projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
eta: don't count TD_SETTING_UP as a running process
[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_udp_close(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 (ntohl(msg->magic) != FIO_LINK_OPEN_CLOSE_MAGIC)
574 return 0;
575 if (ntohl(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_udp_close(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 (ret > 0)
617 break;
618 else if (!ret && (flags & MSG_WAITALL))
619 break;
620
621 ret = poll_wait(td, io_u->file->fd, POLLIN);
622 if (ret <= 0)
623 break;
624 flags |= MSG_WAITALL;
625 } while (1);
626
627 if (is_udp(o) && td->o.verify == VERIFY_NONE)
628 verify_udp_seq(td, nd, io_u);
629
630 return ret;
631}
632
633static int __fio_netio_queue(struct thread_data *td, struct io_u *io_u,
634 enum fio_ddir ddir)
635{
636 struct netio_data *nd = td->io_ops->data;
637 struct netio_options *o = td->eo;
638 int ret;
639
640 if (ddir == DDIR_WRITE) {
641 if (!nd->use_splice || is_udp(o) ||
642 o->proto == FIO_TYPE_UNIX)
643 ret = fio_netio_send(td, io_u);
644 else
645 ret = fio_netio_splice_out(td, io_u);
646 } else if (ddir == DDIR_READ) {
647 if (!nd->use_splice || is_udp(o) ||
648 o->proto == FIO_TYPE_UNIX)
649 ret = fio_netio_recv(td, io_u);
650 else
651 ret = fio_netio_splice_in(td, io_u);
652 } else
653 ret = 0; /* must be a SYNC */
654
655 if (ret != (int) io_u->xfer_buflen) {
656 if (ret > 0) {
657 io_u->resid = io_u->xfer_buflen - ret;
658 io_u->error = 0;
659 return FIO_Q_COMPLETED;
660 } else if (!ret)
661 return FIO_Q_BUSY;
662 else {
663 int err = errno;
664
665 if (ddir == DDIR_WRITE && err == EMSGSIZE)
666 return FIO_Q_BUSY;
667
668 io_u->error = err;
669 }
670 }
671
672 if (io_u->error)
673 td_verror(td, io_u->error, "xfer");
674
675 return FIO_Q_COMPLETED;
676}
677
678static int fio_netio_queue(struct thread_data *td, struct io_u *io_u)
679{
680 struct netio_options *o = td->eo;
681 int ret;
682
683 fio_ro_check(td, io_u);
684
685 ret = __fio_netio_queue(td, io_u, io_u->ddir);
686 if (!o->pingpong || ret != FIO_Q_COMPLETED)
687 return ret;
688
689 /*
690 * For ping-pong mode, receive or send reply as needed
691 */
692 if (td_read(td) && io_u->ddir == DDIR_READ)
693 ret = __fio_netio_queue(td, io_u, DDIR_WRITE);
694 else if (td_write(td) && io_u->ddir == DDIR_WRITE)
695 ret = __fio_netio_queue(td, io_u, DDIR_READ);
696
697 return ret;
698}
699
700static int fio_netio_connect(struct thread_data *td, struct fio_file *f)
701{
702 struct netio_data *nd = td->io_ops->data;
703 struct netio_options *o = td->eo;
704 int type, domain;
705
706 if (o->proto == FIO_TYPE_TCP) {
707 domain = AF_INET;
708 type = SOCK_STREAM;
709 } else if (o->proto == FIO_TYPE_TCP_V6) {
710 domain = AF_INET6;
711 type = SOCK_STREAM;
712 } else if (o->proto == FIO_TYPE_UDP) {
713 domain = AF_INET;
714 type = SOCK_DGRAM;
715 } else if (o->proto == FIO_TYPE_UDP_V6) {
716 domain = AF_INET6;
717 type = SOCK_DGRAM;
718 } else if (o->proto == FIO_TYPE_UNIX) {
719 domain = AF_UNIX;
720 type = SOCK_STREAM;
721 } else {
722 log_err("fio: bad network type %d\n", o->proto);
723 f->fd = -1;
724 return 1;
725 }
726
727 f->fd = socket(domain, type, 0);
728 if (f->fd < 0) {
729 td_verror(td, errno, "socket");
730 return 1;
731 }
732
733#ifdef CONFIG_TCP_NODELAY
734 if (o->nodelay && is_tcp(o)) {
735 int optval = 1;
736
737 if (setsockopt(f->fd, IPPROTO_TCP, TCP_NODELAY, (void *) &optval, sizeof(int)) < 0) {
738 log_err("fio: cannot set TCP_NODELAY option on socket (%s), disable with 'nodelay=0'\n", strerror(errno));
739 return 1;
740 }
741 }
742#endif
743
744 if (set_window_size(td, f->fd)) {
745 close(f->fd);
746 return 1;
747 }
748 if (set_mss(td, f->fd)) {
749 close(f->fd);
750 return 1;
751 }
752
753 if (is_udp(o)) {
754 if (!fio_netio_is_multicast(td->o.filename))
755 return 0;
756 if (is_ipv6(o)) {
757 log_err("fio: multicast not supported on IPv6\n");
758 close(f->fd);
759 return 1;
760 }
761
762 if (o->intfc) {
763 struct in_addr interface_addr;
764
765 if (inet_aton(o->intfc, &interface_addr) == 0) {
766 log_err("fio: interface not valid interface IP\n");
767 close(f->fd);
768 return 1;
769 }
770 if (setsockopt(f->fd, IPPROTO_IP, IP_MULTICAST_IF, (const char*)&interface_addr, sizeof(interface_addr)) < 0) {
771 td_verror(td, errno, "setsockopt IP_MULTICAST_IF");
772 close(f->fd);
773 return 1;
774 }
775 }
776 if (setsockopt(f->fd, IPPROTO_IP, IP_MULTICAST_TTL, (const char*)&o->ttl, sizeof(o->ttl)) < 0) {
777 td_verror(td, errno, "setsockopt IP_MULTICAST_TTL");
778 close(f->fd);
779 return 1;
780 }
781 return 0;
782 } else if (o->proto == FIO_TYPE_TCP) {
783 socklen_t len = sizeof(nd->addr);
784
785 if (connect(f->fd, (struct sockaddr *) &nd->addr, len) < 0) {
786 td_verror(td, errno, "connect");
787 close(f->fd);
788 return 1;
789 }
790 } else if (o->proto == FIO_TYPE_TCP_V6) {
791 socklen_t len = sizeof(nd->addr6);
792
793 if (connect(f->fd, (struct sockaddr *) &nd->addr6, len) < 0) {
794 td_verror(td, errno, "connect");
795 close(f->fd);
796 return 1;
797 }
798
799 } else {
800 struct sockaddr_un *addr = &nd->addr_un;
801 socklen_t len;
802
803 len = sizeof(addr->sun_family) + strlen(addr->sun_path) + 1;
804
805 if (connect(f->fd, (struct sockaddr *) addr, len) < 0) {
806 td_verror(td, errno, "connect");
807 close(f->fd);
808 return 1;
809 }
810 }
811
812 return 0;
813}
814
815static int fio_netio_accept(struct thread_data *td, struct fio_file *f)
816{
817 struct netio_data *nd = td->io_ops->data;
818 struct netio_options *o = td->eo;
819 socklen_t socklen;
820 int state;
821
822 if (is_udp(o)) {
823 f->fd = nd->listenfd;
824 return 0;
825 }
826
827 state = td->runstate;
828 td_set_runstate(td, TD_SETTING_UP);
829
830 log_info("fio: waiting for connection\n");
831
832 if (poll_wait(td, nd->listenfd, POLLIN) < 0)
833 goto err;
834
835 if (o->proto == FIO_TYPE_TCP) {
836 socklen = sizeof(nd->addr);
837 f->fd = accept(nd->listenfd, (struct sockaddr *) &nd->addr, &socklen);
838 } else {
839 socklen = sizeof(nd->addr6);
840 f->fd = accept(nd->listenfd, (struct sockaddr *) &nd->addr6, &socklen);
841 }
842
843 if (f->fd < 0) {
844 td_verror(td, errno, "accept");
845 goto err;
846 }
847
848#ifdef CONFIG_TCP_NODELAY
849 if (o->nodelay && is_tcp(o)) {
850 int optval = 1;
851
852 if (setsockopt(f->fd, IPPROTO_TCP, TCP_NODELAY, (void *) &optval, sizeof(int)) < 0) {
853 log_err("fio: cannot set TCP_NODELAY option on socket (%s), disable with 'nodelay=0'\n", strerror(errno));
854 return 1;
855 }
856 }
857#endif
858
859 reset_all_stats(td);
860 td_set_runstate(td, state);
861 return 0;
862err:
863 td_set_runstate(td, state);
864 return 1;
865}
866
867static void fio_netio_udp_close(struct thread_data *td, struct fio_file *f)
868{
869 struct netio_data *nd = td->io_ops->data;
870 struct netio_options *o = td->eo;
871 struct udp_close_msg msg;
872 struct sockaddr *to;
873 socklen_t len;
874 int ret;
875
876 if (is_ipv6(o)) {
877 to = (struct sockaddr *) &nd->addr6;
878 len = sizeof(nd->addr6);
879 } else {
880 to = (struct sockaddr *) &nd->addr;
881 len = sizeof(nd->addr);
882 }
883
884 msg.magic = htonl(FIO_LINK_OPEN_CLOSE_MAGIC);
885 msg.cmd = htonl(FIO_LINK_CLOSE);
886
887 ret = sendto(f->fd, (void *) &msg, sizeof(msg), MSG_WAITALL, to, len);
888 if (ret < 0)
889 td_verror(td, errno, "sendto udp link close");
890}
891
892static int fio_netio_close_file(struct thread_data *td, struct fio_file *f)
893{
894 struct netio_options *o = td->eo;
895
896 /*
897 * If this is an UDP connection, notify the receiver that we are
898 * closing down the link
899 */
900 if (is_udp(o))
901 fio_netio_udp_close(td, f);
902
903 return generic_close_file(td, f);
904}
905
906static int fio_netio_udp_recv_open(struct thread_data *td, struct fio_file *f)
907{
908 struct netio_data *nd = td->io_ops->data;
909 struct netio_options *o = td->eo;
910 struct udp_close_msg msg;
911 struct sockaddr *to;
912 socklen_t len;
913 int ret;
914
915 if (is_ipv6(o)) {
916 len = sizeof(nd->addr6);
917 to = (struct sockaddr *) &nd->addr6;
918 } else {
919 len = sizeof(nd->addr);
920 to = (struct sockaddr *) &nd->addr;
921 }
922
923 ret = recvfrom(f->fd, (void *) &msg, sizeof(msg), MSG_WAITALL, to, &len);
924 if (ret < 0) {
925 td_verror(td, errno, "recvfrom udp link open");
926 return ret;
927 }
928
929 if (ntohl(msg.magic) != FIO_LINK_OPEN_CLOSE_MAGIC ||
930 ntohl(msg.cmd) != FIO_LINK_OPEN) {
931 log_err("fio: bad udp open magic %x/%x\n", ntohl(msg.magic),
932 ntohl(msg.cmd));
933 return -1;
934 }
935
936 fio_gettime(&td->start, NULL);
937 return 0;
938}
939
940static int fio_netio_udp_send_open(struct thread_data *td, struct fio_file *f)
941{
942 struct netio_data *nd = td->io_ops->data;
943 struct netio_options *o = td->eo;
944 struct udp_close_msg msg;
945 struct sockaddr *to;
946 socklen_t len;
947 int ret;
948
949 if (is_ipv6(o)) {
950 len = sizeof(nd->addr6);
951 to = (struct sockaddr *) &nd->addr6;
952 } else {
953 len = sizeof(nd->addr);
954 to = (struct sockaddr *) &nd->addr;
955 }
956
957 msg.magic = htonl(FIO_LINK_OPEN_CLOSE_MAGIC);
958 msg.cmd = htonl(FIO_LINK_OPEN);
959
960 ret = sendto(f->fd, (void *) &msg, sizeof(msg), MSG_WAITALL, to, len);
961 if (ret < 0) {
962 td_verror(td, errno, "sendto udp link open");
963 return ret;
964 }
965
966 return 0;
967}
968
969static int fio_netio_open_file(struct thread_data *td, struct fio_file *f)
970{
971 int ret;
972 struct netio_options *o = td->eo;
973
974 if (o->listen)
975 ret = fio_netio_accept(td, f);
976 else
977 ret = fio_netio_connect(td, f);
978
979 if (ret) {
980 f->fd = -1;
981 return ret;
982 }
983
984 if (is_udp(o)) {
985 if (td_write(td))
986 ret = fio_netio_udp_send_open(td, f);
987 else {
988 int state;
989
990 state = td->runstate;
991 td_set_runstate(td, TD_SETTING_UP);
992 ret = fio_netio_udp_recv_open(td, f);
993 td_set_runstate(td, state);
994 }
995 }
996
997 if (ret)
998 fio_netio_close_file(td, f);
999
1000 return ret;
1001}
1002
1003static int fio_fill_addr(struct thread_data *td, const char *host, int af,
1004 void *dst, struct addrinfo **res)
1005{
1006 struct netio_options *o = td->eo;
1007 struct addrinfo hints;
1008 int ret;
1009
1010 if (inet_pton(af, host, dst))
1011 return 0;
1012
1013 memset(&hints, 0, sizeof(hints));
1014
1015 if (is_tcp(o))
1016 hints.ai_socktype = SOCK_STREAM;
1017 else
1018 hints.ai_socktype = SOCK_DGRAM;
1019
1020 if (is_ipv6(o))
1021 hints.ai_family = AF_INET6;
1022 else
1023 hints.ai_family = AF_INET;
1024
1025 ret = getaddrinfo(host, NULL, &hints, res);
1026 if (ret) {
1027 int e = EINVAL;
1028 char str[128];
1029
1030 if (ret == EAI_SYSTEM)
1031 e = errno;
1032
1033 snprintf(str, sizeof(str), "getaddrinfo: %s", gai_strerror(ret));
1034 td_verror(td, e, str);
1035 return 1;
1036 }
1037
1038 return 0;
1039}
1040
1041static int fio_netio_setup_connect_inet(struct thread_data *td,
1042 const char *host, unsigned short port)
1043{
1044 struct netio_data *nd = td->io_ops->data;
1045 struct netio_options *o = td->eo;
1046 struct addrinfo *res = NULL;
1047 void *dst, *src;
1048 int af, len;
1049
1050 if (!host) {
1051 log_err("fio: connect with no host to connect to.\n");
1052 if (td_read(td))
1053 log_err("fio: did you forget to set 'listen'?\n");
1054
1055 td_verror(td, EINVAL, "no hostname= set");
1056 return 1;
1057 }
1058
1059 nd->addr.sin_family = AF_INET;
1060 nd->addr.sin_port = htons(port);
1061 nd->addr6.sin6_family = AF_INET6;
1062 nd->addr6.sin6_port = htons(port);
1063
1064 if (is_ipv6(o)) {
1065 af = AF_INET6;
1066 dst = &nd->addr6.sin6_addr;
1067 } else {
1068 af = AF_INET;
1069 dst = &nd->addr.sin_addr;
1070 }
1071
1072 if (fio_fill_addr(td, host, af, dst, &res))
1073 return 1;
1074
1075 if (!res)
1076 return 0;
1077
1078 if (is_ipv6(o)) {
1079 len = sizeof(nd->addr6.sin6_addr);
1080 src = &((struct sockaddr_in6 *) res->ai_addr)->sin6_addr;
1081 } else {
1082 len = sizeof(nd->addr.sin_addr);
1083 src = &((struct sockaddr_in *) res->ai_addr)->sin_addr;
1084 }
1085
1086 memcpy(dst, src, len);
1087 freeaddrinfo(res);
1088 return 0;
1089}
1090
1091static int fio_netio_setup_connect_unix(struct thread_data *td,
1092 const char *path)
1093{
1094 struct netio_data *nd = td->io_ops->data;
1095 struct sockaddr_un *soun = &nd->addr_un;
1096
1097 soun->sun_family = AF_UNIX;
1098 memset(soun->sun_path, 0, sizeof(soun->sun_path));
1099 strncpy(soun->sun_path, path, sizeof(soun->sun_path) - 1);
1100 return 0;
1101}
1102
1103static int fio_netio_setup_connect(struct thread_data *td)
1104{
1105 struct netio_options *o = td->eo;
1106
1107 if (is_udp(o) || is_tcp(o))
1108 return fio_netio_setup_connect_inet(td, td->o.filename,o->port);
1109 else
1110 return fio_netio_setup_connect_unix(td, td->o.filename);
1111}
1112
1113static int fio_netio_setup_listen_unix(struct thread_data *td, const char *path)
1114{
1115 struct netio_data *nd = td->io_ops->data;
1116 struct sockaddr_un *addr = &nd->addr_un;
1117 mode_t mode;
1118 int len, fd;
1119
1120 fd = socket(AF_UNIX, SOCK_STREAM, 0);
1121 if (fd < 0) {
1122 log_err("fio: socket: %s\n", strerror(errno));
1123 return -1;
1124 }
1125
1126 mode = umask(000);
1127
1128 memset(addr, 0, sizeof(*addr));
1129 addr->sun_family = AF_UNIX;
1130 strncpy(addr->sun_path, path, sizeof(addr->sun_path) - 1);
1131 unlink(path);
1132
1133 len = sizeof(addr->sun_family) + strlen(path) + 1;
1134
1135 if (bind(fd, (struct sockaddr *) addr, len) < 0) {
1136 log_err("fio: bind: %s\n", strerror(errno));
1137 close(fd);
1138 return -1;
1139 }
1140
1141 umask(mode);
1142 nd->listenfd = fd;
1143 return 0;
1144}
1145
1146static int fio_netio_setup_listen_inet(struct thread_data *td, short port)
1147{
1148 struct netio_data *nd = td->io_ops->data;
1149 struct netio_options *o = td->eo;
1150 struct ip_mreq mr;
1151 struct sockaddr_in sin;
1152 struct sockaddr *saddr;
1153 int fd, opt, type, domain;
1154 socklen_t len;
1155
1156 memset(&sin, 0, sizeof(sin));
1157
1158 if (o->proto == FIO_TYPE_TCP) {
1159 type = SOCK_STREAM;
1160 domain = AF_INET;
1161 } else if (o->proto == FIO_TYPE_TCP_V6) {
1162 type = SOCK_STREAM;
1163 domain = AF_INET6;
1164 } else if (o->proto == FIO_TYPE_UDP) {
1165 type = SOCK_DGRAM;
1166 domain = AF_INET;
1167 } else if (o->proto == FIO_TYPE_UDP_V6) {
1168 type = SOCK_DGRAM;
1169 domain = AF_INET6;
1170 } else {
1171 log_err("fio: unknown proto %d\n", o->proto);
1172 return 1;
1173 }
1174
1175 fd = socket(domain, type, 0);
1176 if (fd < 0) {
1177 td_verror(td, errno, "socket");
1178 return 1;
1179 }
1180
1181 opt = 1;
1182 if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, (void *) &opt, sizeof(opt)) < 0) {
1183 td_verror(td, errno, "setsockopt");
1184 close(fd);
1185 return 1;
1186 }
1187#ifdef SO_REUSEPORT
1188 if (setsockopt(fd, SOL_SOCKET, SO_REUSEPORT, (void *) &opt, sizeof(opt)) < 0) {
1189 td_verror(td, errno, "setsockopt");
1190 close(fd);
1191 return 1;
1192 }
1193#endif
1194
1195 if (set_window_size(td, fd)) {
1196 close(fd);
1197 return 1;
1198 }
1199 if (set_mss(td, fd)) {
1200 close(fd);
1201 return 1;
1202 }
1203
1204 if (td->o.filename) {
1205 if (!is_udp(o) || !fio_netio_is_multicast(td->o.filename)) {
1206 log_err("fio: hostname not valid for non-multicast inbound network IO\n");
1207 close(fd);
1208 return 1;
1209 }
1210 if (is_ipv6(o)) {
1211 log_err("fio: IPv6 not supported for multicast network IO");
1212 close(fd);
1213 return 1;
1214 }
1215
1216 inet_aton(td->o.filename, &sin.sin_addr);
1217
1218 mr.imr_multiaddr = sin.sin_addr;
1219 if (o->intfc) {
1220 if (inet_aton(o->intfc, &mr.imr_interface) == 0) {
1221 log_err("fio: interface not valid interface IP\n");
1222 close(fd);
1223 return 1;
1224 }
1225 } else {
1226 mr.imr_interface.s_addr = htonl(INADDR_ANY);
1227 }
1228
1229 if (setsockopt(fd, IPPROTO_IP, IP_ADD_MEMBERSHIP, (const char*)&mr, sizeof(mr)) < 0) {
1230 td_verror(td, errno, "setsockopt IP_ADD_MEMBERSHIP");
1231 close(fd);
1232 return 1;
1233 }
1234 }
1235
1236 if (!is_ipv6(o)) {
1237 saddr = (struct sockaddr *) &nd->addr;
1238 len = sizeof(nd->addr);
1239
1240 nd->addr.sin_family = AF_INET;
1241 nd->addr.sin_addr.s_addr = sin.sin_addr.s_addr ? sin.sin_addr.s_addr : htonl(INADDR_ANY);
1242 nd->addr.sin_port = htons(port);
1243 } else {
1244 saddr = (struct sockaddr *) &nd->addr6;
1245 len = sizeof(nd->addr6);
1246
1247 nd->addr6.sin6_family = AF_INET6;
1248 nd->addr6.sin6_addr = in6addr_any;
1249 nd->addr6.sin6_port = htons(port);
1250 }
1251
1252 if (bind(fd, saddr, len) < 0) {
1253 close(fd);
1254 td_verror(td, errno, "bind");
1255 return 1;
1256 }
1257
1258 nd->listenfd = fd;
1259 return 0;
1260}
1261
1262static int fio_netio_setup_listen(struct thread_data *td)
1263{
1264 struct netio_data *nd = td->io_ops->data;
1265 struct netio_options *o = td->eo;
1266 int ret;
1267
1268 if (is_udp(o) || is_tcp(o))
1269 ret = fio_netio_setup_listen_inet(td, o->port);
1270 else
1271 ret = fio_netio_setup_listen_unix(td, td->o.filename);
1272
1273 if (ret)
1274 return ret;
1275 if (is_udp(o))
1276 return 0;
1277
1278 if (listen(nd->listenfd, 10) < 0) {
1279 td_verror(td, errno, "listen");
1280 nd->listenfd = -1;
1281 return 1;
1282 }
1283
1284 return 0;
1285}
1286
1287static int fio_netio_init(struct thread_data *td)
1288{
1289 struct netio_options *o = td->eo;
1290 int ret;
1291
1292#ifdef WIN32
1293 WSADATA wsd;
1294 WSAStartup(MAKEWORD(2,2), &wsd);
1295#endif
1296
1297 if (td_random(td)) {
1298 log_err("fio: network IO can't be random\n");
1299 return 1;
1300 }
1301
1302 if (o->proto == FIO_TYPE_UNIX && o->port) {
1303 log_err("fio: network IO port not valid with unix socket\n");
1304 return 1;
1305 } else if (o->proto != FIO_TYPE_UNIX && !o->port) {
1306 log_err("fio: network IO requires port for tcp or udp\n");
1307 return 1;
1308 }
1309
1310 if (!is_tcp(o)) {
1311 if (o->listen) {
1312 log_err("fio: listen only valid for TCP proto IO\n");
1313 return 1;
1314 }
1315 if (td_rw(td)) {
1316 log_err("fio: datagram network connections must be"
1317 " read OR write\n");
1318 return 1;
1319 }
1320 if (o->proto == FIO_TYPE_UNIX && !td->o.filename) {
1321 log_err("fio: UNIX sockets need host/filename\n");
1322 return 1;
1323 }
1324 o->listen = td_read(td);
1325 }
1326
1327 if (o->listen)
1328 ret = fio_netio_setup_listen(td);
1329 else
1330 ret = fio_netio_setup_connect(td);
1331
1332 return ret;
1333}
1334
1335static void fio_netio_cleanup(struct thread_data *td)
1336{
1337 struct netio_data *nd = td->io_ops->data;
1338
1339 if (nd) {
1340 if (nd->listenfd != -1)
1341 close(nd->listenfd);
1342 if (nd->pipes[0] != -1)
1343 close(nd->pipes[0]);
1344 if (nd->pipes[1] != -1)
1345 close(nd->pipes[1]);
1346
1347 free(nd);
1348 }
1349}
1350
1351static int fio_netio_setup(struct thread_data *td)
1352{
1353 struct netio_data *nd;
1354
1355 if (!td->files_index) {
1356 add_file(td, td->o.filename ?: "net", 0, 0);
1357 td->o.nr_files = td->o.nr_files ?: 1;
1358 td->o.open_files++;
1359 }
1360
1361 if (!td->io_ops->data) {
1362 nd = malloc(sizeof(*nd));;
1363
1364 memset(nd, 0, sizeof(*nd));
1365 nd->listenfd = -1;
1366 nd->pipes[0] = nd->pipes[1] = -1;
1367 td->io_ops->data = nd;
1368 }
1369
1370 return 0;
1371}
1372
1373static void fio_netio_terminate(struct thread_data *td)
1374{
1375 kill(td->pid, SIGTERM);
1376}
1377
1378#ifdef CONFIG_LINUX_SPLICE
1379static int fio_netio_setup_splice(struct thread_data *td)
1380{
1381 struct netio_data *nd;
1382
1383 fio_netio_setup(td);
1384
1385 nd = td->io_ops->data;
1386 if (nd) {
1387 if (pipe(nd->pipes) < 0)
1388 return 1;
1389
1390 nd->use_splice = 1;
1391 return 0;
1392 }
1393
1394 return 1;
1395}
1396
1397static struct ioengine_ops ioengine_splice = {
1398 .name = "netsplice",
1399 .version = FIO_IOOPS_VERSION,
1400 .prep = fio_netio_prep,
1401 .queue = fio_netio_queue,
1402 .setup = fio_netio_setup_splice,
1403 .init = fio_netio_init,
1404 .cleanup = fio_netio_cleanup,
1405 .open_file = fio_netio_open_file,
1406 .close_file = fio_netio_close_file,
1407 .terminate = fio_netio_terminate,
1408 .options = options,
1409 .option_struct_size = sizeof(struct netio_options),
1410 .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR |
1411 FIO_PIPEIO,
1412};
1413#endif
1414
1415static struct ioengine_ops ioengine_rw = {
1416 .name = "net",
1417 .version = FIO_IOOPS_VERSION,
1418 .prep = fio_netio_prep,
1419 .queue = fio_netio_queue,
1420 .setup = fio_netio_setup,
1421 .init = fio_netio_init,
1422 .cleanup = fio_netio_cleanup,
1423 .open_file = fio_netio_open_file,
1424 .close_file = fio_netio_close_file,
1425 .terminate = fio_netio_terminate,
1426 .options = options,
1427 .option_struct_size = sizeof(struct netio_options),
1428 .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_UNIDIR |
1429 FIO_PIPEIO | FIO_BIT_BASED,
1430};
1431
1432static int str_hostname_cb(void *data, const char *input)
1433{
1434 struct netio_options *o = data;
1435
1436 if (o->td->o.filename)
1437 free(o->td->o.filename);
1438 o->td->o.filename = strdup(input);
1439 return 0;
1440}
1441
1442static void fio_init fio_netio_register(void)
1443{
1444 register_ioengine(&ioengine_rw);
1445#ifdef CONFIG_LINUX_SPLICE
1446 register_ioengine(&ioengine_splice);
1447#endif
1448}
1449
1450static void fio_exit fio_netio_unregister(void)
1451{
1452 unregister_ioengine(&ioengine_rw);
1453#ifdef CONFIG_LINUX_SPLICE
1454 unregister_ioengine(&ioengine_splice);
1455#endif
1456}
fio - Flexible IO Tester