X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=blktrace.c;h=d0d271f9add5106b4481321058787f1ba1aa82a4;hb=d025d6c67760a52e34e6c7352f33ad5dbbb4f6f4;hp=24b48d5b2a214ecc9d86222a6263366afbfbe2b7;hpb=e076d33bfa9a5ef334e8db6e143c624d15a7c30f;p=blktrace.git

diff --git a/blktrace.c b/blktrace.c
index 24b48d5..d0d271f 100644
--- a/blktrace.c
+++ b/blktrace.c
@@ -2,6 +2,10 @@
  * block queue tracing application
  *
  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
+ * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
+ *
+ * Rewrite to have a single thread per CPU (managing all devices on that CPU)
+ *	Alan D. Brunelle <alan.brunelle@hp.com> - January 2009
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -18,48 +22,314 @@
  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  *
  */
-#include <pthread.h>
-#include <sys/types.h>
-#include <sys/stat.h>
+
+#include <errno.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <fcntl.h>
+#include <getopt.h>
+#include <sched.h>
 #include <unistd.h>
-#include <locale.h>
+#include <poll.h>
 #include <signal.h>
-#include <fcntl.h>
-#include <string.h>
+#include <pthread.h>
+#include <locale.h>
 #include <sys/ioctl.h>
-#include <sys/param.h>
-#include <sys/statfs.h>
-#include <sys/poll.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <sys/vfs.h>
 #include <sys/mman.h>
+#include <sys/param.h>
+#include <sys/time.h>
+#include <sys/resource.h>
 #include <sys/socket.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <sched.h>
-#include <ctype.h>
-#include <getopt.h>
-#include <errno.h>
 #include <netinet/in.h>
 #include <arpa/inet.h>
 #include <netdb.h>
 #include <sys/sendfile.h>
 
+#include "btt/list.h"
 #include "blktrace.h"
-#include "barrier.h"
-
-static char blktrace_version[] = "0.99";
 
 /*
  * You may want to increase this even more, if you are logging at a high
  * rate and see skipped/missed events
  */
-#define BUF_SIZE	(512 * 1024)
-#define BUF_NR		(4)
+#define BUF_SIZE		(512 * 1024)
+#define BUF_NR			(4)
+
+#define FILE_VBUF_SIZE		(128 * 1024)
+
+#define DEBUGFS_TYPE		(0x64626720)
+#define TRACE_NET_PORT		(8462)
+
+enum {
+	Net_none = 0,
+	Net_server,
+	Net_client,
+};
+
+enum thread_status {
+	Th_running,
+	Th_leaving,
+	Th_error
+};
+
+/*
+ * Generic stats collected: nevents can be _roughly_ estimated by data_read
+ * (discounting pdu...)
+ *
+ * These fields are updated w/ pdc_dr_update & pdc_nev_update below.
+ */
+struct pdc_stats {
+	unsigned long long data_read;
+	unsigned long long nevents;
+};
+
+struct devpath {
+	struct list_head head;
+	char *path;			/* path to device special file */
+	char *buts_name;		/* name returned from bt kernel code */
+	struct pdc_stats *stats;
+	int fd, ncpus;
+	unsigned long long drops;
+
+	/*
+	 * For piped output only:
+	 *
+	 * Each tracer will have a tracer_devpath_head that it will add new
+	 * data onto. It's list is protected above (tracer_devpath_head.mutex)
+	 * and it will signal the processing thread using the dp_cond,
+	 * dp_mutex & dp_entries variables above.
+	 */
+	struct tracer_devpath_head *heads;
+
+	/*
+	 * For network server mode only:
+	 */
+	struct cl_host *ch;
+	u32 cl_id;
+	time_t cl_connect_time;
+	int setup_done;	/* ioctl BLKTRACESETUP done */
+	struct io_info *ios;
+};
+
+/*
+ * For piped output to stdout we will have each tracer thread (one per dev)
+ * tack buffers read from the relay queues on a per-device list.
+ *
+ * The main thread will then collect trace buffers from each of lists in turn.
+ *
+ * We will use a mutex to guard each of the trace_buf list. The tracers
+ * can then signal the main thread using <dp_cond,dp_mutex> and
+ * dp_entries. (When dp_entries is 0, and a tracer adds an entry it will
+ * signal. When dp_entries is 0, the main thread will wait for that condition
+ * to be signalled.)
+ *
+ * adb: It may be better just to have a large buffer per tracer per dev,
+ * and then use it as a ring-buffer. This would certainly cut down a lot
+ * of malloc/free thrashing, at the cost of more memory movements (potentially).
+ */
+struct trace_buf {
+	struct list_head head;
+	struct devpath *dpp;
+	void *buf;
+	int cpu, len;
+};
+
+struct tracer_devpath_head {
+	pthread_mutex_t mutex;
+	struct list_head head;
+	struct trace_buf *prev;
+};
+
+/*
+ * Used to handle the mmap() interfaces for output file (containing traces)
+ */
+struct mmap_info {
+	void *fs_buf;
+	unsigned long long fs_size, fs_max_size, fs_off, fs_buf_len;
+	unsigned long buf_size, buf_nr;
+	int pagesize;
+};
+
+/*
+ * Each thread doing work on a (client) side of blktrace will have one
+ * of these. The ios array contains input/output information, pfds holds
+ * poll() data. The volatile's provide flags to/from the main executing
+ * thread.
+ */
+struct tracer {
+	struct list_head head;
+	struct io_info *ios;
+	struct pollfd *pfds;
+	pthread_t thread;
+	int cpu, nios;
+	volatile int status, is_done;
+};
+
+/*
+ * networking stuff follows. we include a magic number so we know whether
+ * to endianness convert or not.
+ *
+ * The len field is overloaded:
+ *	0 - Indicates an "open" - allowing the server to set up for a dev/cpu
+ *	1 - Indicates a "close" - Shut down connection orderly
+ *
+ * The cpu field is overloaded on close: it will contain the number of drops.
+ */
+struct blktrace_net_hdr {
+	u32 magic;		/* same as trace magic */
+	char buts_name[32];	/* trace name */
+	u32 cpu;		/* for which cpu */
+	u32 max_cpus;
+	u32 len;		/* length of following trace data */
+	u32 cl_id;		/* id for set of client per-cpu connections */
+	u32 buf_size;		/* client buf_size for this trace  */
+	u32 buf_nr;		/* client buf_nr for this trace  */
+	u32 page_size;		/* client page_size for this trace  */
+};
+
+/*
+ * Each host encountered has one of these. The head is used to link this
+ * on to the network server's ch_list. Connections associated with this
+ * host are linked on conn_list, and any devices traced on that host
+ * are connected on the devpaths list.
+ */
+struct cl_host {
+	struct list_head head;
+	struct list_head conn_list;
+	struct list_head devpaths;
+	struct net_server_s *ns;
+	char *hostname;
+	struct in_addr cl_in_addr;
+	int connects, ndevs, cl_opens;
+};
+
+/*
+ * Each connection (client to server socket ('fd')) has one of these. A
+ * back reference to the host ('ch'), and lists headers (for the host
+ * list, and the network server conn_list) are also included.
+ */
+struct cl_conn {
+	struct list_head ch_head, ns_head;
+	struct cl_host *ch;
+	int fd, ncpus;
+	time_t connect_time;
+};
+
+/*
+ * The network server requires some poll structures to be maintained -
+ * one per conection currently on conn_list. The nchs/ch_list values
+ * are for each host connected to this server. The addr field is used
+ * for scratch as new connections are established.
+ */
+struct net_server_s {
+	struct list_head conn_list;
+	struct list_head ch_list;
+	struct pollfd *pfds;
+	int listen_fd, connects, nchs;
+	struct sockaddr_in addr;
+};
+
+/*
+ * This structure is (generically) used to providide information
+ * for a read-to-write set of values.
+ *
+ * ifn & ifd represent input information
+ *
+ * ofn, ofd, ofp, obuf & mmap_info are used for output file (optionally).
+ */
+struct io_info {
+	struct devpath *dpp;
+	FILE *ofp;
+	char *obuf;
+	struct cl_conn *nc;	/* Server network connection */
+
+	/*
+	 * mmap controlled output files
+	 */
+	struct mmap_info mmap_info;
+
+	/*
+	 * Client network fields
+	 */
+	unsigned int ready;
+	unsigned long long data_queued;
+
+	/*
+	 * Input/output file descriptors & names
+	 */
+	int ifd, ofd;
+	char ifn[MAXPATHLEN + 64];
+	char ofn[MAXPATHLEN + 64];
+};
+
+static char blktrace_version[] = "2.0.0";
+
+/*
+ * Linkage to blktrace helper routines (trace conversions)
+ */
+int data_is_native = -1;
+
+static int ndevs;
+static int max_cpus;
+static int ncpus;
+static cpu_set_t *online_cpus;
+static int pagesize;
+static int act_mask = ~0U;
+static int kill_running_trace;
+static int stop_watch;
+static int piped_output;
+
+static char *debugfs_path = "/sys/kernel/debug";
+static char *output_name;
+static char *output_dir;
+
+static unsigned long buf_size = BUF_SIZE;
+static unsigned long buf_nr = BUF_NR;
+
+static FILE *pfp;
+
+static LIST_HEAD(devpaths);
+static LIST_HEAD(tracers);
+
+static volatile int done;
+
+/*
+ * tracer threads add entries, the main thread takes them off and processes
+ * them. These protect the dp_entries variable.
+ */
+static pthread_cond_t dp_cond = PTHREAD_COND_INITIALIZER;
+static pthread_mutex_t dp_mutex = PTHREAD_MUTEX_INITIALIZER;
+static volatile int dp_entries;
+
+/*
+ * These synchronize master / thread interactions.
+ */
+static pthread_cond_t mt_cond = PTHREAD_COND_INITIALIZER;
+static pthread_mutex_t mt_mutex = PTHREAD_MUTEX_INITIALIZER;
+static volatile int nthreads_running;
+static volatile int nthreads_leaving;
+static volatile int nthreads_error;
+static volatile int tracers_run;
 
-#define OFILE_BUF	(128 * 1024)
+/*
+ * network cmd line params
+ */
+static struct sockaddr_in hostname_addr;
+static char hostname[MAXHOSTNAMELEN];
+static int net_port = TRACE_NET_PORT;
+static int net_use_sendfile = 1;
+static int net_mode;
+static int *cl_fds;
 
-#define RELAYFS_TYPE	0xF0B4A981
+static int (*handle_pfds)(struct tracer *, int, int);
+static int (*handle_list)(struct tracer_devpath_head *, struct list_head *);
 
-#define S_OPTS	"d:a:A:r:o:kw:Vb:n:D:lh:p:s"
+#define S_OPTS	"d:a:A:r:o:kw:vVb:n:D:lh:p:sI:"
 static struct option l_opts[] = {
 	{
 		.name = "dev",
@@ -67,6 +337,12 @@ static struct option l_opts[] = {
 		.flag = NULL,
 		.val = 'd'
 	},
+	{
+		.name = "input-devs",
+		.has_arg = required_argument,
+		.flag = NULL,
+		.val = 'I'
+	},
 	{
 		.name = "act-mask",
 		.has_arg = required_argument,
@@ -103,6 +379,12 @@ static struct option l_opts[] = {
 		.flag = NULL,
 		.val = 'w'
 	},
+	{
+		.name = "version",
+		.has_arg = no_argument,
+		.flag = NULL,
+		.val = 'v'
+	},
 	{
 		.name = "version",
 		.has_arg = no_argument,
@@ -146,7 +428,7 @@ static struct option l_opts[] = {
 		.val = 'p'
 	},
 	{
-		.name = "sendfile",
+		.name = "no-sendfile",
 		.has_arg = no_argument,
 		.flag = NULL,
 		.val = 's'
@@ -156,1429 +438,1672 @@ static struct option l_opts[] = {
 	}
 };
 
-struct tip_subbuf {
-	void *buf;
-	unsigned int len;
-	unsigned int max_len;
-};
+static char usage_str[] = "\n\n" \
+	"-d <dev>             | --dev=<dev>\n" \
+        "[ -r <debugfs path>  | --relay=<debugfs path> ]\n" \
+        "[ -o <file>          | --output=<file>]\n" \
+        "[ -D <dir>           | --output-dir=<dir>\n" \
+        "[ -w <time>          | --stopwatch=<time>]\n" \
+        "[ -a <action field>  | --act-mask=<action field>]\n" \
+        "[ -A <action mask>   | --set-mask=<action mask>]\n" \
+        "[ -b <size>          | --buffer-size]\n" \
+        "[ -n <number>        | --num-sub-buffers=<number>]\n" \
+        "[ -l                 | --listen]\n" \
+        "[ -h <hostname>      | --host=<hostname>]\n" \
+        "[ -p <port number>   | --port=<port number>]\n" \
+        "[ -s                 | --no-sendfile]\n" \
+        "[ -I <devs file>     | --input-devs=<devs file>]\n" \
+        "[ -v <version>       | --version]\n" \
+        "[ -V <version>       | --version]\n" \
 
-#define FIFO_SIZE	(1024)	/* should be plenty big! */
-#define CL_SIZE		(128)	/* cache line, any bigger? */
+	"\t-d Use specified device. May also be given last after options\n" \
+	"\t-r Path to mounted debugfs, defaults to /sys/kernel/debug\n" \
+	"\t-o File(s) to send output to\n" \
+	"\t-D Directory to prepend to output file names\n" \
+	"\t-w Stop after defined time, in seconds\n" \
+	"\t-a Only trace specified actions. See documentation\n" \
+	"\t-A Give trace mask as a single value. See documentation\n" \
+	"\t-b Sub buffer size in KiB (default 512)\n" \
+	"\t-n Number of sub buffers (default 4)\n" \
+	"\t-l Run in network listen mode (blktrace server)\n" \
+	"\t-h Run in network client mode, connecting to the given host\n" \
+	"\t-p Network port to use (default 8462)\n" \
+	"\t-s Make the network client NOT use sendfile() to transfer data\n" \
+	"\t-I Add devices found in <devs file>\n" \
+	"\t-v Print program version info\n" \
+	"\t-V Print program version info\n\n";
 
-struct tip_subbuf_fifo {
-	int tail __attribute__((aligned(CL_SIZE)));
-	int head __attribute__((aligned(CL_SIZE)));
-	struct tip_subbuf *q[FIFO_SIZE];
-};
+static void clear_events(struct pollfd *pfd)
+{
+	pfd->events = 0;
+	pfd->revents = 0;
+}
 
-struct thread_information {
-	int cpu;
-	pthread_t thread;
+static inline int net_client_use_sendfile(void)
+{
+	return net_mode == Net_client && net_use_sendfile;
+}
 
-	int fd;
-	void *fd_buf;
-	char fn[MAXPATHLEN + 64];
+static inline int net_client_use_send(void)
+{
+	return net_mode == Net_client && !net_use_sendfile;
+}
 
-	int pfd;
-	size_t *pfd_buf;
+static inline int use_tracer_devpaths(void)
+{
+	return piped_output || net_client_use_send();
+}
 
-	struct in_addr cl_in_addr;
+static inline int in_addr_eq(struct in_addr a, struct in_addr b)
+{
+	return a.s_addr == b.s_addr;
+}
 
-	FILE *ofile;
-	char *ofile_buffer;
-	off_t ofile_offset;
-	int ofile_stdout;
-	int ofile_mmap;
-	volatile int sendfile_pending;
+static inline void pdc_dr_update(struct devpath *dpp, int cpu, int data_read)
+{
+	dpp->stats[cpu].data_read += data_read;
+}
 
-	int (*get_subbuf)(struct thread_information *, unsigned int);
-	int (*flush_subbuf)(struct thread_information *, struct tip_subbuf *);
-	int (*read_data)(struct thread_information *, void *, unsigned int);
+static inline void pdc_nev_update(struct devpath *dpp, int cpu, int nevents)
+{
+	dpp->stats[cpu].nevents += nevents;
+}
 
-	unsigned long events_processed;
-	unsigned long long data_read;
-	struct device_information *device;
+static void show_usage(char *prog)
+{
+	fprintf(stderr, "Usage: %s %s", prog, usage_str);
+}
 
-	int exited;
+/*
+ * Create a timespec 'msec' milliseconds into the future
+ */
+static inline void make_timespec(struct timespec *tsp, long delta_msec)
+{
+	struct timeval now;
 
-	/*
-	 * piped fifo buffers
-	 */
-	struct tip_subbuf_fifo fifo;
-	struct tip_subbuf *leftover_ts;
+	gettimeofday(&now, NULL);
+	tsp->tv_sec = now.tv_sec;
+	tsp->tv_nsec = 1000L * now.tv_usec;
 
-	/*
-	 * mmap controlled output files
-	 */
-	unsigned long long fs_size;
-	unsigned long long fs_max_size;
-	unsigned long fs_off;
-	void *fs_buf;
-	unsigned long fs_buf_len;
-};
+	tsp->tv_nsec += (delta_msec * 1000000L);
+	if (tsp->tv_nsec > 1000000000L) {
+		long secs = tsp->tv_nsec / 1000000000L;
 
-struct device_information {
-	int fd;
-	char *path;
-	char buts_name[32];
-	volatile int trace_started;
-	unsigned long drop_count;
-	struct thread_information *threads;
-};
+		tsp->tv_sec += secs;
+		tsp->tv_nsec -= (secs * 1000000000L);
+	}
+}
 
-static int ncpus;
-static struct thread_information *thread_information;
-static int ndevs;
-static struct device_information *device_information;
+/*
+ * Add a timer to ensure wait ends
+ */
+static void t_pthread_cond_wait(pthread_cond_t *cond, pthread_mutex_t *mutex)
+{
+	struct timespec ts;
 
-/* command line option globals */
-static char *relay_path;
-static char *output_name;
-static char *output_dir;
-static int act_mask = ~0U;
-static int kill_running_trace;
-static unsigned long buf_size = BUF_SIZE;
-static unsigned long buf_nr = BUF_NR;
-static unsigned int page_size;
+	make_timespec(&ts, 50);
+	pthread_cond_timedwait(cond, mutex, &ts);
+}
 
-#define is_done()	(*(volatile int *)(&done))
-static volatile int done;
+static void unblock_tracers(void)
+{
+	pthread_mutex_lock(&mt_mutex);
+	tracers_run = 1;
+	pthread_cond_broadcast(&mt_cond);
+	pthread_mutex_unlock(&mt_mutex);
+}
 
-#define is_trace_stopped()	(*(volatile int *)(&trace_stopped))
-static volatile int trace_stopped;
+static void tracer_wait_unblock(struct tracer *tp)
+{
+	pthread_mutex_lock(&mt_mutex);
+	while (!tp->is_done && !tracers_run)
+		pthread_cond_wait(&mt_cond, &mt_mutex);
+	pthread_mutex_unlock(&mt_mutex);
+}
 
-#define is_stat_shown()	(*(volatile int *)(&stat_shown))
-static volatile int stat_shown;
+static void tracer_signal_ready(struct tracer *tp,
+				enum thread_status th_status,
+				int status)
+{
+	pthread_mutex_lock(&mt_mutex);
+	tp->status = status;
 
-int data_is_native = -1;
+	if (th_status == Th_running)
+		nthreads_running++;
+	else if (th_status == Th_error)
+		nthreads_error++;
+	else
+		nthreads_leaving++;
 
-static void exit_trace(int status);
+	pthread_cond_signal(&mt_cond);
+	pthread_mutex_unlock(&mt_mutex);
+}
 
-#define dip_tracing(dip)	(*(volatile int *)(&(dip)->trace_started))
-#define dip_set_tracing(dip, v)	((dip)->trace_started = (v))
+static void wait_tracers_ready(int ncpus_started)
+{
+	pthread_mutex_lock(&mt_mutex);
+	while ((nthreads_running + nthreads_error) < ncpus_started)
+		t_pthread_cond_wait(&mt_cond, &mt_mutex);
+	pthread_mutex_unlock(&mt_mutex);
+}
 
-#define __for_each_dip(__d, __i, __e)	\
-	for (__i = 0, __d = device_information; __i < __e; __i++, __d++)
+static void wait_tracers_leaving(void)
+{
+	pthread_mutex_lock(&mt_mutex);
+	while (nthreads_leaving < nthreads_running)
+		t_pthread_cond_wait(&mt_cond, &mt_mutex);
+	pthread_mutex_unlock(&mt_mutex);
+}
 
-#define for_each_dip(__d, __i)	__for_each_dip(__d, __i, ndevs)
-#define for_each_tip(__d, __t, __j)	\
-	for (__j = 0, __t = (__d)->threads; __j < ncpus; __j++, __t++)
+static void init_mmap_info(struct mmap_info *mip)
+{
+	mip->buf_size = buf_size;
+	mip->buf_nr = buf_nr;
+	mip->pagesize = pagesize;
+}
 
-/*
- * networking stuff follows. we include a magic number so we know whether
- * to endianness convert or not
- */
-struct blktrace_net_hdr {
-	u32 magic;		/* same as trace magic */
-	char buts_name[32];	/* trace name */
-	u32 cpu;		/* for which cpu */
-	u32 max_cpus;
-	u32 len;		/* length of following trace data */
-};
+static void net_close_connection(int *fd)
+{
+	shutdown(*fd, SHUT_RDWR);
+	close(*fd);
+	*fd = -1;
+}
 
-#define TRACE_NET_PORT		(8462)
+static void dpp_free(struct devpath *dpp)
+{
+	if (dpp->stats)
+		free(dpp->stats);
+	if (dpp->ios)
+		free(dpp->ios);
+	if (dpp->path)
+		free(dpp->path);
+	if (dpp->buts_name)
+		free(dpp->buts_name);
+	free(dpp);
+}
 
-enum {
-	Net_none = 0,
-	Net_server,
-	Net_client,
-};
+static int lock_on_cpu(int cpu)
+{
+	cpu_set_t * cpu_mask;
+	size_t size;
 
-/*
- * network cmd line params
- */
-static char hostname[MAXHOSTNAMELEN];
-static int net_port = TRACE_NET_PORT;
-static int net_mode = 0;
-static int net_use_sendfile;
+	cpu_mask = CPU_ALLOC(max_cpus);
+	size = CPU_ALLOC_SIZE(max_cpus);
 
-static int net_in_fd = -1;
-static int net_out_fd = -1;
+	CPU_ZERO_S(size, cpu_mask);
+	CPU_SET_S(cpu, size, cpu_mask);
+	if (sched_setaffinity(0, size, cpu_mask) < 0) {
+		CPU_FREE(cpu_mask);		
+		return errno;
+	}
 
-static void handle_sigint(__attribute__((__unused__)) int sig)
+	CPU_FREE(cpu_mask);		
+	return 0;
+}
+
+static int increase_limit(int resource, rlim_t increase)
 {
-	struct device_information *dip;
-	int i;
+	struct rlimit rlim;
+	int save_errno = errno;
 
-	/*
-	 * stop trace so we can reap currently produced data
-	 */
-	for_each_dip(dip, i) {
-		if (dip->fd == -1)
-			continue;
-		if (ioctl(dip->fd, BLKTRACESTOP) < 0)
-			perror("BLKTRACESTOP");
+	if (!getrlimit(resource, &rlim)) {
+		rlim.rlim_cur += increase;
+		if (rlim.rlim_cur >= rlim.rlim_max)
+			rlim.rlim_max = rlim.rlim_cur + increase;
+
+		if (!setrlimit(resource, &rlim))
+			return 1;
 	}
 
-	done = 1;
+	errno = save_errno;
+	return 0;
 }
 
-static int get_dropped_count(const char *buts_name)
+static int handle_open_failure(void)
 {
-	int fd;
-	char tmp[MAXPATHLEN + 64];
+	if (errno == ENFILE || errno == EMFILE)
+		return increase_limit(RLIMIT_NOFILE, 16);
+	return 0;
+}
 
-	snprintf(tmp, sizeof(tmp), "%s/block/%s/dropped",
-		 relay_path, buts_name);
+static int handle_mem_failure(size_t length)
+{
+	if (errno == ENFILE)
+		return handle_open_failure();
+	else if (errno == ENOMEM)
+		return increase_limit(RLIMIT_MEMLOCK, 2 * length);
+	return 0;
+}
 
-	fd = open(tmp, O_RDONLY);
-	if (fd < 0) {
-		/*
-		 * this may be ok, if the kernel doesn't support dropped counts
-		 */
-		if (errno == ENOENT)
-			return 0;
+static FILE *my_fopen(const char *path, const char *mode)
+{
+	FILE *fp;
 
-		fprintf(stderr, "Couldn't open dropped file %s\n", tmp);
-		return -1;
-	}
+	do {
+		fp = fopen(path, mode);
+	} while (fp == NULL && handle_open_failure());
 
-	if (read(fd, tmp, sizeof(tmp)) < 0) {
-		perror(tmp);
-		close(fd);
-		return -1;
-	}
+	return fp;
+}
 
-	close(fd);
+static int my_open(const char *path, int flags)
+{
+	int fd;
 
-	return atoi(tmp);
+	do {
+		fd = open(path, flags);
+	} while (fd < 0 && handle_open_failure());
+
+	return fd;
 }
 
-static int start_trace(struct device_information *dip)
+static int my_socket(int domain, int type, int protocol)
 {
-	struct blk_user_trace_setup buts;
+	int fd;
 
-	memset(&buts, 0, sizeof(buts));
-	buts.buf_size = buf_size;
-	buts.buf_nr = buf_nr;
-	buts.act_mask = act_mask;
+	do {
+		fd = socket(domain, type, protocol);
+	} while (fd < 0 && handle_open_failure());
 
-	if (ioctl(dip->fd, BLKTRACESETUP, &buts) < 0) {
-		perror("BLKTRACESETUP");
-		return 1;
-	}
+	return fd;
+}
 
-	if (ioctl(dip->fd, BLKTRACESTART) < 0) {
-		perror("BLKTRACESTART");
-		return 1;
-	}
+static int my_accept(int sockfd, struct sockaddr *addr, socklen_t *addrlen)
+{
+	int fd;
 
-	memcpy(dip->buts_name, buts.name, sizeof(dip->buts_name));
-	dip_set_tracing(dip, 1);
-	return 0;
+	do {
+		fd = accept(sockfd, addr, addrlen);
+	} while (fd < 0 && handle_open_failure());
+
+	return fd;
 }
 
-static void stop_trace(struct device_information *dip)
+static void *my_mmap(void *addr, size_t length, int prot, int flags, int fd,
+		     off_t offset)
 {
-	if (dip_tracing(dip) || kill_running_trace) {
-		dip_set_tracing(dip, 0);
+	void *new;
 
-		/*
-		 * should be stopped, just don't complain if it isn't
-		 */
-		ioctl(dip->fd, BLKTRACESTOP);
+	do {
+		new = mmap(addr, length, prot, flags, fd, offset);
+	} while (new == MAP_FAILED && handle_mem_failure(length));
 
-		if (ioctl(dip->fd, BLKTRACETEARDOWN) < 0)
-			perror("BLKTRACETEARDOWN");
+	return new;
+}
 
-		close(dip->fd);
-		dip->fd = -1;
-	}
+static int my_mlock(struct tracer *tp,
+		    const void *addr, size_t len)
+{
+	int ret, retry = 0;
+
+	do {
+		ret = mlock(addr, len);
+		if ((retry >= 10) && tp && tp->is_done)
+			break;
+		retry++;
+	} while (ret < 0 && handle_mem_failure(len));
+
+	return ret;
 }
 
-static void stop_all_traces(void)
+static int setup_mmap(int fd, unsigned int maxlen,
+		      struct mmap_info *mip,
+		      struct tracer *tp)
 {
-	struct device_information *dip;
-	int i;
+	if (mip->fs_off + maxlen > mip->fs_buf_len) {
+		unsigned long nr = max(16, mip->buf_nr);
+
+		if (mip->fs_buf) {
+			munlock(mip->fs_buf, mip->fs_buf_len);
+			munmap(mip->fs_buf, mip->fs_buf_len);
+			mip->fs_buf = NULL;
+		}
+
+		mip->fs_off = mip->fs_size & (mip->pagesize - 1);
+		mip->fs_buf_len = (nr * mip->buf_size) - mip->fs_off;
+		mip->fs_max_size += mip->fs_buf_len;
 
-	for_each_dip(dip, i) {
-		dip->drop_count = get_dropped_count(dip->buts_name);
-		stop_trace(dip);
+		if (ftruncate(fd, mip->fs_max_size) < 0) {
+			perror("setup_mmap: ftruncate");
+			return 1;
+		}
+
+		mip->fs_buf = my_mmap(NULL, mip->fs_buf_len, PROT_WRITE,
+				      MAP_SHARED, fd,
+				      mip->fs_size - mip->fs_off);
+		if (mip->fs_buf == MAP_FAILED) {
+			perror("setup_mmap: mmap");
+			return 1;
+		}
+		if (my_mlock(tp, mip->fs_buf, mip->fs_buf_len) < 0) {
+			perror("setup_mlock: mlock");
+			return 1;
+		}
 	}
+
+	return 0;
 }
 
-static void wait_for_data(struct thread_information *tip)
+static int __stop_trace(int fd)
 {
-	struct pollfd pfd = { .fd = tip->fd, .events = POLLIN };
-
-	do {
-		poll(&pfd, 1, 100);
-		if (pfd.revents & POLLIN)
-			break;
-		if (tip->ofile_stdout)
-			break;
-	} while (!is_done());
+	/*
+	 * Should be stopped, don't complain if it isn't
+	 */
+	ioctl(fd, BLKTRACESTOP);
+	return ioctl(fd, BLKTRACETEARDOWN);
 }
 
-static int read_data_file(struct thread_information *tip, void *buf,
-			  unsigned int len)
+static int write_data(char *buf, int len)
 {
-	int ret = 0;
+	int ret;
 
-	do {
-		wait_for_data(tip);
+rewrite:
+	ret = fwrite(buf, len, 1, pfp);
+	if (ferror(pfp) || ret != 1) {
+		if (errno == EINTR) {
+			clearerr(pfp);
+			goto rewrite;
+		}
 
-		ret = read(tip->fd, buf, len);
-		if (!ret)
-			continue;
-		else if (ret > 0)
-			return ret;
-		else {
-			if (errno != EAGAIN) {
-				perror(tip->fn);
-				fprintf(stderr,"Thread %d failed read of %s\n",
-					tip->cpu, tip->fn);
-				break;
-			}
-			continue;
+		if (!piped_output || (errno != EPIPE && errno != EBADF)) {
+			fprintf(stderr, "write(%d) failed: %d/%s\n",
+				len, errno, strerror(errno));
 		}
-	} while (!is_done());
+		goto err;
+	}
 
-	return ret;
+	fflush(pfp);
+	return 0;
 
+err:
+	clearerr(pfp);
+	return 1;
 }
 
-static int read_data_net(struct thread_information *tip, void *buf,
-			 unsigned int len)
+/*
+ * Returns the number of bytes read (successfully)
+ */
+static int __net_recv_data(int fd, void *buf, unsigned int len)
 {
 	unsigned int bytes_left = len;
-	int ret = 0;
 
-	do {
-		ret = recv(net_in_fd, buf, bytes_left, MSG_WAITALL);
+	while (bytes_left && !done) {
+		int ret = recv(fd, buf, bytes_left, MSG_WAITALL);
 
-		if (!ret)
-			continue;
+		if (ret == 0)
+			break;
 		else if (ret < 0) {
-			if (errno != EAGAIN) {
-				perror(tip->fn);
-				fprintf(stderr, "server: failed read\n");
-				return 0;
+			if (errno == EAGAIN) {
+				usleep(50);
+				continue;
 			}
-			continue;
+			perror("server: net_recv_data: recv failed");
+			break;
 		} else {
 			buf += ret;
 			bytes_left -= ret;
 		}
-	} while (!is_done() && bytes_left);
+	}
 
 	return len - bytes_left;
 }
 
-static int read_data(struct thread_information *tip, void *buf,
-		     unsigned int len)
+static int net_recv_data(int fd, void *buf, unsigned int len)
 {
-	return tip->read_data(tip, buf, len);
+	return __net_recv_data(fd, buf, len);
 }
 
-static inline struct tip_subbuf *
-subbuf_fifo_dequeue(struct thread_information *tip)
+/*
+ * Returns number of bytes written
+ */
+static int net_send_data(int fd, void *buf, unsigned int buf_len)
 {
-	const int head = tip->fifo.head;
-	const int next = (head + 1) & (FIFO_SIZE - 1);
+	int ret;
+	unsigned int bytes_left = buf_len;
 
-	if (head != tip->fifo.tail) {
-		struct tip_subbuf *ts = tip->fifo.q[head];
+	while (bytes_left) {
+		ret = send(fd, buf, bytes_left, 0);
+		if (ret < 0) {
+			perror("send");
+			break;
+		}
 
-		store_barrier();
-		tip->fifo.head = next;
-		return ts;
+		buf += ret;
+		bytes_left -= ret;
 	}
 
-	return NULL;
+	return buf_len - bytes_left;
 }
 
-static inline int subbuf_fifo_queue(struct thread_information *tip,
-				    struct tip_subbuf *ts)
+static int net_send_header(int fd, int cpu, char *buts_name, int len)
 {
-	const int tail = tip->fifo.tail;
-	const int next = (tail + 1) & (FIFO_SIZE - 1);
+	struct blktrace_net_hdr hdr;
 
-	if (next != tip->fifo.head) {
-		tip->fifo.q[tail] = ts;
-		store_barrier();
-		tip->fifo.tail = next;
-		return 0;
-	}
+	memset(&hdr, 0, sizeof(hdr));
 
-	fprintf(stderr, "fifo too small!\n");
-	return 1;
+	hdr.magic = BLK_IO_TRACE_MAGIC;
+	memset(hdr.buts_name, 0, sizeof(hdr.buts_name));
+	strncpy(hdr.buts_name, buts_name, sizeof(hdr.buts_name));
+	hdr.buts_name[sizeof(hdr.buts_name) - 1] = '\0';
+	hdr.cpu = cpu;
+	hdr.max_cpus = max_cpus;
+	hdr.len = len;
+	hdr.cl_id = getpid();
+	hdr.buf_size = buf_size;
+	hdr.buf_nr = buf_nr;
+	hdr.page_size = pagesize;
+
+	return net_send_data(fd, &hdr, sizeof(hdr)) != sizeof(hdr);
 }
 
-/*
- * For file output, truncate and mmap the file appropriately
- */
-static int mmap_subbuf(struct thread_information *tip, unsigned int maxlen)
+static void net_send_open_close(int fd, int cpu, char *buts_name, int len)
 {
-	int ofd = fileno(tip->ofile);
-	int ret;
+	struct blktrace_net_hdr ret_hdr;
+
+	net_send_header(fd, cpu, buts_name, len);
+	net_recv_data(fd, &ret_hdr, sizeof(ret_hdr));
+}
+
+static void net_send_open(int fd, int cpu, char *buts_name)
+{
+	net_send_open_close(fd, cpu, buts_name, 0);
+}
 
+static void net_send_close(int fd, char *buts_name, int drops)
+{
 	/*
-	 * extend file, if we have to. use chunks of 16 subbuffers.
+	 * Overload CPU w/ number of drops
+	 *
+	 * XXX: Need to clear/set done around call - done=1 (which
+	 * is true here) stops reads from happening... :-(
 	 */
-	if (tip->fs_off + buf_size > tip->fs_buf_len) {
-		if (tip->fs_buf) {
-			munlock(tip->fs_buf, tip->fs_buf_len);
-			munmap(tip->fs_buf, tip->fs_buf_len);
-			tip->fs_buf = NULL;
-		}
+	done = 0;
+	net_send_open_close(fd, drops, buts_name, 1);
+	done = 1;
+}
 
-		tip->fs_off = tip->fs_size & (page_size - 1);
-		tip->fs_buf_len = (16 * buf_size) - tip->fs_off;
-		tip->fs_max_size += tip->fs_buf_len;
+static void ack_open_close(int fd, char *buts_name)
+{
+	net_send_header(fd, 0, buts_name, 2);
+}
 
-		if (ftruncate(ofd, tip->fs_max_size) < 0) {
-			perror("ftruncate");
-			return -1;
-		}
+static void net_send_drops(int fd)
+{
+	struct list_head *p;
 
-		tip->fs_buf = mmap(NULL, tip->fs_buf_len, PROT_WRITE,
-				   MAP_SHARED, ofd, tip->fs_size - tip->fs_off);
-		if (tip->fs_buf == MAP_FAILED) {
-			perror("mmap");
-			return -1;
-		}
-		mlock(tip->fs_buf, tip->fs_buf_len);
-	}
+	__list_for_each(p, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
 
-	ret = read_data(tip, tip->fs_buf + tip->fs_off, maxlen);
-	if (ret >= 0) {
-		tip->data_read += ret;
-		tip->fs_size += ret;
-		tip->fs_off += ret;
-		return 0;
+		net_send_close(fd, dpp->buts_name, dpp->drops);
 	}
-
-	return -1;
 }
 
 /*
- * Use the copy approach for pipes and network
+ * Returns:
+ *	 0: "EOF"
+ *	 1: OK
+ *	-1: Error
  */
-static int get_subbuf(struct thread_information *tip, unsigned int maxlen)
+static int net_get_header(struct cl_conn *nc, struct blktrace_net_hdr *bnh)
 {
-	struct tip_subbuf *ts = malloc(sizeof(*ts));
-	int ret;
-
-	ts->buf = malloc(buf_size);
-	ts->max_len = maxlen;
+	int bytes_read;
+	int fl = fcntl(nc->fd, F_GETFL);
 
-	ret = read_data(tip, ts->buf, ts->max_len);
-	if (ret > 0) {
-		ts->len = ret;
-		tip->data_read += ret;
-		if (subbuf_fifo_queue(tip, ts))
-			return -1;
-	}
+	fcntl(nc->fd, F_SETFL, fl | O_NONBLOCK);
+	bytes_read = __net_recv_data(nc->fd, bnh, sizeof(*bnh));
+	fcntl(nc->fd, F_SETFL, fl & ~O_NONBLOCK);
 
-	return ret;
+	if (bytes_read == sizeof(*bnh))
+		return 1;
+	else if (bytes_read == 0)
+		return 0;
+	else
+		return -1;
 }
 
-static int get_subbuf_sendfile(struct thread_information *tip,
-			       unsigned int maxlen)
+static int net_setup_addr(void)
 {
-	struct tip_subbuf *ts;
+	struct sockaddr_in *addr = &hostname_addr;
 
-	wait_for_data(tip);
+	memset(addr, 0, sizeof(*addr));
+	addr->sin_family = AF_INET;
+	addr->sin_port = htons(net_port);
 
-	/*
-	 * hack to get last data out, we can't use sendfile for that
-	 */
-	if (is_done())
-		return get_subbuf(tip, maxlen);
+	if (inet_aton(hostname, &addr->sin_addr) != 1) {
+		struct hostent *hent;
+retry:
+		hent = gethostbyname(hostname);
+		if (!hent) {
+			if (h_errno == TRY_AGAIN) {
+				usleep(100);
+				goto retry;
+			} else if (h_errno == NO_RECOVERY) {
+				fprintf(stderr, "gethostbyname(%s)"
+					"non-recoverable error encountered\n",
+					hostname);
+			} else {
+				/*
+				 * HOST_NOT_FOUND, NO_ADDRESS or NO_DATA
+				 */
+				fprintf(stderr, "Host %s not found\n",
+					hostname);
+			}
+			return 1;
+		}
 
-	if (tip->sendfile_pending) {
-		usleep(100);
-		return 0;
+		memcpy(&addr->sin_addr, hent->h_addr, 4);
+		memset(hostname, 0, sizeof(hostname));
+		strncpy(hostname, hent->h_name, sizeof(hostname));
+		hostname[sizeof(hostname) - 1] = '\0';
+	}
+
+	return 0;
+}
+
+static int net_setup_client(void)
+{
+	int fd;
+	struct sockaddr_in *addr = &hostname_addr;
+
+	fd = my_socket(AF_INET, SOCK_STREAM, 0);
+	if (fd < 0) {
+		perror("client: socket");
+		return -1;
 	}
 
-	ts = malloc(sizeof(*ts));
-	ts->buf = NULL;
-	ts->max_len = 0;
+	if (connect(fd, (struct sockaddr *)addr, sizeof(*addr)) < 0) {
+		if (errno == ECONNREFUSED)
+			fprintf(stderr,
+				"\nclient: Connection to %s refused, "
+				"perhaps the server is not started?\n\n",
+				hostname);
+		else
+			perror("client: connect");
 
-	if (subbuf_fifo_queue(tip, ts))
+		close(fd);
 		return -1;
+	}
 
-	tip->sendfile_pending++;
-	return buf_size;
+	return fd;
 }
 
-static void close_thread(struct thread_information *tip)
+static int open_client_connections(void)
 {
-	if (tip->fd != -1)
-		close(tip->fd);
-	if (tip->pfd != -1)
-		close(tip->pfd);
-	if (tip->ofile)
-		fclose(tip->ofile);
-	if (tip->ofile_buffer)
-		free(tip->ofile_buffer);
-	if (tip->fd_buf)
-		free(tip->fd_buf);
-	if (tip->pfd_buf)
-		free(tip->pfd_buf);
+	int cpu;
+	size_t alloc_size = CPU_ALLOC_SIZE(max_cpus);
+
+	cl_fds = calloc(ncpus, sizeof(*cl_fds));
+	for (cpu = 0; cpu < max_cpus; cpu++) {
+		if (!CPU_ISSET_S(cpu, alloc_size, online_cpus))
+			continue;
+		cl_fds[cpu] = net_setup_client();
+		if (cl_fds[cpu] < 0)
+			goto err;
+	}
+	return 0;
 
-	tip->fd = -1;
-	tip->pfd = -1;
-	tip->ofile = NULL;
-	tip->ofile_buffer = NULL;
-	tip->fd_buf = NULL;
+err:
+	while (cpu > 0)
+		close(cl_fds[cpu--]);
+	free(cl_fds);
+	return 1;
 }
 
-static void tip_ftrunc_final(struct thread_information *tip)
+static void close_client_connections(void)
 {
-	/*
-	 * truncate to right size and cleanup mmap
-	 */
-	if (tip->ofile_mmap) {
-		int ofd = fileno(tip->ofile);
-
-		if (tip->fs_buf)
-			munmap(tip->fs_buf, tip->fs_buf_len);
-
-		ftruncate(ofd, tip->fs_size);
+	if (cl_fds) {
+		int cpu, *fdp;
+		size_t alloc_size = CPU_ALLOC_SIZE(max_cpus);
+
+		for (cpu = 0, fdp = cl_fds; cpu < max_cpus; cpu++, fdp++) {
+			if (!CPU_ISSET_S(cpu, alloc_size, online_cpus))
+				continue;
+			if (*fdp >= 0) {
+				net_send_drops(*fdp);
+				net_close_connection(fdp);
+			}
+		}
+		free(cl_fds);
 	}
 }
 
-static void *thread_main(void *arg)
+static int setup_buts(void)
 {
-	struct thread_information *tip = arg;
-	pid_t pid = getpid();
-	cpu_set_t cpu_mask;
-
-	CPU_ZERO(&cpu_mask);
-	CPU_SET((tip->cpu), &cpu_mask);
+	struct list_head *p;
+	int ret = 0;
 
-	if (sched_setaffinity(pid, sizeof(cpu_mask), &cpu_mask) == -1) {
-		perror("sched_setaffinity");
-		exit_trace(1);
+	__list_for_each(p, &devpaths) {
+		struct blk_user_trace_setup buts;
+		struct devpath *dpp = list_entry(p, struct devpath, head);
+
+		memset(&buts, 0, sizeof(buts));
+		buts.buf_size = buf_size;
+		buts.buf_nr = buf_nr;
+		buts.act_mask = act_mask;
+
+		if (ioctl(dpp->fd, BLKTRACESETUP, &buts) >= 0) {
+			dpp->ncpus = max_cpus;
+			dpp->buts_name = strdup(buts.name);
+			dpp->setup_done = 1;
+			if (dpp->stats)
+				free(dpp->stats);
+			dpp->stats = calloc(dpp->ncpus, sizeof(*dpp->stats));
+			memset(dpp->stats, 0, dpp->ncpus * sizeof(*dpp->stats));
+		} else {
+			fprintf(stderr, "BLKTRACESETUP(2) %s failed: %d/%s\n",
+				dpp->path, errno, strerror(errno));
+			ret++;
+		}
 	}
 
-	snprintf(tip->fn, sizeof(tip->fn), "%s/block/%s/trace%d",
-			relay_path, tip->device->buts_name, tip->cpu);
-	tip->fd = open(tip->fn, O_RDONLY);
-	if (tip->fd < 0) {
-		perror(tip->fn);
-		fprintf(stderr,"Thread %d failed open of %s\n", tip->cpu,
-			tip->fn);
-		exit_trace(1);
-	}
+	return ret;
+}
 
-	if (net_mode == Net_client && net_use_sendfile) {
-		char tmp[MAXPATHLEN + 64];
+static void start_buts(void)
+{
+	struct list_head *p;
 
-		snprintf(tmp, sizeof(tmp), "%s/block/%s/trace%d.padding",
-			 relay_path, tip->device->buts_name, tip->cpu);
+	__list_for_each(p, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
 
-		tip->pfd = open(tmp, O_RDONLY);
-		if (tip->pfd < 0) {
-			fprintf(stderr, "Couldn't open padding file %s\n", tmp);
-			exit_trace(1);
+		if (ioctl(dpp->fd, BLKTRACESTART) < 0) {
+			fprintf(stderr, "BLKTRACESTART %s failed: %d/%s\n",
+				dpp->path, errno, strerror(errno));
 		}
-
-		tip->pfd_buf = malloc(buf_nr * sizeof(size_t));
 	}
+}
 
-	while (!is_done()) {
-		if (tip->get_subbuf(tip, buf_size) < 0)
-			break;
-	}
+static int get_drops(struct devpath *dpp)
+{
+	int fd, drops = 0;
+	char fn[MAXPATHLEN + 64], tmp[256];
 
-	/*
-	 * trace is stopped, pull data until we get a short read
-	 */
-	while (tip->get_subbuf(tip, buf_size) > 0)
-		;
+	snprintf(fn, sizeof(fn), "%s/block/%s/dropped", debugfs_path,
+		 dpp->buts_name);
 
-	tip_ftrunc_final(tip);
-	tip->exited = 1;
-	return NULL;
+	fd = my_open(fn, O_RDONLY);
+	if (fd < 0) {
+		/*
+		 * This may be ok: the kernel may not support
+		 * dropped counts.
+		 */
+		if (errno != ENOENT)
+			fprintf(stderr, "Could not open %s: %d/%s\n",
+				fn, errno, strerror(errno));
+		return 0;
+	} else if (read(fd, tmp, sizeof(tmp)) < 0) {
+		fprintf(stderr, "Could not read %s: %d/%s\n",
+			fn, errno, strerror(errno));
+	} else
+		drops = atoi(tmp);
+	close(fd);
+
+	return drops;
 }
 
-static int write_data_net(int fd, void *buf, unsigned int buf_len)
+static void get_all_drops(void)
 {
-	unsigned int bytes_left = buf_len;
-	int ret;
+	struct list_head *p;
 
-	while (bytes_left) {
-		ret = send(fd, buf, bytes_left, 0);
-		if (ret < 0) {
-			perror("send");
-			return 1;
-		}
+	__list_for_each(p, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
 
-		buf += ret;
-		bytes_left -= ret;
+		dpp->drops = get_drops(dpp);
 	}
-
-	return 0;
 }
 
-static int net_send_header(struct thread_information *tip, unsigned int len)
+static inline struct trace_buf *alloc_trace_buf(int cpu, int bufsize)
 {
-	struct blktrace_net_hdr hdr;
+	struct trace_buf *tbp;
 
-	hdr.magic = BLK_IO_TRACE_MAGIC;
-	strcpy(hdr.buts_name, tip->device->buts_name);
-	hdr.cpu = tip->cpu;
-	hdr.max_cpus = ncpus;
-	hdr.len = len;
+	tbp = malloc(sizeof(*tbp) + bufsize);
+	INIT_LIST_HEAD(&tbp->head);
+	tbp->len = 0;
+	tbp->buf = (void *)(tbp + 1);
+	tbp->cpu = cpu;
+	tbp->dpp = NULL;	/* Will be set when tbp is added */
 
-	return write_data_net(net_out_fd, &hdr, sizeof(hdr));
+	return tbp;
 }
 
-/*
- * send header with 0 length to signal end-of-run
- */
-static void net_client_send_close(void)
+static void free_tracer_heads(struct devpath *dpp)
 {
-	struct blktrace_net_hdr hdr;
+	int cpu;
+	struct tracer_devpath_head *hd;
 
-	hdr.magic = BLK_IO_TRACE_MAGIC;
-	hdr.cpu = 0;
-	hdr.max_cpus = ncpus;
-	hdr.len = 0;
+	for (cpu = 0, hd = dpp->heads; cpu < max_cpus; cpu++, hd++) {
+		if (hd->prev)
+			free(hd->prev);
 
-	write_data_net(net_out_fd, &hdr, sizeof(hdr));
+		pthread_mutex_destroy(&hd->mutex);
+	}
+	free(dpp->heads);
 }
 
-static int flush_subbuf_net(struct thread_information *tip,
-			    struct tip_subbuf *ts)
+static int setup_tracer_devpaths(void)
 {
-	if (net_send_header(tip, ts->len))
-		return 1;
-	if (write_data_net(net_out_fd, ts->buf, ts->len))
-		return 1;
+	struct list_head *p;
 
-	free(ts->buf);
-	free(ts);
-	return 0;
-}
+	if (net_client_use_send())
+		if (open_client_connections())
+			return 1;
 
-static int net_sendfile(struct thread_information *tip, struct tip_subbuf *ts)
-{
-	int ret = sendfile(net_out_fd, tip->fd, NULL, ts->len);
+	__list_for_each(p, &devpaths) {
+		int cpu;
+		struct tracer_devpath_head *hd;
+		struct devpath *dpp = list_entry(p, struct devpath, head);
 
-	if (ret < 0) {
-		perror("sendfile");
-		return 1;
-	} else if (ret < (int) ts->len) {
-		fprintf(stderr, "short sendfile send (%d of %d)\n", ret, ts->len);
-		return 1;
+		dpp->heads = calloc(max_cpus, sizeof(struct tracer_devpath_head));
+		for (cpu = 0, hd = dpp->heads; cpu < max_cpus; cpu++, hd++) {
+			INIT_LIST_HEAD(&hd->head);
+			pthread_mutex_init(&hd->mutex, NULL);
+			hd->prev = NULL;
+		}
 	}
 
 	return 0;
 }
 
-static int get_subbuf_padding(struct thread_information *tip, off_t off)
+static inline void add_trace_buf(struct devpath *dpp, int cpu,
+						struct trace_buf **tbpp)
 {
-	int padding_size = buf_nr * sizeof(size_t);
-	int ret;
+	struct trace_buf *tbp = *tbpp;
+	struct tracer_devpath_head *hd = &dpp->heads[cpu];
 
-	ret = read(tip->pfd, tip->pfd_buf, padding_size);
-	if (ret == padding_size) {
-		int subbuf = (off / buf_size) % buf_nr;
+	tbp->dpp = dpp;
 
-		ret = tip->pfd_buf[subbuf];
-	} else if (ret < 0)
-		perror("tip pad read");
-	else {
-		fprintf(stderr, "bad pad size read\n");
-		ret = -1;
-	}
+	pthread_mutex_lock(&hd->mutex);
+	list_add_tail(&tbp->head, &hd->head);
+	pthread_mutex_unlock(&hd->mutex);
 
-	return ret;
+	*tbpp = alloc_trace_buf(cpu, buf_size);
 }
 
-static int flush_subbuf_sendfile(struct thread_information *tip,
-				 struct tip_subbuf *ts)
+static inline void incr_entries(int entries_handled)
 {
-	int pad, ret = 1;
-
-	/*
-	 * currently we cannot use sendfile() on the last bytes read, as they
-	 * may not be a full subbuffer. get_subbuf_sendfile() falls back to
-	 * the read approach for those, so use send() to ship them out
-	 */
-	if (ts->buf)
-		return flush_subbuf_net(tip, ts);
-	
-	pad = get_subbuf_padding(tip, tip->ofile_offset);
-	if (pad == -1)
-		goto err;
+	pthread_mutex_lock(&dp_mutex);
+	if (dp_entries == 0)
+		pthread_cond_signal(&dp_cond);
+	dp_entries += entries_handled;
+	pthread_mutex_unlock(&dp_mutex);
+}
 
-	ts->len = buf_size - pad;
+static void decr_entries(int handled)
+{
+	pthread_mutex_lock(&dp_mutex);
+	dp_entries -= handled;
+	pthread_mutex_unlock(&dp_mutex);
+}
 
-	if (net_send_header(tip, ts->len))
-		goto err;
-	if (net_sendfile(tip, ts))
-		goto err;
+static int wait_empty_entries(void)
+{
+	pthread_mutex_lock(&dp_mutex);
+	while (!done && dp_entries == 0)
+		t_pthread_cond_wait(&dp_cond, &dp_mutex);
+	pthread_mutex_unlock(&dp_mutex);
 
-	tip->data_read += ts->len;
-	tip->ofile_offset += buf_size;
-	ret = 0;
-err:
-	tip->sendfile_pending--;
-	free(ts);
-	return ret;
+	return !done;
 }
 
-static int write_data(struct thread_information *tip, void *buf,
-		      unsigned int buf_len)
+static int add_devpath(char *path)
 {
-	int ret;
+	int fd;
+	struct devpath *dpp;
+	struct list_head *p;
 
-	if (!buf_len)
-		return 0;
-
-	while (1) {
-		ret = fwrite(buf, buf_len, 1, tip->ofile);
-		if (ret == 1)
-			break;
-
-		if (ret < 0) {
-			perror("write");
-			return 1;
-		}
+	/*
+	 * Verify device is not duplicated
+	 */
+	__list_for_each(p, &devpaths) {
+	       struct devpath *tmp = list_entry(p, struct devpath, head);
+	       if (!strcmp(tmp->path, path))
+		        return 0;
+	}
+	/*
+	 * Verify device is valid before going too far
+	 */
+	fd = my_open(path, O_RDONLY | O_NONBLOCK);
+	if (fd < 0) {
+		fprintf(stderr, "Invalid path %s specified: %d/%s\n",
+			path, errno, strerror(errno));
+		return 1;
 	}
 
-	if (tip->ofile_stdout)
-		fflush(tip->ofile);
+	dpp = malloc(sizeof(*dpp));
+	memset(dpp, 0, sizeof(*dpp));
+	dpp->path = strdup(path);
+	dpp->fd = fd;
+	ndevs++;
+	list_add_tail(&dpp->head, &devpaths);
 
 	return 0;
 }
 
-static int flush_subbuf_file(struct thread_information *tip,
-			     struct tip_subbuf *ts)
+static void rel_devpaths(void)
 {
-	unsigned int offset = 0;
-	struct blk_io_trace *t;
-	int pdu_len, events = 0;
+	struct list_head *p, *q;
 
-	/*
-	 * surplus from last run
-	 */
-	if (tip->leftover_ts) {
-		struct tip_subbuf *prev_ts = tip->leftover_ts;
-
-		if (prev_ts->len + ts->len > prev_ts->max_len) {
-			prev_ts->max_len += ts->len;
-			prev_ts->buf = realloc(prev_ts->buf, prev_ts->max_len);
-		}
+	list_for_each_safe(p, q, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
 
-		memcpy(prev_ts->buf + prev_ts->len, ts->buf, ts->len);
-		prev_ts->len += ts->len;
+		list_del(&dpp->head);
+		if (dpp->setup_done)
+			__stop_trace(dpp->fd);
+		close(dpp->fd);
 
-		free(ts->buf);
-		free(ts);
+		if (dpp->heads)
+			free_tracer_heads(dpp);
 
-		ts = prev_ts;
-		tip->leftover_ts = NULL;
+		dpp_free(dpp);
+		ndevs--;
 	}
+}
 
-	while (offset + sizeof(*t) <= ts->len) {
-		t = ts->buf + offset;
+static int flush_subbuf_net(struct trace_buf *tbp)
+{
+	int fd = cl_fds[tbp->cpu];
+	struct devpath *dpp = tbp->dpp;
 
-		if (verify_trace(t)) {
-			write_data(tip, ts->buf, offset);
-			return -1;
-		}
+	if (net_send_header(fd, tbp->cpu, dpp->buts_name, tbp->len))
+		return 1;
+	else if (net_send_data(fd, tbp->buf, tbp->len) != tbp->len)
+		return 1;
 
-		pdu_len = t->pdu_len;
+	return 0;
+}
 
-		if (offset + sizeof(*t) + pdu_len > ts->len)
-			break;
+static int
+handle_list_net(__attribute__((__unused__))struct tracer_devpath_head *hd,
+		struct list_head *list)
+{
+	struct trace_buf *tbp;
+	struct list_head *p, *q;
+	int entries_handled = 0;
 
-		offset += sizeof(*t) + pdu_len;
-		tip->events_processed++;
-		tip->data_read += sizeof(*t) + pdu_len;
-		events++;
-	}
+	list_for_each_safe(p, q, list) {
+		tbp = list_entry(p, struct trace_buf, head);
 
-	if (write_data(tip, ts->buf, offset))
-		return -1;
+		list_del(&tbp->head);
+		entries_handled++;
 
-	/*
-	 * leftover bytes, save them for next time
-	 */
-	if (offset != ts->len) {
-		tip->leftover_ts = ts;
-		ts->len -= offset;
-		memmove(ts->buf, ts->buf + offset, ts->len);
-	} else {
-		free(ts->buf);
-		free(ts);
+		if (cl_fds[tbp->cpu] >= 0) {
+			if (flush_subbuf_net(tbp)) {
+				close(cl_fds[tbp->cpu]);
+				cl_fds[tbp->cpu] = -1;
+			}
+		}
+
+		free(tbp);
 	}
 
-	return events;
+	return entries_handled;
 }
 
-static int write_tip_events(struct thread_information *tip)
+/*
+ * Tack 'tbp's buf onto the tail of 'prev's buf
+ */
+static struct trace_buf *tb_combine(struct trace_buf *prev,
+				    struct trace_buf *tbp)
 {
-	struct tip_subbuf *ts = subbuf_fifo_dequeue(tip);
+	unsigned long tot_len;
+
+	tot_len = prev->len + tbp->len;
+	if (tot_len > buf_size) {
+		/*
+		 * tbp->head isn't connected (it was 'prev'
+		 * so it had been taken off of the list
+		 * before). Therefore, we can realloc
+		 * the whole structures, as the other fields
+		 * are "static".
+		 */
+		prev = realloc(prev, sizeof(*prev) + tot_len);
+		prev->buf = (void *)(prev + 1);
+	}
 
-	if (ts)
-		return tip->flush_subbuf(tip, ts);
+	memcpy(prev->buf + prev->len, tbp->buf, tbp->len);
+	prev->len = tot_len;
 
-	return 0;
+	free(tbp);
+	return prev;
 }
 
-/*
- * scans the tips we know and writes out the subbuffers we accumulate
- */
-static void get_and_write_events(void)
+static int handle_list_file(struct tracer_devpath_head *hd,
+			    struct list_head *list)
 {
-	struct device_information *dip;
-	struct thread_information *tip;
-	int i, j, events, ret, tips_running;
+	int off, t_len, nevents;
+	struct blk_io_trace *t;
+	struct list_head *p, *q;
+	int entries_handled = 0;
+	struct trace_buf *tbp, *prev;
 
-	while (!is_done()) {
-		events = 0;
+	prev = hd->prev;
+	list_for_each_safe(p, q, list) {
+		tbp = list_entry(p, struct trace_buf, head);
+		list_del(&tbp->head);
+		entries_handled++;
 
-		for_each_dip(dip, i) {
-			for_each_tip(dip, tip, j) {
-				ret = write_tip_events(tip);
-				if (ret > 0)
-					events += ret;
-			}
+		/*
+		 * If there was some leftover before, tack this new
+		 * entry onto the tail of the previous one.
+		 */
+		if (prev)
+			tbp = tb_combine(prev, tbp);
+
+		/*
+		 * See how many whole traces there are - send them
+		 * all out in one go.
+		 */
+		off = 0;
+		nevents = 0;
+		while (off + (int)sizeof(*t) <= tbp->len) {
+			t = (struct blk_io_trace *)(tbp->buf + off);
+			t_len = sizeof(*t) + t->pdu_len;
+			if (off + t_len > tbp->len)
+				break;
+
+			off += t_len;
+			nevents++;
 		}
+		if (nevents)
+			pdc_nev_update(tbp->dpp, tbp->cpu, nevents);
 
-		if (!events)
-			usleep(10);
+		/*
+		 * Write any full set of traces, any remaining data is kept
+		 * for the next pass.
+		 */
+		if (off) {
+			if (write_data(tbp->buf, off) || off == tbp->len) {
+				free(tbp);
+				prev = NULL;
+			}
+			else {
+				/*
+				 * Move valid data to beginning of buffer
+				 */
+				tbp->len -= off;
+				memmove(tbp->buf, tbp->buf + off, tbp->len);
+				prev = tbp;
+			}
+		} else
+			prev = tbp;
 	}
+	hd->prev = prev;
 
-	/*
-	 * reap stored events
-	 */
-	do {
-		events = 0;
-		tips_running = 0;
-		for_each_dip(dip, i) {
-			for_each_tip(dip, tip, j) {
-				ret = write_tip_events(tip);
-				if (ret > 0)
-					events += ret;
-				tips_running += !tip->exited;
+	return entries_handled;
+}
+
+static void __process_trace_bufs(void)
+{
+	int cpu;
+	struct list_head *p;
+	struct list_head list;
+	int handled = 0;
+
+	__list_for_each(p, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
+		struct tracer_devpath_head *hd = dpp->heads;
+
+		for (cpu = 0; cpu < max_cpus; cpu++, hd++) {
+			pthread_mutex_lock(&hd->mutex);
+			if (list_empty(&hd->head)) {
+				pthread_mutex_unlock(&hd->mutex);
+				continue;
 			}
+
+			list_replace_init(&hd->head, &list);
+			pthread_mutex_unlock(&hd->mutex);
+
+			handled += handle_list(hd, &list);
 		}
-		usleep(10);
-	} while (events || tips_running);
+	}
+
+	if (handled)
+		decr_entries(handled);
+}
+
+static void process_trace_bufs(void)
+{
+	while (wait_empty_entries())
+		__process_trace_bufs();
 }
 
-static void wait_for_threads(void)
+static void clean_trace_bufs(void)
 {
 	/*
-	 * for piped or network output, poll and fetch data for writeout.
-	 * for files, we just wait around for trace threads to exit
+	 * No mutex needed here: we're only reading from the lists,
+	 * tracers are done
 	 */
-	if ((output_name && !strcmp(output_name, "-")) ||
-	    net_mode == Net_client)
-		get_and_write_events();
-	else {
-		struct device_information *dip;
-		struct thread_information *tip;
-		int i, j, tips_running;
+	while (dp_entries)
+		__process_trace_bufs();
+}
+
+static inline void read_err(int cpu, char *ifn)
+{
+	if (errno != EAGAIN)
+		fprintf(stderr, "Thread %d failed read of %s: %d/%s\n",
+			cpu, ifn, errno, strerror(errno));
+}
 
-		do {
-			tips_running = 0;
-			usleep(1000);
+static int net_sendfile(struct io_info *iop)
+{
+	int ret;
 
-			for_each_dip(dip, i)
-				for_each_tip(dip, tip, j)
-					tips_running += !tip->exited;
-		} while (tips_running);
+	ret = sendfile(iop->ofd, iop->ifd, NULL, iop->ready);
+	if (ret < 0) {
+		perror("sendfile");
+		return 1;
+	} else if (ret < (int)iop->ready) {
+		fprintf(stderr, "short sendfile send (%d of %d)\n",
+			ret, iop->ready);
+		return 1;
 	}
 
-	if (net_mode == Net_client)
-		net_client_send_close();
+	return 0;
+}
+
+static inline int net_sendfile_data(struct tracer *tp, struct io_info *iop)
+{
+	struct devpath *dpp = iop->dpp;
+
+	if (net_send_header(iop->ofd, tp->cpu, dpp->buts_name, iop->ready))
+		return 1;
+	return net_sendfile(iop);
 }
 
-static int fill_ofname(struct thread_information *tip, char *dst,
-		       char *buts_name)
+static int fill_ofname(char *dst, int dstlen, char *subdir, char *buts_name,
+		       int cpu)
 {
+	int len;
 	struct stat sb;
-	int len = 0;
-	time_t t;
 
 	if (output_dir)
-		len = sprintf(dst, "%s/", output_dir);
+		len = snprintf(dst, dstlen, "%s/", output_dir);
+	else
+		len = snprintf(dst, dstlen, "./");
 
-	if (net_mode == Net_server) {
-		len += sprintf(dst + len, "%s-", inet_ntoa(tip->cl_in_addr));
-		time(&t);
-		len += strftime(dst + len, 64, "%F-%T/", gmtime(&t));
-	}
+	if (subdir)
+		len += snprintf(dst + len, dstlen - len, "%s", subdir);
 
 	if (stat(dst, &sb) < 0) {
 		if (errno != ENOENT) {
-			perror("stat");
+			fprintf(stderr,
+				"Destination dir %s stat failed: %d/%s\n",
+				dst, errno, strerror(errno));
 			return 1;
 		}
-		if (mkdir(dst, 0755) < 0) {
-			perror(dst);
-			fprintf(stderr, "Can't make output dir\n");
+		/*
+		 * There is no synchronization between multiple threads
+		 * trying to create the directory at once.  It's harmless
+		 * to let them try, so just detect the problem and move on.
+		 */
+		if (mkdir(dst, 0755) < 0 && errno != EEXIST) {
+			fprintf(stderr,
+				"Destination dir %s can't be made: %d/%s\n",
+				dst, errno, strerror(errno));
 			return 1;
 		}
 	}
 
 	if (output_name)
-		sprintf(dst + len, "%s.blktrace.%d", output_name, tip->cpu);
+		snprintf(dst + len, dstlen - len, "%s.blktrace.%d",
+			 output_name, cpu);
 	else
-		sprintf(dst + len, "%s.blktrace.%d", buts_name, tip->cpu);
+		snprintf(dst + len, dstlen - len, "%s.blktrace.%d",
+			 buts_name, cpu);
 
 	return 0;
 }
 
-static void fill_ops(struct thread_information *tip)
+static int set_vbuf(struct io_info *iop, int mode, size_t size)
 {
-	/*
-	 * setup ops
-	 */
-	if (net_mode == Net_client) {
-		if (net_use_sendfile) {
-			tip->get_subbuf = get_subbuf_sendfile;
-			tip->flush_subbuf = flush_subbuf_sendfile;
-		} else {
-			tip->get_subbuf = get_subbuf;
-			tip->flush_subbuf = flush_subbuf_net;
-		}
-	} else {
-		if (tip->ofile_mmap)
-			tip->get_subbuf = mmap_subbuf;
-		else
-			tip->get_subbuf = get_subbuf;
-
-		tip->flush_subbuf = flush_subbuf_file;
+	iop->obuf = malloc(size);
+	if (setvbuf(iop->ofp, iop->obuf, mode, size) < 0) {
+		fprintf(stderr, "setvbuf(%s, %d) failed: %d/%s\n",
+			iop->dpp->path, (int)size, errno,
+			strerror(errno));
+		free(iop->obuf);
+		return 1;
 	}
-			
-	if (net_mode == Net_server)
-		tip->read_data = read_data_net;
-	else
-		tip->read_data = read_data_file;
-}
-
-static int tip_open_output(struct device_information *dip,
-			   struct thread_information *tip)
-{
-	int pipeline = output_name && !strcmp(output_name, "-");
-	int mode, vbuf_size;
-	char op[128];
-
-	if (net_mode == Net_client) {
-		tip->ofile = NULL;
-		tip->ofile_stdout = 0;
-		tip->ofile_mmap = 0;
-		goto done;
-	} else if (pipeline) {
-		tip->ofile = fdopen(STDOUT_FILENO, "w");
-		tip->ofile_stdout = 1;
-		tip->ofile_mmap = 0;
-		mode = _IOLBF;
-		vbuf_size = 512;
+
+	return 0;
+}
+
+static int iop_open(struct io_info *iop, int cpu)
+{
+	char hostdir[MAXPATHLEN + 64];
+
+	iop->ofd = -1;
+	if (net_mode == Net_server) {
+		struct cl_conn *nc = iop->nc;
+		int len;
+
+		len = snprintf(hostdir, sizeof(hostdir), "%s-",
+			       nc->ch->hostname);
+		len += strftime(hostdir + len, sizeof(hostdir) - len, "%F-%T/",
+				gmtime(&iop->dpp->cl_connect_time));
 	} else {
-		if (fill_ofname(tip, op, dip->buts_name))
-			return 1;
-		tip->ofile = fopen(op, "w+");
-		tip->ofile_stdout = 0;
-		tip->ofile_mmap = 1;
-		mode = _IOFBF;
-		vbuf_size = OFILE_BUF;
+		hostdir[0] = 0;
 	}
 
-	if (tip->ofile == NULL) {
-		perror(op);
+	if (fill_ofname(iop->ofn, sizeof(iop->ofn), hostdir,
+			iop->dpp->buts_name, cpu))
+		return 1;
+
+	iop->ofp = my_fopen(iop->ofn, "w+");
+	if (iop->ofp == NULL) {
+		fprintf(stderr, "Open output file %s failed: %d/%s\n",
+			iop->ofn, errno, strerror(errno));
 		return 1;
 	}
 
-	tip->ofile_buffer = malloc(vbuf_size);
-	if (setvbuf(tip->ofile, tip->ofile_buffer, mode, vbuf_size)) {
-		perror("setvbuf");
-		close_thread(tip);
+	if (set_vbuf(iop, _IOLBF, FILE_VBUF_SIZE)) {
+		fprintf(stderr, "set_vbuf for file %s failed: %d/%s\n",
+			iop->ofn, errno, strerror(errno));
+		fclose(iop->ofp);
 		return 1;
 	}
 
-done:
-	fill_ops(tip);
+	iop->ofd = fileno(iop->ofp);
 	return 0;
 }
 
-static int start_threads(struct device_information *dip)
+static void close_iop(struct io_info *iop)
 {
-	struct thread_information *tip;
-	int j;
-
-	for_each_tip(dip, tip, j) {
-		tip->cpu = j;
-		tip->device = dip;
-		tip->events_processed = 0;
-		tip->fd = -1;
-		tip->pfd = -1;
-		memset(&tip->fifo, 0, sizeof(tip->fifo));
-		tip->leftover_ts = NULL;
+	struct mmap_info *mip = &iop->mmap_info;
 
-		if (tip_open_output(dip, tip))
-			return 1;
+	if (mip->fs_buf)
+		munmap(mip->fs_buf, mip->fs_buf_len);
 
-		if (pthread_create(&tip->thread, NULL, thread_main, tip)) {
-			perror("pthread_create");
-			close_thread(tip);
-			return 1;
+	if (!piped_output) {
+		if (ftruncate(fileno(iop->ofp), mip->fs_size) < 0) {
+			fprintf(stderr,
+				"Ignoring err: ftruncate(%s): %d/%s\n",
+				iop->ofn, errno, strerror(errno));
 		}
 	}
 
-	return 0;
-}
-
-static void stop_threads(struct device_information *dip)
-{
-	struct thread_information *tip;
-	unsigned long ret;
-	int i;
-
-	for_each_tip(dip, tip, i) {
-		(void) pthread_join(tip->thread, (void *) &ret);
-		close_thread(tip);
-	}
+	if (iop->ofp)
+		fclose(iop->ofp);
+	if (iop->obuf)
+		free(iop->obuf);
 }
 
-static void stop_all_threads(void)
+static void close_ios(struct tracer *tp)
 {
-	struct device_information *dip;
-	int i;
+	while (tp->nios > 0) {
+		struct io_info *iop = &tp->ios[--tp->nios];
 
-	for_each_dip(dip, i)
-		stop_threads(dip);
-}
-
-static void stop_all_tracing(void)
-{
-	struct device_information *dip;
-	int i;
+		iop->dpp->drops = get_drops(iop->dpp);
+		if (iop->ifd >= 0)
+			close(iop->ifd);
 
-	for_each_dip(dip, i)
-		stop_trace(dip);
-}
+		if (iop->ofp)
+			close_iop(iop);
+		else if (iop->ofd >= 0) {
+			struct devpath *dpp = iop->dpp;
 
-static void exit_trace(int status)
-{
-	if (!is_trace_stopped()) {
-		trace_stopped = 1;
-		stop_all_threads();
-		stop_all_tracing();
+			net_send_close(iop->ofd, dpp->buts_name, dpp->drops);
+			net_close_connection(&iop->ofd);
+		}
 	}
 
-	exit(status);
+	free(tp->ios);
+	free(tp->pfds);
 }
 
-static int resize_devices(char *path)
+static int open_ios(struct tracer *tp)
 {
-	int size = (ndevs + 1) * sizeof(struct device_information);
+	struct pollfd *pfd;
+	struct io_info *iop;
+	struct list_head *p;
+
+	tp->ios = calloc(ndevs, sizeof(struct io_info));
+	memset(tp->ios, 0, ndevs * sizeof(struct io_info));
+
+	tp->pfds = calloc(ndevs, sizeof(struct pollfd));
+	memset(tp->pfds, 0, ndevs * sizeof(struct pollfd));
+
+	tp->nios = 0;
+	iop = tp->ios;
+	pfd = tp->pfds;
+	__list_for_each(p, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
+
+		iop->dpp = dpp;
+		iop->ofd = -1;
+		snprintf(iop->ifn, sizeof(iop->ifn), "%s/block/%s/trace%d",
+			debugfs_path, dpp->buts_name, tp->cpu);
+
+		iop->ifd = my_open(iop->ifn, O_RDONLY | O_NONBLOCK);
+		if (iop->ifd < 0) {
+			fprintf(stderr, "Thread %d failed open %s: %d/%s\n",
+				tp->cpu, iop->ifn, errno, strerror(errno));
+			return 1;
+		}
 
-	device_information = realloc(device_information, size);
-	if (!device_information) {
-		fprintf(stderr, "Out of memory, device %s (%d)\n", path, size);
-		return 1;
+		init_mmap_info(&iop->mmap_info);
+
+		pfd->fd = iop->ifd;
+		pfd->events = POLLIN;
+
+		if (piped_output)
+			;
+		else if (net_client_use_sendfile()) {
+			iop->ofd = net_setup_client();
+			if (iop->ofd < 0)
+				goto err;
+			net_send_open(iop->ofd, tp->cpu, dpp->buts_name);
+		} else if (net_mode == Net_none) {
+			if (iop_open(iop, tp->cpu))
+				goto err;
+		} else {
+			/*
+			 * This ensures that the server knows about all
+			 * connections & devices before _any_ closes
+			 */
+			net_send_open(cl_fds[tp->cpu], tp->cpu, dpp->buts_name);
+		}
+
+		pfd++;
+		iop++;
+		tp->nios++;
 	}
-	device_information[ndevs].path = path;
-	ndevs++;
+
 	return 0;
+
+err:
+	close(iop->ifd);	/* tp->nios _not_ bumped */
+	close_ios(tp);
+	return 1;
 }
 
-static int open_devices(void)
+static int handle_pfds_file(struct tracer *tp, int nevs, int force_read)
 {
-	struct device_information *dip;
-	int i;
+	struct mmap_info *mip;
+	int i, ret, nentries = 0;
+	struct pollfd *pfd = tp->pfds;
+	struct io_info *iop = tp->ios;
+
+	for (i = 0; nevs > 0 && i < ndevs; i++, pfd++, iop++) {
+		if (pfd->revents & POLLIN || force_read) {
+			mip = &iop->mmap_info;
+
+			ret = setup_mmap(iop->ofd, buf_size, mip, tp);
+			if (ret < 0) {
+				pfd->events = 0;
+				break;
+			}
 
-	for_each_dip(dip, i) {
-		dip->fd = open(dip->path, O_RDONLY | O_NONBLOCK);
-		if (dip->fd < 0) {
-			perror(dip->path);
-			return 1;
+			ret = read(iop->ifd, mip->fs_buf + mip->fs_off,
+				   buf_size);
+			if (ret > 0) {
+				pdc_dr_update(iop->dpp, tp->cpu, ret);
+				mip->fs_size += ret;
+				mip->fs_off += ret;
+				nentries++;
+			} else if (ret == 0) {
+				/*
+				 * Short reads after we're done stop us
+				 * from trying reads.
+				 */
+				if (tp->is_done)
+					clear_events(pfd);
+			} else {
+				read_err(tp->cpu, iop->ifn);
+				if (errno != EAGAIN || tp->is_done)
+					clear_events(pfd);
+			}
+			nevs--;
 		}
 	}
 
-	return 0;
+	return nentries;
 }
 
-static int start_devices(void)
+static int handle_pfds_netclient(struct tracer *tp, int nevs, int force_read)
 {
-	struct device_information *dip;
-	int i, j, size;
-
-	size = ncpus * sizeof(struct thread_information);
-	thread_information = malloc(size * ndevs);
-	if (!thread_information) {
-		fprintf(stderr, "Out of memory, threads (%d)\n", size * ndevs);
-		return 1;
-	}
-
-	for_each_dip(dip, i) {
-		if (start_trace(dip)) {
-			close(dip->fd);
-			fprintf(stderr, "Failed to start trace on %s\n",
-				dip->path);
-			break;
+	struct stat sb;
+	int i, nentries = 0;
+	struct pollfd *pfd = tp->pfds;
+	struct io_info *iop = tp->ios;
+
+	for (i = 0; i < ndevs; i++, pfd++, iop++) {
+		if (pfd->revents & POLLIN || force_read) {
+			if (fstat(iop->ifd, &sb) < 0) {
+				perror(iop->ifn);
+				pfd->events = 0;
+			} else if (sb.st_size > (off_t)iop->data_queued) {
+				iop->ready = sb.st_size - iop->data_queued;
+				iop->data_queued = sb.st_size;
+
+				if (!net_sendfile_data(tp, iop)) {
+					pdc_dr_update(iop->dpp, tp->cpu,
+						      iop->ready);
+					nentries++;
+				} else
+					clear_events(pfd);
+			}
+			if (--nevs == 0)
+				break;
 		}
 	}
 
-	if (i != ndevs) {
-		__for_each_dip(dip, j, i)
-			stop_trace(dip);
+	if (nentries)
+		incr_entries(nentries);
 
-		return 1;
-	}
+	return nentries;
+}
 
-	for_each_dip(dip, i) {
-		dip->threads = thread_information + (i * ncpus);
-		if (start_threads(dip)) {
-			fprintf(stderr, "Failed to start worker threads\n");
-			break;
+static int handle_pfds_entries(struct tracer *tp, int nevs, int force_read)
+{
+	int i, nentries = 0;
+	struct trace_buf *tbp;
+	struct pollfd *pfd = tp->pfds;
+	struct io_info *iop = tp->ios;
+
+	tbp = alloc_trace_buf(tp->cpu, buf_size);
+	for (i = 0; i < ndevs; i++, pfd++, iop++) {
+		if (pfd->revents & POLLIN || force_read) {
+			tbp->len = read(iop->ifd, tbp->buf, buf_size);
+			if (tbp->len > 0) {
+				pdc_dr_update(iop->dpp, tp->cpu, tbp->len);
+				add_trace_buf(iop->dpp, tp->cpu, &tbp);
+				nentries++;
+			} else if (tbp->len == 0) {
+				/*
+				 * Short reads after we're done stop us
+				 * from trying reads.
+				 */
+				if (tp->is_done)
+					clear_events(pfd);
+			} else {
+				read_err(tp->cpu, iop->ifn);
+				if (errno != EAGAIN || tp->is_done)
+					clear_events(pfd);
+			}
+			if (!piped_output && --nevs == 0)
+				break;
 		}
 	}
+	free(tbp);
 
-	if (i != ndevs) {
-		__for_each_dip(dip, j, i)
-			stop_threads(dip);
-		for_each_dip(dip, i)
-			stop_trace(dip);
-
-		return 1;
-	}
+	if (nentries)
+		incr_entries(nentries);
 
-	return 0;
+	return nentries;
 }
 
-static void show_stats(void)
+static void *thread_main(void *arg)
 {
-	struct device_information *dip;
-	struct thread_information *tip;
-	unsigned long long events_processed, data_read;
-	unsigned long total_drops;
-	int i, j, no_stdout = 0;
-
-	if (is_stat_shown())
-		return;
-
-	if (output_name && !strcmp(output_name, "-"))
-		no_stdout = 1;
+	int ret, ndone, to_val;
+	struct tracer *tp = arg;
 
-	stat_shown = 1;
+	ret = lock_on_cpu(tp->cpu);
+	if (ret)
+		goto err;
 
-	total_drops = 0;
-	for_each_dip(dip, i) {
-		if (!no_stdout)
-			printf("Device: %s\n", dip->path);
-		events_processed = 0;
-		data_read = 0;
-		for_each_tip(dip, tip, j) {
-			if (!no_stdout)
-				printf("  CPU%3d: %20lu events, %8llu KiB data\n",
-			       		tip->cpu, tip->events_processed,
-					(tip->data_read + 1023) >> 10);
-			events_processed += tip->events_processed;
-			data_read += tip->data_read;
-		}
-		total_drops += dip->drop_count;
-		if (!no_stdout)
-			printf("  Total:  %20llu events (dropped %lu), %8llu KiB data\n",
-					events_processed, dip->drop_count,
-					(data_read + 1023) >> 10);
-	}
+	ret = open_ios(tp);
+	if (ret)
+		goto err;
 
-	if (total_drops)
-		fprintf(stderr, "You have dropped events, consider using a larger buffer size (-b)\n");
-}
+	if (piped_output)
+		to_val = 50;		/* Frequent partial handles */
+	else
+		to_val = 500;		/* 1/2 second intervals */
 
-static struct device_information *net_get_dip(char *buts_name,
-					      struct in_addr *cl_in_addr)
-{
-	struct device_information *dip;
-	int i;
 
-	for (i = 0; i < ndevs; i++) {
-		dip = &device_information[i];
+	tracer_signal_ready(tp, Th_running, 0);
+	tracer_wait_unblock(tp);
 
-		if (!strcmp(dip->buts_name, buts_name))
-			return dip;
+	while (!tp->is_done) {
+		ndone = poll(tp->pfds, ndevs, to_val);
+		if (ndone || piped_output)
+			(void)handle_pfds(tp, ndone, piped_output);
+		else if (ndone < 0 && errno != EINTR)
+			fprintf(stderr, "Thread %d poll failed: %d/%s\n",
+				tp->cpu, errno, strerror(errno));
 	}
 
-	device_information = realloc(device_information, (ndevs + 1) * sizeof(*dip));
-	dip = &device_information[ndevs];
-	memset(dip, 0, sizeof(*dip));
-	dip->fd = -1;
-	strcpy(dip->buts_name, buts_name);
-	dip->path = strdup(buts_name);
-	ndevs++;
-	dip->threads = malloc(ncpus * sizeof(struct thread_information));
-	memset(dip->threads, 0, ncpus * sizeof(struct thread_information));
-
 	/*
-	 * open all files
+	 * Trace is stopped, pull data until we get a short read
 	 */
-	for (i = 0; i < ncpus; i++) {
-		struct thread_information *tip = &dip->threads[i];
-
-		tip->cpu = i;
-		tip->device = dip;
-		tip->fd = -1;
-		tip->pfd = -1;
-		tip->cl_in_addr = *cl_in_addr;
+	while (handle_pfds(tp, ndevs, 1) > 0)
+		;
 
-		if (tip_open_output(dip, tip))
-			return NULL;
-	}
+	close_ios(tp);
+	tracer_signal_ready(tp, Th_leaving, 0);
+	return NULL;
 
-	return dip;
+err:
+	tracer_signal_ready(tp, Th_error, ret);
+	return NULL;
 }
 
-static struct thread_information *net_get_tip(struct blktrace_net_hdr *bnh,
-					      struct in_addr *cl_in_addr)
+static int start_tracer(int cpu)
 {
-	struct device_information *dip;
+	struct tracer *tp;
 
-	ncpus = bnh->max_cpus;
-	dip = net_get_dip(bnh->buts_name, cl_in_addr);
-	return &dip->threads[bnh->cpu];
+	tp = malloc(sizeof(*tp));
+	memset(tp, 0, sizeof(*tp));
+
+	INIT_LIST_HEAD(&tp->head);
+	tp->status = 0;
+	tp->cpu = cpu;
+
+	if (pthread_create(&tp->thread, NULL, thread_main, tp)) {
+		fprintf(stderr, "FAILED to start thread on CPU %d: %d/%s\n",
+			cpu, errno, strerror(errno));
+		free(tp);
+		return 1;
+	}
+
+	list_add_tail(&tp->head, &tracers);
+	return 0;
 }
 
-static int net_get_header(struct blktrace_net_hdr *bnh)
+static int create_output_files(int cpu)
 {
-	int fl = fcntl(net_in_fd, F_GETFL);
-	int bytes_left, ret;
-	void *p = bnh;
+	char fname[MAXPATHLEN + 64];
+	struct list_head *p;
+	FILE *f;
 
-	fcntl(net_in_fd, F_SETFL, fl | O_NONBLOCK);
-	bytes_left = sizeof(*bnh);
-	while (bytes_left && !is_done()) {
-		ret = recv(net_in_fd, p, bytes_left, MSG_WAITALL);
-		if (ret < 0) {
-			if (errno != EAGAIN) {
-				perror("recv header");
-				return 1;
-			}
-			usleep(100);
-			continue;
-		} else if (!ret) {
-			usleep(100);
-			continue;
-		} else {
-			p += ret;
-			bytes_left -= ret;
-		}
+	__list_for_each(p, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
+
+		if (fill_ofname(fname, sizeof(fname), NULL, dpp->buts_name,
+				cpu))
+			return 1;
+		f = my_fopen(fname, "w+");
+		if (!f)
+			return 1;
+		fclose(f);
 	}
-	fcntl(net_in_fd, F_SETFL, fl & ~O_NONBLOCK);
-	return bytes_left;
+	return 0;
 }
 
-static int net_server_loop(struct in_addr *cl_in_addr)
+static void start_tracers(void)
 {
-	struct thread_information *tip;
-	struct blktrace_net_hdr bnh;
+	int cpu, started = 0;
+	struct list_head *p;
+	size_t alloc_size = CPU_ALLOC_SIZE(max_cpus);
+
+	for (cpu = 0; cpu < max_cpus; cpu++) {
+		if (!CPU_ISSET_S(cpu, alloc_size, online_cpus)) {
+			/*
+			 * Create fake empty output files so that other tools
+			 * like blkparse don't have to bother with sparse CPU
+			 * number space.
+			 */
+			if (create_output_files(cpu))
+				break;
+			continue;
+		}
+		if (start_tracer(cpu))
+			break;
+		started++;
+	}
 
-	if (net_get_header(&bnh))
-		return 1;
+	wait_tracers_ready(started);
 
-	if (data_is_native == -1 && check_data_endianness(bnh.magic)) {
-		fprintf(stderr, "server: received data is bad\n");
-		return 1;
+	__list_for_each(p, &tracers) {
+		struct tracer *tp = list_entry(p, struct tracer, head);
+		if (tp->status)
+			fprintf(stderr,
+				"FAILED to start thread on CPU %d: %d/%s\n",
+				tp->cpu, tp->status, strerror(tp->status));
 	}
+}
 
-	if (!data_is_native) {
-		bnh.magic = be32_to_cpu(bnh.magic);
-		bnh.cpu = be32_to_cpu(bnh.cpu);
-		bnh.len = be32_to_cpu(bnh.len);
-	}
+static void stop_tracers(void)
+{
+	struct list_head *p;
 
-	if ((bnh.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
-		fprintf(stderr, "server: bad data magic\n");
-		return 1;
+	/*
+	 * Stop the tracing - makes the tracer threads clean up quicker.
+	 */
+	__list_for_each(p, &devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
+		(void)ioctl(dpp->fd, BLKTRACESTOP);
 	}
 
 	/*
-	 * len == 0 means that the other end signalled end-of-run
+	 * Tell each tracer to quit
 	 */
-	if (!bnh.len) {
-		fprintf(stderr, "server: end of run\n");
-		return 1;
+	__list_for_each(p, &tracers) {
+		struct tracer *tp = list_entry(p, struct tracer, head);
+		tp->is_done = 1;
 	}
+	pthread_cond_broadcast(&mt_cond);
+}
 
-	tip = net_get_tip(&bnh, cl_in_addr);
-	if (!tip)
-		return 1;
+static void del_tracers(void)
+{
+	struct list_head *p, *q;
 
-	if (mmap_subbuf(tip, bnh.len))
-		return 1;
+	list_for_each_safe(p, q, &tracers) {
+		struct tracer *tp = list_entry(p, struct tracer, head);
 
-	return 0;
+		list_del(&tp->head);
+		free(tp);
+	}
 }
 
-/*
- * Start here when we are in server mode - just fetch data from the network
- * and dump to files
- */
-static int net_server(void)
+static void wait_tracers(void)
 {
-	struct device_information *dip;
-	struct thread_information *tip;
-	struct sockaddr_in addr;
-	socklen_t socklen;
-	int fd, opt, i, j;
-
-	fd = socket(AF_INET, SOCK_STREAM, 0);
-	if (fd < 0) {
-		perror("server: socket");
-		return 1;
-	}
+	struct list_head *p;
 
-	opt = 1;
-	if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
-		perror("setsockopt");
-		return 1;
-	}
+	if (use_tracer_devpaths())
+		process_trace_bufs();
 
-	memset(&addr, 0, sizeof(addr));
-	addr.sin_family = AF_INET;
-	addr.sin_addr.s_addr = htonl(INADDR_ANY);
-	addr.sin_port = htons(net_port);
+	wait_tracers_leaving();
 
-	if (bind(fd, (struct sockaddr *) &addr, sizeof(addr)) < 0) {
-		perror("bind");
-		return 1;
-	}
+	__list_for_each(p, &tracers) {
+		int ret;
+		struct tracer *tp = list_entry(p, struct tracer, head);
 
-	if (listen(fd, 1) < 0) {
-		perror("listen");
-		return 1;
+		ret = pthread_join(tp->thread, NULL);
+		if (ret)
+			fprintf(stderr, "Thread join %d failed %d\n",
+				tp->cpu, ret);
 	}
 
-repeat:
-	signal(SIGINT, NULL);
-	signal(SIGHUP, NULL);
-	signal(SIGTERM, NULL);
-	signal(SIGALRM, NULL);
+	if (use_tracer_devpaths())
+		clean_trace_bufs();
 
-	printf("blktrace: waiting for incoming connection...\n");
+	get_all_drops();
+}
 
-	socklen = sizeof(addr);
-	net_in_fd = accept(fd, (struct sockaddr *) &addr, &socklen);
-	if (net_in_fd < 0) {
-		perror("accept");
-		return 1;
-	}
+static void exit_tracing(void)
+{
+	signal(SIGINT, SIG_IGN);
+	signal(SIGHUP, SIG_IGN);
+	signal(SIGTERM, SIG_IGN);
+	signal(SIGALRM, SIG_IGN);
+
+	stop_tracers();
+	wait_tracers();
+	del_tracers();
+	rel_devpaths();
+}
 
-	signal(SIGINT, handle_sigint);
-	signal(SIGHUP, handle_sigint);
-	signal(SIGTERM, handle_sigint);
-	signal(SIGALRM, handle_sigint);
-
-	printf("blktrace: connection from %s\n", inet_ntoa(addr.sin_addr));
-
-	while (!is_done()) {
-		if (net_server_loop(&addr.sin_addr))
-			break;
-	}
-
-	for_each_dip(dip, i)
-		for_each_tip(dip, tip, j)
-			tip_ftrunc_final(tip);
+static void handle_sigint(__attribute__((__unused__)) int sig)
+{
+	done = 1;
+	stop_tracers();
+}
 
-	show_stats();
+static void show_stats(struct list_head *devpaths)
+{
+	FILE *ofp;
+	struct list_head *p;
+	unsigned long long nevents, data_read;
+	unsigned long long total_drops = 0;
+	unsigned long long total_events = 0;
+
+	if (piped_output)
+		ofp = my_fopen("/dev/null", "w");
+	else
+		ofp = stdout;
 
-	if (is_done())
-		return 0;
+	__list_for_each(p, devpaths) {
+		int cpu;
+		struct pdc_stats *sp;
+		struct devpath *dpp = list_entry(p, struct devpath, head);
 
-	/*
-	 * cleanup for next run
-	 */
-	for_each_dip(dip, i) {
-		for_each_tip(dip, tip, j)
-			fclose(tip->ofile);
+		if (net_mode == Net_server)
+			printf("server: end of run for %s:%s\n",
+				dpp->ch->hostname, dpp->buts_name);
 
-		free(dip->threads);
-		free(dip->path);
-	}
+		data_read = 0;
+		nevents = 0;
+
+		fprintf(ofp, "=== %s ===\n", dpp->buts_name);
+		for (cpu = 0, sp = dpp->stats; cpu < dpp->ncpus; cpu++, sp++) {
+			/*
+			 * Estimate events if not known...
+			 */
+			if (sp->nevents == 0) {
+				sp->nevents = sp->data_read /
+						sizeof(struct blk_io_trace);
+			}
 
-	free(device_information);
-	device_information = NULL;
-	ncpus = ndevs = 0;
+			fprintf(ofp,
+				"  CPU%3d: %20llu events, %8llu KiB data\n",
+				cpu, sp->nevents, (sp->data_read + 1023) >> 10);
 
-	close(net_in_fd);
-	net_in_fd = -1;
-	stat_shown = 0;
-	goto repeat;
-}
+			data_read += sp->data_read;
+			nevents += sp->nevents;
+		}
 
-/*
- * Setup outgoing network connection where we will transmit data
- */
-static int net_setup_client(void)
-{
-	struct sockaddr_in addr;
-	int fd;
+		fprintf(ofp, "  Total:  %20llu events (dropped %llu),"
+			     " %8llu KiB data\n", nevents,
+			     dpp->drops, (data_read + 1024) >> 10);
 
-	fd = socket(AF_INET, SOCK_STREAM, 0);
-	if (fd < 0) {
-		perror("client: socket");
-		return 1;
+		total_drops += dpp->drops;
+		total_events += (nevents + dpp->drops);
 	}
 
-	memset(&addr, 0, sizeof(addr));
-	addr.sin_family = AF_INET;
-	addr.sin_port = htons(net_port);
-
-	if (inet_aton(hostname, &addr.sin_addr) != 1) {
-		struct hostent *hent = gethostbyname(hostname);
-		if (!hent) {
-			perror("gethostbyname");
-			return 1;
-		}
+	fflush(ofp);
+	if (piped_output)
+		fclose(ofp);
 
-		memcpy(&addr.sin_addr, hent->h_addr, 4);
-		strcpy(hostname, hent->h_name);
-	}
+	if (total_drops) {
+		double drops_ratio = 1.0;
 
-	printf("blktrace: connecting to %s\n", hostname);
+		if (total_events)
+			drops_ratio = (double)total_drops/(double)total_events;
 
-	if (connect(fd, (struct sockaddr *) &addr, sizeof(addr)) < 0) {
-		perror("client: connect");
-		return 1;
+		fprintf(stderr, "\nYou have %llu (%5.1lf%%) dropped events\n"
+				"Consider using a larger buffer size (-b) "
+				"and/or more buffers (-n)\n",
+			total_drops, 100.0 * drops_ratio);
 	}
-
-	printf("blktrace: connected!\n");
-	net_out_fd = fd;
-	return 0;
 }
 
-static char usage_str[] = \
-	"-d <dev> [ -r relay path ] [ -o <output> ] [-k ] [ -w time ]\n" \
-	"[ -a action ] [ -A action mask ] [ -v ]\n\n" \
-	"\t-d Use specified device. May also be given last after options\n" \
-	"\t-r Path to mounted relayfs, defaults to /relay\n" \
-	"\t-o File(s) to send output to\n" \
-	"\t-D Directory to prepend to output file names\n" \
-	"\t-k Kill a running trace\n" \
-	"\t-w Stop after defined time, in seconds\n" \
-	"\t-a Only trace specified actions. See documentation\n" \
-	"\t-A Give trace mask as a single value. See documentation\n" \
-	"\t-b Sub buffer size in KiB\n" \
-	"\t-n Number of sub buffers\n" \
-	"\t-l Run in network listen mode (blktrace server)\n" \
-	"\t-h Run in network client mode, connecting to the given host\n" \
-	"\t-p Network port to use (default 8462)\n" \
-	"\t-s Make the network client use sendfile() to transfer data\n" \
-	"\t-V Print program version info\n\n";
-
-static void show_usage(char *program)
+static int handle_args(int argc, char *argv[])
 {
-	fprintf(stderr, "Usage: %s %s %s",program, blktrace_version, usage_str);
-}
-
-int main(int argc, char *argv[])
-{
-	static char default_relay_path[] = "/relay";
+	int c, i;
 	struct statfs st;
-	int i, c;
-	int stop_watch = 0;
 	int act_mask_tmp = 0;
 
 	while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) >= 0) {
@@ -1586,7 +2111,7 @@ int main(int argc, char *argv[])
 		case 'a':
 			i = find_mask_map(optarg);
 			if (i < 0) {
-				fprintf(stderr,"Invalid action mask %s\n",
+				fprintf(stderr, "Invalid action mask %s\n",
 					optarg);
 				return 1;
 			}
@@ -1594,7 +2119,7 @@ int main(int argc, char *argv[])
 			break;
 
 		case 'A':
-			if ((sscanf(optarg, "%x", &i) != 1) || 
+			if ((sscanf(optarg, "%x", &i) != 1) ||
 							!valid_act_opt(i)) {
 				fprintf(stderr,
 					"Invalid set action mask %s/0x%x\n",
@@ -1605,12 +2130,33 @@ int main(int argc, char *argv[])
 			break;
 
 		case 'd':
-			if (resize_devices(optarg) != 0)
+			if (add_devpath(optarg) != 0)
 				return 1;
 			break;
 
+		case 'I': {
+			char dev_line[256];
+			FILE *ifp = my_fopen(optarg, "r");
+
+			if (!ifp) {
+				fprintf(stderr,
+					"Invalid file for devices %s\n",
+					optarg);
+				return 1;
+			}
+
+			while (fscanf(ifp, "%s\n", dev_line) == 1) {
+				if (add_devpath(dev_line) != 0) {
+					fclose(ifp);
+					return 1;
+				}
+			}
+			fclose(ifp);
+			break;
+		}
+
 		case 'r':
-			relay_path = optarg;
+			debugfs_path = optarg;
 			break;
 
 		case 'o':
@@ -1629,13 +2175,15 @@ int main(int argc, char *argv[])
 			}
 			break;
 		case 'V':
+		case 'v':
 			printf("%s version %s\n", argv[0], blktrace_version);
-			return 0;
+			exit(0);
+			/*NOTREACHED*/
 		case 'b':
 			buf_size = strtoul(optarg, NULL, 10);
 			if (buf_size <= 0 || buf_size > 16*1024) {
-				fprintf(stderr,
-					"Invalid buffer size (%lu)\n",buf_size);
+				fprintf(stderr, "Invalid buffer size (%lu)\n",
+					buf_size);
 				return 1;
 			}
 			buf_size <<= 10;
@@ -1653,7 +2201,9 @@ int main(int argc, char *argv[])
 			break;
 		case 'h':
 			net_mode = Net_client;
-			strcpy(hostname, optarg);
+			memset(hostname, 0, sizeof(hostname));
+			strncpy(hostname, optarg, sizeof(hostname));
+			hostname[sizeof(hostname) - 1] = '\0';
 			break;
 		case 'l':
 			net_mode = Net_server;
@@ -1662,88 +2212,634 @@ int main(int argc, char *argv[])
 			net_port = atoi(optarg);
 			break;
 		case 's':
-			net_use_sendfile = 1;
+			net_use_sendfile = 0;
 			break;
 		default:
 			show_usage(argv[0]);
-			return 1;
+			exit(1);
+			/*NOTREACHED*/
 		}
 	}
 
-	setlocale(LC_NUMERIC, "en_US");
-
-	page_size = getpagesize();
-
-	if (net_mode == Net_server)
-		return net_server();
-
-	while (optind < argc) {
-		if (resize_devices(argv[optind++]) != 0)
+	while (optind < argc)
+		if (add_devpath(argv[optind++]) != 0)
 			return 1;
-	}
 
-	if (ndevs == 0) {
+	if (net_mode != Net_server && ndevs == 0) {
 		show_usage(argv[0]);
 		return 1;
 	}
 
-	if (!relay_path)
-		relay_path = default_relay_path;
+	if (statfs(debugfs_path, &st) < 0) {
+		fprintf(stderr, "Invalid debug path %s: %d/%s\n",
+			debugfs_path, errno, strerror(errno));
+		return 1;
+	}
+
+	if (st.f_type != (long)DEBUGFS_TYPE) {
+		fprintf(stderr, "Debugfs is not mounted at %s\n", debugfs_path);
+		return 1;
+	}
 
 	if (act_mask_tmp != 0)
 		act_mask = act_mask_tmp;
 
-	if (statfs(relay_path, &st) < 0) {
-		perror("statfs");
-		fprintf(stderr,"%s does not appear to be a valid path\n",
-			relay_path);
-		return 1;
-	} else if (st.f_type != (long) RELAYFS_TYPE) {
-		fprintf(stderr,"%s does not appear to be a relay filesystem\n",
-			relay_path);
+	if (net_mode == Net_client && net_setup_addr())
 		return 1;
+
+	/*
+	 * Set up for appropriate PFD handler based upon output name.
+	 */
+	if (net_client_use_sendfile())
+		handle_pfds = handle_pfds_netclient;
+	else if (net_client_use_send())
+		handle_pfds = handle_pfds_entries;
+	else if (output_name && (strcmp(output_name, "-") == 0)) {
+		piped_output = 1;
+		handle_pfds = handle_pfds_entries;
+		pfp = stdout;
+		if (setvbuf(pfp, NULL, _IONBF, 0)) {
+			perror("setvbuf stdout");
+			return 1;
+		}
+	} else
+		handle_pfds = handle_pfds_file;
+	return 0;
+}
+
+static void ch_add_connection(struct net_server_s *ns, struct cl_host *ch,
+			      int fd)
+{
+	struct cl_conn *nc;
+
+	nc = malloc(sizeof(*nc));
+	memset(nc, 0, sizeof(*nc));
+
+	time(&nc->connect_time);
+	nc->ch = ch;
+	nc->fd = fd;
+	nc->ncpus = -1;
+
+	list_add_tail(&nc->ch_head, &ch->conn_list);
+	ch->connects++;
+
+	list_add_tail(&nc->ns_head, &ns->conn_list);
+	ns->connects++;
+	ns->pfds = realloc(ns->pfds, (ns->connects+1) * sizeof(struct pollfd));
+}
+
+static void ch_rem_connection(struct net_server_s *ns, struct cl_host *ch,
+			      struct cl_conn *nc)
+{
+	net_close_connection(&nc->fd);
+
+	list_del(&nc->ch_head);
+	ch->connects--;
+
+	list_del(&nc->ns_head);
+	ns->connects--;
+	ns->pfds = realloc(ns->pfds, (ns->connects+1) * sizeof(struct pollfd));
+
+	free(nc);
+}
+
+static struct cl_host *net_find_client_host(struct net_server_s *ns,
+					    struct in_addr cl_in_addr)
+{
+	struct list_head *p;
+
+	__list_for_each(p, &ns->ch_list) {
+		struct cl_host *ch = list_entry(p, struct cl_host, head);
+
+		if (in_addr_eq(ch->cl_in_addr, cl_in_addr))
+			return ch;
 	}
 
-	if (open_devices() != 0)
-		return 1;
+	return NULL;
+}
 
-	if (kill_running_trace) {
-		stop_all_traces();
+static struct cl_host *net_add_client_host(struct net_server_s *ns,
+					   struct sockaddr_in *addr)
+{
+	struct cl_host *ch;
+
+	ch = malloc(sizeof(*ch));
+	memset(ch, 0, sizeof(*ch));
+
+	ch->ns = ns;
+	ch->cl_in_addr = addr->sin_addr;
+	list_add_tail(&ch->head, &ns->ch_list);
+	ns->nchs++;
+
+	ch->hostname = strdup(inet_ntoa(addr->sin_addr));
+	printf("server: connection from %s\n", ch->hostname);
+
+	INIT_LIST_HEAD(&ch->conn_list);
+	INIT_LIST_HEAD(&ch->devpaths);
+
+	return ch;
+}
+
+static void device_done(struct devpath *dpp, int ncpus)
+{
+	int cpu;
+	struct io_info *iop;
+
+	for (cpu = 0, iop = dpp->ios; cpu < ncpus; cpu++, iop++)
+		close_iop(iop);
+
+	list_del(&dpp->head);
+	dpp_free(dpp);
+}
+
+static void net_ch_remove(struct cl_host *ch, int ncpus)
+{
+	struct list_head *p, *q;
+	struct net_server_s *ns = ch->ns;
+
+	list_for_each_safe(p, q, &ch->devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
+		device_done(dpp, ncpus);
+	}
+
+	list_for_each_safe(p, q, &ch->conn_list) {
+		struct cl_conn *nc = list_entry(p, struct cl_conn, ch_head);
+
+		ch_rem_connection(ns, ch, nc);
+	}
+
+	list_del(&ch->head);
+	ns->nchs--;
+
+	if (ch->hostname)
+		free(ch->hostname);
+	free(ch);
+}
+
+static void net_add_connection(struct net_server_s *ns)
+{
+	int fd;
+	struct cl_host *ch;
+	socklen_t socklen = sizeof(ns->addr);
+
+	fd = my_accept(ns->listen_fd, (struct sockaddr *)&ns->addr, &socklen);
+	if (fd < 0) {
+		/*
+		 * This is OK: we just won't accept this connection,
+		 * nothing fatal.
+		 */
+		perror("accept");
+	} else {
+		ch = net_find_client_host(ns, ns->addr.sin_addr);
+		if (!ch)
+			ch = net_add_client_host(ns, &ns->addr);
+
+		ch_add_connection(ns, ch, fd);
+	}
+}
+
+static struct devpath *nc_add_dpp(struct cl_conn *nc,
+				  struct blktrace_net_hdr *bnh,
+				  time_t connect_time)
+{
+	int cpu;
+	struct io_info *iop;
+	struct devpath *dpp;
+
+	dpp = malloc(sizeof(*dpp));
+	memset(dpp, 0, sizeof(*dpp));
+
+	dpp->buts_name = strdup(bnh->buts_name);
+	dpp->path = strdup(bnh->buts_name);
+	dpp->fd = -1;
+	dpp->ch = nc->ch;
+	dpp->cl_id = bnh->cl_id;
+	dpp->cl_connect_time = connect_time;
+	dpp->ncpus = nc->ncpus;
+	dpp->stats = calloc(dpp->ncpus, sizeof(*dpp->stats));
+	memset(dpp->stats, 0, dpp->ncpus * sizeof(*dpp->stats));
+
+	list_add_tail(&dpp->head, &nc->ch->devpaths);
+	nc->ch->ndevs++;
+
+	dpp->ios = calloc(nc->ncpus, sizeof(*iop));
+	memset(dpp->ios, 0, ndevs * sizeof(*iop));
+
+	for (cpu = 0, iop = dpp->ios; cpu < nc->ncpus; cpu++, iop++) {
+		iop->dpp = dpp;
+		iop->nc = nc;
+		init_mmap_info(&iop->mmap_info);
+
+		if (iop_open(iop, cpu))
+			goto err;
+	}
+
+	return dpp;
+
+err:
+	/*
+	 * Need to unravel what's been done...
+	 */
+	while (cpu >= 0)
+		close_iop(&dpp->ios[cpu--]);
+	dpp_free(dpp);
+
+	return NULL;
+}
+
+static struct devpath *nc_find_dpp(struct cl_conn *nc,
+				   struct blktrace_net_hdr *bnh)
+{
+	struct list_head *p;
+	time_t connect_time = nc->connect_time;
+
+	__list_for_each(p, &nc->ch->devpaths) {
+		struct devpath *dpp = list_entry(p, struct devpath, head);
+
+		if (!strcmp(dpp->buts_name, bnh->buts_name))
+			return dpp;
+
+		if (dpp->cl_id == bnh->cl_id)
+			connect_time = dpp->cl_connect_time;
+	}
+
+	return nc_add_dpp(nc, bnh, connect_time);
+}
+
+static void net_client_read_data(struct cl_conn *nc, struct devpath *dpp,
+				 struct blktrace_net_hdr *bnh)
+{
+	int ret;
+	struct io_info *iop = &dpp->ios[bnh->cpu];
+	struct mmap_info *mip = &iop->mmap_info;
+
+	if (setup_mmap(iop->ofd, bnh->len, &iop->mmap_info, NULL)) {
+		fprintf(stderr, "ncd(%s:%d): mmap failed\n",
+			nc->ch->hostname, nc->fd);
+		exit(1);
+	}
+
+	ret = net_recv_data(nc->fd, mip->fs_buf + mip->fs_off, bnh->len);
+	if (ret > 0) {
+		pdc_dr_update(dpp, bnh->cpu, ret);
+		mip->fs_size += ret;
+		mip->fs_off += ret;
+	} else if (ret < 0)
+		exit(1);
+}
+
+/*
+ * Returns 1 if we closed a host - invalidates other polling information
+ * that may be present.
+ */
+static int net_client_data(struct cl_conn *nc)
+{
+	int ret;
+	struct devpath *dpp;
+	struct blktrace_net_hdr bnh;
+
+	ret = net_get_header(nc, &bnh);
+	if (ret == 0)
 		return 0;
+
+	if (ret < 0) {
+		fprintf(stderr, "ncd(%d): header read failed\n", nc->fd);
+		exit(1);
 	}
 
-	ncpus = sysconf(_SC_NPROCESSORS_ONLN);
-	if (ncpus < 0) {
-		fprintf(stderr, "sysconf(_SC_NPROCESSORS_ONLN) failed\n");
-		return 1;
+	if (data_is_native == -1 && check_data_endianness(bnh.magic)) {
+		fprintf(stderr, "ncd(%d): received data is bad\n", nc->fd);
+		exit(1);
 	}
 
-	signal(SIGINT, handle_sigint);
-	signal(SIGHUP, handle_sigint);
-	signal(SIGTERM, handle_sigint);
-	signal(SIGALRM, handle_sigint);
+	if (!data_is_native) {
+		bnh.magic = be32_to_cpu(bnh.magic);
+		bnh.cpu = be32_to_cpu(bnh.cpu);
+		bnh.max_cpus = be32_to_cpu(bnh.max_cpus);
+		bnh.len = be32_to_cpu(bnh.len);
+		bnh.cl_id = be32_to_cpu(bnh.cl_id);
+		bnh.buf_size = be32_to_cpu(bnh.buf_size);
+		bnh.buf_nr = be32_to_cpu(bnh.buf_nr);
+		bnh.page_size = be32_to_cpu(bnh.page_size);
+	}
 
-	if (net_mode == Net_client && net_setup_client())
-		return 1;
+	if ((bnh.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
+		fprintf(stderr, "ncd(%s:%d): bad data magic\n",
+			nc->ch->hostname, nc->fd);
+		exit(1);
+	}
 
-	if (start_devices() != 0)
-		return 1;
+	if (nc->ncpus == -1)
+		nc->ncpus = bnh.max_cpus;
+
+	/*
+	 * len == 0 means the other end is sending us a new connection/dpp
+	 * len == 1 means that the other end signalled end-of-run
+	 */
+	dpp = nc_find_dpp(nc, &bnh);
+	if (bnh.len == 0) {
+		/*
+		 * Just adding in the dpp above is enough
+		 */
+		ack_open_close(nc->fd, dpp->buts_name);
+		nc->ch->cl_opens++;
+	} else if (bnh.len == 1) {
+		/*
+		 * overload cpu count with dropped events
+		 */
+		dpp->drops = bnh.cpu;
+
+		ack_open_close(nc->fd, dpp->buts_name);
+		if (--nc->ch->cl_opens == 0) {
+			show_stats(&nc->ch->devpaths);
+			net_ch_remove(nc->ch, nc->ncpus);
+			return 1;
+		}
+	} else
+		net_client_read_data(nc, dpp, &bnh);
 
-	atexit(stop_all_tracing);
+	return 0;
+}
 
-	if (stop_watch)
-		alarm(stop_watch);
+static void handle_client_data(struct net_server_s *ns, int events)
+{
+	struct cl_conn *nc;
+	struct pollfd *pfd;
+	struct list_head *p, *q;
 
-	wait_for_threads();
+	pfd = &ns->pfds[1];
+	list_for_each_safe(p, q, &ns->conn_list) {
+		if (pfd->revents & POLLIN) {
+			nc = list_entry(p, struct cl_conn, ns_head);
 
-	if (!is_trace_stopped()) {
-		trace_stopped = 1;
-		stop_all_threads();
-		stop_all_traces();
+			if (net_client_data(nc) || --events == 0)
+				break;
+		}
+		pfd++;
 	}
+}
+
+static void net_setup_pfds(struct net_server_s *ns)
+{
+	struct pollfd *pfd;
+	struct list_head *p;
+
+	ns->pfds[0].fd = ns->listen_fd;
+	ns->pfds[0].events = POLLIN;
 
-	show_stats();
+	pfd = &ns->pfds[1];
+	__list_for_each(p, &ns->conn_list) {
+		struct cl_conn *nc = list_entry(p, struct cl_conn, ns_head);
+
+		pfd->fd = nc->fd;
+		pfd->events = POLLIN;
+		pfd++;
+	}
+}
+
+static int net_server_handle_connections(struct net_server_s *ns)
+{
+	int events;
+
+	printf("server: waiting for connections...\n");
+
+	while (!done) {
+		net_setup_pfds(ns);
+		events = poll(ns->pfds, ns->connects + 1, -1);
+		if (events < 0) {
+			if (errno != EINTR) {
+				perror("FATAL: poll error");
+				return 1;
+			}
+		} else if (events > 0) {
+			if (ns->pfds[0].revents & POLLIN) {
+				net_add_connection(ns);
+				events--;
+			}
+
+			if (events)
+				handle_client_data(ns, events);
+		}
+	}
 
 	return 0;
 }
 
+static int net_server(void)
+{
+	int fd, opt;
+	int ret = 1;
+	struct net_server_s net_server;
+	struct net_server_s *ns = &net_server;
+
+	memset(ns, 0, sizeof(*ns));
+	INIT_LIST_HEAD(&ns->ch_list);
+	INIT_LIST_HEAD(&ns->conn_list);
+	ns->pfds = malloc(sizeof(struct pollfd));
+
+	fd = my_socket(AF_INET, SOCK_STREAM, 0);
+	if (fd < 0) {
+		perror("server: socket");
+		goto out;
+	}
+
+	opt = 1;
+	if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
+		perror("setsockopt");
+		goto out;
+	}
+
+	memset(&ns->addr, 0, sizeof(ns->addr));
+	ns->addr.sin_family = AF_INET;
+	ns->addr.sin_addr.s_addr = htonl(INADDR_ANY);
+	ns->addr.sin_port = htons(net_port);
+
+	if (bind(fd, (struct sockaddr *) &ns->addr, sizeof(ns->addr)) < 0) {
+		perror("bind");
+		goto out;
+	}
+
+	if (listen(fd, 1) < 0) {
+		perror("listen");
+		goto out;
+	}
+
+	/*
+	 * The actual server looping is done here:
+	 */
+	ns->listen_fd = fd;
+	ret = net_server_handle_connections(ns);
+
+	/*
+	 * Clean up and return...
+	 */
+out:
+	free(ns->pfds);
+	return ret;
+}
+
+static int run_tracers(void)
+{
+	atexit(exit_tracing);
+	if (net_mode == Net_client)
+		printf("blktrace: connecting to %s\n", hostname);
+
+	if (setup_buts())
+		return 1;
+
+	if (use_tracer_devpaths()) {
+		if (setup_tracer_devpaths())
+			return 1;
+
+		if (piped_output)
+			handle_list = handle_list_file;
+		else
+			handle_list = handle_list_net;
+	}
+
+	start_tracers();
+	if (nthreads_running == ncpus) {
+		unblock_tracers();
+		start_buts();
+		if (net_mode == Net_client)
+			printf("blktrace: connected!\n");
+		if (stop_watch)
+			alarm(stop_watch);
+	} else
+		stop_tracers();
+
+	wait_tracers();
+	if (nthreads_running == ncpus)
+		show_stats(&devpaths);
+	if (net_client_use_send())
+		close_client_connections();
+	del_tracers();
+
+	return 0;
+}
+
+static cpu_set_t *get_online_cpus(void)
+{
+	FILE *cpus;
+	cpu_set_t *set;
+	size_t alloc_size;
+	int cpuid, prevcpuid = -1;
+	char nextch;
+	int n, ncpu, curcpu = 0;
+	int *cpu_nums;
+
+	ncpu = sysconf(_SC_NPROCESSORS_CONF);
+	if (ncpu < 0)
+		return NULL;
+
+	cpu_nums = malloc(sizeof(int)*ncpu);
+	if (!cpu_nums) {
+		errno = ENOMEM;
+		return NULL;
+	}
+
+	/*
+	 * There is no way to easily get maximum CPU number. So we have to
+	 * parse the file first to find it out and then create appropriate
+	 * cpuset
+	 */
+	cpus = my_fopen("/sys/devices/system/cpu/online", "r");
+	for (;;) {
+		n = fscanf(cpus, "%d%c", &cpuid, &nextch);
+		if (n <= 0)
+			break;
+		if (n == 2 && nextch == '-') {
+			prevcpuid = cpuid;
+			continue;
+		}
+		if (prevcpuid == -1)
+			prevcpuid = cpuid;
+		while (prevcpuid <= cpuid) {
+			/* More CPUs listed than configured? */
+			if (curcpu >= ncpu) {
+				errno = EINVAL;
+				return NULL;
+			}
+			cpu_nums[curcpu++] = prevcpuid++;
+		}
+		prevcpuid = -1;
+	}
+	fclose(cpus);
+
+	ncpu = curcpu;
+	max_cpus = cpu_nums[ncpu - 1] + 1;
+
+	/* Now that we have maximum cpu number, create a cpuset */
+	set = CPU_ALLOC(max_cpus);
+	if (!set) {
+		errno = ENOMEM;
+		return NULL;
+	}
+	alloc_size = CPU_ALLOC_SIZE(max_cpus);
+	CPU_ZERO_S(alloc_size, set);
+
+	for (curcpu = 0; curcpu < ncpu; curcpu++)
+		CPU_SET_S(cpu_nums[curcpu], alloc_size, set);
+
+	free(cpu_nums);
+
+	return set;
+}
+
+int main(int argc, char *argv[])
+{
+	int ret = 0;
+
+	setlocale(LC_NUMERIC, "en_US");
+	pagesize = getpagesize();
+	online_cpus = get_online_cpus();
+	if (!online_cpus) {
+		fprintf(stderr, "cannot get online cpus %d/%s\n",
+			errno, strerror(errno));
+		ret = 1;
+		goto out;
+	} else if (handle_args(argc, argv)) {
+		ret = 1;
+		goto out;
+	}
+
+	ncpus = CPU_COUNT_S(CPU_ALLOC_SIZE(max_cpus), online_cpus);
+	if (ndevs > 1 && output_name && strcmp(output_name, "-") != 0) {
+		fprintf(stderr, "-o not supported with multiple devices\n");
+		ret = 1;
+		goto out;
+	}
+
+	signal(SIGINT, handle_sigint);
+	signal(SIGHUP, handle_sigint);
+	signal(SIGTERM, handle_sigint);
+	signal(SIGALRM, handle_sigint);
+	signal(SIGPIPE, SIG_IGN);
+
+	if (kill_running_trace) {
+		struct devpath *dpp;
+		struct list_head *p;
+
+		__list_for_each(p, &devpaths) {
+			dpp = list_entry(p, struct devpath, head);
+			if (__stop_trace(dpp->fd)) {
+				fprintf(stderr,
+					"BLKTRACETEARDOWN %s failed: %d/%s\n",
+					dpp->path, errno, strerror(errno));
+			}
+		}
+	} else if (net_mode == Net_server) {
+		if (output_name) {
+			fprintf(stderr, "-o ignored in server mode\n");
+			output_name = NULL;
+		}
+		ret = net_server();
+	} else
+		ret = run_tracers();
+
+out:
+	if (pfp)
+		fclose(pfp);
+	rel_devpaths();
+	return ret;
+}