steadystate: instead of including ss_sum_y in thread_stat record whether ss_sum_y...
[fio.git] / server.c
index efdf51f3870a8587fbaa8a0be9eafa03b75bc055..780f09f4ac8d7fb69da798968bd43368ece23b86 100644 (file)
--- a/server.c
+++ b/server.c
@@ -37,23 +37,28 @@ enum {
        SK_F_COPY       = 2,
        SK_F_SIMPLE     = 4,
        SK_F_VEC        = 8,
+       SK_F_INLINE     = 16,
 };
 
 struct sk_entry {
-       struct flist_head list;
-       int opcode;
+       struct flist_head list; /* link on sk_out->list */
+       int flags;              /* SK_F_* */
+       int opcode;             /* Actual command fields */
        void *buf;
        off_t size;
-       uint64_t *tagptr;
-       int flags;
-       struct flist_head next;
+       uint64_t tag;
+       struct flist_head next; /* Other sk_entry's, if linked command */
 };
 
 struct sk_out {
-       int sk;
-       struct fio_mutex *lock;
-       struct flist_head list;
-       struct fio_mutex *wait;
+       unsigned int refs;      /* frees sk_out when it drops to zero.
+                                * protected by below ->lock */
+
+       int sk;                 /* socket fd to talk to client */
+       struct fio_mutex lock;  /* protects ref and below list */
+       struct flist_head list; /* list of pending transmit work */
+       struct fio_mutex wait;  /* wake backend when items added to list */
+       struct fio_mutex xmit;  /* held while sending data */
 };
 
 static char *fio_server_arg;
@@ -109,16 +114,85 @@ static const char *fio_server_ops[FIO_NET_CMD_NR] = {
        "LOAD_FILE",
        "VTRIGGER",
        "SENDFILE",
+       "JOB_OPT",
 };
 
 static void sk_lock(struct sk_out *sk_out)
 {
-       fio_mutex_down(sk_out->lock);
+       fio_mutex_down(&sk_out->lock);
 }
 
 static void sk_unlock(struct sk_out *sk_out)
 {
-       fio_mutex_up(sk_out->lock);
+       fio_mutex_up(&sk_out->lock);
+}
+
+void sk_out_assign(struct sk_out *sk_out)
+{
+       if (!sk_out)
+               return;
+
+       sk_lock(sk_out);
+       sk_out->refs++;
+       sk_unlock(sk_out);
+       pthread_setspecific(sk_out_key, sk_out);
+}
+
+static void sk_out_free(struct sk_out *sk_out)
+{
+       __fio_mutex_remove(&sk_out->lock);
+       __fio_mutex_remove(&sk_out->wait);
+       __fio_mutex_remove(&sk_out->xmit);
+       sfree(sk_out);
+}
+
+static int __sk_out_drop(struct sk_out *sk_out)
+{
+       if (sk_out) {
+               int refs;
+
+               sk_lock(sk_out);
+               assert(sk_out->refs != 0);
+               refs = --sk_out->refs;
+               sk_unlock(sk_out);
+
+               if (!refs) {
+                       sk_out_free(sk_out);
+                       pthread_setspecific(sk_out_key, NULL);
+                       return 0;
+               }
+       }
+
+       return 1;
+}
+
+void sk_out_drop(void)
+{
+       struct sk_out *sk_out;
+
+       sk_out = pthread_getspecific(sk_out_key);
+       __sk_out_drop(sk_out);
+}
+
+static void __fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
+                              uint32_t pdu_len, uint64_t tag)
+{
+       memset(cmd, 0, sizeof(*cmd));
+
+       cmd->version    = __cpu_to_le16(FIO_SERVER_VER);
+       cmd->opcode     = cpu_to_le16(opcode);
+       cmd->tag        = cpu_to_le64(tag);
+       cmd->pdu_len    = cpu_to_le32(pdu_len);
+}
+
+
+static void fio_init_net_cmd(struct fio_net_cmd *cmd, uint16_t opcode,
+                            const void *pdu, uint32_t pdu_len, uint64_t tag)
+{
+       __fio_init_net_cmd(cmd, opcode, pdu_len, tag);
+
+       if (pdu)
+               memcpy(&cmd->payload, pdu, pdu_len);
 }
 
 const char *fio_server_op(unsigned int op)
@@ -180,7 +254,7 @@ static int fio_sendv_data(int sk, struct iovec *iov, int count)
        return 1;
 }
 
-int fio_send_data(int sk, const void *p, unsigned int len)
+static int fio_send_data(int sk, const void *p, unsigned int len)
 {
        struct iovec iov = { .iov_base = (void *) p, .iov_len = len };
 
@@ -189,10 +263,17 @@ int fio_send_data(int sk, const void *p, unsigned int len)
        return fio_sendv_data(sk, &iov, 1);
 }
 
-int fio_recv_data(int sk, void *p, unsigned int len)
+static int fio_recv_data(int sk, void *p, unsigned int len, bool wait)
 {
+       int flags;
+
+       if (wait)
+               flags = MSG_WAITALL;
+       else
+               flags = OS_MSG_DONTWAIT;
+
        do {
-               int ret = recv(sk, p, len, MSG_WAITALL);
+               int ret = recv(sk, p, len, flags);
 
                if (ret > 0) {
                        len -= ret;
@@ -202,9 +283,11 @@ int fio_recv_data(int sk, void *p, unsigned int len)
                        continue;
                } else if (!ret)
                        break;
-               else if (errno == EAGAIN || errno == EINTR)
-                       continue;
-               else
+               else if (errno == EAGAIN || errno == EINTR) {
+                       if (wait)
+                               continue;
+                       break;
+               } else
                        break;
        } while (!exit_backend);
 
@@ -253,7 +336,7 @@ static int verify_convert_cmd(struct fio_net_cmd *cmd)
 /*
  * Read (and defragment, if necessary) incoming commands
  */
-struct fio_net_cmd *fio_net_recv_cmd(int sk)
+struct fio_net_cmd *fio_net_recv_cmd(int sk, bool wait)
 {
        struct fio_net_cmd cmd, *tmp, *cmdret = NULL;
        size_t cmd_size = 0, pdu_offset = 0;
@@ -262,7 +345,7 @@ struct fio_net_cmd *fio_net_recv_cmd(int sk)
        void *pdu = NULL;
 
        do {
-               ret = fio_recv_data(sk, &cmd, sizeof(cmd));
+               ret = fio_recv_data(sk, &cmd, sizeof(cmd), wait);
                if (ret)
                        break;
 
@@ -306,7 +389,7 @@ struct fio_net_cmd *fio_net_recv_cmd(int sk)
 
                /* There's payload, get it */
                pdu = (void *) cmdret->payload + pdu_offset;
-               ret = fio_recv_data(sk, pdu, cmd.pdu_len);
+               ret = fio_recv_data(sk, pdu, cmd.pdu_len, wait);
                if (ret)
                        break;
 
@@ -381,7 +464,7 @@ static void free_reply(uint64_t tag)
        free(reply);
 }
 
-void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
+static void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
 {
        uint32_t pdu_len;
 
@@ -391,7 +474,7 @@ void fio_net_cmd_crc_pdu(struct fio_net_cmd *cmd, const void *pdu)
        cmd->pdu_crc16 = __cpu_to_le16(fio_crc16(pdu, pdu_len));
 }
 
-void fio_net_cmd_crc(struct fio_net_cmd *cmd)
+static void fio_net_cmd_crc(struct fio_net_cmd *cmd)
 {
        fio_net_cmd_crc_pdu(cmd, cmd->payload);
 }
@@ -448,8 +531,9 @@ int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
        return ret;
 }
 
-struct sk_entry *fio_net_prep_cmd(uint16_t opcode, void *buf, off_t size,
-                                 uint64_t *tagptr, int flags)
+static struct sk_entry *fio_net_prep_cmd(uint16_t opcode, void *buf,
+                                        size_t size, uint64_t *tagptr,
+                                        int flags)
 {
        struct sk_entry *entry;
 
@@ -461,22 +545,31 @@ struct sk_entry *fio_net_prep_cmd(uint16_t opcode, void *buf, off_t size,
                memcpy(entry->buf, buf, size);
        } else
                entry->buf = buf;
+
        entry->size = size;
-       entry->tagptr = tagptr;
+       if (tagptr)
+               entry->tag = *tagptr;
+       else
+               entry->tag = 0;
        entry->flags = flags;
-
        return entry;
 }
 
+static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry);
+
 static void fio_net_queue_entry(struct sk_entry *entry)
 {
        struct sk_out *sk_out = pthread_getspecific(sk_out_key);
 
-       sk_lock(sk_out);
-       flist_add_tail(&entry->list, &sk_out->list);
-       sk_unlock(sk_out);
+       if (entry->flags & SK_F_INLINE)
+               handle_sk_entry(sk_out, entry);
+       else {
+               sk_lock(sk_out);
+               flist_add_tail(&entry->list, &sk_out->list);
+               sk_unlock(sk_out);
 
-       fio_mutex_up(sk_out->wait);
+               fio_mutex_up(&sk_out->wait);
+       }
 }
 
 static int fio_net_queue_cmd(uint16_t opcode, void *buf, off_t size,
@@ -672,10 +765,11 @@ static int handle_load_file_cmd(struct fio_net_cmd *cmd)
 static int handle_run_cmd(struct sk_out *sk_out, struct flist_head *job_list,
                          struct fio_net_cmd *cmd)
 {
-       struct backend_data data;
        pid_t pid;
        int ret;
 
+       sk_out_assign(sk_out);
+
        fio_time_init();
        set_genesis_time();
 
@@ -685,11 +779,9 @@ static int handle_run_cmd(struct sk_out *sk_out, struct flist_head *job_list,
                return 0;
        }
 
-       data.key = sk_out_key;
-       data.ptr = sk_out;
-       //pthread_setspecific(sk_out_key, sk_out);
-       ret = fio_backend(&data);
+       ret = fio_backend(sk_out);
        free_threads_shm();
+       sk_out_drop();
        _exit(ret);
 }
 
@@ -880,9 +972,9 @@ static int handle_trigger_cmd(struct fio_net_cmd *cmd)
                struct all_io_list state;
 
                state.threads = cpu_to_le64((uint64_t) 0);
-               fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, &state, sizeof(state), NULL, SK_F_COPY);
+               fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, &state, sizeof(state), NULL, SK_F_COPY | SK_F_INLINE);
        } else
-               fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, rep, sz, NULL, SK_F_FREE);
+               fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, rep, sz, NULL, SK_F_FREE | SK_F_INLINE);
 
        exec_trigger(buf);
        return 0;
@@ -971,17 +1063,35 @@ static int fio_send_cmd_ext_pdu(int sk, uint16_t opcode, const void *buf,
 {
        struct fio_net_cmd cmd;
        struct iovec iov[2];
+       size_t this_len;
+       int ret;
 
        iov[0].iov_base = (void *) &cmd;
        iov[0].iov_len = sizeof(cmd);
-       iov[1].iov_base = (void *) buf;
-       iov[1].iov_len = size;
 
-       __fio_init_net_cmd(&cmd, opcode, size, tag);
-       cmd.flags = __cpu_to_le32(flags);
-       fio_net_cmd_crc_pdu(&cmd, buf);
+       do {
+               uint32_t this_flags = flags;
+
+               this_len = size;
+               if (this_len > FIO_SERVER_MAX_FRAGMENT_PDU)
+                       this_len = FIO_SERVER_MAX_FRAGMENT_PDU;
+
+               if (this_len < size)
+                       this_flags |= FIO_NET_CMD_F_MORE;
+
+               __fio_init_net_cmd(&cmd, opcode, this_len, tag);
+               cmd.flags = __cpu_to_le32(this_flags);
+               fio_net_cmd_crc_pdu(&cmd, buf);
+
+               iov[1].iov_base = (void *) buf;
+               iov[1].iov_len = this_len;
+
+               ret = fio_sendv_data(sk, iov, 2);
+               size -= this_len;
+               buf += this_len;
+       } while (!ret && size);
 
-       return fio_sendv_data(sk, iov, 2);
+       return ret;
 }
 
 static void finish_entry(struct sk_entry *entry)
@@ -994,29 +1104,24 @@ static void finish_entry(struct sk_entry *entry)
        sfree(entry);
 }
 
-static void entry_set_flags_tag(struct sk_entry *entry, struct flist_head *list,
-                               unsigned int *flags, uint64_t *tag)
+static void entry_set_flags(struct sk_entry *entry, struct flist_head *list,
+                           unsigned int *flags)
 {
        if (!flist_empty(list))
                *flags = FIO_NET_CMD_F_MORE;
        else
                *flags = 0;
-
-       if (entry->tagptr)
-               *tag = *entry->tagptr;
-       else
-               *tag = 0;
 }
 
 static int send_vec_entry(struct sk_out *sk_out, struct sk_entry *first)
 {
        unsigned int flags;
-       uint64_t tag;
        int ret;
 
-       entry_set_flags_tag(first, &first->next, &flags, &tag);
+       entry_set_flags(first, &first->next, &flags);
 
-       ret = fio_send_cmd_ext_pdu(sk_out->sk, first->opcode, first->buf, first->size, tag, flags);
+       ret = fio_send_cmd_ext_pdu(sk_out->sk, first->opcode, first->buf,
+                                       first->size, first->tag, flags);
 
        while (!flist_empty(&first->next)) {
                struct sk_entry *next;
@@ -1024,9 +1129,10 @@ static int send_vec_entry(struct sk_out *sk_out, struct sk_entry *first)
                next = flist_first_entry(&first->next, struct sk_entry, list);
                flist_del_init(&next->list);
 
-               entry_set_flags_tag(next, &first->next, &flags, &tag);
+               entry_set_flags(next, &first->next, &flags);
 
-               ret += fio_send_cmd_ext_pdu(sk_out->sk, next->opcode, next->buf, next->size, tag, flags);
+               ret += fio_send_cmd_ext_pdu(sk_out->sk, next->opcode, next->buf,
+                                               next->size, next->tag, flags);
                finish_entry(next);
        }
 
@@ -1037,17 +1143,19 @@ static int handle_sk_entry(struct sk_out *sk_out, struct sk_entry *entry)
 {
        int ret;
 
+       fio_mutex_down(&sk_out->xmit);
+
        if (entry->flags & SK_F_VEC)
                ret = send_vec_entry(sk_out, entry);
-       if (entry->flags & SK_F_SIMPLE) {
-               uint64_t tag = 0;
+       else if (entry->flags & SK_F_SIMPLE) {
+               ret = fio_net_send_simple_cmd(sk_out->sk, entry->opcode,
+                                               entry->tag, NULL);
+       } else {
+               ret = fio_net_send_cmd(sk_out->sk, entry->opcode, entry->buf,
+                                       entry->size, &entry->tag, NULL);
+       }
 
-               if (entry->tagptr)
-                       tag = *entry->tagptr;
-
-               ret = fio_net_send_simple_cmd(sk_out->sk, entry->opcode, tag, NULL);
-       } else
-               ret = fio_net_send_cmd(sk_out->sk, entry->opcode, entry->buf, entry->size, entry->tagptr, NULL);
+       fio_mutex_up(&sk_out->xmit);
 
        if (ret)
                log_err("fio: failed handling cmd %s\n", fio_server_op(entry->opcode));
@@ -1112,7 +1220,7 @@ static int handle_connection(struct sk_out *sk_out)
                                break;
                        } else if (!ret) {
                                fio_server_check_jobs(&job_list);
-                               fio_mutex_down_timeout(sk_out->wait, timeout);
+                               fio_mutex_down_timeout(&sk_out->wait, timeout);
                                continue;
                        }
 
@@ -1129,7 +1237,7 @@ static int handle_connection(struct sk_out *sk_out)
                if (ret < 0)
                        break;
 
-               cmd = fio_net_recv_cmd(sk_out->sk);
+               cmd = fio_net_recv_cmd(sk_out->sk, true);
                if (!cmd) {
                        ret = -1;
                        break;
@@ -1149,13 +1257,15 @@ static int handle_connection(struct sk_out *sk_out)
        handle_xmits(sk_out);
 
        close(sk_out->sk);
+       sk_out->sk = -1;
+       __sk_out_drop(sk_out);
        _exit(ret);
 }
 
 /* get the address on this host bound by the input socket, 
  * whether it is ipv6 or ipv4 */
 
-int get_my_addr_str(int sk)
+static int get_my_addr_str(int sk)
 {
        struct sockaddr_in6 myaddr6 = { 0, };
        struct sockaddr_in myaddr4 = { 0, };
@@ -1196,21 +1306,14 @@ static int accept_loop(int listen_sk)
        socklen_t len = use_ipv6 ? sizeof(addr6) : sizeof(addr);
        struct pollfd pfd;
        int ret = 0, sk, exitval = 0;
-       struct sk_out *sk_out;
        FLIST_HEAD(conn_list);
 
        dprint(FD_NET, "server enter accept loop\n");
 
        fio_set_fd_nonblocking(listen_sk, "server");
 
-       sk_out = smalloc(sizeof(*sk_out));
-       INIT_FLIST_HEAD(&sk_out->list);
-       sk_out->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
-       sk_out->wait = fio_mutex_init(FIO_MUTEX_LOCKED);
-
-       pthread_setspecific(sk_out_key, sk_out);
-
        while (!exit_backend) {
+               struct sk_out *sk_out;
                const char *from;
                char buf[64];
                pid_t pid;
@@ -1260,28 +1363,31 @@ static int accept_loop(int listen_sk)
 
                dprint(FD_NET, "server: connect from %s\n", from);
 
+               sk_out = smalloc(sizeof(*sk_out));
                sk_out->sk = sk;
+               INIT_FLIST_HEAD(&sk_out->list);
+               __fio_mutex_init(&sk_out->lock, FIO_MUTEX_UNLOCKED);
+               __fio_mutex_init(&sk_out->wait, FIO_MUTEX_LOCKED);
+               __fio_mutex_init(&sk_out->xmit, FIO_MUTEX_UNLOCKED);
 
                pid = fork();
                if (pid) {
                        close(sk);
                        fio_server_add_conn_pid(&conn_list, pid);
-                       pthread_setspecific(sk_out_key, sk_out);
                        continue;
                }
 
-               /* exits */
-               get_my_addr_str(sk); /* if error, it's already logged, non-fatal */
+               /* if error, it's already logged, non-fatal */
+               get_my_addr_str(sk);
+
+               /*
+                * Assign sk_out here, it'll be dropped in handle_connection()
+                * since that function calls _exit() when done
+                */
+               sk_out_assign(sk_out);
                handle_connection(sk_out);
        }
 
-#if 0
-       fio_mutex_remove(sk_out->lock);
-       fio_mutex_remove(sk_out->wait);
-       sfree(sk_out);
-       pthread_setspecific(sk_out_key, NULL);
-#endif
-
        return exitval;
 }
 
@@ -1352,6 +1458,8 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
 {
        struct cmd_ts_pdu p;
        int i, j;
+       void *ss_buf;
+       uint64_t *ss_iops, *ss_bw;
 
        dprint(FD_NET, "server sending end stats\n");
 
@@ -1435,9 +1543,37 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
        for (i = 0; i < p.ts.nr_block_infos; i++)
                p.ts.block_infos[i] = le32_to_cpu(ts->block_infos[i]);
 
+       p.ts.ss_dur             = cpu_to_le64(ts->ss_dur);
+       p.ts.ss_state           = cpu_to_le32(ts->ss_state);
+       p.ts.ss_head            = cpu_to_le32(ts->ss_head);
+       p.ts.ss_limit.u.i       = cpu_to_le64(fio_double_to_uint64(ts->ss_limit.u.f));
+       p.ts.ss_slope.u.i       = cpu_to_le64(fio_double_to_uint64(ts->ss_slope.u.f));
+       p.ts.ss_deviation.u.i   = cpu_to_le64(fio_double_to_uint64(ts->ss_deviation.u.f));
+       p.ts.ss_criterion.u.i   = cpu_to_le64(fio_double_to_uint64(ts->ss_criterion.u.f));
+
        convert_gs(&p.rs, rs);
 
-       fio_net_queue_cmd(FIO_NET_CMD_TS, &p, sizeof(p), NULL, SK_F_COPY);
+       dprint(FD_NET, "ts->ss_state = %d\n", ts->ss_state);
+       if (ts->ss_state & __FIO_SS_DATA) {
+               dprint(FD_NET, "server sending steadystate ring buffers\n");
+
+               ss_buf = malloc(sizeof(p) + 2*ts->ss_dur*sizeof(uint64_t));
+
+               memcpy(ss_buf, &p, sizeof(p));
+
+               ss_iops = (uint64_t *) ((struct cmd_ts_pdu *)ss_buf + 1);
+               ss_bw = ss_iops + (int) ts->ss_dur;
+               for (i = 0; i < ts->ss_dur; i++) {
+                       ss_iops[i] = cpu_to_le64(ts->ss_iops_data[i]);
+                       ss_bw[i] = cpu_to_le64(ts->ss_bw_data[i]);
+               }
+
+               fio_net_queue_cmd(FIO_NET_CMD_TS, ss_buf, sizeof(p) + 2*ts->ss_dur*sizeof(uint64_t), NULL, SK_F_COPY);
+
+               free(ss_buf);
+       }
+       else
+               fio_net_queue_cmd(FIO_NET_CMD_TS, &p, sizeof(p), NULL, SK_F_COPY);
 }
 
 void fio_server_send_gs(struct group_run_stats *rs)
@@ -1450,6 +1586,47 @@ void fio_server_send_gs(struct group_run_stats *rs)
        fio_net_queue_cmd(FIO_NET_CMD_GS, &gs, sizeof(gs), NULL, SK_F_COPY);
 }
 
+void fio_server_send_job_options(struct flist_head *opt_list,
+                                unsigned int groupid)
+{
+       struct cmd_job_option pdu;
+       struct flist_head *entry;
+
+       if (flist_empty(opt_list))
+               return;
+
+       flist_for_each(entry, opt_list) {
+               struct print_option *p;
+               size_t len;
+
+               p = flist_entry(entry, struct print_option, list);
+               memset(&pdu, 0, sizeof(pdu));
+
+               if (groupid == -1U) {
+                       pdu.global = __cpu_to_le16(1);
+                       pdu.groupid = 0;
+               } else {
+                       pdu.global = 0;
+                       pdu.groupid = cpu_to_le32(groupid);
+               }
+               len = strlen(p->name);
+               if (len >= sizeof(pdu.name)) {
+                       len = sizeof(pdu.name) - 1;
+                       pdu.truncated = __cpu_to_le16(1);
+               }
+               memcpy(pdu.name, p->name, len);
+               if (p->value) {
+                       len = strlen(p->value);
+                       if (len >= sizeof(pdu.value)) {
+                               len = sizeof(pdu.value) - 1;
+                               pdu.truncated = __cpu_to_le16(1);
+                       }
+                       memcpy(pdu.value, p->value, len);
+               }
+               fio_net_queue_cmd(FIO_NET_CMD_JOB_OPT, &pdu, sizeof(pdu), NULL, SK_F_COPY);
+       }
+}
+
 static void convert_agg(struct disk_util_agg *dst, struct disk_util_agg *src)
 {
        int i;
@@ -1506,107 +1683,214 @@ void fio_server_send_du(void)
        }
 }
 
-static int fio_send_iolog_gz(struct sk_entry *first, struct io_log *log)
-{
-       int ret = 0;
 #ifdef CONFIG_ZLIB
-       struct sk_entry *entry;
-       z_stream stream;
+static int __fio_append_iolog_gz(struct sk_entry *first, struct io_log *log,
+                                struct io_logs *cur_log, z_stream *stream)
+{
+       unsigned int this_len;
        void *out_pdu;
+       int ret;
 
-       /*
-        * 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->next_in = (void *) cur_log->log;
+       stream->avail_in = cur_log->nr_samples * log_entry_sz(log);
+
+       do {
+               struct sk_entry *entry;
+
+               /*
+                * 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->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
+               stream->next_out = out_pdu;
+               ret = deflate(stream, Z_BLOCK);
+               /* may be Z_OK, or Z_STREAM_END */
+               if (ret < 0) {
+                       free(out_pdu);
+                       return 1;
+               }
+
+               this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
+
+               entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
+                                        NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
+               flist_add_tail(&entry->list, &first->next);
+       } while (stream->avail_in);
+
+       return 0;
+}
+
+static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
+{
+       int ret = 0;
+       z_stream stream;
 
+       memset(&stream, 0, sizeof(stream));
        stream.zalloc = Z_NULL;
        stream.zfree = Z_NULL;
        stream.opaque = Z_NULL;
 
-       if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK) {
-               ret = 1;
-               goto err;
+       if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK)
+               return 1;
+
+       while (!flist_empty(&log->io_logs)) {
+               struct io_logs *cur_log;
+
+               cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
+               flist_del_init(&cur_log->list);
+
+               ret = __fio_append_iolog_gz(first, log, cur_log, &stream);
+               if (ret)
+                       break;
        }
 
-       stream.next_in = (void *) log->log;
-       stream.avail_in = log->nr_samples * log_entry_sz(log);
+       ret = deflate(&stream, Z_FINISH);
 
-       do {
+       while (ret != Z_STREAM_END) {
+               struct sk_entry *entry;
                unsigned int this_len;
+               void *out_pdu;
 
+               out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
                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;
+               if (ret < 0) {
+                       free(out_pdu);
+                       break;
+               }
 
                this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream.avail_out;
 
                entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, out_pdu, this_len,
-                                               NULL, SK_F_FREE | SK_F_VEC);
+                                        NULL, SK_F_VEC | SK_F_INLINE | SK_F_FREE);
                flist_add_tail(&entry->list, &first->next);
-       } while (stream.avail_in);
+       } while (ret != Z_STREAM_END);
+
+       ret = deflateEnd(&stream);
+       if (ret == Z_OK)
+               return 0;
 
-err_zlib:
-       deflateEnd(&stream);
-err:
-       free(out_pdu);
+       return 1;
+}
+#else
+static int fio_append_iolog_gz(struct sk_entry *first, struct io_log *log)
+{
+       return 1;
+}
 #endif
-       return ret;
+
+static int fio_append_gz_chunks(struct sk_entry *first, struct io_log *log)
+{
+       struct sk_entry *entry;
+       struct flist_head *node;
+
+       pthread_mutex_lock(&log->chunk_lock);
+       flist_for_each(node, &log->chunk_list) {
+               struct iolog_compress *c;
+
+               c = flist_entry(node, struct iolog_compress, list);
+               entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, c->buf, c->len,
+                                               NULL, SK_F_VEC | SK_F_INLINE);
+               flist_add_tail(&entry->list, &first->next);
+       }
+       pthread_mutex_unlock(&log->chunk_lock);
+
+       return 0;
+}
+
+static int fio_append_text_log(struct sk_entry *first, struct io_log *log)
+{
+       struct sk_entry *entry;
+
+       while (!flist_empty(&log->io_logs)) {
+               struct io_logs *cur_log;
+               size_t size;
+
+               cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
+               flist_del_init(&cur_log->list);
+
+               size = cur_log->nr_samples * log_entry_sz(log);
+
+               entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, cur_log->log, size,
+                                               NULL, SK_F_VEC | SK_F_INLINE);
+               flist_add_tail(&entry->list, &first->next);
+       }
+
+       return 0;
 }
 
 int fio_send_iolog(struct thread_data *td, struct io_log *log, const char *name)
 {
        struct cmd_iolog_pdu pdu;
        struct sk_entry *first;
-       int i, ret = 0;
+       struct flist_head *entry;
+       int ret = 0;
 
-       pdu.nr_samples = cpu_to_le64(log->nr_samples);
+       pdu.nr_samples = cpu_to_le64(iolog_nr_samples(log));
        pdu.thread_number = cpu_to_le32(td->thread_number);
        pdu.log_type = cpu_to_le32(log->log_type);
-       pdu.compressed = cpu_to_le32(use_zlib);
+
+       if (!flist_empty(&log->chunk_list))
+               pdu.compressed = __cpu_to_le32(STORE_COMPRESSED);
+       else if (use_zlib)
+               pdu.compressed = __cpu_to_le32(XMIT_COMPRESSED);
+       else
+               pdu.compressed = 0;
 
        strncpy((char *) pdu.name, name, FIO_NET_NAME_MAX);
        pdu.name[FIO_NET_NAME_MAX - 1] = '\0';
 
-       for (i = 0; i < log->nr_samples; i++) {
-               struct io_sample *s = get_sample(log, i);
+       /*
+        * We can't do this for a pre-compressed log, but for that case,
+        * log->nr_samples is zero anyway.
+        */
+       flist_for_each(entry, &log->io_logs) {
+               struct io_logs *cur_log;
+               int i;
+
+               cur_log = flist_entry(entry, struct io_logs, list);
+
+               for (i = 0; i < cur_log->nr_samples; i++) {
+                       struct io_sample *s = get_sample(log, cur_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);
+                       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);
 
-               if (log->log_offset) {
-                       struct io_sample_offset *so = (void *) s;
+                       if (log->log_offset) {
+                               struct io_sample_offset *so = (void *) s;
 
-                       so->offset = cpu_to_le64(so->offset);
+                               so->offset = cpu_to_le64(so->offset);
+                       }
                }
        }
 
        /*
         * Assemble header entry first
         */
-       first = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, &pdu, sizeof(pdu), NULL, SK_F_COPY | SK_F_VEC);
+       first = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, &pdu, sizeof(pdu), NULL, SK_F_VEC | SK_F_INLINE | SK_F_COPY);
 
        /*
-        * Now append actual log entries. Compress if we can, otherwise just
-        * plain text output.
+        * Now append actual log entries. If log compression was enabled on
+        * the job, just send out the compressed chunks directly. If we
+        * have a plain log, compress if we can, then send. Otherwise, send
+        * the plain text output.
         */
-       if (use_zlib)
-               ret = fio_send_iolog_gz(first, log);
-       else {
-               struct sk_entry *entry;
-
-               entry = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, log->log,
-                                       log->nr_samples * log_entry_sz(log),
-                                       NULL, SK_F_FREE | SK_F_VEC);
-               flist_add_tail(&entry->list, &first->next);
-       }
+       if (!flist_empty(&log->chunk_list))
+               ret = fio_append_gz_chunks(first, log);
+       else if (use_zlib)
+               ret = fio_append_iolog_gz(first, log);
+       else
+               ret = fio_append_text_log(first, log);
 
+       fio_net_queue_entry(first);
        return ret;
 }
 
@@ -1619,7 +1903,8 @@ void fio_server_send_add_job(struct thread_data *td)
        pdu.groupid = cpu_to_le32(td->groupid);
        convert_thread_options_to_net(&pdu.top, &td->o);
 
-       fio_net_queue_cmd(FIO_NET_CMD_ADD_JOB, &pdu, sizeof(pdu), NULL, SK_F_COPY);
+       fio_net_queue_cmd(FIO_NET_CMD_ADD_JOB, &pdu, sizeof(pdu), NULL,
+                               SK_F_COPY);
 }
 
 void fio_server_send_start(struct thread_data *td)
@@ -1632,20 +1917,21 @@ void fio_server_send_start(struct thread_data *td)
 }
 
 int fio_server_get_verify_state(const char *name, int threadnumber,
-                               void **datap, int *version)
+                               void **datap)
 {
        struct thread_io_list *s;
        struct cmd_sendfile out;
        struct cmd_reply *rep;
        uint64_t tag;
        void *data;
+       int ret;
 
        dprint(FD_NET, "server: request verify state\n");
 
        rep = smalloc(sizeof(*rep));
        if (!rep) {
                log_err("fio: smalloc pool too small\n");
-               return 1;
+               return ENOMEM;
        }
 
        __fio_mutex_init(&rep->lock, FIO_MUTEX_LOCKED);
@@ -1655,23 +1941,27 @@ int fio_server_get_verify_state(const char *name, int threadnumber,
        verify_state_gen_name((char *) out.path, sizeof(out.path), name, me,
                                threadnumber);
        tag = (uint64_t) (uintptr_t) rep;
-       fio_net_queue_cmd(FIO_NET_CMD_SENDFILE, &out, sizeof(out), &tag, SK_F_COPY);
+       fio_net_queue_cmd(FIO_NET_CMD_SENDFILE, &out, sizeof(out), &tag,
+                               SK_F_COPY);
 
        /*
         * Wait for the backend to receive the reply
         */
        if (fio_mutex_down_timeout(&rep->lock, 10000)) {
                log_err("fio: timed out waiting for reply\n");
+               ret = ETIMEDOUT;
                goto fail;
        }
 
        if (rep->error) {
-               log_err("fio: failure on receiving state file: %s\n", strerror(rep->error));
+               log_err("fio: failure on receiving state file %s: %s\n",
+                               out.path, strerror(rep->error));
+               ret = rep->error;
 fail:
                *datap = NULL;
                sfree(rep);
                fio_net_queue_quit();
-               return 1;
+               return ret;
        }
 
        /*
@@ -1679,12 +1969,15 @@ fail:
         * the header, and the thread_io_list checksum
         */
        s = rep->data + sizeof(struct verify_state_hdr);
-       if (verify_state_hdr(rep->data, s, version))
+       if (verify_state_hdr(rep->data, s)) {
+               ret = EILSEQ;
                goto fail;
+       }
 
        /*
         * Don't need the header from now, copy just the thread_io_list
         */
+       ret = 0;
        rep->size -= sizeof(struct verify_state_hdr);
        data = malloc(rep->size);
        memcpy(data, s, rep->size);
@@ -1693,7 +1986,7 @@ fail:
        sfree(rep->data);
        __fio_mutex_remove(&rep->lock);
        sfree(rep);
-       return 0;
+       return ret;
 }
 
 static int fio_init_server_ip(void)
@@ -1721,11 +2014,10 @@ static int fio_init_server_ip(void)
                return -1;
        }
 #ifdef SO_REUSEPORT
-       if (setsockopt(sk, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt)) < 0) {
-               log_err("fio: setsockopt(REUSEPORT): %s\n", strerror(errno));
-               close(sk);
-               return -1;
-       }
+       /*
+        * Not fatal if fails, so just ignore it if that happens
+        */
+       setsockopt(sk, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt));
 #endif
 
        if (use_ipv6) {
@@ -1827,7 +2119,7 @@ static int fio_init_server_connection(void)
 
        log_info("fio: server listening on %s\n", bind_str);
 
-       if (listen(sk, 0) < 0) {
+       if (listen(sk, 4) < 0) {
                log_err("fio: listen: %s\n", strerror(errno));
                close(sk);
                return -1;
@@ -2018,6 +2310,22 @@ static void set_sig_handlers(void)
        sigaction(SIGINT, &act, NULL);
 }
 
+void fio_server_destroy_sk_key(void)
+{
+       pthread_key_delete(sk_out_key);
+}
+
+int fio_server_create_sk_key(void)
+{
+       if (pthread_key_create(&sk_out_key, NULL)) {
+               log_err("fio: can't create sk_out backend key\n");
+               return 1;
+       }
+
+       pthread_setspecific(sk_out_key, NULL);
+       return 0;
+}
+
 static int fio_server(void)
 {
        int sk, ret;
@@ -2033,9 +2341,6 @@ static int fio_server(void)
 
        set_sig_handlers();
 
-       if (pthread_key_create(&sk_out_key, NULL))
-               log_err("fio: can't create sk_out backend key\n");
-
        ret = accept_loop(sk);
 
        close(sk);