fio: groundwork for adding slat, lat percentiles
[fio.git] / client.c
1 #include <stdio.h>
2 #include <stdlib.h>
3 #include <unistd.h>
4 #include <errno.h>
5 #include <fcntl.h>
6 #include <poll.h>
7 #include <sys/types.h>
8 #include <sys/stat.h>
9 #include <sys/socket.h>
10 #include <sys/un.h>
11 #include <netinet/in.h>
12 #include <arpa/inet.h>
13 #include <netdb.h>
14 #include <signal.h>
15 #ifdef CONFIG_ZLIB
16 #include <zlib.h>
17 #endif
18
19 #include "fio.h"
20 #include "client.h"
21 #include "server.h"
22 #include "flist.h"
23 #include "hash.h"
24 #include "verify-state.h"
25
26 static void handle_du(struct fio_client *client, struct fio_net_cmd *cmd);
27 static void handle_ts(struct fio_client *client, struct fio_net_cmd *cmd);
28 static void handle_gs(struct fio_client *client, struct fio_net_cmd *cmd);
29 static void handle_probe(struct fio_client *client, struct fio_net_cmd *cmd);
30 static void handle_text(struct fio_client *client, struct fio_net_cmd *cmd);
31 static void handle_stop(struct fio_client *client);
32 static void handle_start(struct fio_client *client, struct fio_net_cmd *cmd);
33
34 static void convert_text(struct fio_net_cmd *cmd);
35 static void client_display_thread_status(struct jobs_eta *je);
36
37 struct client_ops fio_client_ops = {
38         .text           = handle_text,
39         .disk_util      = handle_du,
40         .thread_status  = handle_ts,
41         .group_stats    = handle_gs,
42         .stop           = handle_stop,
43         .start          = handle_start,
44         .eta            = client_display_thread_status,
45         .probe          = handle_probe,
46         .eta_msec       = FIO_CLIENT_DEF_ETA_MSEC,
47         .client_type    = FIO_CLIENT_TYPE_CLI,
48 };
49
50 static struct timespec eta_ts;
51
52 static FLIST_HEAD(client_list);
53 static FLIST_HEAD(eta_list);
54
55 static FLIST_HEAD(arg_list);
56
57 struct thread_stat client_ts;
58 struct group_run_stats client_gs;
59 int sum_stat_clients;
60
61 static int sum_stat_nr;
62 static struct buf_output allclients;
63 static struct json_object *root = NULL;
64 static struct json_object *job_opt_object = NULL;
65 static struct json_array *clients_array = NULL;
66 static struct json_array *du_array = NULL;
67
68 static int error_clients;
69
70 #define FIO_CLIENT_HASH_BITS    7
71 #define FIO_CLIENT_HASH_SZ      (1 << FIO_CLIENT_HASH_BITS)
72 #define FIO_CLIENT_HASH_MASK    (FIO_CLIENT_HASH_SZ - 1)
73 static struct flist_head client_hash[FIO_CLIENT_HASH_SZ];
74
75 static struct cmd_iolog_pdu *convert_iolog(struct fio_net_cmd *, bool *);
76
77 static void fio_client_add_hash(struct fio_client *client)
78 {
79         int bucket = hash_long(client->fd, FIO_CLIENT_HASH_BITS);
80
81         bucket &= FIO_CLIENT_HASH_MASK;
82         flist_add(&client->hash_list, &client_hash[bucket]);
83 }
84
85 static void fio_client_remove_hash(struct fio_client *client)
86 {
87         if (!flist_empty(&client->hash_list))
88                 flist_del_init(&client->hash_list);
89 }
90
91 static void fio_init fio_client_hash_init(void)
92 {
93         int i;
94
95         for (i = 0; i < FIO_CLIENT_HASH_SZ; i++)
96                 INIT_FLIST_HEAD(&client_hash[i]);
97 }
98
99 static int read_data(int fd, void *data, size_t size)
100 {
101         ssize_t ret;
102
103         while (size) {
104                 ret = read(fd, data, size);
105                 if (ret < 0) {
106                         if (errno == EAGAIN || errno == EINTR)
107                                 continue;
108                         break;
109                 } else if (!ret)
110                         break;
111                 else {
112                         data += ret;
113                         size -= ret;
114                 }
115         }
116
117         if (size)
118                 return EAGAIN;
119
120         return 0;
121 }
122
123 static int read_ini_data(int fd, void *data, size_t size)
124 {
125         char *p = data;
126         int ret = 0;
127         FILE *fp;
128         int dupfd;
129
130         dupfd = dup(fd);
131         if (dupfd < 0)
132                 return errno;
133
134         fp = fdopen(dupfd, "r");
135         if (!fp) {
136                 ret = errno;
137                 close(dupfd);
138                 goto out;
139         }
140
141         while (1) {
142                 ssize_t len;
143                 char buf[OPT_LEN_MAX+1], *sub;
144
145                 if (!fgets(buf, sizeof(buf), fp)) {
146                         if (ferror(fp)) {
147                                 if (errno == EAGAIN || errno == EINTR)
148                                         continue;
149                                 ret = errno;
150                         }
151                         break;
152                 }
153
154                 sub = fio_option_dup_subs(buf);
155                 len = strlen(sub);
156                 if (len + 1 > size) {
157                         log_err("fio: no space left to read data\n");
158                         free(sub);
159                         ret = ENOSPC;
160                         break;
161                 }
162
163                 memcpy(p, sub, len);
164                 free(sub);
165                 p += len;
166                 *p = '\0';
167                 size -= len;
168         }
169
170         fclose(fp);
171 out:
172         return ret;
173 }
174
175 static void fio_client_json_init(void)
176 {
177         char time_buf[32];
178         time_t time_p;
179
180         if (!(output_format & FIO_OUTPUT_JSON))
181                 return;
182
183         time(&time_p);
184         os_ctime_r((const time_t *) &time_p, time_buf, sizeof(time_buf));
185         time_buf[strlen(time_buf) - 1] = '\0';
186
187         root = json_create_object();
188         json_object_add_value_string(root, "fio version", fio_version_string);
189         json_object_add_value_int(root, "timestamp", time_p);
190         json_object_add_value_string(root, "time", time_buf);
191
192         job_opt_object = json_create_object();
193         json_object_add_value_object(root, "global options", job_opt_object);
194         clients_array = json_create_array();
195         json_object_add_value_array(root, "client_stats", clients_array);
196         du_array = json_create_array();
197         json_object_add_value_array(root, "disk_util", du_array);
198 }
199
200 static void fio_client_json_fini(void)
201 {
202         struct buf_output out;
203
204         if (!root)
205                 return;
206
207         buf_output_init(&out);
208
209         __log_buf(&out, "\n");
210         json_print_object(root, &out);
211         __log_buf(&out, "\n");
212         log_info_buf(out.buf, out.buflen);
213
214         buf_output_free(&out);
215
216         json_free_object(root);
217         root = NULL;
218         job_opt_object = NULL;
219         clients_array = NULL;
220         du_array = NULL;
221 }
222
223 static struct fio_client *find_client_by_fd(int fd)
224 {
225         int bucket = hash_long(fd, FIO_CLIENT_HASH_BITS) & FIO_CLIENT_HASH_MASK;
226         struct fio_client *client;
227         struct flist_head *entry;
228
229         flist_for_each(entry, &client_hash[bucket]) {
230                 client = flist_entry(entry, struct fio_client, hash_list);
231
232                 if (client->fd == fd) {
233                         client->refs++;
234                         return client;
235                 }
236         }
237
238         return NULL;
239 }
240
241 void fio_put_client(struct fio_client *client)
242 {
243         if (--client->refs)
244                 return;
245
246         log_info_buf(client->buf.buf, client->buf.buflen);
247         buf_output_free(&client->buf);
248
249         free(client->hostname);
250         if (client->argv)
251                 free(client->argv);
252         if (client->name)
253                 free(client->name);
254         while (client->nr_files) {
255                 struct client_file *cf = &client->files[--client->nr_files];
256
257                 free(cf->file);
258         }
259         if (client->files)
260                 free(client->files);
261         if (client->opt_lists)
262                 free(client->opt_lists);
263
264         if (!client->did_stat)
265                 sum_stat_clients--;
266
267         if (client->error)
268                 error_clients++;
269
270         free(client);
271 }
272
273 static int fio_client_dec_jobs_eta(struct client_eta *eta, client_eta_op eta_fn)
274 {
275         if (!--eta->pending) {
276                 eta_fn(&eta->eta);
277                 free(eta);
278                 return 0;
279         }
280
281         return 1;
282 }
283
284 static void fio_drain_client_text(struct fio_client *client)
285 {
286         do {
287                 struct fio_net_cmd *cmd;
288
289                 cmd = fio_net_recv_cmd(client->fd, false);
290                 if (!cmd)
291                         break;
292
293                 if (cmd->opcode == FIO_NET_CMD_TEXT) {
294                         convert_text(cmd);
295                         client->ops->text(client, cmd);
296                 }
297
298                 free(cmd);
299         } while (1);
300 }
301
302 static void remove_client(struct fio_client *client)
303 {
304         assert(client->refs);
305
306         dprint(FD_NET, "client: removed <%s>\n", client->hostname);
307
308         fio_drain_client_text(client);
309
310         if (!flist_empty(&client->list))
311                 flist_del_init(&client->list);
312
313         fio_client_remove_hash(client);
314
315         if (!flist_empty(&client->eta_list)) {
316                 flist_del_init(&client->eta_list);
317                 fio_client_dec_jobs_eta(client->eta_in_flight, client->ops->eta);
318         }
319
320         close(client->fd);
321         client->fd = -1;
322
323         if (client->ops->removed)
324                 client->ops->removed(client);
325
326         nr_clients--;
327         fio_put_client(client);
328 }
329
330 struct fio_client *fio_get_client(struct fio_client *client)
331 {
332         client->refs++;
333         return client;
334 }
335
336 static void __fio_client_add_cmd_option(struct fio_client *client,
337                                         const char *opt)
338 {
339         int index;
340
341         index = client->argc++;
342         client->argv = realloc(client->argv, sizeof(char *) * client->argc);
343         client->argv[index] = strdup(opt);
344         dprint(FD_NET, "client: add cmd %d: %s\n", index, opt);
345 }
346
347 void fio_client_add_cmd_option(void *cookie, const char *opt)
348 {
349         struct fio_client *client = cookie;
350         struct flist_head *entry;
351
352         if (!client || !opt)
353                 return;
354
355         __fio_client_add_cmd_option(client, opt);
356
357         /*
358          * Duplicate arguments to shared client group
359          */
360         flist_for_each(entry, &arg_list) {
361                 client = flist_entry(entry, struct fio_client, arg_list);
362
363                 __fio_client_add_cmd_option(client, opt);
364         }
365 }
366
367 static struct fio_client *get_new_client(void)
368 {
369         struct fio_client *client;
370
371         client = malloc(sizeof(*client));
372         memset(client, 0, sizeof(*client));
373
374         INIT_FLIST_HEAD(&client->list);
375         INIT_FLIST_HEAD(&client->hash_list);
376         INIT_FLIST_HEAD(&client->arg_list);
377         INIT_FLIST_HEAD(&client->eta_list);
378         INIT_FLIST_HEAD(&client->cmd_list);
379
380         buf_output_init(&client->buf);
381
382         return client;
383 }
384
385 struct fio_client *fio_client_add_explicit(struct client_ops *ops,
386                                            const char *hostname, int type,
387                                            int port)
388 {
389         struct fio_client *client;
390
391         client = get_new_client();
392
393         client->hostname = strdup(hostname);
394
395         if (type == Fio_client_socket)
396                 client->is_sock = true;
397         else {
398                 int ipv6;
399
400                 ipv6 = type == Fio_client_ipv6;
401                 if (fio_server_parse_host(hostname, ipv6,
402                                                 &client->addr.sin_addr,
403                                                 &client->addr6.sin6_addr))
404                         goto err;
405
406                 client->port = port;
407         }
408
409         client->fd = -1;
410         client->ops = ops;
411         client->refs = 1;
412         client->type = ops->client_type;
413
414         __fio_client_add_cmd_option(client, "fio");
415
416         flist_add(&client->list, &client_list);
417         nr_clients++;
418         dprint(FD_NET, "client: added <%s>\n", client->hostname);
419         return client;
420 err:
421         free(client);
422         return NULL;
423 }
424
425 int fio_client_add_ini_file(void *cookie, const char *ini_file, bool remote)
426 {
427         struct fio_client *client = cookie;
428         struct client_file *cf;
429         size_t new_size;
430         void *new_files;
431
432         if (!client)
433                 return 1;
434
435         dprint(FD_NET, "client <%s>: add ini %s\n", client->hostname, ini_file);
436
437         new_size = (client->nr_files + 1) * sizeof(struct client_file);
438         new_files = realloc(client->files, new_size);
439         if (!new_files)
440                 return 1;
441
442         client->files = new_files;
443         cf = &client->files[client->nr_files];
444         cf->file = strdup(ini_file);
445         cf->remote = remote;
446         client->nr_files++;
447         return 0;
448 }
449
450 int fio_client_add(struct client_ops *ops, const char *hostname, void **cookie)
451 {
452         struct fio_client *existing = *cookie;
453         struct fio_client *client;
454
455         if (existing) {
456                 /*
457                  * We always add our "exec" name as the option, hence 1
458                  * means empty.
459                  */
460                 if (existing->argc == 1)
461                         flist_add_tail(&existing->arg_list, &arg_list);
462                 else {
463                         while (!flist_empty(&arg_list))
464                                 flist_del_init(arg_list.next);
465                 }
466         }
467
468         client = get_new_client();
469
470         if (fio_server_parse_string(hostname, &client->hostname,
471                                         &client->is_sock, &client->port,
472                                         &client->addr.sin_addr,
473                                         &client->addr6.sin6_addr,
474                                         &client->ipv6))
475                 return -1;
476
477         client->fd = -1;
478         client->ops = ops;
479         client->refs = 1;
480         client->type = ops->client_type;
481
482         __fio_client_add_cmd_option(client, "fio");
483
484         flist_add(&client->list, &client_list);
485         nr_clients++;
486         dprint(FD_NET, "client: added <%s>\n", client->hostname);
487         *cookie = client;
488         return 0;
489 }
490
491 static const char *server_name(struct fio_client *client, char *buf,
492                                size_t bufsize)
493 {
494         const char *from;
495
496         if (client->ipv6)
497                 from = inet_ntop(AF_INET6, (struct sockaddr *) &client->addr6.sin6_addr, buf, bufsize);
498         else if (client->is_sock)
499                 from = "sock";
500         else
501                 from = inet_ntop(AF_INET, (struct sockaddr *) &client->addr.sin_addr, buf, bufsize);
502
503         return from;
504 }
505
506 static void probe_client(struct fio_client *client)
507 {
508         struct cmd_client_probe_pdu pdu;
509         const char *sname;
510         uint64_t tag;
511         char buf[64];
512
513         dprint(FD_NET, "client: send probe\n");
514
515 #ifdef CONFIG_ZLIB
516         pdu.flags = __le64_to_cpu(FIO_PROBE_FLAG_ZLIB);
517 #else
518         pdu.flags = 0;
519 #endif
520
521         sname = server_name(client, buf, sizeof(buf));
522         memset(pdu.server, 0, sizeof(pdu.server));
523         snprintf((char *) pdu.server, sizeof(pdu.server), "%s", sname);
524
525         fio_net_send_cmd(client->fd, FIO_NET_CMD_PROBE, &pdu, sizeof(pdu), &tag, &client->cmd_list);
526 }
527
528 static int fio_client_connect_ip(struct fio_client *client)
529 {
530         struct sockaddr *addr;
531         socklen_t socklen;
532         int fd, domain;
533
534         if (client->ipv6) {
535                 client->addr6.sin6_family = AF_INET6;
536                 client->addr6.sin6_port = htons(client->port);
537                 domain = AF_INET6;
538                 addr = (struct sockaddr *) &client->addr6;
539                 socklen = sizeof(client->addr6);
540         } else {
541                 client->addr.sin_family = AF_INET;
542                 client->addr.sin_port = htons(client->port);
543                 domain = AF_INET;
544                 addr = (struct sockaddr *) &client->addr;
545                 socklen = sizeof(client->addr);
546         }
547
548         fd = socket(domain, SOCK_STREAM, 0);
549         if (fd < 0) {
550                 int ret = -errno;
551
552                 log_err("fio: socket: %s\n", strerror(errno));
553                 return ret;
554         }
555
556         if (connect(fd, addr, socklen) < 0) {
557                 int ret = -errno;
558
559                 log_err("fio: connect: %s\n", strerror(errno));
560                 log_err("fio: failed to connect to %s:%u\n", client->hostname,
561                                                                 client->port);
562                 close(fd);
563                 return ret;
564         }
565
566         return fd;
567 }
568
569 static int fio_client_connect_sock(struct fio_client *client)
570 {
571         struct sockaddr_un *addr = &client->addr_un;
572         socklen_t len;
573         int fd;
574
575         memset(addr, 0, sizeof(*addr));
576         addr->sun_family = AF_UNIX;
577         snprintf(addr->sun_path, sizeof(addr->sun_path), "%s",
578                  client->hostname);
579
580         fd = socket(AF_UNIX, SOCK_STREAM, 0);
581         if (fd < 0) {
582                 int ret = -errno;
583
584                 log_err("fio: socket: %s\n", strerror(errno));
585                 return ret;
586         }
587
588         len = sizeof(addr->sun_family) + strlen(addr->sun_path) + 1;
589         if (connect(fd, (struct sockaddr *) addr, len) < 0) {
590                 int ret = -errno;
591
592                 log_err("fio: connect; %s\n", strerror(errno));
593                 close(fd);
594                 return ret;
595         }
596
597         return fd;
598 }
599
600 int fio_client_connect(struct fio_client *client)
601 {
602         int fd;
603
604         dprint(FD_NET, "client: connect to host %s\n", client->hostname);
605
606         if (client->is_sock)
607                 fd = fio_client_connect_sock(client);
608         else
609                 fd = fio_client_connect_ip(client);
610
611         dprint(FD_NET, "client: %s connected %d\n", client->hostname, fd);
612
613         if (fd < 0)
614                 return fd;
615
616         client->fd = fd;
617         fio_client_add_hash(client);
618         client->state = Client_connected;
619
620         probe_client(client);
621         return 0;
622 }
623
624 int fio_client_terminate(struct fio_client *client)
625 {
626         return fio_net_send_quit(client->fd);
627 }
628
629 static void fio_clients_terminate(void)
630 {
631         struct flist_head *entry;
632         struct fio_client *client;
633
634         dprint(FD_NET, "client: terminate clients\n");
635
636         flist_for_each(entry, &client_list) {
637                 client = flist_entry(entry, struct fio_client, list);
638                 fio_client_terminate(client);
639         }
640 }
641
642 static void sig_int(int sig)
643 {
644         dprint(FD_NET, "client: got signal %d\n", sig);
645         fio_clients_terminate();
646 }
647
648 static void client_signal_handler(void)
649 {
650         struct sigaction act;
651
652         memset(&act, 0, sizeof(act));
653         act.sa_handler = sig_int;
654         act.sa_flags = SA_RESTART;
655         sigaction(SIGINT, &act, NULL);
656
657         memset(&act, 0, sizeof(act));
658         act.sa_handler = sig_int;
659         act.sa_flags = SA_RESTART;
660         sigaction(SIGTERM, &act, NULL);
661
662 /* Windows uses SIGBREAK as a quit signal from other applications */
663 #ifdef WIN32
664         memset(&act, 0, sizeof(act));
665         act.sa_handler = sig_int;
666         act.sa_flags = SA_RESTART;
667         sigaction(SIGBREAK, &act, NULL);
668 #endif
669
670         memset(&act, 0, sizeof(act));
671         act.sa_handler = sig_show_status;
672         act.sa_flags = SA_RESTART;
673         sigaction(SIGUSR1, &act, NULL);
674 }
675
676 static int send_client_cmd_line(struct fio_client *client)
677 {
678         struct cmd_single_line_pdu *cslp;
679         struct cmd_line_pdu *clp;
680         unsigned long offset;
681         unsigned int *lens;
682         void *pdu;
683         size_t mem;
684         int i, ret;
685
686         dprint(FD_NET, "client: send cmdline %d\n", client->argc);
687
688         lens = malloc(client->argc * sizeof(unsigned int));
689
690         /*
691          * Find out how much mem we need
692          */
693         for (i = 0, mem = 0; i < client->argc; i++) {
694                 lens[i] = strlen(client->argv[i]) + 1;
695                 mem += lens[i];
696         }
697
698         /*
699          * We need one cmd_line_pdu, and argc number of cmd_single_line_pdu
700          */
701         mem += sizeof(*clp) + (client->argc * sizeof(*cslp));
702
703         pdu = malloc(mem);
704         clp = pdu;
705         offset = sizeof(*clp);
706
707         for (i = 0; i < client->argc; i++) {
708                 uint16_t arg_len = lens[i];
709
710                 cslp = pdu + offset;
711                 strcpy((char *) cslp->text, client->argv[i]);
712                 cslp->len = cpu_to_le16(arg_len);
713                 offset += sizeof(*cslp) + arg_len;
714         }
715
716         free(lens);
717         clp->lines = cpu_to_le16(client->argc);
718         clp->client_type = __cpu_to_le16(client->type);
719         ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOBLINE, pdu, mem, NULL, NULL);
720         free(pdu);
721         return ret;
722 }
723
724 int fio_clients_connect(void)
725 {
726         struct fio_client *client;
727         struct flist_head *entry, *tmp;
728         int ret;
729
730 #ifdef WIN32
731         WSADATA wsd;
732         WSAStartup(MAKEWORD(2, 2), &wsd);
733 #endif
734
735         dprint(FD_NET, "client: connect all\n");
736
737         client_signal_handler();
738
739         flist_for_each_safe(entry, tmp, &client_list) {
740                 client = flist_entry(entry, struct fio_client, list);
741
742                 ret = fio_client_connect(client);
743                 if (ret) {
744                         remove_client(client);
745                         continue;
746                 }
747
748                 if (client->argc > 1)
749                         send_client_cmd_line(client);
750         }
751
752         return !nr_clients;
753 }
754
755 int fio_start_client(struct fio_client *client)
756 {
757         dprint(FD_NET, "client: start %s\n", client->hostname);
758         return fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_RUN, 0, NULL);
759 }
760
761 int fio_start_all_clients(void)
762 {
763         struct fio_client *client;
764         struct flist_head *entry, *tmp;
765         int ret;
766
767         dprint(FD_NET, "client: start all\n");
768
769         fio_client_json_init();
770
771         flist_for_each_safe(entry, tmp, &client_list) {
772                 client = flist_entry(entry, struct fio_client, list);
773
774                 ret = fio_start_client(client);
775                 if (ret) {
776                         remove_client(client);
777                         continue;
778                 }
779         }
780
781         return flist_empty(&client_list);
782 }
783
784 static int __fio_client_send_remote_ini(struct fio_client *client,
785                                         const char *filename)
786 {
787         struct cmd_load_file_pdu *pdu;
788         size_t p_size;
789         int ret;
790
791         dprint(FD_NET, "send remote ini %s to %s\n", filename, client->hostname);
792
793         p_size = sizeof(*pdu) + strlen(filename) + 1;
794         pdu = malloc(p_size);
795         memset(pdu, 0, p_size);
796         pdu->name_len = strlen(filename);
797         strcpy((char *) pdu->file, filename);
798         pdu->client_type = cpu_to_le16((uint16_t) client->type);
799
800         client->sent_job = true;
801         ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_LOAD_FILE, pdu, p_size,NULL, NULL);
802         free(pdu);
803         return ret;
804 }
805
806 /*
807  * Send file contents to server backend. We could use sendfile(), but to remain
808  * more portable lets just read/write the darn thing.
809  */
810 static int __fio_client_send_local_ini(struct fio_client *client,
811                                        const char *filename)
812 {
813         struct cmd_job_pdu *pdu;
814         size_t p_size;
815         struct stat sb;
816         char *p;
817         void *buf;
818         off_t len;
819         int fd, ret;
820
821         dprint(FD_NET, "send ini %s to %s\n", filename, client->hostname);
822
823         fd = open(filename, O_RDONLY);
824         if (fd < 0) {
825                 ret = -errno;
826                 log_err("fio: job file <%s> open: %s\n", filename, strerror(errno));
827                 return ret;
828         }
829
830         if (fstat(fd, &sb) < 0) {
831                 ret = -errno;
832                 log_err("fio: job file stat: %s\n", strerror(errno));
833                 close(fd);
834                 return ret;
835         }
836
837         /*
838          * Add extra space for variable expansion, but doesn't guarantee.
839          */
840         sb.st_size += OPT_LEN_MAX;
841         p_size = sb.st_size + sizeof(*pdu);
842         pdu = malloc(p_size);
843         buf = pdu->buf;
844
845         len = sb.st_size;
846         p = buf;
847         if (read_ini_data(fd, p, len)) {
848                 log_err("fio: failed reading job file %s\n", filename);
849                 close(fd);
850                 free(pdu);
851                 return 1;
852         }
853
854         pdu->buf_len = __cpu_to_le32(sb.st_size);
855         pdu->client_type = cpu_to_le32(client->type);
856
857         client->sent_job = true;
858         ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOB, pdu, p_size, NULL, NULL);
859         free(pdu);
860         close(fd);
861         return ret;
862 }
863
864 int fio_client_send_ini(struct fio_client *client, const char *filename,
865                         bool remote)
866 {
867         int ret;
868
869         if (!remote)
870                 ret = __fio_client_send_local_ini(client, filename);
871         else
872                 ret = __fio_client_send_remote_ini(client, filename);
873
874         if (!ret)
875                 client->sent_job = true;
876
877         return ret;
878 }
879
880 static int fio_client_send_cf(struct fio_client *client,
881                               struct client_file *cf)
882 {
883         return fio_client_send_ini(client, cf->file, cf->remote);
884 }
885
886 int fio_clients_send_ini(const char *filename)
887 {
888         struct fio_client *client;
889         struct flist_head *entry, *tmp;
890
891         flist_for_each_safe(entry, tmp, &client_list) {
892                 bool failed = false;
893
894                 client = flist_entry(entry, struct fio_client, list);
895
896                 if (client->nr_files) {
897                         int i;
898
899                         for (i = 0; i < client->nr_files; i++) {
900                                 struct client_file *cf;
901
902                                 cf = &client->files[i];
903
904                                 if (fio_client_send_cf(client, cf)) {
905                                         failed = true;
906                                         remove_client(client);
907                                         break;
908                                 }
909                         }
910                 }
911                 if (client->sent_job || failed)
912                         continue;
913                 if (!filename || fio_client_send_ini(client, filename, 0))
914                         remove_client(client);
915         }
916
917         return !nr_clients;
918 }
919
920 int fio_client_update_options(struct fio_client *client,
921                               struct thread_options *o, uint64_t *tag)
922 {
923         struct cmd_add_job_pdu pdu;
924
925         pdu.thread_number = cpu_to_le32(client->thread_number);
926         pdu.groupid = cpu_to_le32(client->groupid);
927         convert_thread_options_to_net(&pdu.top, o);
928
929         return fio_net_send_cmd(client->fd, FIO_NET_CMD_UPDATE_JOB, &pdu, sizeof(pdu), tag, &client->cmd_list);
930 }
931
932 static void convert_io_stat(struct io_stat *dst, struct io_stat *src)
933 {
934         dst->max_val    = le64_to_cpu(src->max_val);
935         dst->min_val    = le64_to_cpu(src->min_val);
936         dst->samples    = le64_to_cpu(src->samples);
937
938         /*
939          * Floats arrive as IEEE 754 encoded uint64_t, convert back to double
940          */
941         dst->mean.u.f   = fio_uint64_to_double(le64_to_cpu(dst->mean.u.i));
942         dst->S.u.f      = fio_uint64_to_double(le64_to_cpu(dst->S.u.i));
943 }
944
945 static void convert_ts(struct thread_stat *dst, struct thread_stat *src)
946 {
947         int i, j, k;
948
949         dst->error              = le32_to_cpu(src->error);
950         dst->thread_number      = le32_to_cpu(src->thread_number);
951         dst->groupid            = le32_to_cpu(src->groupid);
952         dst->pid                = le32_to_cpu(src->pid);
953         dst->members            = le32_to_cpu(src->members);
954         dst->unified_rw_rep     = le32_to_cpu(src->unified_rw_rep);
955
956         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
957                 convert_io_stat(&dst->clat_stat[i], &src->clat_stat[i]);
958                 convert_io_stat(&dst->slat_stat[i], &src->slat_stat[i]);
959                 convert_io_stat(&dst->lat_stat[i], &src->lat_stat[i]);
960                 convert_io_stat(&dst->bw_stat[i], &src->bw_stat[i]);
961                 convert_io_stat(&dst->iops_stat[i], &src->iops_stat[i]);
962         }
963         convert_io_stat(&dst->sync_stat, &src->sync_stat);
964
965         dst->usr_time           = le64_to_cpu(src->usr_time);
966         dst->sys_time           = le64_to_cpu(src->sys_time);
967         dst->ctx                = le64_to_cpu(src->ctx);
968         dst->minf               = le64_to_cpu(src->minf);
969         dst->majf               = le64_to_cpu(src->majf);
970         dst->clat_percentiles   = le32_to_cpu(src->clat_percentiles);
971         dst->lat_percentiles    = le32_to_cpu(src->lat_percentiles);
972         dst->percentile_precision = le64_to_cpu(src->percentile_precision);
973
974         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
975                 fio_fp64_t *fps = &src->percentile_list[i];
976                 fio_fp64_t *fpd = &dst->percentile_list[i];
977
978                 fpd->u.f = fio_uint64_to_double(le64_to_cpu(fps->u.i));
979         }
980
981         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
982                 dst->io_u_map[i]        = le64_to_cpu(src->io_u_map[i]);
983                 dst->io_u_submit[i]     = le64_to_cpu(src->io_u_submit[i]);
984                 dst->io_u_complete[i]   = le64_to_cpu(src->io_u_complete[i]);
985         }
986
987         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
988                 dst->io_u_lat_n[i]      = le64_to_cpu(src->io_u_lat_n[i]);
989         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
990                 dst->io_u_lat_u[i]      = le64_to_cpu(src->io_u_lat_u[i]);
991         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
992                 dst->io_u_lat_m[i]      = le64_to_cpu(src->io_u_lat_m[i]);
993
994         for (i = 0; i < FIO_LAT_CNT; i++)
995                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
996                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
997                                 dst->io_u_plat[i][j][k] = le64_to_cpu(src->io_u_plat[i][j][k]);
998
999         for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1000                 dst->io_u_sync_plat[j] = le64_to_cpu(src->io_u_sync_plat[j]);
1001
1002         for (i = 0; i < DDIR_RWDIR_SYNC_CNT; i++)
1003                 dst->total_io_u[i]      = le64_to_cpu(src->total_io_u[i]);
1004
1005         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1006                 dst->short_io_u[i]      = le64_to_cpu(src->short_io_u[i]);
1007                 dst->drop_io_u[i]       = le64_to_cpu(src->drop_io_u[i]);
1008         }
1009
1010         dst->total_submit       = le64_to_cpu(src->total_submit);
1011         dst->total_complete     = le64_to_cpu(src->total_complete);
1012         dst->nr_zone_resets     = le64_to_cpu(src->nr_zone_resets);
1013
1014         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1015                 dst->io_bytes[i]        = le64_to_cpu(src->io_bytes[i]);
1016                 dst->runtime[i]         = le64_to_cpu(src->runtime[i]);
1017         }
1018
1019         dst->total_run_time     = le64_to_cpu(src->total_run_time);
1020         dst->continue_on_error  = le16_to_cpu(src->continue_on_error);
1021         dst->total_err_count    = le64_to_cpu(src->total_err_count);
1022         dst->first_error        = le32_to_cpu(src->first_error);
1023         dst->kb_base            = le32_to_cpu(src->kb_base);
1024         dst->unit_base          = le32_to_cpu(src->unit_base);
1025
1026         dst->sig_figs           = le32_to_cpu(src->sig_figs);
1027
1028         dst->latency_depth      = le32_to_cpu(src->latency_depth);
1029         dst->latency_target     = le64_to_cpu(src->latency_target);
1030         dst->latency_window     = le64_to_cpu(src->latency_window);
1031         dst->latency_percentile.u.f = fio_uint64_to_double(le64_to_cpu(src->latency_percentile.u.i));
1032
1033         dst->nr_block_infos     = le64_to_cpu(src->nr_block_infos);
1034         for (i = 0; i < dst->nr_block_infos; i++)
1035                 dst->block_infos[i] = le32_to_cpu(src->block_infos[i]);
1036         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1037                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
1038                         dst->io_u_plat_high_prio[i][j] = le64_to_cpu(src->io_u_plat_high_prio[i][j]);
1039                         dst->io_u_plat_prio[i][j] = le64_to_cpu(src->io_u_plat_prio[i][j]);
1040                 }
1041                 convert_io_stat(&dst->clat_high_prio_stat[i], &src->clat_high_prio_stat[i]);
1042                 convert_io_stat(&dst->clat_prio_stat[i], &src->clat_prio_stat[i]);
1043         }
1044
1045         dst->ss_dur             = le64_to_cpu(src->ss_dur);
1046         dst->ss_state           = le32_to_cpu(src->ss_state);
1047         dst->ss_head            = le32_to_cpu(src->ss_head);
1048         dst->ss_limit.u.f       = fio_uint64_to_double(le64_to_cpu(src->ss_limit.u.i));
1049         dst->ss_slope.u.f       = fio_uint64_to_double(le64_to_cpu(src->ss_slope.u.i));
1050         dst->ss_deviation.u.f   = fio_uint64_to_double(le64_to_cpu(src->ss_deviation.u.i));
1051         dst->ss_criterion.u.f   = fio_uint64_to_double(le64_to_cpu(src->ss_criterion.u.i));
1052
1053         if (dst->ss_state & FIO_SS_DATA) {
1054                 for (i = 0; i < dst->ss_dur; i++ ) {
1055                         dst->ss_iops_data[i] = le64_to_cpu(src->ss_iops_data[i]);
1056                         dst->ss_bw_data[i] = le64_to_cpu(src->ss_bw_data[i]);
1057                 }
1058         }
1059
1060         dst->cachehit           = le64_to_cpu(src->cachehit);
1061         dst->cachemiss          = le64_to_cpu(src->cachemiss);
1062 }
1063
1064 static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
1065 {
1066         int i;
1067
1068         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1069                 dst->max_run[i]         = le64_to_cpu(src->max_run[i]);
1070                 dst->min_run[i]         = le64_to_cpu(src->min_run[i]);
1071                 dst->max_bw[i]          = le64_to_cpu(src->max_bw[i]);
1072                 dst->min_bw[i]          = le64_to_cpu(src->min_bw[i]);
1073                 dst->iobytes[i]         = le64_to_cpu(src->iobytes[i]);
1074                 dst->agg[i]             = le64_to_cpu(src->agg[i]);
1075         }
1076
1077         dst->kb_base    = le32_to_cpu(src->kb_base);
1078         dst->unit_base  = le32_to_cpu(src->unit_base);
1079         dst->sig_figs   = le32_to_cpu(src->sig_figs);
1080         dst->groupid    = le32_to_cpu(src->groupid);
1081         dst->unified_rw_rep     = le32_to_cpu(src->unified_rw_rep);
1082 }
1083
1084 static void json_object_add_client_info(struct json_object *obj,
1085                                         struct fio_client *client)
1086 {
1087         const char *hostname = client->hostname ? client->hostname : "";
1088
1089         json_object_add_value_string(obj, "hostname", hostname);
1090         json_object_add_value_int(obj, "port", client->port);
1091 }
1092
1093 static void handle_ts(struct fio_client *client, struct fio_net_cmd *cmd)
1094 {
1095         struct cmd_ts_pdu *p = (struct cmd_ts_pdu *) cmd->payload;
1096         struct flist_head *opt_list = NULL;
1097         struct json_object *tsobj;
1098
1099         if (client->opt_lists && p->ts.thread_number <= client->jobs)
1100                 opt_list = &client->opt_lists[p->ts.thread_number - 1];
1101
1102         tsobj = show_thread_status(&p->ts, &p->rs, opt_list, &client->buf);
1103         client->did_stat = true;
1104         if (tsobj) {
1105                 json_object_add_client_info(tsobj, client);
1106                 json_array_add_value_object(clients_array, tsobj);
1107         }
1108
1109         if (sum_stat_clients <= 1)
1110                 return;
1111
1112         sum_thread_stats(&client_ts, &p->ts, sum_stat_nr == 1);
1113         sum_group_stats(&client_gs, &p->rs);
1114
1115         client_ts.members++;
1116         client_ts.thread_number = p->ts.thread_number;
1117         client_ts.groupid = p->ts.groupid;
1118         client_ts.unified_rw_rep = p->ts.unified_rw_rep;
1119         client_ts.sig_figs = p->ts.sig_figs;
1120
1121         if (++sum_stat_nr == sum_stat_clients) {
1122                 strcpy(client_ts.name, "All clients");
1123                 tsobj = show_thread_status(&client_ts, &client_gs, NULL, &allclients);
1124                 if (tsobj) {
1125                         json_object_add_client_info(tsobj, client);
1126                         json_array_add_value_object(clients_array, tsobj);
1127                 }
1128         }
1129 }
1130
1131 static void handle_gs(struct fio_client *client, struct fio_net_cmd *cmd)
1132 {
1133         struct group_run_stats *gs = (struct group_run_stats *) cmd->payload;
1134
1135         if (output_format & FIO_OUTPUT_NORMAL)
1136                 show_group_stats(gs, &client->buf);
1137 }
1138
1139 static void handle_job_opt(struct fio_client *client, struct fio_net_cmd *cmd)
1140 {
1141         struct cmd_job_option *pdu = (struct cmd_job_option *) cmd->payload;
1142         struct print_option *p;
1143
1144         if (!job_opt_object)
1145                 return;
1146
1147         pdu->global = le16_to_cpu(pdu->global);
1148         pdu->truncated = le16_to_cpu(pdu->truncated);
1149         pdu->groupid = le32_to_cpu(pdu->groupid);
1150
1151         p = malloc(sizeof(*p));
1152         p->name = strdup((char *) pdu->name);
1153         if (pdu->value[0] != '\0')
1154                 p->value = strdup((char *) pdu->value);
1155         else
1156                 p->value = NULL;
1157
1158         if (pdu->global) {
1159                 const char *pos = "";
1160
1161                 if (p->value)
1162                         pos = p->value;
1163
1164                 json_object_add_value_string(job_opt_object, p->name, pos);
1165         } else if (client->opt_lists) {
1166                 struct flist_head *opt_list = &client->opt_lists[pdu->groupid];
1167
1168                 flist_add_tail(&p->list, opt_list);
1169         } else {
1170                 free(p->value);
1171                 free(p->name);
1172                 free(p);
1173         }
1174 }
1175
1176 static void handle_text(struct fio_client *client, struct fio_net_cmd *cmd)
1177 {
1178         struct cmd_text_pdu *pdu = (struct cmd_text_pdu *) cmd->payload;
1179         const char *buf = (const char *) pdu->buf;
1180         const char *name;
1181         int fio_unused ret;
1182         struct buf_output out;
1183
1184         buf_output_init(&out);
1185
1186         name = client->name ? client->name : client->hostname;
1187
1188         if (!client->skip_newline && !(output_format & FIO_OUTPUT_TERSE))
1189                 __log_buf(&out, "<%s> ", name);
1190         __log_buf(&out, "%s", buf);
1191         log_info_buf(out.buf, out.buflen);
1192         buf_output_free(&out);
1193         client->skip_newline = strchr(buf, '\n') == NULL;
1194 }
1195
1196 static void convert_agg(struct disk_util_agg *agg)
1197 {
1198         int i;
1199
1200         for (i = 0; i < 2; i++) {
1201                 agg->ios[i]     = le64_to_cpu(agg->ios[i]);
1202                 agg->merges[i]  = le64_to_cpu(agg->merges[i]);
1203                 agg->sectors[i] = le64_to_cpu(agg->sectors[i]);
1204                 agg->ticks[i]   = le64_to_cpu(agg->ticks[i]);
1205         }
1206
1207         agg->io_ticks           = le64_to_cpu(agg->io_ticks);
1208         agg->time_in_queue      = le64_to_cpu(agg->time_in_queue);
1209         agg->slavecount         = le32_to_cpu(agg->slavecount);
1210         agg->max_util.u.f       = fio_uint64_to_double(le64_to_cpu(agg->max_util.u.i));
1211 }
1212
1213 static void convert_dus(struct disk_util_stat *dus)
1214 {
1215         int i;
1216
1217         for (i = 0; i < 2; i++) {
1218                 dus->s.ios[i]           = le64_to_cpu(dus->s.ios[i]);
1219                 dus->s.merges[i]        = le64_to_cpu(dus->s.merges[i]);
1220                 dus->s.sectors[i]       = le64_to_cpu(dus->s.sectors[i]);
1221                 dus->s.ticks[i]         = le64_to_cpu(dus->s.ticks[i]);
1222         }
1223
1224         dus->s.io_ticks         = le64_to_cpu(dus->s.io_ticks);
1225         dus->s.time_in_queue    = le64_to_cpu(dus->s.time_in_queue);
1226         dus->s.msec             = le64_to_cpu(dus->s.msec);
1227 }
1228
1229 static void handle_du(struct fio_client *client, struct fio_net_cmd *cmd)
1230 {
1231         struct cmd_du_pdu *du = (struct cmd_du_pdu *) cmd->payload;
1232
1233         if (!client->disk_stats_shown)
1234                 client->disk_stats_shown = true;
1235
1236         if (output_format & FIO_OUTPUT_JSON) {
1237                 struct json_object *duobj;
1238
1239                 json_array_add_disk_util(&du->dus, &du->agg, du_array);
1240                 duobj = json_array_last_value_object(du_array);
1241                 json_object_add_client_info(duobj, client);
1242         }
1243         if (output_format & FIO_OUTPUT_NORMAL) {
1244                 __log_buf(&client->buf, "\nDisk stats (read/write):\n");
1245                 print_disk_util(&du->dus, &du->agg, 0, &client->buf);
1246         }
1247         if (output_format & FIO_OUTPUT_TERSE && terse_version >= 3) {
1248                 print_disk_util(&du->dus, &du->agg, 1, &client->buf);
1249                 __log_buf(&client->buf, "\n");
1250         }
1251 }
1252
1253 static void convert_jobs_eta(struct jobs_eta *je)
1254 {
1255         int i;
1256
1257         je->nr_running          = le32_to_cpu(je->nr_running);
1258         je->nr_ramp             = le32_to_cpu(je->nr_ramp);
1259         je->nr_pending          = le32_to_cpu(je->nr_pending);
1260         je->nr_setting_up       = le32_to_cpu(je->nr_setting_up);
1261         je->files_open          = le32_to_cpu(je->files_open);
1262
1263         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1264                 je->m_rate[i]   = le64_to_cpu(je->m_rate[i]);
1265                 je->t_rate[i]   = le64_to_cpu(je->t_rate[i]);
1266                 je->m_iops[i]   = le32_to_cpu(je->m_iops[i]);
1267                 je->t_iops[i]   = le32_to_cpu(je->t_iops[i]);
1268                 je->rate[i]     = le64_to_cpu(je->rate[i]);
1269                 je->iops[i]     = le32_to_cpu(je->iops[i]);
1270         }
1271
1272         je->elapsed_sec         = le64_to_cpu(je->elapsed_sec);
1273         je->eta_sec             = le64_to_cpu(je->eta_sec);
1274         je->nr_threads          = le32_to_cpu(je->nr_threads);
1275         je->is_pow2             = le32_to_cpu(je->is_pow2);
1276         je->unit_base           = le32_to_cpu(je->unit_base);
1277         je->sig_figs            = le32_to_cpu(je->sig_figs);
1278 }
1279
1280 void fio_client_sum_jobs_eta(struct jobs_eta *dst, struct jobs_eta *je)
1281 {
1282         int i;
1283
1284         dst->nr_running         += je->nr_running;
1285         dst->nr_ramp            += je->nr_ramp;
1286         dst->nr_pending         += je->nr_pending;
1287         dst->nr_setting_up      += je->nr_setting_up;
1288         dst->files_open         += je->files_open;
1289
1290         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1291                 dst->m_rate[i]  += je->m_rate[i];
1292                 dst->t_rate[i]  += je->t_rate[i];
1293                 dst->m_iops[i]  += je->m_iops[i];
1294                 dst->t_iops[i]  += je->t_iops[i];
1295                 dst->rate[i]    += je->rate[i];
1296                 dst->iops[i]    += je->iops[i];
1297         }
1298
1299         dst->elapsed_sec        += je->elapsed_sec;
1300
1301         if (je->eta_sec > dst->eta_sec)
1302                 dst->eta_sec = je->eta_sec;
1303
1304         dst->nr_threads         += je->nr_threads;
1305
1306         /*
1307          * This wont be correct for multiple strings, but at least it
1308          * works for the basic cases.
1309          */
1310         strcpy((char *) dst->run_str, (char *) je->run_str);
1311 }
1312
1313 static bool remove_reply_cmd(struct fio_client *client, struct fio_net_cmd *cmd)
1314 {
1315         struct fio_net_cmd_reply *reply = NULL;
1316         struct flist_head *entry;
1317
1318         flist_for_each(entry, &client->cmd_list) {
1319                 reply = flist_entry(entry, struct fio_net_cmd_reply, list);
1320
1321                 if (cmd->tag == (uintptr_t) reply)
1322                         break;
1323
1324                 reply = NULL;
1325         }
1326
1327         if (!reply) {
1328                 log_err("fio: client: unable to find matching tag (%llx)\n", (unsigned long long) cmd->tag);
1329                 return false;
1330         }
1331
1332         flist_del(&reply->list);
1333         cmd->tag = reply->saved_tag;
1334         free(reply);
1335         return true;
1336 }
1337
1338 int fio_client_wait_for_reply(struct fio_client *client, uint64_t tag)
1339 {
1340         do {
1341                 struct fio_net_cmd_reply *reply = NULL;
1342                 struct flist_head *entry;
1343
1344                 flist_for_each(entry, &client->cmd_list) {
1345                         reply = flist_entry(entry, struct fio_net_cmd_reply, list);
1346
1347                         if (tag == (uintptr_t) reply)
1348                                 break;
1349
1350                         reply = NULL;
1351                 }
1352
1353                 if (!reply)
1354                         break;
1355
1356                 usleep(1000);
1357         } while (1);
1358
1359         return 0;
1360 }
1361
1362 static void handle_eta(struct fio_client *client, struct fio_net_cmd *cmd)
1363 {
1364         struct jobs_eta *je = (struct jobs_eta *) cmd->payload;
1365         struct client_eta *eta = (struct client_eta *) (uintptr_t) cmd->tag;
1366
1367         dprint(FD_NET, "client: got eta tag %p, %d\n", eta, eta->pending);
1368
1369         assert(client->eta_in_flight == eta);
1370
1371         client->eta_in_flight = NULL;
1372         flist_del_init(&client->eta_list);
1373         client->eta_timeouts = 0;
1374
1375         if (client->ops->jobs_eta)
1376                 client->ops->jobs_eta(client, je);
1377
1378         fio_client_sum_jobs_eta(&eta->eta, je);
1379         fio_client_dec_jobs_eta(eta, client->ops->eta);
1380 }
1381
1382 static void client_flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
1383                                       uint64_t sample_size)
1384 {
1385         struct io_sample *s;
1386         int log_offset;
1387         uint64_t i, j, nr_samples;
1388         struct io_u_plat_entry *entry;
1389         uint64_t *io_u_plat;
1390
1391         int stride = 1 << hist_coarseness;
1392
1393         if (!sample_size)
1394                 return;
1395
1396         s = __get_sample(samples, 0, 0);
1397         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
1398
1399         nr_samples = sample_size / __log_entry_sz(log_offset);
1400
1401         for (i = 0; i < nr_samples; i++) {
1402
1403                 s = (struct io_sample *)((char *)__get_sample(samples, log_offset, i) +
1404                         i * sizeof(struct io_u_plat_entry));
1405
1406                 entry = s->data.plat_entry;
1407                 io_u_plat = entry->io_u_plat;
1408
1409                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
1410                                                 io_sample_ddir(s), (unsigned long long) s->bs);
1411                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
1412                         fprintf(f, "%llu, ", (unsigned long long)hist_sum(j, stride, io_u_plat, NULL));
1413                 }
1414                 fprintf(f, "%llu\n", (unsigned long long)
1415                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat, NULL));
1416
1417         }
1418 }
1419
1420 static int fio_client_handle_iolog(struct fio_client *client,
1421                                    struct fio_net_cmd *cmd)
1422 {
1423         struct cmd_iolog_pdu *pdu = NULL;
1424         bool store_direct;
1425         char *log_pathname = NULL;
1426         int ret = 0;
1427
1428         pdu = convert_iolog(cmd, &store_direct);
1429         if (!pdu) {
1430                 log_err("fio: failed converting IO log\n");
1431                 ret = 1;
1432                 goto out;
1433         }
1434
1435         /* allocate buffer big enough for next sprintf() call */
1436         log_pathname = malloc(10 + strlen((char *)pdu->name) +
1437                         strlen(client->hostname));
1438         if (!log_pathname) {
1439                 log_err("fio: memory allocation of unique pathname failed\n");
1440                 ret = -1;
1441                 goto out;
1442         }
1443         /* generate a unique pathname for the log file using hostname */
1444         sprintf(log_pathname, "%s.%s", pdu->name, client->hostname);
1445
1446         if (store_direct) {
1447                 ssize_t wrote;
1448                 size_t sz;
1449                 int fd;
1450
1451                 fd = open((const char *) log_pathname,
1452                                 O_WRONLY | O_CREAT | O_TRUNC, 0644);
1453                 if (fd < 0) {
1454                         log_err("fio: open log %s: %s\n",
1455                                 log_pathname, strerror(errno));
1456                         ret = 1;
1457                         goto out;
1458                 }
1459
1460                 sz = cmd->pdu_len - sizeof(*pdu);
1461                 wrote = write(fd, pdu->samples, sz);
1462                 close(fd);
1463
1464                 if (wrote != sz) {
1465                         log_err("fio: short write on compressed log\n");
1466                         ret = 1;
1467                         goto out;
1468                 }
1469
1470                 ret = 0;
1471         } else {
1472                 FILE *f;
1473                 f = fopen((const char *) log_pathname, "w");
1474                 if (!f) {
1475                         log_err("fio: fopen log %s : %s\n",
1476                                 log_pathname, strerror(errno));
1477                         ret = 1;
1478                         goto out;
1479                 }
1480
1481                 if (pdu->log_type == IO_LOG_TYPE_HIST) {
1482                         client_flush_hist_samples(f, pdu->log_hist_coarseness, pdu->samples,
1483                                            pdu->nr_samples * sizeof(struct io_sample));
1484                 } else {
1485                         flush_samples(f, pdu->samples,
1486                                         pdu->nr_samples * sizeof(struct io_sample));
1487                 }
1488                 fclose(f);
1489                 ret = 0;
1490         }
1491
1492 out:
1493         if (pdu && pdu != (void *) cmd->payload)
1494                 free(pdu);
1495
1496         if (log_pathname)
1497                 free(log_pathname);
1498
1499         return ret;
1500 }
1501
1502 static void handle_probe(struct fio_client *client, struct fio_net_cmd *cmd)
1503 {
1504         struct cmd_probe_reply_pdu *probe = (struct cmd_probe_reply_pdu *) cmd->payload;
1505         const char *os, *arch;
1506         char bit[16];
1507
1508         os = fio_get_os_string(probe->os);
1509         if (!os)
1510                 os = "unknown";
1511
1512         arch = fio_get_arch_string(probe->arch);
1513         if (!arch)
1514                 os = "unknown";
1515
1516         sprintf(bit, "%d-bit", probe->bpp * 8);
1517         probe->flags = le64_to_cpu(probe->flags);
1518
1519         if (output_format & FIO_OUTPUT_NORMAL) {
1520                 log_info("hostname=%s, be=%u, %s, os=%s, arch=%s, fio=%s, flags=%lx\n",
1521                         probe->hostname, probe->bigendian, bit, os, arch,
1522                         probe->fio_version, (unsigned long) probe->flags);
1523         }
1524
1525         if (!client->name)
1526                 client->name = strdup((char *) probe->hostname);
1527 }
1528
1529 static void handle_start(struct fio_client *client, struct fio_net_cmd *cmd)
1530 {
1531         struct cmd_start_pdu *pdu = (struct cmd_start_pdu *) cmd->payload;
1532
1533         client->state = Client_started;
1534         client->jobs = le32_to_cpu(pdu->jobs);
1535         client->nr_stat = le32_to_cpu(pdu->stat_outputs);
1536
1537         if (client->jobs) {
1538                 int i;
1539
1540                 if (client->opt_lists)
1541                         free(client->opt_lists);
1542
1543                 client->opt_lists = malloc(client->jobs * sizeof(struct flist_head));
1544                 for (i = 0; i < client->jobs; i++)
1545                         INIT_FLIST_HEAD(&client->opt_lists[i]);
1546         }
1547
1548         sum_stat_clients += client->nr_stat;
1549 }
1550
1551 static void handle_stop(struct fio_client *client)
1552 {
1553         if (client->error)
1554                 log_info("client <%s>: exited with error %d\n", client->hostname, client->error);
1555 }
1556
1557 static void convert_stop(struct fio_net_cmd *cmd)
1558 {
1559         struct cmd_end_pdu *pdu = (struct cmd_end_pdu *) cmd->payload;
1560
1561         pdu->error = le32_to_cpu(pdu->error);
1562 }
1563
1564 static void convert_text(struct fio_net_cmd *cmd)
1565 {
1566         struct cmd_text_pdu *pdu = (struct cmd_text_pdu *) cmd->payload;
1567
1568         pdu->level      = le32_to_cpu(pdu->level);
1569         pdu->buf_len    = le32_to_cpu(pdu->buf_len);
1570         pdu->log_sec    = le64_to_cpu(pdu->log_sec);
1571         pdu->log_usec   = le64_to_cpu(pdu->log_usec);
1572 }
1573
1574 static struct cmd_iolog_pdu *convert_iolog_gz(struct fio_net_cmd *cmd,
1575                                               struct cmd_iolog_pdu *pdu)
1576 {
1577 #ifdef CONFIG_ZLIB
1578         struct cmd_iolog_pdu *ret;
1579         z_stream stream;
1580         uint64_t nr_samples;
1581         size_t total;
1582         char *p;
1583
1584         stream.zalloc = Z_NULL;
1585         stream.zfree = Z_NULL;
1586         stream.opaque = Z_NULL;
1587         stream.avail_in = 0;
1588         stream.next_in = Z_NULL;
1589
1590         if (inflateInit(&stream) != Z_OK)
1591                 return NULL;
1592
1593         /*
1594          * Get header first, it's not compressed
1595          */
1596         nr_samples = le64_to_cpu(pdu->nr_samples);
1597
1598         if (pdu->log_type == IO_LOG_TYPE_HIST)
1599                 total = nr_samples * (__log_entry_sz(le32_to_cpu(pdu->log_offset)) +
1600                                         sizeof(struct io_u_plat_entry));
1601         else
1602                 total = nr_samples * __log_entry_sz(le32_to_cpu(pdu->log_offset));
1603         ret = malloc(total + sizeof(*pdu));
1604         ret->nr_samples = nr_samples;
1605
1606         memcpy(ret, pdu, sizeof(*pdu));
1607
1608         p = (char *) ret + sizeof(*pdu);
1609
1610         stream.avail_in = cmd->pdu_len - sizeof(*pdu);
1611         stream.next_in = (void *)((char *) pdu + sizeof(*pdu));
1612         while (stream.avail_in) {
1613                 unsigned int this_chunk = 65536;
1614                 unsigned int this_len;
1615                 int err;
1616
1617                 if (this_chunk > total)
1618                         this_chunk = total;
1619
1620                 stream.avail_out = this_chunk;
1621                 stream.next_out = (void *)p;
1622                 err = inflate(&stream, Z_NO_FLUSH);
1623                 /* may be Z_OK, or Z_STREAM_END */
1624                 if (err < 0) {
1625                         /*
1626                          * Z_STREAM_ERROR and Z_BUF_ERROR can safely be
1627                          * ignored */
1628                         if (err == Z_STREAM_ERROR || err == Z_BUF_ERROR)
1629                                 break;
1630                         log_err("fio: inflate error %d\n", err);
1631                         free(ret);
1632                         ret = NULL;
1633                         goto err;
1634                 }
1635
1636                 this_len = this_chunk - stream.avail_out;
1637                 p += this_len;
1638                 total -= this_len;
1639         }
1640
1641 err:
1642         inflateEnd(&stream);
1643         return ret;
1644 #else
1645         return NULL;
1646 #endif
1647 }
1648
1649 /*
1650  * This has been compressed on the server side, since it can be big.
1651  * Uncompress here.
1652  */
1653 static struct cmd_iolog_pdu *convert_iolog(struct fio_net_cmd *cmd,
1654                                            bool *store_direct)
1655 {
1656         struct cmd_iolog_pdu *pdu = (struct cmd_iolog_pdu *) cmd->payload;
1657         struct cmd_iolog_pdu *ret;
1658         uint64_t i;
1659         int compressed;
1660         void *samples;
1661
1662         *store_direct = false;
1663
1664         /*
1665          * Convert if compressed and we support it. If it's not
1666          * compressed, we need not do anything.
1667          */
1668         compressed = le32_to_cpu(pdu->compressed);
1669         if (compressed == XMIT_COMPRESSED) {
1670 #ifndef CONFIG_ZLIB
1671                 log_err("fio: server sent compressed data by mistake\n");
1672                 return NULL;
1673 #endif
1674                 ret = convert_iolog_gz(cmd, pdu);
1675                 if (!ret) {
1676                         log_err("fio: failed decompressing log\n");
1677                         return NULL;
1678                 }
1679         } else if (compressed == STORE_COMPRESSED) {
1680                 *store_direct = true;
1681                 ret = pdu;
1682         } else
1683                 ret = pdu;
1684
1685         ret->nr_samples         = le64_to_cpu(ret->nr_samples);
1686         ret->thread_number      = le32_to_cpu(ret->thread_number);
1687         ret->log_type           = le32_to_cpu(ret->log_type);
1688         ret->compressed         = le32_to_cpu(ret->compressed);
1689         ret->log_offset         = le32_to_cpu(ret->log_offset);
1690         ret->log_hist_coarseness = le32_to_cpu(ret->log_hist_coarseness);
1691
1692         if (*store_direct)
1693                 return ret;
1694
1695         samples = &ret->samples[0];
1696         for (i = 0; i < ret->nr_samples; i++) {
1697                 struct io_sample *s;
1698
1699                 s = __get_sample(samples, ret->log_offset, i);
1700                 if (ret->log_type == IO_LOG_TYPE_HIST)
1701                         s = (struct io_sample *)((char *)s + sizeof(struct io_u_plat_entry) * i);
1702
1703                 s->time         = le64_to_cpu(s->time);
1704                 s->data.val     = le64_to_cpu(s->data.val);
1705                 s->__ddir       = __le32_to_cpu(s->__ddir);
1706                 s->bs           = le64_to_cpu(s->bs);
1707
1708                 if (ret->log_offset) {
1709                         struct io_sample_offset *so = (void *) s;
1710
1711                         so->offset = le64_to_cpu(so->offset);
1712                 }
1713
1714                 if (ret->log_type == IO_LOG_TYPE_HIST) {
1715                         s->data.plat_entry = (struct io_u_plat_entry *)(((char *)s) + sizeof(*s));
1716                         s->data.plat_entry->list.next = NULL;
1717                         s->data.plat_entry->list.prev = NULL;
1718                 }
1719         }
1720
1721         return ret;
1722 }
1723
1724 static void sendfile_reply(int fd, struct cmd_sendfile_reply *rep,
1725                            size_t size, uint64_t tag)
1726 {
1727         rep->error = cpu_to_le32(rep->error);
1728         fio_net_send_cmd(fd, FIO_NET_CMD_SENDFILE, rep, size, &tag, NULL);
1729 }
1730
1731 static int fio_send_file(struct fio_client *client, struct cmd_sendfile *pdu,
1732                          uint64_t tag)
1733 {
1734         struct cmd_sendfile_reply *rep;
1735         struct stat sb;
1736         size_t size;
1737         int fd;
1738
1739         size = sizeof(*rep);
1740         rep = malloc(size);
1741
1742         if (stat((char *)pdu->path, &sb) < 0) {
1743 fail:
1744                 rep->error = errno;
1745                 sendfile_reply(client->fd, rep, size, tag);
1746                 free(rep);
1747                 return 1;
1748         }
1749
1750         size += sb.st_size;
1751         rep = realloc(rep, size);
1752         rep->size = cpu_to_le32((uint32_t) sb.st_size);
1753
1754         fd = open((char *)pdu->path, O_RDONLY);
1755         if (fd == -1 )
1756                 goto fail;
1757
1758         rep->error = read_data(fd, &rep->data, sb.st_size);
1759         sendfile_reply(client->fd, rep, size, tag);
1760         free(rep);
1761         close(fd);
1762         return 0;
1763 }
1764
1765 int fio_handle_client(struct fio_client *client)
1766 {
1767         struct client_ops *ops = client->ops;
1768         struct fio_net_cmd *cmd;
1769         int size;
1770
1771         dprint(FD_NET, "client: handle %s\n", client->hostname);
1772
1773         cmd = fio_net_recv_cmd(client->fd, true);
1774         if (!cmd)
1775                 return 0;
1776
1777         dprint(FD_NET, "client: got cmd op %s from %s (pdu=%u)\n",
1778                 fio_server_op(cmd->opcode), client->hostname, cmd->pdu_len);
1779
1780         client->last_cmd = cmd->opcode;
1781
1782         switch (cmd->opcode) {
1783         case FIO_NET_CMD_QUIT:
1784                 if (ops->quit)
1785                         ops->quit(client, cmd);
1786                 remove_client(client);
1787                 break;
1788         case FIO_NET_CMD_TEXT:
1789                 convert_text(cmd);
1790                 ops->text(client, cmd);
1791                 break;
1792         case FIO_NET_CMD_DU: {
1793                 struct cmd_du_pdu *du = (struct cmd_du_pdu *) cmd->payload;
1794
1795                 convert_dus(&du->dus);
1796                 convert_agg(&du->agg);
1797
1798                 ops->disk_util(client, cmd);
1799                 break;
1800                 }
1801         case FIO_NET_CMD_TS: {
1802                 struct cmd_ts_pdu *p = (struct cmd_ts_pdu *) cmd->payload;
1803
1804                 dprint(FD_NET, "client: ts->ss_state = %u\n", (unsigned int) le32_to_cpu(p->ts.ss_state));
1805                 if (le32_to_cpu(p->ts.ss_state) & FIO_SS_DATA) {
1806                         dprint(FD_NET, "client: received steadystate ring buffers\n");
1807
1808                         size = le64_to_cpu(p->ts.ss_dur);
1809                         p->ts.ss_iops_data = (uint64_t *) ((struct cmd_ts_pdu *)cmd->payload + 1);
1810                         p->ts.ss_bw_data = p->ts.ss_iops_data + size;
1811                 }
1812
1813                 convert_ts(&p->ts, &p->ts);
1814                 convert_gs(&p->rs, &p->rs);
1815
1816                 ops->thread_status(client, cmd);
1817                 break;
1818                 }
1819         case FIO_NET_CMD_GS: {
1820                 struct group_run_stats *gs = (struct group_run_stats *) cmd->payload;
1821
1822                 convert_gs(gs, gs);
1823
1824                 ops->group_stats(client, cmd);
1825                 break;
1826                 }
1827         case FIO_NET_CMD_ETA: {
1828                 struct jobs_eta *je = (struct jobs_eta *) cmd->payload;
1829
1830                 if (!remove_reply_cmd(client, cmd))
1831                         break;
1832                 convert_jobs_eta(je);
1833                 handle_eta(client, cmd);
1834                 break;
1835                 }
1836         case FIO_NET_CMD_PROBE:
1837                 remove_reply_cmd(client, cmd);
1838                 ops->probe(client, cmd);
1839                 break;
1840         case FIO_NET_CMD_SERVER_START:
1841                 client->state = Client_running;
1842                 if (ops->job_start)
1843                         ops->job_start(client, cmd);
1844                 break;
1845         case FIO_NET_CMD_START: {
1846                 struct cmd_start_pdu *pdu = (struct cmd_start_pdu *) cmd->payload;
1847
1848                 pdu->jobs = le32_to_cpu(pdu->jobs);
1849                 ops->start(client, cmd);
1850                 break;
1851                 }
1852         case FIO_NET_CMD_STOP: {
1853                 struct cmd_end_pdu *pdu = (struct cmd_end_pdu *) cmd->payload;
1854
1855                 convert_stop(cmd);
1856                 client->state = Client_stopped;
1857                 client->error = le32_to_cpu(pdu->error);
1858                 client->signal = le32_to_cpu(pdu->signal);
1859                 ops->stop(client);
1860                 break;
1861                 }
1862         case FIO_NET_CMD_ADD_JOB: {
1863                 struct cmd_add_job_pdu *pdu = (struct cmd_add_job_pdu *) cmd->payload;
1864
1865                 client->thread_number = le32_to_cpu(pdu->thread_number);
1866                 client->groupid = le32_to_cpu(pdu->groupid);
1867
1868                 if (ops->add_job)
1869                         ops->add_job(client, cmd);
1870                 break;
1871                 }
1872         case FIO_NET_CMD_IOLOG:
1873                 fio_client_handle_iolog(client, cmd);
1874                 break;
1875         case FIO_NET_CMD_UPDATE_JOB:
1876                 ops->update_job(client, cmd);
1877                 remove_reply_cmd(client, cmd);
1878                 break;
1879         case FIO_NET_CMD_VTRIGGER: {
1880                 struct all_io_list *pdu = (struct all_io_list *) cmd->payload;
1881                 char buf[128];
1882                 int off = 0;
1883
1884                 if (aux_path) {
1885                         strcpy(buf, aux_path);
1886                         off = strlen(buf);
1887                 }
1888
1889                 __verify_save_state(pdu, server_name(client, &buf[off], sizeof(buf) - off));
1890                 exec_trigger(trigger_cmd);
1891                 break;
1892                 }
1893         case FIO_NET_CMD_SENDFILE: {
1894                 struct cmd_sendfile *pdu = (struct cmd_sendfile *) cmd->payload;
1895                 fio_send_file(client, pdu, cmd->tag);
1896                 break;
1897                 }
1898         case FIO_NET_CMD_JOB_OPT: {
1899                 handle_job_opt(client, cmd);
1900                 break;
1901         }
1902         default:
1903                 log_err("fio: unknown client op: %s\n", fio_server_op(cmd->opcode));
1904                 break;
1905         }
1906
1907         free(cmd);
1908         return 1;
1909 }
1910
1911 int fio_clients_send_trigger(const char *cmd)
1912 {
1913         struct flist_head *entry;
1914         struct fio_client *client;
1915         size_t slen;
1916
1917         dprint(FD_NET, "client: send vtrigger: %s\n", cmd);
1918
1919         if (!cmd)
1920                 slen = 0;
1921         else
1922                 slen = strlen(cmd);
1923
1924         flist_for_each(entry, &client_list) {
1925                 struct cmd_vtrigger_pdu *pdu;
1926
1927                 client = flist_entry(entry, struct fio_client, list);
1928
1929                 pdu = malloc(sizeof(*pdu) + slen);
1930                 pdu->len = cpu_to_le16((uint16_t) slen);
1931                 if (slen)
1932                         memcpy(pdu->cmd, cmd, slen);
1933                 fio_net_send_cmd(client->fd, FIO_NET_CMD_VTRIGGER, pdu,
1934                                         sizeof(*pdu) + slen, NULL, NULL);
1935                 free(pdu);
1936         }
1937
1938         return 0;
1939 }
1940
1941 static void request_client_etas(struct client_ops *ops)
1942 {
1943         struct fio_client *client;
1944         struct flist_head *entry;
1945         struct client_eta *eta;
1946         int skipped = 0;
1947
1948         if (eta_print == FIO_ETA_NEVER)
1949                 return;
1950
1951         dprint(FD_NET, "client: request eta (%d)\n", nr_clients);
1952
1953         eta = calloc(1, sizeof(*eta) + __THREAD_RUNSTR_SZ(REAL_MAX_JOBS));
1954         eta->pending = nr_clients;
1955
1956         flist_for_each(entry, &client_list) {
1957                 client = flist_entry(entry, struct fio_client, list);
1958
1959                 if (!flist_empty(&client->eta_list)) {
1960                         skipped++;
1961                         continue;
1962                 }
1963                 if (client->state != Client_running)
1964                         continue;
1965
1966                 assert(!client->eta_in_flight);
1967                 flist_add_tail(&client->eta_list, &eta_list);
1968                 client->eta_in_flight = eta;
1969                 fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_SEND_ETA,
1970                                         (uintptr_t) eta, &client->cmd_list);
1971         }
1972
1973         while (skipped--) {
1974                 if (!fio_client_dec_jobs_eta(eta, ops->eta))
1975                         break;
1976         }
1977
1978         dprint(FD_NET, "client: requested eta tag %p\n", eta);
1979 }
1980
1981 /*
1982  * A single SEND_ETA timeout isn't fatal. Attempt to recover.
1983  */
1984 static int handle_cmd_timeout(struct fio_client *client,
1985                               struct fio_net_cmd_reply *reply)
1986 {
1987         uint16_t reply_opcode = reply->opcode;
1988
1989         flist_del(&reply->list);
1990         free(reply);
1991
1992         if (reply_opcode != FIO_NET_CMD_SEND_ETA)
1993                 return 1;
1994
1995         log_info("client <%s>: timeout on SEND_ETA\n", client->hostname);
1996
1997         flist_del_init(&client->eta_list);
1998         if (client->eta_in_flight) {
1999                 fio_client_dec_jobs_eta(client->eta_in_flight, client->ops->eta);
2000                 client->eta_in_flight = NULL;
2001         }
2002
2003         /*
2004          * If we fail 5 in a row, give up...
2005          */
2006         if (client->eta_timeouts++ > 5)
2007                 return 1;
2008
2009         return 0;
2010 }
2011
2012 static int client_check_cmd_timeout(struct fio_client *client,
2013                                     struct timespec *now)
2014 {
2015         struct fio_net_cmd_reply *reply;
2016         struct flist_head *entry, *tmp;
2017         int ret = 0;
2018
2019         flist_for_each_safe(entry, tmp, &client->cmd_list) {
2020                 unsigned int op;
2021
2022                 reply = flist_entry(entry, struct fio_net_cmd_reply, list);
2023
2024                 if (mtime_since(&reply->ts, now) < FIO_NET_CLIENT_TIMEOUT)
2025                         continue;
2026
2027                 op = reply->opcode;
2028                 if (!handle_cmd_timeout(client, reply))
2029                         continue;
2030
2031                 log_err("fio: client %s, timeout on cmd %s\n", client->hostname,
2032                                                 fio_server_op(op));
2033                 ret = 1;
2034         }
2035
2036         return flist_empty(&client->cmd_list) && ret;
2037 }
2038
2039 static int fio_check_clients_timed_out(void)
2040 {
2041         struct fio_client *client;
2042         struct flist_head *entry, *tmp;
2043         struct timespec ts;
2044         int ret = 0;
2045
2046         fio_gettime(&ts, NULL);
2047
2048         flist_for_each_safe(entry, tmp, &client_list) {
2049                 client = flist_entry(entry, struct fio_client, list);
2050
2051                 if (flist_empty(&client->cmd_list))
2052                         continue;
2053
2054                 if (!client_check_cmd_timeout(client, &ts))
2055                         continue;
2056
2057                 if (client->ops->timed_out)
2058                         client->ops->timed_out(client);
2059                 else
2060                         log_err("fio: client %s timed out\n", client->hostname);
2061
2062                 if (client->last_cmd != FIO_NET_CMD_VTRIGGER)
2063                         client->error = ETIMEDOUT;
2064                 else
2065                         log_info("fio: ignoring timeout due to vtrigger\n");
2066                 remove_client(client);
2067                 ret = 1;
2068         }
2069
2070         return ret;
2071 }
2072
2073 int fio_handle_clients(struct client_ops *ops)
2074 {
2075         struct pollfd *pfds;
2076         int i, ret = 0, retval = 0;
2077
2078         fio_gettime(&eta_ts, NULL);
2079
2080         pfds = malloc(nr_clients * sizeof(struct pollfd));
2081
2082         init_thread_stat(&client_ts);
2083         init_group_run_stat(&client_gs);
2084
2085         while (!exit_backend && nr_clients) {
2086                 struct flist_head *entry, *tmp;
2087                 struct fio_client *client;
2088
2089                 i = 0;
2090                 flist_for_each_safe(entry, tmp, &client_list) {
2091                         client = flist_entry(entry, struct fio_client, list);
2092
2093                         if (!client->sent_job && !client->ops->stay_connected &&
2094                             flist_empty(&client->cmd_list)) {
2095                                 remove_client(client);
2096                                 continue;
2097                         }
2098
2099                         pfds[i].fd = client->fd;
2100                         pfds[i].events = POLLIN;
2101                         i++;
2102                 }
2103
2104                 if (!nr_clients)
2105                         break;
2106
2107                 assert(i == nr_clients);
2108
2109                 do {
2110                         struct timespec ts;
2111                         int timeout;
2112
2113                         fio_gettime(&ts, NULL);
2114                         if (eta_time_within_slack(mtime_since(&eta_ts, &ts))) {
2115                                 request_client_etas(ops);
2116                                 memcpy(&eta_ts, &ts, sizeof(ts));
2117
2118                                 if (fio_check_clients_timed_out())
2119                                         break;
2120                         }
2121
2122                         check_trigger_file();
2123
2124                         timeout = min(100u, ops->eta_msec);
2125
2126                         ret = poll(pfds, nr_clients, timeout);
2127                         if (ret < 0) {
2128                                 if (errno == EINTR)
2129                                         continue;
2130                                 log_err("fio: poll clients: %s\n", strerror(errno));
2131                                 break;
2132                         } else if (!ret)
2133                                 continue;
2134                 } while (ret <= 0);
2135
2136                 for (i = 0; i < nr_clients; i++) {
2137                         if (!(pfds[i].revents & POLLIN))
2138                                 continue;
2139
2140                         client = find_client_by_fd(pfds[i].fd);
2141                         if (!client) {
2142                                 log_err("fio: unknown client fd %ld\n", (long) pfds[i].fd);
2143                                 continue;
2144                         }
2145                         if (!fio_handle_client(client)) {
2146                                 log_info("client: host=%s disconnected\n",
2147                                                 client->hostname);
2148                                 remove_client(client);
2149                                 retval = 1;
2150                         } else if (client->error)
2151                                 retval = 1;
2152                         fio_put_client(client);
2153                 }
2154         }
2155
2156         log_info_buf(allclients.buf, allclients.buflen);
2157         buf_output_free(&allclients);
2158
2159         fio_client_json_fini();
2160
2161         free(pfds);
2162         return retval || error_clients;
2163 }
2164
2165 static void client_display_thread_status(struct jobs_eta *je)
2166 {
2167         if (!(output_format & FIO_OUTPUT_JSON))
2168                 display_thread_status(je);
2169 }