[PATCH] blkparse: another stab at stopwatch_end fixing
[blktrace.git] / blkparse.c
index 14b5b022f58acf5784f855affe2263d3abd0eaa6..c9b37593d161ab28d2bf33c1b418c3c762e162bc 100644 (file)
 #include <getopt.h>
 #include <errno.h>
 #include <signal.h>
+#include <locale.h>
+#include <limits.h>
 
 #include "blktrace.h"
 #include "rbtree.h"
+#include "jhash.h"
 
-#define MAX_CPUS       (512)
+static char blkparse_version[] = "0.90";
 
-#define SECONDS(x)     ((unsigned long long)(x) / 1000000000)
-#define NANO_SECONDS(x)        ((unsigned long long)(x) % 1000000000)
+struct per_dev_info {
+       dev_t id;
+       char *name;
 
-static int backwards;
-static unsigned long long genesis_time, last_reported_time;
+       int backwards;
+       unsigned long long events;
+       unsigned long long last_reported_time;
+       unsigned long long last_read_time;
+       struct io_stats io_stats;
+       unsigned long last_sequence;
+       unsigned long skips;
 
-struct per_cpu_info {
-       int cpu;
-       int nelems;
-
-       int fd;
-       char fname[128];
-
-       FILE *ofp;
-       char ofname[128];
+       int nfiles;
+       int ncpus;
+       struct per_cpu_info *cpus;
+};
 
-       unsigned long qreads, qwrites, creads, cwrites, mreads, mwrites;
-       unsigned long ireads, iwrites;
-       unsigned long long qread_kb, qwrite_kb, cread_kb, cwrite_kb;
-       unsigned long long iread_kb, iwrite_kb;
+struct per_process_info {
+       char name[16];
+       __u32 pid;
+       struct io_stats io_stats;
+       struct per_process_info *hash_next, *list_next;
+       int more_than_one;
+
+       /*
+        * individual io stats
+        */
+       unsigned long long longest_allocation_wait[2];
+       unsigned long long longest_dispatch_wait[2];
+       unsigned long long longest_completion_wait[2];
 };
 
-#define S_OPTS "i:o:"
+#define PPI_HASH_SHIFT (8)
+#define PPI_HASH_SIZE  (1 << PPI_HASH_SHIFT)
+#define PPI_HASH_MASK  (PPI_HASH_SIZE - 1)
+static struct per_process_info *ppi_hash_table[PPI_HASH_SIZE];
+static struct per_process_info *ppi_list;
+static int ppi_list_entries;
+
+#define S_OPTS "i:o:b:stqw:f:F:vn"
 static struct option l_opts[] = {
        {
                .name = "input",
-               .has_arg = 1,
+               .has_arg = required_argument,
                .flag = NULL,
                .val = 'i'
        },
        {
                .name = "output",
-               .has_arg = 1,
+               .has_arg = required_argument,
                .flag = NULL,
                .val = 'o'
        },
        {
-               .name = NULL,
-               .has_arg = 0,
+               .name = "batch",
+               .has_arg = required_argument,
                .flag = NULL,
-               .val = 0
-       }
+               .val = 'b'
+       },
+       {
+               .name = "per program stats",
+               .has_arg = no_argument,
+               .flag = NULL,
+               .val = 's'
+       },
+       {
+               .name = "track ios",
+               .has_arg = no_argument,
+               .flag = NULL,
+               .val = 't'
+       },
+       {
+               .name = "quiet",
+               .has_arg = no_argument,
+               .flag = NULL,
+               .val = 'q'
+       },
+       {
+               .name = "stopwatch",
+               .has_arg = required_argument,
+               .flag = NULL,
+               .val = 'w'
+       },
+       {
+               .name = "format",
+               .has_arg = required_argument,
+               .flag = NULL,
+               .val = 'f'
+       },
+       {
+               .name = "format-spec",
+               .has_arg = required_argument,
+               .flag = NULL,
+               .val = 'F'
+       },
+       {
+               .name = "hash by name",
+               .has_arg = no_argument,
+               .flag = NULL,
+               .val = 'n'
+       },
+       {
+               .name = "version",
+               .has_arg = no_argument,
+               .flag = NULL,
+               .val = 'v'
+       },
 };
 
-static struct rb_root rb_root;
-
+/*
+ * for sorting the displayed output
+ */
 struct trace {
        struct blk_io_trace *bit;
        struct rb_node rb_node;
+       struct trace *next;
+       int skipped;
+};
+
+static struct rb_root rb_sort_root;
+static unsigned long rb_sort_entries;
+
+static struct rb_root rb_track_root;
+
+static struct trace *trace_list;
+
+/*
+ * allocation cache
+ */
+static struct blk_io_trace *bit_alloc_list;
+static struct trace *t_alloc_list;
+
+/*
+ * for tracking individual ios
+ */
+struct io_track {
+       struct rb_node rb_node;
+
+       dev_t device;
+       __u64 sector;
+       __u32 pid;
+       char comm[16];
+       unsigned long long allocation_time;
+       unsigned long long queue_time;
+       unsigned long long dispatch_time;
+       unsigned long long completion_time;
 };
 
-static struct per_cpu_info per_cpu_info[MAX_CPUS];
+static int ndevices;
+static struct per_dev_info *devices;
+static char *get_dev_name(struct per_dev_info *, char *, int);
 
-static unsigned long long events;
+FILE *ofp = NULL;
+static char *output_name;
 
-static int max_cpus;
-static int nfiles;
+static unsigned long long genesis_time;
+static unsigned long long last_allowed_time;
+static unsigned long long stopwatch_start;     /* start from zero by default */
+static unsigned long long stopwatch_end = ULONG_LONG_MAX;      /* "infinity" */
 
-static char *dev, *output_name;
+static int per_process_stats;
+static int track_ios;
+static int ppi_hash_by_pid = 1;
+
+#define RB_BATCH_DEFAULT       (512)
+static int rb_batch = RB_BATCH_DEFAULT;
+
+static int pipeline;
 
 #define is_done()      (*(volatile int *)(&done))
 static volatile int done;
 
-static inline void check_time(struct blk_io_trace *bit)
+#define JHASH_RANDOM   (0x3af5f2ee)
+
+static inline int ppi_hash_pid(__u32 pid)
+{
+       return jhash_1word(pid, JHASH_RANDOM) & PPI_HASH_MASK;
+}
+
+static inline int ppi_hash_name(const char *name)
+{
+       return jhash(name, 16, JHASH_RANDOM) & PPI_HASH_MASK;
+}
+
+static inline int ppi_hash(struct per_process_info *ppi)
+{
+       if (ppi_hash_by_pid)
+               return ppi_hash_pid(ppi->pid);
+
+       return ppi_hash_name(ppi->name);
+}
+
+static inline void add_process_to_hash(struct per_process_info *ppi)
+{
+       const int hash_idx = ppi_hash(ppi);
+
+       ppi->hash_next = ppi_hash_table[hash_idx];
+       ppi_hash_table[hash_idx] = ppi;
+}
+
+static inline void add_process_to_list(struct per_process_info *ppi)
+{
+       ppi->list_next = ppi_list;
+       ppi_list = ppi;
+       ppi_list_entries++;
+}
+
+static struct per_process_info *find_process_by_name(char *name)
+{
+       const int hash_idx = ppi_hash_name(name);
+       struct per_process_info *ppi;
+
+       ppi = ppi_hash_table[hash_idx];
+       while (ppi) {
+               if (!strcmp(ppi->name, name))
+                       return ppi;
+
+               ppi = ppi->hash_next;
+       }
+
+       return NULL;
+}
+
+static struct per_process_info *find_process_by_pid(__u32 pid)
+{
+       const int hash_idx = ppi_hash_pid(pid);
+       struct per_process_info *ppi;
+
+       ppi = ppi_hash_table[hash_idx];
+       while (ppi) {
+               if (ppi->pid == pid)
+                       return ppi;
+
+               ppi = ppi->hash_next;
+       }
+
+       return NULL;
+}
+
+static struct per_process_info *find_process(__u32 pid, char *name)
+{
+       struct per_process_info *ppi;
+
+       if (ppi_hash_by_pid)
+               return find_process_by_pid(pid);
+
+       ppi = find_process_by_name(name);
+       if (ppi && ppi->pid != pid)
+               ppi->more_than_one = 1;
+
+       return ppi;
+}
+
+static inline int trace_rb_insert(struct trace *t)
+{
+       struct rb_node **p = &rb_sort_root.rb_node;
+       struct rb_node *parent = NULL;
+       struct trace *__t;
+
+       while (*p) {
+               parent = *p;
+               __t = rb_entry(parent, struct trace, rb_node);
+
+               if (t->bit->time < __t->bit->time)
+                       p = &(*p)->rb_left;
+               else if (t->bit->time > __t->bit->time)
+                       p = &(*p)->rb_right;
+               else if (t->bit->device < __t->bit->device)
+                       p = &(*p)->rb_left;
+               else if (t->bit->device > __t->bit->device)
+                       p = &(*p)->rb_right;
+               else if (t->bit->sequence < __t->bit->sequence)
+                       p = &(*p)->rb_left;
+               else if (t->bit->sequence > __t->bit->sequence)
+                       p = &(*p)->rb_right;
+               else if (t->bit->device == __t->bit->device) {
+                       fprintf(stderr,
+                               "sequence alias (%d) on device %d,%d!\n",
+                               t->bit->sequence,
+                               MAJOR(t->bit->device), MINOR(t->bit->device));
+                       return 1;
+               }
+       }
+
+       rb_sort_entries++;
+       rb_link_node(&t->rb_node, parent, p);
+       rb_insert_color(&t->rb_node, &rb_sort_root);
+       return 0;
+}
+
+static struct trace *trace_rb_find(dev_t device, unsigned long sequence)
+{
+       struct rb_node **p = &rb_sort_root.rb_node;
+       struct rb_node *parent = NULL;
+       struct trace *__t;
+
+       while (*p) {
+               parent = *p;
+               __t = rb_entry(parent, struct trace, rb_node);
+
+               if (device < __t->bit->device)
+                       p = &(*p)->rb_left;
+               else if (device > __t->bit->device)
+                       p = &(*p)->rb_right;
+               else if (sequence < __t->bit->sequence)
+                       p = &(*p)->rb_left;
+               else if (sequence > __t->bit->sequence)
+                       p = &(*p)->rb_right;
+               else
+                       return __t;
+       }
+
+       return NULL;
+}
+
+static inline int track_rb_insert(struct io_track *iot)
+{
+       struct rb_node **p = &rb_track_root.rb_node;
+       struct rb_node *parent = NULL;
+       struct io_track *__iot;
+
+       while (*p) {
+               parent = *p;
+
+               __iot = rb_entry(parent, struct io_track, rb_node);
+
+               if (iot->device < __iot->device)
+                       p = &(*p)->rb_left;
+               else if (iot->device > __iot->device)
+                       p = &(*p)->rb_right;
+               else if (iot->sector < __iot->sector)
+                       p = &(*p)->rb_left;
+               else if (iot->sector > __iot->sector)
+                       p = &(*p)->rb_right;
+               else {
+                       fprintf(stderr,
+                               "sector alias (%Lu) on device %d,%d!\n",
+                               (unsigned long long) iot->sector,
+                               MAJOR(iot->device), MINOR(iot->device));
+                       return 1;
+               }
+       }
+
+       rb_link_node(&iot->rb_node, parent, p);
+       rb_insert_color(&iot->rb_node, &rb_track_root);
+       return 0;
+}
+
+static struct io_track *__find_track(dev_t device, __u64 sector)
+{
+       struct rb_node **p = &rb_track_root.rb_node;
+       struct rb_node *parent = NULL;
+       struct io_track *__iot;
+
+       while (*p) {
+               parent = *p;
+               
+               __iot = rb_entry(parent, struct io_track, rb_node);
+
+               if (device < __iot->device)
+                       p = &(*p)->rb_left;
+               else if (device > __iot->device)
+                       p = &(*p)->rb_right;
+               else if (sector < __iot->sector)
+                       p = &(*p)->rb_left;
+               else if (sector > __iot->sector)
+                       p = &(*p)->rb_right;
+               else
+                       return __iot;
+       }
+
+       return NULL;
+}
+
+static struct io_track *find_track(__u32 pid, char *comm, dev_t device,
+                                  __u64 sector)
+{
+       struct io_track *iot;
+
+       iot = __find_track(device, sector);
+       if (!iot) {
+               iot = malloc(sizeof(*iot));
+               iot->pid = pid;
+               memcpy(iot->comm, comm, sizeof(iot->comm));
+               iot->device = device;
+               iot->sector = sector;
+               track_rb_insert(iot);
+       }
+
+       return iot;
+}
+
+static void log_track_frontmerge(struct blk_io_trace *t)
+{
+       struct io_track *iot;
+
+       if (!track_ios)
+               return;
+
+       iot = __find_track(t->device, t->sector + (t->bytes >> 9));
+       if (!iot) {
+               fprintf(stderr, "failed to find mergeable event\n");
+               return;
+       }
+
+       rb_erase(&iot->rb_node, &rb_track_root);
+       iot->sector -= t->bytes >> 9;
+       track_rb_insert(iot);
+}
+
+static void log_track_getrq(struct blk_io_trace *t)
+{
+       struct io_track *iot;
+
+       if (!track_ios)
+               return;
+
+       iot = find_track(t->pid, t->comm, t->device, t->sector);
+       iot->allocation_time = t->time;
+}
+
+
+/*
+ * return time between rq allocation and insertion
+ */
+static unsigned long long log_track_insert(struct blk_io_trace *t)
+{
+       unsigned long long elapsed;
+       struct io_track *iot;
+
+       if (!track_ios)
+               return -1;
+
+       iot = find_track(t->pid, t->comm, t->device, t->sector);
+       iot->queue_time = t->time;
+       elapsed = iot->queue_time - iot->allocation_time;
+
+       if (per_process_stats) {
+               struct per_process_info *ppi = find_process(iot->pid,iot->comm);
+               int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+
+               if (ppi && elapsed > ppi->longest_allocation_wait[w])
+                       ppi->longest_allocation_wait[w] = elapsed;
+       }
+
+       return elapsed;
+}
+
+/*
+ * return time between queue and issue
+ */
+static unsigned long long log_track_issue(struct blk_io_trace *t)
+{
+       unsigned long long elapsed;
+       struct io_track *iot;
+
+       if (!track_ios)
+               return -1;
+       if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
+               return -1;
+
+       iot = __find_track(t->device, t->sector);
+       if (!iot) {
+               fprintf(stderr, "failed to find issue event\n");
+               return -1;
+       }
+
+       iot->dispatch_time = t->time;
+       elapsed = iot->dispatch_time - iot->queue_time;
+
+       if (per_process_stats) {
+               struct per_process_info *ppi = find_process(iot->pid,iot->comm);
+               int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+
+               if (ppi && elapsed > ppi->longest_dispatch_wait[w])
+                       ppi->longest_dispatch_wait[w] = elapsed;
+       }
+
+       return elapsed;
+}
+
+/*
+ * return time between dispatch and complete
+ */
+static unsigned long long log_track_complete(struct blk_io_trace *t)
+{
+       unsigned long long elapsed;
+       struct io_track *iot;
+
+       if (!track_ios)
+               return -1;
+       if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
+               return -1;
+
+       iot = __find_track(t->device, t->sector);
+       if (!iot) {
+               fprintf(stderr, "failed to find complete event\n");
+               return -1;
+       }
+
+       iot->completion_time = t->time;
+       elapsed = iot->completion_time - iot->dispatch_time;
+
+       if (per_process_stats) {
+               struct per_process_info *ppi = find_process(iot->pid,iot->comm);
+               int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+
+               if (ppi && elapsed > ppi->longest_completion_wait[w])
+                       ppi->longest_completion_wait[w] = elapsed;
+       }
+
+       /*
+        * kill the trace, we don't need it after completion
+        */
+       rb_erase(&iot->rb_node, &rb_track_root);
+       free(iot);
+
+       return elapsed;
+}
+
+
+static struct io_stats *find_process_io_stats(__u32 pid, char *name)
+{
+       struct per_process_info *ppi = find_process(pid, name);
+
+       if (!ppi) {
+               ppi = malloc(sizeof(*ppi));
+               memset(ppi, 0, sizeof(*ppi));
+               memcpy(ppi->name, name, 16);
+               ppi->pid = pid;
+               add_process_to_hash(ppi);
+               add_process_to_list(ppi);
+       }
+
+       return &ppi->io_stats;
+}
+
+static void resize_cpu_info(struct per_dev_info *pdi, int cpu)
+{
+       struct per_cpu_info *cpus = pdi->cpus;
+       int ncpus = pdi->ncpus;
+       int new_count = cpu + 1;
+       int new_space, size;
+       char *new_start;
+
+       size = new_count * sizeof(struct per_cpu_info);
+       cpus = realloc(cpus, size);
+       if (!cpus) {
+               char name[20];
+               fprintf(stderr, "Out of memory, CPU info for device %s (%d)\n",
+                       get_dev_name(pdi, name, sizeof(name)), size);
+               exit(1);
+       }
+
+       new_start = (char *)cpus + (ncpus * sizeof(struct per_cpu_info));
+       new_space = (new_count - ncpus) * sizeof(struct per_cpu_info);
+       memset(new_start, 0, new_space);
+
+       pdi->ncpus = new_count;
+       pdi->cpus = cpus;
+}
+
+static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
+{
+       struct per_cpu_info *pci;
+
+       if (cpu >= pdi->ncpus)
+               resize_cpu_info(pdi, cpu);
+
+       pci = &pdi->cpus[cpu];
+       pci->cpu = cpu;
+       return pci;
+}
+
+
+static int resize_devices(char *name)
+{
+       int size = (ndevices + 1) * sizeof(struct per_dev_info);
+
+       devices = realloc(devices, size);
+       if (!devices) {
+               fprintf(stderr, "Out of memory, device %s (%d)\n", name, size);
+               return 1;
+       }
+       memset(&devices[ndevices], 0, sizeof(struct per_dev_info));
+       devices[ndevices].name = name;
+       ndevices++;
+       return 0;
+}
+
+static struct per_dev_info *get_dev_info(dev_t id)
+{
+       struct per_dev_info *pdi;
+       int i;
+
+       for (i = 0; i < ndevices; i++) {
+               if (!devices[i].id)
+                       devices[i].id = id;
+               if (devices[i].id == id)
+                       return &devices[i];
+       }
+
+       if (resize_devices(NULL) != 0)
+               return NULL;
+
+       pdi = &devices[ndevices - 1];
+       pdi->id = id;
+       pdi->last_sequence = 0;
+       pdi->last_read_time = 0;
+       return pdi;
+}
+
+static char *get_dev_name(struct per_dev_info *pdi, char *buffer, int size)
+{
+       if (pdi->name)
+               snprintf(buffer, size, "%s", pdi->name);
+       else
+               snprintf(buffer, size, "%d,%d", MAJOR(pdi->id), MINOR(pdi->id));
+       return buffer;
+}
+
+static void check_time(struct per_dev_info *pdi, struct blk_io_trace *bit)
 {
        unsigned long long this = bit->time;
-       unsigned long long last = last_reported_time;
+       unsigned long long last = pdi->last_reported_time;
 
-       backwards = (this < last) ? 'B' : ' ';
-       last_reported_time = this;
+       pdi->backwards = (this < last) ? 'B' : ' ';
+       pdi->last_reported_time = this;
 }
 
-static inline void account_m(struct per_cpu_info *pci, int rw,
-                            unsigned int bytes)
+static inline void __account_m(struct io_stats *ios, struct blk_io_trace *t,
+                              int rw)
 {
        if (rw) {
-               pci->mwrites++;
-               pci->qwrite_kb += bytes >> 10;
+               ios->mwrites++;
+               ios->qwrite_kb += t->bytes >> 10;
        } else {
-               pci->mreads++;
-               pci->qread_kb += bytes >> 10;
+               ios->mreads++;
+               ios->qread_kb += t->bytes >> 10;
        }
 }
 
-static inline void account_q(struct per_cpu_info *pci, int rw,
-                            unsigned int bytes)
+static inline void account_m(struct blk_io_trace *t, struct per_cpu_info *pci,
+                            int rw)
 {
-       if (rw) {
-               pci->qwrites++;
-               pci->qwrite_kb += bytes >> 10;
-       } else {
-               pci->qreads++;
-               pci->qread_kb += bytes >> 10;
+       __account_m(&pci->io_stats, t, rw);
+
+       if (per_process_stats) {
+               struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+
+               __account_m(ios, t, rw);
        }
 }
 
-static inline void account_c(struct per_cpu_info *pci, int rw,
-                            unsigned int bytes)
+static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
+                                  int rw)
 {
        if (rw) {
-               pci->cwrites++;
-               pci->cwrite_kb += bytes >> 10;
+               ios->qwrites++;
+               ios->qwrite_kb += t->bytes >> 10;
        } else {
-               pci->creads++;
-               pci->cread_kb += bytes >> 10;
+               ios->qreads++;
+               ios->qread_kb += t->bytes >> 10;
        }
 }
 
-static inline void account_i(struct per_cpu_info *pci, int rw,
-                            unsigned int bytes)
+static inline void account_queue(struct blk_io_trace *t,
+                                struct per_cpu_info *pci, int rw)
+{
+       __account_queue(&pci->io_stats, t, rw);
+
+       if (per_process_stats) {
+               struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+
+               __account_queue(ios, t, rw);
+       }
+}
+
+static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
 {
        if (rw) {
-               pci->iwrites++;
-               pci->iwrite_kb += bytes >> 10;
+               ios->cwrites++;
+               ios->cwrite_kb += bytes >> 10;
        } else {
-               pci->ireads++;
-               pci->iread_kb += bytes >> 10;
+               ios->creads++;
+               ios->cread_kb += bytes >> 10;
        }
 }
 
-static void output(struct per_cpu_info *pci, char *s)
+static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
+                            int rw, int bytes)
 {
-       printf("%s", s);
+       __account_c(&pci->io_stats, rw, bytes);
 
-       if (pci->ofp)
-               fprintf(pci->ofp,"%s",s);
+       if (per_process_stats) {
+               struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+
+               __account_c(ios, rw, bytes);
+       }
 }
 
-static char hstring[256];
-static char tstring[256];
+static inline void __account_issue(struct io_stats *ios, int rw,
+                                  unsigned int bytes)
+{
+       if (rw) {
+               ios->iwrites++;
+               ios->iwrite_kb += bytes >> 10;
+       } else {
+               ios->ireads++;
+               ios->iread_kb += bytes >> 10;
+       }
+}
 
-static inline char *setup_header(struct per_cpu_info *pci,
-                                struct blk_io_trace *t, char act)
+static inline void account_issue(struct blk_io_trace *t,
+                                struct per_cpu_info *pci, int rw)
 {
-       int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
-       int b = t->action & BLK_TC_ACT(BLK_TC_BARRIER);
-       int s = t->action & BLK_TC_ACT(BLK_TC_SYNC);
-       char rwbs[4];
-       int i = 0;
+       __account_issue(&pci->io_stats, rw, t->bytes);
+
+       if (per_process_stats) {
+               struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+
+               __account_issue(ios, rw, t->bytes);
+       }
+}
 
-       if (w)
-               rwbs[i++] = 'W';
+static inline void __account_unplug(struct io_stats *ios, int timer)
+{
+       if (timer)
+               ios->timer_unplugs++;
        else
-               rwbs[i++] = 'R';
-       if (b)
-               rwbs[i++] = 'B';
-       if (s)
-               rwbs[i++] = 'S';
+               ios->io_unplugs++;
+}
 
-       rwbs[i] = '\0';
+static inline void account_unplug(struct blk_io_trace *t,
+                                 struct per_cpu_info *pci, int timer)
+{
+       __account_unplug(&pci->io_stats, timer);
 
-       sprintf(hstring, "%c %3d %15ld %5Lu.%09Lu %5u %c %3s", backwards,
-               pci->cpu,
-               (unsigned long)t->sequence, SECONDS(t->time), 
-               NANO_SECONDS(t->time), t->pid, act, rwbs);
+       if (per_process_stats) {
+               struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
 
-       return hstring;
+               __account_unplug(ios, timer);
+       }
 }
 
 static void log_complete(struct per_cpu_info *pci, struct blk_io_trace *t,
-                        char act)
+                        char *act)
 {
-       sprintf(tstring,"%s %Lu + %u [%d]\n", setup_header(pci, t, act),
-               (unsigned long long)t->sector, t->bytes >> 9, t->error);
-       output(pci, tstring);
+       process_fmt(act, pci, t, log_track_complete(t), 0, NULL);
+}
+
+static void log_insert(struct per_cpu_info *pci, struct blk_io_trace *t,
+                     char *act)
+{
+       process_fmt(act, pci, t, log_track_insert(t), 0, NULL);
 }
 
 static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
-                     char act)
+                     char *act)
 {
-       sprintf(tstring,"%s %Lu + %u\n", setup_header(pci, t, act),
-               (unsigned long long)t->sector, t->bytes >> 9);
-       output(pci, tstring);
+       process_fmt(act, pci, t, -1, 0, NULL);
 }
 
 static void log_issue(struct per_cpu_info *pci, struct blk_io_trace *t,
-                     char act)
+                     char *act)
 {
-       sprintf(tstring,"%s %Lu + %u\n", setup_header(pci, t, act),
-               (unsigned long long)t->sector, t->bytes >> 9);
-       output(pci, tstring);
+       process_fmt(act, pci, t, log_track_issue(t), 0, NULL);
 }
 
 static void log_merge(struct per_cpu_info *pci, struct blk_io_trace *t,
-                     char act)
+                     char *act)
 {
-       sprintf(tstring,"%s   %Lu + %u\n", setup_header(pci, t, act),
-               (unsigned long long)t->sector, t->bytes >> 9);
-       output(pci, tstring);
+       if (act[0] == 'F')
+               log_track_frontmerge(t);
+
+       process_fmt(act, pci, t, -1ULL, 0, NULL);
 }
 
-static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
-                       char act)
+static void log_action(struct per_cpu_info *pci, struct blk_io_trace *t,
+                       char *act)
 {
-       sprintf(tstring,"%s %Lu + %u\n", setup_header(pci, t, act),
-               (unsigned long long)t->sector, t->bytes >> 9);
-       output(pci, tstring);
+       process_fmt(act, pci, t, -1ULL, 0, NULL);
 }
 
-static int log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char act)
+static void log_generic(struct per_cpu_info *pci, struct blk_io_trace *t,
+                       char *act)
 {
-       unsigned char *buf;
-       int i;
+       process_fmt(act, pci, t, -1ULL, 0, NULL);
+}
 
-       sprintf(tstring,"%s ", setup_header(pci, t, act));
-       output(pci, tstring);
+static void log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
+                     char *act)
+{
+       process_fmt(act, pci, t, -1ULL, 0, NULL);
+}
 
-       buf = (unsigned char *) t + sizeof(*t);
-       for (i = 0; i < t->pdu_len; i++) {
-               sprintf(tstring,"%02x ", buf[i]);
-               output(pci, tstring);
-       }
+static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
+                     char *act)
+{
+       process_fmt(act, pci, t, -1ULL, 0, NULL);
+}
 
-       if (act == 'C') {
-               sprintf(tstring,"[%d]", t->error);
-               output(pci, tstring);
-       }
+static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
+{
+       unsigned char *buf = (unsigned char *) t + sizeof(*t);
 
-       printf("\n");
-       return 0;
+       process_fmt(act, pci, t, -1ULL, t->pdu_len, buf);
 }
 
-static int dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
+static void dump_trace_pc(struct blk_io_trace *t, struct per_cpu_info *pci)
 {
-       int ret = 0;
+       int act = t->action & 0xffff;
 
-       switch (t->action & 0xffff) {
+       switch (act) {
                case __BLK_TA_QUEUE:
-                       log_generic(pci, t, 'Q');
+                       log_generic(pci, t, "Q");
                        break;
                case __BLK_TA_GETRQ:
-                       log_generic(pci, t, 'G');
+                       log_generic(pci, t, "G");
                        break;
                case __BLK_TA_SLEEPRQ:
-                       log_generic(pci, t, 'S');
+                       log_generic(pci, t, "S");
                        break;
                case __BLK_TA_REQUEUE:
-                       log_generic(pci, t, 'R');
+                       log_generic(pci, t, "R");
                        break;
                case __BLK_TA_ISSUE:
-                       ret = log_pc(pci, t, 'D');
+                       log_pc(pci, t, "D");
                        break;
                case __BLK_TA_COMPLETE:
-                       log_pc(pci, t, 'C');
+                       log_pc(pci, t, "C");
+                       break;
+               case __BLK_TA_INSERT:
+                       log_pc(pci, t, "I");
                        break;
                default:
-                       fprintf(stderr, "Bad pc action %x\n", t->action);
-                       ret = 1;
+                       fprintf(stderr, "Bad pc action %x\n", act);
                        break;
        }
-       
-       return ret;
 }
 
 static void dump_trace_fs(struct blk_io_trace *t, struct per_cpu_info *pci)
 {
        int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
+       int act = t->action & 0xffff;
 
-       switch (t->action & 0xffff) {
+       switch (act) {
                case __BLK_TA_QUEUE:
-                       account_q(pci, w, t->bytes);
-                       log_queue(pci, t, 'Q');
+                       account_queue(t, pci, w);
+                       log_queue(pci, t, "Q");
+                       break;
+               case __BLK_TA_INSERT:
+                       log_insert(pci, t, "I");
                        break;
                case __BLK_TA_BACKMERGE:
-                       account_m(pci, w, t->bytes);
-                       log_merge(pci, t, 'M');
+                       account_m(t, pci, w);
+                       log_merge(pci, t, "M");
                        break;
                case __BLK_TA_FRONTMERGE:
-                       account_m(pci, w, t->bytes);
-                       log_merge(pci, t, 'F');
+                       account_m(t, pci, w);
+                       log_merge(pci, t, "F");
                        break;
                case __BLK_TA_GETRQ:
-                       log_generic(pci, t, 'G');
+                       log_track_getrq(t);
+                       log_generic(pci, t, "G");
                        break;
                case __BLK_TA_SLEEPRQ:
-                       log_generic(pci, t, 'S');
+                       log_generic(pci, t, "S");
                        break;
                case __BLK_TA_REQUEUE:
-                       account_c(pci, w, -t->bytes);
-                       log_queue(pci, t, 'R');
+                       account_c(t, pci, w, -t->bytes);
+                       log_queue(pci, t, "R");
                        break;
                case __BLK_TA_ISSUE:
-                       account_i(pci, w, t->bytes);
-                       log_issue(pci, t, 'D');
+                       account_issue(t, pci, w);
+                       log_issue(pci, t, "D");
                        break;
                case __BLK_TA_COMPLETE:
-                       account_c(pci, w, t->bytes);
-                       log_complete(pci, t, 'C');
+                       account_c(t, pci, w, t->bytes);
+                       log_complete(pci, t, "C");
+                       break;
+               case __BLK_TA_PLUG:
+                       log_action(pci, t, "P");
+                       break;
+               case __BLK_TA_UNPLUG_IO:
+                       account_unplug(t, pci, 0);
+                       log_unplug(pci, t, "U");
+                       break;
+               case __BLK_TA_UNPLUG_TIMER:
+                       account_unplug(t, pci, 1);
+                       log_unplug(pci, t, "UT");
+                       break;
+               case __BLK_TA_SPLIT:
+                       log_split(pci, t, "X");
+                       break;
+               case __BLK_TA_BOUNCE:
+                       log_generic(pci, t, "B");
                        break;
                default:
                        fprintf(stderr, "Bad fs action %x\n", t->action);
-                       return;
+                       break;
        }
 }
 
-static int dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci)
+static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
+                      struct per_dev_info *pdi)
 {
-       int ret = 0;
-
        if (t->action & BLK_TC_ACT(BLK_TC_PC))
-               ret = dump_trace_pc(t, pci);
+               dump_trace_pc(t, pci);
        else
                dump_trace_fs(t, pci);
 
-       events++;
-       return ret;
+       pdi->events++;
 }
 
-static void dump_pci_stats(struct per_cpu_info *pci)
+static void dump_io_stats(struct io_stats *ios, char *msg)
 {
-       printf("\tReads:\n");
-       printf("\t\tQueued:    %'8lu, %'8LuKiB\n", pci->qreads, pci->qread_kb);
-       printf("\t\tDispatched %'8lu, %'8LuKiB\n", pci->ireads, pci->iread_kb);
-       printf("\t\tCompleted: %'8lu, %'8LuKiB\n", pci->creads, pci->cread_kb);
-       printf("\t\tMerges:    %'8lu\n", pci->mreads);
+       fprintf(ofp, "%s\n", msg);
+
+       fprintf(ofp, " Reads Queued:    %'8lu, %'8LuKiB\t", ios->qreads, ios->qread_kb);
+       fprintf(ofp, " Writes Queued:    %'8lu, %'8LuKiB\n", ios->qwrites,ios->qwrite_kb);
+
+       fprintf(ofp, " Read Dispatches: %'8lu, %'8LuKiB\t", ios->ireads, ios->iread_kb);
+       fprintf(ofp, " Write Dispatches: %'8lu, %'8LuKiB\n", ios->iwrites,ios->iwrite_kb);
+       fprintf(ofp, " Reads Completed: %'8lu, %'8LuKiB\t", ios->creads, ios->cread_kb);
+       fprintf(ofp, " Writes Completed: %'8lu, %'8LuKiB\n", ios->cwrites,ios->cwrite_kb);
+       fprintf(ofp, " Read Merges:     %'8lu%8c\t", ios->mreads, ' ');
+       fprintf(ofp, " Write Merges:     %'8lu\n", ios->mwrites);
+       fprintf(ofp, " IO unplugs:      %'8lu%8c\t", ios->io_unplugs, ' ');
+       fprintf(ofp, " Timer unplugs:    %'8lu\n", ios->timer_unplugs);
+}
 
-       printf("\tWrites:\n");
-       printf("\t\tQueued:    %'8lu, %'8LuKiB\n", pci->qwrites,pci->qwrite_kb);
-       printf("\t\tDispatched %'8lu, %'8LuKiB\n", pci->iwrites,pci->iwrite_kb);
-       printf("\t\tCompleted: %'8lu, %'8LuKiB\n", pci->cwrites,pci->cwrite_kb);
-       printf("\t\tMerges:    %'8lu\n", pci->mwrites);
+static void dump_wait_stats(struct per_process_info *ppi)
+{
+       unsigned long rawait = ppi->longest_allocation_wait[0] / 1000;
+       unsigned long rdwait = ppi->longest_dispatch_wait[0] / 1000;
+       unsigned long rcwait = ppi->longest_completion_wait[0] / 1000;
+       unsigned long wawait = ppi->longest_allocation_wait[1] / 1000;
+       unsigned long wdwait = ppi->longest_dispatch_wait[1] / 1000;
+       unsigned long wcwait = ppi->longest_completion_wait[1] / 1000;
+
+       fprintf(ofp, " Allocation wait: %'8lu%8c\t", rawait, ' ');
+       fprintf(ofp, " Allocation wait:  %'8lu\n", wawait);
+       fprintf(ofp, " Dispatch wait:   %'8lu%8c\t", rdwait, ' ');
+       fprintf(ofp, " Dispatch wait:    %'8lu\n", wdwait);
+       fprintf(ofp, " Completion wait: %'8lu%8c\t", rcwait, ' ');
+       fprintf(ofp, " Completion wait:  %'8lu\n", wcwait);
 }
 
-static void show_stats(void)
+static int ppi_name_compare(const void *p1, const void *p2)
 {
-       struct per_cpu_info foo, *pci;
-       int i, pci_events = 0;
+       struct per_process_info *ppi1 = *((struct per_process_info **) p1);
+       struct per_process_info *ppi2 = *((struct per_process_info **) p2);
+       int res;
 
-       memset(&foo, 0, sizeof(foo));
+       res = strverscmp(ppi1->name, ppi2->name);
+       if (!res)
+               res = ppi1->pid > ppi2->pid;
 
-       for (i = 0; i < MAX_CPUS; i++) {
-               pci = &per_cpu_info[i];
+       return res;
+}
 
-               if (!pci->nelems)
-                       continue;
+static void sort_process_list(void)
+{
+       struct per_process_info **ppis;
+       struct per_process_info *ppi;
+       int i = 0;
 
-               foo.qreads += pci->qreads;
-               foo.qwrites += pci->qwrites;
-               foo.creads += pci->creads;
-               foo.cwrites += pci->cwrites;
-               foo.mreads += pci->mreads;
-               foo.mwrites += pci->mwrites;
-               foo.qread_kb += pci->qread_kb;
-               foo.qwrite_kb += pci->qwrite_kb;
-               foo.cread_kb += pci->cread_kb;
-               foo.cwrite_kb += pci->cwrite_kb;
+       ppis = malloc(ppi_list_entries * sizeof(struct per_process_info *));
 
-               printf("CPU%d:\n", i);
-               dump_pci_stats(pci);
-               pci_events++;
+       ppi = ppi_list;
+       while (ppi) {
+               ppis[i++] = ppi;
+               ppi = ppi->list_next;
        }
 
-       if (pci_events > 1) {
-               printf("Total:\n");
-               dump_pci_stats(&foo);
+       qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_name_compare);
+
+       i = ppi_list_entries - 1;
+       ppi_list = NULL;
+       while (i >= 0) {
+               ppi = ppis[i];
+
+               ppi->list_next = ppi_list;
+               ppi_list = ppi;
+               i--;
        }
 
-       printf("Events: %'Lu\n", events);
+       free(ppis);
 }
 
-static inline int trace_rb_insert(struct trace *t)
+static void show_process_stats(void)
 {
-       struct rb_node **p = &rb_root.rb_node;
-       struct rb_node *parent = NULL;
-       struct trace *__t;
+       struct per_process_info *ppi;
 
-       while (*p) {
-               parent = *p;
-               __t = rb_entry(parent, struct trace, rb_node);
+       sort_process_list();
 
-               if (t->bit->sequence < __t->bit->sequence)
-                       p = &(*p)->rb_left;
-               else if (t->bit->sequence > __t->bit->sequence)
-                       p = &(*p)->rb_right;
-               else {
-                       fprintf(stderr, "sequence alias!\n");
-                       return 1;
-               }
+       ppi = ppi_list;
+       while (ppi) {
+               char name[64];
+
+               if (ppi->more_than_one)
+                       sprintf(name, "%s (%u, ...)", ppi->name, ppi->pid);
+               else
+                       sprintf(name, "%s (%u)", ppi->name, ppi->pid);
+
+               dump_io_stats(&ppi->io_stats, name);
+               dump_wait_stats(ppi);
+               ppi = ppi->list_next;
        }
 
-       rb_link_node(&t->rb_node, parent, p);
-       rb_insert_color(&t->rb_node, &rb_root);
-       return 0;
+       fprintf(ofp, "\n");
 }
 
-static int sort_entries(void *traces, unsigned long offset, int nr)
+static void show_device_and_cpu_stats(void)
 {
+       struct per_dev_info *pdi;
        struct per_cpu_info *pci;
-       struct blk_io_trace *bit;
-       struct trace *t;
-       void *start = traces;
-       int nelems = 0;
-
-       while (traces - start <= offset - sizeof(*bit)) {
-               if (!nr)
-                       break;
-
-               bit = traces;
+       struct io_stats total, *ios;
+       int i, j, pci_events;
+       char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
+       char name[32];
+
+       for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
+
+               memset(&total, 0, sizeof(total));
+               pci_events = 0;
+
+               if (i > 0)
+                       fprintf(ofp, "\n");
+
+               for (pci = pdi->cpus, j = 0; j < pdi->ncpus; j++, pci++) {
+                       if (!pci->nelems)
+                               continue;
+
+                       ios = &pci->io_stats;
+                       total.qreads += ios->qreads;
+                       total.qwrites += ios->qwrites;
+                       total.creads += ios->creads;
+                       total.cwrites += ios->cwrites;
+                       total.mreads += ios->mreads;
+                       total.mwrites += ios->mwrites;
+                       total.ireads += ios->ireads;
+                       total.iwrites += ios->iwrites;
+                       total.qread_kb += ios->qread_kb;
+                       total.qwrite_kb += ios->qwrite_kb;
+                       total.cread_kb += ios->cread_kb;
+                       total.cwrite_kb += ios->cwrite_kb;
+                       total.iread_kb += ios->iread_kb;
+                       total.iwrite_kb += ios->iwrite_kb;
+                       total.timer_unplugs += ios->timer_unplugs;
+                       total.io_unplugs += ios->io_unplugs;
+
+                       snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
+                                j, get_dev_name(pdi, name, sizeof(name)));
+                       dump_io_stats(ios, line);
+                       pci_events++;
+               }
 
-               t = malloc(sizeof(*t));
-               t->bit = bit;
-               memset(&t->rb_node, 0, sizeof(t->rb_node));
+               if (pci_events > 1) {
+                       fprintf(ofp, "\n");
+                       snprintf(line, sizeof(line) - 1, "Total (%s):",
+                                get_dev_name(pdi, name, sizeof(name)));
+                       dump_io_stats(&total, line);
+               }
 
-               trace_to_cpu(bit);
+               fprintf(ofp, "\nEvents (%s): %'Lu entries, %'lu skips\n",
+                       get_dev_name(pdi, line, sizeof(line)), pdi->events,
+                       pdi->skips);
+       }
+}
 
-               if (verify_trace(bit))
-                       break;
+/*
+ * struct trace and blktrace allocation cache, we do potentially
+ * millions of mallocs for these structures while only using at most
+ * a few thousand at the time
+ */
+static inline void t_free(struct trace *t)
+{
+       t->next = t_alloc_list;
+       t_alloc_list = t;
+}
 
-               pci = &per_cpu_info[bit->cpu];
+static inline struct trace *t_alloc(void)
+{
+       struct trace *t = t_alloc_list;
 
-               if (output_name && !pci->ofp) {
-                       snprintf(pci->ofname, sizeof(pci->ofname) - 1,
-                                       "%s_log.%d", output_name, bit->cpu);
+       if (t) {
+               t_alloc_list = t->next;
+               return t;
+       }
 
-                       pci->ofp = fopen(pci->ofname, "w");
-                       if (pci->ofp == NULL) {
-                               perror(pci->ofname);
-                               break;
-                       }
-               }
+       return malloc(sizeof(*t));
+}
 
-               pci->nelems++;
+static inline void bit_free(struct blk_io_trace *bit)
+{
+       /*
+        * abuse a 64-bit field for a next pointer for the free item
+        */
+       bit->time = (__u64) (unsigned long) bit_alloc_list;
+       bit_alloc_list = (struct blk_io_trace *) bit;
+}
 
-               if (trace_rb_insert(t))
-                       return -1;
+static inline struct blk_io_trace *bit_alloc(void)
+{
+       struct blk_io_trace *bit = bit_alloc_list;
 
-               traces += sizeof(*bit) + bit->pdu_len;
-               nelems++;
-               nr--;
+       if (bit) {
+               bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
+                                bit->time;
+               return bit;
        }
 
-       return nelems;
+       return malloc(sizeof(*bit));
 }
 
-static void free_entries_rb(void)
+static void find_genesis(void)
 {
-       struct rb_node *n;
+       struct trace *t = trace_list;
 
-       while ((n = rb_first(&rb_root)) != NULL) {
-               struct trace *t = rb_entry(n, struct trace, rb_node);
+       genesis_time = -1ULL;
+       while (t != NULL) {
+               if (t->bit->time < genesis_time)
+                       genesis_time = t->bit->time;
 
-               rb_erase(&t->rb_node, &rb_root);
-               free(t);
+               t = t->next;
        }
 }
 
-static void show_entries_rb(void)
+static inline int check_stopwatch(struct blk_io_trace *bit)
+{
+       if (bit->time < stopwatch_end &&
+           bit->time >= stopwatch_start)
+               return 0;
+
+       return 1;
+}
+
+static int sort_entries(void)
 {
-       struct blk_io_trace *bit;
-       struct rb_node *n;
        struct trace *t;
-       int cpu;
+       int nr = 0;
 
-       n = rb_first(&rb_root);
-       if (!n)
-               return;
+       if (!genesis_time)
+               find_genesis();
 
-       do {
-               t = rb_entry(n, struct trace, rb_node);
-               bit = t->bit;
+       while ((t = trace_list) != NULL) {
+               struct blk_io_trace *bit = t->bit;
 
-               cpu = bit->cpu;
-               if (cpu > max_cpus) {
-                       fprintf(stderr, "CPU number too large (%d)\n", cpu);
-                       break;
-               }
+               trace_list = t->next;
+
+               if (verify_trace(bit))
+                       continue;
 
-               if (genesis_time == 0)
-                       genesis_time = bit->time;
                bit->time -= genesis_time;
 
-               check_time(bit);
+               if (check_stopwatch(bit)) {
+                       bit_free(bit);
+                       t_free(t);
+                       continue;
+               }
 
-               if (dump_trace(bit, &per_cpu_info[cpu]))
+               if (trace_rb_insert(t))
                        break;
 
-       } while ((n = rb_next(n)) != NULL);
+               nr++;
+       }
+
+       return nr;
 }
 
-static int do_file(void)
+static void show_entries_rb(int force)
 {
-       int i, ret;
-
-       for (max_cpus = 0, i = 0; i < MAX_CPUS; i++, nfiles++, max_cpus++) {
-               struct per_cpu_info *pci = &per_cpu_info[i];
-               struct stat st;
-               void *tb;
+       struct per_dev_info *pdi = NULL;
+       struct per_cpu_info *pci = NULL;
+       struct blk_io_trace *bit;
+       struct rb_node *n;
+       struct trace *t;
 
-               pci->cpu = i;
-               pci->ofp = NULL;
+       while ((n = rb_first(&rb_sort_root)) != NULL) {
 
-               snprintf(pci->fname, sizeof(pci->fname)-1,"%s_out.%d", dev, i);
-               if (stat(pci->fname, &st) < 0)
+               if (done)
                        break;
-               if (!st.st_size)
-                       continue;
 
-               printf("Processing %s\n", pci->fname);
+               t = rb_entry(n, struct trace, rb_node);
+               bit = t->bit;
 
-               tb = malloc(st.st_size);
+               if (!pdi || pdi->id != bit->device)
+                       pdi = get_dev_info(bit->device);
 
-               pci->fd = open(pci->fname, O_RDONLY);
-               if (pci->fd < 0) {
-                       perror(pci->fname);
+               if (!pdi) {
+                       fprintf(stderr, "Unknown device ID? (%d,%d)\n",
+                               MAJOR(bit->device), MINOR(bit->device));
                        break;
                }
 
-               if (read(pci->fd, tb, st.st_size) != st.st_size) {
-                       fprintf(stderr, "error reading\n");
+               if (bit->cpu > pdi->ncpus) {
+                       fprintf(stderr, "Unknown CPU ID? (%d, device %d,%d)\n",
+                               bit->cpu, MAJOR(bit->device),
+                               MINOR(bit->device));
                        break;
                }
 
-               ret = sort_entries(tb, st.st_size, ~0U);
-               if (ret == -1)
+               /*
+                * back off displaying more info if we are out of sync
+                * on SMP systems. to prevent stalling on lost events,
+                * only allow an event to skip us a few times
+                */
+               if (bit->sequence > (pdi->last_sequence + 1) && !force) {
+                       struct trace *__t;
+
+                       /*
+                        * the wanted sequence is really there, continue
+                        * because this means that the log time is earlier
+                        * on the trace we have now
+                        */
+                       __t = trace_rb_find(pdi->id, pdi->last_sequence + 1);
+                       if (__t)
+                               goto ok;
+
+                       if (t->skipped < 5) {
+                               t->skipped++;
+                               break;
+                       } else
+                               pdi->skips++;
+               }
+
+ok:
+               if (!force && bit->time > last_allowed_time)
                        break;
 
-               close(pci->fd);
-               printf("\t%2d %10s %15d\n", i, pci->fname, pci->nelems);
+               pdi->last_sequence = bit->sequence;
 
-       }
+               check_time(pdi, bit);
 
-       if (!nfiles) {
-               fprintf(stderr, "No files found\n");
-               return 1;
-       }
+               if (!pci || pci->cpu != bit->cpu)
+                       pci = get_cpu_info(pdi, bit->cpu);
 
-       show_entries_rb();
-       return 0;
+               dump_trace(bit, pci, pdi);
+
+               rb_erase(&t->rb_node, &rb_sort_root);
+               rb_sort_entries--;
+               bit_free(bit);
+               t_free(t);
+       }
 }
 
 static int read_data(int fd, void *buffer, int bytes, int block)
@@ -585,6 +1306,7 @@ static int read_data(int fd, void *buffer, int bytes, int block)
                else if (ret < 0) {
                        if (errno != EAGAIN)
                                perror("read");
+
                        return -1;
                } else {
                        p += ret;
@@ -595,141 +1317,348 @@ static int read_data(int fd, void *buffer, int bytes, int block)
        return 0;
 }
 
-static void resize_buffer(void **buffer, long *old_size)
+static int read_events(int fd, int always_block)
 {
-       long cur_size = *old_size;
-       void *ptr;
+       struct per_dev_info *pdi = NULL;
+       int events = 0;
 
-       *old_size *= 2;
-       ptr = malloc(*old_size);
-       memcpy(ptr, *buffer, cur_size);
-       free(*buffer);
-       *buffer = ptr;
-}
+       while (!is_done() && events < rb_batch) {
+               struct blk_io_trace *bit;
+               struct trace *t;
+               int pdu_len;
+               __u32 magic;
 
-static int read_sort_events(int fd, void **buffer)
-{
-       long offset, max_offset;
-       int events;
+               bit = bit_alloc();
 
-       max_offset = 128 * sizeof(struct blk_io_trace);
-       *buffer = malloc(max_offset);
-       events = 0;
-       offset = 0;
+               if (read_data(fd, bit, sizeof(*bit), !events || always_block))
+                       break;
 
-       do {
-               struct blk_io_trace *t;
-               int pdu_len;
+               magic = be32_to_cpu(bit->magic);
+               if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
+                       fprintf(stderr, "Bad magic %x\n", magic);
+                       break;
+               }
 
-               if (max_offset - offset < sizeof(*t))
-                       resize_buffer(buffer, &max_offset);
+               pdu_len = be16_to_cpu(bit->pdu_len);
+               if (pdu_len) {
+                       void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
 
-               if (read_data(fd, *buffer + offset, sizeof(*t), !events)) {
-                       if (events)
+                       if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1))
                                break;
 
-                       usleep(1000);
-                       continue;
+                       bit = ptr;
                }
 
-               t = *buffer + offset;
-               offset += sizeof(*t);
+               t = t_alloc();
+               memset(t, 0, sizeof(*t));
+               t->bit = bit;
 
-               pdu_len = be16_to_cpu(t->pdu_len);
+               trace_to_cpu(bit);
 
-               if (max_offset - offset < pdu_len)
-                       resize_buffer(buffer, &max_offset);
+               t->next = trace_list;
+               trace_list = t;
 
-               if (read_data(fd, *buffer + offset, pdu_len, 1))
-                       break;
+               if (!pdi || pdi->id != bit->device)
+                       pdi = get_dev_info(bit->device);
+
+               if (bit->time > pdi->last_read_time)
+                       pdi->last_read_time = bit->time;
 
-               offset += pdu_len;
                events++;
-       } while (!is_done());
+       }
 
        return events;
 }
 
+static int do_file(void)
+{
+       struct per_cpu_info *pci;
+       struct per_dev_info *pdi;
+       int i, j, events, events_added;
+
+       /*
+        * first prepare all files for reading
+        */
+       for (i = 0; i < ndevices; i++) {
+               pdi = &devices[i];
+               pdi->nfiles = 0;
+               pdi->last_sequence = 0;
+
+               for (j = 0;; j++) {
+                       struct stat st;
+
+                       pci = get_cpu_info(pdi, j);
+                       pci->cpu = j;
+                       pci->fd = -1;
+
+                       snprintf(pci->fname, sizeof(pci->fname)-1,
+                                "%s.blktrace.%d", pdi->name, pci->cpu);
+                       if (stat(pci->fname, &st) < 0)
+                               break;
+                       if (st.st_size) {
+                               pci->fd = open(pci->fname, O_RDONLY);
+                               if (pci->fd < 0) {
+                                       perror(pci->fname);
+                                       continue;
+                               }
+                       }
+
+                       printf("Input file %s added\n", pci->fname);
+                       pdi->nfiles++;
+               }
+       }
+
+       /*
+        * now loop over the files reading in the data
+        */
+       do {
+               events_added = 0;
+               last_allowed_time = -1ULL;
+
+               for (i = 0; i < ndevices; i++) {
+                       pdi = &devices[i];
+
+                       for (j = 0; j < pdi->nfiles; j++) {
+
+                               pci = get_cpu_info(pdi, j);
+
+                               if (pci->fd == -1)
+                                       continue;
+
+                               events = read_events(pci->fd, 1);
+                               if (!events) {
+                                       close(pci->fd);
+                                       pci->fd = -1;
+                                       continue;
+                               }
+
+                               if (pdi->last_read_time < last_allowed_time)
+                                       last_allowed_time = pdi->last_read_time;
+
+                               events_added += events;
+                       }
+               }
+
+               if (sort_entries() == -1)
+                       break;
+
+               show_entries_rb(0);
+
+       } while (events_added);
+
+       if (rb_sort_entries)
+               show_entries_rb(1);
+
+       return 0;
+}
+
 static int do_stdin(void)
 {
        int fd;
-       void *ptr;
 
-       fd = dup(0);
+       last_allowed_time = -1ULL;
+       fd = dup(STDIN_FILENO);
        do {
                int events;
 
-               events = read_sort_events(fd, &ptr);
+               events = read_events(fd, 0);
                if (!events)
                        break;
        
-               sort_entries(ptr, ~0UL, events);
-               show_entries_rb();
-               free_entries_rb();
+               if (sort_entries() == -1)
+                       break;
+
+               show_entries_rb(0);
        } while (1);
 
+       if (rb_sort_entries)
+               show_entries_rb(1);
+
        close(fd);
-       free(ptr);
        return 0;
 }
 
-void flush_output(void)
+static void flush_output(void)
 {
-       int i;
+       fflush(ofp);
+}
 
-       for (i = 0; i < MAX_CPUS; i++) {
-               struct per_cpu_info *pci = &per_cpu_info[i];
+static void handle_sigint(int sig)
+{
+       done = 1;
+       flush_output();
+}
 
-               if (pci->ofp) {
-                       fflush(pci->ofp);
-                       fclose(pci->ofp);
-                       pci->ofp = NULL;
+/*
+ * Extract start and duration times from a string, allowing
+ * us to specify a time interval of interest within a trace.
+ * Format: "duration" (start is zero) or "start:duration".
+ */
+static int find_stopwatch_interval(char *string)
+{
+       double value;
+       char *sp;
+
+       value = strtod(string, &sp);
+       if (sp == string) {
+               fprintf(stderr,"Invalid stopwatch timer: %s\n", string);
+               return 1;
+       }
+       if (*sp == ':') {
+               stopwatch_start = DOUBLE_TO_NANO_ULL(value);
+               string = sp + 1;
+               value = strtod(string, &sp);
+               if (sp == string || *sp != '\0') {
+                       fprintf(stderr,"Invalid stopwatch duration time: %s\n",
+                               string);
+                       return 1;
                }
+       } else if (*sp != '\0') {
+               fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
+               return 1;
        }
+       stopwatch_end = DOUBLE_TO_NANO_ULL(value);
+       if (stopwatch_end <= stopwatch_start) {
+               fprintf(stderr, "Invalid stopwatch interval: %Lu -> %Lu\n",
+                       stopwatch_start, stopwatch_end);
+               return 1;
+       }
+
+       return 0;
 }
 
-void handle_sigint(int sig)
+static char usage_str[] = \
+       "[ -i <input name> ] [-o <output name> [ -s ] [ -t ] [ -q ]\n" \
+       "[ -w start:stop ] [ -f output format ] [ -F format spec ] [ -v] \n\n" \
+       "\t-i Input file containing trace data, or '-' for stdin\n" \
+       "\t-o Output file. If not given, output is stdout\n" \
+       "\t-b stdin read batching\n" \
+       "\t-s Show per-program io statistics\n" \
+       "\t-n Hash processes by name, not pid\n" \
+       "\t-t Track individual ios. Will tell you the time a request took\n" \
+       "\t   to get queued, to get dispatched, and to get completed\n" \
+       "\t-q Quiet. Don't display any stats at the end of the trace\n" \
+       "\t-w Only parse data between the given time interval in seconds.\n" \
+       "\t   If 'start' isn't given, blkparse defaults the start time to 0\n" \
+       "\t -f Output format. Customize the output format. The format field\n" \
+       "\t    identifies can be found in the documentation\n" \
+       "\t-F Format specification. Can be found in the documentation\n" \
+       "\t-v Print program version info\n\n";
+
+static void usage(char *prog)
 {
-       done = 1;
-       flush_output();
+       fprintf(stderr, "Usage: %s %s %s", prog, blkparse_version, usage_str);
 }
 
 int main(int argc, char *argv[])
 {
-       int c, ret;
+       char *ofp_buffer;
+       int c, ret, mode;
+       int per_device_and_cpu_stats = 1;
 
        while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
                switch (c) {
                case 'i':
-                       dev = strdup(optarg);
+                       if (!strcmp(optarg, "-") && !pipeline)
+                               pipeline = 1;
+                       else if (resize_devices(optarg) != 0)
+                               return 1;
                        break;
                case 'o':
-                       output_name = strdup(optarg);
+                       output_name = optarg;
+                       break;
+               case 'b':
+                       rb_batch = atoi(optarg);
+                       if (rb_batch <= 0)
+                               rb_batch = RB_BATCH_DEFAULT;
+                       break;
+               case 's':
+                       per_process_stats = 1;
                        break;
+               case 't':
+                       track_ios = 1;
+                       break;
+               case 'q':
+                       per_device_and_cpu_stats = 0;
+                       break;
+               case 'w':
+                       if (find_stopwatch_interval(optarg) != 0)
+                               return 1;
+                       break;
+               case 'f':
+                       set_all_format_specs(optarg);
+                       break;
+               case 'F':
+                       if (add_format_spec(optarg) != 0)
+                               return 1;
+                       break;
+               case 'n':
+                       ppi_hash_by_pid = 0;
+                       break;
+               case 'v':
+                       printf("%s version %s\n", argv[0], blkparse_version);
+                       return 0;
                default:
-                       fprintf(stderr, "Usage: %s -i <dev>\n", argv[0]);
+                       usage(argv[0]);
                        return 1;
                }
        }
 
-       if (!dev) {
-               fprintf(stderr, "Usage: %s -i <dev>\n", argv[0]);
+       while (optind < argc) {
+               if (!strcmp(argv[optind], "-") && !pipeline)
+                       pipeline = 1;
+               else if (resize_devices(argv[optind]) != 0)
+                       return 1;
+               optind++;
+       }
+
+       if (!pipeline && !ndevices) {
+               usage(argv[0]);
                return 1;
        }
 
-       memset(per_cpu_info, 0, sizeof(per_cpu_info));
-       memset(&rb_root, 0, sizeof(rb_root));
+       memset(&rb_sort_root, 0, sizeof(rb_sort_root));
+       memset(&rb_track_root, 0, sizeof(rb_track_root));
 
        signal(SIGINT, handle_sigint);
        signal(SIGHUP, handle_sigint);
        signal(SIGTERM, handle_sigint);
 
-       if (!strcmp(dev, "-"))
+       setlocale(LC_NUMERIC, "en_US");
+
+       if (!output_name) {
+               ofp = fdopen(STDOUT_FILENO, "w");
+               mode = _IOLBF;
+       } else {
+               char ofname[128];
+
+               snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
+               ofp = fopen(ofname, "w");
+               mode = _IOFBF;
+       }
+
+       if (!ofp) {
+               perror("fopen");
+               return 1;
+       }
+
+       ofp_buffer = malloc(4096);      
+       if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
+               perror("setvbuf");
+               return 1;
+       }
+
+       if (pipeline)
                ret = do_stdin();
        else
                ret = do_file();
 
-       show_stats();
+       if (per_process_stats)
+               show_process_stats();
+
+       if (per_device_and_cpu_stats)
+               show_device_and_cpu_stats();
+
        flush_output();
        return ret;
 }