client: cleanup json output
[fio.git] / client.c
1 #include <stdio.h>
2 #include <stdlib.h>
3 #include <unistd.h>
4 #include <limits.h>
5 #include <errno.h>
6 #include <fcntl.h>
7 #include <sys/poll.h>
8 #include <sys/types.h>
9 #include <sys/stat.h>
10 #include <sys/wait.h>
11 #include <sys/socket.h>
12 #include <sys/un.h>
13 #include <netinet/in.h>
14 #include <arpa/inet.h>
15 #include <netdb.h>
16 #include <signal.h>
17 #ifdef CONFIG_ZLIB
18 #include <zlib.h>
19 #endif
20
21 #include "fio.h"
22 #include "client.h"
23 #include "server.h"
24 #include "flist.h"
25 #include "hash.h"
26 #include "verify.h"
27
28 static void handle_du(struct fio_client *client, struct fio_net_cmd *cmd);
29 static void handle_ts(struct fio_client *client, struct fio_net_cmd *cmd);
30 static void handle_gs(struct fio_client *client, struct fio_net_cmd *cmd);
31 static void handle_probe(struct fio_client *client, struct fio_net_cmd *cmd);
32 static void handle_text(struct fio_client *client, struct fio_net_cmd *cmd);
33 static void handle_stop(struct fio_client *client, struct fio_net_cmd *cmd);
34 static void handle_start(struct fio_client *client, struct fio_net_cmd *cmd);
35
36 struct client_ops fio_client_ops = {
37         .text           = handle_text,
38         .disk_util      = handle_du,
39         .thread_status  = handle_ts,
40         .group_stats    = handle_gs,
41         .stop           = handle_stop,
42         .start          = handle_start,
43         .eta            = display_thread_status,
44         .probe          = handle_probe,
45         .eta_msec       = FIO_CLIENT_DEF_ETA_MSEC,
46         .client_type    = FIO_CLIENT_TYPE_CLI,
47 };
48
49 static struct timeval eta_tv;
50
51 static FLIST_HEAD(client_list);
52 static FLIST_HEAD(eta_list);
53
54 static FLIST_HEAD(arg_list);
55
56 struct thread_stat client_ts;
57 struct group_run_stats client_gs;
58 int sum_stat_clients;
59
60 static int sum_stat_nr;
61 static struct json_object *root = NULL;
62 static struct json_array *clients_array = NULL;
63 static struct json_array *du_array = NULL;
64
65 static int error_clients;
66
67 #define FIO_CLIENT_HASH_BITS    7
68 #define FIO_CLIENT_HASH_SZ      (1 << FIO_CLIENT_HASH_BITS)
69 #define FIO_CLIENT_HASH_MASK    (FIO_CLIENT_HASH_SZ - 1)
70 static struct flist_head client_hash[FIO_CLIENT_HASH_SZ];
71
72 static void fio_client_add_hash(struct fio_client *client)
73 {
74         int bucket = hash_long(client->fd, FIO_CLIENT_HASH_BITS);
75
76         bucket &= FIO_CLIENT_HASH_MASK;
77         flist_add(&client->hash_list, &client_hash[bucket]);
78 }
79
80 static void fio_client_remove_hash(struct fio_client *client)
81 {
82         if (!flist_empty(&client->hash_list))
83                 flist_del_init(&client->hash_list);
84 }
85
86 static void fio_init fio_client_hash_init(void)
87 {
88         int i;
89
90         for (i = 0; i < FIO_CLIENT_HASH_SZ; i++)
91                 INIT_FLIST_HEAD(&client_hash[i]);
92 }
93
94 static int read_data(int fd, void *data, size_t size)
95 {
96         ssize_t ret;
97
98         while (size) {
99                 ret = read(fd, data, size);
100                 if (ret < 0) {
101                         if (errno == EAGAIN || errno == EINTR)
102                                 continue;
103                         break;
104                 } else if (!ret)
105                         break;
106                 else {
107                         data += ret;
108                         size -= ret;
109                 }
110         }
111
112         if (size)
113                 return EAGAIN;
114
115         return 0;
116 }
117
118 static void fio_client_json_init(void)
119 {
120         if (!(output_format & FIO_OUTPUT_JSON))
121                 return;
122
123         root = json_create_object();
124         json_object_add_value_string(root, "fio version", fio_version_string);
125         clients_array = json_create_array();
126         json_object_add_value_array(root, "client_stats", clients_array);
127         du_array = json_create_array();
128         json_object_add_value_array(root, "disk_util", du_array);
129 }
130
131 static void fio_client_json_fini(void)
132 {
133         if (!(output_format & FIO_OUTPUT_JSON))
134                 return;
135         log_info("\n");
136         json_print_object(root, NULL);
137         log_info("\n");
138         json_free_object(root);
139         root = NULL;
140         clients_array = NULL;
141         du_array = NULL;
142 }
143
144 static struct fio_client *find_client_by_fd(int fd)
145 {
146         int bucket = hash_long(fd, FIO_CLIENT_HASH_BITS) & FIO_CLIENT_HASH_MASK;
147         struct fio_client *client;
148         struct flist_head *entry;
149
150         flist_for_each(entry, &client_hash[bucket]) {
151                 client = flist_entry(entry, struct fio_client, hash_list);
152
153                 if (client->fd == fd) {
154                         client->refs++;
155                         return client;
156                 }
157         }
158
159         return NULL;
160 }
161
162 void fio_put_client(struct fio_client *client)
163 {
164         if (--client->refs)
165                 return;
166
167         free(client->hostname);
168         if (client->argv)
169                 free(client->argv);
170         if (client->name)
171                 free(client->name);
172         while (client->nr_files) {
173                 struct client_file *cf = &client->files[--client->nr_files];
174
175                 free(cf->file);
176         }
177         if (client->files)
178                 free(client->files);
179
180         if (!client->did_stat)
181                 sum_stat_clients--;
182
183         if (client->error)
184                 error_clients++;
185
186         free(client);
187 }
188
189 static int fio_client_dec_jobs_eta(struct client_eta *eta, client_eta_op eta_fn)
190 {
191         if (!--eta->pending) {
192                 eta_fn(&eta->eta);
193                 free(eta);
194                 return 0;
195         }
196
197         return 1;
198 }
199
200 static void remove_client(struct fio_client *client)
201 {
202         assert(client->refs);
203
204         dprint(FD_NET, "client: removed <%s>\n", client->hostname);
205
206         if (!flist_empty(&client->list))
207                 flist_del_init(&client->list);
208
209         fio_client_remove_hash(client);
210
211         if (!flist_empty(&client->eta_list)) {
212                 flist_del_init(&client->eta_list);
213                 fio_client_dec_jobs_eta(client->eta_in_flight, client->ops->eta);
214         }
215
216         close(client->fd);
217         client->fd = -1;
218
219         if (client->ops->removed)
220                 client->ops->removed(client);
221
222         nr_clients--;
223         fio_put_client(client);
224 }
225
226 struct fio_client *fio_get_client(struct fio_client *client)
227 {
228         client->refs++;
229         return client;
230 }
231
232 static void __fio_client_add_cmd_option(struct fio_client *client,
233                                         const char *opt)
234 {
235         int index;
236
237         index = client->argc++;
238         client->argv = realloc(client->argv, sizeof(char *) * client->argc);
239         client->argv[index] = strdup(opt);
240         dprint(FD_NET, "client: add cmd %d: %s\n", index, opt);
241 }
242
243 void fio_client_add_cmd_option(void *cookie, const char *opt)
244 {
245         struct fio_client *client = cookie;
246         struct flist_head *entry;
247
248         if (!client || !opt)
249                 return;
250
251         __fio_client_add_cmd_option(client, opt);
252
253         /*
254          * Duplicate arguments to shared client group
255          */
256         flist_for_each(entry, &arg_list) {
257                 client = flist_entry(entry, struct fio_client, arg_list);
258
259                 __fio_client_add_cmd_option(client, opt);
260         }
261 }
262
263 struct fio_client *fio_client_add_explicit(struct client_ops *ops,
264                                            const char *hostname, int type,
265                                            int port)
266 {
267         struct fio_client *client;
268
269         client = malloc(sizeof(*client));
270         memset(client, 0, sizeof(*client));
271
272         INIT_FLIST_HEAD(&client->list);
273         INIT_FLIST_HEAD(&client->hash_list);
274         INIT_FLIST_HEAD(&client->arg_list);
275         INIT_FLIST_HEAD(&client->eta_list);
276         INIT_FLIST_HEAD(&client->cmd_list);
277
278         client->hostname = strdup(hostname);
279
280         if (type == Fio_client_socket)
281                 client->is_sock = 1;
282         else {
283                 int ipv6;
284
285                 ipv6 = type == Fio_client_ipv6;
286                 if (fio_server_parse_host(hostname, ipv6,
287                                                 &client->addr.sin_addr,
288                                                 &client->addr6.sin6_addr))
289                         goto err;
290
291                 client->port = port;
292         }
293
294         client->fd = -1;
295         client->ops = ops;
296         client->refs = 1;
297         client->type = ops->client_type;
298
299         __fio_client_add_cmd_option(client, "fio");
300
301         flist_add(&client->list, &client_list);
302         nr_clients++;
303         dprint(FD_NET, "client: added <%s>\n", client->hostname);
304         return client;
305 err:
306         free(client);
307         return NULL;
308 }
309
310 int fio_client_add_ini_file(void *cookie, const char *ini_file, int remote)
311 {
312         struct fio_client *client = cookie;
313         struct client_file *cf;
314         size_t new_size;
315         void *new_files;
316
317         if (!client)
318                 return 1;
319
320         dprint(FD_NET, "client <%s>: add ini %s\n", client->hostname, ini_file);
321
322         new_size = (client->nr_files + 1) * sizeof(struct client_file);
323         new_files = realloc(client->files, new_size);
324         if (!new_files)
325                 return 1;
326
327         client->files = new_files;
328         cf = &client->files[client->nr_files];
329         cf->file = strdup(ini_file);
330         cf->remote = remote;
331         client->nr_files++;
332         return 0;
333 }
334
335 int fio_client_add(struct client_ops *ops, const char *hostname, void **cookie)
336 {
337         struct fio_client *existing = *cookie;
338         struct fio_client *client;
339
340         if (existing) {
341                 /*
342                  * We always add our "exec" name as the option, hence 1
343                  * means empty.
344                  */
345                 if (existing->argc == 1)
346                         flist_add_tail(&existing->arg_list, &arg_list);
347                 else {
348                         while (!flist_empty(&arg_list))
349                                 flist_del_init(arg_list.next);
350                 }
351         }
352
353         client = malloc(sizeof(*client));
354         memset(client, 0, sizeof(*client));
355
356         INIT_FLIST_HEAD(&client->list);
357         INIT_FLIST_HEAD(&client->hash_list);
358         INIT_FLIST_HEAD(&client->arg_list);
359         INIT_FLIST_HEAD(&client->eta_list);
360         INIT_FLIST_HEAD(&client->cmd_list);
361
362         if (fio_server_parse_string(hostname, &client->hostname,
363                                         &client->is_sock, &client->port,
364                                         &client->addr.sin_addr,
365                                         &client->addr6.sin6_addr,
366                                         &client->ipv6))
367                 return -1;
368
369         client->fd = -1;
370         client->ops = ops;
371         client->refs = 1;
372         client->type = ops->client_type;
373
374         __fio_client_add_cmd_option(client, "fio");
375
376         flist_add(&client->list, &client_list);
377         nr_clients++;
378         dprint(FD_NET, "client: added <%s>\n", client->hostname);
379         *cookie = client;
380         return 0;
381 }
382
383 static const char *server_name(struct fio_client *client, char *buf,
384                                size_t bufsize)
385 {
386         const char *from;
387
388         if (client->ipv6)
389                 from = inet_ntop(AF_INET6, (struct sockaddr *) &client->addr6.sin6_addr, buf, bufsize);
390         else if (client->is_sock)
391                 from = "sock";
392         else
393                 from = inet_ntop(AF_INET, (struct sockaddr *) &client->addr.sin_addr, buf, bufsize);
394
395         return from;
396 }
397
398 static void probe_client(struct fio_client *client)
399 {
400         struct cmd_client_probe_pdu pdu;
401         const char *sname;
402         uint64_t tag;
403         char buf[64];
404
405         dprint(FD_NET, "client: send probe\n");
406
407 #ifdef CONFIG_ZLIB
408         pdu.flags = __le64_to_cpu(FIO_PROBE_FLAG_ZLIB);
409 #else
410         pdu.flags = 0;
411 #endif
412
413         sname = server_name(client, buf, sizeof(buf));
414         memset(pdu.server, 0, sizeof(pdu.server));
415         strncpy((char *) pdu.server, sname, sizeof(pdu.server) - 1);
416
417         fio_net_send_cmd(client->fd, FIO_NET_CMD_PROBE, &pdu, sizeof(pdu), &tag, &client->cmd_list);
418 }
419
420 static int fio_client_connect_ip(struct fio_client *client)
421 {
422         struct sockaddr *addr;
423         socklen_t socklen;
424         int fd, domain;
425
426         if (client->ipv6) {
427                 client->addr6.sin6_family = AF_INET6;
428                 client->addr6.sin6_port = htons(client->port);
429                 domain = AF_INET6;
430                 addr = (struct sockaddr *) &client->addr6;
431                 socklen = sizeof(client->addr6);
432         } else {
433                 client->addr.sin_family = AF_INET;
434                 client->addr.sin_port = htons(client->port);
435                 domain = AF_INET;
436                 addr = (struct sockaddr *) &client->addr;
437                 socklen = sizeof(client->addr);
438         }
439
440         fd = socket(domain, SOCK_STREAM, 0);
441         if (fd < 0) {
442                 int ret = -errno;
443
444                 log_err("fio: socket: %s\n", strerror(errno));
445                 return ret;
446         }
447
448         if (connect(fd, addr, socklen) < 0) {
449                 int ret = -errno;
450
451                 log_err("fio: connect: %s\n", strerror(errno));
452                 log_err("fio: failed to connect to %s:%u\n", client->hostname,
453                                                                 client->port);
454                 close(fd);
455                 return ret;
456         }
457
458         return fd;
459 }
460
461 static int fio_client_connect_sock(struct fio_client *client)
462 {
463         struct sockaddr_un *addr = &client->addr_un;
464         socklen_t len;
465         int fd;
466
467         memset(addr, 0, sizeof(*addr));
468         addr->sun_family = AF_UNIX;
469         strncpy(addr->sun_path, client->hostname, sizeof(addr->sun_path) - 1);
470
471         fd = socket(AF_UNIX, SOCK_STREAM, 0);
472         if (fd < 0) {
473                 int ret = -errno;
474
475                 log_err("fio: socket: %s\n", strerror(errno));
476                 return ret;
477         }
478
479         len = sizeof(addr->sun_family) + strlen(addr->sun_path) + 1;
480         if (connect(fd, (struct sockaddr *) addr, len) < 0) {
481                 int ret = -errno;
482
483                 log_err("fio: connect; %s\n", strerror(errno));
484                 close(fd);
485                 return ret;
486         }
487
488         return fd;
489 }
490
491 int fio_client_connect(struct fio_client *client)
492 {
493         int fd;
494
495         dprint(FD_NET, "client: connect to host %s\n", client->hostname);
496
497         if (client->is_sock)
498                 fd = fio_client_connect_sock(client);
499         else
500                 fd = fio_client_connect_ip(client);
501
502         dprint(FD_NET, "client: %s connected %d\n", client->hostname, fd);
503
504         if (fd < 0)
505                 return fd;
506
507         client->fd = fd;
508         fio_client_add_hash(client);
509         client->state = Client_connected;
510
511         probe_client(client);
512         return 0;
513 }
514
515 int fio_client_terminate(struct fio_client *client)
516 {
517         return fio_net_send_quit(client->fd);
518 }
519
520 void fio_clients_terminate(void)
521 {
522         struct flist_head *entry;
523         struct fio_client *client;
524
525         dprint(FD_NET, "client: terminate clients\n");
526
527         flist_for_each(entry, &client_list) {
528                 client = flist_entry(entry, struct fio_client, list);
529                 fio_client_terminate(client);
530         }
531 }
532
533 static void sig_int(int sig)
534 {
535         dprint(FD_NET, "client: got signal %d\n", sig);
536         fio_clients_terminate();
537 }
538
539 static void client_signal_handler(void)
540 {
541         struct sigaction act;
542
543         memset(&act, 0, sizeof(act));
544         act.sa_handler = sig_int;
545         act.sa_flags = SA_RESTART;
546         sigaction(SIGINT, &act, NULL);
547
548         memset(&act, 0, sizeof(act));
549         act.sa_handler = sig_int;
550         act.sa_flags = SA_RESTART;
551         sigaction(SIGTERM, &act, NULL);
552
553 /* Windows uses SIGBREAK as a quit signal from other applications */
554 #ifdef WIN32
555         memset(&act, 0, sizeof(act));
556         act.sa_handler = sig_int;
557         act.sa_flags = SA_RESTART;
558         sigaction(SIGBREAK, &act, NULL);
559 #endif
560
561         memset(&act, 0, sizeof(act));
562         act.sa_handler = sig_show_status;
563         act.sa_flags = SA_RESTART;
564         sigaction(SIGUSR1, &act, NULL);
565 }
566
567 static int send_client_cmd_line(struct fio_client *client)
568 {
569         struct cmd_single_line_pdu *cslp;
570         struct cmd_line_pdu *clp;
571         unsigned long offset;
572         unsigned int *lens;
573         void *pdu;
574         size_t mem;
575         int i, ret;
576
577         dprint(FD_NET, "client: send cmdline %d\n", client->argc);
578
579         lens = malloc(client->argc * sizeof(unsigned int));
580
581         /*
582          * Find out how much mem we need
583          */
584         for (i = 0, mem = 0; i < client->argc; i++) {
585                 lens[i] = strlen(client->argv[i]) + 1;
586                 mem += lens[i];
587         }
588
589         /*
590          * We need one cmd_line_pdu, and argc number of cmd_single_line_pdu
591          */
592         mem += sizeof(*clp) + (client->argc * sizeof(*cslp));
593
594         pdu = malloc(mem);
595         clp = pdu;
596         offset = sizeof(*clp);
597
598         for (i = 0; i < client->argc; i++) {
599                 uint16_t arg_len = lens[i];
600
601                 cslp = pdu + offset;
602                 strcpy((char *) cslp->text, client->argv[i]);
603                 cslp->len = cpu_to_le16(arg_len);
604                 offset += sizeof(*cslp) + arg_len;
605         }
606
607         free(lens);
608         clp->lines = cpu_to_le16(client->argc);
609         clp->client_type = __cpu_to_le16(client->type);
610         ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOBLINE, pdu, mem, NULL, NULL);
611         free(pdu);
612         return ret;
613 }
614
615 int fio_clients_connect(void)
616 {
617         struct fio_client *client;
618         struct flist_head *entry, *tmp;
619         int ret;
620
621 #ifdef WIN32
622         WSADATA wsd;
623         WSAStartup(MAKEWORD(2, 2), &wsd);
624 #endif
625
626         dprint(FD_NET, "client: connect all\n");
627
628         client_signal_handler();
629
630         flist_for_each_safe(entry, tmp, &client_list) {
631                 client = flist_entry(entry, struct fio_client, list);
632
633                 ret = fio_client_connect(client);
634                 if (ret) {
635                         remove_client(client);
636                         continue;
637                 }
638
639                 if (client->argc > 1)
640                         send_client_cmd_line(client);
641         }
642
643         return !nr_clients;
644 }
645
646 int fio_start_client(struct fio_client *client)
647 {
648         dprint(FD_NET, "client: start %s\n", client->hostname);
649         return fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_RUN, 0, NULL);
650 }
651
652 int fio_start_all_clients(void)
653 {
654         struct fio_client *client;
655         struct flist_head *entry, *tmp;
656         int ret;
657
658         dprint(FD_NET, "client: start all\n");
659
660         fio_client_json_init();
661
662         flist_for_each_safe(entry, tmp, &client_list) {
663                 client = flist_entry(entry, struct fio_client, list);
664
665                 ret = fio_start_client(client);
666                 if (ret) {
667                         remove_client(client);
668                         continue;
669                 }
670         }
671
672         return flist_empty(&client_list);
673 }
674
675 static int __fio_client_send_remote_ini(struct fio_client *client,
676                                         const char *filename)
677 {
678         struct cmd_load_file_pdu *pdu;
679         size_t p_size;
680         int ret;
681
682         dprint(FD_NET, "send remote ini %s to %s\n", filename, client->hostname);
683
684         p_size = sizeof(*pdu) + strlen(filename) + 1;
685         pdu = malloc(p_size);
686         memset(pdu, 0, p_size);
687         pdu->name_len = strlen(filename);
688         strcpy((char *) pdu->file, filename);
689         pdu->client_type = cpu_to_le16((uint16_t) client->type);
690
691         client->sent_job = 1;
692         ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_LOAD_FILE, pdu, p_size,NULL, NULL);
693         free(pdu);
694         return ret;
695 }
696
697 /*
698  * Send file contents to server backend. We could use sendfile(), but to remain
699  * more portable lets just read/write the darn thing.
700  */
701 static int __fio_client_send_local_ini(struct fio_client *client,
702                                        const char *filename)
703 {
704         struct cmd_job_pdu *pdu;
705         size_t p_size;
706         struct stat sb;
707         char *p;
708         void *buf;
709         off_t len;
710         int fd, ret;
711
712         dprint(FD_NET, "send ini %s to %s\n", filename, client->hostname);
713
714         fd = open(filename, O_RDONLY);
715         if (fd < 0) {
716                 ret = -errno;
717                 log_err("fio: job file <%s> open: %s\n", filename, strerror(errno));
718                 return ret;
719         }
720
721         if (fstat(fd, &sb) < 0) {
722                 ret = -errno;
723                 log_err("fio: job file stat: %s\n", strerror(errno));
724                 close(fd);
725                 return ret;
726         }
727
728         p_size = sb.st_size + sizeof(*pdu);
729         pdu = malloc(p_size);
730         buf = pdu->buf;
731
732         len = sb.st_size;
733         p = buf;
734         if (read_data(fd, p, len)) {
735                 log_err("fio: failed reading job file %s\n", filename);
736                 close(fd);
737                 free(pdu);
738                 return 1;
739         }
740
741         pdu->buf_len = __cpu_to_le32(sb.st_size);
742         pdu->client_type = cpu_to_le32(client->type);
743
744         client->sent_job = 1;
745         ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOB, pdu, p_size, NULL, NULL);
746         free(pdu);
747         close(fd);
748         return ret;
749 }
750
751 int fio_client_send_ini(struct fio_client *client, const char *filename,
752                         int remote)
753 {
754         int ret;
755
756         if (!remote)
757                 ret = __fio_client_send_local_ini(client, filename);
758         else
759                 ret = __fio_client_send_remote_ini(client, filename);
760
761         if (!ret)
762                 client->sent_job = 1;
763
764         return ret;
765 }
766
767 static int fio_client_send_cf(struct fio_client *client,
768                               struct client_file *cf)
769 {
770         return fio_client_send_ini(client, cf->file, cf->remote);
771 }
772
773 int fio_clients_send_ini(const char *filename)
774 {
775         struct fio_client *client;
776         struct flist_head *entry, *tmp;
777
778         flist_for_each_safe(entry, tmp, &client_list) {
779                 client = flist_entry(entry, struct fio_client, list);
780
781                 if (client->nr_files) {
782                         int i;
783
784                         for (i = 0; i < client->nr_files; i++) {
785                                 struct client_file *cf;
786
787                                 cf = &client->files[i];
788
789                                 if (fio_client_send_cf(client, cf)) {
790                                         remove_client(client);
791                                         break;
792                                 }
793                         }
794                 }
795                 if (client->sent_job)
796                         continue;
797                 if (!filename || fio_client_send_ini(client, filename, 0))
798                         remove_client(client);
799         }
800
801         return !nr_clients;
802 }
803
804 int fio_client_update_options(struct fio_client *client,
805                               struct thread_options *o, uint64_t *tag)
806 {
807         struct cmd_add_job_pdu pdu;
808
809         pdu.thread_number = cpu_to_le32(client->thread_number);
810         pdu.groupid = cpu_to_le32(client->groupid);
811         convert_thread_options_to_net(&pdu.top, o);
812
813         return fio_net_send_cmd(client->fd, FIO_NET_CMD_UPDATE_JOB, &pdu, sizeof(pdu), tag, &client->cmd_list);
814 }
815
816 static void convert_io_stat(struct io_stat *dst, struct io_stat *src)
817 {
818         dst->max_val    = le64_to_cpu(src->max_val);
819         dst->min_val    = le64_to_cpu(src->min_val);
820         dst->samples    = le64_to_cpu(src->samples);
821
822         /*
823          * Floats arrive as IEEE 754 encoded uint64_t, convert back to double
824          */
825         dst->mean.u.f   = fio_uint64_to_double(le64_to_cpu(dst->mean.u.i));
826         dst->S.u.f      = fio_uint64_to_double(le64_to_cpu(dst->S.u.i));
827 }
828
829 static void convert_ts(struct thread_stat *dst, struct thread_stat *src)
830 {
831         int i, j;
832
833         dst->error              = le32_to_cpu(src->error);
834         dst->thread_number      = le32_to_cpu(src->thread_number);
835         dst->groupid            = le32_to_cpu(src->groupid);
836         dst->pid                = le32_to_cpu(src->pid);
837         dst->members            = le32_to_cpu(src->members);
838         dst->unified_rw_rep     = le32_to_cpu(src->unified_rw_rep);
839
840         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
841                 convert_io_stat(&dst->clat_stat[i], &src->clat_stat[i]);
842                 convert_io_stat(&dst->slat_stat[i], &src->slat_stat[i]);
843                 convert_io_stat(&dst->lat_stat[i], &src->lat_stat[i]);
844                 convert_io_stat(&dst->bw_stat[i], &src->bw_stat[i]);
845         }
846
847         dst->usr_time           = le64_to_cpu(src->usr_time);
848         dst->sys_time           = le64_to_cpu(src->sys_time);
849         dst->ctx                = le64_to_cpu(src->ctx);
850         dst->minf               = le64_to_cpu(src->minf);
851         dst->majf               = le64_to_cpu(src->majf);
852         dst->clat_percentiles   = le64_to_cpu(src->clat_percentiles);
853         dst->percentile_precision = le64_to_cpu(src->percentile_precision);
854
855         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
856                 fio_fp64_t *fps = &src->percentile_list[i];
857                 fio_fp64_t *fpd = &dst->percentile_list[i];
858
859                 fpd->u.f = fio_uint64_to_double(le64_to_cpu(fps->u.i));
860         }
861
862         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
863                 dst->io_u_map[i]        = le32_to_cpu(src->io_u_map[i]);
864                 dst->io_u_submit[i]     = le32_to_cpu(src->io_u_submit[i]);
865                 dst->io_u_complete[i]   = le32_to_cpu(src->io_u_complete[i]);
866         }
867
868         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++) {
869                 dst->io_u_lat_u[i]      = le32_to_cpu(src->io_u_lat_u[i]);
870                 dst->io_u_lat_m[i]      = le32_to_cpu(src->io_u_lat_m[i]);
871         }
872
873         for (i = 0; i < DDIR_RWDIR_CNT; i++)
874                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
875                         dst->io_u_plat[i][j] = le32_to_cpu(src->io_u_plat[i][j]);
876
877         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
878                 dst->total_io_u[i]      = le64_to_cpu(src->total_io_u[i]);
879                 dst->short_io_u[i]      = le64_to_cpu(src->short_io_u[i]);
880                 dst->drop_io_u[i]       = le64_to_cpu(src->drop_io_u[i]);
881         }
882
883         dst->total_submit       = le64_to_cpu(src->total_submit);
884         dst->total_complete     = le64_to_cpu(src->total_complete);
885
886         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
887                 dst->io_bytes[i]        = le64_to_cpu(src->io_bytes[i]);
888                 dst->runtime[i]         = le64_to_cpu(src->runtime[i]);
889         }
890
891         dst->total_run_time     = le64_to_cpu(src->total_run_time);
892         dst->continue_on_error  = le16_to_cpu(src->continue_on_error);
893         dst->total_err_count    = le64_to_cpu(src->total_err_count);
894         dst->first_error        = le32_to_cpu(src->first_error);
895         dst->kb_base            = le32_to_cpu(src->kb_base);
896         dst->unit_base          = le32_to_cpu(src->unit_base);
897
898         dst->latency_depth      = le32_to_cpu(src->latency_depth);
899         dst->latency_target     = le64_to_cpu(src->latency_target);
900         dst->latency_window     = le64_to_cpu(src->latency_window);
901         dst->latency_percentile.u.f = fio_uint64_to_double(le64_to_cpu(src->latency_percentile.u.i));
902
903         dst->nr_block_infos     = le64_to_cpu(src->nr_block_infos);
904         for (i = 0; i < dst->nr_block_infos; i++)
905                 dst->block_infos[i] = le32_to_cpu(src->block_infos[i]);
906 }
907
908 static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
909 {
910         int i;
911
912         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
913                 dst->max_run[i]         = le64_to_cpu(src->max_run[i]);
914                 dst->min_run[i]         = le64_to_cpu(src->min_run[i]);
915                 dst->max_bw[i]          = le64_to_cpu(src->max_bw[i]);
916                 dst->min_bw[i]          = le64_to_cpu(src->min_bw[i]);
917                 dst->io_kb[i]           = le64_to_cpu(src->io_kb[i]);
918                 dst->agg[i]             = le64_to_cpu(src->agg[i]);
919         }
920
921         dst->kb_base    = le32_to_cpu(src->kb_base);
922         dst->unit_base  = le32_to_cpu(src->unit_base);
923         dst->groupid    = le32_to_cpu(src->groupid);
924         dst->unified_rw_rep     = le32_to_cpu(src->unified_rw_rep);
925 }
926
927 static void json_object_add_client_info(struct json_object *obj,
928                                         struct fio_client *client)
929 {
930         const char *hostname = client->hostname ? client->hostname : "";
931
932         json_object_add_value_string(obj, "hostname", hostname);
933         json_object_add_value_int(obj, "port", client->port);
934 }
935
936 static void handle_ts(struct fio_client *client, struct fio_net_cmd *cmd)
937 {
938         struct cmd_ts_pdu *p = (struct cmd_ts_pdu *) cmd->payload;
939         struct json_object *tsobj;
940
941         tsobj = show_thread_status(&p->ts, &p->rs, NULL);
942         client->did_stat = 1;
943         if (tsobj) {
944                 json_object_add_client_info(tsobj, client);
945                 json_array_add_value_object(clients_array, tsobj);
946         }
947
948         if (sum_stat_clients <= 1)
949                 return;
950
951         sum_thread_stats(&client_ts, &p->ts, sum_stat_nr == 1);
952         sum_group_stats(&client_gs, &p->rs);
953
954         client_ts.members++;
955         client_ts.thread_number = p->ts.thread_number;
956         client_ts.groupid = p->ts.groupid;
957         client_ts.unified_rw_rep = p->ts.unified_rw_rep;
958
959         if (++sum_stat_nr == sum_stat_clients) {
960                 strcpy(client_ts.name, "All clients");
961                 tsobj = show_thread_status(&client_ts, &client_gs, NULL);
962                 if (tsobj) {
963                         json_object_add_client_info(tsobj, client);
964                         json_array_add_value_object(clients_array, tsobj);
965                 }
966         }
967 }
968
969 static void handle_gs(struct fio_client *client, struct fio_net_cmd *cmd)
970 {
971         struct group_run_stats *gs = (struct group_run_stats *) cmd->payload;
972
973         if (output_format & FIO_OUTPUT_NORMAL)
974                 show_group_stats(gs, NULL);
975 }
976
977 static void handle_text(struct fio_client *client, struct fio_net_cmd *cmd)
978 {
979         struct cmd_text_pdu *pdu = (struct cmd_text_pdu *) cmd->payload;
980         const char *buf = (const char *) pdu->buf;
981         const char *name;
982         int fio_unused ret;
983
984         name = client->name ? client->name : client->hostname;
985
986         if (!client->skip_newline)
987                 fprintf(f_out, "<%s> ", name);
988         ret = fwrite(buf, pdu->buf_len, 1, f_out);
989         fflush(f_out);
990         client->skip_newline = strchr(buf, '\n') == NULL;
991 }
992
993 static void convert_agg(struct disk_util_agg *agg)
994 {
995         int i;
996
997         for (i = 0; i < 2; i++) {
998                 agg->ios[i]     = le64_to_cpu(agg->ios[i]);
999                 agg->merges[i]  = le64_to_cpu(agg->merges[i]);
1000                 agg->sectors[i] = le64_to_cpu(agg->sectors[i]);
1001                 agg->ticks[i]   = le64_to_cpu(agg->ticks[i]);
1002         }
1003
1004         agg->io_ticks           = le64_to_cpu(agg->io_ticks);
1005         agg->time_in_queue      = le64_to_cpu(agg->time_in_queue);
1006         agg->slavecount         = le32_to_cpu(agg->slavecount);
1007         agg->max_util.u.f       = fio_uint64_to_double(le64_to_cpu(agg->max_util.u.i));
1008 }
1009
1010 static void convert_dus(struct disk_util_stat *dus)
1011 {
1012         int i;
1013
1014         for (i = 0; i < 2; i++) {
1015                 dus->s.ios[i]           = le64_to_cpu(dus->s.ios[i]);
1016                 dus->s.merges[i]        = le64_to_cpu(dus->s.merges[i]);
1017                 dus->s.sectors[i]       = le64_to_cpu(dus->s.sectors[i]);
1018                 dus->s.ticks[i]         = le64_to_cpu(dus->s.ticks[i]);
1019         }
1020
1021         dus->s.io_ticks         = le64_to_cpu(dus->s.io_ticks);
1022         dus->s.time_in_queue    = le64_to_cpu(dus->s.time_in_queue);
1023         dus->s.msec             = le64_to_cpu(dus->s.msec);
1024 }
1025
1026 static void handle_du(struct fio_client *client, struct fio_net_cmd *cmd)
1027 {
1028         struct cmd_du_pdu *du = (struct cmd_du_pdu *) cmd->payload;
1029
1030         if (!client->disk_stats_shown) {
1031                 client->disk_stats_shown = 1;
1032                 log_info("\nDisk stats (read/write):\n");
1033         }
1034
1035         if (output_format & FIO_OUTPUT_JSON) {
1036                 struct json_object *duobj;
1037                 json_array_add_disk_util(&du->dus, &du->agg, du_array);
1038                 duobj = json_array_last_value_object(du_array);
1039                 json_object_add_client_info(duobj, client);
1040         }
1041         if (output_format & FIO_OUTPUT_TERSE)
1042                 print_disk_util(&du->dus, &du->agg, 1, NULL);
1043         if (output_format & FIO_OUTPUT_NORMAL)
1044                 print_disk_util(&du->dus, &du->agg, 0, NULL);
1045 }
1046
1047 static void convert_jobs_eta(struct jobs_eta *je)
1048 {
1049         int i;
1050
1051         je->nr_running          = le32_to_cpu(je->nr_running);
1052         je->nr_ramp             = le32_to_cpu(je->nr_ramp);
1053         je->nr_pending          = le32_to_cpu(je->nr_pending);
1054         je->nr_setting_up       = le32_to_cpu(je->nr_setting_up);
1055         je->files_open          = le32_to_cpu(je->files_open);
1056
1057         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1058                 je->m_rate[i]   = le32_to_cpu(je->m_rate[i]);
1059                 je->t_rate[i]   = le32_to_cpu(je->t_rate[i]);
1060                 je->m_iops[i]   = le32_to_cpu(je->m_iops[i]);
1061                 je->t_iops[i]   = le32_to_cpu(je->t_iops[i]);
1062                 je->rate[i]     = le32_to_cpu(je->rate[i]);
1063                 je->iops[i]     = le32_to_cpu(je->iops[i]);
1064         }
1065
1066         je->elapsed_sec         = le64_to_cpu(je->elapsed_sec);
1067         je->eta_sec             = le64_to_cpu(je->eta_sec);
1068         je->nr_threads          = le32_to_cpu(je->nr_threads);
1069         je->is_pow2             = le32_to_cpu(je->is_pow2);
1070         je->unit_base           = le32_to_cpu(je->unit_base);
1071 }
1072
1073 void fio_client_sum_jobs_eta(struct jobs_eta *dst, struct jobs_eta *je)
1074 {
1075         int i;
1076
1077         dst->nr_running         += je->nr_running;
1078         dst->nr_ramp            += je->nr_ramp;
1079         dst->nr_pending         += je->nr_pending;
1080         dst->nr_setting_up      += je->nr_setting_up;
1081         dst->files_open         += je->files_open;
1082
1083         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1084                 dst->m_rate[i]  += je->m_rate[i];
1085                 dst->t_rate[i]  += je->t_rate[i];
1086                 dst->m_iops[i]  += je->m_iops[i];
1087                 dst->t_iops[i]  += je->t_iops[i];
1088                 dst->rate[i]    += je->rate[i];
1089                 dst->iops[i]    += je->iops[i];
1090         }
1091
1092         dst->elapsed_sec        += je->elapsed_sec;
1093
1094         if (je->eta_sec > dst->eta_sec)
1095                 dst->eta_sec = je->eta_sec;
1096
1097         dst->nr_threads         += je->nr_threads;
1098
1099         /*
1100          * This wont be correct for multiple strings, but at least it
1101          * works for the basic cases.
1102          */
1103         strcpy((char *) dst->run_str, (char *) je->run_str);
1104 }
1105
1106 static void remove_reply_cmd(struct fio_client *client, struct fio_net_cmd *cmd)
1107 {
1108         struct fio_net_cmd_reply *reply = NULL;
1109         struct flist_head *entry;
1110
1111         flist_for_each(entry, &client->cmd_list) {
1112                 reply = flist_entry(entry, struct fio_net_cmd_reply, list);
1113
1114                 if (cmd->tag == (uintptr_t) reply)
1115                         break;
1116
1117                 reply = NULL;
1118         }
1119
1120         if (!reply) {
1121                 log_err("fio: client: unable to find matching tag (%llx)\n", (unsigned long long) cmd->tag);
1122                 return;
1123         }
1124
1125         flist_del(&reply->list);
1126         cmd->tag = reply->saved_tag;
1127         free(reply);
1128 }
1129
1130 int fio_client_wait_for_reply(struct fio_client *client, uint64_t tag)
1131 {
1132         do {
1133                 struct fio_net_cmd_reply *reply = NULL;
1134                 struct flist_head *entry;
1135
1136                 flist_for_each(entry, &client->cmd_list) {
1137                         reply = flist_entry(entry, struct fio_net_cmd_reply, list);
1138
1139                         if (tag == (uintptr_t) reply)
1140                                 break;
1141
1142                         reply = NULL;
1143                 }
1144
1145                 if (!reply)
1146                         break;
1147
1148                 usleep(1000);
1149         } while (1);
1150
1151         return 0;
1152 }
1153
1154 static void handle_eta(struct fio_client *client, struct fio_net_cmd *cmd)
1155 {
1156         struct jobs_eta *je = (struct jobs_eta *) cmd->payload;
1157         struct client_eta *eta = (struct client_eta *) (uintptr_t) cmd->tag;
1158
1159         dprint(FD_NET, "client: got eta tag %p, %d\n", eta, eta->pending);
1160
1161         assert(client->eta_in_flight == eta);
1162
1163         client->eta_in_flight = NULL;
1164         flist_del_init(&client->eta_list);
1165         client->eta_timeouts = 0;
1166
1167         if (client->ops->jobs_eta)
1168                 client->ops->jobs_eta(client, je);
1169
1170         fio_client_sum_jobs_eta(&eta->eta, je);
1171         fio_client_dec_jobs_eta(eta, client->ops->eta);
1172 }
1173
1174 static void handle_probe(struct fio_client *client, struct fio_net_cmd *cmd)
1175 {
1176         struct cmd_probe_reply_pdu *probe = (struct cmd_probe_reply_pdu *) cmd->payload;
1177         const char *os, *arch;
1178         char bit[16];
1179
1180         os = fio_get_os_string(probe->os);
1181         if (!os)
1182                 os = "unknown";
1183
1184         arch = fio_get_arch_string(probe->arch);
1185         if (!arch)
1186                 os = "unknown";
1187
1188         sprintf(bit, "%d-bit", probe->bpp * 8);
1189         probe->flags = le64_to_cpu(probe->flags);
1190
1191         log_info("hostname=%s, be=%u, %s, os=%s, arch=%s, fio=%s, flags=%lx\n",
1192                 probe->hostname, probe->bigendian, bit, os, arch,
1193                 probe->fio_version, (unsigned long) probe->flags);
1194
1195         if (!client->name)
1196                 client->name = strdup((char *) probe->hostname);
1197 }
1198
1199 static void handle_start(struct fio_client *client, struct fio_net_cmd *cmd)
1200 {
1201         struct cmd_start_pdu *pdu = (struct cmd_start_pdu *) cmd->payload;
1202
1203         client->state = Client_started;
1204         client->jobs = le32_to_cpu(pdu->jobs);
1205         client->nr_stat = le32_to_cpu(pdu->stat_outputs);
1206
1207         sum_stat_clients += client->nr_stat;
1208 }
1209
1210 static void handle_stop(struct fio_client *client, struct fio_net_cmd *cmd)
1211 {
1212         if (client->error)
1213                 log_info("client <%s>: exited with error %d\n", client->hostname, client->error);
1214 }
1215
1216 static void convert_stop(struct fio_net_cmd *cmd)
1217 {
1218         struct cmd_end_pdu *pdu = (struct cmd_end_pdu *) cmd->payload;
1219
1220         pdu->error = le32_to_cpu(pdu->error);
1221 }
1222
1223 static void convert_text(struct fio_net_cmd *cmd)
1224 {
1225         struct cmd_text_pdu *pdu = (struct cmd_text_pdu *) cmd->payload;
1226
1227         pdu->level      = le32_to_cpu(pdu->level);
1228         pdu->buf_len    = le32_to_cpu(pdu->buf_len);
1229         pdu->log_sec    = le64_to_cpu(pdu->log_sec);
1230         pdu->log_usec   = le64_to_cpu(pdu->log_usec);
1231 }
1232
1233 static struct cmd_iolog_pdu *convert_iolog_gz(struct fio_net_cmd *cmd,
1234                                               struct cmd_iolog_pdu *pdu)
1235 {
1236 #ifdef CONFIG_ZLIB
1237         struct cmd_iolog_pdu *ret;
1238         z_stream stream;
1239         uint32_t nr_samples;
1240         size_t total;
1241         void *p;
1242
1243         stream.zalloc = Z_NULL;
1244         stream.zfree = Z_NULL;
1245         stream.opaque = Z_NULL;
1246         stream.avail_in = 0;
1247         stream.next_in = Z_NULL;
1248
1249         if (inflateInit(&stream) != Z_OK)
1250                 return NULL;
1251
1252         /*
1253          * Get header first, it's not compressed
1254          */
1255         nr_samples = le64_to_cpu(pdu->nr_samples);
1256
1257         total = nr_samples * __log_entry_sz(le32_to_cpu(pdu->log_offset));
1258         ret = malloc(total + sizeof(*pdu));
1259         ret->nr_samples = nr_samples;
1260
1261         memcpy(ret, pdu, sizeof(*pdu));
1262
1263         p = (void *) ret + sizeof(*pdu);
1264
1265         stream.avail_in = cmd->pdu_len - sizeof(*pdu);
1266         stream.next_in = (void *) pdu + sizeof(*pdu);
1267         while (stream.avail_in) {
1268                 unsigned int this_chunk = 65536;
1269                 unsigned int this_len;
1270                 int err;
1271
1272                 if (this_chunk > total)
1273                         this_chunk = total;
1274
1275                 stream.avail_out = this_chunk;
1276                 stream.next_out = p;
1277                 err = inflate(&stream, Z_NO_FLUSH);
1278                 /* may be Z_OK, or Z_STREAM_END */
1279                 if (err < 0) {
1280                         log_err("fio: inflate error %d\n", err);
1281                         free(ret);
1282                         ret = NULL;
1283                         goto err;
1284                 }
1285
1286                 this_len = this_chunk - stream.avail_out;
1287                 p += this_len;
1288                 total -= this_len;
1289         }
1290
1291 err:
1292         inflateEnd(&stream);
1293         return ret;
1294 #else
1295         return NULL;
1296 #endif
1297 }
1298
1299 /*
1300  * This has been compressed on the server side, since it can be big.
1301  * Uncompress here.
1302  */
1303 static struct cmd_iolog_pdu *convert_iolog(struct fio_net_cmd *cmd)
1304 {
1305         struct cmd_iolog_pdu *pdu = (struct cmd_iolog_pdu *) cmd->payload;
1306         struct cmd_iolog_pdu *ret;
1307         uint64_t i;
1308         void *samples;
1309
1310         /*
1311          * Convert if compressed and we support it. If it's not
1312          * compressed, we need not do anything.
1313          */
1314         if (le32_to_cpu(pdu->compressed)) {
1315 #ifndef CONFIG_ZLIB
1316                 log_err("fio: server sent compressed data by mistake\n");
1317                 return NULL;
1318 #endif
1319                 ret = convert_iolog_gz(cmd, pdu);
1320                 if (!ret) {
1321                         log_err("fio: failed decompressing log\n");
1322                         return NULL;
1323                 }
1324         } else
1325                 ret = pdu;
1326
1327         ret->nr_samples         = le64_to_cpu(ret->nr_samples);
1328         ret->thread_number      = le32_to_cpu(ret->thread_number);
1329         ret->log_type           = le32_to_cpu(ret->log_type);
1330         ret->compressed         = le32_to_cpu(ret->compressed);
1331         ret->log_offset         = le32_to_cpu(ret->log_offset);
1332
1333         samples = &ret->samples[0];
1334         for (i = 0; i < ret->nr_samples; i++) {
1335                 struct io_sample *s;
1336
1337                 s = __get_sample(samples, ret->log_offset, i);
1338                 s->time         = le64_to_cpu(s->time);
1339                 s->val          = le64_to_cpu(s->val);
1340                 s->__ddir       = le32_to_cpu(s->__ddir);
1341                 s->bs           = le32_to_cpu(s->bs);
1342
1343                 if (ret->log_offset) {
1344                         struct io_sample_offset *so = (void *) s;
1345
1346                         so->offset = le64_to_cpu(so->offset);
1347                 }
1348         }
1349
1350         return ret;
1351 }
1352
1353 static void sendfile_reply(int fd, struct cmd_sendfile_reply *rep,
1354                            size_t size, uint64_t tag)
1355 {
1356         rep->error = cpu_to_le32(rep->error);
1357         fio_net_send_cmd(fd, FIO_NET_CMD_SENDFILE, rep, size, &tag, NULL);
1358 }
1359
1360 static int send_file(struct fio_client *client, struct cmd_sendfile *pdu,
1361                      uint64_t tag)
1362 {
1363         struct cmd_sendfile_reply *rep;
1364         struct stat sb;
1365         size_t size;
1366         int fd;
1367
1368         size = sizeof(*rep);
1369         rep = malloc(size);
1370
1371         if (stat((char *)pdu->path, &sb) < 0) {
1372 fail:
1373                 rep->error = errno;
1374                 sendfile_reply(client->fd, rep, size, tag);
1375                 free(rep);
1376                 return 1;
1377         }
1378
1379         size += sb.st_size;
1380         rep = realloc(rep, size);
1381         rep->size = cpu_to_le32((uint32_t) sb.st_size);
1382
1383         fd = open((char *)pdu->path, O_RDONLY);
1384         if (fd == -1 )
1385                 goto fail;
1386
1387         rep->error = read_data(fd, &rep->data, sb.st_size);
1388         sendfile_reply(client->fd, rep, size, tag);
1389         free(rep);
1390         close(fd);
1391         return 0;
1392 }
1393
1394 int fio_handle_client(struct fio_client *client)
1395 {
1396         struct client_ops *ops = client->ops;
1397         struct fio_net_cmd *cmd;
1398
1399         dprint(FD_NET, "client: handle %s\n", client->hostname);
1400
1401         cmd = fio_net_recv_cmd(client->fd);
1402         if (!cmd)
1403                 return 0;
1404
1405         dprint(FD_NET, "client: got cmd op %s from %s (pdu=%u)\n",
1406                 fio_server_op(cmd->opcode), client->hostname, cmd->pdu_len);
1407
1408         switch (cmd->opcode) {
1409         case FIO_NET_CMD_QUIT:
1410                 if (ops->quit)
1411                         ops->quit(client, cmd);
1412                 remove_client(client);
1413                 break;
1414         case FIO_NET_CMD_TEXT:
1415                 convert_text(cmd);
1416                 ops->text(client, cmd);
1417                 break;
1418         case FIO_NET_CMD_DU: {
1419                 struct cmd_du_pdu *du = (struct cmd_du_pdu *) cmd->payload;
1420
1421                 convert_dus(&du->dus);
1422                 convert_agg(&du->agg);
1423
1424                 ops->disk_util(client, cmd);
1425                 break;
1426                 }
1427         case FIO_NET_CMD_TS: {
1428                 struct cmd_ts_pdu *p = (struct cmd_ts_pdu *) cmd->payload;
1429
1430                 convert_ts(&p->ts, &p->ts);
1431                 convert_gs(&p->rs, &p->rs);
1432
1433                 ops->thread_status(client, cmd);
1434                 break;
1435                 }
1436         case FIO_NET_CMD_GS: {
1437                 struct group_run_stats *gs = (struct group_run_stats *) cmd->payload;
1438
1439                 convert_gs(gs, gs);
1440
1441                 ops->group_stats(client, cmd);
1442                 break;
1443                 }
1444         case FIO_NET_CMD_ETA: {
1445                 struct jobs_eta *je = (struct jobs_eta *) cmd->payload;
1446
1447                 remove_reply_cmd(client, cmd);
1448                 convert_jobs_eta(je);
1449                 handle_eta(client, cmd);
1450                 break;
1451                 }
1452         case FIO_NET_CMD_PROBE:
1453                 remove_reply_cmd(client, cmd);
1454                 ops->probe(client, cmd);
1455                 break;
1456         case FIO_NET_CMD_SERVER_START:
1457                 client->state = Client_running;
1458                 if (ops->job_start)
1459                         ops->job_start(client, cmd);
1460                 break;
1461         case FIO_NET_CMD_START: {
1462                 struct cmd_start_pdu *pdu = (struct cmd_start_pdu *) cmd->payload;
1463
1464                 pdu->jobs = le32_to_cpu(pdu->jobs);
1465                 ops->start(client, cmd);
1466                 break;
1467                 }
1468         case FIO_NET_CMD_STOP: {
1469                 struct cmd_end_pdu *pdu = (struct cmd_end_pdu *) cmd->payload;
1470
1471                 convert_stop(cmd);
1472                 client->state = Client_stopped;
1473                 client->error = le32_to_cpu(pdu->error);
1474                 client->signal = le32_to_cpu(pdu->signal);
1475                 ops->stop(client, cmd);
1476                 break;
1477                 }
1478         case FIO_NET_CMD_ADD_JOB: {
1479                 struct cmd_add_job_pdu *pdu = (struct cmd_add_job_pdu *) cmd->payload;
1480
1481                 client->thread_number = le32_to_cpu(pdu->thread_number);
1482                 client->groupid = le32_to_cpu(pdu->groupid);
1483
1484                 if (ops->add_job)
1485                         ops->add_job(client, cmd);
1486                 break;
1487                 }
1488         case FIO_NET_CMD_IOLOG:
1489                 if (ops->iolog) {
1490                         struct cmd_iolog_pdu *pdu;
1491
1492                         pdu = convert_iolog(cmd);
1493                         ops->iolog(client, pdu);
1494                 }
1495                 break;
1496         case FIO_NET_CMD_UPDATE_JOB:
1497                 ops->update_job(client, cmd);
1498                 remove_reply_cmd(client, cmd);
1499                 break;
1500         case FIO_NET_CMD_VTRIGGER: {
1501                 struct all_io_list *pdu = (struct all_io_list *) cmd->payload;
1502                 char buf[128];
1503                 int off = 0;
1504
1505                 if (aux_path) {
1506                         strcpy(buf, aux_path);
1507                         off = strlen(buf);
1508                 }
1509
1510                 __verify_save_state(pdu, server_name(client, &buf[off], sizeof(buf) - off));
1511                 exec_trigger(trigger_cmd);
1512                 break;
1513                 }
1514         case FIO_NET_CMD_SENDFILE: {
1515                 struct cmd_sendfile *pdu = (struct cmd_sendfile *) cmd->payload;
1516                 send_file(client, pdu, cmd->tag);
1517                 break;
1518                 }
1519         default:
1520                 log_err("fio: unknown client op: %s\n", fio_server_op(cmd->opcode));
1521                 break;
1522         }
1523
1524         free(cmd);
1525         return 1;
1526 }
1527
1528 int fio_clients_send_trigger(const char *cmd)
1529 {
1530         struct flist_head *entry;
1531         struct fio_client *client;
1532         size_t slen;
1533
1534         dprint(FD_NET, "client: send vtrigger: %s\n", cmd);
1535
1536         if (!cmd)
1537                 slen = 0;
1538         else
1539                 slen = strlen(cmd);
1540
1541         flist_for_each(entry, &client_list) {
1542                 struct cmd_vtrigger_pdu *pdu;
1543
1544                 client = flist_entry(entry, struct fio_client, list);
1545
1546                 pdu = malloc(sizeof(*pdu) + slen);
1547                 pdu->len = cpu_to_le16((uint16_t) slen);
1548                 if (slen)
1549                         memcpy(pdu->cmd, cmd, slen);
1550                 fio_net_send_cmd(client->fd, FIO_NET_CMD_VTRIGGER, pdu,
1551                                         sizeof(*pdu) + slen, NULL, NULL);
1552                 free(pdu);
1553         }
1554
1555         return 0;
1556 }
1557
1558 static void request_client_etas(struct client_ops *ops)
1559 {
1560         struct fio_client *client;
1561         struct flist_head *entry;
1562         struct client_eta *eta;
1563         int skipped = 0;
1564
1565         dprint(FD_NET, "client: request eta (%d)\n", nr_clients);
1566
1567         eta = calloc(1, sizeof(*eta) + __THREAD_RUNSTR_SZ(REAL_MAX_JOBS));
1568         eta->pending = nr_clients;
1569
1570         flist_for_each(entry, &client_list) {
1571                 client = flist_entry(entry, struct fio_client, list);
1572
1573                 if (!flist_empty(&client->eta_list)) {
1574                         skipped++;
1575                         continue;
1576                 }
1577                 if (client->state != Client_running)
1578                         continue;
1579
1580                 assert(!client->eta_in_flight);
1581                 flist_add_tail(&client->eta_list, &eta_list);
1582                 client->eta_in_flight = eta;
1583                 fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_SEND_ETA,
1584                                         (uintptr_t) eta, &client->cmd_list);
1585         }
1586
1587         while (skipped--) {
1588                 if (!fio_client_dec_jobs_eta(eta, ops->eta))
1589                         break;
1590         }
1591
1592         dprint(FD_NET, "client: requested eta tag %p\n", eta);
1593 }
1594
1595 /*
1596  * A single SEND_ETA timeout isn't fatal. Attempt to recover.
1597  */
1598 static int handle_cmd_timeout(struct fio_client *client,
1599                               struct fio_net_cmd_reply *reply)
1600 {
1601         if (reply->opcode != FIO_NET_CMD_SEND_ETA)
1602                 return 1;
1603
1604         log_info("client <%s>: timeout on SEND_ETA\n", client->hostname);
1605         flist_del(&reply->list);
1606         free(reply);
1607
1608         flist_del_init(&client->eta_list);
1609         if (client->eta_in_flight) {
1610                 fio_client_dec_jobs_eta(client->eta_in_flight, client->ops->eta);
1611                 client->eta_in_flight = NULL;
1612         }
1613
1614         /*
1615          * If we fail 5 in a row, give up...
1616          */
1617         if (client->eta_timeouts++ > 5)
1618                 return 1;
1619
1620         return 0;
1621 }
1622
1623 static int client_check_cmd_timeout(struct fio_client *client,
1624                                     struct timeval *now)
1625 {
1626         struct fio_net_cmd_reply *reply;
1627         struct flist_head *entry, *tmp;
1628         int ret = 0;
1629
1630         flist_for_each_safe(entry, tmp, &client->cmd_list) {
1631                 reply = flist_entry(entry, struct fio_net_cmd_reply, list);
1632
1633                 if (mtime_since(&reply->tv, now) < FIO_NET_CLIENT_TIMEOUT)
1634                         continue;
1635
1636                 if (!handle_cmd_timeout(client, reply))
1637                         continue;
1638
1639                 log_err("fio: client %s, timeout on cmd %s\n", client->hostname,
1640                                                 fio_server_op(reply->opcode));
1641                 flist_del(&reply->list);
1642                 free(reply);
1643                 ret = 1;
1644         }
1645
1646         return flist_empty(&client->cmd_list) && ret;
1647 }
1648
1649 static int fio_check_clients_timed_out(void)
1650 {
1651         struct fio_client *client;
1652         struct flist_head *entry, *tmp;
1653         struct timeval tv;
1654         int ret = 0;
1655
1656         fio_gettime(&tv, NULL);
1657
1658         flist_for_each_safe(entry, tmp, &client_list) {
1659                 client = flist_entry(entry, struct fio_client, list);
1660
1661                 if (flist_empty(&client->cmd_list))
1662                         continue;
1663
1664                 if (!client_check_cmd_timeout(client, &tv))
1665                         continue;
1666
1667                 if (client->ops->timed_out)
1668                         client->ops->timed_out(client);
1669                 else
1670                         log_err("fio: client %s timed out\n", client->hostname);
1671
1672                 client->error = ETIMEDOUT;
1673                 remove_client(client);
1674                 ret = 1;
1675         }
1676
1677         return ret;
1678 }
1679
1680 int fio_handle_clients(struct client_ops *ops)
1681 {
1682         struct pollfd *pfds;
1683         int i, ret = 0, retval = 0;
1684
1685         fio_gettime(&eta_tv, NULL);
1686
1687         pfds = malloc(nr_clients * sizeof(struct pollfd));
1688
1689         init_thread_stat(&client_ts);
1690         init_group_run_stat(&client_gs);
1691
1692         while (!exit_backend && nr_clients) {
1693                 struct flist_head *entry, *tmp;
1694                 struct fio_client *client;
1695
1696                 i = 0;
1697                 flist_for_each_safe(entry, tmp, &client_list) {
1698                         client = flist_entry(entry, struct fio_client, list);
1699
1700                         if (!client->sent_job && !client->ops->stay_connected &&
1701                             flist_empty(&client->cmd_list)) {
1702                                 remove_client(client);
1703                                 continue;
1704                         }
1705
1706                         pfds[i].fd = client->fd;
1707                         pfds[i].events = POLLIN;
1708                         i++;
1709                 }
1710
1711                 if (!nr_clients)
1712                         break;
1713
1714                 assert(i == nr_clients);
1715
1716                 do {
1717                         struct timeval tv;
1718                         int timeout;
1719
1720                         fio_gettime(&tv, NULL);
1721                         if (mtime_since(&eta_tv, &tv) >= 900) {
1722                                 request_client_etas(ops);
1723                                 memcpy(&eta_tv, &tv, sizeof(tv));
1724
1725                                 if (fio_check_clients_timed_out())
1726                                         break;
1727                         }
1728
1729                         check_trigger_file();
1730
1731                         timeout = min(100u, ops->eta_msec);
1732
1733                         ret = poll(pfds, nr_clients, timeout);
1734                         if (ret < 0) {
1735                                 if (errno == EINTR)
1736                                         continue;
1737                                 log_err("fio: poll clients: %s\n", strerror(errno));
1738                                 break;
1739                         } else if (!ret)
1740                                 continue;
1741                 } while (ret <= 0);
1742
1743                 for (i = 0; i < nr_clients; i++) {
1744                         if (!(pfds[i].revents & POLLIN))
1745                                 continue;
1746
1747                         client = find_client_by_fd(pfds[i].fd);
1748                         if (!client) {
1749                                 log_err("fio: unknown client fd %ld\n", (long) pfds[i].fd);
1750                                 continue;
1751                         }
1752                         if (!fio_handle_client(client)) {
1753                                 log_info("client: host=%s disconnected\n",
1754                                                 client->hostname);
1755                                 remove_client(client);
1756                                 retval = 1;
1757                         } else if (client->error)
1758                                 retval = 1;
1759                         fio_put_client(client);
1760                 }
1761         }
1762
1763         fio_client_json_fini();
1764
1765         free(pfds);
1766         return retval || error_clients;
1767 }