X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=server.c;h=6d5d4ea36ebaf9a774753d6d62c5c50d519c55f7;hp=ab3bece1f61f47e3f7e8dd704a5acf64ac50223d;hb=c58390119ccf711ab66e53aff6aa31507c509900;hpb=8d7b618a261e7642769b8f0bf2cf6649bb383330

diff --git a/server.c b/server.c
index ab3bece1..6d5d4ea3 100644
--- a/server.c
+++ b/server.c
@@ -4,7 +4,6 @@
 #include <unistd.h>
 #include <limits.h>
 #include <errno.h>
-#include <fcntl.h>
 #include <sys/poll.h>
 #include <sys/types.h>
 #include <sys/wait.h>
@@ -22,15 +21,46 @@
 #endif
 
 #include "fio.h"
+#include "options.h"
 #include "server.h"
 #include "crc/crc16.h"
 #include "lib/ieee754.h"
+#include "verify.h"
+#include "smalloc.h"
 
 int fio_net_port = FIO_NET_PORT;
 
 int exit_backend = 0;
 
-static int server_fd = -1;
+enum {
+	SK_F_FREE	= 1,
+	SK_F_COPY	= 2,
+	SK_F_SIMPLE	= 4,
+	SK_F_VEC	= 8,
+	SK_F_INLINE	= 16,
+};
+
+struct sk_entry {
+	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 tag;
+	struct flist_head next;	/* Other sk_entry's, if linked command */
+};
+
+struct sk_out {
+	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;
 static char *bind_sock;
 static struct sockaddr_in saddr_in;
@@ -42,6 +72,9 @@ static unsigned int has_zlib = 1;
 static unsigned int has_zlib = 0;
 #endif
 static unsigned int use_zlib;
+static char me[128];
+
+static pthread_key_t sk_out_key;
 
 struct fio_fork_item {
 	struct flist_head list;
@@ -51,6 +84,13 @@ struct fio_fork_item {
 	pid_t pid;
 };
 
+struct cmd_reply {
+	struct fio_mutex lock;
+	void *data;
+	size_t size;
+	int error;
+};
+
 static const char *fio_server_ops[FIO_NET_CMD_NR] = {
 	"",
 	"QUIT",
@@ -68,10 +108,93 @@ static const char *fio_server_ops[FIO_NET_CMD_NR] = {
 	"DISK_UTIL",
 	"SERVER_START",
 	"ADD_JOB",
-	"CMD_RUN",
-	"CMD_IOLOG",
+	"RUN",
+	"IOLOG",
+	"UPDATE_JOB",
+	"LOAD_FILE",
+	"VTRIGGER",
+	"SENDFILE",
+	"JOB_OPT",
 };
 
+static void sk_lock(struct sk_out *sk_out)
+{
+	fio_mutex_down(&sk_out->lock);
+}
+
+static void sk_unlock(struct sk_out *sk_out)
+{
+	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)
 {
 	static char buf[32];
@@ -128,13 +251,10 @@ static int fio_sendv_data(int sk, struct iovec *iov, int count)
 	if (!total_len)
 		return 0;
 
-	if (errno)
-		return -errno;
-
 	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 };
 
@@ -143,10 +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;
@@ -156,9 +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);
 
@@ -207,16 +336,16 @@ 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, *cmdret = NULL;
+	struct fio_net_cmd cmd, *tmp, *cmdret = NULL;
 	size_t cmd_size = 0, pdu_offset = 0;
 	uint16_t crc;
 	int ret, first = 1;
 	void *pdu = NULL;
 
 	do {
-		ret = fio_recv_data(sk, &cmd, sizeof(cmd));
+		ret = fio_recv_data(sk, &cmd, sizeof(cmd), wait);
 		if (ret)
 			break;
 
@@ -232,7 +361,19 @@ struct fio_net_cmd *fio_net_recv_cmd(int sk)
 		} else
 			cmd_size += cmd.pdu_len;
 
-		cmdret = realloc(cmdret, cmd_size);
+		if (cmd_size / 1024 > FIO_SERVER_MAX_CMD_MB * 1024) {
+			log_err("fio: cmd+pdu too large (%llu)\n", (unsigned long long) cmd_size);
+			ret = 1;
+			break;
+		}
+
+		tmp = realloc(cmdret, cmd_size);
+		if (!tmp) {
+			log_err("fio: server failed allocating cmd\n");
+			ret = 1;
+			break;
+		}
+		cmdret = tmp;
 
 		if (first)
 			memcpy(cmdret, &cmd, sizeof(cmd));
@@ -248,7 +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;
 
@@ -274,14 +415,14 @@ struct fio_net_cmd *fio_net_recv_cmd(int sk)
 		/* zero-terminate text input */
 		if (cmdret->pdu_len) {
 			if (cmdret->opcode == FIO_NET_CMD_TEXT) {
-				struct cmd_text_pdu *pdu = (struct cmd_text_pdu *) cmdret->payload;
-				char *buf = (char *) pdu->buf;
+				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) {
-				struct cmd_job_pdu *pdu = (struct cmd_job_pdu *) cmdret->payload;
-				char *buf = (char *) pdu->buf;
-				int len = le32_to_cpu(pdu->buf_len);
+				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[len] = '\0';
 			}
@@ -308,7 +449,7 @@ static uint64_t alloc_reply(uint64_t tag, uint16_t opcode)
 
 	reply = calloc(1, sizeof(*reply));
 	INIT_FLIST_HEAD(&reply->list);
-	gettimeofday(&reply->tv, NULL);
+	fio_gettime(&reply->tv, NULL);
 	reply->saved_tag = tag;
 	reply->opcode = opcode;
 
@@ -323,7 +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;
 
@@ -333,7 +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);
 }
@@ -390,6 +531,61 @@ int fio_net_send_cmd(int fd, uint16_t opcode, const void *buf, off_t size,
 	return ret;
 }
 
+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;
+
+	entry = smalloc(sizeof(*entry));
+	INIT_FLIST_HEAD(&entry->next);
+	entry->opcode = opcode;
+	if (flags & SK_F_COPY) {
+		entry->buf = smalloc(size);
+		memcpy(entry->buf, buf, size);
+	} else
+		entry->buf = buf;
+
+	entry->size = size;
+	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);
+
+	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);
+	}
+}
+
+static int fio_net_queue_cmd(uint16_t opcode, void *buf, off_t size,
+			     uint64_t *tagptr, int flags)
+{
+	struct sk_entry *entry;
+
+	entry = fio_net_prep_cmd(opcode, buf, size, tagptr, flags);
+	if (entry) {
+		fio_net_queue_entry(entry);
+		return 0;
+	}
+
+	return 1;
+}
+
 static int fio_net_send_simple_stack_cmd(int sk, uint16_t opcode, uint64_t tag)
 {
 	struct fio_net_cmd cmd;
@@ -426,6 +622,13 @@ int fio_net_send_simple_cmd(int sk, uint16_t opcode, uint64_t tag,
 	return 0;
 }
 
+static int fio_net_queue_quit(void)
+{
+	dprint(FD_NET, "server: sending quit\n");
+
+	return fio_net_queue_cmd(FIO_NET_CMD_QUIT, NULL, 0, NULL, SK_F_SIMPLE);
+}
+
 int fio_net_send_quit(int sk)
 {
 	dprint(FD_NET, "server: sending quit\n");
@@ -433,8 +636,7 @@ int fio_net_send_quit(int sk)
 	return fio_net_send_simple_cmd(sk, FIO_NET_CMD_QUIT, 0, NULL);
 }
 
-static int fio_net_send_ack(int sk, struct fio_net_cmd *cmd, int error,
-			    int signal)
+static int fio_net_send_ack(struct fio_net_cmd *cmd, int error, int signal)
 {
 	struct cmd_end_pdu epdu;
 	uint64_t tag = 0;
@@ -444,13 +646,13 @@ static int fio_net_send_ack(int sk, struct fio_net_cmd *cmd, int error,
 
 	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);
+	return fio_net_queue_cmd(FIO_NET_CMD_STOP, &epdu, sizeof(epdu), &tag, SK_F_COPY);
 }
 
-int fio_net_send_stop(int sk, int error, int signal)
+static int fio_net_queue_stop(int error, int signal)
 {
 	dprint(FD_NET, "server: sending stop (%d, %d)\n", error, signal);
-	return fio_net_send_ack(sk, NULL, error, signal);
+	return fio_net_send_ack(NULL, error, signal);
 }
 
 static void fio_server_add_fork_item(pid_t pid, struct flist_head *list)
@@ -501,20 +703,23 @@ static void fio_server_check_fork_item(struct fio_fork_item *ffi)
 	}
 }
 
-static void fio_server_fork_item_done(struct fio_fork_item *ffi)
+static void fio_server_fork_item_done(struct fio_fork_item *ffi, bool stop)
 {
 	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);
+	if (stop) {
+		fio_net_queue_stop(ffi->exitval, ffi->signal);
+		fio_net_queue_quit();
+	}
+
 	flist_del(&ffi->list);
 	free(ffi);
 }
 
-static void fio_server_check_fork_items(struct flist_head *list)
+static void fio_server_check_fork_items(struct flist_head *list, bool stop)
 {
 	struct flist_head *entry, *tmp;
 	struct fio_fork_item *ffi;
@@ -525,25 +730,51 @@ static void fio_server_check_fork_items(struct flist_head *list)
 		fio_server_check_fork_item(ffi);
 
 		if (ffi->exited)
-			fio_server_fork_item_done(ffi);
+			fio_server_fork_item_done(ffi, stop);
 	}
 }
 
 static void fio_server_check_jobs(struct flist_head *job_list)
 {
-	fio_server_check_fork_items(job_list);
+	fio_server_check_fork_items(job_list, true);
 }
 
 static void fio_server_check_conns(struct flist_head *conn_list)
 {
-	fio_server_check_fork_items(conn_list);
+	fio_server_check_fork_items(conn_list, false);
 }
 
-static int handle_run_cmd(struct flist_head *job_list, struct fio_net_cmd *cmd)
+static int handle_load_file_cmd(struct fio_net_cmd *cmd)
+{
+	struct cmd_load_file_pdu *pdu = (struct cmd_load_file_pdu *) cmd->payload;
+	void *file_name = pdu->file;
+	struct cmd_start_pdu spdu;
+
+	dprint(FD_NET, "server: loading local file %s\n", (char *) file_name);
+
+	pdu->name_len = le16_to_cpu(pdu->name_len);
+	pdu->client_type = le16_to_cpu(pdu->client_type);
+
+	if (parse_jobs_ini(file_name, 0, 0, pdu->client_type)) {
+		fio_net_queue_quit();
+		return -1;
+	}
+
+	spdu.jobs = cpu_to_le32(thread_number);
+	spdu.stat_outputs = cpu_to_le32(stat_number);
+	fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
+	return 0;
+}
+
+static int handle_run_cmd(struct sk_out *sk_out, struct flist_head *job_list,
+			  struct fio_net_cmd *cmd)
 {
 	pid_t pid;
 	int ret;
 
+	sk_out_assign(sk_out);
+
+	fio_time_init();
 	set_genesis_time();
 
 	pid = fork();
@@ -552,8 +783,9 @@ static int handle_run_cmd(struct flist_head *job_list, struct fio_net_cmd *cmd)
 		return 0;
 	}
 
-	ret = fio_backend();
+	ret = fio_backend(sk_out);
 	free_threads_shm();
+	sk_out_drop();
 	_exit(ret);
 }
 
@@ -567,13 +799,14 @@ static int handle_job_cmd(struct fio_net_cmd *cmd)
 	pdu->client_type = le32_to_cpu(pdu->client_type);
 
 	if (parse_jobs_ini(buf, 1, 0, pdu->client_type)) {
-		fio_net_send_quit(server_fd);
+		fio_net_queue_quit();
 		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);
+
+	fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
 	return 0;
 }
 
@@ -604,7 +837,7 @@ static int handle_jobline_cmd(struct fio_net_cmd *cmd)
 	}
 
 	if (parse_cmd_line(clp->lines, argv, clp->client_type)) {
-		fio_net_send_quit(server_fd);
+		fio_net_queue_quit();
 		free(argv);
 		return -1;
 	}
@@ -613,7 +846,8 @@ static int handle_jobline_cmd(struct fio_net_cmd *cmd)
 
 	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);
+
+	fio_net_queue_cmd(FIO_NET_CMD_START, &spdu, sizeof(spdu), NULL, SK_F_COPY);
 	return 0;
 }
 
@@ -625,6 +859,8 @@ static int handle_probe_cmd(struct fio_net_cmd *cmd)
 
 	dprint(FD_NET, "server: sending probe reply\n");
 
+	strcpy(me, (char *) pdu->server);
+
 	memset(&probe, 0, sizeof(probe));
 	gethostname((char *) probe.hostname, sizeof(probe.hostname));
 #ifdef CONFIG_BIG_ENDIAN
@@ -648,63 +884,60 @@ static int handle_probe_cmd(struct fio_net_cmd *cmd)
 		use_zlib = 0;
 	}
 
-	return fio_net_send_cmd(server_fd, FIO_NET_CMD_PROBE, &probe, sizeof(probe), &tag, NULL);
+	return fio_net_queue_cmd(FIO_NET_CMD_PROBE, &probe, sizeof(probe), &tag, SK_F_COPY);
 }
 
 static int handle_send_eta_cmd(struct fio_net_cmd *cmd)
 {
 	struct jobs_eta *je;
-	size_t size;
 	uint64_t tag = cmd->tag;
+	size_t size;
 	int i;
 
-	if (!thread_number)
-		return 0;
+	dprint(FD_NET, "server sending status\n");
 
-	size = sizeof(*je) + thread_number * sizeof(char) + 1;
-	je = malloc(size);
-	memset(je, 0, size);
+	/*
+	 * Fake ETA return if we don't have a local one, otherwise the client
+	 * will end up timing out waiting for a response to the ETA request
+	 */
+	je = get_jobs_eta(true, &size);
+	if (!je) {
+		size = sizeof(*je);
+		je = calloc(1, size);
+	} else {
+		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 < DDIR_RWDIR_CNT; i++) {
+			je->m_rate[i]	= cpu_to_le64(je->m_rate[i]);
+			je->t_rate[i]	= cpu_to_le64(je->t_rate[i]);
+			je->m_iops[i]	= cpu_to_le32(je->m_iops[i]);
+			je->t_iops[i]	= cpu_to_le32(je->t_iops[i]);
+			je->rate[i]	= cpu_to_le64(je->rate[i]);
+			je->iops[i]	= cpu_to_le32(je->iops[i]);
+		}
 
-	if (!calc_thread_status(je, 1)) {
-		free(je);
-		return 0;
+		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);
 	}
 
-	dprint(FD_NET, "server sending status\n");
-
-	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 < 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]);
-		je->t_iops[i]	= cpu_to_le32(je->t_iops[i]);
-		je->rate[i]	= cpu_to_le32(je->rate[i]);
-		je->iops[i]	= cpu_to_le32(je->iops[i]);
-	}
-
-	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, &tag, NULL);
-	free(je);
+	fio_net_queue_cmd(FIO_NET_CMD_ETA, je, size, &tag, SK_F_FREE);
 	return 0;
 }
 
-static int send_update_job_reply(int fd, uint64_t __tag, int error)
+static int send_update_job_reply(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);
+	return fio_net_queue_cmd(FIO_NET_CMD_UPDATE_JOB, &pdu_error, sizeof(pdu_error), &tag, SK_F_COPY);
 }
 
 static int handle_update_job_cmd(struct fio_net_cmd *cmd)
@@ -718,17 +951,41 @@ static int handle_update_job_cmd(struct fio_net_cmd *cmd)
 	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);
+		send_update_job_reply(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);
+	send_update_job_reply(cmd->tag, 0);
 	return 0;
 }
 
-static int handle_command(struct flist_head *job_list, struct fio_net_cmd *cmd)
+static int handle_trigger_cmd(struct fio_net_cmd *cmd)
+{
+	struct cmd_vtrigger_pdu *pdu = (struct cmd_vtrigger_pdu *) cmd->payload;
+	char *buf = (char *) pdu->cmd;
+	struct all_io_list *rep;
+	size_t sz;
+
+	pdu->len = le16_to_cpu(pdu->len);
+	buf[pdu->len] = '\0';
+
+	rep = get_all_io_list(IO_LIST_ALL, &sz);
+	if (!rep) {
+		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 | SK_F_INLINE);
+	} else
+		fio_net_queue_cmd(FIO_NET_CMD_VTRIGGER, rep, sz, NULL, SK_F_FREE | SK_F_INLINE);
+
+	exec_trigger(buf);
+	return 0;
+}
+
+static int handle_command(struct sk_out *sk_out, struct flist_head *job_list,
+			  struct fio_net_cmd *cmd)
 {
 	int ret;
 
@@ -739,10 +996,14 @@ static int handle_command(struct flist_head *job_list, struct fio_net_cmd *cmd)
 	switch (cmd->opcode) {
 	case FIO_NET_CMD_QUIT:
 		fio_terminate_threads(TERMINATE_ALL);
-		return -1;
+		ret = 0;
+		break;
 	case FIO_NET_CMD_EXIT:
 		exit_backend = 1;
 		return -1;
+	case FIO_NET_CMD_LOAD_FILE:
+		ret = handle_load_file_cmd(cmd);
+		break;
 	case FIO_NET_CMD_JOB:
 		ret = handle_job_cmd(cmd);
 		break;
@@ -756,11 +1017,40 @@ static int handle_command(struct flist_head *job_list, struct fio_net_cmd *cmd)
 		ret = handle_send_eta_cmd(cmd);
 		break;
 	case FIO_NET_CMD_RUN:
-		ret = handle_run_cmd(job_list, cmd);
+		ret = handle_run_cmd(sk_out, job_list, cmd);
 		break;
 	case FIO_NET_CMD_UPDATE_JOB:
 		ret = handle_update_job_cmd(cmd);
 		break;
+	case FIO_NET_CMD_VTRIGGER:
+		ret = handle_trigger_cmd(cmd);
+		break;
+	case FIO_NET_CMD_SENDFILE: {
+		struct cmd_sendfile_reply *in;
+		struct cmd_reply *rep;
+
+		rep = (struct cmd_reply *) (uintptr_t) cmd->tag;
+
+		in = (struct cmd_sendfile_reply *) cmd->payload;
+		in->size = le32_to_cpu(in->size);
+		in->error = le32_to_cpu(in->error);
+		if (in->error) {
+			ret = 1;
+			rep->error = in->error;
+		} else {
+			ret = 0;
+			rep->data = smalloc(in->size);
+			if (!rep->data) {
+				ret = 1;
+				rep->error = ENOMEM;
+			} else {
+				rep->size = in->size;
+				memcpy(rep->data, in->data, in->size);
+			}
+		}
+		fio_mutex_up(&rep->lock);
+		break;
+		}
 	default:
 		log_err("fio: unknown opcode: %s\n", fio_server_op(cmd->opcode));
 		ret = 1;
@@ -769,19 +1059,151 @@ static int handle_command(struct flist_head *job_list, struct fio_net_cmd *cmd)
 	return ret;
 }
 
-static int handle_connection(int sk)
+/*
+ * 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 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);
+
+	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 ret;
+}
+
+static void finish_entry(struct sk_entry *entry)
+{
+	if (entry->flags & SK_F_FREE)
+		free(entry->buf);
+	else if (entry->flags & SK_F_COPY)
+		sfree(entry->buf);
+
+	sfree(entry);
+}
+
+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;
+}
+
+static int send_vec_entry(struct sk_out *sk_out, struct sk_entry *first)
+{
+	unsigned int flags;
+	int ret;
+
+	entry_set_flags(first, &first->next, &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;
+
+		next = flist_first_entry(&first->next, struct sk_entry, list);
+		flist_del_init(&next->list);
+
+		entry_set_flags(next, &first->next, &flags);
+
+		ret += fio_send_cmd_ext_pdu(sk_out->sk, next->opcode, next->buf,
+						next->size, next->tag, flags);
+		finish_entry(next);
+	}
+
+	return ret;
+}
+
+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);
+	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);
+	}
+
+	fio_mutex_up(&sk_out->xmit);
+
+	if (ret)
+		log_err("fio: failed handling cmd %s\n", fio_server_op(entry->opcode));
+
+	finish_entry(entry);
+	return ret;
+}
+
+static int handle_xmits(struct sk_out *sk_out)
+{
+	struct sk_entry *entry;
+	FLIST_HEAD(list);
+	int ret = 0;
+
+	sk_lock(sk_out);
+	if (flist_empty(&sk_out->list)) {
+		sk_unlock(sk_out);
+		return 0;
+	}
+
+	flist_splice_init(&sk_out->list, &list);
+	sk_unlock(sk_out);
+
+	while (!flist_empty(&list)) {
+		entry = flist_entry(list.next, struct sk_entry, list);
+		flist_del(&entry->list);
+		ret += handle_sk_entry(sk_out, entry);
+	}
+
+	return ret;
+}
+
+static int handle_connection(struct sk_out *sk_out)
 {
 	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 = {
-			.fd	= sk,
+			.fd	= sk_out->sk,
 			.events	= POLLIN,
 		};
 
@@ -792,7 +1214,9 @@ static int handle_connection(int sk)
 			if (!flist_empty(&job_list))
 				timeout = 100;
 
-			ret = poll(&pfd, 1, timeout);
+			handle_xmits(sk_out);
+
+			ret = poll(&pfd, 1, 0);
 			if (ret < 0) {
 				if (errno == EINTR)
 					break;
@@ -800,6 +1224,7 @@ static int handle_connection(int sk)
 				break;
 			} else if (!ret) {
 				fio_server_check_jobs(&job_list);
+				fio_mutex_down_timeout(&sk_out->wait, timeout);
 				continue;
 			}
 
@@ -816,13 +1241,13 @@ static int handle_connection(int sk)
 		if (ret < 0)
 			break;
 
-		cmd = fio_net_recv_cmd(sk);
+		cmd = fio_net_recv_cmd(sk_out->sk, true);
 		if (!cmd) {
 			ret = -1;
 			break;
 		}
 
-		ret = handle_command(&job_list, cmd);
+		ret = handle_command(sk_out, &job_list, cmd);
 		if (ret)
 			break;
 
@@ -833,26 +1258,66 @@ static int handle_connection(int sk)
 	if (cmd)
 		free(cmd);
 
-	close(sk);
+	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 */
+
+static int get_my_addr_str(int sk)
+{
+	struct sockaddr_in6 myaddr6 = { 0, };
+	struct sockaddr_in myaddr4 = { 0, };
+	struct sockaddr *sockaddr_p;
+	char *net_addr;
+	socklen_t len;
+	int ret;
+
+	if (use_ipv6) {
+		len = sizeof(myaddr6);
+		sockaddr_p = (struct sockaddr * )&myaddr6;
+		net_addr = (char * )&myaddr6.sin6_addr;
+	} else {
+		len = sizeof(myaddr4);
+		sockaddr_p = (struct sockaddr * )&myaddr4;
+		net_addr = (char * )&myaddr4.sin_addr;
+	}
+
+	ret = getsockname(sk, sockaddr_p, &len);
+	if (ret) {
+		log_err("fio: getsockaddr: %s\n", strerror(errno));
+		return -1;
+	}
+
+	if (!inet_ntop(use_ipv6?AF_INET6:AF_INET, net_addr, client_sockaddr_str, INET6_ADDRSTRLEN - 1)) {
+		log_err("inet_ntop: failed to convert addr to string\n");
+		return -1;
+	}
+
+	dprint(FD_NET, "fio server bound to addr %s\n", client_sockaddr_str);
+	return 0;
+}
+
 static int accept_loop(int listen_sk)
 {
 	struct sockaddr_in addr;
 	struct sockaddr_in6 addr6;
 	socklen_t len = use_ipv6 ? sizeof(addr6) : sizeof(addr);
 	struct pollfd pfd;
-	int ret = 0, sk, flags, exitval = 0;
+	int ret = 0, sk, 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);
+	fio_set_fd_nonblocking(listen_sk, "server");
 
 	while (!exit_backend) {
+		struct sk_out *sk_out;
 		const char *from;
 		char buf[64];
 		pid_t pid;
@@ -902,6 +1367,13 @@ 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);
@@ -909,8 +1381,15 @@ static int accept_loop(int listen_sk)
 			continue;
 		}
 
-		/* exits */
-		handle_connection(sk);
+		/* 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);
 	}
 
 	return exitval;
@@ -918,12 +1397,13 @@ static int accept_loop(int listen_sk)
 
 int fio_server_text_output(int level, const char *buf, size_t len)
 {
+	struct sk_out *sk_out = pthread_getspecific(sk_out_key);
 	struct cmd_text_pdu *pdu;
 	unsigned int tlen;
 	struct timeval tv;
 
-	if (server_fd == -1)
-		return log_local_buf(buf, len);
+	if (!sk_out || sk_out->sk == -1)
+		return -1;
 
 	tlen = sizeof(*pdu) + len;
 	pdu = malloc(tlen);
@@ -937,7 +1417,7 @@ 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, NULL, NULL);
+	fio_net_queue_cmd(FIO_NET_CMD_TEXT, pdu, tlen, NULL, SK_F_COPY);
 	free(pdu);
 	return len;
 }
@@ -951,8 +1431,8 @@ static void convert_io_stat(struct io_stat *dst, struct io_stat *src)
 	/*
 	 * Encode to IEEE 754 for network transfer
 	 */
-	dst->mean.u.i	= __cpu_to_le64(fio_double_to_uint64(src->mean.u.f));
-	dst->S.u.i	= __cpu_to_le64(fio_double_to_uint64(src->S.u.f));
+	dst->mean.u.i	= cpu_to_le64(fio_double_to_uint64(src->mean.u.f));
+	dst->S.u.i	= cpu_to_le64(fio_double_to_uint64(src->S.u.f));
 }
 
 static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
@@ -964,7 +1444,7 @@ static void convert_gs(struct group_run_stats *dst, struct group_run_stats *src)
 		dst->min_run[i]		= cpu_to_le64(src->min_run[i]);
 		dst->max_bw[i]		= cpu_to_le64(src->max_bw[i]);
 		dst->min_bw[i]		= cpu_to_le64(src->min_bw[i]);
-		dst->io_kb[i]		= cpu_to_le64(src->io_kb[i]);
+		dst->iobytes[i]		= cpu_to_le64(src->iobytes[i]);
 		dst->agg[i]		= cpu_to_le64(src->agg[i]);
 	}
 
@@ -982,6 +1462,8 @@ 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");
 
@@ -1011,12 +1493,13 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
 	p.ts.minf		= cpu_to_le64(ts->minf);
 	p.ts.majf		= cpu_to_le64(ts->majf);
 	p.ts.clat_percentiles	= cpu_to_le64(ts->clat_percentiles);
+	p.ts.percentile_precision = cpu_to_le64(ts->percentile_precision);
 
 	for (i = 0; i < FIO_IO_U_LIST_MAX_LEN; i++) {
 		fio_fp64_t *src = &ts->percentile_list[i];
 		fio_fp64_t *dst = &p.ts.percentile_list[i];
 
-		dst->u.i = __cpu_to_le64(fio_double_to_uint64(src->u.f));
+		dst->u.i = cpu_to_le64(fio_double_to_uint64(src->u.f));
 	}
 
 	for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
@@ -1037,6 +1520,7 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
 	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]);
+		p.ts.drop_io_u[i]	= cpu_to_le64(ts->drop_io_u[i]);
 	}
 
 	p.ts.total_submit	= cpu_to_le64(ts->total_submit);
@@ -1057,11 +1541,43 @@ void fio_server_send_ts(struct thread_stat *ts, struct group_run_stats *rs)
 	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));
+	p.ts.latency_percentile.u.i = cpu_to_le64(fio_double_to_uint64(ts->latency_percentile.u.f));
+
+	p.ts.nr_block_infos	= cpu_to_le64(ts->nr_block_infos);
+	for (i = 0; i < p.ts.nr_block_infos; i++)
+		p.ts.block_infos[i] = cpu_to_le32(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_send_cmd(server_fd, FIO_NET_CMD_TS, &p, sizeof(p), NULL, NULL);
+	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)
@@ -1071,7 +1587,48 @@ 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), NULL, NULL);
+	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)
@@ -1079,16 +1636,16 @@ static void convert_agg(struct disk_util_agg *dst, struct disk_util_agg *src)
 	int i;
 
 	for (i = 0; i < 2; i++) {
-		dst->ios[i]	= cpu_to_le32(src->ios[i]);
-		dst->merges[i]	= cpu_to_le32(src->merges[i]);
+		dst->ios[i]	= cpu_to_le64(src->ios[i]);
+		dst->merges[i]	= cpu_to_le64(src->merges[i]);
 		dst->sectors[i]	= cpu_to_le64(src->sectors[i]);
-		dst->ticks[i]	= cpu_to_le32(src->ticks[i]);
+		dst->ticks[i]	= cpu_to_le64(src->ticks[i]);
 	}
 
-	dst->io_ticks		= cpu_to_le32(src->io_ticks);
-	dst->time_in_queue	= cpu_to_le32(src->time_in_queue);
+	dst->io_ticks		= cpu_to_le64(src->io_ticks);
+	dst->time_in_queue	= cpu_to_le64(src->time_in_queue);
 	dst->slavecount		= cpu_to_le32(src->slavecount);
-	dst->max_util.u.i	= __cpu_to_le64(fio_double_to_uint64(src->max_util.u.f));
+	dst->max_util.u.i	= cpu_to_le64(fio_double_to_uint64(src->max_util.u.f));
 }
 
 static void convert_dus(struct disk_util_stat *dst, struct disk_util_stat *src)
@@ -1099,14 +1656,14 @@ static void convert_dus(struct disk_util_stat *dst, struct disk_util_stat *src)
 	strncpy((char *) dst->name, (char *) src->name, FIO_DU_NAME_SZ - 1);
 
 	for (i = 0; i < 2; i++) {
-		dst->s.ios[i]		= cpu_to_le32(src->s.ios[i]);
-		dst->s.merges[i]	= cpu_to_le32(src->s.merges[i]);
+		dst->s.ios[i]		= cpu_to_le64(src->s.ios[i]);
+		dst->s.merges[i]	= cpu_to_le64(src->s.merges[i]);
 		dst->s.sectors[i]	= cpu_to_le64(src->s.sectors[i]);
-		dst->s.ticks[i]		= cpu_to_le32(src->s.ticks[i]);
+		dst->s.ticks[i]		= cpu_to_le64(src->s.ticks[i]);
 	}
 
-	dst->s.io_ticks		= cpu_to_le32(src->s.io_ticks);
-	dst->s.time_in_queue	= cpu_to_le32(src->s.time_in_queue);
+	dst->s.io_ticks		= cpu_to_le64(src->s.io_ticks);
+	dst->s.time_in_queue	= cpu_to_le64(src->s.time_in_queue);
 	dst->s.msec		= cpu_to_le64(src->s.msec);
 }
 
@@ -1126,123 +1683,319 @@ 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), NULL, NULL);
+		fio_net_queue_cmd(FIO_NET_CMD_DU, &pdu, sizeof(pdu), NULL, SK_F_COPY);
 	}
 }
 
+#ifdef CONFIG_ZLIB
+
+static inline void __fio_net_prep_tail(z_stream *stream, void *out_pdu,
+					struct sk_entry **last_entry,
+					struct sk_entry *first)
+{
+	unsigned int this_len = FIO_SERVER_MAX_FRAGMENT_PDU - stream->avail_out;
+
+	*last_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(&(*last_entry)->list, &first->next);
+
+}
+
 /*
- * Send a command with a separate PDU, not inlined in the command
+ * Deflates the next input given, creating as many new packets in the
+ * linked list as necessary.
  */
-static int fio_send_cmd_ext_pdu(int sk, uint16_t opcode, const void *buf,
-				off_t size, uint64_t tag, uint32_t flags)
+static int __deflate_pdu_buffer(void *next_in, unsigned int next_sz, void **out_pdu,
+				struct sk_entry **last_entry, z_stream *stream,
+				struct sk_entry *first)
 {
-	struct fio_net_cmd cmd;
-	struct iovec iov[2];
+	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;
+	stream->next_in = next_in;
+	stream->avail_in = next_sz;
+	do {
+		if (! stream->avail_out) {
+
+			__fio_net_prep_tail(stream, *out_pdu, last_entry, first);
+
+			*out_pdu = malloc(FIO_SERVER_MAX_FRAGMENT_PDU);
+
+			stream->avail_out = FIO_SERVER_MAX_FRAGMENT_PDU;
+			stream->next_out = *out_pdu;
+		}
 
-	__fio_init_net_cmd(&cmd, opcode, size, tag);
-	cmd.flags = __cpu_to_le32(flags);
-	fio_net_cmd_crc_pdu(&cmd, buf);
+		ret = deflate(stream, Z_BLOCK);
 
-	return fio_sendv_data(sk, iov, 2);
+		if (ret < 0) {
+			free(*out_pdu);
+			return 1;
+		}
+	} while (stream->avail_in);
+
+	return 0;
 }
 
-static int fio_send_iolog_gz(struct cmd_iolog_pdu *pdu, struct io_log *log)
+static int __fio_append_iolog_gz_hist(struct sk_entry *first, struct io_log *log,
+				      struct io_logs *cur_log, z_stream *stream)
 {
-	int ret = 0;
-#ifdef CONFIG_ZLIB
-	z_stream stream;
+	struct sk_entry *entry;
 	void *out_pdu;
+	int ret, i, j;
+	int sample_sz = log_entry_sz(log);
 
-	/*
-	 * 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;
+
+	for (i = 0; i < cur_log->nr_samples; i++) {
+		struct io_sample *s;
+		struct io_u_plat_entry *cur_plat_entry, *prev_plat_entry;
+		unsigned int *cur_plat, *prev_plat;
+
+		s = get_sample(log, cur_log, i);
+		ret = __deflate_pdu_buffer(s, sample_sz, &out_pdu, &entry, stream, first);
+		if (ret)
+			return ret;
+
+		/* Do the subtraction on server side so that client doesn't have to
+		 * reconstruct our linked list from packets.
+		 */
+		cur_plat_entry  = s->data.plat_entry;
+		prev_plat_entry = flist_first_entry(&cur_plat_entry->list, struct io_u_plat_entry, list);
+		cur_plat  = cur_plat_entry->io_u_plat;
+		prev_plat = prev_plat_entry->io_u_plat;
+
+		for (j = 0; j < FIO_IO_U_PLAT_NR; j++) {
+			cur_plat[j] -= prev_plat[j];
+		}
+
+		flist_del(&prev_plat_entry->list);
+		free(prev_plat_entry);
+
+		ret = __deflate_pdu_buffer(cur_plat_entry, sizeof(*cur_plat_entry),
+					   &out_pdu, &entry, stream, first);
+
+		if (ret)
+			return ret;
+	}
+
+	__fio_net_prep_tail(stream, out_pdu, &entry, first);
+
+	return 0;
+}
+
+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;
+
+	if (log->log_type == IO_LOG_TYPE_HIST)
+		return __fio_append_iolog_gz_hist(first, log, cur_log, stream);
+
+	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 * sizeof(struct io_sample);
+	ret = deflate(&stream, Z_FINISH);
 
-	do {
-		unsigned int this_len, flags = 0;
-		int ret;
+	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;
 
-		if (stream.avail_in)
-			flags = FIO_NET_CMD_F_MORE;
+		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 (ret != Z_STREAM_END);
 
-		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);
+	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;
-	int i, ret = 0;
+	struct sk_entry *first;
+	struct flist_head *entry;
+	int ret = 0;
 
+	pdu.nr_samples = cpu_to_le64(iolog_nr_samples(log));
 	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);
+	pdu.log_hist_coarseness = cpu_to_le32(log->hist_coarseness);
+
+	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;
 
-	for (i = 0; i < log->nr_samples; i++) {
-		struct io_sample *s = &log->log[i];
+	strncpy((char *) pdu.name, name, FIO_NET_NAME_MAX);
+	pdu.name[FIO_NET_NAME_MAX - 1] = '\0';
+
+	/*
+	 * 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;
 
-		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);
+		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->data.val	= cpu_to_le64(s->data.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;
+
+				so->offset = cpu_to_le64(so->offset);
+			}
+		}
 	}
 
 	/*
-	 * Send header first, it's not compressed.
+	 * Assemble header entry first
 	 */
-	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;
+	first = fio_net_prep_cmd(FIO_NET_CMD_IOLOG, &pdu, sizeof(pdu), NULL, SK_F_VEC | SK_F_INLINE | SK_F_COPY);
 
 	/*
-	 * Now send actual log, compress if we can, otherwise just plain
+	 * 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)
-		return fio_send_iolog_gz(&pdu, log);
+	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);
 
-	return fio_send_cmd_ext_pdu(server_fd, FIO_NET_CMD_IOLOG, log->log,
-			log->nr_samples * sizeof(struct io_sample), 0, 0);
+	fio_net_queue_entry(first);
+	return ret;
 }
 
 void fio_server_send_add_job(struct thread_data *td)
@@ -1254,14 +2007,88 @@ 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_send_cmd(server_fd, FIO_NET_CMD_ADD_JOB, &pdu, sizeof(pdu), NULL, NULL);
+	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)
 {
-	assert(server_fd != -1);
+	struct sk_out *sk_out = pthread_getspecific(sk_out_key);
+
+	assert(sk_out->sk != -1);
 
-	fio_net_send_simple_cmd(server_fd, FIO_NET_CMD_SERVER_START, 0, NULL);
+	fio_net_queue_cmd(FIO_NET_CMD_SERVER_START, NULL, 0, NULL, SK_F_SIMPLE);
+}
+
+int fio_server_get_verify_state(const char *name, int threadnumber,
+				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)
+		return ENOMEM;
+
+	__fio_mutex_init(&rep->lock, FIO_MUTEX_LOCKED);
+	rep->data = NULL;
+	rep->error = 0;
+
+	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);
+
+	/*
+	 * 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: %s\n",
+				out.path, strerror(rep->error));
+		ret = rep->error;
+fail:
+		*datap = NULL;
+		sfree(rep);
+		fio_net_queue_quit();
+		return ret;
+	}
+
+	/*
+	 * The format is verify_state_hdr, then thread_io_list. Verify
+	 * the header, and the thread_io_list checksum
+	 */
+	s = rep->data + sizeof(struct verify_state_hdr);
+	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);
+	*datap = data;
+
+	sfree(rep->data);
+	__fio_mutex_remove(&rep->lock);
+	sfree(rep);
+	return ret;
 }
 
 static int fio_init_server_ip(void)
@@ -1284,16 +2111,15 @@ static int fio_init_server_ip(void)
 
 	opt = 1;
 	if (setsockopt(sk, SOL_SOCKET, SO_REUSEADDR, (void *)&opt, sizeof(opt)) < 0) {
-		log_err("fio: setsockopt: %s\n", strerror(errno));
+		log_err("fio: setsockopt(REUSEADDR): %s\n", strerror(errno));
 		close(sk);
 		return -1;
 	}
 #ifdef SO_REUSEPORT
-	if (setsockopt(sk, SOL_SOCKET, SO_REUSEPORT, &opt, sizeof(opt)) < 0) {
-		log_err("fio: setsockopt: %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) {
@@ -1395,7 +2221,7 @@ 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;
@@ -1586,6 +2412,22 @@ 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;
@@ -1617,8 +2459,12 @@ static int fio_server(void)
 
 void fio_server_got_signal(int signal)
 {
+	struct sk_out *sk_out = pthread_getspecific(sk_out_key);
+
+	assert(sk_out);
+
 	if (signal == SIGPIPE)
-		server_fd = -1;
+		sk_out->sk = -1;
 	else {
 		log_info("\nfio: terminating on signal %d\n", signal);
 		exit_backend = 1;
@@ -1692,14 +2538,13 @@ int fio_start_server(char *pidfile)
 
 	pid = fork();
 	if (pid < 0) {
-		log_err("fio: failed server fork: %s", strerror(errno));
+		log_err("fio: failed server fork: %s\n", strerror(errno));
 		free(pidfile);
 		return -1;
 	} else if (pid) {
-		int ret = write_pid(pid, pidfile);
-
+		ret = write_pid(pid, pidfile);
 		free(pidfile);
-		exit(ret);
+		_exit(ret);
 	}
 
 	setsid();