[PATCH] blktrace: need to free ts->buf for networked transfer
[blktrace.git] / blktrace.c
index 0207dec0e47340bf9fef25a0d6901976db81acfd..f34588fb2a87eb331cba32b4a29eaf8873d6c4f8 100644 (file)
 #include <sys/param.h>
 #include <sys/statfs.h>
 #include <sys/poll.h>
+#include <sys/mman.h>
+#include <sys/socket.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <sched.h>
 #include <ctype.h>
 #include <getopt.h>
 #include <errno.h>
-#include <assert.h>
+#include <netinet/in.h>
+#include <arpa/inet.h>
+#include <netdb.h>
+#include <sys/sendfile.h>
 
 #include "blktrace.h"
-#include "list.h"
+#include "barrier.h"
 
 static char blktrace_version[] = "0.99";
 
@@ -54,10 +59,7 @@ static char blktrace_version[] = "0.99";
 
 #define RELAYFS_TYPE   0xF0B4A981
 
-#define RING_INIT_NR   (2)
-#define RING_MAX_NR    (16UL)
-
-#define S_OPTS "d:a:A:r:o:kw:Vb:n:D:"
+#define S_OPTS "d:a:A:r:o:kw:Vb:n:D:lh:p:"
 static struct option l_opts[] = {
        {
                .name = "dev",
@@ -125,43 +127,77 @@ static struct option l_opts[] = {
                .flag = NULL,
                .val = 'D'
        },
+       {
+               .name = "listen",
+               .has_arg = no_argument,
+               .flag = NULL,
+               .val = 'l'
+       },
+       {
+               .name = "host",
+               .has_arg = required_argument,
+               .flag = NULL,
+               .val = 'h'
+       },
+       {
+               .name = "port",
+               .has_arg = required_argument,
+               .flag = NULL,
+               .val = 'p'
+       },
        {
                .name = NULL,
        }
 };
 
 struct tip_subbuf {
-       struct list_head list;
        void *buf;
        unsigned int len;
        unsigned int max_len;
 };
 
+#define FIFO_SIZE      (1024)  /* should be plenty big! */
+#define CL_SIZE                (128)   /* cache line, any bigger? */
+
+struct tip_subbuf_fifo {
+       int tail __attribute__((aligned(CL_SIZE)));
+       int head __attribute__((aligned(CL_SIZE)));
+       struct tip_subbuf *q[FIFO_SIZE];
+};
+
 struct thread_information {
        int cpu;
        pthread_t thread;
 
        int fd;
        void *fd_buf;
-       unsigned long fd_off;
-       unsigned long fd_size;
-       unsigned long fd_max_size;
        char fn[MAXPATHLEN + 64];
 
        FILE *ofile;
        char *ofile_buffer;
        int ofile_stdout;
+       int ofile_mmap;
 
        unsigned long events_processed;
+       unsigned long long data_read;
        struct device_information *device;
 
        int exited;
 
-       pthread_mutex_t lock;
-       struct list_head subbuf_list;
-       void *leftover_ts;
-       int leftover_ts_len;
-       int leftover_ts_max;
+       /*
+        * piped fifo buffers
+        */
+       struct tip_subbuf_fifo fifo;
+       struct tip_subbuf *leftover_ts;
+
+       /*
+        * 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;
 };
 
 struct device_information {
@@ -186,6 +222,7 @@ 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;
 
 #define is_done()      (*(volatile int *)(&done))
 static volatile int done;
@@ -196,6 +233,8 @@ static volatile int trace_stopped;
 #define is_stat_shown()        (*(volatile int *)(&stat_shown))
 static volatile int stat_shown;
 
+int data_is_native = -1;
+
 static void exit_trace(int status);
 
 #define dip_tracing(dip)       (*(volatile int *)(&(dip)->trace_started))
@@ -208,6 +247,41 @@ static void exit_trace(int status);
 #define for_each_tip(__d, __t, __j)    \
        for (__j = 0, __t = (__d)->threads; __j < ncpus; __j++, __t++)
 
+/*
+ * 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 */
+};
+
+#define TRACE_NET_PORT         (8462)
+
+enum {
+       Net_none = 0,
+       Net_server,
+       Net_client,
+};
+
+/*
+ * network cmd line params
+ */
+static char hostname[MAXHOSTNAMELEN];
+static int net_port = TRACE_NET_PORT;
+static int net_mode = 0;
+
+static int net_in_fd = -1;
+static int net_out_fd = -1;
+
+static void handle_sigint(__attribute__((__unused__)) int sig)
+{
+       done = 1;
+}
+
 static int get_dropped_count(const char *buts_name)
 {
        int fd;
@@ -248,8 +322,13 @@ static int start_trace(struct device_information *dip)
        buts.buf_nr = buf_nr;
        buts.act_mask = act_mask;
 
-       if (ioctl(dip->fd, BLKSTARTTRACE, &buts) < 0) {
-               perror("BLKSTARTTRACE");
+       if (ioctl(dip->fd, BLKTRACESETUP, &buts) < 0) {
+               perror("BLKTRACESETUP");
+               return 1;
+       }
+
+       if (ioctl(dip->fd, BLKTRACESTART) < 0) {
+               perror("BLKTRACESTART");
                return 1;
        }
 
@@ -263,8 +342,10 @@ static void stop_trace(struct device_information *dip)
        if (dip_tracing(dip) || kill_running_trace) {
                dip_set_tracing(dip, 0);
 
-               if (ioctl(dip->fd, BLKSTOPTRACE) < 0)
-                       perror("BLKSTOPTRACE");
+               if (ioctl(dip->fd, BLKTRACESTOP) < 0)
+                       perror("BLKTRACESTOP");
+               if (ioctl(dip->fd, BLKTRACETEARDOWN) < 0)
+                       perror("BLKTRACETEARDOWN");
 
                close(dip->fd);
                dip->fd = -1;
@@ -287,7 +368,7 @@ static void wait_for_data(struct thread_information *tip)
        struct pollfd pfd = { .fd = tip->fd, .events = POLLIN };
 
        do {
-               poll(&pfd, 1, 10);
+               poll(&pfd, 1, 100);
                if (pfd.revents & POLLIN)
                        break;
                if (tip->ofile_stdout)
@@ -295,7 +376,7 @@ static void wait_for_data(struct thread_information *tip)
        } while (!is_done());
 }
 
-static int read_data(struct thread_information *tip, void *buf, int len)
+static int read_data_file(struct thread_information *tip, void *buf, int len)
 {
        int ret = 0;
 
@@ -319,49 +400,144 @@ static int read_data(struct thread_information *tip, void *buf, int len)
        } while (!is_done());
 
        return ret;
+
 }
 
-static inline void tip_fd_unlock(struct thread_information *tip)
+static int read_data_net(struct thread_information *tip, void *buf, int len)
 {
-       pthread_mutex_unlock(&tip->lock);
+       unsigned int bytes_left = len;
+       int ret = 0;
+
+       do {
+               ret = recv(net_in_fd, buf, bytes_left, MSG_WAITALL);
+
+               if (!ret)
+                       continue;
+               else if (ret < 0) {
+                       if (errno != EAGAIN) {
+                               perror(tip->fn);
+                               fprintf(stderr, "server: failed read\n");
+                               return 0;
+                       }
+                       continue;
+               } else {
+                       buf += ret;
+                       bytes_left -= ret;
+               }
+       } while (!is_done() && bytes_left);
+
+       return len;
 }
 
-static inline void tip_fd_lock(struct thread_information *tip)
+static int read_data(struct thread_information *tip, void *buf, int len)
 {
-       pthread_mutex_lock(&tip->lock);
+       if (net_mode == Net_server)
+               return read_data_net(tip, buf, len);
+
+       return read_data_file(tip, buf, len);
 }
 
-static int get_subbuf(struct thread_information *tip)
+static inline struct tip_subbuf *
+subbuf_fifo_dequeue(struct thread_information *tip)
+{
+       const int head = tip->fifo.head;
+       const int next = (head + 1) & (FIFO_SIZE - 1);
+
+       if (head != tip->fifo.tail) {
+               struct tip_subbuf *ts = tip->fifo.q[head];
+
+               store_barrier();
+               tip->fifo.head = next;
+               return ts;
+       }
+
+       return NULL;
+}
+
+static inline int subbuf_fifo_queue(struct thread_information *tip,
+                                   struct tip_subbuf *ts)
+{
+       const int tail = tip->fifo.tail;
+       const int next = (tail + 1) & (FIFO_SIZE - 1);
+
+       if (next != tip->fifo.head) {
+               tip->fifo.q[tail] = ts;
+               store_barrier();
+               tip->fifo.tail = next;
+               return 0;
+       }
+
+       fprintf(stderr, "fifo too small!\n");
+       return 1;
+}
+
+/*
+ * For file output, truncate and mmap the file appropriately
+ */
+static int mmap_subbuf(struct thread_information *tip, unsigned int maxlen)
 {
-       struct tip_subbuf *ts;
-       int ts_size, ret;
+       int ofd = fileno(tip->ofile);
+       int ret;
 
        /*
-        * live tracing should get a lower count to make it more "realtime"
+        * extend file, if we have to. use chunks of 16 subbuffers.
         */
-       if (tip->ofile_stdout)
-               ts_size = 1024;
-       else
-               ts_size = buf_size;
+       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;
+               }
 
-       ts = malloc(sizeof(*ts));
-       ts->buf = malloc(ts_size);
-       ts->max_len = ts_size;
+               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;
 
-       ret = read_data(tip, ts->buf, ts_size);
-       if (ret > 0) {
-               ts->len = ret;
-               tip_fd_lock(tip);
-               list_add_tail(&ts->list, &tip->subbuf_list);
-               tip_fd_unlock(tip);
+               if (ftruncate(ofd, tip->fs_max_size) < 0) {
+                       perror("ftruncate");
+                       return -1;
+               }
+
+               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);
+       }
+
+       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;
        }
 
-       free(ts->buf);
-       free(ts);
        return -1;
 }
 
+/*
+ * Use the copy approach for pipes and network
+ */
+static int get_subbuf(struct thread_information *tip)
+{
+       struct tip_subbuf *ts = malloc(sizeof(*ts));
+       int ret;
+
+       ts->buf = malloc(buf_size);
+       ts->max_len = buf_size;
+
+       ret = read_data(tip, ts->buf, ts->max_len);
+       if (ret > 0) {
+               ts->len = ret;
+               return subbuf_fifo_queue(tip, ts);
+       }
+
+       return ret;
+}
+
 static void close_thread(struct thread_information *tip)
 {
        if (tip->fd != -1)
@@ -379,6 +555,21 @@ static void close_thread(struct thread_information *tip)
        tip->fd_buf = NULL;
 }
 
+static void tip_ftrunc_final(struct thread_information *tip)
+{
+       /*
+        * 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);
+       }
+}
+
 static void *thread_main(void *arg)
 {
        struct thread_information *tip = arg;
@@ -403,20 +594,70 @@ static void *thread_main(void *arg)
                exit_trace(1);
        }
 
-       for (;;) {
-               if (get_subbuf(tip))
-                       break;
+       while (!is_done()) {
+               if (tip->ofile_mmap && net_mode != Net_client) {
+                       if (mmap_subbuf(tip, buf_size))
+                               break;
+               } else {
+                       if (get_subbuf(tip))
+                               break;
+               }
        }
 
+       tip_ftrunc_final(tip);
        tip->exited = 1;
        return NULL;
 }
 
-static int write_data(struct thread_information *tip,
-                     void *buf, unsigned int buf_len)
+static int write_data_net(int fd, void *buf, unsigned int buf_len)
 {
+       unsigned int bytes_left = buf_len;
        int ret;
 
+       while (bytes_left) {
+               ret = send(fd, buf, bytes_left, 0);
+               if (ret < 0) {
+                       perror("send");
+                       return 1;
+               }
+
+               buf += ret;
+               bytes_left -= ret;
+       }
+
+       return 0;
+}
+
+static int flush_subbuf_net(struct thread_information *tip,
+                           struct tip_subbuf *ts)
+{
+       struct blktrace_net_hdr hdr;
+
+       hdr.magic = BLK_IO_TRACE_MAGIC;
+       strcpy(hdr.buts_name, tip->device->buts_name);
+       hdr.cpu = tip->cpu;
+       hdr.max_cpus = ncpus;
+       hdr.len = ts->len;
+
+       if (write_data_net(net_out_fd, &hdr, sizeof(hdr)))
+               return 1;
+
+       if (write_data_net(net_out_fd, ts->buf, ts->len))
+               return 1;
+
+       free(ts->buf);
+       free(ts);
+       return 0;
+}
+
+static int write_data(struct thread_information *tip, void *buf,
+                     unsigned int buf_len)
+{
+       int ret;
+
+       if (!buf_len)
+               return 0;
+
        while (1) {
                ret = fwrite(buf, buf_len, 1, tip->ofile);
                if (ret == 1)
@@ -434,80 +675,81 @@ static int write_data(struct thread_information *tip,
        return 0;
 }
 
-static int flush_subbuf(struct thread_information *tip, struct tip_subbuf *ts)
+static int flush_subbuf_file(struct thread_information *tip,
+                            struct tip_subbuf *ts)
 {
        unsigned int offset = 0;
        struct blk_io_trace *t;
        int pdu_len, events = 0;
 
        /*
-        * surplus from last run, a little ugly...
+        * surplus from last run
         */
-       if (tip->leftover_ts_len) {
-               if (ts->len + tip->leftover_ts_len > ts->max_len)
-                       ts->buf = realloc(ts->buf, ts->len + tip->leftover_ts_len);
+       if (tip->leftover_ts) {
+               struct tip_subbuf *prev_ts = tip->leftover_ts;
 
-               memmove(ts->buf + tip->leftover_ts_len, ts->buf, ts->len);
-               memcpy(ts->buf, tip->leftover_ts, tip->leftover_ts_len);
-               ts->len += tip->leftover_ts_len;
-               tip->leftover_ts_len = 0;
+               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);
+               }
+
+               memcpy(prev_ts->buf + prev_ts->len, ts->buf, ts->len);
+               prev_ts->len += ts->len;
+
+               free(ts->buf);
+               free(ts);
+
+               ts = prev_ts;
+               tip->leftover_ts = NULL;
        }
 
        while (offset + sizeof(*t) <= ts->len) {
                t = ts->buf + offset;
 
-               if (verify_trace(t))
+               if (verify_trace(t)) {
+                       write_data(tip, ts->buf, offset);
                        return -1;
+               }
 
                pdu_len = t->pdu_len;
 
                if (offset + sizeof(*t) + pdu_len > ts->len)
                        break;
 
-               trace_to_be(t);
-
-               if (write_data(tip, t, sizeof(*t) + pdu_len))
-                       return -1;
-
                offset += sizeof(*t) + pdu_len;
                tip->events_processed++;
+               tip->data_read += sizeof(*t) + pdu_len;
                events++;
        }
 
+       if (write_data(tip, ts->buf, offset))
+               return -1;
+
        /*
         * leftover bytes, save them for next time
         */
        if (offset != ts->len) {
-               int surplus = ts->len - offset;
-
-               t = ts->buf + offset;
-               if (surplus > tip->leftover_ts_max) {
-                       tip->leftover_ts = realloc(tip->leftover_ts, surplus);
-                       tip->leftover_ts_max = surplus;
-               }
-
-               memcpy(tip->leftover_ts, ts->buf + offset, surplus);
-               tip->leftover_ts_len = surplus;
+               tip->leftover_ts = ts;
+               ts->len -= offset;
+               memmove(ts->buf, ts->buf + offset, ts->len);
+       } else {
+               free(ts->buf);
+               free(ts);
        }
 
-       free(ts->buf);
-       free(ts);
        return events;
 }
 
 static int write_tip_events(struct thread_information *tip)
 {
-       struct tip_subbuf *ts = NULL;
+       struct tip_subbuf *ts = subbuf_fifo_dequeue(tip);
 
-       tip_fd_lock(tip);
-       if (!list_empty(&tip->subbuf_list)) {
-               ts = list_entry(tip->subbuf_list.next, struct tip_subbuf, list);
-               list_del(&ts->list);
-       }
-       tip_fd_unlock(tip);
+       if (ts) {
+               if (net_mode == Net_client)
+                       return flush_subbuf_net(tip, ts);
 
-       if (ts)
-               return flush_subbuf(tip, ts);
+               return flush_subbuf_file(tip, ts);
+       }
 
        return 0;
 }
@@ -519,7 +761,7 @@ static void get_and_write_events(void)
 {
        struct device_information *dip;
        struct thread_information *tip;
-       int i, j, events, ret, all_exited;
+       int i, j, events, ret, tips_running;
 
        while (!is_done()) {
                events = 0;
@@ -541,55 +783,82 @@ static void get_and_write_events(void)
         */
        do {
                events = 0;
-               all_exited = 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;
-                               all_exited += !tip->exited;
+                               tips_running += !tip->exited;
                        }
                }
                usleep(10);
-       } while (events && !all_exited);
+       } while (events || tips_running);
+}
+
+static void wait_for_threads(void)
+{
+       /*
+        * for piped or network output, poll and fetch data for writeout.
+        * for files, we just wait around for trace threads to exit
+        */
+       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;
+
+               do {
+                       tips_running = 0;
+                       usleep(1000);
+
+                       for_each_dip(dip, i)
+                               for_each_tip(dip, tip, j)
+                                       tips_running += !tip->exited;
+               } while (tips_running);
+       }
+}
+
+static void fill_ofname(char *dst, char *buts_name, int cpu)
+{
+       int len = 0;
+
+       if (output_dir)
+               len = sprintf(dst, "%s/", output_dir);
+
+       if (output_name)
+               sprintf(dst + len, "%s.blktrace.%d", output_name, cpu);
+       else
+               sprintf(dst + len, "%s.blktrace.%d", buts_name, cpu);
 }
 
 static int start_threads(struct device_information *dip)
 {
        struct thread_information *tip;
-       char op[64];
        int j, pipeline = output_name && !strcmp(output_name, "-");
-       int len, mode, vbuf_size;
+       int mode, vbuf_size;
+       char op[64];
 
        for_each_tip(dip, tip, j) {
                tip->cpu = j;
                tip->device = dip;
                tip->events_processed = 0;
-               INIT_LIST_HEAD(&tip->subbuf_list);
-               tip->leftover_ts_len = 0;
-               tip->leftover_ts_max = 0;
+               memset(&tip->fifo, 0, sizeof(tip->fifo));
                tip->leftover_ts = NULL;
 
                if (pipeline) {
                        tip->ofile = fdopen(STDOUT_FILENO, "w");
                        tip->ofile_stdout = 1;
+                       tip->ofile_mmap = 0;
                        mode = _IOLBF;
                        vbuf_size = 512;
                } else {
-                       len = 0;
-
-                       if (output_dir)
-                               len = sprintf(op, "%s/", output_dir);
-
-                       if (output_name) {
-                               sprintf(op + len, "%s.blktrace.%d", output_name,
-                                       tip->cpu);
-                       } else {
-                               sprintf(op + len, "%s.blktrace.%d",
-                                       dip->buts_name, tip->cpu);
-                       }
-                       tip->ofile = fopen(op, "w");
+                       fill_ofname(op, dip->buts_name, tip->cpu);
+                       tip->ofile = fopen(op, "w+");
                        tip->ofile_stdout = 0;
+                       tip->ofile_mmap = 1;
                        mode = _IOFBF;
                        vbuf_size = OFILE_BUF;
                }
@@ -739,7 +1008,7 @@ static void show_stats(void)
 {
        struct device_information *dip;
        struct thread_information *tip;
-       unsigned long long events_processed;
+       unsigned long long events_processed, data_read;
        unsigned long total_drops;
        int i, j, no_stdout = 0;
 
@@ -756,22 +1025,244 @@ static void show_stats(void)
                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: %20ld events\n",
-                                       tip->cpu, tip->events_processed);
+                               printf("  CPU%3d: %20lu events, %8llu KiB data\n",
+                                       tip->cpu, tip->events_processed,
+                                       tip->data_read >> 10);
                        events_processed += tip->events_processed;
+                       data_read += tip->data_read;
                }
                total_drops += dip->drop_count;
                if (!no_stdout)
-                       printf("  Total:  %20lld events (dropped %lu)\n",
-                                       events_processed, dip->drop_count);
+                       printf("  Total:  %20llu events (dropped %lu), %8llu KiB data\n",
+                                       events_processed, dip->drop_count,
+                                       data_read >> 10);
        }
 
        if (total_drops)
                fprintf(stderr, "You have dropped events, consider using a larger buffer size (-b)\n");
 }
 
+static struct device_information *net_get_dip(char *buts_name)
+{
+       struct device_information *dip;
+       int i;
+
+       for (i = 0; i < ndevs; i++) {
+               dip = &device_information[i];
+
+               if (!strcmp(dip->buts_name, buts_name))
+                       return dip;
+       }
+
+       device_information = realloc(device_information, (ndevs + 1) * sizeof(*dip));
+       dip = &device_information[ndevs];
+       strcpy(dip->buts_name, buts_name);
+       ndevs++;
+       dip->threads = malloc(ncpus * sizeof(struct thread_information));
+       memset(dip->threads, 0, ncpus * sizeof(struct thread_information));
+
+       /*
+        * open all files
+        */
+       for (i = 0; i < ncpus; i++) {
+               struct thread_information *tip = &dip->threads[i];
+               char op[64];
+
+               tip->cpu = i;
+               tip->ofile_stdout = 0;
+               tip->ofile_mmap = 1;
+               tip->device = dip;
+
+               fill_ofname(op, dip->buts_name, tip->cpu);
+
+               tip->ofile = fopen(op, "w+");
+               if (!tip->ofile) {
+                       perror("fopen");
+                       return NULL;
+               }
+       }
+
+       return dip;
+}
+
+static struct thread_information *net_get_tip(struct blktrace_net_hdr *bnh)
+{
+       struct device_information *dip;
+
+       ncpus = bnh->max_cpus;
+       dip = net_get_dip(bnh->buts_name);
+       return &dip->threads[bnh->cpu];
+}
+
+static int net_get_header(struct blktrace_net_hdr *bnh)
+{
+       int fl = fcntl(net_in_fd, F_GETFL);
+       int bytes_left, ret;
+       void *p = bnh;
+
+       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;
+               }
+       }
+       fcntl(net_in_fd, F_SETFL, fl & ~O_NONBLOCK);
+       return 0;
+}
+
+static int net_server_loop(void)
+{
+       struct thread_information *tip;
+       struct blktrace_net_hdr bnh;
+
+       if (net_get_header(&bnh))
+               return 1;
+
+       if (data_is_native == -1 && check_data_endianness(bnh.magic)) {
+               fprintf(stderr, "server: received data is bad\n");
+               return 1;
+       }
+
+       if (!data_is_native) {
+               bnh.cpu = be32_to_cpu(bnh.cpu);
+               bnh.len = be32_to_cpu(bnh.len);
+       }
+
+       tip = net_get_tip(&bnh);
+       if (!tip)
+               return 1;
+
+       if (mmap_subbuf(tip, bnh.len))
+               return 1;
+
+       return 0;
+}
+
+/*
+ * Start here when we are in server mode - just fetch data from the network
+ * and dump to files
+ */
+static int net_server(void)
+{
+       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;
+       }
+
+       opt = 1;
+       if (setsockopt(fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)) < 0) {
+               perror("setsockopt");
+               return 1;
+       }
+
+       memset(&addr, 0, sizeof(addr));
+       addr.sin_family = AF_INET;
+       addr.sin_addr.s_addr = htonl(INADDR_ANY);
+       addr.sin_port = htons(net_port);
+
+       if (bind(fd, (struct sockaddr *) &addr, sizeof(addr)) < 0) {
+               perror("bind");
+               return 1;
+       }
+
+       if (listen(fd, 1) < 0) {
+               perror("listen");
+               return 1;
+       }
+
+       printf("blktrace: waiting for incoming connection...\n");
+
+       socklen = sizeof(addr);
+       net_in_fd = accept(fd, (struct sockaddr *) &addr, &socklen);
+       if (net_in_fd < 0) {
+               perror("accept");
+               return 1;
+       }
+
+       signal(SIGINT, handle_sigint);
+       signal(SIGHUP, handle_sigint);
+       signal(SIGTERM, handle_sigint);
+       signal(SIGALRM, handle_sigint);
+
+       printf("blktrace: connected!\n");
+
+       while (!is_done()) {
+               if (net_server_loop())
+                       break;
+       }
+
+       for (i = 0; i < ndevs; i++) {
+               struct device_information *dip = &device_information[i];
+
+               for (j = 0; j < ncpus; j++)
+                       tip_ftrunc_final(&dip->threads[j]);
+       }
+
+       return 0;
+}
+
+/*
+ * Setup outgoing network connection where we will transmit data
+ */
+static int net_setup_client(void)
+{
+       struct sockaddr_in addr;
+       int fd;
+
+       fd = socket(AF_INET, SOCK_STREAM, 0);
+       if (fd < 0) {
+               perror("client: socket");
+               return 1;
+       }
+
+       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;
+               }
+
+               memcpy(&addr.sin_addr, hent->h_addr, 4);
+               strcpy(hostname, hent->h_name);
+       }
+
+       printf("blktrace: connecting to %s\n", hostname);
+
+       if (connect(fd, (struct sockaddr *) &addr, sizeof(addr)) < 0) {
+               perror("client: connect");
+               return 1;
+       }
+
+       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" \
@@ -791,10 +1282,6 @@ static void show_usage(char *program)
 {
        fprintf(stderr, "Usage: %s %s %s",program, blktrace_version, usage_str);
 }
-static void handle_sigint(__attribute__((__unused__)) int sig)
-{
-       done = 1;
-}
 
 int main(int argc, char *argv[])
 {
@@ -874,12 +1361,29 @@ int main(int argc, char *argv[])
                case 'D':
                        output_dir = optarg;
                        break;
+               case 'h':
+                       net_mode = Net_client;
+                       strcpy(hostname, optarg);
+                       break;
+               case 'l':
+                       net_mode = Net_server;
+                       break;
+               case 'p':
+                       net_port = atoi(optarg);
+                       break;
                default:
                        show_usage(argv[0]);
                        return 1;
                }
        }
 
+       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)
                        return 1;
@@ -915,28 +1419,29 @@ int main(int argc, char *argv[])
                return 0;
        }
 
-       setlocale(LC_NUMERIC, "en_US");
-
        ncpus = sysconf(_SC_NPROCESSORS_ONLN);
        if (ncpus < 0) {
                fprintf(stderr, "sysconf(_SC_NPROCESSORS_ONLN) failed\n");
                return 1;
        }
 
-       if (start_devices() != 0)
-               return 1;
-
        signal(SIGINT, handle_sigint);
        signal(SIGHUP, handle_sigint);
        signal(SIGTERM, handle_sigint);
        signal(SIGALRM, handle_sigint);
 
+       if (net_mode == Net_client && net_setup_client())
+               return 1;
+
+       if (start_devices() != 0)
+               return 1;
+
        atexit(stop_all_tracing);
 
        if (stop_watch)
                alarm(stop_watch);
 
-       get_and_write_events();
+       wait_for_threads();
 
        if (!is_trace_stopped()) {
                trace_stopped = 1;