Add more job info
[fio.git] / client.c
index 874f0aecd3b80ccd1b8bed646be783410a4e99e0..a84b92c4739ff7f83492b49d8567a70776893226 100644 (file)
--- a/client.c
+++ b/client.c
 #include <signal.h>
 
 #include "fio.h"
+#include "client.h"
 #include "server.h"
 #include "flist.h"
 #include "hash.h"
 
-struct fio_client {
-       struct flist_head list;
-       struct flist_head hash_list;
-       struct sockaddr_in addr;
-       struct sockaddr_un addr_un;
-       char *hostname;
-       int port;
-       int fd;
+static void fio_client_text_op(struct fio_client *client,
+               FILE *f, __u16 pdu_len, const char *buf)
+{
+       const char *name;
+       int fio_unused ret;
 
-       char *name;
+       name = client->name ? client->name : client->hostname;
 
-       int state;
-       int skip_newline;
-       int is_sock;
+       if (!client->skip_newline)
+               fprintf(f, "<%s> ", name);
+       ret = fwrite(buf, pdu_len, 1, f);
+       fflush(f);
+       client->skip_newline = strchr(buf, '\n') == NULL;
+}
 
-       uint16_t argc;
-       char **argv;
+static void handle_du(struct fio_client *client, struct fio_net_cmd *cmd);
+static void handle_ts(struct fio_net_cmd *cmd);
+static void handle_gs(struct fio_net_cmd *cmd);
+static void handle_eta(struct fio_client *client, struct fio_net_cmd *cmd);
+static void handle_probe(struct fio_client *client, struct fio_net_cmd *cmd);
+
+struct client_ops fio_client_ops = {
+       .text_op        = fio_client_text_op,
+       .disk_util      = handle_du,
+       .thread_status  = handle_ts,
+       .group_stats    = handle_gs,
+       .eta            = handle_eta,
+       .probe          = handle_probe,
 };
 
+static struct timeval eta_tv;
+
 enum {
        Client_created          = 0,
        Client_connected        = 1,
        Client_started          = 2,
-       Client_stopped          = 3,
-       Client_exited           = 4,
+       Client_running          = 3,
+       Client_stopped          = 4,
+       Client_exited           = 5,
 };
 
 static FLIST_HEAD(client_list);
+static FLIST_HEAD(eta_list);
+
+static FLIST_HEAD(arg_list);
+
+static struct thread_stat client_ts;
+static struct group_run_stats client_gs;
+static int sum_stat_clients;
+static int sum_stat_nr;
 
 #define FIO_CLIENT_HASH_BITS   7
 #define FIO_CLIENT_HASH_SZ     (1 << FIO_CLIENT_HASH_BITS)
 #define FIO_CLIENT_HASH_MASK   (FIO_CLIENT_HASH_SZ - 1)
 static struct flist_head client_hash[FIO_CLIENT_HASH_SZ];
 
-static int handle_client(struct fio_client *client);
-
 static void fio_client_add_hash(struct fio_client *client)
 {
        int bucket = hash_long(client->fd, FIO_CLIENT_HASH_BITS);
@@ -101,6 +122,11 @@ static void remove_client(struct fio_client *client)
 
        fio_client_remove_hash(client);
 
+       if (!flist_empty(&client->eta_list)) {
+               flist_del_init(&client->eta_list);
+               fio_client_dec_jobs_eta(client->eta_in_flight, display_thread_status);
+       }
+
        free(client->hostname);
        if (client->argv)
                free(client->argv);
@@ -109,6 +135,7 @@ static void remove_client(struct fio_client *client)
 
        free(client);
        nr_clients--;
+       sum_stat_clients--;
 }
 
 static void __fio_client_add_cmd_option(struct fio_client *client,
@@ -125,26 +152,98 @@ static void __fio_client_add_cmd_option(struct fio_client *client,
 void fio_client_add_cmd_option(void *cookie, const char *opt)
 {
        struct fio_client *client = cookie;
+       struct flist_head *entry;
 
        if (!client || !opt)
                return;
 
        __fio_client_add_cmd_option(client, opt);
+
+       /*
+        * Duplicate arguments to shared client group
+        */
+       flist_for_each(entry, &arg_list) {
+               client = flist_entry(entry, struct fio_client, arg_list);
+
+               __fio_client_add_cmd_option(client, opt);
+       }
+}
+
+struct fio_client *fio_client_add_explicit(const char *hostname, int type,
+                                          int port)
+{
+       struct fio_client *client;
+
+       client = malloc(sizeof(*client));
+       memset(client, 0, sizeof(*client));
+
+       INIT_FLIST_HEAD(&client->list);
+       INIT_FLIST_HEAD(&client->hash_list);
+       INIT_FLIST_HEAD(&client->arg_list);
+       INIT_FLIST_HEAD(&client->eta_list);
+       INIT_FLIST_HEAD(&client->cmd_list);
+
+       client->hostname = strdup(hostname);
+
+       if (type == Fio_client_socket)
+               client->is_sock = 1;
+       else {
+               int ipv6;
+
+               ipv6 = type == Fio_client_ipv6;
+               if (fio_server_parse_host(hostname, &ipv6,
+                                               &client->addr.sin_addr,
+                                               &client->addr6.sin6_addr))
+                       goto err;
+
+               client->port = port;
+       }
+
+       client->fd = -1;
+
+       __fio_client_add_cmd_option(client, "fio");
+
+       flist_add(&client->list, &client_list);
+       nr_clients++;
+       dprint(FD_NET, "client: added <%s>\n", client->hostname);
+       return client;
+err:
+       free(client);
+       return NULL;
 }
 
 int fio_client_add(const char *hostname, void **cookie)
 {
+       struct fio_client *existing = *cookie;
        struct fio_client *client;
 
+       if (existing) {
+               /*
+                * We always add our "exec" name as the option, hence 1
+                * means empty.
+                */
+               if (existing->argc == 1)
+                       flist_add_tail(&existing->arg_list, &arg_list);
+               else {
+                       while (!flist_empty(&arg_list))
+                               flist_del_init(arg_list.next);
+               }
+       }
+
        client = malloc(sizeof(*client));
        memset(client, 0, sizeof(*client));
 
        INIT_FLIST_HEAD(&client->list);
        INIT_FLIST_HEAD(&client->hash_list);
+       INIT_FLIST_HEAD(&client->arg_list);
+       INIT_FLIST_HEAD(&client->eta_list);
+       INIT_FLIST_HEAD(&client->cmd_list);
 
        if (fio_server_parse_string(hostname, &client->hostname,
                                        &client->is_sock, &client->port,
-                                       &client->addr.sin_addr))
+                                       &client->addr.sin_addr,
+                                       &client->addr6.sin6_addr,
+                                       &client->ipv6))
                return -1;
 
        client->fd = -1;
@@ -160,18 +259,31 @@ int fio_client_add(const char *hostname, void **cookie)
 
 static int fio_client_connect_ip(struct fio_client *client)
 {
-       int fd;
-
-       client->addr.sin_family = AF_INET;
-       client->addr.sin_port = htons(client->port);
+       struct sockaddr *addr;
+       fio_socklen_t socklen;
+       int fd, domain;
+
+       if (client->ipv6) {
+               client->addr6.sin6_family = AF_INET6;
+               client->addr6.sin6_port = htons(client->port);
+               domain = AF_INET6;
+               addr = (struct sockaddr *) &client->addr6;
+               socklen = sizeof(client->addr6);
+       } else {
+               client->addr.sin_family = AF_INET;
+               client->addr.sin_port = htons(client->port);
+               domain = AF_INET;
+               addr = (struct sockaddr *) &client->addr;
+               socklen = sizeof(client->addr);
+       }
 
-       fd = socket(AF_INET, SOCK_STREAM, 0);
+       fd = socket(domain, SOCK_STREAM, 0);
        if (fd < 0) {
                log_err("fio: socket: %s\n", strerror(errno));
                return -1;
        }
 
-       if (connect(fd, (struct sockaddr *) &client->addr, sizeof(client->addr)) < 0) {
+       if (connect(fd, addr, socklen) < 0) {
                log_err("fio: connect: %s\n", strerror(errno));
                log_err("fio: failed to connect to %s:%u\n", client->hostname,
                                                                client->port);
@@ -219,6 +331,8 @@ static int fio_client_connect(struct fio_client *client)
        else
                fd = fio_client_connect_ip(client);
 
+       dprint(FD_NET, "client: %s connected %d\n", client->hostname, fd);
+
        if (fd < 0)
                return 1;
 
@@ -238,7 +352,7 @@ void fio_clients_terminate(void)
        flist_for_each(entry, &client_list) {
                client = flist_entry(entry, struct fio_client, list);
 
-               fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_QUIT, 0);
+               fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_QUIT, 0, NULL);
        }
 }
 
@@ -267,8 +381,7 @@ static void probe_client(struct fio_client *client)
 {
        dprint(FD_NET, "client: send probe\n");
 
-       fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_PROBE, 0);
-       handle_client(client);
+       fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_PROBE, 0, &client->cmd_list);
 }
 
 static int send_client_cmd_line(struct fio_client *client)
@@ -276,17 +389,22 @@ static int send_client_cmd_line(struct fio_client *client)
        struct cmd_single_line_pdu *cslp;
        struct cmd_line_pdu *clp;
        unsigned long offset;
+       unsigned int *lens;
        void *pdu;
        size_t mem;
        int i, ret;
 
        dprint(FD_NET, "client: send cmdline %d\n", client->argc);
 
+       lens = malloc(client->argc * sizeof(unsigned int));
+
        /*
         * Find out how much mem we need
         */
-       for (i = 0, mem = 0; i < client->argc; i++)
-               mem += strlen(client->argv[i]) + 1;
+       for (i = 0, mem = 0; i < client->argc; i++) {
+               lens[i] = strlen(client->argv[i]) + 1;
+               mem += lens[i];
+       }
 
        /*
         * We need one cmd_line_pdu, and argc number of cmd_single_line_pdu
@@ -298,7 +416,7 @@ static int send_client_cmd_line(struct fio_client *client)
        offset = sizeof(*clp);
 
        for (i = 0; i < client->argc; i++) {
-               uint16_t arg_len = strlen(client->argv[i]) + 1;
+               uint16_t arg_len = lens[i];
 
                cslp = pdu + offset;
                strcpy((char *) cslp->text, client->argv[i]);
@@ -306,8 +424,9 @@ static int send_client_cmd_line(struct fio_client *client)
                offset += sizeof(*cslp) + arg_len;
        }
 
+       free(lens);
        clp->lines = cpu_to_le16(client->argc);
-       ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOBLINE, pdu, mem);
+       ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOBLINE, pdu, mem, 0);
        free(pdu);
        return ret;
 }
@@ -318,6 +437,11 @@ int fio_clients_connect(void)
        struct flist_head *entry, *tmp;
        int ret;
 
+#ifdef WIN32
+       WSADATA wsd;
+       WSAStartup(MAKEWORD(2,2), &wsd);
+#endif
+
        dprint(FD_NET, "client: connect all\n");
 
        client_signal_handler();
@@ -390,7 +514,8 @@ static int fio_client_send_ini(struct fio_client *client, const char *filename)
                return 1;
        }
 
-       ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOB, buf, sb.st_size);
+       client->sent_job = 1;
+       ret = fio_net_send_cmd(client->fd, FIO_NET_CMD_JOB, buf, sb.st_size, 0);
        free(buf);
        close(fd);
        return ret;
@@ -406,6 +531,8 @@ int fio_clients_send_ini(const char *filename)
 
                if (fio_client_send_ini(client, filename))
                        remove_client(client);
+
+               client->sent_job = 1;
        }
 
        return !nr_clients;
@@ -514,6 +641,20 @@ static void handle_ts(struct fio_net_cmd *cmd)
        convert_gs(&p->rs, &p->rs);
 
        show_thread_status(&p->ts, &p->rs);
+
+       if (sum_stat_clients == 1)
+               return;
+
+       sum_thread_stats(&client_ts, &p->ts, sum_stat_nr);
+       sum_group_stats(&client_gs, &p->rs);
+
+       client_ts.members++;
+       client_ts.groupid = p->ts.groupid;
+
+       if (++sum_stat_nr == sum_stat_clients) {
+               strcpy(client_ts.name, "All clients");
+               show_thread_status(&client_ts, &client_gs);
+       }
 }
 
 static void handle_gs(struct fio_net_cmd *cmd)
@@ -524,35 +665,154 @@ static void handle_gs(struct fio_net_cmd *cmd)
        show_group_stats(gs);
 }
 
-static void handle_eta(struct fio_net_cmd *cmd)
+static void convert_agg(struct disk_util_agg *agg)
+{
+       int i;
+
+       for (i = 0; i < 2; i++) {
+               agg->ios[i]     = le32_to_cpu(agg->ios[i]);
+               agg->merges[i]  = le32_to_cpu(agg->merges[i]);
+               agg->sectors[i] = le64_to_cpu(agg->sectors[i]);
+               agg->ticks[i]   = le32_to_cpu(agg->ticks[i]);
+       }
+
+       agg->io_ticks           = le32_to_cpu(agg->io_ticks);
+       agg->time_in_queue      = le32_to_cpu(agg->time_in_queue);
+       agg->slavecount         = le32_to_cpu(agg->slavecount);
+       agg->max_util.u.f       = fio_uint64_to_double(__le64_to_cpu(agg->max_util.u.i));
+}
+
+static void convert_dus(struct disk_util_stat *dus)
+{
+       int i;
+
+       for (i = 0; i < 2; i++) {
+               dus->ios[i]     = le32_to_cpu(dus->ios[i]);
+               dus->merges[i]  = le32_to_cpu(dus->merges[i]);
+               dus->sectors[i] = le64_to_cpu(dus->sectors[i]);
+               dus->ticks[i]   = le32_to_cpu(dus->ticks[i]);
+       }
+
+       dus->io_ticks           = le32_to_cpu(dus->io_ticks);
+       dus->time_in_queue      = le32_to_cpu(dus->time_in_queue);
+       dus->msec               = le64_to_cpu(dus->msec);
+}
+
+static void handle_du(struct fio_client *client, struct fio_net_cmd *cmd)
+{
+       struct cmd_du_pdu *du = (struct cmd_du_pdu *) cmd->payload;
+
+       convert_dus(&du->dus);
+       convert_agg(&du->agg);
+
+       if (!client->disk_stats_shown) {
+               client->disk_stats_shown = 1;
+               log_info("\nDisk stats (read/write):\n");
+       }
+
+       print_disk_util(&du->dus, &du->agg, terse_output);
+}
+
+void fio_client_convert_jobs_eta(struct jobs_eta *je)
 {
-       struct jobs_eta *je = (struct jobs_eta *) cmd->payload;
        int i;
 
        je->nr_running          = le32_to_cpu(je->nr_running);
        je->nr_ramp             = le32_to_cpu(je->nr_ramp);
        je->nr_pending          = le32_to_cpu(je->nr_pending);
        je->files_open          = le32_to_cpu(je->files_open);
-       je->m_rate              = le32_to_cpu(je->m_rate);
-       je->t_rate              = le32_to_cpu(je->t_rate);
-       je->m_iops              = le32_to_cpu(je->m_iops);
-       je->t_iops              = le32_to_cpu(je->t_iops);
 
        for (i = 0; i < 2; i++) {
+               je->m_rate[i]           = le32_to_cpu(je->m_rate[i]);
+               je->t_rate[i]           = le32_to_cpu(je->t_rate[i]);
+               je->m_iops[i]           = le32_to_cpu(je->m_iops[i]);
+               je->t_iops[i]           = le32_to_cpu(je->t_iops[i]);
                je->rate[i]     = le32_to_cpu(je->rate[i]);
                je->iops[i]     = le32_to_cpu(je->iops[i]);
        }
 
-       je->elapsed_sec         = le32_to_cpu(je->nr_running);
+       je->elapsed_sec         = le64_to_cpu(je->elapsed_sec);
        je->eta_sec             = le64_to_cpu(je->eta_sec);
+}
+
+void fio_client_sum_jobs_eta(struct jobs_eta *dst, struct jobs_eta *je)
+{
+       int i;
 
-       display_thread_status(je);
+       dst->nr_running         += je->nr_running;
+       dst->nr_ramp            += je->nr_ramp;
+       dst->nr_pending         += je->nr_pending;
+       dst->files_open         += je->files_open;
+
+       for (i = 0; i < 2; i++) {
+               dst->m_rate[i]  += je->m_rate[i];
+               dst->t_rate[i]  += je->t_rate[i];
+               dst->m_iops[i]  += je->m_iops[i];
+               dst->t_iops[i]  += je->t_iops[i];
+               dst->rate[i]    += je->rate[i];
+               dst->iops[i]    += je->iops[i];
+       }
+
+       dst->elapsed_sec        += je->elapsed_sec;
+
+       if (je->eta_sec > dst->eta_sec)
+               dst->eta_sec = je->eta_sec;
+}
+
+void fio_client_dec_jobs_eta(struct client_eta *eta, void (*fn)(struct jobs_eta *))
+{
+       if (!--eta->pending) {
+               fn(&eta->eta);
+               free(eta);
+       }
+}
+
+static void remove_reply_cmd(struct fio_client *client, struct fio_net_cmd *cmd)
+{
+       struct fio_net_int_cmd *icmd = NULL;
+       struct flist_head *entry;
+
+       flist_for_each(entry, &client->cmd_list) {
+               icmd = flist_entry(entry, struct fio_net_int_cmd, list);
+
+               if (cmd->tag == (uintptr_t) icmd)
+                       break;
+
+               icmd = NULL;
+       }
+
+       if (!icmd) {
+               log_err("fio: client: unable to find matching tag\n");
+               return;
+       }
+
+       flist_del(&icmd->list);
+       cmd->tag = icmd->saved_tag;
+       free(icmd);
+}
+
+static void handle_eta(struct fio_client *client, struct fio_net_cmd *cmd)
+{
+       struct jobs_eta *je = (struct jobs_eta *) cmd->payload;
+       struct client_eta *eta = (struct client_eta *) (uintptr_t) cmd->tag;
+
+       dprint(FD_NET, "client: got eta tag %p, %d\n", eta, eta->pending);
+
+       assert(client->eta_in_flight == eta);
+
+       client->eta_in_flight = NULL;
+       flist_del_init(&client->eta_list);
+
+       fio_client_convert_jobs_eta(je);
+       fio_client_sum_jobs_eta(&eta->eta, je);
+       fio_client_dec_jobs_eta(eta, display_thread_status);
 }
 
 static void handle_probe(struct fio_client *client, struct fio_net_cmd *cmd)
 {
        struct cmd_probe_pdu *probe = (struct cmd_probe_pdu *) cmd->payload;
        const char *os, *arch;
+       char bit[16];
 
        os = fio_get_os_string(probe->os);
        if (!os)
@@ -562,15 +822,36 @@ static void handle_probe(struct fio_client *client, struct fio_net_cmd *cmd)
        if (!arch)
                os = "unknown";
 
-       log_info("hostname=%s, be=%u, os=%s, arch=%s, fio=%u.%u.%u\n",
-               probe->hostname, probe->bigendian, os, arch, probe->fio_major,
-               probe->fio_minor, probe->fio_patch);
+       sprintf(bit, "%d-bit", probe->bpp * 8);
+
+       log_info("hostname=%s, be=%u, %s, os=%s, arch=%s, fio=%u.%u.%u\n",
+               probe->hostname, probe->bigendian, bit, os, arch,
+               probe->fio_major, probe->fio_minor, probe->fio_patch);
 
        if (!client->name)
                client->name = strdup((char *) probe->hostname);
 }
 
-static int handle_client(struct fio_client *client)
+static void handle_start(struct fio_client *client, struct fio_net_cmd *cmd)
+{
+       struct cmd_start_pdu *pdu = (struct cmd_start_pdu *) cmd->payload;
+
+       client->state = Client_started;
+       client->jobs = le32_to_cpu(pdu->jobs);
+}
+
+static void handle_stop(struct fio_client *client, struct fio_net_cmd *cmd)
+{
+       struct cmd_end_pdu *pdu = (struct cmd_end_pdu *) cmd->payload;
+
+       client->state = Client_stopped;
+       client->error = le32_to_cpu(pdu->error);
+
+       if (client->error)
+               log_info("client <%s>: exited with error %d\n", client->hostname, client->error);
+}
+
+int fio_handle_client(struct fio_client *client, struct client_ops *ops)
 {
        struct fio_net_cmd *cmd;
 
@@ -580,55 +861,63 @@ static int handle_client(struct fio_client *client)
        if (!cmd)
                return 0;
 
-       dprint(FD_NET, "client: got cmd op %d from %s\n",
-                                       cmd->opcode, client->hostname);
+       dprint(FD_NET, "client: got cmd op %s from %s\n",
+                               fio_server_op(cmd->opcode), client->hostname);
 
        switch (cmd->opcode) {
        case FIO_NET_CMD_QUIT:
+               if (ops->quit)
+                       ops->quit(client);
                remove_client(client);
                free(cmd);
                break;
        case FIO_NET_CMD_TEXT: {
                const char *buf = (const char *) cmd->payload;
-               const char *name;
-               int fio_unused ret;
-
-               name = client->name ? client->name : client->hostname;
-
-               if (!client->skip_newline)
-                       fprintf(f_out, "<%s> ", name);
-               ret = fwrite(buf, cmd->pdu_len, 1, f_out);
-               fflush(f_out);
-               client->skip_newline = strchr(buf, '\n') == NULL;
+               ops->text_op(client, f_out, cmd->pdu_len, buf);
                free(cmd);
                break;
                }
+       case FIO_NET_CMD_DU:
+               ops->disk_util(client, cmd);
+               free(cmd);
+               break;
        case FIO_NET_CMD_TS:
-               handle_ts(cmd);
+               ops->thread_status(cmd);
                free(cmd);
                break;
        case FIO_NET_CMD_GS:
-               handle_gs(cmd);
+               ops->group_stats(cmd);
                free(cmd);
                break;
        case FIO_NET_CMD_ETA:
-               handle_eta(cmd);
+               remove_reply_cmd(client, cmd);
+               ops->eta(client, cmd);
                free(cmd);
                break;
        case FIO_NET_CMD_PROBE:
-               handle_probe(client, cmd);
+               remove_reply_cmd(client, cmd);
+               ops->probe(client, cmd);
+               free(cmd);
+               break;
+       case FIO_NET_CMD_RUN:
+               client->state = Client_running;
                free(cmd);
                break;
        case FIO_NET_CMD_START:
-               client->state = Client_started;
+               handle_start(client, cmd);
                free(cmd);
                break;
        case FIO_NET_CMD_STOP:
-               client->state = Client_stopped;
+               handle_stop(client, cmd);
+               free(cmd);
+               break;
+       case FIO_NET_CMD_ADD_JOB:
+               if (ops->add_job)
+                       ops->add_job(client, cmd);
                free(cmd);
                break;
        default:
-               log_err("fio: unknown client op: %d\n", cmd->opcode);
+               log_err("fio: unknown client op: %s\n", fio_server_op(cmd->opcode));
                free(cmd);
                break;
        }
@@ -636,28 +925,140 @@ static int handle_client(struct fio_client *client)
        return 1;
 }
 
-int fio_handle_clients(void)
+static void request_client_etas(void)
 {
        struct fio_client *client;
        struct flist_head *entry;
+       struct client_eta *eta;
+       int skipped = 0;
+
+       dprint(FD_NET, "client: request eta (%d)\n", nr_clients);
+
+       eta = malloc(sizeof(*eta));
+       memset(&eta->eta, 0, sizeof(eta->eta));
+       eta->pending = nr_clients;
+
+       flist_for_each(entry, &client_list) {
+               client = flist_entry(entry, struct fio_client, list);
+
+               if (!flist_empty(&client->eta_list)) {
+                       skipped++;
+                       continue;
+               }
+               if (client->state != Client_running)
+                       continue;
+
+               assert(!client->eta_in_flight);
+               flist_add_tail(&client->eta_list, &eta_list);
+               client->eta_in_flight = eta;
+               fio_net_send_simple_cmd(client->fd, FIO_NET_CMD_SEND_ETA,
+                                       (uintptr_t) eta, &client->cmd_list);
+       }
+
+       while (skipped--)
+               fio_client_dec_jobs_eta(eta, display_thread_status);
+
+       dprint(FD_NET, "client: requested eta tag %p\n", eta);
+}
+
+static int client_check_cmd_timeout(struct fio_client *client,
+                                   struct timeval *now)
+{
+       struct fio_net_int_cmd *cmd;
+       struct flist_head *entry, *tmp;
+       int ret = 0;
+
+       flist_for_each_safe(entry, tmp, &client->cmd_list) {
+               cmd = flist_entry(entry, struct fio_net_int_cmd, list);
+
+               if (mtime_since(&cmd->tv, now) < FIO_NET_CLIENT_TIMEOUT)
+                       continue;
+
+               log_err("fio: client %s, timeout on cmd %s\n", client->hostname,
+                                               fio_server_op(cmd->cmd.opcode));
+               flist_del(&cmd->list);
+               free(cmd);
+               ret = 1;
+       }
+
+       return flist_empty(&client->cmd_list) && ret;
+}
+
+static int fio_client_timed_out(void)
+{
+       struct fio_client *client;
+       struct flist_head *entry, *tmp;
+       struct timeval tv;
+       int ret = 0;
+
+       gettimeofday(&tv, NULL);
+
+       flist_for_each_safe(entry, tmp, &client_list) {
+               client = flist_entry(entry, struct fio_client, list);
+
+               if (flist_empty(&client->cmd_list))
+                       continue;
+
+               if (!client_check_cmd_timeout(client, &tv))
+                       continue;
+
+               log_err("fio: client %s timed out\n", client->hostname);
+               remove_client(client);
+               ret = 1;
+       }
+
+       return ret;
+}
+
+int fio_handle_clients(struct client_ops *ops)
+{
        struct pollfd *pfds;
-       int i, ret = 0;
+       int i, ret = 0, retval = 0;
+
+       gettimeofday(&eta_tv, NULL);
 
        pfds = malloc(nr_clients * sizeof(struct pollfd));
 
+       sum_stat_clients = nr_clients;
+       init_thread_stat(&client_ts);
+       init_group_run_stat(&client_gs);
+
        while (!exit_backend && nr_clients) {
+               struct flist_head *entry, *tmp;
+               struct fio_client *client;
+
                i = 0;
-               flist_for_each(entry, &client_list) {
+               flist_for_each_safe(entry, tmp, &client_list) {
                        client = flist_entry(entry, struct fio_client, list);
 
+                       if (!client->sent_job && !ops->stay_connected &&
+                           flist_empty(&client->cmd_list)) {
+                               remove_client(client);
+                               continue;
+                       }
+
                        pfds[i].fd = client->fd;
                        pfds[i].events = POLLIN;
                        i++;
                }
 
+               if (!nr_clients)
+                       break;
+
                assert(i == nr_clients);
 
                do {
+                       struct timeval tv;
+
+                       gettimeofday(&tv, NULL);
+                       if (mtime_since(&eta_tv, &tv) >= 900) {
+                               request_client_etas();
+                               memcpy(&eta_tv, &tv, sizeof(tv));
+
+                               if (fio_client_timed_out())
+                                       break;
+                       }
+
                        ret = poll(pfds, nr_clients, 100);
                        if (ret < 0) {
                                if (errno == EINTR)
@@ -677,14 +1078,16 @@ int fio_handle_clients(void)
                                log_err("fio: unknown client fd %d\n", pfds[i].fd);
                                continue;
                        }
-                       if (!handle_client(client)) {
+                       if (!fio_handle_client(client, ops)) {
                                log_info("client: host=%s disconnected\n",
                                                client->hostname);
                                remove_client(client);
-                       }
+                               retval = 1;
+                       } else if (client->error)
+                               retval = 1;
                }
        }
 
        free(pfds);
-       return 0;
+       return retval;
 }