1#include <stdio.h>
2#include <stdlib.h>
3#include <stdarg.h>
4#include <unistd.h>
5#include <limits.h>
6#include <errno.h>
7#include <sys/poll.h>
8#include <sys/types.h>
9#include <sys/wait.h>
10#include <sys/socket.h>
11#include <sys/stat.h>
12#include <sys/un.h>
13#include <sys/uio.h>
14#include <netinet/in.h>
15#include <arpa/inet.h>
16#include <netdb.h>
17#include <syslog.h>
18#include <signal.h>
19#ifdef CONFIG_ZLIB
20#include <zlib.h>
21#endif
22
23#include "fio.h"
24#include "options.h"
25#include "server.h"
26#include "crc/crc16.h"
27#include "lib/ieee754.h"
28#include "verify.h"
29#include "smalloc.h"
30
31int fio_net_port = FIO_NET_PORT;
32
33int exit_backend = 0;
34
35enum {
36	SK_F_FREE	= 1,
37	SK_F_COPY	= 2,
38	SK_F_SIMPLE	= 4,
39	SK_F_VEC	= 8,
40	SK_F_INLINE	= 16,
41};
42
43struct sk_entry {
44	struct flist_head list;	/* link on sk_out->list */
45	int flags;		/* SK_F_* */
46	int opcode;		/* Actual command fields */
47	void *buf;
48	off_t size;
49	uint64_t tag;
50	struct flist_head next;	/* Other sk_entry's, if linked command */
51};
52
53static char *fio_server_arg;
54static char *bind_sock;
55static struct sockaddr_in saddr_in;
56static struct sockaddr_in6 saddr_in6;
57static int use_ipv6;
58#ifdef CONFIG_ZLIB
59static unsigned int has_zlib = 1;
60#else
61static unsigned int has_zlib = 0;
62#endif
63static unsigned int use_zlib;
64static char me[128];
65
66static pthread_key_t sk_out_key;
67
68struct fio_fork_item {
69	struct flist_head list;
70	int exitval;
71	int signal;
72	int exited;
73	pid_t pid;
74};
75
76struct cmd_reply {
77	struct fio_mutex lock;
78	void *data;
79	size_t size;
80	int error;
81};
82
83static const char *fio_server_ops[FIO_NET_CMD_NR] = {
84	"",
85	"QUIT",
86	"EXIT",
87	"JOB",
88	"JOBLINE",
89	"TEXT",
90	"TS",
91	"GS",
92	"SEND_ETA",
93	"ETA",
94	"PROBE",
95	"START",
96	"STOP",
97	"DISK_UTIL",
98	"SERVER_START",
99	"ADD_JOB",
100	"RUN",
101	"IOLOG",
102	"UPDATE_JOB",
103	"LOAD_FILE",
104	"VTRIGGER",
105	"SENDFILE",
106	"JOB_OPT",
107};
108
109static void sk_lock(struct sk_out *sk_out)
110{
111	fio_mutex_down(&sk_out->lock);
112}
113
114static void sk_unlock(struct sk_out *sk_out)
115{
116	fio_mutex_up(&sk_out->lock);
117}
118
119void sk_out_assign(struct sk_out *sk_out)
120{
121	if (!sk_out)
122		return;
123
124	sk_lock(sk_out);
125	sk_out->refs++;
126	sk_unlock(sk_out);
127	pthread_setspecific(sk_out_key, sk_out);
128}
129
130static void sk_out_free(struct sk_out *sk_out)
131{
132	__fio_mutex_remove(&sk_out->lock);
133	__fio_mutex_remove(&sk_out->wait);
134	__fio_mutex_remove(&sk_out->xmit);
135	sfree(sk_out);
136}
137
138static int __sk_out_drop(struct sk_out *sk_out)
139{
140	if (sk_out) {
141		int refs;
142
143		sk_lock(sk_out);
144		assert(sk_out->refs != 0);
145		refs = --sk_out->refs;
146		sk_unlock(sk_out);
147
148		if (!refs) {
149			sk_out_free(sk_out);
150			pthread_setspecific(sk_out_key, NULL);
151			return 0;
152		}
153	}
154
155	return 1;
156}
157
158void sk_out_drop(void)
159{
160	struct sk_out *sk_out;
161
162	sk_out = pthread_getspecific(sk_out_key);
163	__sk_out_drop(sk_out);
164}
165
166static void __fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
167			       uint32_t pdu_len, uint64_t tag)
168{
169	memset(cmd, 0, sizeof(*cmd));
170
171	cmd->version	= __cpu_to_le16(FIO_SERVER_VER);
172	cmd->opcode	= cpu_to_le16(opcode);
173	cmd->tag	= cpu_to_le64(tag);
174	cmd->pdu_len	= cpu_to_le32(pdu_len);
175}
176
177
178static void fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
179			     const void *pdu, uint32_t pdu_len, uint64_t tag)
180{
181	__fio_init_net_cmd(cmd, opcode, pdu_len, tag);
182
183	if (pdu)
184		memcpy(&cmd->payload, pdu, pdu_len);
185}
186
187const char *fio_server_op(unsigned int op)
188{
189	static char buf[32];
190
191	if (op < FIO_NET_CMD_NR)
192		return fio_server_ops[op];
193
194	sprintf(buf, "UNKNOWN/%d", op);
195	return buf;
196}
197
198static ssize_t iov_total_len(const struct iovec *iov, int count)
199{
200	ssize_t ret = 0;
201
202	while (count--) {
203		ret += iov->iov_len;
204		iov++;
205	}
206
207	return ret;
208}
209
210static int fio_sendv_data(int sk, struct iovec *iov, int count)
211{
212	ssize_t total_len = iov_total_len(iov, count);
213	ssize_t ret;
214
215	do {
216		ret = writev(sk, iov, count);
217		if (ret > 0) {
218			total_len -= ret;
219			if (!total_len)
220				break;
221
222			while (ret) {
223				if (ret >= iov->iov_len) {
224					ret -= iov->iov_len;
225					iov++;
226					continue;
227				}
228				iov->iov_base += ret;
229				iov->iov_len -= ret;
230				ret = 0;
231			}
232		} else if (!ret)
233			break;
234		else if (errno == EAGAIN || errno == EINTR)
235			continue;
236		else
237			break;
238	} while (!exit_backend);
239
240	if (!total_len)
241		return 0;
242
243	return 1;
244}
245
246static int fio_send_data(int sk, const void *p, unsigned int len)
247{
248	struct iovec iov = { .iov_base = (void *) p, .iov_len = len };
249
250	assert(len <= sizeof(struct fio_net_cmd) + FIO_SERVER_MAX_FRAGMENT_PDU);
251
252	return fio_sendv_data(sk, &iov, 1);
253}
254
255static int fio_recv_data(int sk, void *p, unsigned int len, bool wait)
256{
257	int flags;
258
259	if (wait)
260		flags = MSG_WAITALL;
261	else
262		flags = OS_MSG_DONTWAIT;
263
264	do {
265		int ret = recv(sk, p, len, flags);
266
267		if (ret > 0) {
268			len -= ret;
269			if (!len)
270				break;
271			p += ret;
272			continue;
273		} else if (!ret)
274			break;
275		else if (errno == EAGAIN || errno == EINTR) {
276			if (wait)
277				continue;
278			break;
279		} else
280			break;
281	} while (!exit_backend);
282
283	if (!len)
284		return 0;
285
286	return -1;
287}
288
289static int verify_convert_cmd(struct fio_net_cmd *cmd)
290{
291	uint16_t crc;
292
293	cmd->cmd_crc16 = le16_to_cpu(cmd->cmd_crc16);
294	cmd->pdu_crc16 = le16_to_cpu(cmd->pdu_crc16);
295
296	crc = fio_crc16(cmd, FIO_NET_CMD_CRC_SZ);
297	if (crc != cmd->cmd_crc16) {
298		log_err("fio: server bad crc on command (got %x, wanted %x)\n",
299				cmd->cmd_crc16, crc);
300		return 1;
301	}
302
303	cmd->version	= le16_to_cpu(cmd->version);
304	cmd->opcode	= le16_to_cpu(cmd->opcode);
305	cmd->flags	= le32_to_cpu(cmd->flags);
306	cmd->tag	= le64_to_cpu(cmd->tag);
307	cmd->pdu_len	= le32_to_cpu(cmd->pdu_len);
308
309	switch (cmd->version) {
310	case FIO_SERVER_VER:
311		break;
312	default:
313		log_err("fio: bad server cmd version %d\n", cmd->version);
314		return 1;
315	}
316
317	if (cmd->pdu_len > FIO_SERVER_MAX_FRAGMENT_PDU) {
318		log_err("fio: command payload too large: %u\n", cmd->pdu_len);
319		return 1;
320	}
321
322	return 0;
323}
324
325/*
326 * Read (and defragment, if necessary) incoming commands
327 */
328struct fio_net_cmd *fio_net_recv_cmd(int sk, bool wait)
329{
330	struct fio_net_cmd cmd, *tmp, *cmdret = NULL;
331	size_t cmd_size = 0, pdu_offset = 0;
332	uint16_t crc;
333	int ret, first = 1;
334	void *pdu = NULL;
335
336	do {
337		ret = fio_recv_data(sk, &cmd, sizeof(cmd), wait);
338		if (ret)
339			break;
340
341		/* We have a command, verify it and swap if need be */
342		ret = verify_convert_cmd(&cmd);
343		if (ret)
344			break;
345
346		if (first) {
347			/* if this is text, add room for \0 at the end */
348			cmd_size = sizeof(cmd) + cmd.pdu_len + 1;
349			assert(!cmdret);
350		} else
351			cmd_size += cmd.pdu_len;
352
353		if (cmd_size / 1024 > FIO_SERVER_MAX_CMD_MB * 1024) {
354			log_err("fio: cmd+pdu too large (%llu)\n", (unsigned long long) cmd_size);
355			ret = 1;
356			break;
357		}
358
359		tmp = realloc(cmdret, cmd_size);
360		if (!tmp) {
361			log_err("fio: server failed allocating cmd\n");
362			ret = 1;
363			break;
364		}
365		cmdret = tmp;
366
367		if (first)
368			memcpy(cmdret, &cmd, sizeof(cmd));
369		else if (cmdret->opcode != cmd.opcode) {
370			log_err("fio: fragment opcode mismatch (%d != %d)\n",
371					cmdret->opcode, cmd.opcode);
372			ret = 1;
373			break;
374		}
375
376		if (!cmd.pdu_len)
377			break;
378
379		/* There's payload, get it */
380		pdu = (void *) cmdret->payload + pdu_offset;
381		ret = fio_recv_data(sk, pdu, cmd.pdu_len, wait);
382		if (ret)
383			break;
384
385		/* Verify payload crc */
386		crc = fio_crc16(pdu, cmd.pdu_len);
387		if (crc != cmd.pdu_crc16) {
388			log_err("fio: server bad crc on payload ");
389			log_err("(got %x, wanted %x)\n", cmd.pdu_crc16, crc);
390			ret = 1;
391			break;
392		}
393
394		pdu_offset += cmd.pdu_len;
395		if (!first)
396			cmdret->pdu_len += cmd.pdu_len;
397		first = 0;
398	} while (cmd.flags & FIO_NET_CMD_F_MORE);
399
400	if (ret) {
401		free(cmdret);
402		cmdret = NULL;
403	} else if (cmdret) {
404		/* zero-terminate text input */
405		if (cmdret->pdu_len) {
406			if (cmdret->opcode == FIO_NET_CMD_TEXT) {
407				struct cmd_text_pdu *__pdu = (struct cmd_text_pdu *) cmdret->payload;
408				char *buf = (char *) __pdu->buf;
409
410				buf[__pdu->buf_len] = '\0';
411			} else if (cmdret->opcode == FIO_NET_CMD_JOB) {
412				struct cmd_job_pdu *__pdu = (struct cmd_job_pdu *) cmdret->payload;
413				char *buf = (char *) __pdu->buf;
414				int len = le32_to_cpu(__pdu->buf_len);
415
416				buf[len] = '\0';
417			}
418		}
419
420		/* frag flag is internal */
421		cmdret->flags &= ~FIO_NET_CMD_F_MORE;
422	}
423
424	return cmdret;
425}
426
427static void add_reply(uint64_t tag, struct flist_head *list)
428{
429	struct fio_net_cmd_reply *reply;
430
431	reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
432	flist_add_tail(&reply->list, list);
433}
434
435static uint64_t alloc_reply(uint64_t tag, uint16_t opcode)
436{
437	struct fio_net_cmd_reply *reply;
438
439	reply = calloc(1, sizeof(*reply));
440	INIT_FLIST_HEAD(&reply->list);
441	fio_gettime(&reply->tv, NULL);
442	reply->saved_tag = tag;
443	reply->opcode = opcode;
444
445	return (uintptr_t) reply;
446}
447
448static void free_reply(uint64_t tag)
449{
450	struct fio_net_cmd_reply *reply;
451
452	reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
453	free(reply);
454}
455
456static void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
457{
458	uint32_t pdu_len;
459
460	cmd->cmd_crc16 = __cpu_to_le16(fio_crc16(cmd, FIO_NET_CMD_CRC_SZ));
461
462	pdu_len = le32_to_cpu(cmd->pdu_len);
463	cmd->pdu_crc16 = __cpu_to_le16(fio_crc16(pdu, pdu_len));
464}
465
466static void fio_net_cmd_crc(struct fio_net_cmd *cmd)
467{
468	fio_net_cmd_crc_pdu(cmd, cmd->payload);
469}
470
471int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
472		     uint64_t *tagptr, struct flist_head *list)
473{
474	struct fio_net_cmd *cmd = NULL;
475	size_t this_len, cur_len = 0;
476	uint64_t tag;
477	int ret;
478
479	if (list) {
480		assert(tagptr);
481		tag = *tagptr = alloc_reply(*tagptr, opcode);
482	} else
483		tag = tagptr ? *tagptr : 0;
484
485	do {
486		this_len = size;
487		if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
488			this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
489
490		if (!cmd || cur_len < sizeof(*cmd) + this_len) {
491			if (cmd)
492				free(cmd);
493
494			cur_len = sizeof(*cmd) + this_len;
495			cmd = malloc(cur_len);
496		}
497
498		fio_init_net_cmd(cmd, opcode, buf, this_len, tag);
499
500		if (this_len < size)
501			cmd->flags = __cpu_to_le32(FIO_NET_CMD_F_MORE);
502
503		fio_net_cmd_crc(cmd);
504
505		ret = fio_send_data(fd, cmd, sizeof(*cmd) + this_len);
506		size -= this_len;
507		buf += this_len;
508	} while (!ret && size);
509
510	if (list) {
511		if (ret)
512			free_reply(tag);
513		else
514			add_reply(tag, list);
515	}
516
517	if (cmd)
518		free(cmd);
519
520	return ret;
521}
522
523static struct sk_entry *fio_net_prep_cmd(uint16_t opcode, void *buf,
524					 size_t size, uint64_t *tagptr,
525					 int flags)
526{
527	struct sk_entry *entry;
528
529	entry = smalloc(sizeof(*entry));
530	INIT_FLIST_HEAD(&entry->next);
531	entry->opcode = opcode;
532	if (flags & SK_F_COPY) {
533		entry->buf = smalloc(size);
534		memcpy(entry->buf, buf, size);
535	} else
536		entry->buf = buf;
537
538	entry->size = size;
539	if (tagptr)
540		entry->tag = *tagptr;
541	else
542		entry->tag = 0;
543	entry->flags = flags;
544	return entry;
545}
546
547static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry);
548
549static void fio_net_queue_entry(struct sk_entry *entry)
550{
551	struct sk_out *sk_out = pthread_getspecific(sk_out_key);
552
553	if (entry->flags & SK_F_INLINE)
554		handle_sk_entry(sk_out, entry);
555	else {
556		sk_lock(sk_out);
557		flist_add_tail(&entry->list, &sk_out->list);
558		sk_unlock(sk_out);
559
560		fio_mutex_up(&sk_out->wait);
561	}
562}
563
564static int fio_net_queue_cmd(uint16_t opcode, void *buf, off_t size,
565			     uint64_t *tagptr, int flags)
566{
567	struct sk_entry *entry;
568
569	entry = fio_net_prep_cmd(opcode, buf, size, tagptr, flags);
570	if (entry) {
571		fio_net_queue_entry(entry);
572		return 0;
573	}
574
575	return 1;
576}
577
578static int fio_net_send_simple_stack_cmd(int sk, uint16_t opcode, uint64_t tag)
579{
580	struct fio_net_cmd cmd;
581
582	fio_init_net_cmd(&cmd, opcode, NULL, 0, tag);
583	fio_net_cmd_crc(&cmd);
584
585	return fio_send_data(sk, &cmd, sizeof(cmd));
586}
587
588/*
589 * If 'list' is non-NULL, then allocate and store the sent command for
590 * later verification.
591 */
592int fio_net_send_simple_cmd(int sk, uint16_t opcode, uint64_t tag,
593			    struct flist_head *list)
594{
595	int ret;
596
597	if (list)
598		tag = alloc_reply(tag, opcode);
599
600	ret = fio_net_send_simple_stack_cmd(sk, opcode, tag);
601	if (ret) {
602		if (list)
603			free_reply(tag);
604
605		return ret;
606	}
607
608	if (list)
609		add_reply(tag, list);
610
611	return 0;
612}
613
614static int fio_net_queue_quit(void)
615{
616	dprint(FD_NET, "server: sending quit\n");
617
618	return fio_net_queue_cmd(FIO_NET_CMD_QUIT, NULL, 0, NULL, SK_F_SIMPLE);
619}
620
621int fio_net_send_quit(int sk)
622{
623	dprint(FD_NET, "server: sending quit\n");
624
625	return fio_net_send_simple_cmd(sk, FIO_NET_CMD_QUIT, 0, NULL);
626}
627
628static int fio_net_send_ack(struct fio_net_cmd *cmd, int error, int signal)
629{
630	struct cmd_end_pdu epdu;
631	uint64_t tag = 0;
632
633	if (cmd)
634		tag = cmd->tag;
635
636	epdu.error = __cpu_to_le32(error);
637	epdu.signal = __cpu_to_le32(signal);
638	return fio_net_queue_cmd(FIO_NET_CMD_STOP, &epdu, sizeof(epdu), &tag, SK_F_COPY);
639}
640
641static int fio_net_queue_stop(int error, int signal)
642{
643	dprint(FD_NET, "server: sending stop (%d, %d)\n", error, signal);
644	return fio_net_send_ack(NULL, error, signal);
645}
646
647static void fio_server_add_fork_item(pid_t pid, struct flist_head *list)
648{
649	struct fio_fork_item *ffi;
650
651	ffi = malloc(sizeof(*ffi));
652	ffi->exitval = 0;
653	ffi->signal = 0;
654	ffi->exited = 0;
655	ffi->pid = pid;
656	flist_add_tail(&ffi->list, list);
657}
658
659static void fio_server_add_conn_pid(struct flist_head *conn_list, pid_t pid)
660{
661	dprint(FD_NET, "server: forked off connection job (pid=%u)\n", (int) pid);
662	fio_server_add_fork_item(pid, conn_list);
663}
664
665static void fio_server_add_job_pid(struct flist_head *job_list, pid_t pid)
666{
667	dprint(FD_NET, "server: forked off job job (pid=%u)\n", (int) pid);
668	fio_server_add_fork_item(pid, job_list);
669}
670
671static void fio_server_check_fork_item(struct fio_fork_item *ffi)
672{
673	int ret, status;
674
675	ret = waitpid(ffi->pid, &status, WNOHANG);
676	if (ret < 0) {
677		if (errno == ECHILD) {
678			log_err("fio: connection pid %u disappeared\n", (int) ffi->pid);
679			ffi->exited = 1;
680		} else
681			log_err("fio: waitpid: %s\n", strerror(errno));
682	} else if (ret == ffi->pid) {
683		if (WIFSIGNALED(status)) {
684			ffi->signal = WTERMSIG(status);
685			ffi->exited = 1;
686		}
687		if (WIFEXITED(status)) {
688			if (WEXITSTATUS(status))
689				ffi->exitval = WEXITSTATUS(status);
690			ffi->exited = 1;
691		}
692	}
693}
694
695static void fio_server_fork_item_done(struct fio_fork_item *ffi, bool stop)
696{
697	dprint(FD_NET, "pid %u exited, sig=%u, exitval=%d\n", (int) ffi->pid, ffi->signal, ffi->exitval);
698
699	/*
700	 * Fold STOP and QUIT...
701	 */
702	if (stop) {
703		fio_net_queue_stop(ffi->exitval, ffi->signal);
704		fio_net_queue_quit();
705	}
706
707	flist_del(&ffi->list);
708	free(ffi);
709}
710
711static void fio_server_check_fork_items(struct flist_head *list, bool stop)
712{
713	struct flist_head *entry, *tmp;
714	struct fio_fork_item *ffi;
715
716	flist_for_each_safe(entry, tmp, list) {
717		ffi = flist_entry(entry, struct fio_fork_item, list);
718
719		fio_server_check_fork_item(ffi);
720
721		if (ffi->exited)
722			fio_server_fork_item_done(ffi, stop);
723	}
724}
725
726static void fio_server_check_jobs(struct flist_head *job_list)
727{
728	fio_server_check_fork_items(job_list, true);
729}
730
731static void fio_server_check_conns(struct flist_head *conn_list)
732{
733	fio_server_check_fork_items(conn_list, false);
734}
735
736static int handle_load_file_cmd(struct fio_net_cmd *cmd)
737{
738	struct cmd_load_file_pdu *pdu = (struct cmd_load_file_pdu *) cmd->payload;
739	void *file_name = pdu->file;
740	struct cmd_start_pdu spdu;
741
742	dprint(FD_NET, "server: loading local file %s\n", (char *) file_name);
743
744	pdu->name_len = le16_to_cpu(pdu->name_len);
745	pdu->client_type = le16_to_cpu(pdu->client_type);
746
747	if (parse_jobs_ini(file_name, 0, 0, pdu->client_type)) {
748		fio_net_queue_quit();
749		return -1;
750	}
751
752	spdu.jobs = cpu_to_le32(thread_number);
753	spdu.stat_outputs = cpu_to_le32(stat_number);
754	fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
755	return 0;
756}
757
758static int handle_run_cmd(struct sk_out *sk_out, struct flist_head *job_list,
759			  struct fio_net_cmd *cmd)
760{
761	pid_t pid;
762	int ret;
763
764	sk_out_assign(sk_out);
765
766	fio_time_init();
767	set_genesis_time();
768
769	pid = fork();
770	if (pid) {
771		fio_server_add_job_pid(job_list, pid);
772		return 0;
773	}
774
775	ret = fio_backend(sk_out);
776	free_threads_shm();
777	sk_out_drop();
778	_exit(ret);
779}
780
781static int handle_job_cmd(struct fio_net_cmd *cmd)
782{
783	struct cmd_job_pdu *pdu = (struct cmd_job_pdu *) cmd->payload;
784	void *buf = pdu->buf;
785	struct cmd_start_pdu spdu;
786
787	pdu->buf_len = le32_to_cpu(pdu->buf_len);
788	pdu->client_type = le32_to_cpu(pdu->client_type);
789
790	if (parse_jobs_ini(buf, 1, 0, pdu->client_type)) {
791		fio_net_queue_quit();
792		return -1;
793	}
794
795	spdu.jobs = cpu_to_le32(thread_number);
796	spdu.stat_outputs = cpu_to_le32(stat_number);
797
798	fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
799	return 0;
800}
801
802static int handle_jobline_cmd(struct fio_net_cmd *cmd)
803{
804	void *pdu = cmd->payload;
805	struct cmd_single_line_pdu *cslp;
806	struct cmd_line_pdu *clp;
807	unsigned long offset;
808	struct cmd_start_pdu spdu;
809	char **argv;
810	int i;
811
812	clp = pdu;
813	clp->lines = le16_to_cpu(clp->lines);
814	clp->client_type = le16_to_cpu(clp->client_type);
815	argv = malloc(clp->lines * sizeof(char *));
816	offset = sizeof(*clp);
817
818	dprint(FD_NET, "server: %d command line args\n", clp->lines);
819
820	for (i = 0; i < clp->lines; i++) {
821		cslp = pdu + offset;
822		argv[i] = (char *) cslp->text;
823
824		offset += sizeof(*cslp) + le16_to_cpu(cslp->len);
825		dprint(FD_NET, "server: %d: %s\n", i, argv[i]);
826	}
827
828	if (parse_cmd_line(clp->lines, argv, clp->client_type)) {
829		fio_net_queue_quit();
830		free(argv);
831		return -1;
832	}
833
834	free(argv);
835
836	spdu.jobs = cpu_to_le32(thread_number);
837	spdu.stat_outputs = cpu_to_le32(stat_number);
838
839	fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
840	return 0;
841}
842
843static int handle_probe_cmd(struct fio_net_cmd *cmd)
844{
845	struct cmd_client_probe_pdu *pdu = (struct cmd_client_probe_pdu *) cmd->payload;
846	struct cmd_probe_reply_pdu probe;
847	uint64_t tag = cmd->tag;
848
849	dprint(FD_NET, "server: sending probe reply\n");
850
851	strcpy(me, (char *) pdu->server);
852
853	memset(&probe, 0, sizeof(probe));
854	gethostname((char *) probe.hostname, sizeof(probe.hostname));
855#ifdef CONFIG_BIG_ENDIAN
856	probe.bigendian = 1;
857#endif
858	strncpy((char *) probe.fio_version, fio_version_string, sizeof(probe.fio_version));
859
860	probe.os	= FIO_OS;
861	probe.arch	= FIO_ARCH;
862	probe.bpp	= sizeof(void *);
863	probe.cpus	= __cpu_to_le32(cpus_online());
864
865	/*
866	 * If the client supports compression and we do too, then enable it
867	 */
868	if (has_zlib && le64_to_cpu(pdu->flags) & FIO_PROBE_FLAG_ZLIB) {
869		probe.flags = __cpu_to_le64(FIO_PROBE_FLAG_ZLIB);
870		use_zlib = 1;
871	} else {
872		probe.flags = 0;
873		use_zlib = 0;
874	}
875
876	return fio_net_queue_cmd(FIO_NET_CMD_PROBE, &probe, sizeof(probe), &tag, SK_F_COPY);
877}
878
879static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
880{
881	struct jobs_eta *je;
882	uint64_t tag = cmd->tag;
883	size_t size;
884	int i;
885
886	dprint(FD_NET, "server sending status\n");
887
888	/*
889	 * Fake ETA return if we don't have a local one, otherwise the client
890	 * will end up timing out waiting for a response to the ETA request
891	 */
892	je = get_jobs_eta(true, &size);
893	if (!je) {
894		size = sizeof(*je);
895		je = calloc(1, size);
896	} else {
897		je->nr_running		= cpu_to_le32(je->nr_running);
898		je->nr_ramp		= cpu_to_le32(je->nr_ramp);
899		je->nr_pending		= cpu_to_le32(je->nr_pending);
900		je->nr_setting_up	= cpu_to_le32(je->nr_setting_up);
901		je->files_open		= cpu_to_le32(je->files_open);
902
903		for (i = 0; i < DDIR_RWDIR_CNT; i++) {
904			je->m_rate[i]	= cpu_to_le64(je->m_rate[i]);
905			je->t_rate[i]	= cpu_to_le64(je->t_rate[i]);
906			je->m_iops[i]	= cpu_to_le32(je->m_iops[i]);
907			je->t_iops[i]	= cpu_to_le32(je->t_iops[i]);
908			je->rate[i]	= cpu_to_le64(je->rate[i]);
909			je->iops[i]	= cpu_to_le32(je->iops[i]);
910		}
911
912		je->elapsed_sec		= cpu_to_le64(je->elapsed_sec);
913		je->eta_sec		= cpu_to_le64(je->eta_sec);
914		je->nr_threads		= cpu_to_le32(je->nr_threads);
915		je->is_pow2		= cpu_to_le32(je->is_pow2);
916		je->unit_base		= cpu_to_le32(je->unit_base);
917	}
918
919	fio_net_queue_cmd(FIO_NET_CMD_ETA, je, size, &tag, SK_F_FREE);
920	return 0;
921}
922
923static int send_update_job_reply(uint64_t __tag, int error)
924{
925	uint64_t tag = __tag;
926	uint32_t pdu_error;
927
928	pdu_error = __cpu_to_le32(error);
929	return fio_net_queue_cmd(FIO_NET_CMD_UPDATE_JOB, &pdu_error, sizeof(pdu_error), &tag, SK_F_COPY);
930}
931
932static int handle_update_job_cmd(struct fio_net_cmd *cmd)
933{
934	struct cmd_add_job_pdu *pdu = (struct cmd_add_job_pdu *) cmd->payload;
935	struct thread_data *td;
936	uint32_t tnumber;
937
938	tnumber = le32_to_cpu(pdu->thread_number);
939
940	dprint(FD_NET, "server: updating options for job %u\n", tnumber);
941
942	if (!tnumber || tnumber > thread_number) {
943		send_update_job_reply(cmd->tag, ENODEV);
944		return 0;
945	}
946
947	td = &threads[tnumber - 1];
948	convert_thread_options_to_cpu(&td->o, &pdu->top);
949	send_update_job_reply(cmd->tag, 0);
950	return 0;
951}
952
953static int handle_trigger_cmd(struct fio_net_cmd *cmd)
954{
955	struct cmd_vtrigger_pdu *pdu = (struct cmd_vtrigger_pdu *) cmd->payload;
956	char *buf = (char *) pdu->cmd;
957	struct all_io_list *rep;
958	size_t sz;
959
960	pdu->len = le16_to_cpu(pdu->len);
961	buf[pdu->len] = '\0';
962
963	rep = get_all_io_list(IO_LIST_ALL, &sz);
964	if (!rep) {
965		struct all_io_list state;
966
967		state.threads = cpu_to_le64((uint64_t) 0);
968		fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, &state, sizeof(state), NULL, SK_F_COPY | SK_F_INLINE);
969	} else
970		fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, rep, sz, NULL, SK_F_FREE | SK_F_INLINE);
971
972	exec_trigger(buf);
973	return 0;
974}
975
976static int handle_command(struct sk_out *sk_out, struct flist_head *job_list,
977			  struct fio_net_cmd *cmd)
978{
979	int ret;
980
981	dprint(FD_NET, "server: got op [%s], pdu=%u, tag=%llx\n",
982			fio_server_op(cmd->opcode), cmd->pdu_len,
983			(unsigned long long) cmd->tag);
984
985	switch (cmd->opcode) {
986	case FIO_NET_CMD_QUIT:
987		fio_terminate_threads(TERMINATE_ALL);
988		ret = 0;
989		break;
990	case FIO_NET_CMD_EXIT:
991		exit_backend = 1;
992		return -1;
993	case FIO_NET_CMD_LOAD_FILE:
994		ret = handle_load_file_cmd(cmd);
995		break;
996	case FIO_NET_CMD_JOB:
997		ret = handle_job_cmd(cmd);
998		break;
999	case FIO_NET_CMD_JOBLINE:
1000		ret = handle_jobline_cmd(cmd);
1001		break;
1002	case FIO_NET_CMD_PROBE:
1003		ret = handle_probe_cmd(cmd);
1004		break;
1005	case FIO_NET_CMD_SEND_ETA:
1006		ret = handle_send_eta_cmd(cmd);
1007		break;
1008	case FIO_NET_CMD_RUN:
1009		ret = handle_run_cmd(sk_out, job_list, cmd);
1010		break;
1011	case FIO_NET_CMD_UPDATE_JOB:
1012		ret = handle_update_job_cmd(cmd);
1013		break;
1014	case FIO_NET_CMD_VTRIGGER:
1015		ret = handle_trigger_cmd(cmd);
1016		break;
1017	case FIO_NET_CMD_SENDFILE: {
1018		struct cmd_sendfile_reply *in;
1019		struct cmd_reply *rep;
1020
1021		rep = (struct cmd_reply *) (uintptr_t) cmd->tag;
1022
1023		in = (struct cmd_sendfile_reply *) cmd->payload;
1024		in->size = le32_to_cpu(in->size);
1025		in->error = le32_to_cpu(in->error);
1026		if (in->error) {
1027			ret = 1;
1028			rep->error = in->error;
1029		} else {
1030			ret = 0;
1031			rep->data = smalloc(in->size);
1032			if (!rep->data) {
1033				ret = 1;
1034				rep->error = ENOMEM;
1035			} else {
1036				rep->size = in->size;
1037				memcpy(rep->data, in->data, in->size);
1038			}
1039		}
1040		fio_mutex_up(&rep->lock);
1041		break;
1042		}
1043	default:
1044		log_err("fio: unknown opcode: %s\n", fio_server_op(cmd->opcode));
1045		ret = 1;
1046	}
1047
1048	return ret;
1049}
1050
1051/*
1052 * Send a command with a separate PDU, not inlined in the command
1053 */
1054static int fio_send_cmd_ext_pdu(int sk, uint16_t opcode, const void *buf,
1055				off_t size, uint64_t tag, uint32_t flags)
1056{
1057	struct fio_net_cmd cmd;
1058	struct iovec iov[2];
1059	size_t this_len;
1060	int ret;
1061
1062	iov[0].iov_base = (void *) &cmd;
1063	iov[0].iov_len = sizeof(cmd);
1064
1065	do {
1066		uint32_t this_flags = flags;
1067
1068		this_len = size;
1069		if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
1070			this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
1071
1072		if (this_len < size)
1073			this_flags |= FIO_NET_CMD_F_MORE;
1074
1075		__fio_init_net_cmd(&cmd, opcode, this_len, tag);
1076		cmd.flags = __cpu_to_le32(this_flags);
1077		fio_net_cmd_crc_pdu(&cmd, buf);
1078
1079		iov[1].iov_base = (void *) buf;
1080		iov[1].iov_len = this_len;
1081
1082		ret = fio_sendv_data(sk, iov, 2);
1083		size -= this_len;
1084		buf += this_len;
1085	} while (!ret && size);
1086
1087	return ret;
1088}
1089
1090static void finish_entry(struct sk_entry *entry)
1091{
1092	if (entry->flags & SK_F_FREE)
1093		free(entry->buf);
1094	else if (entry->flags & SK_F_COPY)
1095		sfree(entry->buf);
1096
1097	sfree(entry);
1098}
1099
1100static void entry_set_flags(struct sk_entry *entry, struct flist_head *list,
1101			    unsigned int *flags)
1102{
1103	if (!flist_empty(list))
1104		*flags = FIO_NET_CMD_F_MORE;
1105	else
1106		*flags = 0;
1107}
1108
1109static int send_vec_entry(struct sk_out *sk_out, struct sk_entry *first)
1110{
1111	unsigned int flags;
1112	int ret;
1113
1114	entry_set_flags(first, &first->next, &flags);
1115
1116	ret = fio_send_cmd_ext_pdu(sk_out->sk, first->opcode, first->buf,
1117					first->size, first->tag, flags);
1118
1119	while (!flist_empty(&first->next)) {
1120		struct sk_entry *next;
1121
1122		next = flist_first_entry(&first->next, struct sk_entry, list);
1123		flist_del_init(&next->list);
1124
1125		entry_set_flags(next, &first->next, &flags);
1126
1127		ret += fio_send_cmd_ext_pdu(sk_out->sk, next->opcode, next->buf,
1128						next->size, next->tag, flags);
1129		finish_entry(next);
1130	}
1131
1132	return ret;
1133}
1134
1135static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry)
1136{
1137	int ret;
1138
1139	fio_mutex_down(&sk_out->xmit);
1140
1141	if (entry->flags & SK_F_VEC)
1142		ret = send_vec_entry(sk_out, entry);
1143	else if (entry->flags & SK_F_SIMPLE) {
1144		ret = fio_net_send_simple_cmd(sk_out->sk, entry->opcode,
1145						entry->tag, NULL);
1146	} else {
1147		ret = fio_net_send_cmd(sk_out->sk, entry->opcode, entry->buf,
1148					entry->size, &entry->tag, NULL);
1149	}
1150
1151	fio_mutex_up(&sk_out->xmit);
1152
1153	if (ret)
1154		log_err("fio: failed handling cmd %s\n", fio_server_op(entry->opcode));
1155
1156	finish_entry(entry);
1157	return ret;
1158}
1159
1160static int handle_xmits(struct sk_out *sk_out)
1161{
1162	struct sk_entry *entry;
1163	FLIST_HEAD(list);
1164	int ret = 0;
1165
1166	sk_lock(sk_out);
1167	if (flist_empty(&sk_out->list)) {
1168		sk_unlock(sk_out);
1169		return 0;
1170	}
1171
1172	flist_splice_init(&sk_out->list, &list);
1173	sk_unlock(sk_out);
1174
1175	while (!flist_empty(&list)) {
1176		entry = flist_entry(list.next, struct sk_entry, list);
1177		flist_del(&entry->list);
1178		ret += handle_sk_entry(sk_out, entry);
1179	}
1180
1181	return ret;
1182}
1183
1184static int handle_connection(struct sk_out *sk_out)
1185{
1186	struct fio_net_cmd *cmd = NULL;
1187	FLIST_HEAD(job_list);
1188	int ret = 0;
1189
1190	reset_fio_state();
1191
1192	/* read forever */
1193	while (!exit_backend) {
1194		struct pollfd pfd = {
1195			.fd	= sk_out->sk,
1196			.events	= POLLIN,
1197		};
1198
1199		ret = 0;
1200		do {
1201			int timeout = 1000;
1202
1203			if (!flist_empty(&job_list))
1204				timeout = 100;
1205
1206			handle_xmits(sk_out);
1207
1208			ret = poll(&pfd, 1, 0);
1209			if (ret < 0) {
1210				if (errno == EINTR)
1211					break;
1212				log_err("fio: poll: %s\n", strerror(errno));
1213				break;
1214			} else if (!ret) {
1215				fio_server_check_jobs(&job_list);
1216				fio_mutex_down_timeout(&sk_out->wait, timeout);
1217				continue;
1218			}
1219
1220			if (pfd.revents & POLLIN)
1221				break;
1222			if (pfd.revents & (POLLERR|POLLHUP)) {
1223				ret = 1;
1224				break;
1225			}
1226		} while (!exit_backend);
1227
1228		fio_server_check_jobs(&job_list);
1229
1230		if (ret < 0)
1231			break;
1232
1233		cmd = fio_net_recv_cmd(sk_out->sk, true);
1234		if (!cmd) {
1235			ret = -1;
1236			break;
1237		}
1238
1239		ret = handle_command(sk_out, &job_list, cmd);
1240		if (ret)
1241			break;
1242
1243		free(cmd);
1244		cmd = NULL;
1245	}
1246
1247	if (cmd)
1248		free(cmd);
1249
1250	handle_xmits(sk_out);
1251
1252	close(sk_out->sk);
1253	sk_out->sk = -1;
1254	__sk_out_drop(sk_out);
1255	_exit(ret);
1256}
1257
1258/* get the address on this host bound by the input socket,
1259 * whether it is ipv6 or ipv4 */
1260
1261static int get_my_addr_str(int sk)
1262{
1263	struct sockaddr_in6 myaddr6 = { 0, };
1264	struct sockaddr_in myaddr4 = { 0, };
1265	struct sockaddr *sockaddr_p;
1266	char *net_addr;
1267	socklen_t len;
1268	int ret;
1269
1270	if (use_ipv6) {
1271		len = sizeof(myaddr6);
1272		sockaddr_p = (struct sockaddr * )&myaddr6;
1273		net_addr = (char * )&myaddr6.sin6_addr;
1274	} else {
1275		len = sizeof(myaddr4);
1276		sockaddr_p = (struct sockaddr * )&myaddr4;
1277		net_addr = (char * )&myaddr4.sin_addr;
1278	}
1279
1280	ret = getsockname(sk, sockaddr_p, &len);
1281	if (ret) {
1282		log_err("fio: getsockaddr: %s\n", strerror(errno));
1283		return -1;
1284	}
1285
1286	if (!inet_ntop(use_ipv6?AF_INET6:AF_INET, net_addr, client_sockaddr_str, INET6_ADDRSTRLEN - 1)) {
1287		log_err("inet_ntop: failed to convert addr to string\n");
1288		return -1;
1289	}
1290
1291	dprint(FD_NET, "fio server bound to addr %s\n", client_sockaddr_str);
1292	return 0;
1293}
1294
1295static int accept_loop(int listen_sk)
1296{
1297	struct sockaddr_in addr;
1298	struct sockaddr_in6 addr6;
1299	socklen_t len = use_ipv6 ? sizeof(addr6) : sizeof(addr);
1300	struct pollfd pfd;
1301	int ret = 0, sk, exitval = 0;
1302	FLIST_HEAD(conn_list);
1303
1304	dprint(FD_NET, "server enter accept loop\n");
1305
1306	fio_set_fd_nonblocking(listen_sk, "server");
1307
1308	while (!exit_backend) {
1309		struct sk_out *sk_out;
1310		const char *from;
1311		char buf[64];
1312		pid_t pid;
1313
1314		pfd.fd = listen_sk;
1315		pfd.events = POLLIN;
1316		do {
1317			int timeout = 1000;
1318
1319			if (!flist_empty(&conn_list))
1320				timeout = 100;
1321
1322			ret = poll(&pfd, 1, timeout);
1323			if (ret < 0) {
1324				if (errno == EINTR)
1325					break;
1326				log_err("fio: poll: %s\n", strerror(errno));
1327				break;
1328			} else if (!ret) {
1329				fio_server_check_conns(&conn_list);
1330				continue;
1331			}
1332
1333			if (pfd.revents & POLLIN)
1334				break;
1335		} while (!exit_backend);
1336
1337		fio_server_check_conns(&conn_list);
1338
1339		if (exit_backend || ret < 0)
1340			break;
1341
1342		if (use_ipv6)
1343			sk = accept(listen_sk, (struct sockaddr *) &addr6, &len);
1344		else
1345			sk = accept(listen_sk, (struct sockaddr *) &addr, &len);
1346
1347		if (sk < 0) {
1348			log_err("fio: accept: %s\n", strerror(errno));
1349			return -1;
1350		}
1351
1352		if (use_ipv6)
1353			from = inet_ntop(AF_INET6, (struct sockaddr *) &addr6.sin6_addr, buf, sizeof(buf));
1354		else
1355			from = inet_ntop(AF_INET, (struct sockaddr *) &addr.sin_addr, buf, sizeof(buf));
1356
1357		dprint(FD_NET, "server: connect from %s\n", from);
1358
1359		sk_out = smalloc(sizeof(*sk_out));
1360		sk_out->sk = sk;
1361		INIT_FLIST_HEAD(&sk_out->list);
1362		__fio_mutex_init(&sk_out->lock, FIO_MUTEX_UNLOCKED);
1363		__fio_mutex_init(&sk_out->wait, FIO_MUTEX_LOCKED);
1364		__fio_mutex_init(&sk_out->xmit, FIO_MUTEX_UNLOCKED);
1365
1366		pid = fork();
1367		if (pid) {
1368			close(sk);
1369			fio_server_add_conn_pid(&conn_list, pid);
1370			continue;
1371		}
1372
1373		/* if error, it's already logged, non-fatal */
1374		get_my_addr_str(sk);
1375
1376		/*
1377		 * Assign sk_out here, it'll be dropped in handle_connection()
1378		 * since that function calls _exit() when done
1379		 */
1380		sk_out_assign(sk_out);
1381		handle_connection(sk_out);
1382	}
1383
1384	return exitval;
1385}
1386
1387int fio_server_text_output(int level, const char *buf, size_t len)
1388{
1389	struct sk_out *sk_out = pthread_getspecific(sk_out_key);
1390	struct cmd_text_pdu *pdu;
1391	unsigned int tlen;
1392	struct timeval tv;
1393
1394	if (!sk_out || sk_out->sk == -1)
1395		return -1;
1396
1397	tlen = sizeof(*pdu) + len;
1398	pdu = malloc(tlen);
1399
1400	pdu->level	= __cpu_to_le32(level);
1401	pdu->buf_len	= __cpu_to_le32(len);
1402
1403	gettimeofday(&tv, NULL);
1404	pdu->log_sec	= __cpu_to_le64(tv.tv_sec);
1405	pdu->log_usec	= __cpu_to_le64(tv.tv_usec);
1406
1407	memcpy(pdu->buf, buf, len);
1408
1409	fio_net_queue_cmd(FIO_NET_CMD_TEXT, pdu, tlen, NULL, SK_F_COPY);
1410	free(pdu);
1411	return len;
1412}
1413
1414static void convert_io_stat(struct io_stat *dst, struct io_stat *src)
1415{
1416	dst->max_val	= cpu_to_le64(src->max_val);
1417	dst->min_val	= cpu_to_le64(src->min_val);
1418	dst->samples	= cpu_to_le64(src->samples);
1419
1420	/*
1421	 * Encode to IEEE 754 for network transfer
1422	 */
1423	dst->mean.u.i	= cpu_to_le64(fio_double_to_uint64(src->mean.u.f));
1424	dst->S.u.i	= cpu_to_le64(fio_double_to_uint64(src->S.u.f));
1425}
1426
1427static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
1428{
1429	int i;
1430
1431	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1432		dst->max_run[i]		= cpu_to_le64(src->max_run[i]);
1433		dst->min_run[i]		= cpu_to_le64(src->min_run[i]);
1434		dst->max_bw[i]		= cpu_to_le64(src->max_bw[i]);
1435		dst->min_bw[i]		= cpu_to_le64(src->min_bw[i]);
1436		dst->iobytes[i]		= cpu_to_le64(src->iobytes[i]);
1437		dst->agg[i]		= cpu_to_le64(src->agg[i]);
1438	}
1439
1440	dst->kb_base	= cpu_to_le32(src->kb_base);
1441	dst->unit_base	= cpu_to_le32(src->unit_base);
1442	dst->groupid	= cpu_to_le32(src->groupid);
1443	dst->unified_rw_rep	= cpu_to_le32(src->unified_rw_rep);
1444}
1445
1446/*
1447 * Send a CMD_TS, which packs struct thread_stat and group_run_stats
1448 * into a single payload.
1449 */
1450void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
1451{
1452	struct cmd_ts_pdu p;
1453	int i, j;
1454	void *ss_buf;
1455	uint64_t *ss_iops, *ss_bw;
1456
1457	dprint(FD_NET, "server sending end stats\n");
1458
1459	memset(&p, 0, sizeof(p));
1460
1461	strncpy(p.ts.name, ts->name, FIO_JOBNAME_SIZE - 1);
1462	strncpy(p.ts.verror, ts->verror, FIO_VERROR_SIZE - 1);
1463	strncpy(p.ts.description, ts->description, FIO_JOBDESC_SIZE - 1);
1464
1465	p.ts.error		= cpu_to_le32(ts->error);
1466	p.ts.thread_number	= cpu_to_le32(ts->thread_number);
1467	p.ts.groupid		= cpu_to_le32(ts->groupid);
1468	p.ts.pid		= cpu_to_le32(ts->pid);
1469	p.ts.members		= cpu_to_le32(ts->members);
1470	p.ts.unified_rw_rep	= cpu_to_le32(ts->unified_rw_rep);
1471
1472	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1473		convert_io_stat(&p.ts.clat_stat[i], &ts->clat_stat[i]);
1474		convert_io_stat(&p.ts.slat_stat[i], &ts->slat_stat[i]);
1475		convert_io_stat(&p.ts.lat_stat[i], &ts->lat_stat[i]);
1476		convert_io_stat(&p.ts.bw_stat[i], &ts->bw_stat[i]);
1477	}
1478
1479	p.ts.usr_time		= cpu_to_le64(ts->usr_time);
1480	p.ts.sys_time		= cpu_to_le64(ts->sys_time);
1481	p.ts.ctx		= cpu_to_le64(ts->ctx);
1482	p.ts.minf		= cpu_to_le64(ts->minf);
1483	p.ts.majf		= cpu_to_le64(ts->majf);
1484	p.ts.clat_percentiles	= cpu_to_le64(ts->clat_percentiles);
1485	p.ts.percentile_precision = cpu_to_le64(ts->percentile_precision);
1486
1487	for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1488		fio_fp64_t *src = &ts->percentile_list[i];
1489		fio_fp64_t *dst = &p.ts.percentile_list[i];
1490
1491		dst->u.i = cpu_to_le64(fio_double_to_uint64(src->u.f));
1492	}
1493
1494	for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1495		p.ts.io_u_map[i]	= cpu_to_le32(ts->io_u_map[i]);
1496		p.ts.io_u_submit[i]	= cpu_to_le32(ts->io_u_submit[i]);
1497		p.ts.io_u_complete[i]	= cpu_to_le32(ts->io_u_complete[i]);
1498	}
1499
1500	for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1501		p.ts.io_u_lat_u[i]	= cpu_to_le32(ts->io_u_lat_u[i]);
1502	for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1503		p.ts.io_u_lat_m[i]	= cpu_to_le32(ts->io_u_lat_m[i]);
1504
1505	for (i = 0; i < DDIR_RWDIR_CNT; i++)
1506		for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1507			p.ts.io_u_plat[i][j] = cpu_to_le32(ts->io_u_plat[i][j]);
1508
1509	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1510		p.ts.total_io_u[i]	= cpu_to_le64(ts->total_io_u[i]);
1511		p.ts.short_io_u[i]	= cpu_to_le64(ts->short_io_u[i]);
1512		p.ts.drop_io_u[i]	= cpu_to_le64(ts->drop_io_u[i]);
1513	}
1514
1515	p.ts.total_submit	= cpu_to_le64(ts->total_submit);
1516	p.ts.total_complete	= cpu_to_le64(ts->total_complete);
1517
1518	for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1519		p.ts.io_bytes[i]	= cpu_to_le64(ts->io_bytes[i]);
1520		p.ts.runtime[i]		= cpu_to_le64(ts->runtime[i]);
1521	}
1522
1523	p.ts.total_run_time	= cpu_to_le64(ts->total_run_time);
1524	p.ts.continue_on_error	= cpu_to_le16(ts->continue_on_error);
1525	p.ts.total_err_count	= cpu_to_le64(ts->total_err_count);
1526	p.ts.first_error	= cpu_to_le32(ts->first_error);
1527	p.ts.kb_base		= cpu_to_le32(ts->kb_base);
1528	p.ts.unit_base		= cpu_to_le32(ts->unit_base);
1529
1530	p.ts.latency_depth	= cpu_to_le32(ts->latency_depth);
1531	p.ts.latency_target	= cpu_to_le64(ts->latency_target);
1532	p.ts.latency_window	= cpu_to_le64(ts->latency_window);
1533	p.ts.latency_percentile.u.i = cpu_to_le64(fio_double_to_uint64(ts->latency_percentile.u.f));
1534
1535	p.ts.nr_block_infos	= cpu_to_le64(ts->nr_block_infos);
1536	for (i = 0; i < p.ts.nr_block_infos; i++)
1537		p.ts.block_infos[i] = cpu_to_le32(ts->block_infos[i]);
1538
1539	p.ts.ss_dur		= cpu_to_le64(ts->ss_dur);
1540	p.ts.ss_state		= cpu_to_le32(ts->ss_state);
1541	p.ts.ss_head		= cpu_to_le32(ts->ss_head);
1542	p.ts.ss_limit.u.i	= cpu_to_le64(fio_double_to_uint64(ts->ss_limit.u.f));
1543	p.ts.ss_slope.u.i	= cpu_to_le64(fio_double_to_uint64(ts->ss_slope.u.f));
1544	p.ts.ss_deviation.u.i	= cpu_to_le64(fio_double_to_uint64(ts->ss_deviation.u.f));
1545	p.ts.ss_criterion.u.i	= cpu_to_le64(fio_double_to_uint64(ts->ss_criterion.u.f));
1546
1547	convert_gs(&p.rs, rs);
1548
1549	dprint(FD_NET, "ts->ss_state = %d\n", ts->ss_state);
1550	if (ts->ss_state & __FIO_SS_DATA) {
1551		dprint(FD_NET, "server sending steadystate ring buffers\n");
1552
1553		ss_buf = malloc(sizeof(p) + 2*ts->ss_dur*sizeof(uint64_t));
1554
1555		memcpy(ss_buf, &p, sizeof(p));
1556
1557		ss_iops = (uint64_t *) ((struct cmd_ts_pdu *)ss_buf + 1);
1558		ss_bw = ss_iops + (int) ts->ss_dur;
1559		for (i = 0; i < ts->ss_dur; i++) {
1560			ss_iops[i] = cpu_to_le64(ts->ss_iops_data[i]);
1561			ss_bw[i] = cpu_to_le64(ts->ss_bw_data[i]);
1562		}
1563
1564		fio_net_queue_cmd(FIO_NET_CMD_TS, ss_buf, sizeof(p) + 2*ts->ss_dur*sizeof(uint64_t), NULL, SK_F_COPY);
1565
1566		free(ss_buf);
1567	}
1568	else
1569		fio_net_queue_cmd(FIO_NET_CMD_TS, &p, sizeof(p), NULL, SK_F_COPY);
1570}
1571
1572void fio_server_send_gs(struct group_run_stats *rs)
1573{
1574	struct group_run_stats gs;
1575
1576	dprint(FD_NET, "server sending group run stats\n");
1577
1578	convert_gs(&gs, rs);
1579	fio_net_queue_cmd(FIO_NET_CMD_GS, &gs, sizeof(gs), NULL, SK_F_COPY);
1580}
1581
1582void fio_server_send_job_options(struct flist_head *opt_list,
1583				 unsigned int groupid)
1584{
1585	struct cmd_job_option pdu;
1586	struct flist_head *entry;
1587
1588	if (flist_empty(opt_list))
1589		return;
1590
1591	flist_for_each(entry, opt_list) {
1592		struct print_option *p;
1593		size_t len;
1594
1595		p = flist_entry(entry, struct print_option, list);
1596		memset(&pdu, 0, sizeof(pdu));
1597
1598		if (groupid == -1U) {
1599			pdu.global = __cpu_to_le16(1);
1600			pdu.groupid = 0;
1601		} else {
1602			pdu.global = 0;
1603			pdu.groupid = cpu_to_le32(groupid);
1604		}
1605		len = strlen(p->name);
1606		if (len >= sizeof(pdu.name)) {
1607			len = sizeof(pdu.name) - 1;
1608			pdu.truncated = __cpu_to_le16(1);
1609		}
1610		memcpy(pdu.name, p->name, len);
1611		if (p->value) {
1612			len = strlen(p->value);
1613			if (len >= sizeof(pdu.value)) {
1614				len = sizeof(pdu.value) - 1;
1615				pdu.truncated = __cpu_to_le16(1);
1616			}
1617			memcpy(pdu.value, p->value, len);
1618		}
1619		fio_net_queue_cmd(FIO_NET_CMD_JOB_OPT, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1620	}
1621}
1622
1623static void convert_agg(struct disk_util_agg *dst, struct disk_util_agg *src)
1624{
1625	int i;
1626
1627	for (i = 0; i < 2; i++) {
1628		dst->ios[i]	= cpu_to_le64(src->ios[i]);
1629		dst->merges[i]	= cpu_to_le64(src->merges[i]);
1630		dst->sectors[i]	= cpu_to_le64(src->sectors[i]);
1631		dst->ticks[i]	= cpu_to_le64(src->ticks[i]);
1632	}
1633
1634	dst->io_ticks		= cpu_to_le64(src->io_ticks);
1635	dst->time_in_queue	= cpu_to_le64(src->time_in_queue);
1636	dst->slavecount		= cpu_to_le32(src->slavecount);
1637	dst->max_util.u.i	= cpu_to_le64(fio_double_to_uint64(src->max_util.u.f));
1638}
1639
1640static void convert_dus(struct disk_util_stat *dst, struct disk_util_stat *src)
1641{
1642	int i;
1643
1644	dst->name[FIO_DU_NAME_SZ - 1] = '\0';
1645	strncpy((char *) dst->name, (char *) src->name, FIO_DU_NAME_SZ - 1);
1646
1647	for (i = 0; i < 2; i++) {
1648		dst->s.ios[i]		= cpu_to_le64(src->s.ios[i]);
1649		dst->s.merges[i]	= cpu_to_le64(src->s.merges[i]);
1650		dst->s.sectors[i]	= cpu_to_le64(src->s.sectors[i]);
1651		dst->s.ticks[i]		= cpu_to_le64(src->s.ticks[i]);
1652	}
1653
1654	dst->s.io_ticks		= cpu_to_le64(src->s.io_ticks);
1655	dst->s.time_in_queue	= cpu_to_le64(src->s.time_in_queue);
1656	dst->s.msec		= cpu_to_le64(src->s.msec);
1657}
1658
1659void fio_server_send_du(void)
1660{
1661	struct disk_util *du;
1662	struct flist_head *entry;
1663	struct cmd_du_pdu pdu;
1664
1665	dprint(FD_NET, "server: sending disk_util %d\n", !flist_empty(&disk_list));
1666
1667	memset(&pdu, 0, sizeof(pdu));
1668
1669	flist_for_each(entry, &disk_list) {
1670		du = flist_entry(entry, struct disk_util, list);
1671
1672		convert_dus(&pdu.dus, &du->dus);
1673		convert_agg(&pdu.agg, &du->agg);
1674
1675		fio_net_queue_cmd(FIO_NET_CMD_DU, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1676	}
1677}
1678
1679#ifdef CONFIG_ZLIB
1680
1681static inline void __fio_net_prep_tail(z_stream *stream, void *out_pdu,
1682					struct sk_entry **last_entry,
1683					struct sk_entry *first)
1684{
1685	unsigned int this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
1686
1687	*last_entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
1688				 NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
1689	flist_add_tail(&(*last_entry)->list, &first->next);
1690
1691}
1692
1693/*
1694 * Deflates the next input given, creating as many new packets in the
1695 * linked list as necessary.
1696 */
1697static int __deflate_pdu_buffer(void *next_in, unsigned int next_sz, void **out_pdu,
1698				struct sk_entry **last_entry, z_stream *stream,
1699				struct sk_entry *first)
1700{
1701	int ret;
1702
1703	stream->next_in = next_in;
1704	stream->avail_in = next_sz;
1705	do {
1706		if (! stream->avail_out) {
1707
1708			__fio_net_prep_tail(stream, *out_pdu, last_entry, first);
1709
1710			*out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1711
1712			stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1713			stream->next_out = *out_pdu;
1714		}
1715
1716		ret = deflate(stream, Z_BLOCK);
1717
1718		if (ret < 0) {
1719			free(*out_pdu);
1720			return 1;
1721		}
1722	} while (stream->avail_in);
1723
1724	return 0;
1725}
1726
1727static int __fio_append_iolog_gz_hist(struct sk_entry *first, struct io_log *log,
1728				      struct io_logs *cur_log, z_stream *stream)
1729{
1730	struct sk_entry *entry;
1731	void *out_pdu;
1732	int ret, i, j;
1733	int sample_sz = log_entry_sz(log);
1734
1735	out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1736	stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1737	stream->next_out = out_pdu;
1738
1739	for (i = 0; i < cur_log->nr_samples; i++) {
1740		struct io_sample *s;
1741		struct io_u_plat_entry *cur_plat_entry, *prev_plat_entry;
1742		unsigned int *cur_plat, *prev_plat;
1743
1744		s = get_sample(log, cur_log, i);
1745		ret = __deflate_pdu_buffer(s, sample_sz, &out_pdu, &entry, stream, first);
1746		if (ret)
1747			return ret;
1748
1749		/* Do the subtraction on server side so that client doesn't have to
1750		 * reconstruct our linked list from packets.
1751		 */
1752		cur_plat_entry  = s->data.plat_entry;
1753		prev_plat_entry = flist_first_entry(&cur_plat_entry->list, struct io_u_plat_entry, list);
1754		cur_plat  = cur_plat_entry->io_u_plat;
1755		prev_plat = prev_plat_entry->io_u_plat;
1756
1757		for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
1758			cur_plat[j] -= prev_plat[j];
1759		}
1760
1761		flist_del(&prev_plat_entry->list);
1762		free(prev_plat_entry);
1763
1764		ret = __deflate_pdu_buffer(cur_plat_entry, sizeof(*cur_plat_entry),
1765					   &out_pdu, &entry, stream, first);
1766
1767		if (ret)
1768			return ret;
1769	}
1770
1771	__fio_net_prep_tail(stream, out_pdu, &entry, first);
1772
1773	return 0;
1774}
1775
1776static int __fio_append_iolog_gz(struct sk_entry *first, struct io_log *log,
1777				 struct io_logs *cur_log, z_stream *stream)
1778{
1779	unsigned int this_len;
1780	void *out_pdu;
1781	int ret;
1782
1783	if (log->log_type == IO_LOG_TYPE_HIST)
1784		return __fio_append_iolog_gz_hist(first, log, cur_log, stream);
1785
1786	stream->next_in = (void *) cur_log->log;
1787	stream->avail_in = cur_log->nr_samples * log_entry_sz(log);
1788
1789	do {
1790		struct sk_entry *entry;
1791
1792		/*
1793		 * Dirty - since the log is potentially huge, compress it into
1794		 * FIO_SERVER_MAX_FRAGMENT_PDU chunks and let the receiving
1795		 * side defragment it.
1796		 */
1797		out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1798
1799		stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1800		stream->next_out = out_pdu;
1801		ret = deflate(stream, Z_BLOCK);
1802		/* may be Z_OK, or Z_STREAM_END */
1803		if (ret < 0) {
1804			free(out_pdu);
1805			return 1;
1806		}
1807
1808		this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
1809
1810		entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
1811					 NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
1812		flist_add_tail(&entry->list, &first->next);
1813	} while (stream->avail_in);
1814
1815	return 0;
1816}
1817
1818static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
1819{
1820	int ret = 0;
1821	z_stream stream;
1822
1823	memset(&stream, 0, sizeof(stream));
1824	stream.zalloc = Z_NULL;
1825	stream.zfree = Z_NULL;
1826	stream.opaque = Z_NULL;
1827
1828	if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK)
1829		return 1;
1830
1831	while (!flist_empty(&log->io_logs)) {
1832		struct io_logs *cur_log;
1833
1834		cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1835		flist_del_init(&cur_log->list);
1836
1837		ret = __fio_append_iolog_gz(first, log, cur_log, &stream);
1838		if (ret)
1839			break;
1840	}
1841
1842	ret = deflate(&stream, Z_FINISH);
1843
1844	while (ret != Z_STREAM_END) {
1845		struct sk_entry *entry;
1846		unsigned int this_len;
1847		void *out_pdu;
1848
1849		out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1850		stream.avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1851		stream.next_out = out_pdu;
1852
1853		ret = deflate(&stream, Z_FINISH);
1854		/* may be Z_OK, or Z_STREAM_END */
1855		if (ret < 0) {
1856			free(out_pdu);
1857			break;
1858		}
1859
1860		this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream.avail_out;
1861
1862		entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
1863					 NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
1864		flist_add_tail(&entry->list, &first->next);
1865	} while (ret != Z_STREAM_END);
1866
1867	ret = deflateEnd(&stream);
1868	if (ret == Z_OK)
1869		return 0;
1870
1871	return 1;
1872}
1873#else
1874static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
1875{
1876	return 1;
1877}
1878#endif
1879
1880static int fio_append_gz_chunks(struct sk_entry *first, struct io_log *log)
1881{
1882	struct sk_entry *entry;
1883	struct flist_head *node;
1884
1885	pthread_mutex_lock(&log->chunk_lock);
1886	flist_for_each(node, &log->chunk_list) {
1887		struct iolog_compress *c;
1888
1889		c = flist_entry(node, struct iolog_compress, list);
1890		entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, c->buf, c->len,
1891						NULL, SK_F_VEC | SK_F_INLINE);
1892		flist_add_tail(&entry->list, &first->next);
1893	}
1894	pthread_mutex_unlock(&log->chunk_lock);
1895
1896	return 0;
1897}
1898
1899static int fio_append_text_log(struct sk_entry *first, struct io_log *log)
1900{
1901	struct sk_entry *entry;
1902
1903	while (!flist_empty(&log->io_logs)) {
1904		struct io_logs *cur_log;
1905		size_t size;
1906
1907		cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1908		flist_del_init(&cur_log->list);
1909
1910		size = cur_log->nr_samples * log_entry_sz(log);
1911
1912		entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, cur_log->log, size,
1913						NULL, SK_F_VEC | SK_F_INLINE);
1914		flist_add_tail(&entry->list, &first->next);
1915	}
1916
1917	return 0;
1918}
1919
1920int fio_send_iolog(struct thread_data *td, struct io_log *log, const char *name)
1921{
1922	struct cmd_iolog_pdu pdu;
1923	struct sk_entry *first;
1924	struct flist_head *entry;
1925	int ret = 0;
1926
1927	pdu.nr_samples = cpu_to_le64(iolog_nr_samples(log));
1928	pdu.thread_number = cpu_to_le32(td->thread_number);
1929	pdu.log_type = cpu_to_le32(log->log_type);
1930	pdu.log_hist_coarseness = cpu_to_le32(log->hist_coarseness);
1931
1932	if (!flist_empty(&log->chunk_list))
1933		pdu.compressed = __cpu_to_le32(STORE_COMPRESSED);
1934	else if (use_zlib)
1935		pdu.compressed = __cpu_to_le32(XMIT_COMPRESSED);
1936	else
1937		pdu.compressed = 0;
1938
1939	strncpy((char *) pdu.name, name, FIO_NET_NAME_MAX);
1940	pdu.name[FIO_NET_NAME_MAX - 1] = '\0';
1941
1942	/*
1943	 * We can't do this for a pre-compressed log, but for that case,
1944	 * log->nr_samples is zero anyway.
1945	 */
1946	flist_for_each(entry, &log->io_logs) {
1947		struct io_logs *cur_log;
1948		int i;
1949
1950		cur_log = flist_entry(entry, struct io_logs, list);
1951
1952		for (i = 0; i < cur_log->nr_samples; i++) {
1953			struct io_sample *s = get_sample(log, cur_log, i);
1954
1955			s->time		= cpu_to_le64(s->time);
1956			s->data.val	= cpu_to_le64(s->data.val);
1957			s->__ddir	= cpu_to_le32(s->__ddir);
1958			s->bs		= cpu_to_le32(s->bs);
1959
1960			if (log->log_offset) {
1961				struct io_sample_offset *so = (void *) s;
1962
1963				so->offset = cpu_to_le64(so->offset);
1964			}
1965		}
1966	}
1967
1968	/*
1969	 * Assemble header entry first
1970	 */
1971	first = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, &pdu, sizeof(pdu), NULL, SK_F_VEC | SK_F_INLINE | SK_F_COPY);
1972
1973	/*
1974	 * Now append actual log entries. If log compression was enabled on
1975	 * the job, just send out the compressed chunks directly. If we
1976	 * have a plain log, compress if we can, then send. Otherwise, send
1977	 * the plain text output.
1978	 */
1979	if (!flist_empty(&log->chunk_list))
1980		ret = fio_append_gz_chunks(first, log);
1981	else if (use_zlib)
1982		ret = fio_append_iolog_gz(first, log);
1983	else
1984		ret = fio_append_text_log(first, log);
1985
1986	fio_net_queue_entry(first);
1987	return ret;
1988}
1989
1990void fio_server_send_add_job(struct thread_data *td)
1991{
1992	struct cmd_add_job_pdu pdu;
1993
1994	memset(&pdu, 0, sizeof(pdu));
1995	pdu.thread_number = cpu_to_le32(td->thread_number);
1996	pdu.groupid = cpu_to_le32(td->groupid);
1997	convert_thread_options_to_net(&pdu.top, &td->o);
1998
1999	fio_net_queue_cmd(FIO_NET_CMD_ADD_JOB, &pdu, sizeof(pdu), NULL,
2000				SK_F_COPY);
2001}
2002
2003void fio_server_send_start(struct thread_data *td)
2004{
2005	struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2006
2007	assert(sk_out->sk != -1);
2008
2009	fio_net_queue_cmd(FIO_NET_CMD_SERVER_START, NULL, 0, NULL, SK_F_SIMPLE);
2010}
2011
2012int fio_server_get_verify_state(const char *name, int threadnumber,
2013				void **datap)
2014{
2015	struct thread_io_list *s;
2016	struct cmd_sendfile out;
2017	struct cmd_reply *rep;
2018	uint64_t tag;
2019	void *data;
2020	int ret;
2021
2022	dprint(FD_NET, "server: request verify state\n");
2023
2024	rep = smalloc(sizeof(*rep));
2025	if (!rep)
2026		return ENOMEM;
2027
2028	__fio_mutex_init(&rep->lock, FIO_MUTEX_LOCKED);
2029	rep->data = NULL;
2030	rep->error = 0;
2031
2032	verify_state_gen_name((char *) out.path, sizeof(out.path), name, me,
2033				threadnumber);
2034	tag = (uint64_t) (uintptr_t) rep;
2035	fio_net_queue_cmd(FIO_NET_CMD_SENDFILE, &out, sizeof(out), &tag,
2036				SK_F_COPY);
2037
2038	/*
2039	 * Wait for the backend to receive the reply
2040	 */
2041	if (fio_mutex_down_timeout(&rep->lock, 10000)) {
2042		log_err("fio: timed out waiting for reply\n");
2043		ret = ETIMEDOUT;
2044		goto fail;
2045	}
2046
2047	if (rep->error) {
2048		log_err("fio: failure on receiving state file %s: %s\n",
2049				out.path, strerror(rep->error));
2050		ret = rep->error;
2051fail:
2052		*datap = NULL;
2053		sfree(rep);
2054		fio_net_queue_quit();
2055		return ret;
2056	}
2057
2058	/*
2059	 * The format is verify_state_hdr, then thread_io_list. Verify
2060	 * the header, and the thread_io_list checksum
2061	 */
2062	s = rep->data + sizeof(struct verify_state_hdr);
2063	if (verify_state_hdr(rep->data, s)) {
2064		ret = EILSEQ;
2065		goto fail;
2066	}
2067
2068	/*
2069	 * Don't need the header from now, copy just the thread_io_list
2070	 */
2071	ret = 0;
2072	rep->size -= sizeof(struct verify_state_hdr);
2073	data = malloc(rep->size);
2074	memcpy(data, s, rep->size);
2075	*datap = data;
2076
2077	sfree(rep->data);
2078	__fio_mutex_remove(&rep->lock);
2079	sfree(rep);
2080	return ret;
2081}
2082
2083static int fio_init_server_ip(void)
2084{
2085	struct sockaddr *addr;
2086	socklen_t socklen;
2087	char buf[80];
2088	const char *str;
2089	int sk, opt;
2090
2091	if (use_ipv6)
2092		sk = socket(AF_INET6, SOCK_STREAM, 0);
2093	else
2094		sk = socket(AF_INET, SOCK_STREAM, 0);
2095
2096	if (sk < 0) {
2097		log_err("fio: socket: %s\n", strerror(errno));
2098		return -1;
2099	}
2100
2101	opt = 1;
2102	if (setsockopt(sk, SOL_SOCKET, SO_REUSEADDR, (void *)&opt, sizeof(opt)) < 0) {
2103		log_err("fio: setsockopt(REUSEADDR): %s\n", strerror(errno));
2104		close(sk);
2105		return -1;
2106	}
2107#ifdef SO_REUSEPORT
2108	/*
2109	 * Not fatal if fails, so just ignore it if that happens
2110	 */
2111	setsockopt(sk, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt));
2112#endif
2113
2114	if (use_ipv6) {
2115		const void *src = &saddr_in6.sin6_addr;
2116
2117		addr = (struct sockaddr *) &saddr_in6;
2118		socklen = sizeof(saddr_in6);
2119		saddr_in6.sin6_family = AF_INET6;
2120		str = inet_ntop(AF_INET6, src, buf, sizeof(buf));
2121	} else {
2122		const void *src = &saddr_in.sin_addr;
2123
2124		addr = (struct sockaddr *) &saddr_in;
2125		socklen = sizeof(saddr_in);
2126		saddr_in.sin_family = AF_INET;
2127		str = inet_ntop(AF_INET, src, buf, sizeof(buf));
2128	}
2129
2130	if (bind(sk, addr, socklen) < 0) {
2131		log_err("fio: bind: %s\n", strerror(errno));
2132		log_info("fio: failed with IPv%c %s\n", use_ipv6 ? '6' : '4', str);
2133		close(sk);
2134		return -1;
2135	}
2136
2137	return sk;
2138}
2139
2140static int fio_init_server_sock(void)
2141{
2142	struct sockaddr_un addr;
2143	socklen_t len;
2144	mode_t mode;
2145	int sk;
2146
2147	sk = socket(AF_UNIX, SOCK_STREAM, 0);
2148	if (sk < 0) {
2149		log_err("fio: socket: %s\n", strerror(errno));
2150		return -1;
2151	}
2152
2153	mode = umask(000);
2154
2155	memset(&addr, 0, sizeof(addr));
2156	addr.sun_family = AF_UNIX;
2157	strncpy(addr.sun_path, bind_sock, sizeof(addr.sun_path) - 1);
2158
2159	len = sizeof(addr.sun_family) + strlen(bind_sock) + 1;
2160
2161	if (bind(sk, (struct sockaddr *) &addr, len) < 0) {
2162		log_err("fio: bind: %s\n", strerror(errno));
2163		close(sk);
2164		return -1;
2165	}
2166
2167	umask(mode);
2168	return sk;
2169}
2170
2171static int fio_init_server_connection(void)
2172{
2173	char bind_str[128];
2174	int sk;
2175
2176	dprint(FD_NET, "starting server\n");
2177
2178	if (!bind_sock)
2179		sk = fio_init_server_ip();
2180	else
2181		sk = fio_init_server_sock();
2182
2183	if (sk < 0)
2184		return sk;
2185
2186	memset(bind_str, 0, sizeof(bind_str));
2187
2188	if (!bind_sock) {
2189		char *p, port[16];
2190		const void *src;
2191		int af;
2192
2193		if (use_ipv6) {
2194			af = AF_INET6;
2195			src = &saddr_in6.sin6_addr;
2196		} else {
2197			af = AF_INET;
2198			src = &saddr_in.sin_addr;
2199		}
2200
2201		p = (char *) inet_ntop(af, src, bind_str, sizeof(bind_str));
2202
2203		sprintf(port, ",%u", fio_net_port);
2204		if (p)
2205			strcat(p, port);
2206		else
2207			strncpy(bind_str, port, sizeof(bind_str) - 1);
2208	} else
2209		strncpy(bind_str, bind_sock, sizeof(bind_str) - 1);
2210
2211	log_info("fio: server listening on %s\n", bind_str);
2212
2213	if (listen(sk, 4) < 0) {
2214		log_err("fio: listen: %s\n", strerror(errno));
2215		close(sk);
2216		return -1;
2217	}
2218
2219	return sk;
2220}
2221
2222int fio_server_parse_host(const char *host, int ipv6, struct in_addr *inp,
2223			  struct in6_addr *inp6)
2224
2225{
2226	int ret = 0;
2227
2228	if (ipv6)
2229		ret = inet_pton(AF_INET6, host, inp6);
2230	else
2231		ret = inet_pton(AF_INET, host, inp);
2232
2233	if (ret != 1) {
2234		struct addrinfo hints, *res;
2235
2236		memset(&hints, 0, sizeof(hints));
2237		hints.ai_family = ipv6 ? AF_INET6 : AF_INET;
2238		hints.ai_socktype = SOCK_STREAM;
2239
2240		ret = getaddrinfo(host, NULL, &hints, &res);
2241		if (ret) {
2242			log_err("fio: failed to resolve <%s> (%s)\n", host,
2243					gai_strerror(ret));
2244			return 1;
2245		}
2246
2247		if (ipv6)
2248			memcpy(inp6, &((struct sockaddr_in6 *) res->ai_addr)->sin6_addr, sizeof(*inp6));
2249		else
2250			memcpy(inp, &((struct sockaddr_in *) res->ai_addr)->sin_addr, sizeof(*inp));
2251
2252		ret = 1;
2253		freeaddrinfo(res);
2254	}
2255
2256	return !(ret == 1);
2257}
2258
2259/*
2260 * Parse a host/ip/port string. Reads from 'str'.
2261 *
2262 * Outputs:
2263 *
2264 * For IPv4:
2265 *	*ptr is the host, *port is the port, inp is the destination.
2266 * For IPv6:
2267 *	*ptr is the host, *port is the port, inp6 is the dest, and *ipv6 is 1.
2268 * For local domain sockets:
2269 *	*ptr is the filename, *is_sock is 1.
2270 */
2271int fio_server_parse_string(const char *str, char **ptr, int *is_sock,
2272			    int *port, struct in_addr *inp,
2273			    struct in6_addr *inp6, int *ipv6)
2274{
2275	const char *host = str;
2276	char *portp;
2277	int lport = 0;
2278
2279	*ptr = NULL;
2280	*is_sock = 0;
2281	*port = fio_net_port;
2282	*ipv6 = 0;
2283
2284	if (!strncmp(str, "sock:", 5)) {
2285		*ptr = strdup(str + 5);
2286		*is_sock = 1;
2287
2288		return 0;
2289	}
2290
2291	/*
2292	 * Is it ip:<ip or host>:port
2293	 */
2294	if (!strncmp(host, "ip:", 3))
2295		host += 3;
2296	else if (!strncmp(host, "ip4:", 4))
2297		host += 4;
2298	else if (!strncmp(host, "ip6:", 4)) {
2299		host += 4;
2300		*ipv6 = 1;
2301	} else if (host[0] == ':') {
2302		/* String is :port */
2303		host++;
2304		lport = atoi(host);
2305		if (!lport || lport > 65535) {
2306			log_err("fio: bad server port %u\n", lport);
2307			return 1;
2308		}
2309		/* no hostname given, we are done */
2310		*port = lport;
2311		return 0;
2312	}
2313
2314	/*
2315	 * If no port seen yet, check if there's a last ',' at the end
2316	 */
2317	if (!lport) {
2318		portp = strchr(host, ',');
2319		if (portp) {
2320			*portp = '\0';
2321			portp++;
2322			lport = atoi(portp);
2323			if (!lport || lport > 65535) {
2324				log_err("fio: bad server port %u\n", lport);
2325				return 1;
2326			}
2327		}
2328	}
2329
2330	if (lport)
2331		*port = lport;
2332
2333	if (!strlen(host))
2334		return 0;
2335
2336	*ptr = strdup(host);
2337
2338	if (fio_server_parse_host(*ptr, *ipv6, inp, inp6)) {
2339		free(*ptr);
2340		*ptr = NULL;
2341		return 1;
2342	}
2343
2344	if (*port == 0)
2345		*port = fio_net_port;
2346
2347	return 0;
2348}
2349
2350/*
2351 * Server arg should be one of:
2352 *
2353 * sock:/path/to/socket
2354 *   ip:1.2.3.4
2355 *      1.2.3.4
2356 *
2357 * Where sock uses unix domain sockets, and ip binds the server to
2358 * a specific interface. If no arguments are given to the server, it
2359 * uses IP and binds to 0.0.0.0.
2360 *
2361 */
2362static int fio_handle_server_arg(void)
2363{
2364	int port = fio_net_port;
2365	int is_sock, ret = 0;
2366
2367	saddr_in.sin_addr.s_addr = htonl(INADDR_ANY);
2368
2369	if (!fio_server_arg)
2370		goto out;
2371
2372	ret = fio_server_parse_string(fio_server_arg, &bind_sock, &is_sock,
2373					&port, &saddr_in.sin_addr,
2374					&saddr_in6.sin6_addr, &use_ipv6);
2375
2376	if (!is_sock && bind_sock) {
2377		free(bind_sock);
2378		bind_sock = NULL;
2379	}
2380
2381out:
2382	fio_net_port = port;
2383	saddr_in.sin_port = htons(port);
2384	saddr_in6.sin6_port = htons(port);
2385	return ret;
2386}
2387
2388static void sig_int(int sig)
2389{
2390	if (bind_sock)
2391		unlink(bind_sock);
2392}
2393
2394static void set_sig_handlers(void)
2395{
2396	struct sigaction act;
2397
2398	memset(&act, 0, sizeof(act));
2399	act.sa_handler = sig_int;
2400	act.sa_flags = SA_RESTART;
2401	sigaction(SIGINT, &act, NULL);
2402}
2403
2404void fio_server_destroy_sk_key(void)
2405{
2406	pthread_key_delete(sk_out_key);
2407}
2408
2409int fio_server_create_sk_key(void)
2410{
2411	if (pthread_key_create(&sk_out_key, NULL)) {
2412		log_err("fio: can't create sk_out backend key\n");
2413		return 1;
2414	}
2415
2416	pthread_setspecific(sk_out_key, NULL);
2417	return 0;
2418}
2419
2420static int fio_server(void)
2421{
2422	int sk, ret;
2423
2424	dprint(FD_NET, "starting server\n");
2425
2426	if (fio_handle_server_arg())
2427		return -1;
2428
2429	sk = fio_init_server_connection();
2430	if (sk < 0)
2431		return -1;
2432
2433	set_sig_handlers();
2434
2435	ret = accept_loop(sk);
2436
2437	close(sk);
2438
2439	if (fio_server_arg) {
2440		free(fio_server_arg);
2441		fio_server_arg = NULL;
2442	}
2443	if (bind_sock)
2444		free(bind_sock);
2445
2446	return ret;
2447}
2448
2449void fio_server_got_signal(int signal)
2450{
2451	struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2452
2453	assert(sk_out);
2454
2455	if (signal == SIGPIPE)
2456		sk_out->sk = -1;
2457	else {
2458		log_info("\nfio: terminating on signal %d\n", signal);
2459		exit_backend = 1;
2460	}
2461}
2462
2463static int check_existing_pidfile(const char *pidfile)
2464{
2465	struct stat sb;
2466	char buf[16];
2467	pid_t pid;
2468	FILE *f;
2469
2470	if (stat(pidfile, &sb))
2471		return 0;
2472
2473	f = fopen(pidfile, "r");
2474	if (!f)
2475		return 0;
2476
2477	if (fread(buf, sb.st_size, 1, f) <= 0) {
2478		fclose(f);
2479		return 1;
2480	}
2481	fclose(f);
2482
2483	pid = atoi(buf);
2484	if (kill(pid, SIGCONT) < 0)
2485		return errno != ESRCH;
2486
2487	return 1;
2488}
2489
2490static int write_pid(pid_t pid, const char *pidfile)
2491{
2492	FILE *fpid;
2493
2494	fpid = fopen(pidfile, "w");
2495	if (!fpid) {
2496		log_err("fio: failed opening pid file %s\n", pidfile);
2497		return 1;
2498	}
2499
2500	fprintf(fpid, "%u\n", (unsigned int) pid);
2501	fclose(fpid);
2502	return 0;
2503}
2504
2505/*
2506 * If pidfile is specified, background us.
2507 */
2508int fio_start_server(char *pidfile)
2509{
2510	pid_t pid;
2511	int ret;
2512
2513#if defined(WIN32)
2514	WSADATA wsd;
2515	WSAStartup(MAKEWORD(2, 2), &wsd);
2516#endif
2517
2518	if (!pidfile)
2519		return fio_server();
2520
2521	if (check_existing_pidfile(pidfile)) {
2522		log_err("fio: pidfile %s exists and server appears alive\n",
2523								pidfile);
2524		free(pidfile);
2525		return -1;
2526	}
2527
2528	pid = fork();
2529	if (pid < 0) {
2530		log_err("fio: failed server fork: %s\n", strerror(errno));
2531		free(pidfile);
2532		return -1;
2533	} else if (pid) {
2534		ret = write_pid(pid, pidfile);
2535		free(pidfile);
2536		_exit(ret);
2537	}
2538
2539	setsid();
2540	openlog("fio", LOG_NDELAY|LOG_NOWAIT|LOG_PID, LOG_USER);
2541	log_syslog = 1;
2542	close(STDIN_FILENO);
2543	close(STDOUT_FILENO);
2544	close(STDERR_FILENO);
2545	f_out = NULL;
2546	f_err = NULL;
2547
2548	ret = fio_server();
2549
2550	closelog();
2551	unlink(pidfile);
2552	free(pidfile);
2553	return ret;
2554}
2555
2556void fio_server_set_arg(const char *arg)
2557{
2558	fio_server_arg = strdup(arg);
2559}
2560