8 #include <sys/socket.h>
12 #include <netinet/in.h>
13 #include <arpa/inet.h>
24 #include "crc/crc16.h"
25 #include "lib/ieee754.h"
26 #include "verify-state.h"
29 int fio_net_port = FIO_NET_PORT;
42 struct flist_head list; /* link on sk_out->list */
43 int flags; /* SK_F_* */
44 int opcode; /* Actual command fields */
48 struct flist_head next; /* Other sk_entry's, if linked command */
51 static char *fio_server_arg;
52 static char *bind_sock;
53 static struct sockaddr_in saddr_in;
54 static struct sockaddr_in6 saddr_in6;
57 static unsigned int has_zlib = 1;
59 static unsigned int has_zlib = 0;
61 static unsigned int use_zlib;
64 static pthread_key_t sk_out_key;
66 struct fio_fork_item {
67 struct flist_head list;
81 static const char *fio_server_ops[FIO_NET_CMD_NR] = {
107 static void sk_lock(struct sk_out *sk_out)
109 fio_sem_down(&sk_out->lock);
112 static void sk_unlock(struct sk_out *sk_out)
114 fio_sem_up(&sk_out->lock);
117 void sk_out_assign(struct sk_out *sk_out)
125 pthread_setspecific(sk_out_key, sk_out);
128 static void sk_out_free(struct sk_out *sk_out)
130 __fio_sem_remove(&sk_out->lock);
131 __fio_sem_remove(&sk_out->wait);
132 __fio_sem_remove(&sk_out->xmit);
136 static int __sk_out_drop(struct sk_out *sk_out)
142 assert(sk_out->refs != 0);
143 refs = --sk_out->refs;
148 pthread_setspecific(sk_out_key, NULL);
156 void sk_out_drop(void)
158 struct sk_out *sk_out;
160 sk_out = pthread_getspecific(sk_out_key);
161 __sk_out_drop(sk_out);
164 static void __fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
165 uint32_t pdu_len, uint64_t tag)
167 memset(cmd, 0, sizeof(*cmd));
169 cmd->version = __cpu_to_le16(FIO_SERVER_VER);
170 cmd->opcode = cpu_to_le16(opcode);
171 cmd->tag = cpu_to_le64(tag);
172 cmd->pdu_len = cpu_to_le32(pdu_len);
176 static void fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
177 const void *pdu, uint32_t pdu_len, uint64_t tag)
179 __fio_init_net_cmd(cmd, opcode, pdu_len, tag);
182 memcpy(&cmd->payload, pdu, pdu_len);
185 const char *fio_server_op(unsigned int op)
189 if (op < FIO_NET_CMD_NR)
190 return fio_server_ops[op];
192 sprintf(buf, "UNKNOWN/%d", op);
196 static ssize_t iov_total_len(const struct iovec *iov, int count)
208 static int fio_sendv_data(int sk, struct iovec *iov, int count)
210 ssize_t total_len = iov_total_len(iov, count);
214 ret = writev(sk, iov, count);
221 if (ret >= iov->iov_len) {
226 iov->iov_base += ret;
232 else if (errno == EAGAIN || errno == EINTR)
236 } while (!exit_backend);
244 static int fio_send_data(int sk, const void *p, unsigned int len)
246 struct iovec iov = { .iov_base = (void *) p, .iov_len = len };
248 assert(len <= sizeof(struct fio_net_cmd) + FIO_SERVER_MAX_FRAGMENT_PDU);
250 return fio_sendv_data(sk, &iov, 1);
253 static int fio_recv_data(int sk, void *buf, unsigned int len, bool wait)
261 flags = OS_MSG_DONTWAIT;
264 int ret = recv(sk, p, len, flags);
274 else if (errno == EAGAIN || errno == EINTR) {
280 } while (!exit_backend);
288 static int verify_convert_cmd(struct fio_net_cmd *cmd)
292 cmd->cmd_crc16 = le16_to_cpu(cmd->cmd_crc16);
293 cmd->pdu_crc16 = le16_to_cpu(cmd->pdu_crc16);
295 crc = fio_crc16(cmd, FIO_NET_CMD_CRC_SZ);
296 if (crc != cmd->cmd_crc16) {
297 log_err("fio: server bad crc on command (got %x, wanted %x)\n",
298 cmd->cmd_crc16, crc);
299 fprintf(f_err, "fio: server bad crc on command (got %x, wanted %x)\n",
300 cmd->cmd_crc16, crc);
304 cmd->version = le16_to_cpu(cmd->version);
305 cmd->opcode = le16_to_cpu(cmd->opcode);
306 cmd->flags = le32_to_cpu(cmd->flags);
307 cmd->tag = le64_to_cpu(cmd->tag);
308 cmd->pdu_len = le32_to_cpu(cmd->pdu_len);
310 switch (cmd->version) {
314 log_err("fio: bad server cmd version %d\n", cmd->version);
315 fprintf(f_err, "fio: client/server version mismatch (%d != %d)\n",
316 cmd->version, FIO_SERVER_VER);
320 if (cmd->pdu_len > FIO_SERVER_MAX_FRAGMENT_PDU) {
321 log_err("fio: command payload too large: %u\n", cmd->pdu_len);
329 * Read (and defragment, if necessary) incoming commands
331 struct fio_net_cmd *fio_net_recv_cmd(int sk, bool wait)
333 struct fio_net_cmd cmd, *tmp, *cmdret = NULL;
334 size_t cmd_size = 0, pdu_offset = 0;
340 ret = fio_recv_data(sk, &cmd, sizeof(cmd), wait);
344 /* We have a command, verify it and swap if need be */
345 ret = verify_convert_cmd(&cmd);
350 /* if this is text, add room for \0 at the end */
351 cmd_size = sizeof(cmd) + cmd.pdu_len + 1;
354 cmd_size += cmd.pdu_len;
356 if (cmd_size / 1024 > FIO_SERVER_MAX_CMD_MB * 1024) {
357 log_err("fio: cmd+pdu too large (%llu)\n", (unsigned long long) cmd_size);
362 tmp = realloc(cmdret, cmd_size);
364 log_err("fio: server failed allocating cmd\n");
371 memcpy(cmdret, &cmd, sizeof(cmd));
372 else if (cmdret->opcode != cmd.opcode) {
373 log_err("fio: fragment opcode mismatch (%d != %d)\n",
374 cmdret->opcode, cmd.opcode);
382 /* There's payload, get it */
383 pdu = (char *) cmdret->payload + pdu_offset;
384 ret = fio_recv_data(sk, pdu, cmd.pdu_len, wait);
388 /* Verify payload crc */
389 crc = fio_crc16(pdu, cmd.pdu_len);
390 if (crc != cmd.pdu_crc16) {
391 log_err("fio: server bad crc on payload ");
392 log_err("(got %x, wanted %x)\n", cmd.pdu_crc16, crc);
397 pdu_offset += cmd.pdu_len;
399 cmdret->pdu_len += cmd.pdu_len;
401 } while (cmd.flags & FIO_NET_CMD_F_MORE);
407 /* zero-terminate text input */
408 if (cmdret->pdu_len) {
409 if (cmdret->opcode == FIO_NET_CMD_TEXT) {
410 struct cmd_text_pdu *__pdu = (struct cmd_text_pdu *) cmdret->payload;
411 char *buf = (char *) __pdu->buf;
413 buf[__pdu->buf_len] = '\0';
414 } else if (cmdret->opcode == FIO_NET_CMD_JOB) {
415 struct cmd_job_pdu *__pdu = (struct cmd_job_pdu *) cmdret->payload;
416 char *buf = (char *) __pdu->buf;
417 int len = le32_to_cpu(__pdu->buf_len);
423 /* frag flag is internal */
424 cmdret->flags &= ~FIO_NET_CMD_F_MORE;
430 static void add_reply(uint64_t tag, struct flist_head *list)
432 struct fio_net_cmd_reply *reply;
434 reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
435 flist_add_tail(&reply->list, list);
438 static uint64_t alloc_reply(uint64_t tag, uint16_t opcode)
440 struct fio_net_cmd_reply *reply;
442 reply = calloc(1, sizeof(*reply));
443 INIT_FLIST_HEAD(&reply->list);
444 fio_gettime(&reply->ts, NULL);
445 reply->saved_tag = tag;
446 reply->opcode = opcode;
448 return (uintptr_t) reply;
451 static void free_reply(uint64_t tag)
453 struct fio_net_cmd_reply *reply;
455 reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
459 static void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
463 cmd->cmd_crc16 = __cpu_to_le16(fio_crc16(cmd, FIO_NET_CMD_CRC_SZ));
465 pdu_len = le32_to_cpu(cmd->pdu_len);
466 cmd->pdu_crc16 = __cpu_to_le16(fio_crc16(pdu, pdu_len));
469 static void fio_net_cmd_crc(struct fio_net_cmd *cmd)
471 fio_net_cmd_crc_pdu(cmd, cmd->payload);
474 int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
475 uint64_t *tagptr, struct flist_head *list)
477 struct fio_net_cmd *cmd = NULL;
478 size_t this_len, cur_len = 0;
484 tag = *tagptr = alloc_reply(*tagptr, opcode);
486 tag = tagptr ? *tagptr : 0;
490 if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
491 this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
493 if (!cmd || cur_len < sizeof(*cmd) + this_len) {
497 cur_len = sizeof(*cmd) + this_len;
498 cmd = malloc(cur_len);
501 fio_init_net_cmd(cmd, opcode, buf, this_len, tag);
504 cmd->flags = __cpu_to_le32(FIO_NET_CMD_F_MORE);
506 fio_net_cmd_crc(cmd);
508 ret = fio_send_data(fd, cmd, sizeof(*cmd) + this_len);
511 } while (!ret && size);
517 add_reply(tag, list);
526 static struct sk_entry *fio_net_prep_cmd(uint16_t opcode, void *buf,
527 size_t size, uint64_t *tagptr,
530 struct sk_entry *entry;
532 entry = smalloc(sizeof(*entry));
536 INIT_FLIST_HEAD(&entry->next);
537 entry->opcode = opcode;
538 if (flags & SK_F_COPY) {
539 entry->buf = smalloc(size);
540 memcpy(entry->buf, buf, size);
546 entry->tag = *tagptr;
549 entry->flags = flags;
553 static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry);
555 static void fio_net_queue_entry(struct sk_entry *entry)
557 struct sk_out *sk_out = pthread_getspecific(sk_out_key);
559 if (entry->flags & SK_F_INLINE)
560 handle_sk_entry(sk_out, entry);
563 flist_add_tail(&entry->list, &sk_out->list);
566 fio_sem_up(&sk_out->wait);
570 static int fio_net_queue_cmd(uint16_t opcode, void *buf, off_t size,
571 uint64_t *tagptr, int flags)
573 struct sk_entry *entry;
575 entry = fio_net_prep_cmd(opcode, buf, size, tagptr, flags);
577 fio_net_queue_entry(entry);
584 static int fio_net_send_simple_stack_cmd(int sk, uint16_t opcode, uint64_t tag)
586 struct fio_net_cmd cmd;
588 fio_init_net_cmd(&cmd, opcode, NULL, 0, tag);
589 fio_net_cmd_crc(&cmd);
591 return fio_send_data(sk, &cmd, sizeof(cmd));
595 * If 'list' is non-NULL, then allocate and store the sent command for
596 * later verification.
598 int fio_net_send_simple_cmd(int sk, uint16_t opcode, uint64_t tag,
599 struct flist_head *list)
604 tag = alloc_reply(tag, opcode);
606 ret = fio_net_send_simple_stack_cmd(sk, opcode, tag);
615 add_reply(tag, list);
620 static int fio_net_queue_quit(void)
622 dprint(FD_NET, "server: sending quit\n");
624 return fio_net_queue_cmd(FIO_NET_CMD_QUIT, NULL, 0, NULL, SK_F_SIMPLE);
627 int fio_net_send_quit(int sk)
629 dprint(FD_NET, "server: sending quit\n");
631 return fio_net_send_simple_cmd(sk, FIO_NET_CMD_QUIT, 0, NULL);
634 static int fio_net_send_ack(struct fio_net_cmd *cmd, int error, int signal)
636 struct cmd_end_pdu epdu;
642 epdu.error = __cpu_to_le32(error);
643 epdu.signal = __cpu_to_le32(signal);
644 return fio_net_queue_cmd(FIO_NET_CMD_STOP, &epdu, sizeof(epdu), &tag, SK_F_COPY);
647 static int fio_net_queue_stop(int error, int signal)
649 dprint(FD_NET, "server: sending stop (%d, %d)\n", error, signal);
650 return fio_net_send_ack(NULL, error, signal);
653 static void fio_server_add_fork_item(pid_t pid, struct flist_head *list)
655 struct fio_fork_item *ffi;
657 ffi = malloc(sizeof(*ffi));
662 flist_add_tail(&ffi->list, list);
665 static void fio_server_add_conn_pid(struct flist_head *conn_list, pid_t pid)
667 dprint(FD_NET, "server: forked off connection job (pid=%u)\n", (int) pid);
668 fio_server_add_fork_item(pid, conn_list);
671 static void fio_server_add_job_pid(struct flist_head *job_list, pid_t pid)
673 dprint(FD_NET, "server: forked off job job (pid=%u)\n", (int) pid);
674 fio_server_add_fork_item(pid, job_list);
677 static void fio_server_check_fork_item(struct fio_fork_item *ffi)
681 ret = waitpid(ffi->pid, &status, WNOHANG);
683 if (errno == ECHILD) {
684 log_err("fio: connection pid %u disappeared\n", (int) ffi->pid);
687 log_err("fio: waitpid: %s\n", strerror(errno));
688 } else if (ret == ffi->pid) {
689 if (WIFSIGNALED(status)) {
690 ffi->signal = WTERMSIG(status);
693 if (WIFEXITED(status)) {
694 if (WEXITSTATUS(status))
695 ffi->exitval = WEXITSTATUS(status);
701 static void fio_server_fork_item_done(struct fio_fork_item *ffi, bool stop)
703 dprint(FD_NET, "pid %u exited, sig=%u, exitval=%d\n", (int) ffi->pid, ffi->signal, ffi->exitval);
706 * Fold STOP and QUIT...
709 fio_net_queue_stop(ffi->exitval, ffi->signal);
710 fio_net_queue_quit();
713 flist_del(&ffi->list);
717 static void fio_server_check_fork_items(struct flist_head *list, bool stop)
719 struct flist_head *entry, *tmp;
720 struct fio_fork_item *ffi;
722 flist_for_each_safe(entry, tmp, list) {
723 ffi = flist_entry(entry, struct fio_fork_item, list);
725 fio_server_check_fork_item(ffi);
728 fio_server_fork_item_done(ffi, stop);
732 static void fio_server_check_jobs(struct flist_head *job_list)
734 fio_server_check_fork_items(job_list, true);
737 static void fio_server_check_conns(struct flist_head *conn_list)
739 fio_server_check_fork_items(conn_list, false);
742 static int handle_load_file_cmd(struct fio_net_cmd *cmd)
744 struct cmd_load_file_pdu *pdu = (struct cmd_load_file_pdu *) cmd->payload;
745 void *file_name = pdu->file;
746 struct cmd_start_pdu spdu;
748 dprint(FD_NET, "server: loading local file %s\n", (char *) file_name);
750 pdu->name_len = le16_to_cpu(pdu->name_len);
751 pdu->client_type = le16_to_cpu(pdu->client_type);
753 if (parse_jobs_ini(file_name, 0, 0, pdu->client_type)) {
754 fio_net_queue_quit();
758 spdu.jobs = cpu_to_le32(thread_number);
759 spdu.stat_outputs = cpu_to_le32(stat_number);
760 fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
764 static int handle_run_cmd(struct sk_out *sk_out, struct flist_head *job_list,
765 struct fio_net_cmd *cmd)
770 sk_out_assign(sk_out);
777 fio_server_add_job_pid(job_list, pid);
781 ret = fio_backend(sk_out);
787 static int handle_job_cmd(struct fio_net_cmd *cmd)
789 struct cmd_job_pdu *pdu = (struct cmd_job_pdu *) cmd->payload;
790 void *buf = pdu->buf;
791 struct cmd_start_pdu spdu;
793 pdu->buf_len = le32_to_cpu(pdu->buf_len);
794 pdu->client_type = le32_to_cpu(pdu->client_type);
796 if (parse_jobs_ini(buf, 1, 0, pdu->client_type)) {
797 fio_net_queue_quit();
801 spdu.jobs = cpu_to_le32(thread_number);
802 spdu.stat_outputs = cpu_to_le32(stat_number);
804 fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
808 static int handle_jobline_cmd(struct fio_net_cmd *cmd)
810 void *pdu = cmd->payload;
811 struct cmd_single_line_pdu *cslp;
812 struct cmd_line_pdu *clp;
813 unsigned long offset;
814 struct cmd_start_pdu spdu;
819 clp->lines = le16_to_cpu(clp->lines);
820 clp->client_type = le16_to_cpu(clp->client_type);
821 argv = malloc(clp->lines * sizeof(char *));
822 offset = sizeof(*clp);
824 dprint(FD_NET, "server: %d command line args\n", clp->lines);
826 for (i = 0; i < clp->lines; i++) {
828 argv[i] = (char *) cslp->text;
830 offset += sizeof(*cslp) + le16_to_cpu(cslp->len);
831 dprint(FD_NET, "server: %d: %s\n", i, argv[i]);
834 if (parse_cmd_line(clp->lines, argv, clp->client_type)) {
835 fio_net_queue_quit();
842 spdu.jobs = cpu_to_le32(thread_number);
843 spdu.stat_outputs = cpu_to_le32(stat_number);
845 fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
849 static int handle_probe_cmd(struct fio_net_cmd *cmd)
851 struct cmd_client_probe_pdu *pdu = (struct cmd_client_probe_pdu *) cmd->payload;
852 uint64_t tag = cmd->tag;
853 struct cmd_probe_reply_pdu probe = {
854 #ifdef CONFIG_BIG_ENDIAN
859 .bpp = sizeof(void *),
860 .cpus = __cpu_to_le32(cpus_online()),
863 dprint(FD_NET, "server: sending probe reply\n");
865 strcpy(me, (char *) pdu->server);
867 gethostname((char *) probe.hostname, sizeof(probe.hostname));
868 strncpy((char *) probe.fio_version, fio_version_string, sizeof(probe.fio_version) - 1);
871 * If the client supports compression and we do too, then enable it
873 if (has_zlib && le64_to_cpu(pdu->flags) & FIO_PROBE_FLAG_ZLIB) {
874 probe.flags = __cpu_to_le64(FIO_PROBE_FLAG_ZLIB);
881 return fio_net_queue_cmd(FIO_NET_CMD_PROBE, &probe, sizeof(probe), &tag, SK_F_COPY);
884 static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
887 uint64_t tag = cmd->tag;
891 dprint(FD_NET, "server sending status\n");
894 * Fake ETA return if we don't have a local one, otherwise the client
895 * will end up timing out waiting for a response to the ETA request
897 je = get_jobs_eta(true, &size);
900 je = calloc(1, size);
902 je->nr_running = cpu_to_le32(je->nr_running);
903 je->nr_ramp = cpu_to_le32(je->nr_ramp);
904 je->nr_pending = cpu_to_le32(je->nr_pending);
905 je->nr_setting_up = cpu_to_le32(je->nr_setting_up);
906 je->files_open = cpu_to_le32(je->files_open);
908 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
909 je->m_rate[i] = cpu_to_le64(je->m_rate[i]);
910 je->t_rate[i] = cpu_to_le64(je->t_rate[i]);
911 je->m_iops[i] = cpu_to_le32(je->m_iops[i]);
912 je->t_iops[i] = cpu_to_le32(je->t_iops[i]);
913 je->rate[i] = cpu_to_le64(je->rate[i]);
914 je->iops[i] = cpu_to_le32(je->iops[i]);
917 je->elapsed_sec = cpu_to_le64(je->elapsed_sec);
918 je->eta_sec = cpu_to_le64(je->eta_sec);
919 je->nr_threads = cpu_to_le32(je->nr_threads);
920 je->is_pow2 = cpu_to_le32(je->is_pow2);
921 je->unit_base = cpu_to_le32(je->unit_base);
924 fio_net_queue_cmd(FIO_NET_CMD_ETA, je, size, &tag, SK_F_FREE);
928 static int send_update_job_reply(uint64_t __tag, int error)
930 uint64_t tag = __tag;
933 pdu_error = __cpu_to_le32(error);
934 return fio_net_queue_cmd(FIO_NET_CMD_UPDATE_JOB, &pdu_error, sizeof(pdu_error), &tag, SK_F_COPY);
937 static int handle_update_job_cmd(struct fio_net_cmd *cmd)
939 struct cmd_add_job_pdu *pdu = (struct cmd_add_job_pdu *) cmd->payload;
940 struct thread_data *td;
943 tnumber = le32_to_cpu(pdu->thread_number);
945 dprint(FD_NET, "server: updating options for job %u\n", tnumber);
947 if (!tnumber || tnumber > thread_number) {
948 send_update_job_reply(cmd->tag, ENODEV);
952 td = &threads[tnumber - 1];
953 convert_thread_options_to_cpu(&td->o, &pdu->top);
954 send_update_job_reply(cmd->tag, 0);
958 static int handle_trigger_cmd(struct fio_net_cmd *cmd, struct flist_head *job_list)
960 struct cmd_vtrigger_pdu *pdu = (struct cmd_vtrigger_pdu *) cmd->payload;
961 char *buf = (char *) pdu->cmd;
962 struct all_io_list *rep;
965 pdu->len = le16_to_cpu(pdu->len);
966 buf[pdu->len] = '\0';
968 rep = get_all_io_list(IO_LIST_ALL, &sz);
970 struct all_io_list state;
972 state.threads = cpu_to_le64((uint64_t) 0);
973 fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, &state, sizeof(state), NULL, SK_F_COPY | SK_F_INLINE);
975 fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, rep, sz, NULL, SK_F_FREE | SK_F_INLINE);
977 fio_terminate_threads(TERMINATE_ALL);
978 fio_server_check_jobs(job_list);
983 static int handle_command(struct sk_out *sk_out, struct flist_head *job_list,
984 struct fio_net_cmd *cmd)
988 dprint(FD_NET, "server: got op [%s], pdu=%u, tag=%llx\n",
989 fio_server_op(cmd->opcode), cmd->pdu_len,
990 (unsigned long long) cmd->tag);
992 switch (cmd->opcode) {
993 case FIO_NET_CMD_QUIT:
994 fio_terminate_threads(TERMINATE_ALL);
997 case FIO_NET_CMD_EXIT:
1000 case FIO_NET_CMD_LOAD_FILE:
1001 ret = handle_load_file_cmd(cmd);
1003 case FIO_NET_CMD_JOB:
1004 ret = handle_job_cmd(cmd);
1006 case FIO_NET_CMD_JOBLINE:
1007 ret = handle_jobline_cmd(cmd);
1009 case FIO_NET_CMD_PROBE:
1010 ret = handle_probe_cmd(cmd);
1012 case FIO_NET_CMD_SEND_ETA:
1013 ret = handle_send_eta_cmd(cmd);
1015 case FIO_NET_CMD_RUN:
1016 ret = handle_run_cmd(sk_out, job_list, cmd);
1018 case FIO_NET_CMD_UPDATE_JOB:
1019 ret = handle_update_job_cmd(cmd);
1021 case FIO_NET_CMD_VTRIGGER:
1022 ret = handle_trigger_cmd(cmd, job_list);
1024 case FIO_NET_CMD_SENDFILE: {
1025 struct cmd_sendfile_reply *in;
1026 struct cmd_reply *rep;
1028 rep = (struct cmd_reply *) (uintptr_t) cmd->tag;
1030 in = (struct cmd_sendfile_reply *) cmd->payload;
1031 in->size = le32_to_cpu(in->size);
1032 in->error = le32_to_cpu(in->error);
1035 rep->error = in->error;
1038 rep->data = smalloc(in->size);
1041 rep->error = ENOMEM;
1043 rep->size = in->size;
1044 memcpy(rep->data, in->data, in->size);
1047 fio_sem_up(&rep->lock);
1051 log_err("fio: unknown opcode: %s\n", fio_server_op(cmd->opcode));
1059 * Send a command with a separate PDU, not inlined in the command
1061 static int fio_send_cmd_ext_pdu(int sk, uint16_t opcode, const void *buf,
1062 off_t size, uint64_t tag, uint32_t flags)
1064 struct fio_net_cmd cmd;
1065 struct iovec iov[2];
1069 iov[0].iov_base = (void *) &cmd;
1070 iov[0].iov_len = sizeof(cmd);
1073 uint32_t this_flags = flags;
1076 if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
1077 this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
1079 if (this_len < size)
1080 this_flags |= FIO_NET_CMD_F_MORE;
1082 __fio_init_net_cmd(&cmd, opcode, this_len, tag);
1083 cmd.flags = __cpu_to_le32(this_flags);
1084 fio_net_cmd_crc_pdu(&cmd, buf);
1086 iov[1].iov_base = (void *) buf;
1087 iov[1].iov_len = this_len;
1089 ret = fio_sendv_data(sk, iov, 2);
1092 } while (!ret && size);
1097 static void finish_entry(struct sk_entry *entry)
1099 if (entry->flags & SK_F_FREE)
1101 else if (entry->flags & SK_F_COPY)
1107 static void entry_set_flags(struct sk_entry *entry, struct flist_head *list,
1108 unsigned int *flags)
1110 if (!flist_empty(list))
1111 *flags = FIO_NET_CMD_F_MORE;
1116 static int send_vec_entry(struct sk_out *sk_out, struct sk_entry *first)
1121 entry_set_flags(first, &first->next, &flags);
1123 ret = fio_send_cmd_ext_pdu(sk_out->sk, first->opcode, first->buf,
1124 first->size, first->tag, flags);
1126 while (!flist_empty(&first->next)) {
1127 struct sk_entry *next;
1129 next = flist_first_entry(&first->next, struct sk_entry, list);
1130 flist_del_init(&next->list);
1132 entry_set_flags(next, &first->next, &flags);
1134 ret += fio_send_cmd_ext_pdu(sk_out->sk, next->opcode, next->buf,
1135 next->size, next->tag, flags);
1142 static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry)
1146 fio_sem_down(&sk_out->xmit);
1148 if (entry->flags & SK_F_VEC)
1149 ret = send_vec_entry(sk_out, entry);
1150 else if (entry->flags & SK_F_SIMPLE) {
1151 ret = fio_net_send_simple_cmd(sk_out->sk, entry->opcode,
1154 ret = fio_net_send_cmd(sk_out->sk, entry->opcode, entry->buf,
1155 entry->size, &entry->tag, NULL);
1158 fio_sem_up(&sk_out->xmit);
1161 log_err("fio: failed handling cmd %s\n", fio_server_op(entry->opcode));
1163 finish_entry(entry);
1167 static int handle_xmits(struct sk_out *sk_out)
1169 struct sk_entry *entry;
1174 if (flist_empty(&sk_out->list)) {
1179 flist_splice_init(&sk_out->list, &list);
1182 while (!flist_empty(&list)) {
1183 entry = flist_entry(list.next, struct sk_entry, list);
1184 flist_del(&entry->list);
1185 ret += handle_sk_entry(sk_out, entry);
1191 static int handle_connection(struct sk_out *sk_out)
1193 struct fio_net_cmd *cmd = NULL;
1194 FLIST_HEAD(job_list);
1200 while (!exit_backend) {
1201 struct pollfd pfd = {
1209 if (!flist_empty(&job_list))
1212 handle_xmits(sk_out);
1214 ret = poll(&pfd, 1, 0);
1218 log_err("fio: poll: %s\n", strerror(errno));
1221 fio_server_check_jobs(&job_list);
1222 fio_sem_down_timeout(&sk_out->wait, timeout);
1226 if (pfd.revents & POLLIN)
1228 if (pfd.revents & (POLLERR|POLLHUP)) {
1232 } while (!exit_backend);
1234 fio_server_check_jobs(&job_list);
1239 cmd = fio_net_recv_cmd(sk_out->sk, true);
1245 ret = handle_command(sk_out, &job_list, cmd);
1256 handle_xmits(sk_out);
1260 __sk_out_drop(sk_out);
1264 /* get the address on this host bound by the input socket,
1265 * whether it is ipv6 or ipv4 */
1267 static int get_my_addr_str(int sk)
1269 struct sockaddr_in6 myaddr6 = { 0, };
1270 struct sockaddr_in myaddr4 = { 0, };
1271 struct sockaddr *sockaddr_p;
1277 len = sizeof(myaddr6);
1278 sockaddr_p = (struct sockaddr * )&myaddr6;
1279 net_addr = (char * )&myaddr6.sin6_addr;
1281 len = sizeof(myaddr4);
1282 sockaddr_p = (struct sockaddr * )&myaddr4;
1283 net_addr = (char * )&myaddr4.sin_addr;
1286 ret = getsockname(sk, sockaddr_p, &len);
1288 log_err("fio: getsockname: %s\n", strerror(errno));
1292 if (!inet_ntop(use_ipv6?AF_INET6:AF_INET, net_addr, client_sockaddr_str, INET6_ADDRSTRLEN - 1)) {
1293 log_err("inet_ntop: failed to convert addr to string\n");
1297 dprint(FD_NET, "fio server bound to addr %s\n", client_sockaddr_str);
1301 static int accept_loop(int listen_sk)
1303 struct sockaddr_in addr;
1304 struct sockaddr_in6 addr6;
1305 socklen_t len = use_ipv6 ? sizeof(addr6) : sizeof(addr);
1307 int ret = 0, sk, exitval = 0;
1308 FLIST_HEAD(conn_list);
1310 dprint(FD_NET, "server enter accept loop\n");
1312 fio_set_fd_nonblocking(listen_sk, "server");
1314 while (!exit_backend) {
1315 struct sk_out *sk_out;
1321 pfd.events = POLLIN;
1325 if (!flist_empty(&conn_list))
1328 ret = poll(&pfd, 1, timeout);
1332 log_err("fio: poll: %s\n", strerror(errno));
1335 fio_server_check_conns(&conn_list);
1339 if (pfd.revents & POLLIN)
1341 } while (!exit_backend);
1343 fio_server_check_conns(&conn_list);
1345 if (exit_backend || ret < 0)
1349 sk = accept(listen_sk, (struct sockaddr *) &addr6, &len);
1351 sk = accept(listen_sk, (struct sockaddr *) &addr, &len);
1354 log_err("fio: accept: %s\n", strerror(errno));
1359 from = inet_ntop(AF_INET6, (struct sockaddr *) &addr6.sin6_addr, buf, sizeof(buf));
1361 from = inet_ntop(AF_INET, (struct sockaddr *) &addr.sin_addr, buf, sizeof(buf));
1363 dprint(FD_NET, "server: connect from %s\n", from);
1365 sk_out = scalloc(1, sizeof(*sk_out));
1372 INIT_FLIST_HEAD(&sk_out->list);
1373 __fio_sem_init(&sk_out->lock, FIO_SEM_UNLOCKED);
1374 __fio_sem_init(&sk_out->wait, FIO_SEM_LOCKED);
1375 __fio_sem_init(&sk_out->xmit, FIO_SEM_UNLOCKED);
1380 fio_server_add_conn_pid(&conn_list, pid);
1384 /* if error, it's already logged, non-fatal */
1385 get_my_addr_str(sk);
1388 * Assign sk_out here, it'll be dropped in handle_connection()
1389 * since that function calls _exit() when done
1391 sk_out_assign(sk_out);
1392 handle_connection(sk_out);
1398 int fio_server_text_output(int level, const char *buf, size_t len)
1400 struct sk_out *sk_out = pthread_getspecific(sk_out_key);
1401 struct cmd_text_pdu *pdu;
1405 if (!sk_out || sk_out->sk == -1)
1408 tlen = sizeof(*pdu) + len;
1411 pdu->level = __cpu_to_le32(level);
1412 pdu->buf_len = __cpu_to_le32(len);
1414 gettimeofday(&tv, NULL);
1415 pdu->log_sec = __cpu_to_le64(tv.tv_sec);
1416 pdu->log_usec = __cpu_to_le64(tv.tv_usec);
1418 memcpy(pdu->buf, buf, len);
1420 fio_net_queue_cmd(FIO_NET_CMD_TEXT, pdu, tlen, NULL, SK_F_COPY);
1425 static void convert_io_stat(struct io_stat *dst, struct io_stat *src)
1427 dst->max_val = cpu_to_le64(src->max_val);
1428 dst->min_val = cpu_to_le64(src->min_val);
1429 dst->samples = cpu_to_le64(src->samples);
1432 * Encode to IEEE 754 for network transfer
1434 dst->mean.u.i = cpu_to_le64(fio_double_to_uint64(src->mean.u.f));
1435 dst->S.u.i = cpu_to_le64(fio_double_to_uint64(src->S.u.f));
1438 static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
1442 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1443 dst->max_run[i] = cpu_to_le64(src->max_run[i]);
1444 dst->min_run[i] = cpu_to_le64(src->min_run[i]);
1445 dst->max_bw[i] = cpu_to_le64(src->max_bw[i]);
1446 dst->min_bw[i] = cpu_to_le64(src->min_bw[i]);
1447 dst->iobytes[i] = cpu_to_le64(src->iobytes[i]);
1448 dst->agg[i] = cpu_to_le64(src->agg[i]);
1451 dst->kb_base = cpu_to_le32(src->kb_base);
1452 dst->unit_base = cpu_to_le32(src->unit_base);
1453 dst->groupid = cpu_to_le32(src->groupid);
1454 dst->unified_rw_rep = cpu_to_le32(src->unified_rw_rep);
1455 dst->sig_figs = cpu_to_le32(src->sig_figs);
1459 * Send a CMD_TS, which packs struct thread_stat and group_run_stats
1460 * into a single payload.
1462 void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
1464 struct cmd_ts_pdu p;
1467 uint64_t *ss_iops, *ss_bw;
1469 dprint(FD_NET, "server sending end stats\n");
1471 memset(&p, 0, sizeof(p));
1473 strncpy(p.ts.name, ts->name, FIO_JOBNAME_SIZE - 1);
1474 strncpy(p.ts.verror, ts->verror, FIO_VERROR_SIZE - 1);
1475 strncpy(p.ts.description, ts->description, FIO_JOBDESC_SIZE - 1);
1477 p.ts.error = cpu_to_le32(ts->error);
1478 p.ts.thread_number = cpu_to_le32(ts->thread_number);
1479 p.ts.groupid = cpu_to_le32(ts->groupid);
1480 p.ts.pid = cpu_to_le32(ts->pid);
1481 p.ts.members = cpu_to_le32(ts->members);
1482 p.ts.unified_rw_rep = cpu_to_le32(ts->unified_rw_rep);
1484 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1485 convert_io_stat(&p.ts.clat_stat[i], &ts->clat_stat[i]);
1486 convert_io_stat(&p.ts.slat_stat[i], &ts->slat_stat[i]);
1487 convert_io_stat(&p.ts.lat_stat[i], &ts->lat_stat[i]);
1488 convert_io_stat(&p.ts.bw_stat[i], &ts->bw_stat[i]);
1489 convert_io_stat(&p.ts.iops_stat[i], &ts->iops_stat[i]);
1492 p.ts.usr_time = cpu_to_le64(ts->usr_time);
1493 p.ts.sys_time = cpu_to_le64(ts->sys_time);
1494 p.ts.ctx = cpu_to_le64(ts->ctx);
1495 p.ts.minf = cpu_to_le64(ts->minf);
1496 p.ts.majf = cpu_to_le64(ts->majf);
1497 p.ts.clat_percentiles = cpu_to_le32(ts->clat_percentiles);
1498 p.ts.lat_percentiles = cpu_to_le32(ts->lat_percentiles);
1499 p.ts.percentile_precision = cpu_to_le64(ts->percentile_precision);
1501 for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
1502 fio_fp64_t *src = &ts->percentile_list[i];
1503 fio_fp64_t *dst = &p.ts.percentile_list[i];
1505 dst->u.i = cpu_to_le64(fio_double_to_uint64(src->u.f));
1508 for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
1509 p.ts.io_u_map[i] = cpu_to_le64(ts->io_u_map[i]);
1510 p.ts.io_u_submit[i] = cpu_to_le64(ts->io_u_submit[i]);
1511 p.ts.io_u_complete[i] = cpu_to_le64(ts->io_u_complete[i]);
1514 for (i = 0; i < FIO_IO_U_LAT_N_NR; i++)
1515 p.ts.io_u_lat_n[i] = cpu_to_le64(ts->io_u_lat_n[i]);
1516 for (i = 0; i < FIO_IO_U_LAT_U_NR; i++)
1517 p.ts.io_u_lat_u[i] = cpu_to_le64(ts->io_u_lat_u[i]);
1518 for (i = 0; i < FIO_IO_U_LAT_M_NR; i++)
1519 p.ts.io_u_lat_m[i] = cpu_to_le64(ts->io_u_lat_m[i]);
1521 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1522 for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
1523 p.ts.io_u_plat[i][j] = cpu_to_le64(ts->io_u_plat[i][j]);
1525 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1526 p.ts.total_io_u[i] = cpu_to_le64(ts->total_io_u[i]);
1527 p.ts.short_io_u[i] = cpu_to_le64(ts->short_io_u[i]);
1528 p.ts.drop_io_u[i] = cpu_to_le64(ts->drop_io_u[i]);
1531 p.ts.total_submit = cpu_to_le64(ts->total_submit);
1532 p.ts.total_complete = cpu_to_le64(ts->total_complete);
1534 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1535 p.ts.io_bytes[i] = cpu_to_le64(ts->io_bytes[i]);
1536 p.ts.runtime[i] = cpu_to_le64(ts->runtime[i]);
1539 p.ts.total_run_time = cpu_to_le64(ts->total_run_time);
1540 p.ts.continue_on_error = cpu_to_le16(ts->continue_on_error);
1541 p.ts.total_err_count = cpu_to_le64(ts->total_err_count);
1542 p.ts.first_error = cpu_to_le32(ts->first_error);
1543 p.ts.kb_base = cpu_to_le32(ts->kb_base);
1544 p.ts.unit_base = cpu_to_le32(ts->unit_base);
1546 p.ts.latency_depth = cpu_to_le32(ts->latency_depth);
1547 p.ts.latency_target = cpu_to_le64(ts->latency_target);
1548 p.ts.latency_window = cpu_to_le64(ts->latency_window);
1549 p.ts.latency_percentile.u.i = cpu_to_le64(fio_double_to_uint64(ts->latency_percentile.u.f));
1551 p.ts.sig_figs = cpu_to_le32(ts->sig_figs);
1553 p.ts.nr_block_infos = cpu_to_le64(ts->nr_block_infos);
1554 for (i = 0; i < p.ts.nr_block_infos; i++)
1555 p.ts.block_infos[i] = cpu_to_le32(ts->block_infos[i]);
1557 p.ts.ss_dur = cpu_to_le64(ts->ss_dur);
1558 p.ts.ss_state = cpu_to_le32(ts->ss_state);
1559 p.ts.ss_head = cpu_to_le32(ts->ss_head);
1560 p.ts.ss_limit.u.i = cpu_to_le64(fio_double_to_uint64(ts->ss_limit.u.f));
1561 p.ts.ss_slope.u.i = cpu_to_le64(fio_double_to_uint64(ts->ss_slope.u.f));
1562 p.ts.ss_deviation.u.i = cpu_to_le64(fio_double_to_uint64(ts->ss_deviation.u.f));
1563 p.ts.ss_criterion.u.i = cpu_to_le64(fio_double_to_uint64(ts->ss_criterion.u.f));
1565 convert_gs(&p.rs, rs);
1567 dprint(FD_NET, "ts->ss_state = %d\n", ts->ss_state);
1568 if (ts->ss_state & FIO_SS_DATA) {
1569 dprint(FD_NET, "server sending steadystate ring buffers\n");
1571 ss_buf = malloc(sizeof(p) + 2*ts->ss_dur*sizeof(uint64_t));
1573 memcpy(ss_buf, &p, sizeof(p));
1575 ss_iops = (uint64_t *) ((struct cmd_ts_pdu *)ss_buf + 1);
1576 ss_bw = ss_iops + (int) ts->ss_dur;
1577 for (i = 0; i < ts->ss_dur; i++) {
1578 ss_iops[i] = cpu_to_le64(ts->ss_iops_data[i]);
1579 ss_bw[i] = cpu_to_le64(ts->ss_bw_data[i]);
1582 fio_net_queue_cmd(FIO_NET_CMD_TS, ss_buf, sizeof(p) + 2*ts->ss_dur*sizeof(uint64_t), NULL, SK_F_COPY);
1587 fio_net_queue_cmd(FIO_NET_CMD_TS, &p, sizeof(p), NULL, SK_F_COPY);
1590 void fio_server_send_gs(struct group_run_stats *rs)
1592 struct group_run_stats gs;
1594 dprint(FD_NET, "server sending group run stats\n");
1596 convert_gs(&gs, rs);
1597 fio_net_queue_cmd(FIO_NET_CMD_GS, &gs, sizeof(gs), NULL, SK_F_COPY);
1600 void fio_server_send_job_options(struct flist_head *opt_list,
1603 struct cmd_job_option pdu;
1604 struct flist_head *entry;
1606 if (flist_empty(opt_list))
1609 flist_for_each(entry, opt_list) {
1610 struct print_option *p;
1613 p = flist_entry(entry, struct print_option, list);
1614 memset(&pdu, 0, sizeof(pdu));
1617 pdu.global = __cpu_to_le16(1);
1621 pdu.groupid = cpu_to_le32(gid);
1623 len = strlen(p->name);
1624 if (len >= sizeof(pdu.name)) {
1625 len = sizeof(pdu.name) - 1;
1626 pdu.truncated = __cpu_to_le16(1);
1628 memcpy(pdu.name, p->name, len);
1630 len = strlen(p->value);
1631 if (len >= sizeof(pdu.value)) {
1632 len = sizeof(pdu.value) - 1;
1633 pdu.truncated = __cpu_to_le16(1);
1635 memcpy(pdu.value, p->value, len);
1637 fio_net_queue_cmd(FIO_NET_CMD_JOB_OPT, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1641 static void convert_agg(struct disk_util_agg *dst, struct disk_util_agg *src)
1645 for (i = 0; i < 2; i++) {
1646 dst->ios[i] = cpu_to_le64(src->ios[i]);
1647 dst->merges[i] = cpu_to_le64(src->merges[i]);
1648 dst->sectors[i] = cpu_to_le64(src->sectors[i]);
1649 dst->ticks[i] = cpu_to_le64(src->ticks[i]);
1652 dst->io_ticks = cpu_to_le64(src->io_ticks);
1653 dst->time_in_queue = cpu_to_le64(src->time_in_queue);
1654 dst->slavecount = cpu_to_le32(src->slavecount);
1655 dst->max_util.u.i = cpu_to_le64(fio_double_to_uint64(src->max_util.u.f));
1658 static void convert_dus(struct disk_util_stat *dst, struct disk_util_stat *src)
1662 dst->name[FIO_DU_NAME_SZ - 1] = '\0';
1663 strncpy((char *) dst->name, (char *) src->name, FIO_DU_NAME_SZ - 1);
1665 for (i = 0; i < 2; i++) {
1666 dst->s.ios[i] = cpu_to_le64(src->s.ios[i]);
1667 dst->s.merges[i] = cpu_to_le64(src->s.merges[i]);
1668 dst->s.sectors[i] = cpu_to_le64(src->s.sectors[i]);
1669 dst->s.ticks[i] = cpu_to_le64(src->s.ticks[i]);
1672 dst->s.io_ticks = cpu_to_le64(src->s.io_ticks);
1673 dst->s.time_in_queue = cpu_to_le64(src->s.time_in_queue);
1674 dst->s.msec = cpu_to_le64(src->s.msec);
1677 void fio_server_send_du(void)
1679 struct disk_util *du;
1680 struct flist_head *entry;
1681 struct cmd_du_pdu pdu;
1683 dprint(FD_NET, "server: sending disk_util %d\n", !flist_empty(&disk_list));
1685 memset(&pdu, 0, sizeof(pdu));
1687 flist_for_each(entry, &disk_list) {
1688 du = flist_entry(entry, struct disk_util, list);
1690 convert_dus(&pdu.dus, &du->dus);
1691 convert_agg(&pdu.agg, &du->agg);
1693 fio_net_queue_cmd(FIO_NET_CMD_DU, &pdu, sizeof(pdu), NULL, SK_F_COPY);
1699 static inline void __fio_net_prep_tail(z_stream *stream, void *out_pdu,
1700 struct sk_entry **last_entry,
1701 struct sk_entry *first)
1703 unsigned int this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
1705 *last_entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
1706 NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
1708 flist_add_tail(&(*last_entry)->list, &first->next);
1712 * Deflates the next input given, creating as many new packets in the
1713 * linked list as necessary.
1715 static int __deflate_pdu_buffer(void *next_in, unsigned int next_sz, void **out_pdu,
1716 struct sk_entry **last_entry, z_stream *stream,
1717 struct sk_entry *first)
1721 stream->next_in = next_in;
1722 stream->avail_in = next_sz;
1724 if (!stream->avail_out) {
1725 __fio_net_prep_tail(stream, *out_pdu, last_entry, first);
1726 if (*last_entry == NULL)
1729 *out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1731 stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1732 stream->next_out = *out_pdu;
1735 ret = deflate(stream, Z_BLOCK);
1741 } while (stream->avail_in);
1746 static int __fio_append_iolog_gz_hist(struct sk_entry *first, struct io_log *log,
1747 struct io_logs *cur_log, z_stream *stream)
1749 struct sk_entry *entry;
1752 int sample_sz = log_entry_sz(log);
1754 out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1755 stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1756 stream->next_out = out_pdu;
1758 for (i = 0; i < cur_log->nr_samples; i++) {
1759 struct io_sample *s;
1760 struct io_u_plat_entry *cur_plat_entry, *prev_plat_entry;
1761 uint64_t *cur_plat, *prev_plat;
1763 s = get_sample(log, cur_log, i);
1764 ret = __deflate_pdu_buffer(s, sample_sz, &out_pdu, &entry, stream, first);
1768 /* Do the subtraction on server side so that client doesn't have to
1769 * reconstruct our linked list from packets.
1771 cur_plat_entry = s->data.plat_entry;
1772 prev_plat_entry = flist_first_entry(&cur_plat_entry->list, struct io_u_plat_entry, list);
1773 cur_plat = cur_plat_entry->io_u_plat;
1774 prev_plat = prev_plat_entry->io_u_plat;
1776 for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
1777 cur_plat[j] -= prev_plat[j];
1780 flist_del(&prev_plat_entry->list);
1781 free(prev_plat_entry);
1783 ret = __deflate_pdu_buffer(cur_plat_entry, sizeof(*cur_plat_entry),
1784 &out_pdu, &entry, stream, first);
1790 __fio_net_prep_tail(stream, out_pdu, &entry, first);
1791 return entry == NULL;
1794 static int __fio_append_iolog_gz(struct sk_entry *first, struct io_log *log,
1795 struct io_logs *cur_log, z_stream *stream)
1797 unsigned int this_len;
1801 if (log->log_type == IO_LOG_TYPE_HIST)
1802 return __fio_append_iolog_gz_hist(first, log, cur_log, stream);
1804 stream->next_in = (void *) cur_log->log;
1805 stream->avail_in = cur_log->nr_samples * log_entry_sz(log);
1808 struct sk_entry *entry;
1811 * Dirty - since the log is potentially huge, compress it into
1812 * FIO_SERVER_MAX_FRAGMENT_PDU chunks and let the receiving
1813 * side defragment it.
1815 out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1817 stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1818 stream->next_out = out_pdu;
1819 ret = deflate(stream, Z_BLOCK);
1820 /* may be Z_OK, or Z_STREAM_END */
1826 this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
1828 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
1829 NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
1834 flist_add_tail(&entry->list, &first->next);
1835 } while (stream->avail_in);
1840 static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
1849 if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK)
1852 while (!flist_empty(&log->io_logs)) {
1853 struct io_logs *cur_log;
1855 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1856 flist_del_init(&cur_log->list);
1858 ret = __fio_append_iolog_gz(first, log, cur_log, &stream);
1863 ret = deflate(&stream, Z_FINISH);
1865 while (ret != Z_STREAM_END) {
1866 struct sk_entry *entry;
1867 unsigned int this_len;
1870 out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
1871 stream.avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
1872 stream.next_out = out_pdu;
1874 ret = deflate(&stream, Z_FINISH);
1875 /* may be Z_OK, or Z_STREAM_END */
1881 this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream.avail_out;
1883 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
1884 NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
1889 flist_add_tail(&entry->list, &first->next);
1890 } while (ret != Z_STREAM_END);
1892 ret = deflateEnd(&stream);
1899 static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
1905 static int fio_append_gz_chunks(struct sk_entry *first, struct io_log *log)
1907 struct sk_entry *entry;
1908 struct flist_head *node;
1911 pthread_mutex_lock(&log->chunk_lock);
1912 flist_for_each(node, &log->chunk_list) {
1913 struct iolog_compress *c;
1915 c = flist_entry(node, struct iolog_compress, list);
1916 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, c->buf, c->len,
1917 NULL, SK_F_VEC | SK_F_INLINE);
1922 flist_add_tail(&entry->list, &first->next);
1924 pthread_mutex_unlock(&log->chunk_lock);
1928 static int fio_append_text_log(struct sk_entry *first, struct io_log *log)
1930 struct sk_entry *entry;
1933 while (!flist_empty(&log->io_logs)) {
1934 struct io_logs *cur_log;
1937 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1938 flist_del_init(&cur_log->list);
1940 size = cur_log->nr_samples * log_entry_sz(log);
1942 entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, cur_log->log, size,
1943 NULL, SK_F_VEC | SK_F_INLINE);
1948 flist_add_tail(&entry->list, &first->next);
1954 int fio_send_iolog(struct thread_data *td, struct io_log *log, const char *name)
1956 struct cmd_iolog_pdu pdu = {
1957 .nr_samples = cpu_to_le64(iolog_nr_samples(log)),
1958 .thread_number = cpu_to_le32(td->thread_number),
1959 .log_type = cpu_to_le32(log->log_type),
1960 .log_hist_coarseness = cpu_to_le32(log->hist_coarseness),
1962 struct sk_entry *first;
1963 struct flist_head *entry;
1966 if (!flist_empty(&log->chunk_list))
1967 pdu.compressed = __cpu_to_le32(STORE_COMPRESSED);
1969 pdu.compressed = __cpu_to_le32(XMIT_COMPRESSED);
1973 strncpy((char *) pdu.name, name, FIO_NET_NAME_MAX);
1974 pdu.name[FIO_NET_NAME_MAX - 1] = '\0';
1977 * We can't do this for a pre-compressed log, but for that case,
1978 * log->nr_samples is zero anyway.
1980 flist_for_each(entry, &log->io_logs) {
1981 struct io_logs *cur_log;
1984 cur_log = flist_entry(entry, struct io_logs, list);
1986 for (i = 0; i < cur_log->nr_samples; i++) {
1987 struct io_sample *s = get_sample(log, cur_log, i);
1989 s->time = cpu_to_le64(s->time);
1990 s->data.val = cpu_to_le64(s->data.val);
1991 s->__ddir = cpu_to_le32(s->__ddir);
1992 s->bs = cpu_to_le64(s->bs);
1994 if (log->log_offset) {
1995 struct io_sample_offset *so = (void *) s;
1997 so->offset = cpu_to_le64(so->offset);
2003 * Assemble header entry first
2005 first = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, &pdu, sizeof(pdu), NULL, SK_F_VEC | SK_F_INLINE | SK_F_COPY);
2010 * Now append actual log entries. If log compression was enabled on
2011 * the job, just send out the compressed chunks directly. If we
2012 * have a plain log, compress if we can, then send. Otherwise, send
2013 * the plain text output.
2015 if (!flist_empty(&log->chunk_list))
2016 ret = fio_append_gz_chunks(first, log);
2018 ret = fio_append_iolog_gz(first, log);
2020 ret = fio_append_text_log(first, log);
2022 fio_net_queue_entry(first);
2026 void fio_server_send_add_job(struct thread_data *td)
2028 struct cmd_add_job_pdu pdu = {
2029 .thread_number = cpu_to_le32(td->thread_number),
2030 .groupid = cpu_to_le32(td->groupid),
2033 convert_thread_options_to_net(&pdu.top, &td->o);
2035 fio_net_queue_cmd(FIO_NET_CMD_ADD_JOB, &pdu, sizeof(pdu), NULL,
2039 void fio_server_send_start(struct thread_data *td)
2041 struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2043 assert(sk_out->sk != -1);
2045 fio_net_queue_cmd(FIO_NET_CMD_SERVER_START, NULL, 0, NULL, SK_F_SIMPLE);
2048 int fio_server_get_verify_state(const char *name, int threadnumber,
2051 struct thread_io_list *s;
2052 struct cmd_sendfile out;
2053 struct cmd_reply *rep;
2058 dprint(FD_NET, "server: request verify state\n");
2060 rep = smalloc(sizeof(*rep));
2064 __fio_sem_init(&rep->lock, FIO_SEM_LOCKED);
2068 verify_state_gen_name((char *) out.path, sizeof(out.path), name, me,
2070 tag = (uint64_t) (uintptr_t) rep;
2071 fio_net_queue_cmd(FIO_NET_CMD_SENDFILE, &out, sizeof(out), &tag,
2075 * Wait for the backend to receive the reply
2077 if (fio_sem_down_timeout(&rep->lock, 10000)) {
2078 log_err("fio: timed out waiting for reply\n");
2084 log_err("fio: failure on receiving state file %s: %s\n",
2085 out.path, strerror(rep->error));
2090 fio_net_queue_quit();
2095 * The format is verify_state_hdr, then thread_io_list. Verify
2096 * the header, and the thread_io_list checksum
2098 s = rep->data + sizeof(struct verify_state_hdr);
2099 if (verify_state_hdr(rep->data, s)) {
2105 * Don't need the header from now, copy just the thread_io_list
2108 rep->size -= sizeof(struct verify_state_hdr);
2109 data = malloc(rep->size);
2110 memcpy(data, s, rep->size);
2114 __fio_sem_remove(&rep->lock);
2119 static int fio_init_server_ip(void)
2121 struct sockaddr *addr;
2128 sk = socket(AF_INET6, SOCK_STREAM, 0);
2130 sk = socket(AF_INET, SOCK_STREAM, 0);
2133 log_err("fio: socket: %s\n", strerror(errno));
2138 if (setsockopt(sk, SOL_SOCKET, SO_REUSEADDR, (void *)&opt, sizeof(opt)) < 0) {
2139 log_err("fio: setsockopt(REUSEADDR): %s\n", strerror(errno));
2145 * Not fatal if fails, so just ignore it if that happens
2147 setsockopt(sk, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt));
2151 void *src = &saddr_in6.sin6_addr;
2153 addr = (struct sockaddr *) &saddr_in6;
2154 socklen = sizeof(saddr_in6);
2155 saddr_in6.sin6_family = AF_INET6;
2156 str = inet_ntop(AF_INET6, src, buf, sizeof(buf));
2158 void *src = &saddr_in.sin_addr;
2160 addr = (struct sockaddr *) &saddr_in;
2161 socklen = sizeof(saddr_in);
2162 saddr_in.sin_family = AF_INET;
2163 str = inet_ntop(AF_INET, src, buf, sizeof(buf));
2166 if (bind(sk, addr, socklen) < 0) {
2167 log_err("fio: bind: %s\n", strerror(errno));
2168 log_info("fio: failed with IPv%c %s\n", use_ipv6 ? '6' : '4', str);
2176 static int fio_init_server_sock(void)
2178 struct sockaddr_un addr;
2183 sk = socket(AF_UNIX, SOCK_STREAM, 0);
2185 log_err("fio: socket: %s\n", strerror(errno));
2191 memset(&addr, 0, sizeof(addr));
2192 addr.sun_family = AF_UNIX;
2193 strncpy(addr.sun_path, bind_sock, sizeof(addr.sun_path) - 1);
2195 len = sizeof(addr.sun_family) + strlen(bind_sock) + 1;
2197 if (bind(sk, (struct sockaddr *) &addr, len) < 0) {
2198 log_err("fio: bind: %s\n", strerror(errno));
2207 static int fio_init_server_connection(void)
2212 dprint(FD_NET, "starting server\n");
2215 sk = fio_init_server_ip();
2217 sk = fio_init_server_sock();
2222 memset(bind_str, 0, sizeof(bind_str));
2231 src = &saddr_in6.sin6_addr;
2234 src = &saddr_in.sin_addr;
2237 p = (char *) inet_ntop(af, src, bind_str, sizeof(bind_str));
2239 sprintf(port, ",%u", fio_net_port);
2243 strncpy(bind_str, port, sizeof(bind_str) - 1);
2245 strncpy(bind_str, bind_sock, sizeof(bind_str) - 1);
2247 log_info("fio: server listening on %s\n", bind_str);
2249 if (listen(sk, 4) < 0) {
2250 log_err("fio: listen: %s\n", strerror(errno));
2258 int fio_server_parse_host(const char *host, int ipv6, struct in_addr *inp,
2259 struct in6_addr *inp6)
2265 ret = inet_pton(AF_INET6, host, inp6);
2267 ret = inet_pton(AF_INET, host, inp);
2270 struct addrinfo *res, hints = {
2271 .ai_family = ipv6 ? AF_INET6 : AF_INET,
2272 .ai_socktype = SOCK_STREAM,
2275 ret = getaddrinfo(host, NULL, &hints, &res);
2277 log_err("fio: failed to resolve <%s> (%s)\n", host,
2283 memcpy(inp6, &((struct sockaddr_in6 *) res->ai_addr)->sin6_addr, sizeof(*inp6));
2285 memcpy(inp, &((struct sockaddr_in *) res->ai_addr)->sin_addr, sizeof(*inp));
2295 * Parse a host/ip/port string. Reads from 'str'.
2300 * *ptr is the host, *port is the port, inp is the destination.
2302 * *ptr is the host, *port is the port, inp6 is the dest, and *ipv6 is 1.
2303 * For local domain sockets:
2304 * *ptr is the filename, *is_sock is 1.
2306 int fio_server_parse_string(const char *str, char **ptr, bool *is_sock,
2307 int *port, struct in_addr *inp,
2308 struct in6_addr *inp6, int *ipv6)
2310 const char *host = str;
2316 *port = fio_net_port;
2319 if (!strncmp(str, "sock:", 5)) {
2320 *ptr = strdup(str + 5);
2327 * Is it ip:<ip or host>:port
2329 if (!strncmp(host, "ip:", 3))
2331 else if (!strncmp(host, "ip4:", 4))
2333 else if (!strncmp(host, "ip6:", 4)) {
2336 } else if (host[0] == ':') {
2337 /* String is :port */
2340 if (!lport || lport > 65535) {
2341 log_err("fio: bad server port %u\n", lport);
2344 /* no hostname given, we are done */
2350 * If no port seen yet, check if there's a last ',' at the end
2353 portp = strchr(host, ',');
2357 lport = atoi(portp);
2358 if (!lport || lport > 65535) {
2359 log_err("fio: bad server port %u\n", lport);
2371 *ptr = strdup(host);
2373 if (fio_server_parse_host(*ptr, *ipv6, inp, inp6)) {
2380 *port = fio_net_port;
2386 * Server arg should be one of:
2388 * sock:/path/to/socket
2392 * Where sock uses unix domain sockets, and ip binds the server to
2393 * a specific interface. If no arguments are given to the server, it
2394 * uses IP and binds to 0.0.0.0.
2397 static int fio_handle_server_arg(void)
2399 int port = fio_net_port;
2403 saddr_in.sin_addr.s_addr = htonl(INADDR_ANY);
2405 if (!fio_server_arg)
2408 ret = fio_server_parse_string(fio_server_arg, &bind_sock, &is_sock,
2409 &port, &saddr_in.sin_addr,
2410 &saddr_in6.sin6_addr, &use_ipv6);
2412 if (!is_sock && bind_sock) {
2418 fio_net_port = port;
2419 saddr_in.sin_port = htons(port);
2420 saddr_in6.sin6_port = htons(port);
2424 static void sig_int(int sig)
2430 static void set_sig_handlers(void)
2432 struct sigaction act = {
2433 .sa_handler = sig_int,
2434 .sa_flags = SA_RESTART,
2437 sigaction(SIGINT, &act, NULL);
2440 void fio_server_destroy_sk_key(void)
2442 pthread_key_delete(sk_out_key);
2445 int fio_server_create_sk_key(void)
2447 if (pthread_key_create(&sk_out_key, NULL)) {
2448 log_err("fio: can't create sk_out backend key\n");
2452 pthread_setspecific(sk_out_key, NULL);
2456 static int fio_server(void)
2460 dprint(FD_NET, "starting server\n");
2462 if (fio_handle_server_arg())
2465 sk = fio_init_server_connection();
2471 ret = accept_loop(sk);
2475 if (fio_server_arg) {
2476 free(fio_server_arg);
2477 fio_server_arg = NULL;
2485 void fio_server_got_signal(int signal)
2487 struct sk_out *sk_out = pthread_getspecific(sk_out_key);
2491 if (signal == SIGPIPE)
2494 log_info("\nfio: terminating on signal %d\n", signal);
2499 static int check_existing_pidfile(const char *pidfile)
2506 if (stat(pidfile, &sb))
2509 f = fopen(pidfile, "r");
2513 if (fread(buf, sb.st_size, 1, f) <= 0) {
2520 if (kill(pid, SIGCONT) < 0)
2521 return errno != ESRCH;
2526 static int write_pid(pid_t pid, const char *pidfile)
2530 fpid = fopen(pidfile, "w");
2532 log_err("fio: failed opening pid file %s\n", pidfile);
2536 fprintf(fpid, "%u\n", (unsigned int) pid);
2542 * If pidfile is specified, background us.
2544 int fio_start_server(char *pidfile)
2551 WSAStartup(MAKEWORD(2, 2), &wsd);
2555 return fio_server();
2557 if (check_existing_pidfile(pidfile)) {
2558 log_err("fio: pidfile %s exists and server appears alive\n",
2566 log_err("fio: failed server fork: %s\n", strerror(errno));
2570 ret = write_pid(pid, pidfile);
2576 openlog("fio", LOG_NDELAY|LOG_NOWAIT|LOG_PID, LOG_USER);
2578 close(STDIN_FILENO);
2579 close(STDOUT_FILENO);
2580 close(STDERR_FILENO);
2592 void fio_server_set_arg(const char *arg)
2594 fio_server_arg = strdup(arg);