Merge branch 'master' of https://github.com/celestinechen/fio
[fio.git] / server.c
1 #include <stdio.h>
2 #include <stdlib.h>
3 #include <unistd.h>
4 #include <errno.h>
5 #include <poll.h>
6 #include <sys/types.h>
7 #include <sys/wait.h>
8 #include <sys/socket.h>
9 #include <sys/stat.h>
10 #include <sys/un.h>
11 #include <sys/uio.h>
12 #include <netinet/in.h>
13 #include <arpa/inet.h>
14 #include <netdb.h>
15 #include <syslog.h>
16 #include <signal.h>
17 #ifdef CONFIG_ZLIB
18 #include <zlib.h>
19 #endif
20
21 #include "fio.h"
22 #include "options.h"
23 #include "server.h"
24 #include "crc/crc16.h"
25 #include "lib/ieee754.h"
26 #include "verify-state.h"
27 #include "smalloc.h"
28
29 int fio_net_port = FIO_NET_PORT;
30
31 bool exit_backend = false;
32
33 enum {
34         SK_F_FREE       = 1,
35         SK_F_COPY       = 2,
36         SK_F_SIMPLE     = 4,
37         SK_F_VEC        = 8,
38         SK_F_INLINE     = 16,
39 };
40
41 struct sk_entry {
42         struct flist_head list; /* link on sk_out->list */
43         int flags;              /* SK_F_* */
44         int opcode;             /* Actual command fields */
45         void *buf;
46         off_t size;
47         uint64_t tag;
48         struct flist_head next; /* Other sk_entry's, if linked command */
49 };
50
51 static char *fio_server_arg;
52 static char *bind_sock;
53 static struct sockaddr_in saddr_in;
54 static struct sockaddr_in6 saddr_in6;
55 static int use_ipv6;
56 #ifdef CONFIG_ZLIB
57 static unsigned int has_zlib = 1;
58 #else
59 static unsigned int has_zlib = 0;
60 #endif
61 static unsigned int use_zlib;
62 static char me[128];
63
64 static pthread_key_t sk_out_key;
65
66 #ifdef WIN32
67 static char *fio_server_pipe_name  = NULL;
68 static HANDLE hjob = INVALID_HANDLE_VALUE;
69 struct ffi_element {
70         union {
71                 pthread_t thread;
72                 HANDLE hProcess;
73         };
74         bool is_thread;
75 };
76 #endif
77
78 struct fio_fork_item {
79         struct flist_head list;
80         int exitval;
81         int signal;
82         int exited;
83 #ifdef WIN32
84         struct ffi_element element;
85 #else
86         pid_t pid;
87 #endif
88 };
89
90 struct cmd_reply {
91         struct fio_sem lock;
92         void *data;
93         size_t size;
94         int error;
95 };
96
97 static const char *fio_server_ops[FIO_NET_CMD_NR] = {
98         "",
99         "QUIT",
100         "EXIT",
101         "JOB",
102         "JOBLINE",
103         "TEXT",
104         "TS",
105         "GS",
106         "SEND_ETA",
107         "ETA",
108         "PROBE",
109         "START",
110         "STOP",
111         "DISK_UTIL",
112         "SERVER_START",
113         "ADD_JOB",
114         "RUN",
115         "IOLOG",
116         "UPDATE_JOB",
117         "LOAD_FILE",
118         "VTRIGGER",
119         "SENDFILE",
120         "JOB_OPT",
121 };
122
123 static void sk_lock(struct sk_out *sk_out)
124 {
125         fio_sem_down(&sk_out->lock);
126 }
127
128 static void sk_unlock(struct sk_out *sk_out)
129 {
130         fio_sem_up(&sk_out->lock);
131 }
132
133 void sk_out_assign(struct sk_out *sk_out)
134 {
135         if (!sk_out)
136                 return;
137
138         sk_lock(sk_out);
139         sk_out->refs++;
140         sk_unlock(sk_out);
141         pthread_setspecific(sk_out_key, sk_out);
142 }
143
144 static void sk_out_free(struct sk_out *sk_out)
145 {
146         __fio_sem_remove(&sk_out->lock);
147         __fio_sem_remove(&sk_out->wait);
148         __fio_sem_remove(&sk_out->xmit);
149         sfree(sk_out);
150 }
151
152 static int __sk_out_drop(struct sk_out *sk_out)
153 {
154         if (sk_out) {
155                 int refs;
156
157                 sk_lock(sk_out);
158                 assert(sk_out->refs != 0);
159                 refs = --sk_out->refs;
160                 sk_unlock(sk_out);
161
162                 if (!refs) {
163                         sk_out_free(sk_out);
164                         pthread_setspecific(sk_out_key, NULL);
165                         return 0;
166                 }
167         }
168
169         return 1;
170 }
171
172 void sk_out_drop(void)
173 {
174         struct sk_out *sk_out;
175
176         sk_out = pthread_getspecific(sk_out_key);
177         __sk_out_drop(sk_out);
178 }
179
180 static void __fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
181                                uint32_t pdu_len, uint64_t tag)
182 {
183         memset(cmd, 0, sizeof(*cmd));
184
185         cmd->version    = __cpu_to_le16(FIO_SERVER_VER);
186         cmd->opcode     = cpu_to_le16(opcode);
187         cmd->tag        = cpu_to_le64(tag);
188         cmd->pdu_len    = cpu_to_le32(pdu_len);
189 }
190
191
192 static void fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
193                              const void *pdu, uint32_t pdu_len, uint64_t tag)
194 {
195         __fio_init_net_cmd(cmd, opcode, pdu_len, tag);
196
197         if (pdu)
198                 memcpy(&cmd->payload, pdu, pdu_len);
199 }
200
201 const char *fio_server_op(unsigned int op)
202 {
203         static char buf[32];
204
205         if (op < FIO_NET_CMD_NR)
206                 return fio_server_ops[op];
207
208         sprintf(buf, "UNKNOWN/%d", op);
209         return buf;
210 }
211
212 static ssize_t iov_total_len(const struct iovec *iov, int count)
213 {
214         ssize_t ret = 0;
215
216         while (count--) {
217                 ret += iov->iov_len;
218                 iov++;
219         }
220
221         return ret;
222 }
223
224 static int fio_sendv_data(int sk, struct iovec *iov, int count)
225 {
226         ssize_t total_len = iov_total_len(iov, count);
227         ssize_t ret;
228
229         do {
230                 ret = writev(sk, iov, count);
231                 if (ret > 0) {
232                         total_len -= ret;
233                         if (!total_len)
234                                 break;
235
236                         while (ret) {
237                                 if (ret >= iov->iov_len) {
238                                         ret -= iov->iov_len;
239                                         iov++;
240                                         continue;
241                                 }
242                                 iov->iov_base += ret;
243                                 iov->iov_len -= ret;
244                                 ret = 0;
245                         }
246                 } else if (!ret)
247                         break;
248                 else if (errno == EAGAIN || errno == EINTR)
249                         continue;
250                 else
251                         break;
252         } while (!exit_backend);
253
254         if (!total_len)
255                 return 0;
256
257         return 1;
258 }
259
260 static int fio_send_data(int sk, const void *p, unsigned int len)
261 {
262         struct iovec iov = { .iov_base = (void *) p, .iov_len = len };
263
264         assert(len <= sizeof(struct fio_net_cmd) + FIO_SERVER_MAX_FRAGMENT_PDU);
265
266         return fio_sendv_data(sk, &iov, 1);
267 }
268
269 bool fio_server_poll_fd(int fd, short events, int timeout)
270 {
271         struct pollfd pfd = {
272                 .fd     = fd,
273                 .events = events,
274         };
275         int ret;
276
277         ret = poll(&pfd, 1, timeout);
278         if (ret < 0) {
279                 if (errno == EINTR)
280                         return false;
281                 log_err("fio: poll: %s\n", strerror(errno));
282                 return false;
283         } else if (!ret) {
284                 return false;
285         }
286         if (pfd.revents & events)
287                 return true;
288         return false;
289 }
290
291 static int fio_recv_data(int sk, void *buf, unsigned int len, bool wait)
292 {
293         int flags;
294         char *p = buf;
295
296         if (wait)
297                 flags = MSG_WAITALL;
298         else
299                 flags = OS_MSG_DONTWAIT;
300
301         do {
302                 int ret = recv(sk, p, len, flags);
303
304                 if (ret > 0) {
305                         len -= ret;
306                         if (!len)
307                                 break;
308                         p += ret;
309                         continue;
310                 } else if (!ret)
311                         break;
312                 else if (errno == EAGAIN || errno == EINTR) {
313                         if (wait)
314                                 continue;
315                         break;
316                 } else
317                         break;
318         } while (!exit_backend);
319
320         if (!len)
321                 return 0;
322
323         return -1;
324 }
325
326 static int verify_convert_cmd(struct fio_net_cmd *cmd)
327 {
328         uint16_t crc;
329
330         cmd->cmd_crc16 = le16_to_cpu(cmd->cmd_crc16);
331         cmd->pdu_crc16 = le16_to_cpu(cmd->pdu_crc16);
332
333         crc = fio_crc16(cmd, FIO_NET_CMD_CRC_SZ);
334         if (crc != cmd->cmd_crc16) {
335                 log_err("fio: server bad crc on command (got %x, wanted %x)\n",
336                                 cmd->cmd_crc16, crc);
337                 fprintf(f_err, "fio: server bad crc on command (got %x, wanted %x)\n",
338                                 cmd->cmd_crc16, crc);
339                 return 1;
340         }
341
342         cmd->version    = le16_to_cpu(cmd->version);
343         cmd->opcode     = le16_to_cpu(cmd->opcode);
344         cmd->flags      = le32_to_cpu(cmd->flags);
345         cmd->tag        = le64_to_cpu(cmd->tag);
346         cmd->pdu_len    = le32_to_cpu(cmd->pdu_len);
347
348         switch (cmd->version) {
349         case FIO_SERVER_VER:
350                 break;
351         default:
352                 log_err("fio: bad server cmd version %d\n", cmd->version);
353                 fprintf(f_err, "fio: client/server version mismatch (%d != %d)\n",
354                                 cmd->version, FIO_SERVER_VER);
355                 return 1;
356         }
357
358         if (cmd->pdu_len > FIO_SERVER_MAX_FRAGMENT_PDU) {
359                 log_err("fio: command payload too large: %u\n", cmd->pdu_len);
360                 return 1;
361         }
362
363         return 0;
364 }
365
366 /*
367  * Read (and defragment, if necessary) incoming commands
368  */
369 struct fio_net_cmd *fio_net_recv_cmd(int sk, bool wait)
370 {
371         struct fio_net_cmd cmd, *tmp, *cmdret = NULL;
372         size_t cmd_size = 0, pdu_offset = 0;
373         uint16_t crc;
374         int ret, first = 1;
375         void *pdu = NULL;
376
377         do {
378                 ret = fio_recv_data(sk, &cmd, sizeof(cmd), wait);
379                 if (ret)
380                         break;
381
382                 /* We have a command, verify it and swap if need be */
383                 ret = verify_convert_cmd(&cmd);
384                 if (ret)
385                         break;
386
387                 if (first) {
388                         /* if this is text, add room for \0 at the end */
389                         cmd_size = sizeof(cmd) + cmd.pdu_len + 1;
390                         assert(!cmdret);
391                 } else
392                         cmd_size += cmd.pdu_len;
393
394                 if (cmd_size / 1024 > FIO_SERVER_MAX_CMD_MB * 1024) {
395                         log_err("fio: cmd+pdu too large (%llu)\n", (unsigned long long) cmd_size);
396                         ret = 1;
397                         break;
398                 }
399
400                 tmp = realloc(cmdret, cmd_size);
401                 if (!tmp) {
402                         log_err("fio: server failed allocating cmd\n");
403                         ret = 1;
404                         break;
405                 }
406                 cmdret = tmp;
407
408                 if (first)
409                         memcpy(cmdret, &cmd, sizeof(cmd));
410                 else if (cmdret->opcode != cmd.opcode) {
411                         log_err("fio: fragment opcode mismatch (%d != %d)\n",
412                                         cmdret->opcode, cmd.opcode);
413                         ret = 1;
414                         break;
415                 }
416
417                 if (!cmd.pdu_len)
418                         break;
419
420                 /* There's payload, get it */
421                 pdu = (char *) cmdret->payload + pdu_offset;
422                 ret = fio_recv_data(sk, pdu, cmd.pdu_len, wait);
423                 if (ret)
424                         break;
425
426                 /* Verify payload crc */
427                 crc = fio_crc16(pdu, cmd.pdu_len);
428                 if (crc != cmd.pdu_crc16) {
429                         log_err("fio: server bad crc on payload ");
430                         log_err("(got %x, wanted %x)\n", cmd.pdu_crc16, crc);
431                         ret = 1;
432                         break;
433                 }
434
435                 pdu_offset += cmd.pdu_len;
436                 if (!first)
437                         cmdret->pdu_len += cmd.pdu_len;
438                 first = 0;
439         } while (cmd.flags & FIO_NET_CMD_F_MORE);
440
441         if (ret) {
442                 free(cmdret);
443                 cmdret = NULL;
444         } else if (cmdret) {
445                 /* zero-terminate text input */
446                 if (cmdret->pdu_len) {
447                         if (cmdret->opcode == FIO_NET_CMD_TEXT) {
448                                 struct cmd_text_pdu *__pdu = (struct cmd_text_pdu *) cmdret->payload;
449                                 char *buf = (char *) __pdu->buf;
450                                 int len = le32_to_cpu(__pdu->buf_len);
451
452                                 buf[len] = '\0';
453                         } else if (cmdret->opcode == FIO_NET_CMD_JOB) {
454                                 struct cmd_job_pdu *__pdu = (struct cmd_job_pdu *) cmdret->payload;
455                                 char *buf = (char *) __pdu->buf;
456                                 int len = le32_to_cpu(__pdu->buf_len);
457
458                                 buf[len] = '\0';
459                         }
460                 }
461
462                 /* frag flag is internal */
463                 cmdret->flags &= ~FIO_NET_CMD_F_MORE;
464         }
465
466         return cmdret;
467 }
468
469 static void add_reply(uint64_t tag, struct flist_head *list)
470 {
471         struct fio_net_cmd_reply *reply;
472
473         reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
474         flist_add_tail(&reply->list, list);
475 }
476
477 static uint64_t alloc_reply(uint64_t tag, uint16_t opcode)
478 {
479         struct fio_net_cmd_reply *reply;
480
481         reply = calloc(1, sizeof(*reply));
482         INIT_FLIST_HEAD(&reply->list);
483         fio_gettime(&reply->ts, NULL);
484         reply->saved_tag = tag;
485         reply->opcode = opcode;
486
487         return (uintptr_t) reply;
488 }
489
490 static void free_reply(uint64_t tag)
491 {
492         struct fio_net_cmd_reply *reply;
493
494         reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
495         free(reply);
496 }
497
498 static void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
499 {
500         uint32_t pdu_len;
501
502         cmd->cmd_crc16 = __cpu_to_le16(fio_crc16(cmd, FIO_NET_CMD_CRC_SZ));
503
504         pdu_len = le32_to_cpu(cmd->pdu_len);
505         cmd->pdu_crc16 = __cpu_to_le16(fio_crc16(pdu, pdu_len));
506 }
507
508 static void fio_net_cmd_crc(struct fio_net_cmd *cmd)
509 {
510         fio_net_cmd_crc_pdu(cmd, cmd->payload);
511 }
512
513 int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
514                      uint64_t *tagptr, struct flist_head *list)
515 {
516         struct fio_net_cmd *cmd = NULL;
517         size_t this_len, cur_len = 0;
518         uint64_t tag;
519         int ret;
520
521         if (list) {
522                 assert(tagptr);
523                 tag = *tagptr = alloc_reply(*tagptr, opcode);
524         } else
525                 tag = tagptr ? *tagptr : 0;
526
527         do {
528                 this_len = size;
529                 if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
530                         this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
531
532                 if (!cmd || cur_len < sizeof(*cmd) + this_len) {
533                         if (cmd)
534                                 free(cmd);
535
536                         cur_len = sizeof(*cmd) + this_len;
537                         cmd = malloc(cur_len);
538                 }
539
540                 fio_init_net_cmd(cmd, opcode, buf, this_len, tag);
541
542                 if (this_len < size)
543                         cmd->flags = __cpu_to_le32(FIO_NET_CMD_F_MORE);
544
545                 fio_net_cmd_crc(cmd);
546
547                 ret = fio_send_data(fd, cmd, sizeof(*cmd) + this_len);
548                 size -= this_len;
549                 buf += this_len;
550         } while (!ret && size);
551
552         if (list) {
553                 if (ret)
554                         free_reply(tag);
555                 else
556                         add_reply(tag, list);
557         }
558
559         if (cmd)
560                 free(cmd);
561
562         return ret;
563 }
564
565 static struct sk_entry *fio_net_prep_cmd(uint16_t opcode, void *buf,
566                                          size_t size, uint64_t *tagptr,
567                                          int flags)
568 {
569         struct sk_entry *entry;
570
571         entry = smalloc(sizeof(*entry));
572         if (!entry)
573                 return NULL;
574
575         INIT_FLIST_HEAD(&entry->next);
576         entry->opcode = opcode;
577         if (flags & SK_F_COPY) {
578                 entry->buf = smalloc(size);
579                 memcpy(entry->buf, buf, size);
580         } else
581                 entry->buf = buf;
582
583         entry->size = size;
584         if (tagptr)
585                 entry->tag = *tagptr;
586         else
587                 entry->tag = 0;
588         entry->flags = flags;
589         return entry;
590 }
591
592 static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry);
593
594 static void fio_net_queue_entry(struct sk_entry *entry)
595 {
596         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
597
598         if (entry->flags & SK_F_INLINE)
599                 handle_sk_entry(sk_out, entry);
600         else {
601                 sk_lock(sk_out);
602                 flist_add_tail(&entry->list, &sk_out->list);
603                 sk_unlock(sk_out);
604
605                 fio_sem_up(&sk_out->wait);
606         }
607 }
608
609 static int fio_net_queue_cmd(uint16_t opcode, void *buf, off_t size,
610                              uint64_t *tagptr, int flags)
611 {
612         struct sk_entry *entry;
613
614         entry = fio_net_prep_cmd(opcode, buf, size, tagptr, flags);
615         if (entry) {
616                 fio_net_queue_entry(entry);
617                 return 0;
618         }
619
620         return 1;
621 }
622
623 static int fio_net_send_simple_stack_cmd(int sk, uint16_t opcode, uint64_t tag)
624 {
625         struct fio_net_cmd cmd;
626
627         fio_init_net_cmd(&cmd, opcode, NULL, 0, tag);
628         fio_net_cmd_crc(&cmd);
629
630         return fio_send_data(sk, &cmd, sizeof(cmd));
631 }
632
633 /*
634  * If 'list' is non-NULL, then allocate and store the sent command for
635  * later verification.
636  */
637 int fio_net_send_simple_cmd(int sk, uint16_t opcode, uint64_t tag,
638                             struct flist_head *list)
639 {
640         int ret;
641
642         if (list)
643                 tag = alloc_reply(tag, opcode);
644
645         ret = fio_net_send_simple_stack_cmd(sk, opcode, tag);
646         if (ret) {
647                 if (list)
648                         free_reply(tag);
649
650                 return ret;
651         }
652
653         if (list)
654                 add_reply(tag, list);
655
656         return 0;
657 }
658
659 static int fio_net_queue_quit(void)
660 {
661         dprint(FD_NET, "server: sending quit\n");
662
663         return fio_net_queue_cmd(FIO_NET_CMD_QUIT, NULL, 0, NULL, SK_F_SIMPLE);
664 }
665
666 int fio_net_send_quit(int sk)
667 {
668         dprint(FD_NET, "server: sending quit\n");
669
670         return fio_net_send_simple_cmd(sk, FIO_NET_CMD_QUIT, 0, NULL);
671 }
672
673 static int fio_net_send_ack(struct fio_net_cmd *cmd, int error, int signal)
674 {
675         struct cmd_end_pdu epdu;
676         uint64_t tag = 0;
677
678         if (cmd)
679                 tag = cmd->tag;
680
681         epdu.error = __cpu_to_le32(error);
682         epdu.signal = __cpu_to_le32(signal);
683         return fio_net_queue_cmd(FIO_NET_CMD_STOP, &epdu, sizeof(epdu), &tag, SK_F_COPY);
684 }
685
686 static int fio_net_queue_stop(int error, int signal)
687 {
688         dprint(FD_NET, "server: sending stop (%d, %d)\n", error, signal);
689         return fio_net_send_ack(NULL, error, signal);
690 }
691
692 #ifdef WIN32
693 static void fio_server_add_fork_item(struct ffi_element *element, struct flist_head *list)
694 {
695         struct fio_fork_item *ffi;
696
697         ffi = malloc(sizeof(*ffi));
698         ffi->exitval = 0;
699         ffi->signal = 0;
700         ffi->exited = 0;
701         ffi->element = *element;
702         flist_add_tail(&ffi->list, list);
703 }
704
705 static void fio_server_add_conn_pid(struct flist_head *conn_list, HANDLE hProcess)
706 {
707         struct ffi_element element = {.hProcess = hProcess, .is_thread=FALSE};
708         dprint(FD_NET, "server: forked off connection job (tid=%u)\n", (int) element.thread);
709
710         fio_server_add_fork_item(&element, conn_list);
711 }
712
713 static void fio_server_add_job_pid(struct flist_head *job_list, pthread_t thread)
714 {
715         struct ffi_element element = {.thread = thread, .is_thread=TRUE};
716         dprint(FD_NET, "server: forked off job job (tid=%u)\n", (int) element.thread);
717         fio_server_add_fork_item(&element, job_list);
718 }
719
720 static void fio_server_check_fork_item(struct fio_fork_item *ffi)
721 {
722         int ret;
723
724         if (ffi->element.is_thread) {
725
726                 ret = pthread_kill(ffi->element.thread, 0);
727                 if (ret) {
728                         int rev_val;
729                         pthread_join(ffi->element.thread, (void**) &rev_val); /*if the thread is dead, then join it to get status*/
730
731                         ffi->exitval = rev_val;
732                         if (ffi->exitval)
733                                 log_err("thread (tid=%u) exited with %x\n", (int) ffi->element.thread, (int) ffi->exitval);
734                         dprint(FD_PROCESS, "thread (tid=%u) exited with %x\n", (int) ffi->element.thread, (int) ffi->exitval);
735                         ffi->exited = 1;
736                 }
737         } else {
738                 DWORD exit_val;
739                 GetExitCodeProcess(ffi->element.hProcess, &exit_val);
740
741                 if (exit_val != STILL_ACTIVE) {
742                         dprint(FD_PROCESS, "process %u exited with %d\n", GetProcessId(ffi->element.hProcess), exit_val);
743                         ffi->exited = 1;
744                         ffi->exitval = exit_val;
745                 }
746         }
747 }
748 #else
749 static void fio_server_add_fork_item(pid_t pid, struct flist_head *list)
750 {
751         struct fio_fork_item *ffi;
752
753         ffi = malloc(sizeof(*ffi));
754         ffi->exitval = 0;
755         ffi->signal = 0;
756         ffi->exited = 0;
757         ffi->pid = pid;
758         flist_add_tail(&ffi->list, list);
759 }
760
761 static void fio_server_add_conn_pid(struct flist_head *conn_list, pid_t pid)
762 {
763         dprint(FD_NET, "server: forked off connection job (pid=%u)\n", (int) pid);
764         fio_server_add_fork_item(pid, conn_list);
765 }
766
767 static void fio_server_add_job_pid(struct flist_head *job_list, pid_t pid)
768 {
769         dprint(FD_NET, "server: forked off job job (pid=%u)\n", (int) pid);
770         fio_server_add_fork_item(pid, job_list);
771 }
772
773 static void fio_server_check_fork_item(struct fio_fork_item *ffi)
774 {
775         int ret, status;
776
777         ret = waitpid(ffi->pid, &status, WNOHANG);
778         if (ret < 0) {
779                 if (errno == ECHILD) {
780                         log_err("fio: connection pid %u disappeared\n", (int) ffi->pid);
781                         ffi->exited = 1;
782                 } else
783                         log_err("fio: waitpid: %s\n", strerror(errno));
784         } else if (ret == ffi->pid) {
785                 if (WIFSIGNALED(status)) {
786                         ffi->signal = WTERMSIG(status);
787                         ffi->exited = 1;
788                 }
789                 if (WIFEXITED(status)) {
790                         if (WEXITSTATUS(status))
791                                 ffi->exitval = WEXITSTATUS(status);
792                         ffi->exited = 1;
793                 }
794         }
795 }
796 #endif
797
798 static void fio_server_fork_item_done(struct fio_fork_item *ffi, bool stop)
799 {
800 #ifdef WIN32
801         if (ffi->element.is_thread)
802                 dprint(FD_NET, "tid %u exited, sig=%u, exitval=%d\n", (int) ffi->element.thread, ffi->signal, ffi->exitval);
803         else {
804                 dprint(FD_NET, "pid %u exited, sig=%u, exitval=%d\n", (int)  GetProcessId(ffi->element.hProcess), ffi->signal, ffi->exitval);
805                 CloseHandle(ffi->element.hProcess);
806                 ffi->element.hProcess = INVALID_HANDLE_VALUE;
807         }
808 #else
809         dprint(FD_NET, "pid %u exited, sig=%u, exitval=%d\n", (int) ffi->pid, ffi->signal, ffi->exitval);
810 #endif
811
812         /*
813          * Fold STOP and QUIT...
814          */
815         if (stop) {
816                 fio_net_queue_stop(ffi->exitval, ffi->signal);
817                 fio_net_queue_quit();
818         }
819
820         flist_del(&ffi->list);
821         free(ffi);
822 }
823
824 static void fio_server_check_fork_items(struct flist_head *list, bool stop)
825 {
826         struct flist_head *entry, *tmp;
827         struct fio_fork_item *ffi;
828
829         flist_for_each_safe(entry, tmp, list) {
830                 ffi = flist_entry(entry, struct fio_fork_item, list);
831
832                 fio_server_check_fork_item(ffi);
833
834                 if (ffi->exited)
835                         fio_server_fork_item_done(ffi, stop);
836         }
837 }
838
839 static void fio_server_check_jobs(struct flist_head *job_list)
840 {
841         fio_server_check_fork_items(job_list, true);
842 }
843
844 static void fio_server_check_conns(struct flist_head *conn_list)
845 {
846         fio_server_check_fork_items(conn_list, false);
847 }
848
849 static int handle_load_file_cmd(struct fio_net_cmd *cmd)
850 {
851         struct cmd_load_file_pdu *pdu = (struct cmd_load_file_pdu *) cmd->payload;
852         void *file_name = pdu->file;
853         struct cmd_start_pdu spdu;
854
855         dprint(FD_NET, "server: loading local file %s\n", (char *) file_name);
856
857         pdu->name_len = le16_to_cpu(pdu->name_len);
858         pdu->client_type = le16_to_cpu(pdu->client_type);
859
860         if (parse_jobs_ini(file_name, 0, 0, pdu->client_type)) {
861                 fio_net_queue_quit();
862                 return -1;
863         }
864
865         spdu.jobs = cpu_to_le32(thread_number);
866         spdu.stat_outputs = cpu_to_le32(stat_number);
867         fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
868         return 0;
869 }
870
871 #ifdef WIN32
872 static void *fio_backend_thread(void *data)
873 {
874         int ret;
875         struct sk_out *sk_out = (struct sk_out *) data;
876
877         sk_out_assign(sk_out);
878
879         ret = fio_backend(sk_out);
880         sk_out_drop();
881
882         pthread_exit((void*) (intptr_t) ret);
883         return NULL;
884 }
885 #endif
886
887 static int handle_run_cmd(struct sk_out *sk_out, struct flist_head *job_list,
888                           struct fio_net_cmd *cmd)
889 {
890         int ret;
891
892         fio_time_init();
893         set_genesis_time();
894
895 #ifdef WIN32
896         {
897                 pthread_t thread;
898                 /* both this thread and backend_thread call sk_out_assign() to double increment
899                  * the ref count.  This ensures struct is valid until both threads are done with it
900                  */
901                 sk_out_assign(sk_out);
902                 ret = pthread_create(&thread, NULL,     fio_backend_thread, sk_out);
903                 if (ret) {
904                         log_err("pthread_create: %s\n", strerror(ret));
905                         return ret;
906                 }
907
908                 fio_server_add_job_pid(job_list, thread);
909                 return ret;
910         }
911 #else
912     {
913                 pid_t pid;
914                 sk_out_assign(sk_out);
915                 pid = fork();
916                 if (pid) {
917                         fio_server_add_job_pid(job_list, pid);
918                         return 0;
919                 }
920
921                 ret = fio_backend(sk_out);
922                 free_threads_shm();
923                 sk_out_drop();
924                 _exit(ret);
925         }
926 #endif
927 }
928
929 static int handle_job_cmd(struct fio_net_cmd *cmd)
930 {
931         struct cmd_job_pdu *pdu = (struct cmd_job_pdu *) cmd->payload;
932         void *buf = pdu->buf;
933         struct cmd_start_pdu spdu;
934
935         pdu->buf_len = le32_to_cpu(pdu->buf_len);
936         pdu->client_type = le32_to_cpu(pdu->client_type);
937
938         if (parse_jobs_ini(buf, 1, 0, pdu->client_type)) {
939                 fio_net_queue_quit();
940                 return -1;
941         }
942
943         spdu.jobs = cpu_to_le32(thread_number);
944         spdu.stat_outputs = cpu_to_le32(stat_number);
945
946         fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
947         return 0;
948 }
949
950 static int handle_jobline_cmd(struct fio_net_cmd *cmd)
951 {
952         void *pdu = cmd->payload;
953         struct cmd_single_line_pdu *cslp;
954         struct cmd_line_pdu *clp;
955         unsigned long offset;
956         struct cmd_start_pdu spdu;
957         char **argv;
958         int i;
959
960         clp = pdu;
961         clp->lines = le16_to_cpu(clp->lines);
962         clp->client_type = le16_to_cpu(clp->client_type);
963         argv = malloc(clp->lines * sizeof(char *));
964         offset = sizeof(*clp);
965
966         dprint(FD_NET, "server: %d command line args\n", clp->lines);
967
968         for (i = 0; i < clp->lines; i++) {
969                 cslp = pdu + offset;
970                 argv[i] = (char *) cslp->text;
971
972                 offset += sizeof(*cslp) + le16_to_cpu(cslp->len);
973                 dprint(FD_NET, "server: %d: %s\n", i, argv[i]);
974         }
975
976         if (parse_cmd_line(clp->lines, argv, clp->client_type)) {
977                 fio_net_queue_quit();
978                 free(argv);
979                 return -1;
980         }
981
982         free(argv);
983
984         spdu.jobs = cpu_to_le32(thread_number);
985         spdu.stat_outputs = cpu_to_le32(stat_number);
986
987         fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
988         return 0;
989 }
990
991 static int handle_probe_cmd(struct fio_net_cmd *cmd)
992 {
993         struct cmd_client_probe_pdu *pdu = (struct cmd_client_probe_pdu *) cmd->payload;
994         uint64_t tag = cmd->tag;
995         struct cmd_probe_reply_pdu probe = {
996 #ifdef CONFIG_BIG_ENDIAN
997                 .bigendian      = 1,
998 #endif
999                 .os             = FIO_OS,
1000                 .arch           = FIO_ARCH,
1001                 .bpp            = sizeof(void *),
1002                 .cpus           = __cpu_to_le32(cpus_configured()),
1003         };
1004
1005         dprint(FD_NET, "server: sending probe reply\n");
1006
1007         strcpy(me, (char *) pdu->server);
1008
1009         gethostname((char *) probe.hostname, sizeof(probe.hostname));
1010         snprintf((char *) probe.fio_version, sizeof(probe.fio_version), "%s",
1011                  fio_version_string);
1012
1013         /*
1014          * If the client supports compression and we do too, then enable it
1015          */
1016         if (has_zlib && le64_to_cpu(pdu->flags) & FIO_PROBE_FLAG_ZLIB) {
1017                 probe.flags = __cpu_to_le64(FIO_PROBE_FLAG_ZLIB);
1018                 use_zlib = 1;
1019         } else {
1020                 probe.flags = 0;
1021                 use_zlib = 0;
1022         }
1023
1024         return fio_net_queue_cmd(FIO_NET_CMD_PROBE, &probe, sizeof(probe), &tag, SK_F_COPY);
1025 }
1026
1027 static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
1028 {
1029         struct jobs_eta *je;
1030         uint64_t tag = cmd->tag;
1031         size_t size;
1032         int i;
1033
1034         dprint(FD_NET, "server sending status\n");
1035
1036         /*
1037          * Fake ETA return if we don't have a local one, otherwise the client
1038          * will end up timing out waiting for a response to the ETA request
1039          */
1040         je = get_jobs_eta(true, &size);
1041         if (!je) {
1042                 size = sizeof(*je);
1043                 je = calloc(1, size);
1044         } else {
1045                 je->nr_running          = cpu_to_le32(je->nr_running);
1046                 je->nr_ramp             = cpu_to_le32(je->nr_ramp);
1047                 je->nr_pending          = cpu_to_le32(je->nr_pending);
1048                 je->nr_setting_up       = cpu_to_le32(je->nr_setting_up);
1049                 je->files_open          = cpu_to_le32(je->files_open);
1050
1051                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1052                         je->m_rate[i]   = cpu_to_le64(je->m_rate[i]);
1053                         je->t_rate[i]   = cpu_to_le64(je->t_rate[i]);
1054                         je->m_iops[i]   = cpu_to_le32(je->m_iops[i]);
1055                         je->t_iops[i]   = cpu_to_le32(je->t_iops[i]);
1056                         je->rate[i]     = cpu_to_le64(je->rate[i]);
1057                         je->iops[i]     = cpu_to_le32(je->iops[i]);
1058                 }
1059
1060                 je->elapsed_sec         = cpu_to_le64(je->elapsed_sec);
1061                 je->eta_sec             = cpu_to_le64(je->eta_sec);
1062                 je->nr_threads          = cpu_to_le32(je->nr_threads);
1063                 je->is_pow2             = cpu_to_le32(je->is_pow2);
1064                 je->unit_base           = cpu_to_le32(je->unit_base);
1065         }
1066
1067         fio_net_queue_cmd(FIO_NET_CMD_ETA, je, size, &tag, SK_F_FREE);
1068         return 0;
1069 }
1070
1071 static int send_update_job_reply(uint64_t __tag, int error)
1072 {
1073         uint64_t tag = __tag;
1074         uint32_t pdu_error;
1075
1076         pdu_error = __cpu_to_le32(error);
1077         return fio_net_queue_cmd(FIO_NET_CMD_UPDATE_JOB, &pdu_error, sizeof(pdu_error), &tag, SK_F_COPY);
1078 }
1079
1080 static int handle_update_job_cmd(struct fio_net_cmd *cmd)
1081 {
1082         struct cmd_add_job_pdu *pdu = (struct cmd_add_job_pdu *) cmd->payload;
1083         struct thread_data *td;
1084         uint32_t tnumber;
1085         int ret;
1086
1087         tnumber = le32_to_cpu(pdu->thread_number);
1088
1089         dprint(FD_NET, "server: updating options for job %u\n", tnumber);
1090
1091         if (!tnumber || tnumber > thread_number) {
1092                 send_update_job_reply(cmd->tag, ENODEV);
1093                 return 0;
1094         }
1095
1096         td = tnumber_to_td(tnumber);
1097         ret = convert_thread_options_to_cpu(&td->o, &pdu->top,
1098                         cmd->pdu_len - offsetof(struct cmd_add_job_pdu, top));
1099         send_update_job_reply(cmd->tag, ret);
1100         return 0;
1101 }
1102
1103 static int handle_trigger_cmd(struct fio_net_cmd *cmd, struct flist_head *job_list)
1104 {
1105         struct cmd_vtrigger_pdu *pdu = (struct cmd_vtrigger_pdu *) cmd->payload;
1106         char *buf = (char *) pdu->cmd;
1107         struct all_io_list *rep;
1108         size_t sz;
1109
1110         pdu->len = le16_to_cpu(pdu->len);
1111         buf[pdu->len] = '\0';
1112
1113         rep = get_all_io_list(IO_LIST_ALL, &sz);
1114         if (!rep) {
1115                 struct all_io_list state;
1116
1117                 state.threads = cpu_to_le64((uint64_t) 0);
1118                 fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, &state, sizeof(state), NULL, SK_F_COPY | SK_F_INLINE);
1119         } else
1120                 fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, rep, sz, NULL, SK_F_FREE | SK_F_INLINE);
1121
1122         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
1123         fio_server_check_jobs(job_list);
1124         exec_trigger(buf);
1125         return 0;
1126 }
1127
1128 static int handle_command(struct sk_out *sk_out, struct flist_head *job_list,
1129                           struct fio_net_cmd *cmd)
1130 {
1131         int ret;
1132
1133         dprint(FD_NET, "server: got op [%s], pdu=%u, tag=%llx\n",
1134                         fio_server_op(cmd->opcode), cmd->pdu_len,
1135                         (unsigned long long) cmd->tag);
1136
1137         switch (cmd->opcode) {
1138         case FIO_NET_CMD_QUIT:
1139                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
1140                 ret = 0;
1141                 break;
1142         case FIO_NET_CMD_EXIT:
1143                 exit_backend = true;
1144                 return -1;
1145         case FIO_NET_CMD_LOAD_FILE:
1146                 ret = handle_load_file_cmd(cmd);
1147                 break;
1148         case FIO_NET_CMD_JOB:
1149                 ret = handle_job_cmd(cmd);
1150                 break;
1151         case FIO_NET_CMD_JOBLINE:
1152                 ret = handle_jobline_cmd(cmd);
1153                 break;
1154         case FIO_NET_CMD_PROBE:
1155                 ret = handle_probe_cmd(cmd);
1156                 break;
1157         case FIO_NET_CMD_SEND_ETA:
1158                 ret = handle_send_eta_cmd(cmd);
1159                 break;
1160         case FIO_NET_CMD_RUN:
1161                 ret = handle_run_cmd(sk_out, job_list, cmd);
1162                 break;
1163         case FIO_NET_CMD_UPDATE_JOB:
1164                 ret = handle_update_job_cmd(cmd);
1165                 break;
1166         case FIO_NET_CMD_VTRIGGER:
1167                 ret = handle_trigger_cmd(cmd, job_list);
1168                 break;
1169         case FIO_NET_CMD_SENDFILE: {
1170                 struct cmd_sendfile_reply *in;
1171                 struct cmd_reply *rep;
1172
1173                 rep = (struct cmd_reply *) (uintptr_t) cmd->tag;
1174
1175                 in = (struct cmd_sendfile_reply *) cmd->payload;
1176                 in->size = le32_to_cpu(in->size);
1177                 in->error = le32_to_cpu(in->error);
1178                 if (in->error) {
1179                         ret = 1;
1180                         rep->error = in->error;
1181                 } else {
1182                         ret = 0;
1183                         rep->data = smalloc(in->size);
1184                         if (!rep->data) {
1185                                 ret = 1;
1186                                 rep->error = ENOMEM;
1187                         } else {
1188                                 rep->size = in->size;
1189                                 memcpy(rep->data, in->data, in->size);
1190                         }
1191                 }
1192                 fio_sem_up(&rep->lock);
1193                 break;
1194                 }
1195         default:
1196                 log_err("fio: unknown opcode: %s\n", fio_server_op(cmd->opcode));
1197                 ret = 1;
1198         }
1199
1200         return ret;
1201 }
1202
1203 /*
1204  * Send a command with a separate PDU, not inlined in the command
1205  */
1206 static int fio_send_cmd_ext_pdu(int sk, uint16_t opcode, const void *buf,
1207                                 off_t size, uint64_t tag, uint32_t flags)
1208 {
1209         struct fio_net_cmd cmd;
1210         struct iovec iov[2];
1211         size_t this_len;
1212         int ret;
1213
1214         iov[0].iov_base = (void *) &cmd;
1215         iov[0].iov_len = sizeof(cmd);
1216
1217         do {
1218                 uint32_t this_flags = flags;
1219
1220                 this_len = size;
1221                 if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
1222                         this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
1223
1224                 if (this_len < size)
1225                         this_flags |= FIO_NET_CMD_F_MORE;
1226
1227                 __fio_init_net_cmd(&cmd, opcode, this_len, tag);
1228                 cmd.flags = __cpu_to_le32(this_flags);
1229                 fio_net_cmd_crc_pdu(&cmd, buf);
1230
1231                 iov[1].iov_base = (void *) buf;
1232                 iov[1].iov_len = this_len;
1233
1234                 ret = fio_sendv_data(sk, iov, 2);
1235                 size -= this_len;
1236                 buf += this_len;
1237         } while (!ret && size);
1238
1239         return ret;
1240 }
1241
1242 static void finish_entry(struct sk_entry *entry)
1243 {
1244         if (entry->flags & SK_F_FREE)
1245                 free(entry->buf);
1246         else if (entry->flags & SK_F_COPY)
1247                 sfree(entry->buf);
1248
1249         sfree(entry);
1250 }
1251
1252 static void entry_set_flags(struct sk_entry *entry, struct flist_head *list,
1253                             unsigned int *flags)
1254 {
1255         if (!flist_empty(list))
1256                 *flags = FIO_NET_CMD_F_MORE;
1257         else
1258                 *flags = 0;
1259 }
1260
1261 static int send_vec_entry(struct sk_out *sk_out, struct sk_entry *first)
1262 {
1263         unsigned int flags;
1264         int ret;
1265
1266         entry_set_flags(first, &first->next, &flags);
1267
1268         ret = fio_send_cmd_ext_pdu(sk_out->sk, first->opcode, first->buf,
1269                                         first->size, first->tag, flags);
1270
1271         while (!flist_empty(&first->next)) {
1272                 struct sk_entry *next;
1273
1274                 next = flist_first_entry(&first->next, struct sk_entry, list);
1275                 flist_del_init(&next->list);
1276
1277                 entry_set_flags(next, &first->next, &flags);
1278
1279                 ret += fio_send_cmd_ext_pdu(sk_out->sk, next->opcode, next->buf,
1280                                                 next->size, next->tag, flags);
1281                 finish_entry(next);
1282         }
1283
1284         return ret;
1285 }
1286
1287 static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry)
1288 {
1289         int ret;
1290
1291         fio_sem_down(&sk_out->xmit);
1292
1293         if (entry->flags & SK_F_VEC)
1294                 ret = send_vec_entry(sk_out, entry);
1295         else if (entry->flags & SK_F_SIMPLE) {
1296                 ret = fio_net_send_simple_cmd(sk_out->sk, entry->opcode,
1297                                                 entry->tag, NULL);
1298         } else {
1299                 ret = fio_net_send_cmd(sk_out->sk, entry->opcode, entry->buf,
1300                                         entry->size, &entry->tag, NULL);
1301         }
1302
1303         fio_sem_up(&sk_out->xmit);
1304
1305         if (ret)
1306                 log_err("fio: failed handling cmd %s\n", fio_server_op(entry->opcode));
1307
1308         finish_entry(entry);
1309         return ret;
1310 }
1311
1312 static int handle_xmits(struct sk_out *sk_out)
1313 {
1314         struct sk_entry *entry;
1315         FLIST_HEAD(list);
1316         int ret = 0;
1317
1318         sk_lock(sk_out);
1319         if (flist_empty(&sk_out->list)) {
1320                 sk_unlock(sk_out);
1321                 return 0;
1322         }
1323
1324         flist_splice_init(&sk_out->list, &list);
1325         sk_unlock(sk_out);
1326
1327         while (!flist_empty(&list)) {
1328                 entry = flist_first_entry(&list, struct sk_entry, list);
1329                 flist_del(&entry->list);
1330                 ret += handle_sk_entry(sk_out, entry);
1331         }
1332
1333         return ret;
1334 }
1335
1336 static int handle_connection(struct sk_out *sk_out)
1337 {
1338         struct fio_net_cmd *cmd = NULL;
1339         FLIST_HEAD(job_list);
1340         int ret = 0;
1341
1342         reset_fio_state();
1343
1344         /* read forever */
1345         while (!exit_backend) {
1346                 struct pollfd pfd = {
1347                         .fd     = sk_out->sk,
1348                         .events = POLLIN,
1349                 };
1350
1351                 do {
1352                         int timeout = 1000;
1353
1354                         if (!flist_empty(&job_list))
1355                                 timeout = 100;
1356
1357                         handle_xmits(sk_out);
1358
1359                         ret = poll(&pfd, 1, 0);
1360                         if (ret < 0) {
1361                                 if (errno == EINTR)
1362                                         break;
1363                                 log_err("fio: poll: %s\n", strerror(errno));
1364                                 break;
1365                         } else if (!ret) {
1366                                 fio_server_check_jobs(&job_list);
1367                                 fio_sem_down_timeout(&sk_out->wait, timeout);
1368                                 continue;
1369                         }
1370
1371                         if (pfd.revents & POLLIN)
1372                                 break;
1373                         if (pfd.revents & (POLLERR|POLLHUP)) {
1374                                 ret = 1;
1375                                 break;
1376                         }
1377                 } while (!exit_backend);
1378
1379                 fio_server_check_jobs(&job_list);
1380
1381                 if (ret < 0)
1382                         break;
1383
1384                 if (pfd.revents & POLLIN)
1385                         cmd = fio_net_recv_cmd(sk_out->sk, true);
1386                 if (!cmd) {
1387                         ret = -1;
1388                         break;
1389                 }
1390
1391                 ret = handle_command(sk_out, &job_list, cmd);
1392                 if (ret)
1393                         break;
1394
1395                 free(cmd);
1396                 cmd = NULL;
1397         }
1398
1399         if (cmd)
1400                 free(cmd);
1401
1402         handle_xmits(sk_out);
1403
1404         close(sk_out->sk);
1405         sk_out->sk = -1;
1406         __sk_out_drop(sk_out);
1407         _exit(ret);
1408 }
1409
1410 /* get the address on this host bound by the input socket,
1411  * whether it is ipv6 or ipv4 */
1412
1413 static int get_my_addr_str(int sk)
1414 {
1415         struct sockaddr_in6 myaddr6 = { 0, };
1416         struct sockaddr_in myaddr4 = { 0, };
1417         struct sockaddr *sockaddr_p;
1418         char *net_addr;
1419         socklen_t len;
1420         int ret;
1421
1422         if (use_ipv6) {
1423                 len = sizeof(myaddr6);
1424                 sockaddr_p = (struct sockaddr * )&myaddr6;
1425                 net_addr = (char * )&myaddr6.sin6_addr;
1426         } else {
1427                 len = sizeof(myaddr4);
1428                 sockaddr_p = (struct sockaddr * )&myaddr4;
1429                 net_addr = (char * )&myaddr4.sin_addr;
1430         }
1431
1432         ret = getsockname(sk, sockaddr_p, &len);
1433         if (ret) {
1434                 log_err("fio: getsockname: %s\n", strerror(errno));
1435                 return -1;
1436         }
1437
1438         if (!inet_ntop(use_ipv6?AF_INET6:AF_INET, net_addr, client_sockaddr_str, INET6_ADDRSTRLEN - 1)) {
1439                 log_err("inet_ntop: failed to convert addr to string\n");
1440                 return -1;
1441         }
1442
1443         dprint(FD_NET, "fio server bound to addr %s\n", client_sockaddr_str);
1444         return 0;
1445 }
1446
1447 #ifdef WIN32
1448 static int handle_connection_process(void)
1449 {
1450         WSAPROTOCOL_INFO protocol_info;
1451         DWORD bytes_read;
1452         HANDLE hpipe;
1453         int sk;
1454         struct sk_out *sk_out;
1455         int ret;
1456         char *msg = (char *) "connected";
1457
1458         log_info("server enter accept loop.  ProcessID %d\n", GetCurrentProcessId());
1459
1460         hpipe = CreateFile(
1461                                         fio_server_pipe_name,
1462                                         GENERIC_READ | GENERIC_WRITE,
1463                                         0, NULL,
1464                                         OPEN_EXISTING,
1465                                         0, NULL);
1466
1467         if (hpipe == INVALID_HANDLE_VALUE) {
1468                 log_err("couldnt open pipe %s error %lu\n",
1469                                 fio_server_pipe_name, GetLastError());
1470                 return -1;
1471         }
1472
1473         if (!ReadFile(hpipe, &protocol_info, sizeof(protocol_info), &bytes_read, NULL)) {
1474                 log_err("couldnt read pi from pipe %s error %lu\n", fio_server_pipe_name,
1475                                 GetLastError());
1476         }
1477
1478         if (use_ipv6) /* use protocol_info to create a duplicate of parents socket */
1479                 sk = WSASocket(AF_INET6, SOCK_STREAM, 0, &protocol_info, 0, 0);
1480         else
1481                 sk = WSASocket(AF_INET,  SOCK_STREAM, 0, &protocol_info, 0, 0);
1482
1483         sk_out = scalloc(1, sizeof(*sk_out));
1484         if (!sk_out) {
1485                 CloseHandle(hpipe);
1486                 close(sk);
1487                 return -1;
1488         }
1489
1490         sk_out->sk = sk;
1491         sk_out->hProcess = INVALID_HANDLE_VALUE;
1492         INIT_FLIST_HEAD(&sk_out->list);
1493         __fio_sem_init(&sk_out->lock, FIO_SEM_UNLOCKED);
1494         __fio_sem_init(&sk_out->wait, FIO_SEM_LOCKED);
1495         __fio_sem_init(&sk_out->xmit, FIO_SEM_UNLOCKED);
1496
1497         get_my_addr_str(sk);
1498
1499         if (!WriteFile(hpipe, msg, strlen(msg), NULL, NULL)) {
1500                 log_err("couldnt write pipe\n");
1501                 close(sk);
1502                 return -1;
1503         }
1504         CloseHandle(hpipe);
1505
1506         sk_out_assign(sk_out);
1507
1508         ret = handle_connection(sk_out);
1509         __sk_out_drop(sk_out);
1510         return ret;
1511 }
1512 #endif
1513
1514 static int accept_loop(int listen_sk)
1515 {
1516         struct sockaddr_in addr;
1517         struct sockaddr_in6 addr6;
1518         socklen_t len = use_ipv6 ? sizeof(addr6) : sizeof(addr);
1519         struct pollfd pfd;
1520         int ret = 0, sk, exitval = 0;
1521         FLIST_HEAD(conn_list);
1522
1523         dprint(FD_NET, "server enter accept loop\n");
1524
1525         fio_set_fd_nonblocking(listen_sk, "server");
1526
1527         while (!exit_backend) {
1528                 struct sk_out *sk_out;
1529                 const char *from;
1530                 char buf[64];
1531 #ifdef WIN32
1532                 HANDLE hProcess;
1533 #else
1534                 pid_t pid;
1535 #endif
1536                 pfd.fd = listen_sk;
1537                 pfd.events = POLLIN;
1538                 do {
1539                         int timeout = 1000;
1540
1541                         if (!flist_empty(&conn_list))
1542                                 timeout = 100;
1543
1544                         ret = poll(&pfd, 1, timeout);
1545                         if (ret < 0) {
1546                                 if (errno == EINTR)
1547                                         break;
1548                                 log_err("fio: poll: %s\n", strerror(errno));
1549                                 break;
1550                         } else if (!ret) {
1551                                 fio_server_check_conns(&conn_list);
1552                                 continue;
1553                         }
1554
1555                         if (pfd.revents & POLLIN)
1556                                 break;
1557                 } while (!exit_backend);
1558
1559                 fio_server_check_conns(&conn_list);
1560
1561                 if (exit_backend || ret < 0)
1562                         break;
1563
1564                 if (use_ipv6)
1565                         sk = accept(listen_sk, (struct sockaddr *) &addr6, &len);
1566                 else
1567                         sk = accept(listen_sk, (struct sockaddr *) &addr, &len);
1568
1569                 if (sk < 0) {
1570                         log_err("fio: accept: %s\n", strerror(errno));
1571                         return -1;
1572                 }
1573
1574                 if (use_ipv6)
1575                         from = inet_ntop(AF_INET6, (struct sockaddr *) &addr6.sin6_addr, buf, sizeof(buf));
1576                 else
1577                         from = inet_ntop(AF_INET, (struct sockaddr *) &addr.sin_addr, buf, sizeof(buf));
1578
1579                 dprint(FD_NET, "server: connect from %s\n", from);
1580
1581                 sk_out = scalloc(1, sizeof(*sk_out));
1582                 if (!sk_out) {
1583                         close(sk);
1584                         return -1;
1585                 }
1586
1587                 sk_out->sk = sk;
1588                 INIT_FLIST_HEAD(&sk_out->list);
1589                 __fio_sem_init(&sk_out->lock, FIO_SEM_UNLOCKED);
1590                 __fio_sem_init(&sk_out->wait, FIO_SEM_LOCKED);
1591                 __fio_sem_init(&sk_out->xmit, FIO_SEM_UNLOCKED);
1592
1593 #ifdef WIN32
1594                 hProcess = windows_handle_connection(hjob, sk);
1595                 if (hProcess == INVALID_HANDLE_VALUE)
1596                         return -1;
1597                 sk_out->hProcess = hProcess;
1598                 fio_server_add_conn_pid(&conn_list, hProcess);
1599 #else
1600                 pid = fork();
1601                 if (pid) {
1602                         close(sk);
1603                         fio_server_add_conn_pid(&conn_list, pid);
1604                         continue;
1605                 }
1606
1607                 /* if error, it's already logged, non-fatal */
1608                 get_my_addr_str(sk);
1609
1610                 /*
1611                  * Assign sk_out here, it'll be dropped in handle_connection()
1612                  * since that function calls _exit() when done
1613                  */
1614                 sk_out_assign(sk_out);
1615                 handle_connection(sk_out);
1616 #endif
1617         }
1618
1619         return exitval;
1620 }
1621
1622 int fio_server_text_output(int level, const char *buf, size_t len)
1623 {
1624         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
1625         struct cmd_text_pdu *pdu;
1626         unsigned int tlen;
1627         struct timeval tv;
1628
1629         if (!sk_out || sk_out->sk == -1)
1630                 return -1;
1631
1632         tlen = sizeof(*pdu) + len;
1633         pdu = malloc(tlen);
1634
1635         pdu->level      = __cpu_to_le32(level);
1636         pdu->buf_len    = __cpu_to_le32(len);
1637
1638         gettimeofday(&tv, NULL);
1639         pdu->log_sec    = __cpu_to_le64(tv.tv_sec);
1640         pdu->log_usec   = __cpu_to_le64(tv.tv_usec);
1641
1642         memcpy(pdu->buf, buf, len);
1643
1644         fio_net_queue_cmd(FIO_NET_CMD_TEXT, pdu, tlen, NULL, SK_F_COPY);
1645         free(pdu);
1646         return len;
1647 }
1648
1649 static void convert_io_stat(struct io_stat *dst, struct io_stat *src)
1650 {
1651         dst->max_val    = cpu_to_le64(src->max_val);
1652         dst->min_val    = cpu_to_le64(src->min_val);
1653         dst->samples    = cpu_to_le64(src->samples);
1654
1655         /*
1656          * Encode to IEEE 754 for network transfer
1657          */
1658         dst->mean.u.i   = cpu_to_le64(fio_double_to_uint64(src->mean.u.f));
1659         dst->S.u.i      = cpu_to_le64(fio_double_to_uint64(src->S.u.f));
1660 }
1661
1662 static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
1663 {
1664         int i;
1665
1666         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1667                 dst->max_run[i]         = cpu_to_le64(src->max_run[i]);
1668                 dst->min_run[i]         = cpu_to_le64(src->min_run[i]);
1669                 dst->max_bw[i]          = cpu_to_le64(src->max_bw[i]);
1670                 dst->min_bw[i]          = cpu_to_le64(src->min_bw[i]);
1671                 dst->iobytes[i]         = cpu_to_le64(src->iobytes[i]);
1672                 dst->agg[i]             = cpu_to_le64(src->agg[i]);
1673         }
1674
1675         dst->kb_base    = cpu_to_le32(src->kb_base);
1676         dst->unit_base  = cpu_to_le32(src->unit_base);
1677         dst->groupid    = cpu_to_le32(src->groupid);
1678         dst->unified_rw_rep     = cpu_to_le32(src->unified_rw_rep);
1679         dst->sig_figs   = cpu_to_le32(src->sig_figs);
1680 }
1681
1682 /*
1683  * Send a CMD_TS, which packs struct thread_stat and group_run_stats
1684  * into a single payload.
1685  */
1686 void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
1687 {
1688         struct cmd_ts_pdu p;
1689         int i, j, k;
1690         size_t clat_prio_stats_extra_size = 0;
1691         size_t ss_extra_size = 0;
1692         size_t extended_buf_size = 0;
1693         void *extended_buf;
1694         void *extended_buf_wp;
1695
1696         dprint(FD_NET, "server sending end stats\n");
1697
1698         memset(&p, 0, sizeof(p));
1699
1700         snprintf(p.ts.name, sizeof(p.ts.name), "%s", ts->name);
1701         snprintf(p.ts.verror, sizeof(p.ts.verror), "%s", ts->verror);
1702         snprintf(p.ts.description, sizeof(p.ts.description), "%s",
1703                  ts->description);
1704
1705         p.ts.error              = cpu_to_le32(ts->error);
1706         p.ts.thread_number      = cpu_to_le32(ts->thread_number);
1707         p.ts.groupid            = cpu_to_le32(ts->groupid);
1708         p.ts.pid                = cpu_to_le32(ts->pid);
1709         p.ts.members            = cpu_to_le32(ts->members);
1710         p.ts.unified_rw_rep     = cpu_to_le32(ts->unified_rw_rep);
1711         p.ts.ioprio             = cpu_to_le32(ts->ioprio);
1712         p.ts.disable_prio_stat  = cpu_to_le32(ts->disable_prio_stat);
1713
1714         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1715                 convert_io_stat(&p.ts.clat_stat[i], &ts->clat_stat[i]);
1716                 convert_io_stat(&p.ts.slat_stat[i], &ts->slat_stat[i]);
1717                 convert_io_stat(&p.ts.lat_stat[i], &ts->lat_stat[i]);
1718                 convert_io_stat(&p.ts.bw_stat[i], &ts->bw_stat[i]);
1719                 convert_io_stat(&p.ts.iops_stat[i], &ts->iops_stat[i]);
1720         }
1721         convert_io_stat(&p.ts.sync_stat, &ts->sync_stat);
1722
1723         p.ts.usr_time           = cpu_to_le64(ts->usr_time);
1724         p.ts.sys_time           = cpu_to_le64(ts->sys_time);
1725         p.ts.ctx                = cpu_to_le64(ts->ctx);
1726         p.ts.minf               = cpu_to_le64(ts->minf);
1727         p.ts.majf               = cpu_to_le64(ts->majf);
1728         p.ts.clat_percentiles   = cpu_to_le32(ts->clat_percentiles);
1729         p.ts.lat_percentiles    = cpu_to_le32(ts->lat_percentiles);
1730         p.ts.slat_percentiles   = cpu_to_le32(ts->slat_percentiles);
1731         p.ts.percentile_precision = cpu_to_le64(ts->percentile_precision);
1732
1733         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1734                 fio_fp64_t *src = &ts->percentile_list[i];
1735                 fio_fp64_t *dst = &p.ts.percentile_list[i];
1736
1737                 dst->u.i = cpu_to_le64(fio_double_to_uint64(src->u.f));
1738         }
1739
1740         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1741                 p.ts.io_u_map[i]        = cpu_to_le64(ts->io_u_map[i]);
1742                 p.ts.io_u_submit[i]     = cpu_to_le64(ts->io_u_submit[i]);
1743                 p.ts.io_u_complete[i]   = cpu_to_le64(ts->io_u_complete[i]);
1744         }
1745
1746         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
1747                 p.ts.io_u_lat_n[i]      = cpu_to_le64(ts->io_u_lat_n[i]);
1748         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1749                 p.ts.io_u_lat_u[i]      = cpu_to_le64(ts->io_u_lat_u[i]);
1750         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1751                 p.ts.io_u_lat_m[i]      = cpu_to_le64(ts->io_u_lat_m[i]);
1752
1753         for (i = 0; i < FIO_LAT_CNT; i++)
1754                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
1755                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1756                                 p.ts.io_u_plat[i][j][k] = cpu_to_le64(ts->io_u_plat[i][j][k]);
1757
1758         for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1759                 p.ts.io_u_sync_plat[j] = cpu_to_le64(ts->io_u_sync_plat[j]);
1760
1761         for (i = 0; i < DDIR_RWDIR_SYNC_CNT; i++)
1762                 p.ts.total_io_u[i]      = cpu_to_le64(ts->total_io_u[i]);
1763
1764         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1765                 p.ts.short_io_u[i]      = cpu_to_le64(ts->short_io_u[i]);
1766                 p.ts.drop_io_u[i]       = cpu_to_le64(ts->drop_io_u[i]);
1767         }
1768
1769         p.ts.total_submit       = cpu_to_le64(ts->total_submit);
1770         p.ts.total_complete     = cpu_to_le64(ts->total_complete);
1771         p.ts.nr_zone_resets     = cpu_to_le64(ts->nr_zone_resets);
1772
1773         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1774                 p.ts.io_bytes[i]        = cpu_to_le64(ts->io_bytes[i]);
1775                 p.ts.runtime[i]         = cpu_to_le64(ts->runtime[i]);
1776         }
1777
1778         p.ts.total_run_time     = cpu_to_le64(ts->total_run_time);
1779         p.ts.continue_on_error  = cpu_to_le16(ts->continue_on_error);
1780         p.ts.total_err_count    = cpu_to_le64(ts->total_err_count);
1781         p.ts.first_error        = cpu_to_le32(ts->first_error);
1782         p.ts.kb_base            = cpu_to_le32(ts->kb_base);
1783         p.ts.unit_base          = cpu_to_le32(ts->unit_base);
1784
1785         p.ts.latency_depth      = cpu_to_le32(ts->latency_depth);
1786         p.ts.latency_target     = cpu_to_le64(ts->latency_target);
1787         p.ts.latency_window     = cpu_to_le64(ts->latency_window);
1788         p.ts.latency_percentile.u.i = cpu_to_le64(fio_double_to_uint64(ts->latency_percentile.u.f));
1789
1790         p.ts.sig_figs           = cpu_to_le32(ts->sig_figs);
1791
1792         p.ts.nr_block_infos     = cpu_to_le64(ts->nr_block_infos);
1793         for (i = 0; i < p.ts.nr_block_infos; i++)
1794                 p.ts.block_infos[i] = cpu_to_le32(ts->block_infos[i]);
1795
1796         p.ts.ss_dur             = cpu_to_le64(ts->ss_dur);
1797         p.ts.ss_state           = cpu_to_le32(ts->ss_state);
1798         p.ts.ss_head            = cpu_to_le32(ts->ss_head);
1799         p.ts.ss_limit.u.i       = cpu_to_le64(fio_double_to_uint64(ts->ss_limit.u.f));
1800         p.ts.ss_slope.u.i       = cpu_to_le64(fio_double_to_uint64(ts->ss_slope.u.f));
1801         p.ts.ss_deviation.u.i   = cpu_to_le64(fio_double_to_uint64(ts->ss_deviation.u.f));
1802         p.ts.ss_criterion.u.i   = cpu_to_le64(fio_double_to_uint64(ts->ss_criterion.u.f));
1803
1804         p.ts.cachehit           = cpu_to_le64(ts->cachehit);
1805         p.ts.cachemiss          = cpu_to_le64(ts->cachemiss);
1806
1807         convert_gs(&p.rs, rs);
1808
1809         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1810                 if (ts->nr_clat_prio[i])
1811                         clat_prio_stats_extra_size += ts->nr_clat_prio[i] * sizeof(*ts->clat_prio[i]);
1812         }
1813         extended_buf_size += clat_prio_stats_extra_size;
1814
1815         dprint(FD_NET, "ts->ss_state = %d\n", ts->ss_state);
1816         if (ts->ss_state & FIO_SS_DATA)
1817                 ss_extra_size = 2 * ts->ss_dur * sizeof(uint64_t);
1818
1819         extended_buf_size += ss_extra_size;
1820         if (!extended_buf_size) {
1821                 fio_net_queue_cmd(FIO_NET_CMD_TS, &p, sizeof(p), NULL, SK_F_COPY);
1822                 return;
1823         }
1824
1825         extended_buf_size += sizeof(p);
1826         extended_buf = calloc(1, extended_buf_size);
1827         if (!extended_buf) {
1828                 log_err("fio: failed to allocate FIO_NET_CMD_TS buffer\n");
1829                 return;
1830         }
1831
1832         memcpy(extended_buf, &p, sizeof(p));
1833         extended_buf_wp = (struct cmd_ts_pdu *)extended_buf + 1;
1834
1835         if (clat_prio_stats_extra_size) {
1836                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1837                         struct clat_prio_stat *prio = (struct clat_prio_stat *) extended_buf_wp;
1838
1839                         for (j = 0; j < ts->nr_clat_prio[i]; j++) {
1840                                 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1841                                         prio->io_u_plat[k] =
1842                                                 cpu_to_le64(ts->clat_prio[i][j].io_u_plat[k]);
1843                                 convert_io_stat(&prio->clat_stat,
1844                                                 &ts->clat_prio[i][j].clat_stat);
1845                                 prio->ioprio = cpu_to_le32(ts->clat_prio[i][j].ioprio);
1846                                 prio++;
1847                         }
1848
1849                         if (ts->nr_clat_prio[i]) {
1850                                 uint64_t offset = (char *)extended_buf_wp - (char *)extended_buf;
1851                                 struct cmd_ts_pdu *ptr = extended_buf;
1852
1853                                 ptr->ts.clat_prio_offset[i] = cpu_to_le64(offset);
1854                                 ptr->ts.nr_clat_prio[i] = cpu_to_le32(ts->nr_clat_prio[i]);
1855                         }
1856
1857                         extended_buf_wp = prio;
1858                 }
1859         }
1860
1861         if (ss_extra_size) {
1862                 uint64_t *ss_iops, *ss_bw;
1863                 uint64_t offset;
1864                 struct cmd_ts_pdu *ptr = extended_buf;
1865
1866                 dprint(FD_NET, "server sending steadystate ring buffers\n");
1867
1868                 /* ss iops */
1869                 ss_iops = (uint64_t *) extended_buf_wp;
1870                 for (i = 0; i < ts->ss_dur; i++)
1871                         ss_iops[i] = cpu_to_le64(ts->ss_iops_data[i]);
1872
1873                 offset = (char *)extended_buf_wp - (char *)extended_buf;
1874                 ptr->ts.ss_iops_data_offset = cpu_to_le64(offset);
1875                 extended_buf_wp = ss_iops + (int) ts->ss_dur;
1876
1877                 /* ss bw */
1878                 ss_bw = extended_buf_wp;
1879                 for (i = 0; i < ts->ss_dur; i++)
1880                         ss_bw[i] = cpu_to_le64(ts->ss_bw_data[i]);
1881
1882                 offset = (char *)extended_buf_wp - (char *)extended_buf;
1883                 ptr->ts.ss_bw_data_offset = cpu_to_le64(offset);
1884                 extended_buf_wp = ss_bw + (int) ts->ss_dur;
1885         }
1886
1887         fio_net_queue_cmd(FIO_NET_CMD_TS, extended_buf, extended_buf_size, NULL, SK_F_COPY);
1888         free(extended_buf);
1889 }
1890
1891 void fio_server_send_gs(struct group_run_stats *rs)
1892 {
1893         struct group_run_stats gs;
1894
1895         dprint(FD_NET, "server sending group run stats\n");
1896
1897         convert_gs(&gs, rs);
1898         fio_net_queue_cmd(FIO_NET_CMD_GS, &gs, sizeof(gs), NULL, SK_F_COPY);
1899 }
1900
1901 void fio_server_send_job_options(struct flist_head *opt_list,
1902                                  unsigned int gid)
1903 {
1904         struct cmd_job_option pdu;
1905         struct flist_head *entry;
1906
1907         if (flist_empty(opt_list))
1908                 return;
1909
1910         flist_for_each(entry, opt_list) {
1911                 struct print_option *p;
1912                 size_t len;
1913
1914                 p = flist_entry(entry, struct print_option, list);
1915                 memset(&pdu, 0, sizeof(pdu));
1916
1917                 if (gid == -1U) {
1918                         pdu.global = __cpu_to_le16(1);
1919                         pdu.groupid = 0;
1920                 } else {
1921                         pdu.global = 0;
1922                         pdu.groupid = cpu_to_le32(gid);
1923                 }
1924                 len = strlen(p->name);
1925                 if (len >= sizeof(pdu.name)) {
1926                         len = sizeof(pdu.name) - 1;
1927                         pdu.truncated = __cpu_to_le16(1);
1928                 }
1929                 memcpy(pdu.name, p->name, len);
1930                 if (p->value) {
1931                         len = strlen(p->value);
1932                         if (len >= sizeof(pdu.value)) {
1933                                 len = sizeof(pdu.value) - 1;
1934                                 pdu.truncated = __cpu_to_le16(1);
1935                         }
1936                         memcpy(pdu.value, p->value, len);
1937                 }
1938                 fio_net_queue_cmd(FIO_NET_CMD_JOB_OPT, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1939         }
1940 }
1941
1942 static void convert_agg(struct disk_util_agg *dst, struct disk_util_agg *src)
1943 {
1944         int i;
1945
1946         for (i = 0; i < 2; i++) {
1947                 dst->ios[i]     = cpu_to_le64(src->ios[i]);
1948                 dst->merges[i]  = cpu_to_le64(src->merges[i]);
1949                 dst->sectors[i] = cpu_to_le64(src->sectors[i]);
1950                 dst->ticks[i]   = cpu_to_le64(src->ticks[i]);
1951         }
1952
1953         dst->io_ticks           = cpu_to_le64(src->io_ticks);
1954         dst->time_in_queue      = cpu_to_le64(src->time_in_queue);
1955         dst->slavecount         = cpu_to_le32(src->slavecount);
1956         dst->max_util.u.i       = cpu_to_le64(fio_double_to_uint64(src->max_util.u.f));
1957 }
1958
1959 static void convert_dus(struct disk_util_stat *dst, struct disk_util_stat *src)
1960 {
1961         int i;
1962
1963         snprintf((char *) dst->name, sizeof(dst->name), "%s", src->name);
1964
1965         for (i = 0; i < 2; i++) {
1966                 dst->s.ios[i]           = cpu_to_le64(src->s.ios[i]);
1967                 dst->s.merges[i]        = cpu_to_le64(src->s.merges[i]);
1968                 dst->s.sectors[i]       = cpu_to_le64(src->s.sectors[i]);
1969                 dst->s.ticks[i]         = cpu_to_le64(src->s.ticks[i]);
1970         }
1971
1972         dst->s.io_ticks         = cpu_to_le64(src->s.io_ticks);
1973         dst->s.time_in_queue    = cpu_to_le64(src->s.time_in_queue);
1974         dst->s.msec             = cpu_to_le64(src->s.msec);
1975 }
1976
1977 void fio_server_send_du(void)
1978 {
1979         struct disk_util *du;
1980         struct flist_head *entry;
1981         struct cmd_du_pdu pdu;
1982
1983         dprint(FD_NET, "server: sending disk_util %d\n", !flist_empty(&disk_list));
1984
1985         memset(&pdu, 0, sizeof(pdu));
1986
1987         flist_for_each(entry, &disk_list) {
1988                 du = flist_entry(entry, struct disk_util, list);
1989
1990                 convert_dus(&pdu.dus, &du->dus);
1991                 convert_agg(&pdu.agg, &du->agg);
1992
1993                 fio_net_queue_cmd(FIO_NET_CMD_DU, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1994         }
1995 }
1996
1997 #ifdef CONFIG_ZLIB
1998
1999 static inline void __fio_net_prep_tail(z_stream *stream, void *out_pdu,
2000                                         struct sk_entry **last_entry,
2001                                         struct sk_entry *first)
2002 {
2003         unsigned int this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
2004
2005         *last_entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
2006                                  NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
2007         if (*last_entry)
2008                 flist_add_tail(&(*last_entry)->list, &first->next);
2009 }
2010
2011 /*
2012  * Deflates the next input given, creating as many new packets in the
2013  * linked list as necessary.
2014  */
2015 static int __deflate_pdu_buffer(void *next_in, unsigned int next_sz, void **out_pdu,
2016                                 struct sk_entry **last_entry, z_stream *stream,
2017                                 struct sk_entry *first)
2018 {
2019         int ret;
2020
2021         stream->next_in = next_in;
2022         stream->avail_in = next_sz;
2023         do {
2024                 if (!stream->avail_out) {
2025                         __fio_net_prep_tail(stream, *out_pdu, last_entry, first);
2026                         if (*last_entry == NULL)
2027                                 return 1;
2028
2029                         *out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2030
2031                         stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2032                         stream->next_out = *out_pdu;
2033                 }
2034
2035                 ret = deflate(stream, Z_BLOCK);
2036
2037                 if (ret < 0) {
2038                         free(*out_pdu);
2039                         return 1;
2040                 }
2041         } while (stream->avail_in);
2042
2043         return 0;
2044 }
2045
2046 static int __fio_append_iolog_gz_hist(struct sk_entry *first, struct io_log *log,
2047                                       struct io_logs *cur_log, z_stream *stream)
2048 {
2049         struct sk_entry *entry;
2050         void *out_pdu;
2051         int ret, i, j;
2052         int sample_sz = log_entry_sz(log);
2053
2054         out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2055         stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2056         stream->next_out = out_pdu;
2057
2058         for (i = 0; i < cur_log->nr_samples; i++) {
2059                 struct io_sample *s;
2060                 struct io_u_plat_entry *cur_plat_entry, *prev_plat_entry;
2061                 uint64_t *cur_plat, *prev_plat;
2062
2063                 s = get_sample(log, cur_log, i);
2064                 ret = __deflate_pdu_buffer(s, sample_sz, &out_pdu, &entry, stream, first);
2065                 if (ret)
2066                         return ret;
2067
2068                 /* Do the subtraction on server side so that client doesn't have to
2069                  * reconstruct our linked list from packets.
2070                  */
2071                 cur_plat_entry  = s->data.plat_entry;
2072                 prev_plat_entry = flist_first_entry(&cur_plat_entry->list, struct io_u_plat_entry, list);
2073                 cur_plat  = cur_plat_entry->io_u_plat;
2074                 prev_plat = prev_plat_entry->io_u_plat;
2075
2076                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2077                         cur_plat[j] -= prev_plat[j];
2078                 }
2079
2080                 flist_del(&prev_plat_entry->list);
2081                 free(prev_plat_entry);
2082
2083                 ret = __deflate_pdu_buffer(cur_plat_entry, sizeof(*cur_plat_entry),
2084                                            &out_pdu, &entry, stream, first);
2085
2086                 if (ret)
2087                         return ret;
2088         }
2089
2090         __fio_net_prep_tail(stream, out_pdu, &entry, first);
2091         return entry == NULL;
2092 }
2093
2094 static int __fio_append_iolog_gz(struct sk_entry *first, struct io_log *log,
2095                                  struct io_logs *cur_log, z_stream *stream)
2096 {
2097         unsigned int this_len;
2098         void *out_pdu;
2099         int ret;
2100
2101         if (log->log_type == IO_LOG_TYPE_HIST)
2102                 return __fio_append_iolog_gz_hist(first, log, cur_log, stream);
2103
2104         stream->next_in = (void *) cur_log->log;
2105         stream->avail_in = cur_log->nr_samples * log_entry_sz(log);
2106
2107         do {
2108                 struct sk_entry *entry;
2109
2110                 /*
2111                  * Dirty - since the log is potentially huge, compress it into
2112                  * FIO_SERVER_MAX_FRAGMENT_PDU chunks and let the receiving
2113                  * side defragment it.
2114                  */
2115                 out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2116
2117                 stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2118                 stream->next_out = out_pdu;
2119                 ret = deflate(stream, Z_BLOCK);
2120                 /* may be Z_OK, or Z_STREAM_END */
2121                 if (ret < 0) {
2122                         free(out_pdu);
2123                         return 1;
2124                 }
2125
2126                 this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
2127
2128                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
2129                                          NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
2130                 if (!entry) {
2131                         free(out_pdu);
2132                         return 1;
2133                 }
2134                 flist_add_tail(&entry->list, &first->next);
2135         } while (stream->avail_in);
2136
2137         return 0;
2138 }
2139
2140 static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
2141 {
2142         z_stream stream = {
2143                 .zalloc = Z_NULL,
2144                 .zfree  = Z_NULL,
2145                 .opaque = Z_NULL,
2146         };
2147         int ret = 0;
2148
2149         if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK)
2150                 return 1;
2151
2152         while (!flist_empty(&log->io_logs)) {
2153                 struct io_logs *cur_log;
2154
2155                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
2156                 flist_del_init(&cur_log->list);
2157
2158                 ret = __fio_append_iolog_gz(first, log, cur_log, &stream);
2159                 if (ret)
2160                         break;
2161         }
2162
2163         ret = deflate(&stream, Z_FINISH);
2164
2165         while (ret != Z_STREAM_END) {
2166                 struct sk_entry *entry;
2167                 unsigned int this_len;
2168                 void *out_pdu;
2169
2170                 out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2171                 stream.avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2172                 stream.next_out = out_pdu;
2173
2174                 ret = deflate(&stream, Z_FINISH);
2175                 /* may be Z_OK, or Z_STREAM_END */
2176                 if (ret < 0) {
2177                         free(out_pdu);
2178                         break;
2179                 }
2180
2181                 this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream.avail_out;
2182
2183                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
2184                                          NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
2185                 if (!entry) {
2186                         free(out_pdu);
2187                         break;
2188                 }
2189                 flist_add_tail(&entry->list, &first->next);
2190         }
2191
2192         ret = deflateEnd(&stream);
2193         if (ret == Z_OK)
2194                 return 0;
2195
2196         return 1;
2197 }
2198 #else
2199 static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
2200 {
2201         return 1;
2202 }
2203 #endif
2204
2205 static int fio_append_gz_chunks(struct sk_entry *first, struct io_log *log)
2206 {
2207         struct sk_entry *entry;
2208         struct flist_head *node;
2209         int ret = 0;
2210
2211         pthread_mutex_lock(&log->chunk_lock);
2212         flist_for_each(node, &log->chunk_list) {
2213                 struct iolog_compress *c;
2214
2215                 c = flist_entry(node, struct iolog_compress, list);
2216                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, c->buf, c->len,
2217                                                 NULL, SK_F_VEC | SK_F_INLINE);
2218                 if (!entry) {
2219                         ret = 1;
2220                         break;
2221                 }
2222                 flist_add_tail(&entry->list, &first->next);
2223         }
2224         pthread_mutex_unlock(&log->chunk_lock);
2225         return ret;
2226 }
2227
2228 static int fio_append_text_log(struct sk_entry *first, struct io_log *log)
2229 {
2230         struct sk_entry *entry;
2231         int ret = 0;
2232
2233         while (!flist_empty(&log->io_logs)) {
2234                 struct io_logs *cur_log;
2235                 size_t size;
2236
2237                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
2238                 flist_del_init(&cur_log->list);
2239
2240                 size = cur_log->nr_samples * log_entry_sz(log);
2241
2242                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, cur_log->log, size,
2243                                                 NULL, SK_F_VEC | SK_F_INLINE);
2244                 if (!entry) {
2245                         ret = 1;
2246                         break;
2247                 }
2248                 flist_add_tail(&entry->list, &first->next);
2249         }
2250
2251         return ret;
2252 }
2253
2254 int fio_send_iolog(struct thread_data *td, struct io_log *log, const char *name)
2255 {
2256         struct cmd_iolog_pdu pdu = {
2257                 .nr_samples             = cpu_to_le64(iolog_nr_samples(log)),
2258                 .thread_number          = cpu_to_le32(td->thread_number),
2259                 .log_type               = cpu_to_le32(log->log_type),
2260                 .log_hist_coarseness    = cpu_to_le32(log->hist_coarseness),
2261         };
2262         struct sk_entry *first;
2263         struct flist_head *entry;
2264         int ret = 0;
2265
2266         if (!flist_empty(&log->chunk_list))
2267                 pdu.compressed = __cpu_to_le32(STORE_COMPRESSED);
2268         else if (use_zlib)
2269                 pdu.compressed = __cpu_to_le32(XMIT_COMPRESSED);
2270         else
2271                 pdu.compressed = 0;
2272
2273         snprintf((char *) pdu.name, sizeof(pdu.name), "%s", name);
2274
2275         /*
2276          * We can't do this for a pre-compressed log, but for that case,
2277          * log->nr_samples is zero anyway.
2278          */
2279         flist_for_each(entry, &log->io_logs) {
2280                 struct io_logs *cur_log;
2281                 int i;
2282
2283                 cur_log = flist_entry(entry, struct io_logs, list);
2284
2285                 for (i = 0; i < cur_log->nr_samples; i++) {
2286                         struct io_sample *s = get_sample(log, cur_log, i);
2287
2288                         s->time         = cpu_to_le64(s->time);
2289                         if (log->log_type != IO_LOG_TYPE_HIST)
2290                                 s->data.val     = cpu_to_le64(s->data.val);
2291                         s->__ddir       = __cpu_to_le32(s->__ddir);
2292                         s->bs           = cpu_to_le64(s->bs);
2293
2294                         if (log->log_offset) {
2295                                 struct io_sample_offset *so = (void *) s;
2296
2297                                 so->offset = cpu_to_le64(so->offset);
2298                         }
2299                 }
2300         }
2301
2302         /*
2303          * Assemble header entry first
2304          */
2305         first = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, &pdu, sizeof(pdu), NULL, SK_F_VEC | SK_F_INLINE | SK_F_COPY);
2306         if (!first)
2307                 return 1;
2308
2309         /*
2310          * Now append actual log entries. If log compression was enabled on
2311          * the job, just send out the compressed chunks directly. If we
2312          * have a plain log, compress if we can, then send. Otherwise, send
2313          * the plain text output.
2314          */
2315         if (!flist_empty(&log->chunk_list))
2316                 ret = fio_append_gz_chunks(first, log);
2317         else if (use_zlib)
2318                 ret = fio_append_iolog_gz(first, log);
2319         else
2320                 ret = fio_append_text_log(first, log);
2321
2322         fio_net_queue_entry(first);
2323         return ret;
2324 }
2325
2326 void fio_server_send_add_job(struct thread_data *td)
2327 {
2328         struct cmd_add_job_pdu *pdu;
2329         size_t cmd_sz = offsetof(struct cmd_add_job_pdu, top) +
2330                 thread_options_pack_size(&td->o);
2331
2332         pdu = malloc(cmd_sz);
2333         pdu->thread_number = cpu_to_le32(td->thread_number);
2334         pdu->groupid = cpu_to_le32(td->groupid);
2335
2336         convert_thread_options_to_net(&pdu->top, &td->o);
2337
2338         fio_net_queue_cmd(FIO_NET_CMD_ADD_JOB, pdu, cmd_sz, NULL, SK_F_COPY);
2339         free(pdu);
2340 }
2341
2342 void fio_server_send_start(struct thread_data *td)
2343 {
2344         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2345
2346         assert(sk_out->sk != -1);
2347
2348         fio_net_queue_cmd(FIO_NET_CMD_SERVER_START, NULL, 0, NULL, SK_F_SIMPLE);
2349 }
2350
2351 int fio_server_get_verify_state(const char *name, int threadnumber,
2352                                 void **datap)
2353 {
2354         struct thread_io_list *s;
2355         struct cmd_sendfile out;
2356         struct cmd_reply *rep;
2357         uint64_t tag;
2358         void *data;
2359         int ret;
2360
2361         dprint(FD_NET, "server: request verify state\n");
2362
2363         rep = smalloc(sizeof(*rep));
2364         if (!rep)
2365                 return ENOMEM;
2366
2367         __fio_sem_init(&rep->lock, FIO_SEM_LOCKED);
2368         rep->data = NULL;
2369         rep->error = 0;
2370
2371         verify_state_gen_name((char *) out.path, sizeof(out.path), name, me,
2372                                 threadnumber);
2373         tag = (uint64_t) (uintptr_t) rep;
2374         fio_net_queue_cmd(FIO_NET_CMD_SENDFILE, &out, sizeof(out), &tag,
2375                                 SK_F_COPY);
2376
2377         /*
2378          * Wait for the backend to receive the reply
2379          */
2380         if (fio_sem_down_timeout(&rep->lock, 10000)) {
2381                 log_err("fio: timed out waiting for reply\n");
2382                 ret = ETIMEDOUT;
2383                 goto fail;
2384         }
2385
2386         if (rep->error) {
2387                 log_err("fio: failure on receiving state file %s: %s\n",
2388                                 out.path, strerror(rep->error));
2389                 ret = rep->error;
2390 fail:
2391                 *datap = NULL;
2392                 sfree(rep);
2393                 fio_net_queue_quit();
2394                 return ret;
2395         }
2396
2397         /*
2398          * The format is verify_state_hdr, then thread_io_list. Verify
2399          * the header, and the thread_io_list checksum
2400          */
2401         s = rep->data + sizeof(struct verify_state_hdr);
2402         if (verify_state_hdr(rep->data, s)) {
2403                 ret = EILSEQ;
2404                 goto fail;
2405         }
2406
2407         /*
2408          * Don't need the header from now, copy just the thread_io_list
2409          */
2410         ret = 0;
2411         rep->size -= sizeof(struct verify_state_hdr);
2412         data = malloc(rep->size);
2413         memcpy(data, s, rep->size);
2414         *datap = data;
2415
2416         sfree(rep->data);
2417         __fio_sem_remove(&rep->lock);
2418         sfree(rep);
2419         return ret;
2420 }
2421
2422 static int fio_init_server_ip(void)
2423 {
2424         struct sockaddr *addr;
2425         socklen_t socklen;
2426         char buf[80];
2427         const char *str;
2428         int sk, opt;
2429
2430         if (use_ipv6)
2431                 sk = socket(AF_INET6, SOCK_STREAM, 0);
2432         else
2433                 sk = socket(AF_INET, SOCK_STREAM, 0);
2434
2435         if (sk < 0) {
2436                 log_err("fio: socket: %s\n", strerror(errno));
2437                 return -1;
2438         }
2439
2440         opt = 1;
2441         if (setsockopt(sk, SOL_SOCKET, SO_REUSEADDR, (void *)&opt, sizeof(opt)) < 0) {
2442                 log_err("fio: setsockopt(REUSEADDR): %s\n", strerror(errno));
2443                 close(sk);
2444                 return -1;
2445         }
2446 #ifdef SO_REUSEPORT
2447         /*
2448          * Not fatal if fails, so just ignore it if that happens
2449          */
2450         if (setsockopt(sk, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt))) {
2451         }
2452 #endif
2453
2454         if (use_ipv6) {
2455                 void *src = &saddr_in6.sin6_addr;
2456
2457                 addr = (struct sockaddr *) &saddr_in6;
2458                 socklen = sizeof(saddr_in6);
2459                 saddr_in6.sin6_family = AF_INET6;
2460                 str = inet_ntop(AF_INET6, src, buf, sizeof(buf));
2461         } else {
2462                 void *src = &saddr_in.sin_addr;
2463
2464                 addr = (struct sockaddr *) &saddr_in;
2465                 socklen = sizeof(saddr_in);
2466                 saddr_in.sin_family = AF_INET;
2467                 str = inet_ntop(AF_INET, src, buf, sizeof(buf));
2468         }
2469
2470         if (bind(sk, addr, socklen) < 0) {
2471                 log_err("fio: bind: %s\n", strerror(errno));
2472                 log_info("fio: failed with IPv%c %s\n", use_ipv6 ? '6' : '4', str);
2473                 close(sk);
2474                 return -1;
2475         }
2476
2477         return sk;
2478 }
2479
2480 static int fio_init_server_sock(void)
2481 {
2482         struct sockaddr_un addr;
2483         socklen_t len;
2484         mode_t mode;
2485         int sk;
2486
2487         sk = socket(AF_UNIX, SOCK_STREAM, 0);
2488         if (sk < 0) {
2489                 log_err("fio: socket: %s\n", strerror(errno));
2490                 return -1;
2491         }
2492
2493         mode = umask(000);
2494
2495         addr.sun_family = AF_UNIX;
2496         snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", bind_sock);
2497
2498         len = sizeof(addr.sun_family) + strlen(bind_sock) + 1;
2499
2500         if (bind(sk, (struct sockaddr *) &addr, len) < 0) {
2501                 log_err("fio: bind: %s\n", strerror(errno));
2502                 close(sk);
2503                 return -1;
2504         }
2505
2506         umask(mode);
2507         return sk;
2508 }
2509
2510 static int fio_init_server_connection(void)
2511 {
2512         char bind_str[128];
2513         int sk;
2514
2515         dprint(FD_NET, "starting server\n");
2516
2517         if (!bind_sock)
2518                 sk = fio_init_server_ip();
2519         else
2520                 sk = fio_init_server_sock();
2521
2522         if (sk < 0)
2523                 return sk;
2524
2525         memset(bind_str, 0, sizeof(bind_str));
2526
2527         if (!bind_sock) {
2528                 char *p, port[16];
2529                 void *src;
2530                 int af;
2531
2532                 if (use_ipv6) {
2533                         af = AF_INET6;
2534                         src = &saddr_in6.sin6_addr;
2535                 } else {
2536                         af = AF_INET;
2537                         src = &saddr_in.sin_addr;
2538                 }
2539
2540                 p = (char *) inet_ntop(af, src, bind_str, sizeof(bind_str));
2541
2542                 sprintf(port, ",%u", fio_net_port);
2543                 if (p)
2544                         strcat(p, port);
2545                 else
2546                         snprintf(bind_str, sizeof(bind_str), "%s", port);
2547         } else
2548                 snprintf(bind_str, sizeof(bind_str), "%s", bind_sock);
2549
2550         log_info("fio: server listening on %s\n", bind_str);
2551
2552         if (listen(sk, 4) < 0) {
2553                 log_err("fio: listen: %s\n", strerror(errno));
2554                 close(sk);
2555                 return -1;
2556         }
2557
2558         return sk;
2559 }
2560
2561 int fio_server_parse_host(const char *host, int ipv6, struct in_addr *inp,
2562                           struct in6_addr *inp6)
2563
2564 {
2565         int ret = 0;
2566
2567         if (ipv6)
2568                 ret = inet_pton(AF_INET6, host, inp6);
2569         else
2570                 ret = inet_pton(AF_INET, host, inp);
2571
2572         if (ret != 1) {
2573                 struct addrinfo *res, hints = {
2574                         .ai_family = ipv6 ? AF_INET6 : AF_INET,
2575                         .ai_socktype = SOCK_STREAM,
2576                 };
2577
2578                 ret = getaddrinfo(host, NULL, &hints, &res);
2579                 if (ret) {
2580                         log_err("fio: failed to resolve <%s> (%s)\n", host,
2581                                         gai_strerror(ret));
2582                         return 1;
2583                 }
2584
2585                 if (ipv6)
2586                         memcpy(inp6, &((struct sockaddr_in6 *) res->ai_addr)->sin6_addr, sizeof(*inp6));
2587                 else
2588                         memcpy(inp, &((struct sockaddr_in *) res->ai_addr)->sin_addr, sizeof(*inp));
2589
2590                 ret = 1;
2591                 freeaddrinfo(res);
2592         }
2593
2594         return !(ret == 1);
2595 }
2596
2597 /*
2598  * Parse a host/ip/port string. Reads from 'str'.
2599  *
2600  * Outputs:
2601  *
2602  * For IPv4:
2603  *      *ptr is the host, *port is the port, inp is the destination.
2604  * For IPv6:
2605  *      *ptr is the host, *port is the port, inp6 is the dest, and *ipv6 is 1.
2606  * For local domain sockets:
2607  *      *ptr is the filename, *is_sock is 1.
2608  */
2609 int fio_server_parse_string(const char *str, char **ptr, bool *is_sock,
2610                             int *port, struct in_addr *inp,
2611                             struct in6_addr *inp6, int *ipv6)
2612 {
2613         const char *host = str;
2614         char *portp;
2615         int lport = 0;
2616
2617         *ptr = NULL;
2618         *is_sock = false;
2619         *port = fio_net_port;
2620         *ipv6 = 0;
2621
2622         if (!strncmp(str, "sock:", 5)) {
2623                 *ptr = strdup(str + 5);
2624                 *is_sock = true;
2625
2626                 return 0;
2627         }
2628
2629         /*
2630          * Is it ip:<ip or host>:port
2631          */
2632         if (!strncmp(host, "ip:", 3))
2633                 host += 3;
2634         else if (!strncmp(host, "ip4:", 4))
2635                 host += 4;
2636         else if (!strncmp(host, "ip6:", 4)) {
2637                 host += 4;
2638                 *ipv6 = 1;
2639         } else if (host[0] == ':') {
2640                 /* String is :port */
2641                 host++;
2642                 lport = atoi(host);
2643                 if (!lport || lport > 65535) {
2644                         log_err("fio: bad server port %u\n", lport);
2645                         return 1;
2646                 }
2647                 /* no hostname given, we are done */
2648                 *port = lport;
2649                 return 0;
2650         }
2651
2652         /*
2653          * If no port seen yet, check if there's a last ',' at the end
2654          */
2655         if (!lport) {
2656                 portp = strchr(host, ',');
2657                 if (portp) {
2658                         *portp = '\0';
2659                         portp++;
2660                         lport = atoi(portp);
2661                         if (!lport || lport > 65535) {
2662                                 log_err("fio: bad server port %u\n", lport);
2663                                 return 1;
2664                         }
2665                 }
2666         }
2667
2668         if (lport)
2669                 *port = lport;
2670
2671         if (!strlen(host))
2672                 return 0;
2673
2674         *ptr = strdup(host);
2675
2676         if (fio_server_parse_host(*ptr, *ipv6, inp, inp6)) {
2677                 free(*ptr);
2678                 *ptr = NULL;
2679                 return 1;
2680         }
2681
2682         if (*port == 0)
2683                 *port = fio_net_port;
2684
2685         return 0;
2686 }
2687
2688 /*
2689  * Server arg should be one of:
2690  *
2691  * sock:/path/to/socket
2692  *   ip:1.2.3.4
2693  *      1.2.3.4
2694  *
2695  * Where sock uses unix domain sockets, and ip binds the server to
2696  * a specific interface. If no arguments are given to the server, it
2697  * uses IP and binds to 0.0.0.0.
2698  *
2699  */
2700 static int fio_handle_server_arg(void)
2701 {
2702         int port = fio_net_port;
2703         bool is_sock;
2704         int ret = 0;
2705
2706         saddr_in.sin_addr.s_addr = htonl(INADDR_ANY);
2707
2708         if (!fio_server_arg)
2709                 goto out;
2710
2711         ret = fio_server_parse_string(fio_server_arg, &bind_sock, &is_sock,
2712                                         &port, &saddr_in.sin_addr,
2713                                         &saddr_in6.sin6_addr, &use_ipv6);
2714
2715         if (!is_sock && bind_sock) {
2716                 free(bind_sock);
2717                 bind_sock = NULL;
2718         }
2719
2720 out:
2721         fio_net_port = port;
2722         saddr_in.sin_port = htons(port);
2723         saddr_in6.sin6_port = htons(port);
2724         return ret;
2725 }
2726
2727 static void sig_int(int sig)
2728 {
2729         if (bind_sock)
2730                 unlink(bind_sock);
2731 }
2732
2733 static void set_sig_handlers(void)
2734 {
2735         struct sigaction act = {
2736                 .sa_handler = sig_int,
2737                 .sa_flags = SA_RESTART,
2738         };
2739
2740         sigaction(SIGINT, &act, NULL);
2741
2742         /* Windows uses SIGBREAK as a quit signal from other applications */
2743 #ifdef WIN32
2744         sigaction(SIGBREAK, &act, NULL);
2745 #endif
2746 }
2747
2748 void fio_server_destroy_sk_key(void)
2749 {
2750         pthread_key_delete(sk_out_key);
2751 }
2752
2753 int fio_server_create_sk_key(void)
2754 {
2755         if (pthread_key_create(&sk_out_key, NULL)) {
2756                 log_err("fio: can't create sk_out backend key\n");
2757                 return 1;
2758         }
2759
2760         pthread_setspecific(sk_out_key, NULL);
2761         return 0;
2762 }
2763
2764 static int fio_server(void)
2765 {
2766         int sk, ret;
2767
2768         dprint(FD_NET, "starting server\n");
2769
2770         if (fio_handle_server_arg())
2771                 return -1;
2772
2773         set_sig_handlers();
2774
2775 #ifdef WIN32
2776         /* if this is a child process, go handle the connection */
2777         if (fio_server_pipe_name != NULL) {
2778                 ret = handle_connection_process();
2779                 return ret;
2780         }
2781
2782         /* job to link child processes so they terminate together */
2783         hjob = windows_create_job();
2784         if (hjob == INVALID_HANDLE_VALUE)
2785                 return -1;
2786 #endif
2787
2788         sk = fio_init_server_connection();
2789         if (sk < 0)
2790                 return -1;
2791
2792         ret = accept_loop(sk);
2793
2794         close(sk);
2795
2796         if (fio_server_arg) {
2797                 free(fio_server_arg);
2798                 fio_server_arg = NULL;
2799         }
2800         if (bind_sock)
2801                 free(bind_sock);
2802
2803         return ret;
2804 }
2805
2806 void fio_server_got_signal(int signal)
2807 {
2808         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2809
2810         assert(sk_out);
2811
2812         if (signal == SIGPIPE)
2813                 sk_out->sk = -1;
2814         else {
2815                 log_info("\nfio: terminating on signal %d\n", signal);
2816                 exit_backend = true;
2817         }
2818 }
2819
2820 static int check_existing_pidfile(const char *pidfile)
2821 {
2822         struct stat sb;
2823         char buf[16];
2824         pid_t pid;
2825         FILE *f;
2826
2827         if (stat(pidfile, &sb))
2828                 return 0;
2829
2830         f = fopen(pidfile, "r");
2831         if (!f)
2832                 return 0;
2833
2834         if (fread(buf, sb.st_size, 1, f) <= 0) {
2835                 fclose(f);
2836                 return 1;
2837         }
2838         fclose(f);
2839
2840         pid = atoi(buf);
2841         if (kill(pid, SIGCONT) < 0)
2842                 return errno != ESRCH;
2843
2844         return 1;
2845 }
2846
2847 static int write_pid(pid_t pid, const char *pidfile)
2848 {
2849         FILE *fpid;
2850
2851         fpid = fopen(pidfile, "w");
2852         if (!fpid) {
2853                 log_err("fio: failed opening pid file %s\n", pidfile);
2854                 return 1;
2855         }
2856
2857         fprintf(fpid, "%u\n", (unsigned int) pid);
2858         fclose(fpid);
2859         return 0;
2860 }
2861
2862 /*
2863  * If pidfile is specified, background us.
2864  */
2865 int fio_start_server(char *pidfile)
2866 {
2867         FILE *file;
2868         pid_t pid;
2869         int ret;
2870
2871 #if defined(WIN32)
2872         WSADATA wsd;
2873         WSAStartup(MAKEWORD(2, 2), &wsd);
2874 #endif
2875
2876         if (!pidfile)
2877                 return fio_server();
2878
2879         if (check_existing_pidfile(pidfile)) {
2880                 log_err("fio: pidfile %s exists and server appears alive\n",
2881                                                                 pidfile);
2882                 free(pidfile);
2883                 return -1;
2884         }
2885
2886         pid = fork();
2887         if (pid < 0) {
2888                 log_err("fio: failed server fork: %s\n", strerror(errno));
2889                 free(pidfile);
2890                 return -1;
2891         } else if (pid) {
2892                 ret = write_pid(pid, pidfile);
2893                 free(pidfile);
2894                 _exit(ret);
2895         }
2896
2897         setsid();
2898         openlog("fio", LOG_NDELAY|LOG_NOWAIT|LOG_PID, LOG_USER);
2899         log_syslog = true;
2900
2901         file = freopen("/dev/null", "r", stdin);
2902         if (!file)
2903                 perror("freopen");
2904
2905         file = freopen("/dev/null", "w", stdout);
2906         if (!file)
2907                 perror("freopen");
2908
2909         file = freopen("/dev/null", "w", stderr);
2910         if (!file)
2911                 perror("freopen");
2912
2913         f_out = NULL;
2914         f_err = NULL;
2915
2916         ret = fio_server();
2917
2918         fclose(stdin);
2919         fclose(stdout);
2920         fclose(stderr);
2921
2922         closelog();
2923         unlink(pidfile);
2924         free(pidfile);
2925         return ret;
2926 }
2927
2928 void fio_server_set_arg(const char *arg)
2929 {
2930         fio_server_arg = strdup(arg);
2931 }
2932
2933 #ifdef WIN32
2934 void fio_server_internal_set(const char *arg)
2935 {
2936         fio_server_pipe_name = strdup(arg);
2937 }
2938 #endif