server.c

   1 #include <stdio.h>
   2 #include <stdlib.h>
   3 #include <string.h>
   4 #include <unistd.h>
   5 #include <errno.h>
   6 #include <poll.h>
   7 #include <sys/types.h>
   8 #include <sys/wait.h>
   9 #include <sys/socket.h>
  10 #include <sys/stat.h>
  11 #include <sys/un.h>
  12 #include <sys/uio.h>
  13 #include <netinet/in.h>
  14 #include <arpa/inet.h>
  15 #include <netdb.h>
  16 #include <syslog.h>
  17 #include <signal.h>
  18 #ifdef CONFIG_ZLIB
  19 #include <zlib.h>
  20 #endif
  21
  22 #include "fio.h"
  23 #include "options.h"
  24 #include "server.h"
  25 #include "crc/crc16.h"
  26 #include "lib/ieee754.h"
  27 #include "verify-state.h"
  28 #include "smalloc.h"
  29
  30 int fio_net_port = FIO_NET_PORT;
  31
  32 bool exit_backend = false;
  33
  34 enum {
  35         SK_F_FREE       = 1,
  36         SK_F_COPY       = 2,
  37         SK_F_SIMPLE     = 4,
  38         SK_F_VEC        = 8,
  39         SK_F_INLINE     = 16,
  40 };
  41
  42 struct sk_entry {
  43         struct flist_head list; /* link on sk_out->list */
  44         int flags;              /* SK_F_* */
  45         int opcode;             /* Actual command fields */
  46         void *buf;
  47         off_t size;
  48         uint64_t tag;
  49         struct flist_head next; /* Other sk_entry's, if linked command */
  50 };
  51
  52 static char *fio_server_arg;
  53 static char *bind_sock;
  54 static struct sockaddr_in saddr_in;
  55 static struct sockaddr_in6 saddr_in6;
  56 static int use_ipv6;
  57 #ifdef CONFIG_ZLIB
  58 static unsigned int has_zlib = 1;
  59 #else
  60 static unsigned int has_zlib = 0;
  61 #endif
  62 static unsigned int use_zlib;
  63 static char me[128];
  64
  65 static pthread_key_t sk_out_key;
  66
  67 #ifdef WIN32
  68 static char *fio_server_pipe_name  = NULL;
  69 static HANDLE hjob = INVALID_HANDLE_VALUE;
  70 struct ffi_element {
  71         union {
  72                 pthread_t thread;
  73                 HANDLE hProcess;
  74         };
  75         bool is_thread;
  76 };
  77 #endif
  78
  79 struct fio_fork_item {
  80         struct flist_head list;
  81         int exitval;
  82         int signal;
  83         int exited;
  84 #ifdef WIN32
  85         struct ffi_element element;
  86 #else
  87         pid_t pid;
  88 #endif
  89 };
  90
  91 struct cmd_reply {
  92         struct fio_sem lock;
  93         void *data;
  94         size_t size;
  95         int error;
  96 };
  97
  98 static const char *fio_server_ops[FIO_NET_CMD_NR] = {
  99         "",
 100         "QUIT",
 101         "EXIT",
 102         "JOB",
 103         "JOBLINE",
 104         "TEXT",
 105         "TS",
 106         "GS",
 107         "SEND_ETA",
 108         "ETA",
 109         "PROBE",
 110         "START",
 111         "STOP",
 112         "DISK_UTIL",
 113         "SERVER_START",
 114         "ADD_JOB",
 115         "RUN",
 116         "IOLOG",
 117         "UPDATE_JOB",
 118         "LOAD_FILE",
 119         "VTRIGGER",
 120         "SENDFILE",
 121         "JOB_OPT",
 122 };
 123
 124 static void sk_lock(struct sk_out *sk_out)
 125 {
 126         fio_sem_down(&sk_out->lock);
 127 }
 128
 129 static void sk_unlock(struct sk_out *sk_out)
 130 {
 131         fio_sem_up(&sk_out->lock);
 132 }
 133
 134 void sk_out_assign(struct sk_out *sk_out)
 135 {
 136         if (!sk_out)
 137                 return;
 138
 139         sk_lock(sk_out);
 140         sk_out->refs++;
 141         sk_unlock(sk_out);
 142         pthread_setspecific(sk_out_key, sk_out);
 143 }
 144
 145 static void sk_out_free(struct sk_out *sk_out)
 146 {
 147         __fio_sem_remove(&sk_out->lock);
 148         __fio_sem_remove(&sk_out->wait);
 149         __fio_sem_remove(&sk_out->xmit);
 150         sfree(sk_out);
 151 }
 152
 153 static int __sk_out_drop(struct sk_out *sk_out)
 154 {
 155         if (sk_out) {
 156                 int refs;
 157
 158                 sk_lock(sk_out);
 159                 assert(sk_out->refs != 0);
 160                 refs = --sk_out->refs;
 161                 sk_unlock(sk_out);
 162
 163                 if (!refs) {
 164                         sk_out_free(sk_out);
 165                         pthread_setspecific(sk_out_key, NULL);
 166                         return 0;
 167                 }
 168         }
 169
 170         return 1;
 171 }
 172
 173 void sk_out_drop(void)
 174 {
 175         struct sk_out *sk_out;
 176
 177         sk_out = pthread_getspecific(sk_out_key);
 178         __sk_out_drop(sk_out);
 179 }
 180
 181 static void __fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
 182                                uint32_t pdu_len, uint64_t tag)
 183 {
 184         memset(cmd, 0, sizeof(*cmd));
 185
 186         cmd->version    = __cpu_to_le16(FIO_SERVER_VER);
 187         cmd->opcode     = cpu_to_le16(opcode);
 188         cmd->tag        = cpu_to_le64(tag);
 189         cmd->pdu_len    = cpu_to_le32(pdu_len);
 190 }
 191
 192
 193 static void fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
 194                              const void *pdu, uint32_t pdu_len, uint64_t tag)
 195 {
 196         __fio_init_net_cmd(cmd, opcode, pdu_len, tag);
 197
 198         if (pdu)
 199                 memcpy(&cmd->payload, pdu, pdu_len);
 200 }
 201
 202 const char *fio_server_op(unsigned int op)
 203 {
 204         static char buf[32];
 205
 206         if (op < FIO_NET_CMD_NR)
 207                 return fio_server_ops[op];
 208
 209         sprintf(buf, "UNKNOWN/%d", op);
 210         return buf;
 211 }
 212
 213 static ssize_t iov_total_len(const struct iovec *iov, int count)
 214 {
 215         ssize_t ret = 0;
 216
 217         while (count--) {
 218                 ret += iov->iov_len;
 219                 iov++;
 220         }
 221
 222         return ret;
 223 }
 224
 225 static int fio_sendv_data(int sk, struct iovec *iov, int count)
 226 {
 227         ssize_t total_len = iov_total_len(iov, count);
 228         ssize_t ret;
 229
 230         do {
 231                 ret = writev(sk, iov, count);
 232                 if (ret > 0) {
 233                         total_len -= ret;
 234                         if (!total_len)
 235                                 break;
 236
 237                         while (ret) {
 238                                 if (ret >= iov->iov_len) {
 239                                         ret -= iov->iov_len;
 240                                         iov++;
 241                                         continue;
 242                                 }
 243                                 iov->iov_base += ret;
 244                                 iov->iov_len -= ret;
 245                                 ret = 0;
 246                         }
 247                 } else if (!ret)
 248                         break;
 249                 else if (errno == EAGAIN || errno == EINTR)
 250                         continue;
 251                 else
 252                         break;
 253         } while (!exit_backend);
 254
 255         if (!total_len)
 256                 return 0;
 257
 258         return 1;
 259 }
 260
 261 static int fio_send_data(int sk, const void *p, unsigned int len)
 262 {
 263         struct iovec iov = { .iov_base = (void *) p, .iov_len = len };
 264
 265         assert(len <= sizeof(struct fio_net_cmd) + FIO_SERVER_MAX_FRAGMENT_PDU);
 266
 267         return fio_sendv_data(sk, &iov, 1);
 268 }
 269
 270 bool fio_server_poll_fd(int fd, short events, int timeout)
 271 {
 272         struct pollfd pfd = {
 273                 .fd     = fd,
 274                 .events = events,
 275         };
 276         int ret;
 277
 278         ret = poll(&pfd, 1, timeout);
 279         if (ret < 0) {
 280                 if (errno == EINTR)
 281                         return false;
 282                 log_err("fio: poll: %s\n", strerror(errno));
 283                 return false;
 284         } else if (!ret) {
 285                 return false;
 286         }
 287         if (pfd.revents & events)
 288                 return true;
 289         return false;
 290 }
 291
 292 static int fio_recv_data(int sk, void *buf, unsigned int len, bool wait)
 293 {
 294         int flags;
 295         char *p = buf;
 296
 297         if (wait)
 298                 flags = MSG_WAITALL;
 299         else
 300                 flags = OS_MSG_DONTWAIT;
 301
 302         do {
 303                 int ret = recv(sk, p, len, flags);
 304
 305                 if (ret > 0) {
 306                         len -= ret;
 307                         if (!len)
 308                                 break;
 309                         p += ret;
 310                         continue;
 311                 } else if (!ret)
 312                         break;
 313                 else if (errno == EAGAIN || errno == EINTR) {
 314                         if (wait)
 315                                 continue;
 316                         break;
 317                 } else
 318                         break;
 319         } while (!exit_backend);
 320
 321         if (!len)
 322                 return 0;
 323
 324         return -1;
 325 }
 326
 327 static int verify_convert_cmd(struct fio_net_cmd *cmd)
 328 {
 329         uint16_t crc;
 330
 331         cmd->cmd_crc16 = le16_to_cpu(cmd->cmd_crc16);
 332         cmd->pdu_crc16 = le16_to_cpu(cmd->pdu_crc16);
 333
 334         crc = fio_crc16(cmd, FIO_NET_CMD_CRC_SZ);
 335         if (crc != cmd->cmd_crc16) {
 336                 log_err("fio: server bad crc on command (got %x, wanted %x)\n",
 337                                 cmd->cmd_crc16, crc);
 338                 fprintf(f_err, "fio: server bad crc on command (got %x, wanted %x)\n",
 339                                 cmd->cmd_crc16, crc);
 340                 return 1;
 341         }
 342
 343         cmd->version    = le16_to_cpu(cmd->version);
 344         cmd->opcode     = le16_to_cpu(cmd->opcode);
 345         cmd->flags      = le32_to_cpu(cmd->flags);
 346         cmd->tag        = le64_to_cpu(cmd->tag);
 347         cmd->pdu_len    = le32_to_cpu(cmd->pdu_len);
 348
 349         switch (cmd->version) {
 350         case FIO_SERVER_VER:
 351                 break;
 352         default:
 353                 log_err("fio: bad server cmd version %d\n", cmd->version);
 354                 fprintf(f_err, "fio: client/server version mismatch (%d != %d)\n",
 355                                 cmd->version, FIO_SERVER_VER);
 356                 return 1;
 357         }
 358
 359         if (cmd->pdu_len > FIO_SERVER_MAX_FRAGMENT_PDU) {
 360                 log_err("fio: command payload too large: %u\n", cmd->pdu_len);
 361                 return 1;
 362         }
 363
 364         return 0;
 365 }
 366
 367 /*
 368  * Read (and defragment, if necessary) incoming commands
 369  */
 370 struct fio_net_cmd *fio_net_recv_cmd(int sk, bool wait)
 371 {
 372         struct fio_net_cmd cmd, *tmp, *cmdret = NULL;
 373         size_t cmd_size = 0, pdu_offset = 0;
 374         uint16_t crc;
 375         int ret, first = 1;
 376         void *pdu = NULL;
 377
 378         do {
 379                 ret = fio_recv_data(sk, &cmd, sizeof(cmd), wait);
 380                 if (ret)
 381                         break;
 382
 383                 /* We have a command, verify it and swap if need be */
 384                 ret = verify_convert_cmd(&cmd);
 385                 if (ret)
 386                         break;
 387
 388                 if (first) {
 389                         /* if this is text, add room for \0 at the end */
 390                         cmd_size = sizeof(cmd) + cmd.pdu_len + 1;
 391                         assert(!cmdret);
 392                 } else
 393                         cmd_size += cmd.pdu_len;
 394
 395                 if (cmd_size / 1024 > FIO_SERVER_MAX_CMD_MB * 1024) {
 396                         log_err("fio: cmd+pdu too large (%llu)\n", (unsigned long long) cmd_size);
 397                         ret = 1;
 398                         break;
 399                 }
 400
 401                 tmp = realloc(cmdret, cmd_size);
 402                 if (!tmp) {
 403                         log_err("fio: server failed allocating cmd\n");
 404                         ret = 1;
 405                         break;
 406                 }
 407                 cmdret = tmp;
 408
 409                 if (first)
 410                         memcpy(cmdret, &cmd, sizeof(cmd));
 411                 else if (cmdret->opcode != cmd.opcode) {
 412                         log_err("fio: fragment opcode mismatch (%d != %d)\n",
 413                                         cmdret->opcode, cmd.opcode);
 414                         ret = 1;
 415                         break;
 416                 }
 417
 418                 if (!cmd.pdu_len)
 419                         break;
 420
 421                 /* There's payload, get it */
 422                 pdu = (char *) cmdret->payload + pdu_offset;
 423                 ret = fio_recv_data(sk, pdu, cmd.pdu_len, wait);
 424                 if (ret)
 425                         break;
 426
 427                 /* Verify payload crc */
 428                 crc = fio_crc16(pdu, cmd.pdu_len);
 429                 if (crc != cmd.pdu_crc16) {
 430                         log_err("fio: server bad crc on payload ");
 431                         log_err("(got %x, wanted %x)\n", cmd.pdu_crc16, crc);
 432                         ret = 1;
 433                         break;
 434                 }
 435
 436                 pdu_offset += cmd.pdu_len;
 437                 if (!first)
 438                         cmdret->pdu_len += cmd.pdu_len;
 439                 first = 0;
 440         } while (cmd.flags & FIO_NET_CMD_F_MORE);
 441
 442         if (ret) {
 443                 free(cmdret);
 444                 cmdret = NULL;
 445         } else if (cmdret) {
 446                 /* zero-terminate text input */
 447                 if (cmdret->pdu_len) {
 448                         if (cmdret->opcode == FIO_NET_CMD_TEXT) {
 449                                 struct cmd_text_pdu *__pdu = (struct cmd_text_pdu *) cmdret->payload;
 450                                 char *buf = (char *) __pdu->buf;
 451                                 int len = le32_to_cpu(__pdu->buf_len);
 452
 453                                 buf[len] = '\0';
 454                         } else if (cmdret->opcode == FIO_NET_CMD_JOB) {
 455                                 struct cmd_job_pdu *__pdu = (struct cmd_job_pdu *) cmdret->payload;
 456                                 char *buf = (char *) __pdu->buf;
 457                                 int len = le32_to_cpu(__pdu->buf_len);
 458
 459                                 buf[len] = '\0';
 460                         }
 461                 }
 462
 463                 /* frag flag is internal */
 464                 cmdret->flags &= ~FIO_NET_CMD_F_MORE;
 465         }
 466
 467         return cmdret;
 468 }
 469
 470 static void add_reply(uint64_t tag, struct flist_head *list)
 471 {
 472         struct fio_net_cmd_reply *reply;
 473
 474         reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
 475         flist_add_tail(&reply->list, list);
 476 }
 477
 478 static uint64_t alloc_reply(uint64_t tag, uint16_t opcode)
 479 {
 480         struct fio_net_cmd_reply *reply;
 481
 482         reply = calloc(1, sizeof(*reply));
 483         INIT_FLIST_HEAD(&reply->list);
 484         fio_gettime(&reply->ts, NULL);
 485         reply->saved_tag = tag;
 486         reply->opcode = opcode;
 487
 488         return (uintptr_t) reply;
 489 }
 490
 491 static void free_reply(uint64_t tag)
 492 {
 493         struct fio_net_cmd_reply *reply;
 494
 495         reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
 496         free(reply);
 497 }
 498
 499 static void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
 500 {
 501         uint32_t pdu_len;
 502
 503         cmd->cmd_crc16 = __cpu_to_le16(fio_crc16(cmd, FIO_NET_CMD_CRC_SZ));
 504
 505         pdu_len = le32_to_cpu(cmd->pdu_len);
 506         cmd->pdu_crc16 = __cpu_to_le16(fio_crc16(pdu, pdu_len));
 507 }
 508
 509 static void fio_net_cmd_crc(struct fio_net_cmd *cmd)
 510 {
 511         fio_net_cmd_crc_pdu(cmd, cmd->payload);
 512 }
 513
 514 int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
 515                      uint64_t *tagptr, struct flist_head *list)
 516 {
 517         struct fio_net_cmd *cmd = NULL;
 518         size_t this_len, cur_len = 0;
 519         uint64_t tag;
 520         int ret;
 521
 522         if (list) {
 523                 assert(tagptr);
 524                 tag = *tagptr = alloc_reply(*tagptr, opcode);
 525         } else
 526                 tag = tagptr ? *tagptr : 0;
 527
 528         do {
 529                 this_len = size;
 530                 if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
 531                         this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
 532
 533                 if (!cmd || cur_len < sizeof(*cmd) + this_len) {
 534                         if (cmd)
 535                                 free(cmd);
 536
 537                         cur_len = sizeof(*cmd) + this_len;
 538                         cmd = malloc(cur_len);
 539                 }
 540
 541                 fio_init_net_cmd(cmd, opcode, buf, this_len, tag);
 542
 543                 if (this_len < size)
 544                         cmd->flags = __cpu_to_le32(FIO_NET_CMD_F_MORE);
 545
 546                 fio_net_cmd_crc(cmd);
 547
 548                 ret = fio_send_data(fd, cmd, sizeof(*cmd) + this_len);
 549                 size -= this_len;
 550                 buf += this_len;
 551         } while (!ret && size);
 552
 553         if (list) {
 554                 if (ret)
 555                         free_reply(tag);
 556                 else
 557                         add_reply(tag, list);
 558         }
 559
 560         if (cmd)
 561                 free(cmd);
 562
 563         return ret;
 564 }
 565
 566 static struct sk_entry *fio_net_prep_cmd(uint16_t opcode, void *buf,
 567                                          size_t size, uint64_t *tagptr,
 568                                          int flags)
 569 {
 570         struct sk_entry *entry;
 571
 572         entry = smalloc(sizeof(*entry));
 573         if (!entry)
 574                 return NULL;
 575
 576         INIT_FLIST_HEAD(&entry->next);
 577         entry->opcode = opcode;
 578         if (flags & SK_F_COPY) {
 579                 entry->buf = smalloc(size);
 580                 memcpy(entry->buf, buf, size);
 581         } else
 582                 entry->buf = buf;
 583
 584         entry->size = size;
 585         if (tagptr)
 586                 entry->tag = *tagptr;
 587         else
 588                 entry->tag = 0;
 589         entry->flags = flags;
 590         return entry;
 591 }
 592
 593 static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry);
 594
 595 static void fio_net_queue_entry(struct sk_entry *entry)
 596 {
 597         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
 598
 599         if (entry->flags & SK_F_INLINE)
 600                 handle_sk_entry(sk_out, entry);
 601         else {
 602                 sk_lock(sk_out);
 603                 flist_add_tail(&entry->list, &sk_out->list);
 604                 sk_unlock(sk_out);
 605
 606                 fio_sem_up(&sk_out->wait);
 607         }
 608 }
 609
 610 static int fio_net_queue_cmd(uint16_t opcode, void *buf, off_t size,
 611                              uint64_t *tagptr, int flags)
 612 {
 613         struct sk_entry *entry;
 614
 615         entry = fio_net_prep_cmd(opcode, buf, size, tagptr, flags);
 616         if (entry) {
 617                 fio_net_queue_entry(entry);
 618                 return 0;
 619         }
 620
 621         return 1;
 622 }
 623
 624 static int fio_net_send_simple_stack_cmd(int sk, uint16_t opcode, uint64_t tag)
 625 {
 626         struct fio_net_cmd cmd;
 627
 628         fio_init_net_cmd(&cmd, opcode, NULL, 0, tag);
 629         fio_net_cmd_crc(&cmd);
 630
 631         return fio_send_data(sk, &cmd, sizeof(cmd));
 632 }
 633
 634 /*
 635  * If 'list' is non-NULL, then allocate and store the sent command for
 636  * later verification.
 637  */
 638 int fio_net_send_simple_cmd(int sk, uint16_t opcode, uint64_t tag,
 639                             struct flist_head *list)
 640 {
 641         int ret;
 642
 643         if (list)
 644                 tag = alloc_reply(tag, opcode);
 645
 646         ret = fio_net_send_simple_stack_cmd(sk, opcode, tag);
 647         if (ret) {
 648                 if (list)
 649                         free_reply(tag);
 650
 651                 return ret;
 652         }
 653
 654         if (list)
 655                 add_reply(tag, list);
 656
 657         return 0;
 658 }
 659
 660 static int fio_net_queue_quit(void)
 661 {
 662         dprint(FD_NET, "server: sending quit\n");
 663
 664         return fio_net_queue_cmd(FIO_NET_CMD_QUIT, NULL, 0, NULL, SK_F_SIMPLE);
 665 }
 666
 667 int fio_net_send_quit(int sk)
 668 {
 669         dprint(FD_NET, "server: sending quit\n");
 670
 671         return fio_net_send_simple_cmd(sk, FIO_NET_CMD_QUIT, 0, NULL);
 672 }
 673
 674 static int fio_net_send_ack(struct fio_net_cmd *cmd, int error, int signal)
 675 {
 676         struct cmd_end_pdu epdu;
 677         uint64_t tag = 0;
 678
 679         if (cmd)
 680                 tag = cmd->tag;
 681
 682         epdu.error = __cpu_to_le32(error);
 683         epdu.signal = __cpu_to_le32(signal);
 684         return fio_net_queue_cmd(FIO_NET_CMD_STOP, &epdu, sizeof(epdu), &tag, SK_F_COPY);
 685 }
 686
 687 static int fio_net_queue_stop(int error, int signal)
 688 {
 689         dprint(FD_NET, "server: sending stop (%d, %d)\n", error, signal);
 690         return fio_net_send_ack(NULL, error, signal);
 691 }
 692
 693 #ifdef WIN32
 694 static void fio_server_add_fork_item(struct ffi_element *element, struct flist_head *list)
 695 {
 696         struct fio_fork_item *ffi;
 697
 698         ffi = malloc(sizeof(*ffi));
 699         ffi->exitval = 0;
 700         ffi->signal = 0;
 701         ffi->exited = 0;
 702         ffi->element = *element;
 703         flist_add_tail(&ffi->list, list);
 704 }
 705
 706 static void fio_server_add_conn_pid(struct flist_head *conn_list, HANDLE hProcess)
 707 {
 708         struct ffi_element element = {.hProcess = hProcess, .is_thread=FALSE};
 709         dprint(FD_NET, "server: forked off connection job (tid=%u)\n", (int) element.thread);
 710
 711         fio_server_add_fork_item(&element, conn_list);
 712 }
 713
 714 static void fio_server_add_job_pid(struct flist_head *job_list, pthread_t thread)
 715 {
 716         struct ffi_element element = {.thread = thread, .is_thread=TRUE};
 717         dprint(FD_NET, "server: forked off job job (tid=%u)\n", (int) element.thread);
 718         fio_server_add_fork_item(&element, job_list);
 719 }
 720
 721 static void fio_server_check_fork_item(struct fio_fork_item *ffi)
 722 {
 723         int ret;
 724
 725         if (ffi->element.is_thread) {
 726
 727                 ret = pthread_kill(ffi->element.thread, 0);
 728                 if (ret) {
 729                         int rev_val;
 730                         pthread_join(ffi->element.thread, (void**) &rev_val); /*if the thread is dead, then join it to get status*/
 731
 732                         ffi->exitval = rev_val;
 733                         if (ffi->exitval)
 734                                 log_err("thread (tid=%u) exited with %x\n", (int) ffi->element.thread, (int) ffi->exitval);
 735                         dprint(FD_PROCESS, "thread (tid=%u) exited with %x\n", (int) ffi->element.thread, (int) ffi->exitval);
 736                         ffi->exited = 1;
 737                 }
 738         } else {
 739                 DWORD exit_val;
 740                 GetExitCodeProcess(ffi->element.hProcess, &exit_val);
 741
 742                 if (exit_val != STILL_ACTIVE) {
 743                         dprint(FD_PROCESS, "process %u exited with %d\n", GetProcessId(ffi->element.hProcess), exit_val);
 744                         ffi->exited = 1;
 745                         ffi->exitval = exit_val;
 746                 }
 747         }
 748 }
 749 #else
 750 static void fio_server_add_fork_item(pid_t pid, struct flist_head *list)
 751 {
 752         struct fio_fork_item *ffi;
 753
 754         ffi = malloc(sizeof(*ffi));
 755         ffi->exitval = 0;
 756         ffi->signal = 0;
 757         ffi->exited = 0;
 758         ffi->pid = pid;
 759         flist_add_tail(&ffi->list, list);
 760 }
 761
 762 static void fio_server_add_conn_pid(struct flist_head *conn_list, pid_t pid)
 763 {
 764         dprint(FD_NET, "server: forked off connection job (pid=%u)\n", (int) pid);
 765         fio_server_add_fork_item(pid, conn_list);
 766 }
 767
 768 static void fio_server_add_job_pid(struct flist_head *job_list, pid_t pid)
 769 {
 770         dprint(FD_NET, "server: forked off job job (pid=%u)\n", (int) pid);
 771         fio_server_add_fork_item(pid, job_list);
 772 }
 773
 774 static void fio_server_check_fork_item(struct fio_fork_item *ffi)
 775 {
 776         int ret, status;
 777
 778         ret = waitpid(ffi->pid, &status, WNOHANG);
 779         if (ret < 0) {
 780                 if (errno == ECHILD) {
 781                         log_err("fio: connection pid %u disappeared\n", (int) ffi->pid);
 782                         ffi->exited = 1;
 783                 } else
 784                         log_err("fio: waitpid: %s\n", strerror(errno));
 785         } else if (ret == ffi->pid) {
 786                 if (WIFSIGNALED(status)) {
 787                         ffi->signal = WTERMSIG(status);
 788                         ffi->exited = 1;
 789                 }
 790                 if (WIFEXITED(status)) {
 791                         if (WEXITSTATUS(status))
 792                                 ffi->exitval = WEXITSTATUS(status);
 793                         ffi->exited = 1;
 794                 }
 795         }
 796 }
 797 #endif
 798
 799 static void fio_server_fork_item_done(struct fio_fork_item *ffi, bool stop)
 800 {
 801 #ifdef WIN32
 802         if (ffi->element.is_thread)
 803                 dprint(FD_NET, "tid %u exited, sig=%u, exitval=%d\n", (int) ffi->element.thread, ffi->signal, ffi->exitval);
 804         else {
 805                 dprint(FD_NET, "pid %u exited, sig=%u, exitval=%d\n", (int)  GetProcessId(ffi->element.hProcess), ffi->signal, ffi->exitval);
 806                 CloseHandle(ffi->element.hProcess);
 807                 ffi->element.hProcess = INVALID_HANDLE_VALUE;
 808         }
 809 #else
 810         dprint(FD_NET, "pid %u exited, sig=%u, exitval=%d\n", (int) ffi->pid, ffi->signal, ffi->exitval);
 811 #endif
 812
 813         /*
 814          * Fold STOP and QUIT...
 815          */
 816         if (stop) {
 817                 fio_net_queue_stop(ffi->exitval, ffi->signal);
 818                 fio_net_queue_quit();
 819         }
 820
 821         flist_del(&ffi->list);
 822         free(ffi);
 823 }
 824
 825 static void fio_server_check_fork_items(struct flist_head *list, bool stop)
 826 {
 827         struct flist_head *entry, *tmp;
 828         struct fio_fork_item *ffi;
 829
 830         flist_for_each_safe(entry, tmp, list) {
 831                 ffi = flist_entry(entry, struct fio_fork_item, list);
 832
 833                 fio_server_check_fork_item(ffi);
 834
 835                 if (ffi->exited)
 836                         fio_server_fork_item_done(ffi, stop);
 837         }
 838 }
 839
 840 static void fio_server_check_jobs(struct flist_head *job_list)
 841 {
 842         fio_server_check_fork_items(job_list, true);
 843 }
 844
 845 static void fio_server_check_conns(struct flist_head *conn_list)
 846 {
 847         fio_server_check_fork_items(conn_list, false);
 848 }
 849
 850 static int handle_load_file_cmd(struct fio_net_cmd *cmd)
 851 {
 852         struct cmd_load_file_pdu *pdu = (struct cmd_load_file_pdu *) cmd->payload;
 853         void *file_name = pdu->file;
 854         struct cmd_start_pdu spdu;
 855
 856         dprint(FD_NET, "server: loading local file %s\n", (char *) file_name);
 857
 858         pdu->name_len = le16_to_cpu(pdu->name_len);
 859         pdu->client_type = le16_to_cpu(pdu->client_type);
 860
 861         if (parse_jobs_ini(file_name, 0, 0, pdu->client_type)) {
 862                 fio_net_queue_quit();
 863                 return -1;
 864         }
 865
 866         spdu.jobs = cpu_to_le32(thread_number);
 867         spdu.stat_outputs = cpu_to_le32(stat_number);
 868         fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
 869         return 0;
 870 }
 871
 872 #ifdef WIN32
 873 static void *fio_backend_thread(void *data)
 874 {
 875         int ret;
 876         struct sk_out *sk_out = (struct sk_out *) data;
 877
 878         sk_out_assign(sk_out);
 879
 880         ret = fio_backend(sk_out);
 881         sk_out_drop();
 882
 883         pthread_exit((void*) (intptr_t) ret);
 884         return NULL;
 885 }
 886 #endif
 887
 888 static int handle_run_cmd(struct sk_out *sk_out, struct flist_head *job_list,
 889                           struct fio_net_cmd *cmd)
 890 {
 891         int ret;
 892
 893         fio_time_init();
 894         set_genesis_time();
 895
 896 #ifdef WIN32
 897         {
 898                 pthread_t thread;
 899                 /* both this thread and backend_thread call sk_out_assign() to double increment
 900                  * the ref count.  This ensures struct is valid until both threads are done with it
 901                  */
 902                 sk_out_assign(sk_out);
 903                 ret = pthread_create(&thread, NULL,     fio_backend_thread, sk_out);
 904                 if (ret) {
 905                         log_err("pthread_create: %s\n", strerror(ret));
 906                         return ret;
 907                 }
 908
 909                 fio_server_add_job_pid(job_list, thread);
 910                 return ret;
 911         }
 912 #else
 913     {
 914                 pid_t pid;
 915                 sk_out_assign(sk_out);
 916                 pid = fork();
 917                 if (pid) {
 918                         fio_server_add_job_pid(job_list, pid);
 919                         return 0;
 920                 }
 921
 922                 ret = fio_backend(sk_out);
 923                 free_threads_shm();
 924                 sk_out_drop();
 925                 _exit(ret);
 926         }
 927 #endif
 928 }
 929
 930 static int handle_job_cmd(struct fio_net_cmd *cmd)
 931 {
 932         struct cmd_job_pdu *pdu = (struct cmd_job_pdu *) cmd->payload;
 933         void *buf = pdu->buf;
 934         struct cmd_start_pdu spdu;
 935
 936         pdu->buf_len = le32_to_cpu(pdu->buf_len);
 937         pdu->client_type = le32_to_cpu(pdu->client_type);
 938
 939         if (parse_jobs_ini(buf, 1, 0, pdu->client_type)) {
 940                 fio_net_queue_quit();
 941                 return -1;
 942         }
 943
 944         spdu.jobs = cpu_to_le32(thread_number);
 945         spdu.stat_outputs = cpu_to_le32(stat_number);
 946
 947         fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
 948         return 0;
 949 }
 950
 951 static int handle_jobline_cmd(struct fio_net_cmd *cmd)
 952 {
 953         void *pdu = cmd->payload;
 954         struct cmd_single_line_pdu *cslp;
 955         struct cmd_line_pdu *clp;
 956         unsigned long offset;
 957         struct cmd_start_pdu spdu;
 958         char **argv;
 959         int i;
 960
 961         clp = pdu;
 962         clp->lines = le16_to_cpu(clp->lines);
 963         clp->client_type = le16_to_cpu(clp->client_type);
 964         argv = malloc(clp->lines * sizeof(char *));
 965         offset = sizeof(*clp);
 966
 967         dprint(FD_NET, "server: %d command line args\n", clp->lines);
 968
 969         for (i = 0; i < clp->lines; i++) {
 970                 cslp = pdu + offset;
 971                 argv[i] = (char *) cslp->text;
 972
 973                 offset += sizeof(*cslp) + le16_to_cpu(cslp->len);
 974                 dprint(FD_NET, "server: %d: %s\n", i, argv[i]);
 975         }
 976
 977         if (parse_cmd_line(clp->lines, argv, clp->client_type)) {
 978                 fio_net_queue_quit();
 979                 free(argv);
 980                 return -1;
 981         }
 982
 983         free(argv);
 984
 985         spdu.jobs = cpu_to_le32(thread_number);
 986         spdu.stat_outputs = cpu_to_le32(stat_number);
 987
 988         fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
 989         return 0;
 990 }
 991
 992 static int handle_probe_cmd(struct fio_net_cmd *cmd)
 993 {
 994         struct cmd_client_probe_pdu *pdu = (struct cmd_client_probe_pdu *) cmd->payload;
 995         uint64_t tag = cmd->tag;
 996         struct cmd_probe_reply_pdu probe = {
 997 #ifdef CONFIG_BIG_ENDIAN
 998                 .bigendian      = 1,
 999 #endif
1000                 .os             = FIO_OS,
1001                 .arch           = FIO_ARCH,
1002                 .bpp            = sizeof(void *),
1003                 .cpus           = __cpu_to_le32(cpus_configured()),
1004         };
1005
1006         dprint(FD_NET, "server: sending probe reply\n");
1007
1008         strcpy(me, (char *) pdu->server);
1009
1010         gethostname((char *) probe.hostname, sizeof(probe.hostname));
1011         snprintf((char *) probe.fio_version, sizeof(probe.fio_version), "%s",
1012                  fio_version_string);
1013
1014         /*
1015          * If the client supports compression and we do too, then enable it
1016          */
1017         if (has_zlib && le64_to_cpu(pdu->flags) & FIO_PROBE_FLAG_ZLIB) {
1018                 probe.flags = __cpu_to_le64(FIO_PROBE_FLAG_ZLIB);
1019                 use_zlib = 1;
1020         } else {
1021                 probe.flags = 0;
1022                 use_zlib = 0;
1023         }
1024
1025         return fio_net_queue_cmd(FIO_NET_CMD_PROBE, &probe, sizeof(probe), &tag, SK_F_COPY);
1026 }
1027
1028 static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
1029 {
1030         struct jobs_eta *je;
1031         uint64_t tag = cmd->tag;
1032         size_t size;
1033         int i;
1034
1035         dprint(FD_NET, "server sending status\n");
1036
1037         /*
1038          * Fake ETA return if we don't have a local one, otherwise the client
1039          * will end up timing out waiting for a response to the ETA request
1040          */
1041         je = get_jobs_eta(true, &size);
1042         if (!je) {
1043                 size = sizeof(*je);
1044                 je = calloc(1, size);
1045         } else {
1046                 je->nr_running          = cpu_to_le32(je->nr_running);
1047                 je->nr_ramp             = cpu_to_le32(je->nr_ramp);
1048                 je->nr_pending          = cpu_to_le32(je->nr_pending);
1049                 je->nr_setting_up       = cpu_to_le32(je->nr_setting_up);
1050                 je->files_open          = cpu_to_le32(je->files_open);
1051
1052                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1053                         je->m_rate[i]   = cpu_to_le64(je->m_rate[i]);
1054                         je->t_rate[i]   = cpu_to_le64(je->t_rate[i]);
1055                         je->m_iops[i]   = cpu_to_le32(je->m_iops[i]);
1056                         je->t_iops[i]   = cpu_to_le32(je->t_iops[i]);
1057                         je->rate[i]     = cpu_to_le64(je->rate[i]);
1058                         je->iops[i]     = cpu_to_le32(je->iops[i]);
1059                 }
1060
1061                 je->elapsed_sec         = cpu_to_le64(je->elapsed_sec);
1062                 je->eta_sec             = cpu_to_le64(je->eta_sec);
1063                 je->nr_threads          = cpu_to_le32(je->nr_threads);
1064                 je->is_pow2             = cpu_to_le32(je->is_pow2);
1065                 je->unit_base           = cpu_to_le32(je->unit_base);
1066         }
1067
1068         fio_net_queue_cmd(FIO_NET_CMD_ETA, je, size, &tag, SK_F_FREE);
1069         return 0;
1070 }
1071
1072 static int send_update_job_reply(uint64_t __tag, int error)
1073 {
1074         uint64_t tag = __tag;
1075         uint32_t pdu_error;
1076
1077         pdu_error = __cpu_to_le32(error);
1078         return fio_net_queue_cmd(FIO_NET_CMD_UPDATE_JOB, &pdu_error, sizeof(pdu_error), &tag, SK_F_COPY);
1079 }
1080
1081 static int handle_update_job_cmd(struct fio_net_cmd *cmd)
1082 {
1083         struct cmd_add_job_pdu *pdu = (struct cmd_add_job_pdu *) cmd->payload;
1084         struct thread_data *td;
1085         uint32_t tnumber;
1086         int ret;
1087
1088         tnumber = le32_to_cpu(pdu->thread_number);
1089
1090         dprint(FD_NET, "server: updating options for job %u\n", tnumber);
1091
1092         if (!tnumber || tnumber > thread_number) {
1093                 send_update_job_reply(cmd->tag, ENODEV);
1094                 return 0;
1095         }
1096
1097         td = tnumber_to_td(tnumber);
1098         ret = convert_thread_options_to_cpu(&td->o, &pdu->top,
1099                         cmd->pdu_len - offsetof(struct cmd_add_job_pdu, top));
1100         send_update_job_reply(cmd->tag, ret);
1101         return 0;
1102 }
1103
1104 static int handle_trigger_cmd(struct fio_net_cmd *cmd, struct flist_head *job_list)
1105 {
1106         struct cmd_vtrigger_pdu *pdu = (struct cmd_vtrigger_pdu *) cmd->payload;
1107         char *buf = (char *) pdu->cmd;
1108         struct all_io_list *rep;
1109         size_t sz;
1110
1111         pdu->len = le16_to_cpu(pdu->len);
1112         buf[pdu->len] = '\0';
1113
1114         rep = get_all_io_list(IO_LIST_ALL, &sz);
1115         if (!rep) {
1116                 struct all_io_list state;
1117
1118                 state.threads = cpu_to_le64((uint64_t) 0);
1119                 fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, &state, sizeof(state), NULL, SK_F_COPY | SK_F_INLINE);
1120         } else
1121                 fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, rep, sz, NULL, SK_F_FREE | SK_F_INLINE);
1122
1123         fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
1124         fio_server_check_jobs(job_list);
1125         exec_trigger(buf);
1126         return 0;
1127 }
1128
1129 static int handle_command(struct sk_out *sk_out, struct flist_head *job_list,
1130                           struct fio_net_cmd *cmd)
1131 {
1132         int ret;
1133
1134         dprint(FD_NET, "server: got op [%s], pdu=%u, tag=%llx\n",
1135                         fio_server_op(cmd->opcode), cmd->pdu_len,
1136                         (unsigned long long) cmd->tag);
1137
1138         switch (cmd->opcode) {
1139         case FIO_NET_CMD_QUIT:
1140                 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
1141                 ret = 0;
1142                 break;
1143         case FIO_NET_CMD_EXIT:
1144                 exit_backend = true;
1145                 return -1;
1146         case FIO_NET_CMD_LOAD_FILE:
1147                 ret = handle_load_file_cmd(cmd);
1148                 break;
1149         case FIO_NET_CMD_JOB:
1150                 ret = handle_job_cmd(cmd);
1151                 break;
1152         case FIO_NET_CMD_JOBLINE:
1153                 ret = handle_jobline_cmd(cmd);
1154                 break;
1155         case FIO_NET_CMD_PROBE:
1156                 ret = handle_probe_cmd(cmd);
1157                 break;
1158         case FIO_NET_CMD_SEND_ETA:
1159                 ret = handle_send_eta_cmd(cmd);
1160                 break;
1161         case FIO_NET_CMD_RUN:
1162                 ret = handle_run_cmd(sk_out, job_list, cmd);
1163                 break;
1164         case FIO_NET_CMD_UPDATE_JOB:
1165                 ret = handle_update_job_cmd(cmd);
1166                 break;
1167         case FIO_NET_CMD_VTRIGGER:
1168                 ret = handle_trigger_cmd(cmd, job_list);
1169                 break;
1170         case FIO_NET_CMD_SENDFILE: {
1171                 struct cmd_sendfile_reply *in;
1172                 struct cmd_reply *rep;
1173
1174                 rep = (struct cmd_reply *) (uintptr_t) cmd->tag;
1175
1176                 in = (struct cmd_sendfile_reply *) cmd->payload;
1177                 in->size = le32_to_cpu(in->size);
1178                 in->error = le32_to_cpu(in->error);
1179                 if (in->error) {
1180                         ret = 1;
1181                         rep->error = in->error;
1182                 } else {
1183                         ret = 0;
1184                         rep->data = smalloc(in->size);
1185                         if (!rep->data) {
1186                                 ret = 1;
1187                                 rep->error = ENOMEM;
1188                         } else {
1189                                 rep->size = in->size;
1190                                 memcpy(rep->data, in->data, in->size);
1191                         }
1192                 }
1193                 fio_sem_up(&rep->lock);
1194                 break;
1195                 }
1196         default:
1197                 log_err("fio: unknown opcode: %s\n", fio_server_op(cmd->opcode));
1198                 ret = 1;
1199         }
1200
1201         return ret;
1202 }
1203
1204 /*
1205  * Send a command with a separate PDU, not inlined in the command
1206  */
1207 static int fio_send_cmd_ext_pdu(int sk, uint16_t opcode, const void *buf,
1208                                 off_t size, uint64_t tag, uint32_t flags)
1209 {
1210         struct fio_net_cmd cmd;
1211         struct iovec iov[2];
1212         size_t this_len;
1213         int ret;
1214
1215         iov[0].iov_base = (void *) &cmd;
1216         iov[0].iov_len = sizeof(cmd);
1217
1218         do {
1219                 uint32_t this_flags = flags;
1220
1221                 this_len = size;
1222                 if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
1223                         this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
1224
1225                 if (this_len < size)
1226                         this_flags |= FIO_NET_CMD_F_MORE;
1227
1228                 __fio_init_net_cmd(&cmd, opcode, this_len, tag);
1229                 cmd.flags = __cpu_to_le32(this_flags);
1230                 fio_net_cmd_crc_pdu(&cmd, buf);
1231
1232                 iov[1].iov_base = (void *) buf;
1233                 iov[1].iov_len = this_len;
1234
1235                 ret = fio_sendv_data(sk, iov, 2);
1236                 size -= this_len;
1237                 buf += this_len;
1238         } while (!ret && size);
1239
1240         return ret;
1241 }
1242
1243 static void finish_entry(struct sk_entry *entry)
1244 {
1245         if (entry->flags & SK_F_FREE)
1246                 free(entry->buf);
1247         else if (entry->flags & SK_F_COPY)
1248                 sfree(entry->buf);
1249
1250         sfree(entry);
1251 }
1252
1253 static void entry_set_flags(struct sk_entry *entry, struct flist_head *list,
1254                             unsigned int *flags)
1255 {
1256         if (!flist_empty(list))
1257                 *flags = FIO_NET_CMD_F_MORE;
1258         else
1259                 *flags = 0;
1260 }
1261
1262 static int send_vec_entry(struct sk_out *sk_out, struct sk_entry *first)
1263 {
1264         unsigned int flags;
1265         int ret;
1266
1267         entry_set_flags(first, &first->next, &flags);
1268
1269         ret = fio_send_cmd_ext_pdu(sk_out->sk, first->opcode, first->buf,
1270                                         first->size, first->tag, flags);
1271
1272         while (!flist_empty(&first->next)) {
1273                 struct sk_entry *next;
1274
1275                 next = flist_first_entry(&first->next, struct sk_entry, list);
1276                 flist_del_init(&next->list);
1277
1278                 entry_set_flags(next, &first->next, &flags);
1279
1280                 ret += fio_send_cmd_ext_pdu(sk_out->sk, next->opcode, next->buf,
1281                                                 next->size, next->tag, flags);
1282                 finish_entry(next);
1283         }
1284
1285         return ret;
1286 }
1287
1288 static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry)
1289 {
1290         int ret;
1291
1292         fio_sem_down(&sk_out->xmit);
1293
1294         if (entry->flags & SK_F_VEC)
1295                 ret = send_vec_entry(sk_out, entry);
1296         else if (entry->flags & SK_F_SIMPLE) {
1297                 ret = fio_net_send_simple_cmd(sk_out->sk, entry->opcode,
1298                                                 entry->tag, NULL);
1299         } else {
1300                 ret = fio_net_send_cmd(sk_out->sk, entry->opcode, entry->buf,
1301                                         entry->size, &entry->tag, NULL);
1302         }
1303
1304         fio_sem_up(&sk_out->xmit);
1305
1306         if (ret)
1307                 log_err("fio: failed handling cmd %s\n", fio_server_op(entry->opcode));
1308
1309         finish_entry(entry);
1310         return ret;
1311 }
1312
1313 static int handle_xmits(struct sk_out *sk_out)
1314 {
1315         struct sk_entry *entry;
1316         FLIST_HEAD(list);
1317         int ret = 0;
1318
1319         sk_lock(sk_out);
1320         if (flist_empty(&sk_out->list)) {
1321                 sk_unlock(sk_out);
1322                 return 0;
1323         }
1324
1325         flist_splice_init(&sk_out->list, &list);
1326         sk_unlock(sk_out);
1327
1328         while (!flist_empty(&list)) {
1329                 entry = flist_first_entry(&list, struct sk_entry, list);
1330                 flist_del(&entry->list);
1331                 ret += handle_sk_entry(sk_out, entry);
1332         }
1333
1334         return ret;
1335 }
1336
1337 static int handle_connection(struct sk_out *sk_out)
1338 {
1339         struct fio_net_cmd *cmd = NULL;
1340         FLIST_HEAD(job_list);
1341         int ret = 0;
1342
1343         reset_fio_state();
1344
1345         /* read forever */
1346         while (!exit_backend) {
1347                 struct pollfd pfd = {
1348                         .fd     = sk_out->sk,
1349                         .events = POLLIN,
1350                 };
1351
1352                 do {
1353                         int timeout = 1000;
1354
1355                         if (!flist_empty(&job_list))
1356                                 timeout = 100;
1357
1358                         handle_xmits(sk_out);
1359
1360                         ret = poll(&pfd, 1, 0);
1361                         if (ret < 0) {
1362                                 if (errno == EINTR)
1363                                         break;
1364                                 log_err("fio: poll: %s\n", strerror(errno));
1365                                 break;
1366                         } else if (!ret) {
1367                                 fio_server_check_jobs(&job_list);
1368                                 fio_sem_down_timeout(&sk_out->wait, timeout);
1369                                 continue;
1370                         }
1371
1372                         if (pfd.revents & POLLIN)
1373                                 break;
1374                         if (pfd.revents & (POLLERR|POLLHUP)) {
1375                                 ret = 1;
1376                                 break;
1377                         }
1378                 } while (!exit_backend);
1379
1380                 fio_server_check_jobs(&job_list);
1381
1382                 if (ret < 0)
1383                         break;
1384
1385                 if (pfd.revents & POLLIN)
1386                         cmd = fio_net_recv_cmd(sk_out->sk, true);
1387                 if (!cmd) {
1388                         ret = -1;
1389                         break;
1390                 }
1391
1392                 ret = handle_command(sk_out, &job_list, cmd);
1393                 if (ret)
1394                         break;
1395
1396                 free(cmd);
1397                 cmd = NULL;
1398         }
1399
1400         if (cmd)
1401                 free(cmd);
1402
1403         handle_xmits(sk_out);
1404
1405         close(sk_out->sk);
1406         sk_out->sk = -1;
1407         __sk_out_drop(sk_out);
1408         _exit(ret);
1409 }
1410
1411 /* get the address on this host bound by the input socket,
1412  * whether it is ipv6 or ipv4 */
1413
1414 static int get_my_addr_str(int sk)
1415 {
1416         struct sockaddr_in6 myaddr6 = { 0, };
1417         struct sockaddr_in myaddr4 = { 0, };
1418         struct sockaddr *sockaddr_p;
1419         char *net_addr;
1420         socklen_t len;
1421         int ret;
1422
1423         if (use_ipv6) {
1424                 len = sizeof(myaddr6);
1425                 sockaddr_p = (struct sockaddr * )&myaddr6;
1426                 net_addr = (char * )&myaddr6.sin6_addr;
1427         } else {
1428                 len = sizeof(myaddr4);
1429                 sockaddr_p = (struct sockaddr * )&myaddr4;
1430                 net_addr = (char * )&myaddr4.sin_addr;
1431         }
1432
1433         ret = getsockname(sk, sockaddr_p, &len);
1434         if (ret) {
1435                 log_err("fio: getsockname: %s\n", strerror(errno));
1436                 return -1;
1437         }
1438
1439         if (!inet_ntop(use_ipv6?AF_INET6:AF_INET, net_addr, client_sockaddr_str, INET6_ADDRSTRLEN - 1)) {
1440                 log_err("inet_ntop: failed to convert addr to string\n");
1441                 return -1;
1442         }
1443
1444         dprint(FD_NET, "fio server bound to addr %s\n", client_sockaddr_str);
1445         return 0;
1446 }
1447
1448 #ifdef WIN32
1449 static int handle_connection_process(void)
1450 {
1451         WSAPROTOCOL_INFO protocol_info;
1452         DWORD bytes_read;
1453         HANDLE hpipe;
1454         int sk;
1455         struct sk_out *sk_out;
1456         int ret;
1457         char *msg = (char *) "connected";
1458
1459         log_info("server enter accept loop.  ProcessID %d\n", GetCurrentProcessId());
1460
1461         hpipe = CreateFile(
1462                                         fio_server_pipe_name,
1463                                         GENERIC_READ | GENERIC_WRITE,
1464                                         0, NULL,
1465                                         OPEN_EXISTING,
1466                                         0, NULL);
1467
1468         if (hpipe == INVALID_HANDLE_VALUE) {
1469                 log_err("couldnt open pipe %s error %lu\n",
1470                                 fio_server_pipe_name, GetLastError());
1471                 return -1;
1472         }
1473
1474         if (!ReadFile(hpipe, &protocol_info, sizeof(protocol_info), &bytes_read, NULL)) {
1475                 log_err("couldnt read pi from pipe %s error %lu\n", fio_server_pipe_name,
1476                                 GetLastError());
1477         }
1478
1479         if (use_ipv6) /* use protocol_info to create a duplicate of parents socket */
1480                 sk = WSASocket(AF_INET6, SOCK_STREAM, 0, &protocol_info, 0, 0);
1481         else
1482                 sk = WSASocket(AF_INET,  SOCK_STREAM, 0, &protocol_info, 0, 0);
1483
1484         sk_out = scalloc(1, sizeof(*sk_out));
1485         if (!sk_out) {
1486                 CloseHandle(hpipe);
1487                 close(sk);
1488                 return -1;
1489         }
1490
1491         sk_out->sk = sk;
1492         sk_out->hProcess = INVALID_HANDLE_VALUE;
1493         INIT_FLIST_HEAD(&sk_out->list);
1494         __fio_sem_init(&sk_out->lock, FIO_SEM_UNLOCKED);
1495         __fio_sem_init(&sk_out->wait, FIO_SEM_LOCKED);
1496         __fio_sem_init(&sk_out->xmit, FIO_SEM_UNLOCKED);
1497
1498         get_my_addr_str(sk);
1499
1500         if (!WriteFile(hpipe, msg, strlen(msg), NULL, NULL)) {
1501                 log_err("couldnt write pipe\n");
1502                 close(sk);
1503                 return -1;
1504         }
1505         CloseHandle(hpipe);
1506
1507         sk_out_assign(sk_out);
1508
1509         ret = handle_connection(sk_out);
1510         __sk_out_drop(sk_out);
1511         return ret;
1512 }
1513 #endif
1514
1515 static int accept_loop(int listen_sk)
1516 {
1517         struct sockaddr_in addr;
1518         struct sockaddr_in6 addr6;
1519         socklen_t len = use_ipv6 ? sizeof(addr6) : sizeof(addr);
1520         struct pollfd pfd;
1521         int ret = 0, sk, exitval = 0;
1522         FLIST_HEAD(conn_list);
1523
1524         dprint(FD_NET, "server enter accept loop\n");
1525
1526         fio_set_fd_nonblocking(listen_sk, "server");
1527
1528         while (!exit_backend) {
1529                 struct sk_out *sk_out;
1530                 const char *from;
1531                 char buf[64];
1532 #ifdef WIN32
1533                 HANDLE hProcess;
1534 #else
1535                 pid_t pid;
1536 #endif
1537                 pfd.fd = listen_sk;
1538                 pfd.events = POLLIN;
1539                 do {
1540                         int timeout = 1000;
1541
1542                         if (!flist_empty(&conn_list))
1543                                 timeout = 100;
1544
1545                         ret = poll(&pfd, 1, timeout);
1546                         if (ret < 0) {
1547                                 if (errno == EINTR)
1548                                         break;
1549                                 log_err("fio: poll: %s\n", strerror(errno));
1550                                 break;
1551                         } else if (!ret) {
1552                                 fio_server_check_conns(&conn_list);
1553                                 continue;
1554                         }
1555
1556                         if (pfd.revents & POLLIN)
1557                                 break;
1558                 } while (!exit_backend);
1559
1560                 fio_server_check_conns(&conn_list);
1561
1562                 if (exit_backend || ret < 0)
1563                         break;
1564
1565                 if (use_ipv6)
1566                         sk = accept(listen_sk, (struct sockaddr *) &addr6, &len);
1567                 else
1568                         sk = accept(listen_sk, (struct sockaddr *) &addr, &len);
1569
1570                 if (sk < 0) {
1571                         log_err("fio: accept: %s\n", strerror(errno));
1572                         return -1;
1573                 }
1574
1575                 if (use_ipv6)
1576                         from = inet_ntop(AF_INET6, (struct sockaddr *) &addr6.sin6_addr, buf, sizeof(buf));
1577                 else
1578                         from = inet_ntop(AF_INET, (struct sockaddr *) &addr.sin_addr, buf, sizeof(buf));
1579
1580                 dprint(FD_NET, "server: connect from %s\n", from);
1581
1582                 sk_out = scalloc(1, sizeof(*sk_out));
1583                 if (!sk_out) {
1584                         close(sk);
1585                         return -1;
1586                 }
1587
1588                 sk_out->sk = sk;
1589                 INIT_FLIST_HEAD(&sk_out->list);
1590                 __fio_sem_init(&sk_out->lock, FIO_SEM_UNLOCKED);
1591                 __fio_sem_init(&sk_out->wait, FIO_SEM_LOCKED);
1592                 __fio_sem_init(&sk_out->xmit, FIO_SEM_UNLOCKED);
1593
1594 #ifdef WIN32
1595                 hProcess = windows_handle_connection(hjob, sk);
1596                 if (hProcess == INVALID_HANDLE_VALUE)
1597                         return -1;
1598                 sk_out->hProcess = hProcess;
1599                 fio_server_add_conn_pid(&conn_list, hProcess);
1600 #else
1601                 pid = fork();
1602                 if (pid) {
1603                         close(sk);
1604                         fio_server_add_conn_pid(&conn_list, pid);
1605                         continue;
1606                 }
1607
1608                 /* if error, it's already logged, non-fatal */
1609                 get_my_addr_str(sk);
1610
1611                 /*
1612                  * Assign sk_out here, it'll be dropped in handle_connection()
1613                  * since that function calls _exit() when done
1614                  */
1615                 sk_out_assign(sk_out);
1616                 handle_connection(sk_out);
1617 #endif
1618         }
1619
1620         return exitval;
1621 }
1622
1623 int fio_server_text_output(int level, const char *buf, size_t len)
1624 {
1625         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
1626         struct cmd_text_pdu *pdu;
1627         unsigned int tlen;
1628         struct timeval tv;
1629
1630         if (!sk_out || sk_out->sk == -1)
1631                 return -1;
1632
1633         tlen = sizeof(*pdu) + len;
1634         pdu = malloc(tlen);
1635
1636         pdu->level      = __cpu_to_le32(level);
1637         pdu->buf_len    = __cpu_to_le32(len);
1638
1639         gettimeofday(&tv, NULL);
1640         pdu->log_sec    = __cpu_to_le64(tv.tv_sec);
1641         pdu->log_usec   = __cpu_to_le64(tv.tv_usec);
1642
1643         memcpy(pdu->buf, buf, len);
1644
1645         fio_net_queue_cmd(FIO_NET_CMD_TEXT, pdu, tlen, NULL, SK_F_COPY);
1646         free(pdu);
1647         return len;
1648 }
1649
1650 static void convert_io_stat(struct io_stat *dst, struct io_stat *src)
1651 {
1652         dst->max_val    = cpu_to_le64(src->max_val);
1653         dst->min_val    = cpu_to_le64(src->min_val);
1654         dst->samples    = cpu_to_le64(src->samples);
1655
1656         /*
1657          * Encode to IEEE 754 for network transfer
1658          */
1659         dst->mean.u.i   = cpu_to_le64(fio_double_to_uint64(src->mean.u.f));
1660         dst->S.u.i      = cpu_to_le64(fio_double_to_uint64(src->S.u.f));
1661 }
1662
1663 static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
1664 {
1665         int i;
1666
1667         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1668                 dst->max_run[i]         = cpu_to_le64(src->max_run[i]);
1669                 dst->min_run[i]         = cpu_to_le64(src->min_run[i]);
1670                 dst->max_bw[i]          = cpu_to_le64(src->max_bw[i]);
1671                 dst->min_bw[i]          = cpu_to_le64(src->min_bw[i]);
1672                 dst->iobytes[i]         = cpu_to_le64(src->iobytes[i]);
1673                 dst->agg[i]             = cpu_to_le64(src->agg[i]);
1674         }
1675
1676         dst->kb_base    = cpu_to_le32(src->kb_base);
1677         dst->unit_base  = cpu_to_le32(src->unit_base);
1678         dst->groupid    = cpu_to_le32(src->groupid);
1679         dst->unified_rw_rep     = cpu_to_le32(src->unified_rw_rep);
1680         dst->sig_figs   = cpu_to_le32(src->sig_figs);
1681 }
1682
1683 /*
1684  * Send a CMD_TS, which packs struct thread_stat and group_run_stats
1685  * into a single payload.
1686  */
1687 void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
1688 {
1689         struct cmd_ts_pdu p;
1690         int i, j, k;
1691         size_t clat_prio_stats_extra_size = 0;
1692         size_t ss_extra_size = 0;
1693         size_t extended_buf_size = 0;
1694         void *extended_buf;
1695         void *extended_buf_wp;
1696
1697         dprint(FD_NET, "server sending end stats\n");
1698
1699         memset(&p, 0, sizeof(p));
1700
1701         snprintf(p.ts.name, sizeof(p.ts.name), "%s", ts->name);
1702         snprintf(p.ts.verror, sizeof(p.ts.verror), "%s", ts->verror);
1703         snprintf(p.ts.description, sizeof(p.ts.description), "%s",
1704                  ts->description);
1705
1706         p.ts.error              = cpu_to_le32(ts->error);
1707         p.ts.thread_number      = cpu_to_le32(ts->thread_number);
1708         p.ts.groupid            = cpu_to_le32(ts->groupid);
1709         p.ts.job_start          = cpu_to_le64(ts->job_start);
1710         p.ts.pid                = cpu_to_le32(ts->pid);
1711         p.ts.members            = cpu_to_le32(ts->members);
1712         p.ts.unified_rw_rep     = cpu_to_le32(ts->unified_rw_rep);
1713         p.ts.ioprio             = cpu_to_le32(ts->ioprio);
1714         p.ts.disable_prio_stat  = cpu_to_le32(ts->disable_prio_stat);
1715
1716         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1717                 convert_io_stat(&p.ts.clat_stat[i], &ts->clat_stat[i]);
1718                 convert_io_stat(&p.ts.slat_stat[i], &ts->slat_stat[i]);
1719                 convert_io_stat(&p.ts.lat_stat[i], &ts->lat_stat[i]);
1720                 convert_io_stat(&p.ts.bw_stat[i], &ts->bw_stat[i]);
1721                 convert_io_stat(&p.ts.iops_stat[i], &ts->iops_stat[i]);
1722         }
1723         convert_io_stat(&p.ts.sync_stat, &ts->sync_stat);
1724
1725         p.ts.usr_time           = cpu_to_le64(ts->usr_time);
1726         p.ts.sys_time           = cpu_to_le64(ts->sys_time);
1727         p.ts.ctx                = cpu_to_le64(ts->ctx);
1728         p.ts.minf               = cpu_to_le64(ts->minf);
1729         p.ts.majf               = cpu_to_le64(ts->majf);
1730         p.ts.clat_percentiles   = cpu_to_le32(ts->clat_percentiles);
1731         p.ts.lat_percentiles    = cpu_to_le32(ts->lat_percentiles);
1732         p.ts.slat_percentiles   = cpu_to_le32(ts->slat_percentiles);
1733         p.ts.percentile_precision = cpu_to_le64(ts->percentile_precision);
1734
1735         for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1736                 fio_fp64_t *src = &ts->percentile_list[i];
1737                 fio_fp64_t *dst = &p.ts.percentile_list[i];
1738
1739                 dst->u.i = cpu_to_le64(fio_double_to_uint64(src->u.f));
1740         }
1741
1742         for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1743                 p.ts.io_u_map[i]        = cpu_to_le64(ts->io_u_map[i]);
1744                 p.ts.io_u_submit[i]     = cpu_to_le64(ts->io_u_submit[i]);
1745                 p.ts.io_u_complete[i]   = cpu_to_le64(ts->io_u_complete[i]);
1746         }
1747
1748         for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
1749                 p.ts.io_u_lat_n[i]      = cpu_to_le64(ts->io_u_lat_n[i]);
1750         for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1751                 p.ts.io_u_lat_u[i]      = cpu_to_le64(ts->io_u_lat_u[i]);
1752         for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1753                 p.ts.io_u_lat_m[i]      = cpu_to_le64(ts->io_u_lat_m[i]);
1754
1755         for (i = 0; i < FIO_LAT_CNT; i++)
1756                 for (j = 0; j < DDIR_RWDIR_CNT; j++)
1757                         for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1758                                 p.ts.io_u_plat[i][j][k] = cpu_to_le64(ts->io_u_plat[i][j][k]);
1759
1760         for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1761                 p.ts.io_u_sync_plat[j] = cpu_to_le64(ts->io_u_sync_plat[j]);
1762
1763         for (i = 0; i < DDIR_RWDIR_SYNC_CNT; i++)
1764                 p.ts.total_io_u[i]      = cpu_to_le64(ts->total_io_u[i]);
1765
1766         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1767                 p.ts.short_io_u[i]      = cpu_to_le64(ts->short_io_u[i]);
1768                 p.ts.drop_io_u[i]       = cpu_to_le64(ts->drop_io_u[i]);
1769         }
1770
1771         p.ts.total_submit       = cpu_to_le64(ts->total_submit);
1772         p.ts.total_complete     = cpu_to_le64(ts->total_complete);
1773         p.ts.nr_zone_resets     = cpu_to_le64(ts->nr_zone_resets);
1774
1775         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1776                 p.ts.io_bytes[i]        = cpu_to_le64(ts->io_bytes[i]);
1777                 p.ts.runtime[i]         = cpu_to_le64(ts->runtime[i]);
1778         }
1779
1780         p.ts.total_run_time     = cpu_to_le64(ts->total_run_time);
1781         p.ts.continue_on_error  = cpu_to_le16(ts->continue_on_error);
1782         p.ts.total_err_count    = cpu_to_le64(ts->total_err_count);
1783         p.ts.first_error        = cpu_to_le32(ts->first_error);
1784         p.ts.kb_base            = cpu_to_le32(ts->kb_base);
1785         p.ts.unit_base          = cpu_to_le32(ts->unit_base);
1786
1787         p.ts.latency_depth      = cpu_to_le32(ts->latency_depth);
1788         p.ts.latency_target     = cpu_to_le64(ts->latency_target);
1789         p.ts.latency_window     = cpu_to_le64(ts->latency_window);
1790         p.ts.latency_percentile.u.i = cpu_to_le64(fio_double_to_uint64(ts->latency_percentile.u.f));
1791
1792         p.ts.sig_figs           = cpu_to_le32(ts->sig_figs);
1793
1794         p.ts.nr_block_infos     = cpu_to_le64(ts->nr_block_infos);
1795         for (i = 0; i < p.ts.nr_block_infos; i++)
1796                 p.ts.block_infos[i] = cpu_to_le32(ts->block_infos[i]);
1797
1798         p.ts.ss_dur             = cpu_to_le64(ts->ss_dur);
1799         p.ts.ss_state           = cpu_to_le32(ts->ss_state);
1800         p.ts.ss_head            = cpu_to_le32(ts->ss_head);
1801         p.ts.ss_limit.u.i       = cpu_to_le64(fio_double_to_uint64(ts->ss_limit.u.f));
1802         p.ts.ss_slope.u.i       = cpu_to_le64(fio_double_to_uint64(ts->ss_slope.u.f));
1803         p.ts.ss_deviation.u.i   = cpu_to_le64(fio_double_to_uint64(ts->ss_deviation.u.f));
1804         p.ts.ss_criterion.u.i   = cpu_to_le64(fio_double_to_uint64(ts->ss_criterion.u.f));
1805
1806         p.ts.cachehit           = cpu_to_le64(ts->cachehit);
1807         p.ts.cachemiss          = cpu_to_le64(ts->cachemiss);
1808
1809         convert_gs(&p.rs, rs);
1810
1811         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1812                 if (ts->nr_clat_prio[i])
1813                         clat_prio_stats_extra_size += ts->nr_clat_prio[i] * sizeof(*ts->clat_prio[i]);
1814         }
1815         extended_buf_size += clat_prio_stats_extra_size;
1816
1817         dprint(FD_NET, "ts->ss_state = %d\n", ts->ss_state);
1818         if (ts->ss_state & FIO_SS_DATA)
1819                 ss_extra_size = 2 * ts->ss_dur * sizeof(uint64_t);
1820
1821         extended_buf_size += ss_extra_size;
1822         if (!extended_buf_size) {
1823                 fio_net_queue_cmd(FIO_NET_CMD_TS, &p, sizeof(p), NULL, SK_F_COPY);
1824                 return;
1825         }
1826
1827         extended_buf_size += sizeof(p);
1828         extended_buf = calloc(1, extended_buf_size);
1829         if (!extended_buf) {
1830                 log_err("fio: failed to allocate FIO_NET_CMD_TS buffer\n");
1831                 return;
1832         }
1833
1834         memcpy(extended_buf, &p, sizeof(p));
1835         extended_buf_wp = (struct cmd_ts_pdu *)extended_buf + 1;
1836
1837         if (clat_prio_stats_extra_size) {
1838                 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1839                         struct clat_prio_stat *prio = (struct clat_prio_stat *) extended_buf_wp;
1840
1841                         for (j = 0; j < ts->nr_clat_prio[i]; j++) {
1842                                 for (k = 0; k < FIO_IO_U_PLAT_NR; k++)
1843                                         prio->io_u_plat[k] =
1844                                                 cpu_to_le64(ts->clat_prio[i][j].io_u_plat[k]);
1845                                 convert_io_stat(&prio->clat_stat,
1846                                                 &ts->clat_prio[i][j].clat_stat);
1847                                 prio->ioprio = cpu_to_le32(ts->clat_prio[i][j].ioprio);
1848                                 prio++;
1849                         }
1850
1851                         if (ts->nr_clat_prio[i]) {
1852                                 uint64_t offset = (char *)extended_buf_wp - (char *)extended_buf;
1853                                 struct cmd_ts_pdu *ptr = extended_buf;
1854
1855                                 ptr->ts.clat_prio_offset[i] = cpu_to_le64(offset);
1856                                 ptr->ts.nr_clat_prio[i] = cpu_to_le32(ts->nr_clat_prio[i]);
1857                         }
1858
1859                         extended_buf_wp = prio;
1860                 }
1861         }
1862
1863         if (ss_extra_size) {
1864                 uint64_t *ss_iops, *ss_bw;
1865                 uint64_t offset;
1866                 struct cmd_ts_pdu *ptr = extended_buf;
1867
1868                 dprint(FD_NET, "server sending steadystate ring buffers\n");
1869
1870                 /* ss iops */
1871                 ss_iops = (uint64_t *) extended_buf_wp;
1872                 for (i = 0; i < ts->ss_dur; i++)
1873                         ss_iops[i] = cpu_to_le64(ts->ss_iops_data[i]);
1874
1875                 offset = (char *)extended_buf_wp - (char *)extended_buf;
1876                 ptr->ts.ss_iops_data_offset = cpu_to_le64(offset);
1877                 extended_buf_wp = ss_iops + (int) ts->ss_dur;
1878
1879                 /* ss bw */
1880                 ss_bw = extended_buf_wp;
1881                 for (i = 0; i < ts->ss_dur; i++)
1882                         ss_bw[i] = cpu_to_le64(ts->ss_bw_data[i]);
1883
1884                 offset = (char *)extended_buf_wp - (char *)extended_buf;
1885                 ptr->ts.ss_bw_data_offset = cpu_to_le64(offset);
1886         }
1887
1888         fio_net_queue_cmd(FIO_NET_CMD_TS, extended_buf, extended_buf_size, NULL, SK_F_COPY);
1889         free(extended_buf);
1890 }
1891
1892 void fio_server_send_gs(struct group_run_stats *rs)
1893 {
1894         struct group_run_stats gs;
1895
1896         dprint(FD_NET, "server sending group run stats\n");
1897
1898         convert_gs(&gs, rs);
1899         fio_net_queue_cmd(FIO_NET_CMD_GS, &gs, sizeof(gs), NULL, SK_F_COPY);
1900 }
1901
1902 void fio_server_send_job_options(struct flist_head *opt_list,
1903                                  unsigned int gid)
1904 {
1905         struct cmd_job_option pdu;
1906         struct flist_head *entry;
1907
1908         if (flist_empty(opt_list))
1909                 return;
1910
1911         flist_for_each(entry, opt_list) {
1912                 struct print_option *p;
1913                 size_t len;
1914
1915                 p = flist_entry(entry, struct print_option, list);
1916                 memset(&pdu, 0, sizeof(pdu));
1917
1918                 if (gid == -1U) {
1919                         pdu.global = __cpu_to_le16(1);
1920                         pdu.groupid = 0;
1921                 } else {
1922                         pdu.global = 0;
1923                         pdu.groupid = cpu_to_le32(gid);
1924                 }
1925                 len = strlen(p->name);
1926                 if (len >= sizeof(pdu.name)) {
1927                         len = sizeof(pdu.name) - 1;
1928                         pdu.truncated = __cpu_to_le16(1);
1929                 }
1930                 memcpy(pdu.name, p->name, len);
1931                 if (p->value) {
1932                         len = strlen(p->value);
1933                         if (len >= sizeof(pdu.value)) {
1934                                 len = sizeof(pdu.value) - 1;
1935                                 pdu.truncated = __cpu_to_le16(1);
1936                         }
1937                         memcpy(pdu.value, p->value, len);
1938                 }
1939                 fio_net_queue_cmd(FIO_NET_CMD_JOB_OPT, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1940         }
1941 }
1942
1943 static void convert_agg(struct disk_util_agg *dst, struct disk_util_agg *src)
1944 {
1945         int i;
1946
1947         for (i = 0; i < 2; i++) {
1948                 dst->ios[i]     = cpu_to_le64(src->ios[i]);
1949                 dst->merges[i]  = cpu_to_le64(src->merges[i]);
1950                 dst->sectors[i] = cpu_to_le64(src->sectors[i]);
1951                 dst->ticks[i]   = cpu_to_le64(src->ticks[i]);
1952         }
1953
1954         dst->io_ticks           = cpu_to_le64(src->io_ticks);
1955         dst->time_in_queue      = cpu_to_le64(src->time_in_queue);
1956         dst->slavecount         = cpu_to_le32(src->slavecount);
1957         dst->max_util.u.i       = cpu_to_le64(fio_double_to_uint64(src->max_util.u.f));
1958 }
1959
1960 static void convert_dus(struct disk_util_stat *dst, struct disk_util_stat *src)
1961 {
1962         int i;
1963
1964         snprintf((char *) dst->name, sizeof(dst->name), "%s", src->name);
1965
1966         for (i = 0; i < 2; i++) {
1967                 dst->s.ios[i]           = cpu_to_le64(src->s.ios[i]);
1968                 dst->s.merges[i]        = cpu_to_le64(src->s.merges[i]);
1969                 dst->s.sectors[i]       = cpu_to_le64(src->s.sectors[i]);
1970                 dst->s.ticks[i]         = cpu_to_le64(src->s.ticks[i]);
1971         }
1972
1973         dst->s.io_ticks         = cpu_to_le64(src->s.io_ticks);
1974         dst->s.time_in_queue    = cpu_to_le64(src->s.time_in_queue);
1975         dst->s.msec             = cpu_to_le64(src->s.msec);
1976 }
1977
1978 void fio_server_send_du(void)
1979 {
1980         struct disk_util *du;
1981         struct flist_head *entry;
1982         struct cmd_du_pdu pdu;
1983
1984         dprint(FD_NET, "server: sending disk_util %d\n", !flist_empty(&disk_list));
1985
1986         memset(&pdu, 0, sizeof(pdu));
1987
1988         flist_for_each(entry, &disk_list) {
1989                 du = flist_entry(entry, struct disk_util, list);
1990
1991                 convert_dus(&pdu.dus, &du->dus);
1992                 convert_agg(&pdu.agg, &du->agg);
1993
1994                 fio_net_queue_cmd(FIO_NET_CMD_DU, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1995         }
1996 }
1997
1998 #ifdef CONFIG_ZLIB
1999
2000 static inline void __fio_net_prep_tail(z_stream *stream, void *out_pdu,
2001                                         struct sk_entry **last_entry,
2002                                         struct sk_entry *first)
2003 {
2004         unsigned int this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
2005
2006         *last_entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
2007                                  NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
2008         if (*last_entry)
2009                 flist_add_tail(&(*last_entry)->list, &first->next);
2010 }
2011
2012 /*
2013  * Deflates the next input given, creating as many new packets in the
2014  * linked list as necessary.
2015  */
2016 static int __deflate_pdu_buffer(void *next_in, unsigned int next_sz, void **out_pdu,
2017                                 struct sk_entry **last_entry, z_stream *stream,
2018                                 struct sk_entry *first)
2019 {
2020         int ret;
2021
2022         stream->next_in = next_in;
2023         stream->avail_in = next_sz;
2024         do {
2025                 if (!stream->avail_out) {
2026                         __fio_net_prep_tail(stream, *out_pdu, last_entry, first);
2027                         if (*last_entry == NULL)
2028                                 return 1;
2029
2030                         *out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2031
2032                         stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2033                         stream->next_out = *out_pdu;
2034                 }
2035
2036                 ret = deflate(stream, Z_BLOCK);
2037
2038                 if (ret < 0) {
2039                         free(*out_pdu);
2040                         return 1;
2041                 }
2042         } while (stream->avail_in);
2043
2044         return 0;
2045 }
2046
2047 static int __fio_append_iolog_gz_hist(struct sk_entry *first, struct io_log *log,
2048                                       struct io_logs *cur_log, z_stream *stream)
2049 {
2050         struct sk_entry *entry;
2051         void *out_pdu;
2052         int ret, i, j;
2053         int sample_sz = log_entry_sz(log);
2054
2055         out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2056         stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2057         stream->next_out = out_pdu;
2058
2059         for (i = 0; i < cur_log->nr_samples; i++) {
2060                 struct io_sample *s;
2061                 struct io_u_plat_entry *cur_plat_entry, *prev_plat_entry;
2062                 uint64_t *cur_plat, *prev_plat;
2063
2064                 s = get_sample(log, cur_log, i);
2065                 ret = __deflate_pdu_buffer(s, sample_sz, &out_pdu, &entry, stream, first);
2066                 if (ret)
2067                         return ret;
2068
2069                 /* Do the subtraction on server side so that client doesn't have to
2070                  * reconstruct our linked list from packets.
2071                  */
2072                 cur_plat_entry  = s->data.plat_entry;
2073                 prev_plat_entry = flist_first_entry(&cur_plat_entry->list, struct io_u_plat_entry, list);
2074                 cur_plat  = cur_plat_entry->io_u_plat;
2075                 prev_plat = prev_plat_entry->io_u_plat;
2076
2077                 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
2078                         cur_plat[j] -= prev_plat[j];
2079                 }
2080
2081                 flist_del(&prev_plat_entry->list);
2082                 free(prev_plat_entry);
2083
2084                 ret = __deflate_pdu_buffer(cur_plat_entry, sizeof(*cur_plat_entry),
2085                                            &out_pdu, &entry, stream, first);
2086
2087                 if (ret)
2088                         return ret;
2089         }
2090
2091         __fio_net_prep_tail(stream, out_pdu, &entry, first);
2092         return entry == NULL;
2093 }
2094
2095 static int __fio_append_iolog_gz(struct sk_entry *first, struct io_log *log,
2096                                  struct io_logs *cur_log, z_stream *stream)
2097 {
2098         unsigned int this_len;
2099         void *out_pdu;
2100         int ret;
2101
2102         if (log->log_type == IO_LOG_TYPE_HIST)
2103                 return __fio_append_iolog_gz_hist(first, log, cur_log, stream);
2104
2105         stream->next_in = (void *) cur_log->log;
2106         stream->avail_in = cur_log->nr_samples * log_entry_sz(log);
2107
2108         do {
2109                 struct sk_entry *entry;
2110
2111                 /*
2112                  * Dirty - since the log is potentially huge, compress it into
2113                  * FIO_SERVER_MAX_FRAGMENT_PDU chunks and let the receiving
2114                  * side defragment it.
2115                  */
2116                 out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2117
2118                 stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2119                 stream->next_out = out_pdu;
2120                 ret = deflate(stream, Z_BLOCK);
2121                 /* may be Z_OK, or Z_STREAM_END */
2122                 if (ret < 0) {
2123                         free(out_pdu);
2124                         return 1;
2125                 }
2126
2127                 this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
2128
2129                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
2130                                          NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
2131                 if (!entry) {
2132                         free(out_pdu);
2133                         return 1;
2134                 }
2135                 flist_add_tail(&entry->list, &first->next);
2136         } while (stream->avail_in);
2137
2138         return 0;
2139 }
2140
2141 static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
2142 {
2143         z_stream stream = {
2144                 .zalloc = Z_NULL,
2145                 .zfree  = Z_NULL,
2146                 .opaque = Z_NULL,
2147         };
2148         int ret = 0;
2149
2150         if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK)
2151                 return 1;
2152
2153         while (!flist_empty(&log->io_logs)) {
2154                 struct io_logs *cur_log;
2155
2156                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
2157                 flist_del_init(&cur_log->list);
2158
2159                 ret = __fio_append_iolog_gz(first, log, cur_log, &stream);
2160                 if (ret)
2161                         break;
2162         }
2163
2164         ret = deflate(&stream, Z_FINISH);
2165
2166         while (ret != Z_STREAM_END) {
2167                 struct sk_entry *entry;
2168                 unsigned int this_len;
2169                 void *out_pdu;
2170
2171                 out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
2172                 stream.avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
2173                 stream.next_out = out_pdu;
2174
2175                 ret = deflate(&stream, Z_FINISH);
2176                 /* may be Z_OK, or Z_STREAM_END */
2177                 if (ret < 0) {
2178                         free(out_pdu);
2179                         break;
2180                 }
2181
2182                 this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream.avail_out;
2183
2184                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
2185                                          NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
2186                 if (!entry) {
2187                         free(out_pdu);
2188                         break;
2189                 }
2190                 flist_add_tail(&entry->list, &first->next);
2191         }
2192
2193         ret = deflateEnd(&stream);
2194         if (ret == Z_OK)
2195                 return 0;
2196
2197         return 1;
2198 }
2199 #else
2200 static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
2201 {
2202         return 1;
2203 }
2204 #endif
2205
2206 static int fio_append_gz_chunks(struct sk_entry *first, struct io_log *log)
2207 {
2208         struct sk_entry *entry;
2209         struct flist_head *node;
2210         int ret = 0;
2211
2212         pthread_mutex_lock(&log->chunk_lock);
2213         flist_for_each(node, &log->chunk_list) {
2214                 struct iolog_compress *c;
2215
2216                 c = flist_entry(node, struct iolog_compress, list);
2217                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, c->buf, c->len,
2218                                                 NULL, SK_F_VEC | SK_F_INLINE);
2219                 if (!entry) {
2220                         ret = 1;
2221                         break;
2222                 }
2223                 flist_add_tail(&entry->list, &first->next);
2224         }
2225         pthread_mutex_unlock(&log->chunk_lock);
2226         return ret;
2227 }
2228
2229 static int fio_append_text_log(struct sk_entry *first, struct io_log *log)
2230 {
2231         struct sk_entry *entry;
2232         int ret = 0;
2233
2234         while (!flist_empty(&log->io_logs)) {
2235                 struct io_logs *cur_log;
2236                 size_t size;
2237
2238                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
2239                 flist_del_init(&cur_log->list);
2240
2241                 size = cur_log->nr_samples * log_entry_sz(log);
2242
2243                 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, cur_log->log, size,
2244                                                 NULL, SK_F_VEC | SK_F_INLINE);
2245                 if (!entry) {
2246                         ret = 1;
2247                         break;
2248                 }
2249                 flist_add_tail(&entry->list, &first->next);
2250         }
2251
2252         return ret;
2253 }
2254
2255 int fio_send_iolog(struct thread_data *td, struct io_log *log, const char *name)
2256 {
2257         struct cmd_iolog_pdu pdu = {
2258                 .nr_samples             = cpu_to_le64(iolog_nr_samples(log)),
2259                 .thread_number          = cpu_to_le32(td->thread_number),
2260                 .log_type               = cpu_to_le32(log->log_type),
2261                 .log_hist_coarseness    = cpu_to_le32(log->hist_coarseness),
2262                 .per_job_logs           = cpu_to_le32(td->o.per_job_logs),
2263         };
2264         struct sk_entry *first;
2265         struct flist_head *entry;
2266         int ret = 0;
2267
2268         if (!flist_empty(&log->chunk_list))
2269                 pdu.compressed = __cpu_to_le32(STORE_COMPRESSED);
2270         else if (use_zlib)
2271                 pdu.compressed = __cpu_to_le32(XMIT_COMPRESSED);
2272         else
2273                 pdu.compressed = 0;
2274
2275         snprintf((char *) pdu.name, sizeof(pdu.name), "%s", name);
2276
2277         /*
2278          * We can't do this for a pre-compressed log, but for that case,
2279          * log->nr_samples is zero anyway.
2280          */
2281         flist_for_each(entry, &log->io_logs) {
2282                 struct io_logs *cur_log;
2283                 int i;
2284
2285                 cur_log = flist_entry(entry, struct io_logs, list);
2286
2287                 for (i = 0; i < cur_log->nr_samples; i++) {
2288                         struct io_sample *s = get_sample(log, cur_log, i);
2289
2290                         s->time         = cpu_to_le64(s->time);
2291                         if (log->log_type != IO_LOG_TYPE_HIST) {
2292                                 s->data.val.val0        = cpu_to_le64(s->data.val.val0);
2293                                 s->data.val.val1        = cpu_to_le64(s->data.val.val1);
2294                         }
2295                         s->__ddir       = __cpu_to_le32(s->__ddir);
2296                         s->bs           = cpu_to_le64(s->bs);
2297
2298                         if (log->log_offset) {
2299                                 struct io_sample_offset *so = (void *) s;
2300
2301                                 so->offset = cpu_to_le64(so->offset);
2302                         }
2303                 }
2304         }
2305
2306         /*
2307          * Assemble header entry first
2308          */
2309         first = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, &pdu, sizeof(pdu), NULL, SK_F_VEC | SK_F_INLINE | SK_F_COPY);
2310         if (!first)
2311                 return 1;
2312
2313         /*
2314          * Now append actual log entries. If log compression was enabled on
2315          * the job, just send out the compressed chunks directly. If we
2316          * have a plain log, compress if we can, then send. Otherwise, send
2317          * the plain text output.
2318          */
2319         if (!flist_empty(&log->chunk_list))
2320                 ret = fio_append_gz_chunks(first, log);
2321         else if (use_zlib)
2322                 ret = fio_append_iolog_gz(first, log);
2323         else
2324                 ret = fio_append_text_log(first, log);
2325
2326         fio_net_queue_entry(first);
2327         return ret;
2328 }
2329
2330 void fio_server_send_add_job(struct thread_data *td)
2331 {
2332         struct cmd_add_job_pdu *pdu;
2333         size_t cmd_sz = offsetof(struct cmd_add_job_pdu, top) +
2334                 thread_options_pack_size(&td->o);
2335
2336         pdu = malloc(cmd_sz);
2337         pdu->thread_number = cpu_to_le32(td->thread_number);
2338         pdu->groupid = cpu_to_le32(td->groupid);
2339
2340         convert_thread_options_to_net(&pdu->top, &td->o);
2341
2342         fio_net_queue_cmd(FIO_NET_CMD_ADD_JOB, pdu, cmd_sz, NULL, SK_F_COPY);
2343         free(pdu);
2344 }
2345
2346 void fio_server_send_start(struct thread_data *td)
2347 {
2348         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2349
2350         if (sk_out->sk == -1) {
2351                 log_err("pthread getting specific for key failed, sk_out %p, sk %i, err: %i:%s",
2352                         sk_out, sk_out->sk, errno, strerror(errno));
2353                 abort();
2354         }
2355
2356         fio_net_queue_cmd(FIO_NET_CMD_SERVER_START, NULL, 0, NULL, SK_F_SIMPLE);
2357 }
2358
2359 int fio_server_get_verify_state(const char *name, int threadnumber,
2360                                 void **datap)
2361 {
2362         struct thread_io_list *s;
2363         struct cmd_sendfile out;
2364         struct cmd_reply *rep;
2365         uint64_t tag;
2366         void *data;
2367         int ret;
2368
2369         dprint(FD_NET, "server: request verify state\n");
2370
2371         rep = smalloc(sizeof(*rep));
2372         if (!rep)
2373                 return ENOMEM;
2374
2375         __fio_sem_init(&rep->lock, FIO_SEM_LOCKED);
2376         rep->data = NULL;
2377         rep->error = 0;
2378
2379         verify_state_gen_name((char *) out.path, sizeof(out.path), name, me,
2380                                 threadnumber);
2381         tag = (uint64_t) (uintptr_t) rep;
2382         fio_net_queue_cmd(FIO_NET_CMD_SENDFILE, &out, sizeof(out), &tag,
2383                                 SK_F_COPY);
2384
2385         /*
2386          * Wait for the backend to receive the reply
2387          */
2388         if (fio_sem_down_timeout(&rep->lock, 10000)) {
2389                 log_err("fio: timed out waiting for reply\n");
2390                 ret = ETIMEDOUT;
2391                 goto fail;
2392         }
2393
2394         if (rep->error) {
2395                 log_err("fio: failure on receiving state file %s: %s\n",
2396                                 out.path, strerror(rep->error));
2397                 ret = rep->error;
2398 fail:
2399                 *datap = NULL;
2400                 sfree(rep);
2401                 fio_net_queue_quit();
2402                 return ret;
2403         }
2404
2405         /*
2406          * The format is verify_state_hdr, then thread_io_list. Verify
2407          * the header, and the thread_io_list checksum
2408          */
2409         s = rep->data + sizeof(struct verify_state_hdr);
2410         if (verify_state_hdr(rep->data, s)) {
2411                 ret = EILSEQ;
2412                 goto fail;
2413         }
2414
2415         /*
2416          * Don't need the header from now, copy just the thread_io_list
2417          */
2418         ret = 0;
2419         rep->size -= sizeof(struct verify_state_hdr);
2420         data = malloc(rep->size);
2421         memcpy(data, s, rep->size);
2422         *datap = data;
2423
2424         sfree(rep->data);
2425         __fio_sem_remove(&rep->lock);
2426         sfree(rep);
2427         return ret;
2428 }
2429
2430 static int fio_init_server_ip(void)
2431 {
2432         struct sockaddr *addr;
2433         socklen_t socklen;
2434         char buf[80];
2435         const char *str;
2436         int sk, opt;
2437
2438         if (use_ipv6)
2439                 sk = socket(AF_INET6, SOCK_STREAM, 0);
2440         else
2441                 sk = socket(AF_INET, SOCK_STREAM, 0);
2442
2443         if (sk < 0) {
2444                 log_err("fio: socket: %s\n", strerror(errno));
2445                 return -1;
2446         }
2447
2448         opt = 1;
2449         if (setsockopt(sk, SOL_SOCKET, SO_REUSEADDR, (void *)&opt, sizeof(opt)) < 0) {
2450                 log_err("fio: setsockopt(REUSEADDR): %s\n", strerror(errno));
2451                 close(sk);
2452                 return -1;
2453         }
2454 #ifdef SO_REUSEPORT
2455         /*
2456          * Not fatal if fails, so just ignore it if that happens
2457          */
2458         if (setsockopt(sk, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt))) {
2459         }
2460 #endif
2461
2462         if (use_ipv6) {
2463                 void *src = &saddr_in6.sin6_addr;
2464
2465                 addr = (struct sockaddr *) &saddr_in6;
2466                 socklen = sizeof(saddr_in6);
2467                 saddr_in6.sin6_family = AF_INET6;
2468                 str = inet_ntop(AF_INET6, src, buf, sizeof(buf));
2469         } else {
2470                 void *src = &saddr_in.sin_addr;
2471
2472                 addr = (struct sockaddr *) &saddr_in;
2473                 socklen = sizeof(saddr_in);
2474                 saddr_in.sin_family = AF_INET;
2475                 str = inet_ntop(AF_INET, src, buf, sizeof(buf));
2476         }
2477
2478         if (bind(sk, addr, socklen) < 0) {
2479                 log_err("fio: bind: %s\n", strerror(errno));
2480                 log_info("fio: failed with IPv%c %s\n", use_ipv6 ? '6' : '4', str);
2481                 close(sk);
2482                 return -1;
2483         }
2484
2485         return sk;
2486 }
2487
2488 static int fio_init_server_sock(void)
2489 {
2490         struct sockaddr_un addr;
2491         socklen_t len;
2492         mode_t mode;
2493         int sk;
2494
2495         sk = socket(AF_UNIX, SOCK_STREAM, 0);
2496         if (sk < 0) {
2497                 log_err("fio: socket: %s\n", strerror(errno));
2498                 return -1;
2499         }
2500
2501         mode = umask(000);
2502
2503         addr.sun_family = AF_UNIX;
2504         snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", bind_sock);
2505
2506         len = sizeof(addr.sun_family) + strlen(bind_sock) + 1;
2507
2508         if (bind(sk, (struct sockaddr *) &addr, len) < 0) {
2509                 log_err("fio: bind: %s\n", strerror(errno));
2510                 close(sk);
2511                 return -1;
2512         }
2513
2514         umask(mode);
2515         return sk;
2516 }
2517
2518 static int fio_init_server_connection(void)
2519 {
2520         char bind_str[128];
2521         int sk;
2522
2523         dprint(FD_NET, "starting server\n");
2524
2525         if (!bind_sock)
2526                 sk = fio_init_server_ip();
2527         else
2528                 sk = fio_init_server_sock();
2529
2530         if (sk < 0)
2531                 return sk;
2532
2533         memset(bind_str, 0, sizeof(bind_str));
2534
2535         if (!bind_sock) {
2536                 char *p, port[16];
2537                 void *src;
2538                 int af;
2539
2540                 if (use_ipv6) {
2541                         af = AF_INET6;
2542                         src = &saddr_in6.sin6_addr;
2543                 } else {
2544                         af = AF_INET;
2545                         src = &saddr_in.sin_addr;
2546                 }
2547
2548                 p = (char *) inet_ntop(af, src, bind_str, sizeof(bind_str));
2549
2550                 sprintf(port, ",%u", fio_net_port);
2551                 if (p)
2552                         strcat(p, port);
2553                 else
2554                         snprintf(bind_str, sizeof(bind_str), "%s", port);
2555         } else
2556                 snprintf(bind_str, sizeof(bind_str), "%s", bind_sock);
2557
2558         log_info("fio: server listening on %s\n", bind_str);
2559
2560         if (listen(sk, 4) < 0) {
2561                 log_err("fio: listen: %s\n", strerror(errno));
2562                 close(sk);
2563                 return -1;
2564         }
2565
2566         return sk;
2567 }
2568
2569 int fio_server_parse_host(const char *host, int ipv6, struct in_addr *inp,
2570                           struct in6_addr *inp6)
2571
2572 {
2573         int ret = 0;
2574
2575         if (ipv6)
2576                 ret = inet_pton(AF_INET6, host, inp6);
2577         else
2578                 ret = inet_pton(AF_INET, host, inp);
2579
2580         if (ret != 1) {
2581                 struct addrinfo *res, hints = {
2582                         .ai_family = ipv6 ? AF_INET6 : AF_INET,
2583                         .ai_socktype = SOCK_STREAM,
2584                 };
2585
2586                 ret = getaddrinfo(host, NULL, &hints, &res);
2587                 if (ret) {
2588                         log_err("fio: failed to resolve <%s> (%s)\n", host,
2589                                         gai_strerror(ret));
2590                         return 1;
2591                 }
2592
2593                 if (ipv6)
2594                         memcpy(inp6, &((struct sockaddr_in6 *) res->ai_addr)->sin6_addr, sizeof(*inp6));
2595                 else
2596                         memcpy(inp, &((struct sockaddr_in *) res->ai_addr)->sin_addr, sizeof(*inp));
2597
2598                 ret = 1;
2599                 freeaddrinfo(res);
2600         }
2601
2602         return !(ret == 1);
2603 }
2604
2605 /*
2606  * Parse a host/ip/port string. Reads from 'str'.
2607  *
2608  * Outputs:
2609  *
2610  * For IPv4:
2611  *      *ptr is the host, *port is the port, inp is the destination.
2612  * For IPv6:
2613  *      *ptr is the host, *port is the port, inp6 is the dest, and *ipv6 is 1.
2614  * For local domain sockets:
2615  *      *ptr is the filename, *is_sock is 1.
2616  */
2617 int fio_server_parse_string(const char *str, char **ptr, bool *is_sock,
2618                             int *port, struct in_addr *inp,
2619                             struct in6_addr *inp6, int *ipv6)
2620 {
2621         const char *host = str;
2622         char *portp;
2623         int lport = 0;
2624
2625         *ptr = NULL;
2626         *is_sock = false;
2627         *port = fio_net_port;
2628         *ipv6 = 0;
2629
2630         if (!strncmp(str, "sock:", 5)) {
2631                 *ptr = strdup(str + 5);
2632                 *is_sock = true;
2633
2634                 return 0;
2635         }
2636
2637         /*
2638          * Is it ip:<ip or host>:port
2639          */
2640         if (!strncmp(host, "ip:", 3))
2641                 host += 3;
2642         else if (!strncmp(host, "ip4:", 4))
2643                 host += 4;
2644         else if (!strncmp(host, "ip6:", 4)) {
2645                 host += 4;
2646                 *ipv6 = 1;
2647         } else if (host[0] == ':') {
2648                 /* String is :port */
2649                 host++;
2650                 lport = atoi(host);
2651                 if (!lport || lport > 65535) {
2652                         log_err("fio: bad server port %u\n", lport);
2653                         return 1;
2654                 }
2655                 /* no hostname given, we are done */
2656                 *port = lport;
2657                 return 0;
2658         }
2659
2660         /*
2661          * If no port seen yet, check if there's a last ',' at the end
2662          */
2663         if (!lport) {
2664                 portp = strchr(host, ',');
2665                 if (portp) {
2666                         *portp = '\0';
2667                         portp++;
2668                         lport = atoi(portp);
2669                         if (!lport || lport > 65535) {
2670                                 log_err("fio: bad server port %u\n", lport);
2671                                 return 1;
2672                         }
2673                 }
2674         }
2675
2676         if (lport)
2677                 *port = lport;
2678
2679         if (!strlen(host))
2680                 return 0;
2681
2682         *ptr = strdup(host);
2683
2684         if (fio_server_parse_host(*ptr, *ipv6, inp, inp6)) {
2685                 free(*ptr);
2686                 *ptr = NULL;
2687                 return 1;
2688         }
2689
2690         if (*port == 0)
2691                 *port = fio_net_port;
2692
2693         return 0;
2694 }
2695
2696 /*
2697  * Server arg should be one of:
2698  *
2699  * sock:/path/to/socket
2700  *   ip:1.2.3.4
2701  *      1.2.3.4
2702  *
2703  * Where sock uses unix domain sockets, and ip binds the server to
2704  * a specific interface. If no arguments are given to the server, it
2705  * uses IP and binds to 0.0.0.0.
2706  *
2707  */
2708 static int fio_handle_server_arg(void)
2709 {
2710         int port = fio_net_port;
2711         bool is_sock;
2712         int ret = 0;
2713
2714         saddr_in.sin_addr.s_addr = htonl(INADDR_ANY);
2715
2716         if (!fio_server_arg)
2717                 goto out;
2718
2719         ret = fio_server_parse_string(fio_server_arg, &bind_sock, &is_sock,
2720                                         &port, &saddr_in.sin_addr,
2721                                         &saddr_in6.sin6_addr, &use_ipv6);
2722
2723         if (!is_sock && bind_sock) {
2724                 free(bind_sock);
2725                 bind_sock = NULL;
2726         }
2727
2728 out:
2729         fio_net_port = port;
2730         saddr_in.sin_port = htons(port);
2731         saddr_in6.sin6_port = htons(port);
2732         return ret;
2733 }
2734
2735 static void sig_int(int sig)
2736 {
2737         if (bind_sock)
2738                 unlink(bind_sock);
2739 }
2740
2741 static void set_sig_handlers(void)
2742 {
2743         struct sigaction act = {
2744                 .sa_handler = sig_int,
2745                 .sa_flags = SA_RESTART,
2746         };
2747
2748         sigaction(SIGINT, &act, NULL);
2749
2750         /* Windows uses SIGBREAK as a quit signal from other applications */
2751 #ifdef WIN32
2752         sigaction(SIGBREAK, &act, NULL);
2753 #endif
2754 }
2755
2756 void fio_server_destroy_sk_key(void)
2757 {
2758         pthread_key_delete(sk_out_key);
2759 }
2760
2761 int fio_server_create_sk_key(void)
2762 {
2763         if (pthread_key_create(&sk_out_key, NULL)) {
2764                 log_err("fio: can't create sk_out backend key\n");
2765                 return 1;
2766         }
2767
2768         pthread_setspecific(sk_out_key, NULL);
2769         return 0;
2770 }
2771
2772 static int fio_server(void)
2773 {
2774         int sk, ret;
2775
2776         dprint(FD_NET, "starting server\n");
2777
2778         if (fio_handle_server_arg())
2779                 return -1;
2780
2781         set_sig_handlers();
2782
2783 #ifdef WIN32
2784         /* if this is a child process, go handle the connection */
2785         if (fio_server_pipe_name != NULL) {
2786                 ret = handle_connection_process();
2787                 return ret;
2788         }
2789
2790         /* job to link child processes so they terminate together */
2791         hjob = windows_create_job();
2792         if (hjob == INVALID_HANDLE_VALUE)
2793                 return -1;
2794 #endif
2795
2796         sk = fio_init_server_connection();
2797         if (sk < 0)
2798                 return -1;
2799
2800         ret = accept_loop(sk);
2801
2802         close(sk);
2803
2804         if (fio_server_arg) {
2805                 free(fio_server_arg);
2806                 fio_server_arg = NULL;
2807         }
2808         if (bind_sock)
2809                 free(bind_sock);
2810
2811         return ret;
2812 }
2813
2814 void fio_server_got_signal(int signal)
2815 {
2816         struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2817
2818         assert(sk_out);
2819
2820         if (signal == SIGPIPE)
2821                 sk_out->sk = -1;
2822         else {
2823                 log_info("\nfio: terminating on signal %d\n", signal);
2824                 exit_backend = true;
2825         }
2826 }
2827
2828 static int check_existing_pidfile(const char *pidfile)
2829 {
2830         struct stat sb;
2831         char buf[16];
2832         pid_t pid;
2833         FILE *f;
2834
2835         if (stat(pidfile, &sb))
2836                 return 0;
2837
2838         f = fopen(pidfile, "r");
2839         if (!f)
2840                 return 0;
2841
2842         if (fread(buf, sb.st_size, 1, f) <= 0) {
2843                 fclose(f);
2844                 return 1;
2845         }
2846         fclose(f);
2847
2848         pid = atoi(buf);
2849         if (kill(pid, SIGCONT) < 0)
2850                 return errno != ESRCH;
2851
2852         return 1;
2853 }
2854
2855 static int write_pid(pid_t pid, const char *pidfile)
2856 {
2857         FILE *fpid;
2858
2859         fpid = fopen(pidfile, "w");
2860         if (!fpid) {
2861                 log_err("fio: failed opening pid file %s\n", pidfile);
2862                 return 1;
2863         }
2864
2865         fprintf(fpid, "%u\n", (unsigned int) pid);
2866         fclose(fpid);
2867         return 0;
2868 }
2869
2870 /*
2871  * If pidfile is specified, background us.
2872  */
2873 int fio_start_server(char *pidfile)
2874 {
2875         FILE *file;
2876         pid_t pid;
2877         int ret;
2878
2879 #if defined(WIN32)
2880         WSADATA wsd;
2881         WSAStartup(MAKEWORD(2, 2), &wsd);
2882 #endif
2883
2884         if (!pidfile)
2885                 return fio_server();
2886
2887         if (check_existing_pidfile(pidfile)) {
2888                 log_err("fio: pidfile %s exists and server appears alive\n",
2889                                                                 pidfile);
2890                 free(pidfile);
2891                 return -1;
2892         }
2893
2894         pid = fork();
2895         if (pid < 0) {
2896                 log_err("fio: failed server fork: %s\n", strerror(errno));
2897                 free(pidfile);
2898                 return -1;
2899         } else if (pid) {
2900                 ret = write_pid(pid, pidfile);
2901                 free(pidfile);
2902                 _exit(ret);
2903         }
2904
2905         setsid();
2906         openlog("fio", LOG_NDELAY|LOG_NOWAIT|LOG_PID, LOG_USER);
2907         log_syslog = true;
2908
2909         file = freopen("/dev/null", "r", stdin);
2910         if (!file)
2911                 perror("freopen");
2912
2913         file = freopen("/dev/null", "w", stdout);
2914         if (!file)
2915                 perror("freopen");
2916
2917         file = freopen("/dev/null", "w", stderr);
2918         if (!file)
2919                 perror("freopen");
2920
2921         f_out = NULL;
2922         f_err = NULL;
2923
2924         ret = fio_server();
2925
2926         fclose(stdin);
2927         fclose(stdout);
2928         fclose(stderr);
2929
2930         closelog();
2931         unlink(pidfile);
2932         free(pidfile);
2933         return ret;
2934 }
2935
2936 void fio_server_set_arg(const char *arg)
2937 {
2938         fio_server_arg = strdup(arg);
2939 }
2940
2941 #ifdef WIN32
2942 void fio_server_internal_set(const char *arg)
2943 {
2944         fio_server_pipe_name = strdup(arg);
2945 }
2946 #endif