#include <errno.h>
#include <signal.h>
#include <locale.h>
+#include <limits.h>
#include "blktrace.h"
#include "rbtree.h"
-#define SECONDS(x) ((unsigned long long)(x) / 1000000000)
-#define NANO_SECONDS(x) ((unsigned long long)(x) % 1000000000)
+#define SECONDS(x) ((unsigned long long)(x) / 1000000000)
+#define NANO_SECONDS(x) ((unsigned long long)(x) % 1000000000)
+#define DOUBLE_TO_NANO_ULL(d) ((unsigned long long)((d) * 1000000000))
-static int backwards;
-static unsigned long long genesis_time, last_reported_time;
+#define MINORBITS 20
+#define MINORMASK ((1U << MINORBITS) - 1)
+#define MAJOR(dev) ((unsigned int) ((dev) >> MINORBITS))
+#define MINOR(dev) ((unsigned int) ((dev) & MINORMASK))
+
+#define min(a, b) ((a) < (b) ? (a) : (b))
+
+struct io_stats {
+ 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_cpu_info {
int cpu;
int fd;
char fname[128];
- FILE *ofp;
- char ofname[128];
+ struct io_stats io_stats;
+};
- 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_dev_info {
+ dev_t id;
+ char *name;
+
+ int backwards;
+ unsigned long long events;
+ unsigned long long last_reported_time;
+ struct io_stats io_stats;
+
+ int ncpus;
+ struct per_cpu_info *cpus;
+};
+
+struct per_process_info {
+ char name[16];
+ __u32 pid;
+ struct io_stats io_stats;
+ struct per_process_info *hash_next, *list_next;
+
+ /*
+ * 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:b:"
+#define PPI_HASH_SHIFT (8)
+static struct per_process_info *ppi_hash[1 << PPI_HASH_SHIFT];
+static struct per_process_info *ppi_list;
+
+#define S_OPTS "i:o:b:stqw:"
static struct option l_opts[] = {
{
.name = "input",
.flag = NULL,
.val = 'b'
},
+ {
+ .name = "per program stats",
+ .has_arg = 0,
+ .flag = NULL,
+ .val = 's'
+ },
+ {
+ .name = "track ios",
+ .has_arg = 0,
+ .flag = NULL,
+ .val = 't'
+ },
+ {
+ .name = "quiet",
+ .has_arg = 0,
+ .flag = NULL,
+ .val = 'q'
+ },
+ {
+ .name = "stopwatch",
+ .has_arg = 1,
+ .flag = NULL,
+ .val = 'w'
+ },
{
.name = NULL,
.has_arg = 0,
}
};
-static struct rb_root rb_root;
+static struct rb_root rb_sort_root;
+static struct rb_root rb_track_root;
+/*
+ * for sorting the displayed output
+ */
struct trace {
struct blk_io_trace *bit;
struct rb_node rb_node;
};
-static int max_cpus;
-static struct per_cpu_info *per_cpu_info;
+/*
+ * for tracking individual ios
+ */
+struct io_track {
+ struct rb_node rb_node;
-static unsigned long long events;
+ dev_t device;
+ __u64 sector;
+ __u32 pid;
+ unsigned long long allocation_time;
+ unsigned long long queue_time;
+ unsigned long long dispatch_time;
+ unsigned long long completion_time;
+};
+
+static int ndevices;
+static struct per_dev_info *devices;
+static char *get_dev_name(struct per_dev_info *, char *, int);
+
+static FILE *ofp;
+static char *output_name;
-static char *dev, *output_name;
+static unsigned long long genesis_time;
+static unsigned long long stopwatch_start; /* start from zero by default */
+static unsigned long long stopwatch_end = ULONG_LONG_MAX; /* "infinity" */
+
+static int per_process_stats;
+static int track_ios;
#define RB_BATCH_DEFAULT (1024)
static int rb_batch = RB_BATCH_DEFAULT;
+static int pipeline;
+
#define is_done() (*(volatile int *)(&done))
static volatile int done;
-static void resize_cpu_info(int cpuid)
+static inline unsigned long hash_long(unsigned long val)
{
- int new_space, new_max = cpuid + 1;
+#if __WORDSIZE == 32
+ val *= 0x9e370001UL;
+#elif __WORDSIZE == 64
+ val *= 0x9e37fffffffc0001UL;
+#else
+#error unknown word size
+#endif
+
+ return val >> (__WORDSIZE - PPI_HASH_SHIFT);
+}
+
+static inline void add_process_to_hash(struct per_process_info *ppi)
+{
+ const int hash_idx = hash_long(ppi->pid);
+
+ ppi->hash_next = ppi_hash[hash_idx];
+ ppi_hash[hash_idx] = ppi;
+}
+
+static inline void add_process_to_list(struct per_process_info *ppi)
+{
+ ppi->list_next = ppi_list;
+ ppi_list = ppi;
+}
+
+static struct per_process_info *find_process_by_pid(__u32 pid)
+{
+ const int hash_idx = hash_long(pid);
+ struct per_process_info *ppi;
+
+ ppi = ppi_hash[hash_idx];
+ while (ppi) {
+ if (ppi->pid == pid)
+ return ppi;
+
+ ppi = ppi->hash_next;
+ }
+
+ return NULL;
+}
+
+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;
+
+ if (genesis_time == 0 || t->bit->time < genesis_time)
+ genesis_time = t->bit->time;
+
+ 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_link_node(&t->rb_node, parent, p);
+ rb_insert_color(&t->rb_node, &rb_sort_root);
+ return 0;
+}
+
+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 (%llu) on device %d,%d!\n",
+ 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, dev_t device, __u64 sector)
+{
+ struct io_track *iot;
+
+ iot = __find_track(device, sector);
+ if (!iot) {
+ iot = malloc(sizeof(*iot));
+ iot->pid = pid;
+ iot->device = device;
+ iot->sector = sector;
+ track_rb_insert(iot);
+ }
+
+ return iot;
+}
+
+static void log_track_merge(struct blk_io_trace *t)
+{
+ struct io_track *iot;
+
+ if (!track_ios)
+ return;
+ if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
+ return;
+
+ iot = __find_track(t->device, t->sector - (t->bytes >> 10));
+ if (!iot) {
+ fprintf(stderr, "Trying to merge on non-existing request\n");
+ return;
+ }
+
+ rb_erase(&iot->rb_node, &rb_track_root);
+ iot->sector -= t->bytes >> 10;
+ 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->device, t->sector);
+ iot->allocation_time = t->time;
+}
+
+
+/*
+ * return time between rq allocation and queue
+ */
+static unsigned long long log_track_queue(struct blk_io_trace *t)
+{
+ unsigned long long elapsed;
+ struct io_track *iot;
+
+ if (!track_ios)
+ return -1;
+
+ iot = find_track(t->pid, 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_by_pid(iot->pid);
+ 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, "Trying to issue on non-existing request\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_by_pid(iot->pid);
+ 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, "Trying to dispatch on non-existing request\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_by_pid(iot->pid);
+ 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_by_pid(pid);
+
+ if (!ppi) {
+ ppi = malloc(sizeof(*ppi));
+ memset(ppi, 0, sizeof(*ppi));
+ strncpy(ppi->name, name, sizeof(ppi->name));
+ 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;
- per_cpu_info = realloc(per_cpu_info, new_max * sizeof(*per_cpu_info));
- if (!per_cpu_info) {
- fprintf(stderr, "Cannot allocate CPU info -- %d x %d bytes\n",
- new_max, (int) sizeof(*per_cpu_info));
+ 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 *)per_cpu_info + (max_cpus * sizeof(*per_cpu_info));
- new_space = (new_max - max_cpus) * sizeof(*per_cpu_info);
+ 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);
- max_cpus = new_max;
+
+ pdi->ncpus = new_count;
+ pdi->cpus = cpus;
+}
+
+static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
+{
+ if (cpu >= pdi->ncpus)
+ resize_cpu_info(pdi, cpu);
+ return &pdi->cpus[cpu];
}
-static struct per_cpu_info *get_cpu_info(int cpu)
+
+static int resize_devices(char *name)
{
- struct per_cpu_info *pci;
+ int size = (ndevices + 1) * sizeof(struct per_dev_info);
- if (cpu >= max_cpus)
- resize_cpu_info(cpu);
+ 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;
+}
- /*
- * ->cpu might already be set, but just set it unconditionally
- */
- pci = &per_cpu_info[cpu];
- pci->cpu = cpu;
+static struct per_dev_info *get_dev_info(dev_t id, int create)
+{
+ int i;
- return pci;
+ for (i = 0; i < ndevices; i++)
+ if (devices[i].id == id)
+ return &devices[i];
+ if (!create)
+ return NULL;
+ if (resize_devices(NULL) != 0)
+ return NULL;
+ return &devices[ndevices-1];
}
-static inline void check_time(struct blk_io_trace *bit)
+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)
+{
+ __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_q(struct io_stats *ios, struct blk_io_trace *t,
+ int rw)
{
if (rw) {
- pci->qwrites++;
- pci->qwrite_kb += bytes >> 10;
+ ios->qwrites++;
+ ios->qwrite_kb += t->bytes >> 10;
} else {
- pci->qreads++;
- pci->qread_kb += bytes >> 10;
+ ios->qreads++;
+ ios->qread_kb += t->bytes >> 10;
+ }
+}
+
+static inline void account_q(struct blk_io_trace *t, struct per_cpu_info *pci,
+ int rw)
+{
+ __account_q(&pci->io_stats, t, rw);
+
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+
+ __account_q(ios, t, rw);
}
}
-static inline void account_c(struct per_cpu_info *pci, int rw,
- unsigned int bytes)
+static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
{
if (rw) {
- pci->cwrites++;
- pci->cwrite_kb += bytes >> 10;
+ ios->cwrites++;
+ ios->cwrite_kb += bytes >> 10;
} else {
- pci->creads++;
- pci->cread_kb += bytes >> 10;
+ ios->creads++;
+ ios->cread_kb += bytes >> 10;
}
}
-static inline void account_i(struct per_cpu_info *pci, int rw,
- unsigned int bytes)
+static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
+ int rw, int bytes)
+{
+ __account_c(&pci->io_stats, rw, bytes);
+
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+
+ __account_c(ios, rw, bytes);
+ }
+}
+
+static inline void __account_i(struct io_stats *ios, int rw, unsigned int bytes)
{
if (rw) {
- pci->iwrites++;
- pci->iwrite_kb += bytes >> 10;
+ ios->iwrites++;
+ ios->iwrite_kb += bytes >> 10;
} else {
- pci->ireads++;
- pci->iread_kb += bytes >> 10;
+ ios->ireads++;
+ ios->iread_kb += bytes >> 10;
}
}
-static void output(struct per_cpu_info *pci, char *s)
+static inline void account_i(struct blk_io_trace *t, struct per_cpu_info *pci,
+ int rw)
{
- printf("%s", s);
+ __account_i(&pci->io_stats, rw, t->bytes);
+
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+
+ __account_i(ios, rw, t->bytes);
+ }
+}
- if (pci->ofp)
- fprintf(pci->ofp, "%s", s);
+static void output(struct per_cpu_info *pci, char *s)
+{
+ fprintf(ofp, "%s", s);
}
static char hstring[256];
rwbs[i] = '\0';
- sprintf(hstring, "%c %3d %15ld %5Lu.%09Lu %5u %c %3s", backwards,
- pci->cpu,
+ sprintf(hstring, "%3d,%-3d %2d %8ld %5Lu.%09Lu %5u %c %3s",
+ MAJOR(t->device), MINOR(t->device), pci->cpu,
(unsigned long)t->sequence, SECONDS(t->time),
NANO_SECONDS(t->time), t->pid, act, rwbs);
static void log_complete(struct per_cpu_info *pci, struct blk_io_trace *t,
char act)
{
- sprintf(tstring,"%s %Lu + %u [%d]\n", setup_header(pci, t, act),
- (unsigned long long)t->sector, t->bytes >> 9, t->error);
+ unsigned long long elapsed = log_track_complete(t);
+
+ if (elapsed != -1ULL) {
+ unsigned long usec = elapsed / 1000;
+
+ sprintf(tstring,"%s %Lu + %u (%8lu) [%d]\n",
+ setup_header(pci, t, act),
+ (unsigned long long)t->sector, t->bytes >> 9,
+ usec, t->error);
+ } else {
+ 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);
}
static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
char act)
{
- sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
- (unsigned long long)t->sector, t->bytes >> 9, t->comm);
+ unsigned long long elapsed = log_track_queue(t);
+
+ if (elapsed != -1ULL) {
+ unsigned long usec = elapsed / 1000;
+
+ sprintf(tstring,"%s %Lu + %u (%8lu) [%s]\n",
+ setup_header(pci, t, act),
+ (unsigned long long)t->sector, t->bytes >> 9,
+ usec, t->comm);
+ } else {
+ sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
+ (unsigned long long)t->sector, t->bytes >> 9, t->comm);
+ }
output(pci, tstring);
}
static void log_issue(struct per_cpu_info *pci, struct blk_io_trace *t,
char act)
{
- sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
- (unsigned long long)t->sector, t->bytes >> 9, t->comm);
+ unsigned long long elapsed = log_track_issue(t);
+
+ if (elapsed != -1ULL) {
+ double usec = (double) elapsed / 1000;
+
+ sprintf(tstring,"%s %Lu + %u (%8.2f) [%s]\n",
+ setup_header(pci, t, act),
+ (unsigned long long)t->sector, t->bytes >> 9,
+ usec, t->comm);
+ } else {
+ sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
+ (unsigned long long)t->sector, t->bytes >> 9, t->comm);
+ }
+
output(pci, tstring);
}
static void log_merge(struct per_cpu_info *pci, struct blk_io_trace *t,
char act)
{
- sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
+ log_track_merge(t);
+
+ sprintf(tstring,"%s %Lu + %u [%s]\n", setup_header(pci, t, act),
(unsigned long long)t->sector, t->bytes >> 9, t->comm);
output(pci, tstring);
}
output(pci, tstring);
}
+static int log_unplug(struct per_cpu_info *pci, struct blk_io_trace *t,
+ char act)
+{
+ __u64 *depth;
+ int len;
+
+ len = sprintf(tstring,"%s ", setup_header(pci, t, act));
+ depth = (__u64 *) t + sizeof(*t);
+ sprintf(tstring + len, "%u\n", (unsigned int) be64_to_cpu(*depth));
+ output(pci, tstring);
+
+ return 0;
+}
+
static int log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char act)
{
unsigned char *buf;
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);
+ account_q(t, pci, w);
log_queue(pci, t, 'Q');
break;
case __BLK_TA_BACKMERGE:
- account_m(pci, w, t->bytes);
+ account_m(t, pci, w);
log_merge(pci, t, 'M');
break;
case __BLK_TA_FRONTMERGE:
- account_m(pci, w, t->bytes);
+ account_m(t, pci, w);
log_merge(pci, t, 'F');
break;
case __BLK_TA_GETRQ:
+ log_track_getrq(t);
log_generic(pci, t, 'G');
break;
case __BLK_TA_SLEEPRQ:
log_generic(pci, t, 'S');
break;
case __BLK_TA_REQUEUE:
- account_c(pci, w, -t->bytes);
+ account_c(t, pci, w, -t->bytes);
log_queue(pci, t, 'R');
break;
case __BLK_TA_ISSUE:
- account_i(pci, w, t->bytes);
+ account_i(t, pci, w);
log_issue(pci, t, 'D');
break;
case __BLK_TA_COMPLETE:
- account_c(pci, w, t->bytes);
+ account_c(t, pci, w, t->bytes);
log_complete(pci, t, 'C');
break;
+ case __BLK_TA_PLUG:
+ log_generic(pci, t, 'P');
+ break;
+ case __BLK_TA_UNPLUG:
+ log_unplug(pci, t, 'U');
+ break;
default:
fprintf(stderr, "Bad fs action %x\n", t->action);
break;
}
}
-static int dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci)
+static int dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
+ struct per_dev_info *pdi)
{
int ret = 0;
- if (output_name && !pci->ofp) {
- snprintf(pci->ofname, sizeof(pci->ofname) - 1,
- "%s_log.%d", output_name, pci->cpu);
-
- pci->ofp = fopen(pci->ofname, "w");
- if (pci->ofp == NULL) {
- perror(pci->ofname);
- return 1;
- }
- }
-
if (t->action & BLK_TC_ACT(BLK_TC_PC))
ret = dump_trace_pc(t, pci);
else
dump_trace_fs(t, pci);
- events++;
+ pdi->events++;
return ret;
}
-static void dump_pci_stats(struct per_cpu_info *pci)
+static void dump_io_stats(struct io_stats *ios, char *msg)
{
- printf(" Reads Queued: %'8lu, %'8LuKiB\t", pci->qreads, pci->qread_kb);
- printf(" Writes Queued: %'8lu, %'8LuKiB\n", pci->qwrites,pci->qwrite_kb);
+ fprintf(ofp, "%s\n", msg);
- printf(" Read Dispatches: %'8lu, %'8LuKiB\t", pci->ireads, pci->iread_kb);
- printf(" Write Dispatches: %'8lu, %'8LuKiB\n", pci->iwrites,pci->iwrite_kb);
- printf(" Reads Completed: %'8lu, %'8LuKiB\t", pci->creads, pci->cread_kb);
- printf(" Writes Completed: %'8lu, %'8LuKiB\n", pci->cwrites,pci->cwrite_kb);
- printf(" Read Merges: %'8lu%8c\t", pci->mreads, ' ');
+ fprintf(ofp, " Reads Queued: %'8lu, %'8LuKiB\t", ios->qreads, ios->qread_kb);
+ fprintf(ofp, " Writes Queued: %'8lu, %'8LuKiB\n", ios->qwrites,ios->qwrite_kb);
- printf(" Write Merges: %'8lu\n", pci->mwrites);
+ 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);
}
-static void show_stats(void)
+static void dump_wait_stats(struct per_process_info *ppi)
{
- struct per_cpu_info foo, *pci;
- int i, pci_events = 0;
-
- memset(&foo, 0, sizeof(foo));
-
- for (i = 0; i < max_cpus; i++) {
- pci = &per_cpu_info[i];
-
- if (!pci->nelems)
- continue;
-
- 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;
+ 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);
+}
- printf("CPU%d:\n", i);
- dump_pci_stats(pci);
- pci_events++;
- }
+static void show_process_stats(void)
+{
+ struct per_process_info *ppi;
- if (pci_events > 1) {
- printf("Total:\n");
- dump_pci_stats(&foo);
+ ppi = ppi_list;
+ while (ppi) {
+ dump_io_stats(&ppi->io_stats, ppi->name);
+ dump_wait_stats(ppi);
+ ppi = ppi->list_next;
}
- printf("Events: %'Lu\n", events);
+ fprintf(ofp, "\n");
}
-static inline int trace_rb_insert(struct trace *t)
+static void show_device_and_cpu_stats(void)
{
- struct rb_node **p = &rb_root.rb_node;
- struct rb_node *parent = NULL;
- struct trace *__t;
-
- while (*p) {
- parent = *p;
- __t = rb_entry(parent, struct trace, rb_node);
+ struct per_dev_info *pdi;
+ struct per_cpu_info *pci;
+ 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;
+
+ snprintf(line, sizeof(line) - 1, "CPU%d (%s):",
+ j, get_dev_name(pdi, name, sizeof(name)));
+ dump_io_stats(ios, line);
+ pci_events++;
+ }
- 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;
+ 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);
}
- }
- rb_link_node(&t->rb_node, parent, p);
- rb_insert_color(&t->rb_node, &rb_root);
- return 0;
+ fprintf(ofp, "Events (%s): %'Lu\n",
+ get_dev_name(pdi, line, sizeof(line)), pdi->events);
+ }
}
-#define min(a, b) ((a) < (b) ? (a) : (b))
-
static struct blk_io_trace *find_trace(void *p, unsigned long offset, int nr)
{
unsigned long max_offset = min(offset,nr * sizeof(struct blk_io_trace));
return NULL;
}
-static int sort_entries(void *traces, unsigned long offset, int nr)
+static int sort_entries(void *traces, unsigned long offset, int nr,
+ struct per_dev_info *fpdi, struct per_cpu_info *fpci)
{
+ 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;
t = malloc(sizeof(*t));
+ if (!t) {
+ fprintf(stderr, "Out of memory, seq %d on dev %d,%d\n",
+ bit->sequence,
+ MAJOR(bit->device), MINOR(bit->device));
+ return -1;
+ }
t->bit = bit;
memset(&t->rb_node, 0, sizeof(t->rb_node));
trace_to_cpu(bit);
- if (verify_trace(bit))
+ if (verify_trace(bit)) {
+ free(t);
break;
+ }
- pci = get_cpu_info(bit->cpu);
+ pdi = fpdi ? fpdi : get_dev_info(bit->device, 1);
+ pdi->id = bit->device;
+ pci = fpci ? fpci : get_cpu_info(pdi, bit->cpu);
+ pci->cpu = bit->cpu;
pci->nelems++;
- if (trace_rb_insert(t))
+ if (trace_rb_insert(t)) {
+ free(t);
return -1;
+ }
traces += sizeof(*bit) + bit->pdu_len;
- nelems++;
nr--;
}
- return nelems;
+ return 0;
}
static void free_entries_rb(void)
{
struct rb_node *n;
- while ((n = rb_first(&rb_root)) != NULL) {
+ while ((n = rb_first(&rb_sort_root)) != NULL) {
struct trace *t = rb_entry(n, struct trace, rb_node);
- rb_erase(&t->rb_node, &rb_root);
+ rb_erase(&t->rb_node, &rb_sort_root);
free(t);
}
}
static void show_entries_rb(void)
{
+ struct per_dev_info *pdi;
struct blk_io_trace *bit;
struct rb_node *n;
struct trace *t;
int cpu;
- n = rb_first(&rb_root);
+ n = rb_first(&rb_sort_root);
if (!n)
return;
t = rb_entry(n, struct trace, rb_node);
bit = t->bit;
+ pdi = get_dev_info(bit->device, 0);
+ if (!pdi) {
+ fprintf(stderr, "Unknown device ID? (%d,%d)\n",
+ MAJOR(bit->device), MINOR(bit->device));
+ break;
+ }
cpu = bit->cpu;
- if (cpu > max_cpus) {
- fprintf(stderr, "CPU number too large (%d)\n", cpu);
+ if (cpu > pdi->ncpus) {
+ fprintf(stderr, "Unknown CPU ID? (%d, device %d,%d)\n",
+ cpu, MAJOR(bit->device), MINOR(bit->device));
break;
}
- if (genesis_time == 0)
- genesis_time = bit->time;
bit->time -= genesis_time;
+ if (bit->time < stopwatch_start)
+ continue;
+ if (bit->time >= stopwatch_end)
+ break;
- check_time(bit);
+ check_time(pdi, bit);
- if (dump_trace(bit, &per_cpu_info[cpu]))
+ if (dump_trace(bit, &pdi->cpus[cpu], pdi))
break;
} while ((n = rb_next(n)) != NULL);
static int do_file(void)
{
- int i, nfiles;
-
- for (i = 0, nfiles = 0;; i++, nfiles++) {
- struct per_cpu_info *pci;
- struct stat st;
- void *tb;
-
- pci = get_cpu_info(i);
- pci->ofp = NULL;
-
- snprintf(pci->fname, sizeof(pci->fname)-1,"%s_out.%d", dev, i);
- if (stat(pci->fname, &st) < 0)
- break;
- if (!st.st_size)
- continue;
+ struct per_dev_info *pdi;
+ int i, j, nfiles = 0;
- printf("Processing %s\n", pci->fname);
+ for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
+ for (j = 0;; j++, nfiles++) {
+ struct per_cpu_info *pci;
+ struct stat st;
+ void *tb;
- tb = malloc(st.st_size);
+ pci = get_cpu_info(pdi, j);
+ pci->cpu = j;
- pci->fd = open(pci->fname, O_RDONLY);
- if (pci->fd < 0) {
- perror(pci->fname);
- break;
+ snprintf(pci->fname, sizeof(pci->fname)-1,
+ "%s_out.%d", pdi->name, j);
+ if (stat(pci->fname, &st) < 0)
+ break;
+ if (!st.st_size)
+ continue;
+
+ printf("Processing %s\n", pci->fname);
+
+ tb = malloc(st.st_size);
+ if (!tb) {
+ fprintf(stderr, "Out of memory, skip file %s\n",
+ pci->fname);
+ continue;
+ }
+
+ pci->fd = open(pci->fname, O_RDONLY);
+ if (pci->fd < 0) {
+ perror(pci->fname);
+ free(tb);
+ continue;
+ }
+
+ if (read_data(pci->fd, tb, st.st_size, 1)) {
+ close(pci->fd);
+ free(tb);
+ continue;
+ }
+
+ if (sort_entries(tb, st.st_size, ~0U, pdi, pci) == -1) {
+ close(pci->fd);
+ free(tb);
+ continue;
+ }
+
+ printf("Completed %s (CPU%d %d, entries)\n",
+ pci->fname, j, pci->nelems);
+ close(pci->fd);
}
-
- if (read_data(pci->fd, tb, st.st_size, 1))
- break;
-
- if (sort_entries(tb, st.st_size, ~0U) == -1)
- break;
-
- close(pci->fd);
- printf("\t%2d %10s %15d\n", i, pci->fname, pci->nelems);
-
}
if (!nfiles) {
{
long old_size = *size;
+ if (*size == 0)
+ *size = 64 * sizeof(struct blk_io_trace);
+
*size *= 2;
*buffer = realloc(*buffer, *size);
- memset(*buffer + offset, 0, *size - old_size);
+
+ if (old_size)
+ memset(*buffer + offset, 0, *size - old_size);
}
-static int read_sort_events(int fd, void **buffer)
+static int read_sort_events(int fd, void **buffer, long *max_offset)
{
- long offset, max_offset;
+ long offset;
int events;
- max_offset = 128 * sizeof(struct blk_io_trace);
- *buffer = malloc(max_offset);
- events = 0;
- offset = 0;
-
+ events = offset = 0;
do {
struct blk_io_trace *t;
int pdu_len;
+ __u32 magic;
- if (max_offset - offset < sizeof(*t))
- resize_buffer(buffer, &max_offset, offset);
+ if (*max_offset - offset < sizeof(*t))
+ resize_buffer(buffer, max_offset, offset);
- if (read_data(fd, *buffer + offset, sizeof(*t), !events)) {
- if (events)
- break;
-
- usleep(1000);
- continue;
- }
+ if (read_data(fd, *buffer + offset, sizeof(*t), !events))
+ break;
t = *buffer + offset;
offset += sizeof(*t);
+ magic = be32_to_cpu(t->magic);
+ if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
+ fprintf(stderr, "Bad magic %x\n", magic);
+ break;
+ }
+
pdu_len = be16_to_cpu(t->pdu_len);
if (pdu_len) {
- if (max_offset - offset < pdu_len)
- resize_buffer(buffer, &max_offset, offset);
+ if (*max_offset - offset <= pdu_len)
+ resize_buffer(buffer, max_offset, offset);
if (read_data(fd, *buffer + offset, pdu_len, 1))
break;
static int do_stdin(void)
{
int fd;
- void *ptr;
+ void *ptr = NULL;
+ long max_offset;
fd = dup(STDIN_FILENO);
+ max_offset = 0;
do {
int events;
- events = read_sort_events(fd, &ptr);
+ events = read_sort_events(fd, &ptr, &max_offset);
if (!events)
break;
- if (sort_entries(ptr, ~0UL, events) == -1)
+ if (sort_entries(ptr, ~0UL, events, NULL, NULL) == -1)
break;
show_entries_rb();
free_entries_rb();
} while (1);
+ if (ptr)
+ free(ptr);
+
close(fd);
- free(ptr);
return 0;
}
static void flush_output(void)
{
- int i;
-
- for (i = 0; i < max_cpus; i++) {
- struct per_cpu_info *pci = &per_cpu_info[i];
-
- if (pci->ofp) {
- fflush(pci->ofp);
- fclose(pci->ofp);
- pci->ofp = NULL;
- }
- }
+ fflush(ofp);
}
static void handle_sigint(int sig)
flush_output();
}
+/*
+ * 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 = stopwatch_start + DOUBLE_TO_NANO_ULL(value);
+ return 0;
+}
+
static void usage(char *prog)
{
- fprintf(stderr, "Usage: %s -i <name> [-o <output>]\n", prog);
+ fprintf(stderr, "Usage: %s "
+ "[-i <name>] [-o <output>] [-s] [-w N[:n]] <name>...\n",
+ prog);
}
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;
default:
usage(argv[0]);
return 1;
}
}
- if (!dev) {
+ 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(&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);
setlocale(LC_NUMERIC, "en_US");
- if (!strcmp(dev, "-"))
+ if (!output_name) {
+ ofp = fdopen(STDOUT_FILENO, "w");
+ mode = _IOLBF;
+ } else {
+ char ofname[128];
+
+ snprintf(ofname, sizeof(ofname) - 1, "%s.log", 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;
}