server: unlink sock file if interrupted early
[fio.git] / server.c
index c504e872027a550cad5a8c159b7a2dba65dda7de..d72835b72690148fc78662fb99599465e2836a55 100644 (file)
--- a/server.c
+++ b/server.c
 #include <sys/socket.h>
 #include <sys/stat.h>
 #include <sys/un.h>
+#include <sys/uio.h>
 #include <netinet/in.h>
 #include <arpa/inet.h>
 #include <netdb.h>
 #include <syslog.h>
 #include <signal.h>
+#ifdef CONFIG_ZLIB
+#include <zlib.h>
+#endif
 
 #include "fio.h"
 #include "server.h"
 #include "crc/crc16.h"
 #include "lib/ieee754.h"
 
-#include "fio_version.h"
-
 int fio_net_port = FIO_NET_PORT;
 
 int exit_backend = 0;
@@ -33,8 +35,21 @@ static char *fio_server_arg;
 static char *bind_sock;
 static struct sockaddr_in saddr_in;
 static struct sockaddr_in6 saddr_in6;
-static int first_cmd_check;
 static int use_ipv6;
+#ifdef CONFIG_ZLIB
+static unsigned int has_zlib = 1;
+#else
+static unsigned int has_zlib = 0;
+#endif
+static unsigned int use_zlib;
+
+struct fio_fork_item {
+       struct flist_head list;
+       int exitval;
+       int signal;
+       int exited;
+       pid_t pid;
+};
 
 static const char *fio_server_ops[FIO_NET_CMD_NR] = {
        "",
@@ -51,8 +66,10 @@ static const char *fio_server_ops[FIO_NET_CMD_NR] = {
        "START",
        "STOP",
        "DISK_UTIL",
-       "RUN",
+       "SERVER_START",
        "ADD_JOB",
+       "CMD_RUN"
+       "CMD_IOLOG",
 };
 
 const char *fio_server_op(unsigned int op)
@@ -66,19 +83,40 @@ const char *fio_server_op(unsigned int op)
        return buf;
 }
 
-int fio_send_data(int sk, const void *p, unsigned int len)
+static ssize_t iov_total_len(const struct iovec *iov, int count)
 {
-       assert(len <= sizeof(struct fio_net_cmd) + FIO_SERVER_MAX_PDU);
+       ssize_t ret = 0;
 
-       do {
-               int ret = send(sk, p, len, 0);
+       while (count--) {
+               ret += iov->iov_len;
+               iov++;
+       }
 
+       return ret;
+}
+
+static int fio_sendv_data(int sk, struct iovec *iov, int count)
+{
+       ssize_t total_len = iov_total_len(iov, count);
+       ssize_t ret;
+
+       do {
+               ret = writev(sk, iov, count);
                if (ret > 0) {
-                       len -= ret;
-                       if (!len)
+                       total_len -= ret;
+                       if (!total_len)
                                break;
-                       p += ret;
-                       continue;
+
+                       while (ret) {
+                               if (ret >= iov->iov_len) {
+                                       ret -= iov->iov_len;
+                                       iov++;
+                                       continue;
+                               }
+                               iov->iov_base += ret;
+                               iov->iov_len -= ret;
+                               ret = 0;
+                       }
                } else if (!ret)
                        break;
                else if (errno == EAGAIN || errno == EINTR)
@@ -87,12 +125,24 @@ int fio_send_data(int sk, const void *p, unsigned int len)
                        break;
        } while (!exit_backend);
 
-       if (!len)
+       if (!total_len)
                return 0;
 
+       if (errno)
+               return -errno;
+
        return 1;
 }
 
+int fio_send_data(int sk, const void *p, unsigned int len)
+{
+       struct iovec iov = { .iov_base = (void *) p, .iov_len = len };
+
+       assert(len <= sizeof(struct fio_net_cmd) + FIO_SERVER_MAX_FRAGMENT_PDU);
+
+       return fio_sendv_data(sk, &iov, 1);
+}
+
 int fio_recv_data(int sk, void *p, unsigned int len)
 {
        do {
@@ -146,7 +196,7 @@ static int verify_convert_cmd(struct fio_net_cmd *cmd)
                return 1;
        }
 
-       if (cmd->pdu_len > FIO_SERVER_MAX_PDU) {
+       if (cmd->pdu_len > FIO_SERVER_MAX_FRAGMENT_PDU) {
                log_err("fio: command payload too large: %u\n", cmd->pdu_len);
                return 1;
        }
@@ -227,11 +277,13 @@ struct fio_net_cmd *fio_net_recv_cmd(int sk)
                                struct cmd_text_pdu *pdu = (struct cmd_text_pdu *) cmdret->payload;
                                char *buf = (char *) pdu->buf;
 
-                               buf[pdu->buf_len ] = '\0';
+                               buf[pdu->buf_len] = '\0';
                        } else if (cmdret->opcode == FIO_NET_CMD_JOB) {
-                               char *buf = (char *) cmdret->payload;
+                               struct cmd_job_pdu *pdu = (struct cmd_job_pdu *) cmdret->payload;
+                               char *buf = (char *) pdu->buf;
+                               int len = le32_to_cpu(pdu->buf_len);
 
-                               buf[cmdret->pdu_len ] = '\0';
+                               buf[len] = '\0';
                        }
                }
 
@@ -242,28 +294,68 @@ struct fio_net_cmd *fio_net_recv_cmd(int sk)
        return cmdret;
 }
 
-void fio_net_cmd_crc(struct fio_net_cmd *cmd)
+static void add_reply(uint64_t tag, struct flist_head *list)
+{
+       struct fio_net_cmd_reply *reply;
+
+       reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
+       flist_add_tail(&reply->list, list);
+}
+
+static uint64_t alloc_reply(uint64_t tag, uint16_t opcode)
+{
+       struct fio_net_cmd_reply *reply;
+
+       reply = calloc(1, sizeof(*reply));
+       INIT_FLIST_HEAD(&reply->list);
+       gettimeofday(&reply->tv, NULL);
+       reply->saved_tag = tag;
+       reply->opcode = opcode;
+
+       return (uintptr_t) reply;
+}
+
+static void free_reply(uint64_t tag)
+{
+       struct fio_net_cmd_reply *reply;
+
+       reply = (struct fio_net_cmd_reply *) (uintptr_t) tag;
+       free(reply);
+}
+
+void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
 {
        uint32_t pdu_len;
 
        cmd->cmd_crc16 = __cpu_to_le16(fio_crc16(cmd, FIO_NET_CMD_CRC_SZ));
 
        pdu_len = le32_to_cpu(cmd->pdu_len);
-       if (pdu_len)
-               cmd->pdu_crc16 = __cpu_to_le16(fio_crc16(cmd->payload, pdu_len));
+       cmd->pdu_crc16 = __cpu_to_le16(fio_crc16(pdu, pdu_len));
+}
+
+void fio_net_cmd_crc(struct fio_net_cmd *cmd)
+{
+       fio_net_cmd_crc_pdu(cmd, cmd->payload);
 }
 
 int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
-                    uint64_t tag)
+                    uint64_t *tagptr, struct flist_head *list)
 {
        struct fio_net_cmd *cmd = NULL;
        size_t this_len, cur_len = 0;
+       uint64_t tag;
        int ret;
 
+       if (list) {
+               assert(tagptr);
+               tag = *tagptr = alloc_reply(*tagptr, opcode);
+       } else
+               tag = tagptr ? *tagptr : 0;
+
        do {
                this_len = size;
-               if (this_len > FIO_SERVER_MAX_PDU)
-                       this_len = FIO_SERVER_MAX_PDU;
+               if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
+                       this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
 
                if (!cmd || cur_len < sizeof(*cmd) + this_len) {
                        if (cmd)
@@ -285,6 +377,13 @@ int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
                buf += this_len;
        } while (!ret && size);
 
+       if (list) {
+               if (ret)
+                       free_reply(tag);
+               else
+                       add_reply(tag, list);
+       }
+
        if (cmd)
                free(cmd);
 
@@ -308,61 +407,174 @@ static int fio_net_send_simple_stack_cmd(int sk, uint16_t opcode, uint64_t tag)
 int fio_net_send_simple_cmd(int sk, uint16_t opcode, uint64_t tag,
                            struct flist_head *list)
 {
-       struct fio_net_int_cmd *cmd;
        int ret;
 
-       if (!list)
-               return fio_net_send_simple_stack_cmd(sk, opcode, tag);
-
-       cmd = malloc(sizeof(*cmd));
-
-       fio_init_net_cmd(&cmd->cmd, opcode, NULL, 0, (uintptr_t) cmd);
-       fio_net_cmd_crc(&cmd->cmd);
+       if (list)
+               tag = alloc_reply(tag, opcode);
 
-       INIT_FLIST_HEAD(&cmd->list);
-       gettimeofday(&cmd->tv, NULL);
-       cmd->saved_tag = tag;
-
-       ret = fio_send_data(sk, &cmd->cmd, sizeof(cmd->cmd));
+       ret = fio_net_send_simple_stack_cmd(sk, opcode, tag);
        if (ret) {
-               free(cmd);
+               if (list)
+                       free_reply(tag);
+
                return ret;
        }
 
-       flist_add_tail(&cmd->list, list);
+       if (list)
+               add_reply(tag, list);
+
        return 0;
 }
 
-static int fio_server_send_quit_cmd(void)
+int fio_net_send_quit(int sk)
 {
        dprint(FD_NET, "server: sending quit\n");
-       return fio_net_send_simple_cmd(server_fd, FIO_NET_CMD_QUIT, 0, NULL);
+
+       return fio_net_send_simple_cmd(sk, FIO_NET_CMD_QUIT, 0, NULL);
 }
 
-static int handle_job_cmd(struct fio_net_cmd *cmd)
+static int fio_net_send_ack(int sk, struct fio_net_cmd *cmd, int error,
+                           int signal)
 {
-       char *buf = (char *) cmd->payload;
-       struct cmd_start_pdu spdu;
        struct cmd_end_pdu epdu;
-       int ret;
+       uint64_t tag = 0;
 
-       if (parse_jobs_ini(buf, 1, 0)) {
-               fio_server_send_quit_cmd();
-               return -1;
+       if (cmd)
+               tag = cmd->tag;
+
+       epdu.error = __cpu_to_le32(error);
+       epdu.signal = __cpu_to_le32(signal);
+       return fio_net_send_cmd(sk, FIO_NET_CMD_STOP, &epdu, sizeof(epdu), &tag, NULL);
+}
+
+int fio_net_send_stop(int sk, int error, int signal)
+{
+       dprint(FD_NET, "server: sending stop (%d, %d)\n", error, signal);
+       return fio_net_send_ack(sk, NULL, error, signal);
+}
+
+static void fio_server_add_fork_item(pid_t pid, struct flist_head *list)
+{
+       struct fio_fork_item *ffi;
+
+       ffi = malloc(sizeof(*ffi));
+       ffi->exitval = 0;
+       ffi->signal = 0;
+       ffi->exited = 0;
+       ffi->pid = pid;
+       flist_add_tail(&ffi->list, list);
+}
+
+static void fio_server_add_conn_pid(struct flist_head *conn_list, pid_t pid)
+{
+       dprint(FD_NET, "server: forked off connection job (pid=%u)\n", (int) pid);
+       fio_server_add_fork_item(pid, conn_list);
+}
+
+static void fio_server_add_job_pid(struct flist_head *job_list, pid_t pid)
+{
+       dprint(FD_NET, "server: forked off job job (pid=%u)\n", (int) pid);
+       fio_server_add_fork_item(pid, job_list);
+}
+
+static void fio_server_check_fork_item(struct fio_fork_item *ffi)
+{
+       int ret, status;
+
+       ret = waitpid(ffi->pid, &status, WNOHANG);
+       if (ret < 0) {
+               if (errno == ECHILD) {
+                       log_err("fio: connection pid %u disappeared\n", (int) ffi->pid);
+                       ffi->exited = 1;
+               } else
+                       log_err("fio: waitpid: %s\n", strerror(errno));
+       } else if (ret == ffi->pid) {
+               if (WIFSIGNALED(status)) {
+                       ffi->signal = WTERMSIG(status);
+                       ffi->exited = 1;
+               }
+               if (WIFEXITED(status)) {
+                       if (WEXITSTATUS(status))
+                               ffi->exitval = WEXITSTATUS(status);
+                       ffi->exited = 1;
+               }
        }
+}
 
-       spdu.jobs = cpu_to_le32(thread_number);
-       fio_net_send_cmd(server_fd, FIO_NET_CMD_START, &spdu, sizeof(spdu), 0);
+static void fio_server_fork_item_done(struct fio_fork_item *ffi)
+{
+       dprint(FD_NET, "pid %u exited, sig=%u, exitval=%d\n", (int) ffi->pid, ffi->signal, ffi->exitval);
+
+       /*
+        * Fold STOP and QUIT...
+        */
+       fio_net_send_stop(server_fd, ffi->exitval, ffi->signal);
+       fio_net_send_quit(server_fd);
+       flist_del(&ffi->list);
+       free(ffi);
+}
+
+static void fio_server_check_fork_items(struct flist_head *list)
+{
+       struct flist_head *entry, *tmp;
+       struct fio_fork_item *ffi;
+
+       flist_for_each_safe(entry, tmp, list) {
+               ffi = flist_entry(entry, struct fio_fork_item, list);
+
+               fio_server_check_fork_item(ffi);
+
+               if (ffi->exited)
+                       fio_server_fork_item_done(ffi);
+       }
+}
+
+static void fio_server_check_jobs(struct flist_head *job_list)
+{
+       fio_server_check_fork_items(job_list);
+}
+
+static void fio_server_check_conns(struct flist_head *conn_list)
+{
+       fio_server_check_fork_items(conn_list);
+}
+
+static int handle_run_cmd(struct flist_head *job_list, struct fio_net_cmd *cmd)
+{
+       pid_t pid;
+       int ret;
+
+       set_genesis_time();
+
+       pid = fork();
+       if (pid) {
+               fio_server_add_job_pid(job_list, pid);
+               return 0;
+       }
 
        ret = fio_backend();
+       free_threads_shm();
+       _exit(ret);
+}
+
+static int handle_job_cmd(struct fio_net_cmd *cmd)
+{
+       struct cmd_job_pdu *pdu = (struct cmd_job_pdu *) cmd->payload;
+       void *buf = pdu->buf;
+       struct cmd_start_pdu spdu;
 
-       epdu.error = ret;
-       fio_net_send_cmd(server_fd, FIO_NET_CMD_STOP, &epdu, sizeof(epdu), 0);
+       pdu->buf_len = le32_to_cpu(pdu->buf_len);
+       pdu->client_type = le32_to_cpu(pdu->client_type);
 
-       fio_server_send_quit_cmd();
-       reset_fio_state();
-       first_cmd_check = 0;
-       return ret;
+       if (parse_jobs_ini(buf, 1, 0, pdu->client_type)) {
+               fio_net_send_quit(server_fd);
+               return -1;
+       }
+
+       spdu.jobs = cpu_to_le32(thread_number);
+       spdu.stat_outputs = cpu_to_le32(stat_number);
+       fio_net_send_cmd(server_fd, FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, NULL);
+       return 0;
 }
 
 static int handle_jobline_cmd(struct fio_net_cmd *cmd)
@@ -371,11 +583,13 @@ static int handle_jobline_cmd(struct fio_net_cmd *cmd)
        struct cmd_single_line_pdu *cslp;
        struct cmd_line_pdu *clp;
        unsigned long offset;
+       struct cmd_start_pdu spdu;
        char **argv;
-       int ret, i;
+       int i;
 
        clp = pdu;
        clp->lines = le16_to_cpu(clp->lines);
+       clp->client_type = le16_to_cpu(clp->client_type);
        argv = malloc(clp->lines * sizeof(char *));
        offset = sizeof(*clp);
 
@@ -389,50 +603,59 @@ static int handle_jobline_cmd(struct fio_net_cmd *cmd)
                dprint(FD_NET, "server: %d: %s\n", i, argv[i]);
        }
 
-       if (parse_cmd_line(clp->lines, argv)) {
-               fio_server_send_quit_cmd();
+       if (parse_cmd_line(clp->lines, argv, clp->client_type)) {
+               fio_net_send_quit(server_fd);
                free(argv);
                return -1;
        }
 
        free(argv);
 
-       fio_net_send_simple_cmd(server_fd, FIO_NET_CMD_START, 0, NULL);
-
-       ret = fio_backend();
-       fio_server_send_quit_cmd();
-       reset_fio_state();
-       first_cmd_check = 0;
-       return ret;
+       spdu.jobs = cpu_to_le32(thread_number);
+       spdu.stat_outputs = cpu_to_le32(stat_number);
+       fio_net_send_cmd(server_fd, FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, NULL);
+       return 0;
 }
 
 static int handle_probe_cmd(struct fio_net_cmd *cmd)
 {
-       struct cmd_probe_pdu probe;
+       struct cmd_client_probe_pdu *pdu = (struct cmd_client_probe_pdu *) cmd->payload;
+       struct cmd_probe_reply_pdu probe;
+       uint64_t tag = cmd->tag;
 
        dprint(FD_NET, "server: sending probe reply\n");
 
        memset(&probe, 0, sizeof(probe));
        gethostname((char *) probe.hostname, sizeof(probe.hostname));
-#ifdef FIO_BIG_ENDIAN
+#ifdef CONFIG_BIG_ENDIAN
        probe.bigendian = 1;
 #endif
-       probe.fio_major = FIO_MAJOR;
-       probe.fio_minor = FIO_MINOR;
-       probe.fio_patch = FIO_PATCH;
+       strncpy((char *) probe.fio_version, fio_version_string, sizeof(probe.fio_version));
 
        probe.os        = FIO_OS;
        probe.arch      = FIO_ARCH;
-
        probe.bpp       = sizeof(void *);
+       probe.cpus      = __cpu_to_le32(cpus_online());
 
-       return fio_net_send_cmd(server_fd, FIO_NET_CMD_PROBE, &probe, sizeof(probe), cmd->tag);
+       /*
+        * If the client supports compression and we do too, then enable it
+        */
+       if (has_zlib && le64_to_cpu(pdu->flags) & FIO_PROBE_FLAG_ZLIB) {
+               probe.flags = __cpu_to_le64(FIO_PROBE_FLAG_ZLIB);
+               use_zlib = 1;
+       } else {
+               probe.flags = 0;
+               use_zlib = 0;
+       }
+
+       return fio_net_send_cmd(server_fd, FIO_NET_CMD_PROBE, &probe, sizeof(probe), &tag, NULL);
 }
 
 static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
 {
        struct jobs_eta *je;
        size_t size;
+       uint64_t tag = cmd->tag;
        int i;
 
        if (!thread_number)
@@ -452,9 +675,10 @@ static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
        je->nr_running          = cpu_to_le32(je->nr_running);
        je->nr_ramp             = cpu_to_le32(je->nr_ramp);
        je->nr_pending          = cpu_to_le32(je->nr_pending);
+       je->nr_setting_up       = cpu_to_le32(je->nr_setting_up);
        je->files_open          = cpu_to_le32(je->files_open);
 
-       for (i = 0; i < 2; i++) {
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                je->m_rate[i]   = cpu_to_le32(je->m_rate[i]);
                je->t_rate[i]   = cpu_to_le32(je->t_rate[i]);
                je->m_iops[i]   = cpu_to_le32(je->m_iops[i]);
@@ -466,18 +690,51 @@ static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
        je->elapsed_sec         = cpu_to_le64(je->elapsed_sec);
        je->eta_sec             = cpu_to_le64(je->eta_sec);
        je->nr_threads          = cpu_to_le32(je->nr_threads);
+       je->is_pow2             = cpu_to_le32(je->is_pow2);
+       je->unit_base           = cpu_to_le32(je->unit_base);
 
-       fio_net_send_cmd(server_fd, FIO_NET_CMD_ETA, je, size, cmd->tag);
+       fio_net_send_cmd(server_fd, FIO_NET_CMD_ETA, je, size, &tag, NULL);
        free(je);
        return 0;
 }
 
-static int handle_command(struct fio_net_cmd *cmd)
+static int send_update_job_reply(int fd, uint64_t __tag, int error)
+{
+       uint64_t tag = __tag;
+       uint32_t pdu_error;
+
+       pdu_error = __cpu_to_le32(error);
+       return fio_net_send_cmd(fd, FIO_NET_CMD_UPDATE_JOB, &pdu_error, sizeof(pdu_error), &tag, NULL);
+}
+
+static int handle_update_job_cmd(struct fio_net_cmd *cmd)
+{
+       struct cmd_add_job_pdu *pdu = (struct cmd_add_job_pdu *) cmd->payload;
+       struct thread_data *td;
+       uint32_t tnumber;
+
+       tnumber = le32_to_cpu(pdu->thread_number);
+
+       dprint(FD_NET, "server: updating options for job %u\n", tnumber);
+
+       if (!tnumber || tnumber > thread_number) {
+               send_update_job_reply(server_fd, cmd->tag, ENODEV);
+               return 0;
+       }
+
+       td = &threads[tnumber - 1];
+       convert_thread_options_to_cpu(&td->o, &pdu->top);
+       send_update_job_reply(server_fd, cmd->tag, 0);
+       return 0;
+}
+
+static int handle_command(struct flist_head *job_list, struct fio_net_cmd *cmd)
 {
        int ret;
 
-       dprint(FD_NET, "server: got op [%s], pdu=%u, tag=%lx\n",
-                       fio_server_op(cmd->opcode), cmd->pdu_len, cmd->tag);
+       dprint(FD_NET, "server: got op [%s], pdu=%u, tag=%llx\n",
+                       fio_server_op(cmd->opcode), cmd->pdu_len,
+                       (unsigned long long) cmd->tag);
 
        switch (cmd->opcode) {
        case FIO_NET_CMD_QUIT:
@@ -498,19 +755,29 @@ static int handle_command(struct fio_net_cmd *cmd)
        case FIO_NET_CMD_SEND_ETA:
                ret = handle_send_eta_cmd(cmd);
                break;
+       case FIO_NET_CMD_RUN:
+               ret = handle_run_cmd(job_list, cmd);
+               break;
+       case FIO_NET_CMD_UPDATE_JOB:
+               ret = handle_update_job_cmd(cmd);
+               break;
        default:
-               log_err("fio: unknown opcode: %s\n",fio_server_op(cmd->opcode));
+               log_err("fio: unknown opcode: %s\n", fio_server_op(cmd->opcode));
                ret = 1;
        }
 
        return ret;
 }
 
-static int handle_connection(int sk, int block)
+static int handle_connection(int sk)
 {
        struct fio_net_cmd *cmd = NULL;
+       FLIST_HEAD(job_list);
        int ret = 0;
 
+       reset_fio_state();
+       server_fd = sk;
+
        /* read forever */
        while (!exit_backend) {
                struct pollfd pfd = {
@@ -520,15 +787,19 @@ static int handle_connection(int sk, int block)
 
                ret = 0;
                do {
-                       ret = poll(&pfd, 1, 100);
+                       int timeout = 1000;
+
+                       if (!flist_empty(&job_list))
+                               timeout = 100;
+
+                       ret = poll(&pfd, 1, timeout);
                        if (ret < 0) {
                                if (errno == EINTR)
                                        break;
                                log_err("fio: poll: %s\n", strerror(errno));
                                break;
                        } else if (!ret) {
-                               if (!block)
-                                       return 0;
+                               fio_server_check_jobs(&job_list);
                                continue;
                        }
 
@@ -540,6 +811,8 @@ static int handle_connection(int sk, int block)
                        }
                } while (!exit_backend);
 
+               fio_server_check_jobs(&job_list);
+
                if (ret < 0)
                        break;
 
@@ -549,7 +822,7 @@ static int handle_connection(int sk, int block)
                        break;
                }
 
-               ret = handle_command(cmd);
+               ret = handle_command(&job_list, cmd);
                if (ret)
                        break;
 
@@ -560,70 +833,86 @@ static int handle_connection(int sk, int block)
        if (cmd)
                free(cmd);
 
-       return ret;
-}
-
-void fio_server_idle_loop(void)
-{
-       if (!first_cmd_check) {
-               fio_net_send_simple_cmd(server_fd, FIO_NET_CMD_RUN, 0, NULL);
-               first_cmd_check = 1;
-       }
-       if (server_fd != -1)
-               handle_connection(server_fd, 0);
+       close(sk);
+       _exit(ret);
 }
 
 static int accept_loop(int listen_sk)
 {
        struct sockaddr_in addr;
-       fio_socklen_t len = sizeof(addr);
+       struct sockaddr_in6 addr6;
+       socklen_t len = use_ipv6 ? sizeof(addr6) : sizeof(addr);
        struct pollfd pfd;
-       int ret, sk, flags, exitval = 0;
+       int ret = 0, sk, flags, exitval = 0;
+       FLIST_HEAD(conn_list);
 
        dprint(FD_NET, "server enter accept loop\n");
 
        flags = fcntl(listen_sk, F_GETFL);
        flags |= O_NONBLOCK;
        fcntl(listen_sk, F_SETFL, flags);
-again:
-       pfd.fd = listen_sk;
-       pfd.events = POLLIN;
-       do {
-               ret = poll(&pfd, 1, 100);
-               if (ret < 0) {
-                       if (errno == EINTR)
+
+       while (!exit_backend) {
+               const char *from;
+               char buf[64];
+               pid_t pid;
+
+               pfd.fd = listen_sk;
+               pfd.events = POLLIN;
+               do {
+                       int timeout = 1000;
+
+                       if (!flist_empty(&conn_list))
+                               timeout = 100;
+
+                       ret = poll(&pfd, 1, timeout);
+                       if (ret < 0) {
+                               if (errno == EINTR)
+                                       break;
+                               log_err("fio: poll: %s\n", strerror(errno));
                                break;
-                       log_err("fio: poll: %s\n", strerror(errno));
-                       goto out;
-               } else if (!ret)
-                       continue;
+                       } else if (!ret) {
+                               fio_server_check_conns(&conn_list);
+                               continue;
+                       }
 
-               if (pfd.revents & POLLIN)
-                       break;
-       } while (!exit_backend);
+                       if (pfd.revents & POLLIN)
+                               break;
+               } while (!exit_backend);
 
-       if (exit_backend)
-               goto out;
+               fio_server_check_conns(&conn_list);
 
-       sk = accept(listen_sk, (struct sockaddr *) &addr, &len);
-       if (sk < 0) {
-               log_err("fio: accept: %s\n", strerror(errno));
-               return -1;
-       }
+               if (exit_backend || ret < 0)
+                       break;
 
-       dprint(FD_NET, "server: connect from %s\n", inet_ntoa(addr.sin_addr));
+               if (use_ipv6)
+                       sk = accept(listen_sk, (struct sockaddr *) &addr6, &len);
+               else
+                       sk = accept(listen_sk, (struct sockaddr *) &addr, &len);
 
-       server_fd = sk;
+               if (sk < 0) {
+                       log_err("fio: accept: %s\n", strerror(errno));
+                       return -1;
+               }
 
-       exitval = handle_connection(sk, 1);
+               if (use_ipv6)
+                       from = inet_ntop(AF_INET6, (struct sockaddr *) &addr6.sin6_addr, buf, sizeof(buf));
+               else
+                       from = inet_ntop(AF_INET, (struct sockaddr *) &addr.sin_addr, buf, sizeof(buf));
 
-       server_fd = -1;
-       close(sk);
+               dprint(FD_NET, "server: connect from %s\n", from);
 
-       if (!exit_backend)
-               goto again;
+               pid = fork();
+               if (pid) {
+                       close(sk);
+                       fio_server_add_conn_pid(&conn_list, pid);
+                       continue;
+               }
+
+               /* exits */
+               handle_connection(sk);
+       }
 
-out:
        return exitval;
 }
 
@@ -648,7 +937,7 @@ int fio_server_text_output(int level, const char *buf, size_t len)
 
        memcpy(pdu->buf, buf, len);
 
-       fio_net_send_cmd(server_fd, FIO_NET_CMD_TEXT, pdu, tlen, 0);
+       fio_net_send_cmd(server_fd, FIO_NET_CMD_TEXT, pdu, tlen, NULL, NULL);
        free(pdu);
        return len;
 }
@@ -670,7 +959,7 @@ static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
 {
        int i;
 
-       for (i = 0; i < 2; i++) {
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                dst->max_run[i]         = cpu_to_le64(src->max_run[i]);
                dst->min_run[i]         = cpu_to_le64(src->min_run[i]);
                dst->max_bw[i]          = cpu_to_le64(src->max_bw[i]);
@@ -680,7 +969,9 @@ static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
        }
 
        dst->kb_base    = cpu_to_le32(src->kb_base);
+       dst->unit_base  = cpu_to_le32(src->unit_base);
        dst->groupid    = cpu_to_le32(src->groupid);
+       dst->unified_rw_rep     = cpu_to_le32(src->unified_rw_rep);
 }
 
 /*
@@ -696,16 +987,18 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
 
        memset(&p, 0, sizeof(p));
 
-       strcpy(p.ts.name, ts->name);
-       strcpy(p.ts.verror, ts->verror);
-       strcpy(p.ts.description, ts->description);
+       strncpy(p.ts.name, ts->name, FIO_JOBNAME_SIZE - 1);
+       strncpy(p.ts.verror, ts->verror, FIO_VERROR_SIZE - 1);
+       strncpy(p.ts.description, ts->description, FIO_JOBDESC_SIZE - 1);
 
-       p.ts.error      = cpu_to_le32(ts->error);
-       p.ts.groupid    = cpu_to_le32(ts->groupid);
-       p.ts.pid        = cpu_to_le32(ts->pid);
-       p.ts.members    = cpu_to_le32(ts->members);
+       p.ts.error              = cpu_to_le32(ts->error);
+       p.ts.thread_number      = cpu_to_le32(ts->thread_number);
+       p.ts.groupid            = cpu_to_le32(ts->groupid);
+       p.ts.pid                = cpu_to_le32(ts->pid);
+       p.ts.members            = cpu_to_le32(ts->members);
+       p.ts.unified_rw_rep     = cpu_to_le32(ts->unified_rw_rep);
 
-       for (i = 0; i < 2; i++) {
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                convert_io_stat(&p.ts.clat_stat[i], &ts->clat_stat[i]);
                convert_io_stat(&p.ts.slat_stat[i], &ts->slat_stat[i]);
                convert_io_stat(&p.ts.lat_stat[i], &ts->lat_stat[i]);
@@ -737,11 +1030,11 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
                p.ts.io_u_lat_m[i]      = cpu_to_le32(ts->io_u_lat_m[i]);
        }
 
-       for (i = 0; i < 2; i++)
+       for (i = 0; i < DDIR_RWDIR_CNT; i++)
                for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
                        p.ts.io_u_plat[i][j] = cpu_to_le32(ts->io_u_plat[i][j]);
 
-       for (i = 0; i < 3; i++) {
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                p.ts.total_io_u[i]      = cpu_to_le64(ts->total_io_u[i]);
                p.ts.short_io_u[i]      = cpu_to_le64(ts->short_io_u[i]);
        }
@@ -749,7 +1042,7 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
        p.ts.total_submit       = cpu_to_le64(ts->total_submit);
        p.ts.total_complete     = cpu_to_le64(ts->total_complete);
 
-       for (i = 0; i < 2; i++) {
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
                p.ts.io_bytes[i]        = cpu_to_le64(ts->io_bytes[i]);
                p.ts.runtime[i]         = cpu_to_le64(ts->runtime[i]);
        }
@@ -759,10 +1052,16 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
        p.ts.total_err_count    = cpu_to_le64(ts->total_err_count);
        p.ts.first_error        = cpu_to_le32(ts->first_error);
        p.ts.kb_base            = cpu_to_le32(ts->kb_base);
+       p.ts.unit_base          = cpu_to_le32(ts->unit_base);
+
+       p.ts.latency_depth      = cpu_to_le32(ts->latency_depth);
+       p.ts.latency_target     = cpu_to_le64(ts->latency_target);
+       p.ts.latency_window     = cpu_to_le64(ts->latency_window);
+       p.ts.latency_percentile.u.i = __cpu_to_le64(fio_double_to_uint64(ts->latency_percentile.u.f));
 
        convert_gs(&p.rs, rs);
 
-       fio_net_send_cmd(server_fd, FIO_NET_CMD_TS, &p, sizeof(p), 0);
+       fio_net_send_cmd(server_fd, FIO_NET_CMD_TS, &p, sizeof(p), NULL, NULL);
 }
 
 void fio_server_send_gs(struct group_run_stats *rs)
@@ -772,7 +1071,7 @@ void fio_server_send_gs(struct group_run_stats *rs)
        dprint(FD_NET, "server sending group run stats\n");
 
        convert_gs(&gs, rs);
-       fio_net_send_cmd(server_fd, FIO_NET_CMD_GS, &gs, sizeof(gs), 0);
+       fio_net_send_cmd(server_fd, FIO_NET_CMD_GS, &gs, sizeof(gs), NULL, NULL);
 }
 
 static void convert_agg(struct disk_util_agg *dst, struct disk_util_agg *src)
@@ -826,36 +1125,150 @@ void fio_server_send_du(void)
                convert_dus(&pdu.dus, &du->dus);
                convert_agg(&pdu.agg, &du->agg);
 
-               fio_net_send_cmd(server_fd, FIO_NET_CMD_DU, &pdu, sizeof(pdu), 0);
+               fio_net_send_cmd(server_fd, FIO_NET_CMD_DU, &pdu, sizeof(pdu), NULL, NULL);
        }
 }
 
-void fio_server_send_add_job(struct thread_options *o, const char *ioengine)
+/*
+ * Send a command with a separate PDU, not inlined in the command
+ */
+static int fio_send_cmd_ext_pdu(int sk, uint16_t opcode, const void *buf,
+                               off_t size, uint64_t tag, uint32_t flags)
 {
-       struct cmd_add_job_pdu pdu;
-       int i;
+       struct fio_net_cmd cmd;
+       struct iovec iov[2];
 
-       strcpy((char *) pdu.jobname, o->name);
-       strcpy((char *) pdu.ioengine, ioengine);
+       iov[0].iov_base = (void *) &cmd;
+       iov[0].iov_len = sizeof(cmd);
+       iov[1].iov_base = (void *) buf;
+       iov[1].iov_len = size;
 
-       pdu.iodepth             = cpu_to_le32(o->iodepth);
-       pdu.rw                  = cpu_to_le32(o->td_ddir);
+       __fio_init_net_cmd(&cmd, opcode, size, tag);
+       cmd.flags = __cpu_to_le32(flags);
+       fio_net_cmd_crc_pdu(&cmd, buf);
 
-       for (i = 0; i < 2; i++) {
-               pdu.min_bs[i]   = cpu_to_le32(o->min_bs[i]);
-               pdu.max_bs[i]   = cpu_to_le32(o->max_bs[i]);
+       return fio_sendv_data(sk, iov, 2);
+}
+
+static int fio_send_iolog_gz(struct cmd_iolog_pdu *pdu, struct io_log *log)
+{
+       int ret = 0;
+#ifdef CONFIG_ZLIB
+       z_stream stream;
+       void *out_pdu;
+
+       /*
+        * Dirty - since the log is potentially huge, compress it into
+        * FIO_SERVER_MAX_FRAGMENT_PDU chunks and let the receiving
+        * side defragment it.
+        */
+       out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
+
+       stream.zalloc = Z_NULL;
+       stream.zfree = Z_NULL;
+       stream.opaque = Z_NULL;
+
+       if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK) {
+               ret = 1;
+               goto err;
        }
 
-       pdu.numjobs             = cpu_to_le32(o->numjobs);
-       pdu.group_reporting     = cpu_to_le32(o->group_reporting);
+       stream.next_in = (void *) log->log;
+       stream.avail_in = log->nr_samples * sizeof(struct io_sample);
 
-       fio_net_send_cmd(server_fd, FIO_NET_CMD_ADD_JOB, &pdu, sizeof(pdu), 0);
+       do {
+               unsigned int this_len, flags = 0;
+               int ret;
+
+               stream.avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
+               stream.next_out = out_pdu;
+               ret = deflate(&stream, Z_FINISH);
+               /* may be Z_OK, or Z_STREAM_END */
+               if (ret < 0)
+                       goto err_zlib;
+
+               this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream.avail_out;
+
+               if (stream.avail_in)
+                       flags = FIO_NET_CMD_F_MORE;
+
+               ret = fio_send_cmd_ext_pdu(server_fd, FIO_NET_CMD_IOLOG,
+                                          out_pdu, this_len, 0, flags);
+               if (ret)
+                       goto err_zlib;
+       } while (stream.avail_in);
+
+err_zlib:
+       deflateEnd(&stream);
+err:
+       free(out_pdu);
+#endif
+       return ret;
+}
+
+int fio_send_iolog(struct thread_data *td, struct io_log *log, const char *name)
+{
+       struct cmd_iolog_pdu pdu;
+       int i, ret = 0;
+
+       pdu.thread_number = cpu_to_le32(td->thread_number);
+       pdu.nr_samples = __cpu_to_le32(log->nr_samples);
+       pdu.log_type = cpu_to_le32(log->log_type);
+       pdu.compressed = cpu_to_le32(use_zlib);
+       strcpy((char *) pdu.name, name);
+
+       for (i = 0; i < log->nr_samples; i++) {
+               struct io_sample *s = &log->log[i];
+
+               s->time = cpu_to_le64(s->time);
+               s->val  = cpu_to_le64(s->val);
+               s->ddir = cpu_to_le32(s->ddir);
+               s->bs   = cpu_to_le32(s->bs);
+       }
+
+       /*
+        * Send header first, it's not compressed.
+        */
+       ret = fio_send_cmd_ext_pdu(server_fd, FIO_NET_CMD_IOLOG, &pdu,
+                                       sizeof(pdu), 0, FIO_NET_CMD_F_MORE);
+       if (ret)
+               return ret;
+
+       /*
+        * Now send actual log, compress if we can, otherwise just plain
+        */
+       if (use_zlib)
+               return fio_send_iolog_gz(&pdu, log);
+
+       return fio_send_cmd_ext_pdu(server_fd, FIO_NET_CMD_IOLOG, log->log,
+                       log->nr_samples * sizeof(struct io_sample), 0, 0);
+}
+
+void fio_server_send_add_job(struct thread_data *td)
+{
+       struct cmd_add_job_pdu pdu;
+
+       memset(&pdu, 0, sizeof(pdu));
+       pdu.thread_number = cpu_to_le32(td->thread_number);
+       pdu.groupid = cpu_to_le32(td->groupid);
+       convert_thread_options_to_net(&pdu.top, &td->o);
+
+       fio_net_send_cmd(server_fd, FIO_NET_CMD_ADD_JOB, &pdu, sizeof(pdu), NULL, NULL);
+}
+
+void fio_server_send_start(struct thread_data *td)
+{
+       assert(server_fd != -1);
+
+       fio_net_send_simple_cmd(server_fd, FIO_NET_CMD_SERVER_START, 0, NULL);
 }
 
 static int fio_init_server_ip(void)
 {
        struct sockaddr *addr;
-       fio_socklen_t socklen;
+       socklen_t socklen;
+       char buf[80];
+       const char *str;
        int sk, opt;
 
        if (use_ipv6)
@@ -869,7 +1282,7 @@ static int fio_init_server_ip(void)
        }
 
        opt = 1;
-       if (setsockopt(sk, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
+       if (setsockopt(sk, SOL_SOCKET, SO_REUSEADDR, (void *)&opt, sizeof(opt)) < 0) {
                log_err("fio: setsockopt: %s\n", strerror(errno));
                close(sk);
                return -1;
@@ -883,17 +1296,24 @@ static int fio_init_server_ip(void)
 #endif
 
        if (use_ipv6) {
+               const void *src = &saddr_in6.sin6_addr;
+
                addr = (struct sockaddr *) &saddr_in6;
                socklen = sizeof(saddr_in6);
                saddr_in6.sin6_family = AF_INET6;
+               str = inet_ntop(AF_INET6, src, buf, sizeof(buf));
        } else {
+               const void *src = &saddr_in.sin_addr;
+
                addr = (struct sockaddr *) &saddr_in;
                socklen = sizeof(saddr_in);
                saddr_in.sin_family = AF_INET;
+               str = inet_ntop(AF_INET, src, buf, sizeof(buf));
        }
 
        if (bind(sk, addr, socklen) < 0) {
                log_err("fio: bind: %s\n", strerror(errno));
+               log_info("fio: failed with IPv%c %s\n", use_ipv6 ? '6' : '4', str);
                close(sk);
                return -1;
        }
@@ -904,7 +1324,7 @@ static int fio_init_server_ip(void)
 static int fio_init_server_sock(void)
 {
        struct sockaddr_un addr;
-       fio_socklen_t len;
+       socklen_t len;
        mode_t mode;
        int sk;
 
@@ -919,7 +1339,6 @@ static int fio_init_server_sock(void)
        memset(&addr, 0, sizeof(addr));
        addr.sun_family = AF_UNIX;
        strcpy(addr.sun_path, bind_sock);
-       unlink(bind_sock);
 
        len = sizeof(addr.sun_family) + strlen(bind_sock) + 1;
 
@@ -981,41 +1400,38 @@ static int fio_init_server_connection(void)
        return sk;
 }
 
-int fio_server_parse_host(const char *host, int *ipv6, struct in_addr *inp,
+int fio_server_parse_host(const char *host, int ipv6, struct in_addr *inp,
                          struct in6_addr *inp6)
 
 {
        int ret = 0;
 
-       if (*ipv6)
+       if (ipv6)
                ret = inet_pton(AF_INET6, host, inp6);
        else
                ret = inet_pton(AF_INET, host, inp);
 
        if (ret != 1) {
-               struct hostent *hent;
+               struct addrinfo hints, *res;
 
-               hent = gethostbyname(host);
-               if (!hent) {
-                       log_err("fio: failed to resolve <%s>\n", host);
-                       return 0;
-               }
+               memset(&hints, 0, sizeof(hints));
+               hints.ai_family = ipv6 ? AF_INET6 : AF_INET;
+               hints.ai_socktype = SOCK_STREAM;
 
-               if (*ipv6) {
-                       if (hent->h_addrtype != AF_INET6) {
-                               log_info("fio: falling back to IPv4\n");
-                               *ipv6 = 0;
-                       } else
-                               memcpy(inp6, hent->h_addr_list[0], 16);
-               }
-               if (!*ipv6) {
-                       if (hent->h_addrtype != AF_INET) {
-                               log_err("fio: lookup type mismatch\n");
-                               return 0;
-                       }
-                       memcpy(inp, hent->h_addr_list[0], 4);
+               ret = getaddrinfo(host, NULL, &hints, &res);
+               if (ret) {
+                       log_err("fio: failed to resolve <%s> (%s)\n", host,
+                                       gai_strerror(ret));
+                       return 1;
                }
+
+               if (ipv6)
+                       memcpy(inp6, &((struct sockaddr_in6 *) res->ai_addr)->sin6_addr, sizeof(*inp6));
+               else
+                       memcpy(inp, &((struct sockaddr_in *) res->ai_addr)->sin_addr, sizeof(*inp));
+
                ret = 1;
+               freeaddrinfo(res);
        }
 
        return !(ret == 1);
@@ -1068,7 +1484,7 @@ int fio_server_parse_string(const char *str, char **ptr, int *is_sock,
                host++;
                lport = atoi(host);
                if (!lport || lport > 65535) {
-                       log_err("fio: bad server port %u\n", port);
+                       log_err("fio: bad server port %u\n", lport);
                        return 1;
                }
                /* no hostname given, we are done */
@@ -1077,7 +1493,7 @@ int fio_server_parse_string(const char *str, char **ptr, int *is_sock,
        }
 
        /*
-        * If no port seen yet, check if there's a last ':' at the end
+        * If no port seen yet, check if there's a last ',' at the end
         */
        if (!lport) {
                portp = strchr(host, ',');
@@ -1086,7 +1502,7 @@ int fio_server_parse_string(const char *str, char **ptr, int *is_sock,
                        portp++;
                        lport = atoi(portp);
                        if (!lport || lport > 65535) {
-                               log_err("fio: bad server port %u\n", port);
+                               log_err("fio: bad server port %u\n", lport);
                                return 1;
                        }
                }
@@ -1100,7 +1516,7 @@ int fio_server_parse_string(const char *str, char **ptr, int *is_sock,
 
        *ptr = strdup(host);
 
-       if (fio_server_parse_host(*ptr, ipv6, inp, inp6)) {
+       if (fio_server_parse_host(*ptr, *ipv6, inp, inp6)) {
                free(*ptr);
                *ptr = NULL;
                return 1;
@@ -1150,6 +1566,22 @@ out:
        return ret;
 }
 
+static void sig_int(int sig)
+{
+       if (bind_sock)
+               unlink(bind_sock);
+}
+
+static void set_sig_handlers(void)
+{
+       struct sigaction act;
+
+       memset(&act, 0, sizeof(act));
+       act.sa_handler = sig_int;
+       act.sa_flags = SA_RESTART;
+       sigaction(SIGINT, &act, NULL);
+}
+
 static int fio_server(void)
 {
        int sk, ret;
@@ -1163,6 +1595,8 @@ static int fio_server(void)
        if (sk < 0)
                return -1;
 
+       set_sig_handlers();
+
        ret = accept_loop(sk);
 
        close(sk);
@@ -1238,8 +1672,8 @@ int fio_start_server(char *pidfile)
        int ret;
 
 #if defined(WIN32)
-    WSADATA wsd;
-    WSAStartup(MAKEWORD(2,2), &wsd);
+       WSADATA wsd;
+       WSAStartup(MAKEWORD(2, 2), &wsd);
 #endif
 
        if (!pidfile)