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