runstate(td, state);
995 }
996 }
997
998 if (ret)
999 fio_netio_close_file(td, f);
1000
1001 return ret;
1002}
1003
1004static int fio_fill_addr(struct thread_data *td, const char *host, int af,
1005 void *dst, struct addrinfo **res)
1006{
1007 struct netio_options *o = td->eo;
1008 struct addrinfo hints;
1009 int ret;
1010
1011 if (inet_pton(af, host, dst))
1012 return 0;
1013
1014 memset(&hints, 0, sizeof(hints));
1015
1016 if (is_tcp(o))
1017 hints.ai_socktype = SOCK_STREAM;
1018 else
1019 hints.ai_socktype = SOCK_DGRAM;
1020
1021 if (is_ipv6(o))
1022 hints.ai_family = AF_INET6;
1023 else
1024 hints.ai_family = AF_INET;
1025
1026 ret = getaddrinfo(host, NULL, &hints, res);
1027 if (ret) {
1028 int e = EINVAL;
1029 char str[128];
1030
1031 if (ret == EAI_SYSTEM)
1032 e = errno;
1033
1034 snprintf(str, sizeof(str), "getaddrinfo: %s", gai_strerror(ret));
1035 td_verror(td, e, str);
1036 return 1;
1037 }
1038
1039 return 0;
1040}
1041
1042static int fio_netio_setup_connect_inet(struct thread_data *td,
1043 const char *host, unsigned short port)
1044{
1045 struct netio_data *nd = td->io_ops->data;
1046 struct netio_options *o = td->eo;
1047 struct addrinfo *res = NULL;
1048 void *dst, *src;
1049 int af, len;
1050
1051 if (!host) {
1052 log_err("fio: connect with no host to connect to.\n");
1053 if (td_read(td))
1054 log_err("fio: did you forget to set 'listen'?\n");
1055
1056 td_verror(td, EINVAL, "no hostname= set");
1057 return 1;
1058 }
1059
1060 nd->addr.sin_family = AF_INET;
1061 nd->addr.sin_port = htons(port);
1062 nd->addr6.sin6_family = AF_INET6;
1063 nd->addr6.sin6_port = htons(port);
1064
1065 if (is_ipv6(o)) {
1066 af = AF_INET6;
1067 dst = &nd->addr6.sin6_addr;
1068 } else {
1069 af = AF_INET;
1070 dst = &nd->addr.sin_addr;
1071 }
1072
1073 if (fio_fill_addr(td, host, af, dst, &res))
1074 return 1;
1075
1076 if (!res)
1077 return 0;
1078
1079 if (is_ipv6(o)) {
1080 len = sizeof(nd->addr6.sin6_addr);
1081 src = &((struct sockaddr_in6 *) res->ai_addr)->sin6_addr;
1082 } else {
1083 len = sizeof(nd->addr.sin_addr);
1084 src = &((struct sockaddr_in *) res->ai_addr)->sin_addr;
1085 }
1086
1087 memcpy(dst, src, len);
1088 freeaddrinfo(res);
1089 return 0;
1090}
1091
1092static int fio_netio_setup_connect_unix(struct thread_data *td,
1093 const char *path)
1094{
1095 struct netio_data *nd = td->io_ops->data;
1096 struct sockaddr_un *soun = &nd->addr_un;
1097
1098 soun->sun_family = AF_UNIX;
1099 memset(soun->sun_path, 0, sizeof(soun->sun_path));
1100 strncpy(soun->sun_path, path, sizeof(soun->sun_path) - 1);
1101 return 0;
1102}
1103
1104static int fio_netio_setup_connect(struct thread_data *td)
1105{
1106 struct netio_options *o = td->eo;
1107
1108 if (is_udp(o) || is_tcp(o))
1109 return fio_netio_setup_connect_inet(td, td->o.filename,o->port);
1110 else
1111 return fio_netio_setup_connect_unix(td, td->o.filename);
1112}
1113
1114static int fio_netio_setup_listen_unix(struct thread_data *td, const char *path)
1115{
1116 struct netio_data *nd = td->io_ops->data;
1117 struct sockaddr_un *addr = &nd->addr_un;
1118 mode_t mode;
1119 int len, fd;
1120
1121 fd = socket(AF_UNIX, SOCK_STREAM, 0);
1122 if (fd < 0) {
1123 log_err("fio: socket: %s\n", strerror(errno));
1124 return -1;
1125 }
1126
1127 mode = umask(000);
1128
1129 memset(addr, 0, sizeof(*addr));
1130 addr->sun_family = AF_UNIX;
1131 strncpy(addr->sun_path, path, sizeof(addr->sun_path) - 1);
1132 unlink(path);
1133
1134 len = sizeof(addr->sun_family) + strlen(path) + 1;
1135
1136 if (bind(fd, (struct sockaddr *) addr, len) < 0) {
1137 log_err("fio: bind: %s\n", strerror(errno));
1138 close(fd);
1139 return -1;
1140 }
1141
1142 umask(mode);
1143 nd->listenfd = fd;
1144 return 0;
1145}
1146
1147static int fio_netio_setup_listen_inet(struct thread_data *td, short port)
1148{
1149 struct netio_data *nd = td->io_ops->data;
1150 struct netio_options *o = td->eo;
1151 struct ip_mreq mr;
1152 struct sockaddr_in sin;
1153 struct sockaddr *saddr;
1154 int fd, opt, type, domain;
1155 socklen_t len;
1156
1157 memset(&sin, 0, sizeof(sin));
1158
1159 if (o->proto == FIO_TYPE_TCP) {
1160 type = SOCK_STREAM;
1161 domain = AF_INET;
1162 } else if (o->proto == FIO_TYPE_TCP_V6) {
1163 type = SOCK_STREAM;
1164 domain = AF_INET6;
1165 } else if (o->proto == FIO_TYPE_UDP) {
1166 type = SOCK_DGRAM;
1167 domain = AF_INET;
1168 } else if (o->proto == FIO_TYPE_UDP_V6) {
1169 type = SOCK_DGRAM;
1170 domain = AF_INET6;
1171 } else {
1172 log_err("fio: unknown proto %d\n", o->proto);
1173 return 1;
1174 }
1175
1176 fd = socket(domain, type, 0);
1177 if (fd < 0) {
1178 td_verror(td, errno, "socket");
1179 return 1;
1180 }
1181
1182 opt = 1;
1183 if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, (void *) &opt, sizeof(opt)) < 0) {
1184 td_verror(td, errno, "setsockopt");
1185 close(fd);
1186 return 1;
1187 }
1188#ifdef SO_REUSEPORT
1189 if (setsockopt(fd, SOL_SOCKET, SO_REUSEPORT, (void *) &opt, sizeof(opt)) < 0) {
1190 td_verror(td, errno, "setsockopt");
1191 close(fd);
1192 return 1;
1193 }
1194#endif
1195
1196 if (set_window_size(td, fd)) {
1197 close(fd);
1198 return 1;
1199 }
1200 if (set_mss(td, fd)) {
1201 close(fd);
1202 return 1;
1203 }
1204
1205 if (td->o.filename) {
1206 if (!is_udp(o) || !fio_netio_is_multicast(td->o.filename)) {
1207 log_err("fio: hostname not valid for non-multicast inbound network IO\n");
1208 close(fd);
1209 return 1;
1210 }
1211 if (is_ipv6(o)) {
1212 log_err("fio: IPv6 not supported for multicast network IO");
1213 close(fd);
1214 return 1;
1215 }
1216
1217 inet_aton(td->o.filename, &sin.sin_addr);
1218
1219 mr.imr_multiaddr = sin.sin_addr;
1220 if (o->intfc) {
1221 if (inet_aton(o->intfc, &mr.imr_interface) == 0) {
1222 log_err("fio: interface not valid interface IP\n");
1223 close(fd);
1224 return 1;
1225 }
1226 } else {
1227 mr.imr_interface.s_addr = htonl(INADDR_ANY);
1228 }
1229
1230 if (setsockopt(fd, IPPROTO_IP, IP_ADD_MEMBERSHIP, (const char*)&mr, sizeof(mr)) < 0) {
1231 td_verror(td, errno, "setsockopt IP_ADD_MEMBERSHIP");
1232 close(fd);
1233 return 1;
1234 }
1235 }
1236
1237 if (!is_ipv6(o)) {
1238 saddr = (struct sockaddr *) &nd->addr;
1239 len = sizeof(nd->addr);
1240
1241 nd->addr.sin_family = AF_INET;
1242 nd->addr.sin_addr.s_addr = sin.sin_addr.s_addr ? sin.sin_addr.s_addr : htonl(INADDR_ANY);
1243 nd->addr.sin_port = htons(port);
1244 } else {
1245 saddr = (struct sockaddr *) &nd->addr6;
1246 len = sizeof(nd->addr6);
1247
1248 nd->addr6.sin6_family = AF_INET6;
1249 nd->addr6.sin6_addr = in6addr_any;
1250 nd->addr6.sin6_port = htons(port);
1251 }
1252
1253 if (bind(fd, saddr, len) < 0) {
1254 close(fd);
1255 td_verror(td, errno, "bind");
1256 return 1;
1257 }
1258
1259 nd->listenfd = fd;
1260 return 0;
1261}
1262
1263static int fio_netio_setup_listen(struct thread_data *td)
1264{
1265 struct netio_data *nd = td->io_ops->data;
1266 struct netio_options *o = td->eo;
1267 int ret;
1268
1269 if (is_udp(o) || is_tcp(o))
1270 ret = fio_netio_setup_listen_inet(td, o->port);
1271 else
1272 ret = fio_netio_setup_listen_unix(td, td->o.filename);
1273
1274 if (ret)
1275 return ret;
1276 if (is_udp(o))
1277 return 0;
1278
1279 if (listen(nd->listenfd, 10) < 0) {
1280 td_verror(td, errno, "listen");
1281 nd->listenfd = -1;
1282 return 1;
1283 }
1284
1285 return 0;
1286}
1287
1288static int fio_netio_init(struct thread_data *td)
1289{
1290 struct netio_options *o = td->eo;
1291 int ret;
1292
1293#ifdef WIN32
1294 WSADATA wsd;
1295 WSAStartup(MAKEWORD(2,2), &wsd);
1296#endif
1297
1298 if (td_random(td)) {
1299 log_err("fio: network IO can't be random\n");
1300 return 1;
1301 }
1302
1303 if (o->proto == FIO_TYPE_UNIX && o->port) {
1304 log_err("fio: network IO port not valid with unix socket\n");
1305 return 1;
1306 } else if (o->proto != FIO_TYPE_UNIX && !o->port) {
1307 log_err("fio: network IO requires port for tcp or udp\n");
1308 return 1;
1309 }
1310
1311 o->port += td->subjob_number;
1312
1313 if (!is_tcp(o)) {
1314 if (o->listen) {
1315 log_err("fio: listen only valid for TCP proto IO\n");
1316 return 1;
1317 }
1318 if (td_rw(td)) {
1319 log_err("fio: datagram network connections must be"
1320 " read OR write\n");
1321 return 1;
1322 }
1323 if (o->proto == FIO_TYPE_UNIX && !td->o.filename) {
1324 log_err("fio: UNIX sockets need host/filename\n");
1325 return 1;
1326 }
1327 o->listen = td_read(td);
1328 }
1329
1330 if (o->listen)
1331 ret = fio_netio_setup_listen(td);
1332 else
1333 ret = fio_netio_setup_connect(td);
1334
1335 return ret;
1336}
1337
1338static void fio_netio_cleanup(struct thread_data *td)
1339{
1340 struct netio_data *nd = td->io_ops->data;
1341
1342 if (nd) {
1343 if (nd->listenfd != -1)
1344 close(nd->listenfd);
1345 if (nd->pipes[0] != -1)
1346 close(nd->pipes[0]);
1347 if (nd->pipes[1] != -1)
1348 close(nd->pipes[1]);
1349
1350 free(nd);
1351 }
1352}
1353
1354static int fio_netio_setup(struct thread_data *td)
1355{
1356 struct netio_data *nd;
1357
1358 if (!td->files_index) {
1359 add_file(td, td->o.filename ?: "net", 0, 0);
1360 td->o.nr_files = td->o.nr_files ?: 1;
1361 td->o.open_files++;
1362 }
1363
1364 if (!td->io_ops->data) {
1365 nd = malloc(sizeof(*nd));;
1366
1367 memset(nd, 0, sizeof(*nd));
1368 nd->listenfd = -1;
1369 nd->pipes[0] = nd->pipes[1] = -1;
1370 td->io_ops->data = nd;
1371 }
1372
1373 return 0;
1374}
1375
1376static void fio_netio_terminate(struct thread_data *td)
1377{
1378 kill(td->pid, SIGTERM);
1379}
1380
1381#ifdef CONFIG_LINUX_SPLICE
1382static int fio_netio_setup_splice(struct thread_data *td)
1383{
1384 struct netio_data *nd;
1385
1386 fio_netio_setup(td);
1387
1388 nd = td->io_ops->data;
1389 if (nd) {
1390 if (pipe(nd->pipes) < 0)
1391 return 1;
1392
1393 nd->use_splice = 1;
1394 return 0;
1395 }
1396
1397 return 1;
1398}
1399
1400static struct ioengine_ops ioengine_splice = {
1401 .name = "netsplice",
1402 .version = FIO_IOOPS_VERSION,
1403 .prep = fio_netio_prep,
1404 .queue = fio_netio_queue,
1405 .setup = fio_netio_setup_splice,
1406 .init = fio_netio_init,
1407 .cleanup = fio_netio_cleanup,
1408 .open_file = fio_netio_open_file,
1409 .close_file = fio_netio_close_file,
1410 .terminate = fio_netio_terminate,
1411 .options = options,
1412 .option_struct_size = sizeof(struct netio_options),
1413 .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR |
1414 FIO_PIPEIO,
1415};
1416#endif
1417
1418static struct ioengine_ops ioengine_rw = {
1419 .name = "net",
1420 .version = FIO_IOOPS_VERSION,
1421 .prep = fio_netio_prep,
1422 .queue = fio_netio_queue,
1423 .setup = fio_netio_setup,
1424 .init = fio_netio_init,
1425 .cleanup = fio_netio_cleanup,
1426 .open_file = fio_netio_open_file,
1427 .close_file = fio_netio_close_file,
1428 .terminate = fio_netio_terminate,
1429 .options = options,
1430 .option_struct_size = sizeof(struct netio_options),
1431 .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR |
1432 FIO_PIPEIO | FIO_BIT_BASED,
1433};
1434
1435static int str_hostname_cb(void *data, const char *input)
1436{
1437 struct netio_options *o = data;
1438
1439 if (o->td->o.filename)
1440 free(o->td->o.filename);
1441 o->td->o.filename = strdup(input);
1442 return 0;
1443}
1444
1445static void fio_init fio_netio_register(void)
1446{
1447 register_ioengine(&ioengine_rw);
1448#ifdef CONFIG_LINUX_SPLICE
1449 register_ioengine(&ioengine_splice);
1450#endif
1451}
1452
1453static void fio_exit fio_netio_unregister(void)
1454{
1455 unregister_ioengine(&ioengine_rw);
1456#ifdef CONFIG_LINUX_SPLICE
1457 unregister_ioengine(&ioengine_splice);
1458#endif
1459}
fio - Flexible IO Tester