* block queue tracing parse application
*
* Copyright (C) 2005 Jens Axboe <axboe@suse.de>
+ * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
#include <errno.h>
#include <signal.h>
#include <locale.h>
-#include <limits.h>
-#include <ctype.h>
+#include <libgen.h>
#include "blktrace.h"
#include "rbtree.h"
+#include "jhash.h"
-#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 char blkparse_version[] = "1.2.0";
-#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))
-
-#define HEADER "%D %2c %8s %5T.%9t %5p %2a %3d "
-
-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;
- unsigned long io_unplugs, timer_unplugs;
-};
-
-struct per_cpu_info {
- int cpu;
- int nelems;
-
- int fd;
- char fname[128];
-
- struct io_stats io_stats;
+struct skip_info {
+ unsigned long start, end;
+ struct skip_info *prev, *next;
};
struct per_dev_info {
- dev_t id;
+ dev_t dev;
char *name;
int backwards;
unsigned long long events;
+ unsigned long long first_reported_time;
unsigned long long last_reported_time;
+ unsigned long long last_read_time;
struct io_stats io_stats;
- unsigned long last_sequence;
unsigned long skips;
+ unsigned long long seq_skips;
+ unsigned int max_depth[2];
+ unsigned int cur_depth[2];
+ struct rb_root rb_track;
+
+ int nfiles;
int ncpus;
+
+ unsigned long *cpu_map;
+ unsigned int cpu_map_max;
+
struct per_cpu_info *cpus;
};
+/*
+ * some duplicated effort here, we can unify this hash and the ppi hash later
+ */
+struct process_pid_map {
+ pid_t pid;
+ char comm[16];
+ struct process_pid_map *hash_next, *list_next;
+};
+
+#define PPM_HASH_SHIFT (8)
+#define PPM_HASH_SIZE (1 << PPM_HASH_SHIFT)
+#define PPM_HASH_MASK (PPM_HASH_SIZE - 1)
+static struct process_pid_map *ppm_hash_table[PPM_HASH_SIZE];
+
struct per_process_info {
- char name[16];
- __u32 pid;
+ struct process_pid_map *ppm;
struct io_stats io_stats;
struct per_process_info *hash_next, *list_next;
+ int more_than_one;
/*
* individual io stats
};
#define PPI_HASH_SHIFT (8)
-static struct per_process_info *ppi_hash[1 << PPI_HASH_SHIFT];
+#define PPI_HASH_SIZE (1 << PPI_HASH_SHIFT)
+#define PPI_HASH_MASK (PPI_HASH_SIZE - 1)
+
+enum {
+ SORT_PROG_EVENT_N, /* Program Name */
+ SORT_PROG_EVENT_QKB, /* KB: Queued read and write */
+ SORT_PROG_EVENT_RKB, /* KB: Queued Read */
+ SORT_PROG_EVENT_WKB, /* KB: Queued Write */
+ SORT_PROG_EVENT_CKB, /* KB: Complete */
+ SORT_PROG_EVENT_QIO, /* IO: Queued read and write */
+ SORT_PROG_EVENT_RIO, /* IO: Queued Read */
+ SORT_PROG_EVENT_WIO, /* IO: Queued Write */
+ SORT_PROG_EVENT_CIO, /* IO: Complete */
+};
+
+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:"
static struct option l_opts[] = {
- {
- .name = "input",
+ {
+ .name = "act-mask",
.has_arg = required_argument,
.flag = NULL,
- .val = 'i'
+ .val = 'a'
},
{
- .name = "output",
+ .name = "set-mask",
.has_arg = required_argument,
.flag = NULL,
- .val = 'o'
+ .val = 'A'
},
{
.name = "batch",
.val = 'b'
},
{
- .name = "per program stats",
+ .name = "input-directory",
+ .has_arg = required_argument,
+ .flag = NULL,
+ .val = 'D'
+ },
+ {
+ .name = "dump-binary",
+ .has_arg = required_argument,
+ .flag = NULL,
+ .val = 'd'
+ },
+ {
+ .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 = 's'
+ .val = 'h'
+ },
+ {
+ .name = "input",
+ .has_arg = required_argument,
+ .flag = NULL,
+ .val = 'i'
},
{
- .name = "track ios",
+ .name = "no-msgs",
.has_arg = no_argument,
.flag = NULL,
- .val = 't'
+ .val = 'M'
+ },
+ {
+ .name = "output",
+ .has_arg = required_argument,
+ .flag = NULL,
+ .val = 'o'
+ },
+ {
+ .name = "no-text-output",
+ .has_arg = no_argument,
+ .flag = NULL,
+ .val = 'O'
},
{
.name = "quiet",
.flag = NULL,
.val = 'q'
},
+ {
+ .name = "per-program-stats",
+ .has_arg = no_argument,
+ .flag = NULL,
+ .val = 's'
+ },
+ {
+ .name = "sort-program-stats",
+ .has_arg = required_argument,
+ .flag = NULL,
+ .val = 'S'
+ },
+ {
+ .name = "track-ios",
+ .has_arg = no_argument,
+ .flag = NULL,
+ .val = 't'
+ },
{
.name = "stopwatch",
.has_arg = required_argument,
.val = 'w'
},
{
- .name = "format",
- .has_arg = required_argument,
+ .name = "verbose",
+ .has_arg = no_argument,
.flag = NULL,
- .val = 'f'
+ .val = 'v'
},
{
- .name = "format-spec",
- .has_arg = required_argument,
+ .name = "version",
+ .has_arg = no_argument,
.flag = NULL,
- .val = 'F'
+ .val = 'V'
},
+ {
+ .name = NULL,
+ }
};
/*
struct blk_io_trace *bit;
struct rb_node rb_node;
struct trace *next;
- int skipped;
+ unsigned long read_sequence;
};
static struct rb_root rb_sort_root;
-static struct rb_root rb_track_root;
+static unsigned long rb_sort_entries;
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;
+ struct process_pid_map *ppm;
__u64 sector;
- __u32 pid;
unsigned long long allocation_time;
unsigned long long queue_time;
unsigned long long dispatch_time;
static int ndevices;
static struct per_dev_info *devices;
static char *get_dev_name(struct per_dev_info *, char *, int);
+static int trace_rb_insert_last(struct per_dev_info *, struct trace *);
-static FILE *ofp;
+FILE *ofp = NULL;
static char *output_name;
+static char *input_dir;
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 unsigned long long stopwatch_end = -1ULL; /* "infinity" */
+static unsigned long read_sequence;
static int per_process_stats;
+static int per_process_stats_event = SORT_PROG_EVENT_N;
+static int per_device_and_cpu_stats = 1;
static int track_ios;
+static int ppi_hash_by_pid = 1;
+static int verbose;
+static unsigned int act_mask = -1U;
+static int stats_printed;
+static int bin_output_msgs = 1;
+int data_is_native = -1;
+
+static FILE *dump_fp;
+static char *dump_binary;
+
+static unsigned int t_alloc_cache;
+static unsigned int bit_alloc_cache;
-#define RB_BATCH_DEFAULT (1024)
-static int rb_batch = RB_BATCH_DEFAULT;
+#define RB_BATCH_DEFAULT (512)
+static unsigned int rb_batch = RB_BATCH_DEFAULT;
static int pipeline;
+static char *pipename;
+
+static int text_output = 1;
#define is_done() (*(volatile int *)(&done))
static volatile int done;
-static inline unsigned long hash_long(unsigned long val)
+struct timespec abs_start_time;
+static unsigned long long start_timestamp;
+
+static int have_drv_data = 0;
+
+#define JHASH_RANDOM (0x3af5f2ee)
+
+#define CPUS_PER_LONG (8 * sizeof(unsigned long))
+#define CPU_IDX(cpu) ((cpu) / CPUS_PER_LONG)
+#define CPU_BIT(cpu) ((cpu) & (CPUS_PER_LONG - 1))
+
+static void output_binary(void *buf, int len)
{
-#if __WORDSIZE == 32
- val *= 0x9e370001UL;
-#elif __WORDSIZE == 64
- val *= 0x9e37fffffffc0001UL;
-#else
-#error unknown word size
-#endif
+ if (dump_binary) {
+ size_t n = fwrite(buf, len, 1, dump_fp);
+ if (n != 1) {
+ perror(dump_binary);
+ fclose(dump_fp);
+ dump_binary = NULL;
+ }
+ }
+}
+
+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;
+
+ for (new_count = 0; new_count < pdi->ncpus; new_count++) {
+ struct per_cpu_info *pci = &pdi->cpus[new_count];
- return val >> (__WORDSIZE - PPI_HASH_SHIFT);
+ if (!pci->fd) {
+ pci->fd = -1;
+ memset(&pci->rb_last, 0, sizeof(pci->rb_last));
+ pci->rb_last_entries = 0;
+ pci->last_sequence = -1;
+ }
+ }
}
-static inline void add_process_to_hash(struct per_process_info *ppi)
+static struct per_cpu_info *get_cpu_info(struct per_dev_info *pdi, int cpu)
{
- const int hash_idx = hash_long(ppi->pid);
+ struct per_cpu_info *pci;
+
+ if (cpu >= pdi->ncpus)
+ resize_cpu_info(pdi, cpu);
- ppi->hash_next = ppi_hash[hash_idx];
- ppi_hash[hash_idx] = ppi;
+ pci = &pdi->cpus[cpu];
+ pci->cpu = cpu;
+ return pci;
}
-static inline void add_process_to_list(struct per_process_info *ppi)
+
+static int resize_devices(char *name)
{
- ppi->list_next = ppi_list;
- ppi_list = ppi;
- ppi_list_entries++;
+ 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_process_info *find_process_by_pid(__u32 pid)
+static struct per_dev_info *get_dev_info(dev_t dev)
{
- const int hash_idx = hash_long(pid);
- struct per_process_info *ppi;
+ struct per_dev_info *pdi;
+ int i;
- ppi = ppi_hash[hash_idx];
- while (ppi) {
- if (ppi->pid == pid)
- return ppi;
+ for (i = 0; i < ndevices; i++) {
+ if (!devices[i].dev)
+ devices[i].dev = dev;
+ if (devices[i].dev == dev)
+ return &devices[i];
+ }
- ppi = ppi->hash_next;
+ if (resize_devices(NULL))
+ return NULL;
+
+ pdi = &devices[ndevices - 1];
+ pdi->dev = dev;
+ pdi->first_reported_time = 0;
+ pdi->last_read_time = 0;
+
+ return pdi;
+}
+
+static void insert_skip(struct per_cpu_info *pci, unsigned long start,
+ unsigned long end)
+{
+ struct skip_info *sip;
+
+ for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
+ if (end == (sip->start - 1)) {
+ sip->start = start;
+ return;
+ } else if (start == (sip->end + 1)) {
+ sip->end = end;
+ return;
+ }
}
- return NULL;
+ sip = malloc(sizeof(struct skip_info));
+ sip->start = start;
+ sip->end = end;
+ sip->prev = sip->next = NULL;
+ if (pci->skips_tail == NULL)
+ pci->skips_head = pci->skips_tail = sip;
+ else {
+ sip->prev = pci->skips_tail;
+ pci->skips_tail->next = sip;
+ pci->skips_tail = sip;
+ }
}
-static inline int trace_rb_insert(struct trace *t)
+static void remove_sip(struct per_cpu_info *pci, struct skip_info *sip)
{
- struct rb_node **p = &rb_sort_root.rb_node;
- struct rb_node *parent = NULL;
- struct trace *__t;
+ if (sip->prev == NULL) {
+ if (sip->next == NULL)
+ pci->skips_head = pci->skips_tail = NULL;
+ else {
+ pci->skips_head = sip->next;
+ sip->next->prev = NULL;
+ }
+ } else if (sip->next == NULL) {
+ pci->skips_tail = sip->prev;
+ sip->prev->next = NULL;
+ } else {
+ sip->prev->next = sip->next;
+ sip->next->prev = sip->prev;
+ }
- if (genesis_time == 0 || t->bit->time < genesis_time)
- genesis_time = t->bit->time;
+ sip->prev = sip->next = NULL;
+ free(sip);
+}
- while (*p) {
- parent = *p;
- __t = rb_entry(parent, struct trace, rb_node);
+#define IN_SKIP(sip,seq) (((sip)->start <= (seq)) && ((seq) <= sip->end))
+static int check_current_skips(struct per_cpu_info *pci, unsigned long seq)
+{
+ struct skip_info *sip;
- 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));
+ for (sip = pci->skips_tail; sip != NULL; sip = sip->prev) {
+ if (IN_SKIP(sip, seq)) {
+ if (sip->start == seq) {
+ if (sip->end == seq)
+ remove_sip(pci, sip);
+ else
+ sip->start += 1;
+ } else if (sip->end == seq)
+ sip->end -= 1;
+ else {
+ sip->end = seq - 1;
+ insert_skip(pci, seq + 1, sip->end);
+ }
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)
+static void collect_pdi_skips(struct per_dev_info *pdi)
{
- struct rb_node **p = &rb_track_root.rb_node;
- struct rb_node *parent = NULL;
- struct io_track *__iot;
+ struct skip_info *sip;
+ int cpu;
- while (*p) {
- parent = *p;
+ pdi->skips = 0;
+ pdi->seq_skips = 0;
- __iot = rb_entry(parent, struct io_track, rb_node);
+ for (cpu = 0; cpu < pdi->ncpus; cpu++) {
+ struct per_cpu_info *pci = &pdi->cpus[cpu];
- 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;
+ for (sip = pci->skips_head; sip != NULL; sip = sip->next) {
+ pdi->skips++;
+ pdi->seq_skips += (sip->end - sip->start + 1);
+ if (verbose)
+ fprintf(stderr,"(%d,%d): skipping %lu -> %lu\n",
+ MAJOR(pdi->dev), MINOR(pdi->dev),
+ sip->start, sip->end);
}
}
-
- 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)
+static void cpu_mark_online(struct per_dev_info *pdi, unsigned int cpu)
{
- struct rb_node **p = &rb_track_root.rb_node;
- struct rb_node *parent = NULL;
- struct io_track *__iot;
+ if (cpu >= pdi->cpu_map_max || !pdi->cpu_map) {
+ int new_max = (cpu + CPUS_PER_LONG) & ~(CPUS_PER_LONG - 1);
+ unsigned long *map = malloc(new_max / sizeof(long));
- while (*p) {
- parent = *p;
-
- __iot = rb_entry(parent, struct io_track, rb_node);
+ memset(map, 0, new_max / sizeof(long));
- 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;
+ if (pdi->cpu_map) {
+ memcpy(map, pdi->cpu_map, pdi->cpu_map_max / sizeof(long));
+ free(pdi->cpu_map);
+ }
+
+ pdi->cpu_map = map;
+ pdi->cpu_map_max = new_max;
}
- return NULL;
+ pdi->cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
}
-static struct io_track *find_track(__u32 pid, dev_t device, __u64 sector)
+static inline void cpu_mark_offline(struct per_dev_info *pdi, int cpu)
{
- struct io_track *iot;
+ pdi->cpu_map[CPU_IDX(cpu)] &= ~(1UL << CPU_BIT(cpu));
+}
- iot = __find_track(device, sector);
- if (!iot) {
- iot = malloc(sizeof(*iot));
- iot->pid = pid;
- iot->device = device;
- iot->sector = sector;
- track_rb_insert(iot);
- }
+static inline int cpu_is_online(struct per_dev_info *pdi, int cpu)
+{
+ return (pdi->cpu_map[CPU_IDX(cpu)] & (1UL << CPU_BIT(cpu))) != 0;
+}
- return iot;
+static inline int ppm_hash_pid(pid_t pid)
+{
+ return jhash_1word(pid, JHASH_RANDOM) & PPM_HASH_MASK;
}
-static void log_track_frontmerge(struct blk_io_trace *t)
+static struct process_pid_map *find_ppm(pid_t pid)
{
- struct io_track *iot;
+ const int hash_idx = ppm_hash_pid(pid);
+ struct process_pid_map *ppm;
- if (!track_ios)
- return;
+ ppm = ppm_hash_table[hash_idx];
+ while (ppm) {
+ if (ppm->pid == pid)
+ return ppm;
- iot = __find_track(t->device, t->sector + (t->bytes >> 9));
- if (!iot) {
- fprintf(stderr, "failed to find mergeable event\n");
- return;
+ ppm = ppm->hash_next;
}
- rb_erase(&iot->rb_node, &rb_track_root);
- iot->sector -= t->bytes >> 9;
- track_rb_insert(iot);
+ return NULL;
}
-static void log_track_getrq(struct blk_io_trace *t)
+static struct process_pid_map *add_ppm_hash(pid_t pid, const char *name)
{
- struct io_track *iot;
+ const int hash_idx = ppm_hash_pid(pid);
+ struct process_pid_map *ppm;
- if (!track_ios)
- return;
+ ppm = find_ppm(pid);
+ if (!ppm) {
+ ppm = malloc(sizeof(*ppm));
+ memset(ppm, 0, sizeof(*ppm));
+ ppm->pid = pid;
+ memset(ppm->comm, 0, sizeof(ppm->comm));
+ strncpy(ppm->comm, name, sizeof(ppm->comm));
+ ppm->comm[sizeof(ppm->comm) - 1] = '\0';
+ ppm->hash_next = ppm_hash_table[hash_idx];
+ ppm_hash_table[hash_idx] = ppm;
+ }
- iot = find_track(t->pid, t->device, t->sector);
- iot->allocation_time = t->time;
+ return ppm;
}
-
-/*
- * return time between rq allocation and queue
- */
-static unsigned long long log_track_queue(struct blk_io_trace *t)
+static void handle_notify(struct blk_io_trace *bit)
+{
+ void *payload = (caddr_t) bit + sizeof(*bit);
+ __u32 two32[2];
+
+ switch (bit->action) {
+ case BLK_TN_PROCESS:
+ add_ppm_hash(bit->pid, payload);
+ break;
+
+ case BLK_TN_TIMESTAMP:
+ if (bit->pdu_len != sizeof(two32))
+ return;
+ memcpy(two32, payload, sizeof(two32));
+ if (!data_is_native) {
+ two32[0] = be32_to_cpu(two32[0]);
+ two32[1] = be32_to_cpu(two32[1]);
+ }
+ start_timestamp = bit->time;
+ abs_start_time.tv_sec = two32[0];
+ abs_start_time.tv_nsec = two32[1];
+ if (abs_start_time.tv_nsec < 0) {
+ abs_start_time.tv_sec--;
+ abs_start_time.tv_nsec += 1000000000;
+ }
+
+ break;
+
+ case BLK_TN_MESSAGE:
+ if (bit->pdu_len > 0) {
+ char msg[bit->pdu_len+1];
+
+ memcpy(msg, (char *)payload, bit->pdu_len);
+ msg[bit->pdu_len] = '\0';
+
+ fprintf(ofp,
+ "%3d,%-3d %2d %8s %5d.%09lu %5u %2s %3s %s\n",
+ MAJOR(bit->device), MINOR(bit->device),
+ bit->cpu, "0", (int) SECONDS(bit->time),
+ (unsigned long) NANO_SECONDS(bit->time),
+ 0, "m", "N", msg);
+ }
+ break;
+
+ default:
+ /* Ignore unknown notify events */
+ ;
+ }
+}
+
+char *find_process_name(pid_t pid)
+{
+ struct process_pid_map *ppm = find_ppm(pid);
+
+ if (ppm)
+ return ppm->comm;
+
+ return NULL;
+}
+
+static inline int ppi_hash_pid(pid_t 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)
+{
+ struct process_pid_map *ppm = ppi->ppm;
+
+ if (ppi_hash_by_pid)
+ return ppi_hash_pid(ppm->pid);
+
+ return ppi_hash_name(ppm->comm);
+}
+
+static inline void add_ppi_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_ppi_to_list(struct per_process_info *ppi)
+{
+ ppi->list_next = ppi_list;
+ ppi_list = ppi;
+ ppi_list_entries++;
+}
+
+static struct per_process_info *find_ppi_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) {
+ struct process_pid_map *ppm = ppi->ppm;
+
+ if (!strcmp(ppm->comm, name))
+ return ppi;
+
+ ppi = ppi->hash_next;
+ }
+
+ return NULL;
+}
+
+static struct per_process_info *find_ppi_by_pid(pid_t pid)
+{
+ const int hash_idx = ppi_hash_pid(pid);
+ struct per_process_info *ppi;
+
+ ppi = ppi_hash_table[hash_idx];
+ while (ppi) {
+ struct process_pid_map *ppm = ppi->ppm;
+
+ if (ppm->pid == pid)
+ return ppi;
+
+ ppi = ppi->hash_next;
+ }
+
+ return NULL;
+}
+
+static struct per_process_info *find_ppi(pid_t pid)
+{
+ struct per_process_info *ppi;
+ char *name;
+
+ if (ppi_hash_by_pid)
+ return find_ppi_by_pid(pid);
+
+ name = find_process_name(pid);
+ if (!name)
+ return NULL;
+
+ ppi = find_ppi_by_name(name);
+ if (ppi && ppi->ppm->pid != pid)
+ ppi->more_than_one = 1;
+
+ return ppi;
+}
+
+/*
+ * 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)
+{
+ if (t_alloc_cache < 1024) {
+ t->next = t_alloc_list;
+ t_alloc_list = t;
+ t_alloc_cache++;
+ } else
+ free(t);
+}
+
+static inline struct trace *t_alloc(void)
+{
+ struct trace *t = t_alloc_list;
+
+ if (t) {
+ t_alloc_list = t->next;
+ t_alloc_cache--;
+ return t;
+ }
+
+ return malloc(sizeof(*t));
+}
+
+static inline void bit_free(struct blk_io_trace *bit)
+{
+ if (bit_alloc_cache < 1024 && !bit->pdu_len) {
+ /*
+ * 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;
+ bit_alloc_cache++;
+ } else
+ free(bit);
+}
+
+static inline struct blk_io_trace *bit_alloc(void)
+{
+ struct blk_io_trace *bit = bit_alloc_list;
+
+ if (bit) {
+ bit_alloc_list = (struct blk_io_trace *) (unsigned long) \
+ bit->time;
+ bit_alloc_cache--;
+ return bit;
+ }
+
+ return malloc(sizeof(*bit));
+}
+
+static inline void __put_trace_last(struct per_dev_info *pdi, struct trace *t)
+{
+ struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
+
+ rb_erase(&t->rb_node, &pci->rb_last);
+ pci->rb_last_entries--;
+
+ bit_free(t->bit);
+ t_free(t);
+}
+
+static void put_trace(struct per_dev_info *pdi, struct trace *t)
+{
+ rb_erase(&t->rb_node, &rb_sort_root);
+ rb_sort_entries--;
+
+ trace_rb_insert_last(pdi, t);
+}
+
+static inline int trace_rb_insert(struct trace *t, struct rb_root *root)
+{
+ struct rb_node **p = &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 /* >= sequence */
+ p = &(*p)->rb_right;
+ }
+
+ rb_link_node(&t->rb_node, parent, p);
+ rb_insert_color(&t->rb_node, root);
+ return 0;
+}
+
+static inline int trace_rb_insert_sort(struct trace *t)
+{
+ if (!trace_rb_insert(t, &rb_sort_root)) {
+ rb_sort_entries++;
+ return 0;
+ }
+
+ return 1;
+}
+
+static int trace_rb_insert_last(struct per_dev_info *pdi, struct trace *t)
+{
+ struct per_cpu_info *pci = get_cpu_info(pdi, t->bit->cpu);
+
+ if (trace_rb_insert(t, &pci->rb_last))
+ return 1;
+
+ pci->rb_last_entries++;
+
+ if (pci->rb_last_entries > rb_batch * pdi->nfiles) {
+ struct rb_node *n = rb_first(&pci->rb_last);
+
+ t = rb_entry(n, struct trace, rb_node);
+ __put_trace_last(pdi, t);
+ }
+
+ return 0;
+}
+
+static struct trace *trace_rb_find(dev_t device, unsigned long sequence,
+ struct rb_root *root, int order)
+{
+ struct rb_node *n = root->rb_node;
+ struct rb_node *prev = NULL;
+ struct trace *__t;
+
+ while (n) {
+ __t = rb_entry(n, struct trace, rb_node);
+ prev = n;
+
+ if (device < __t->bit->device)
+ n = n->rb_left;
+ else if (device > __t->bit->device)
+ n = n->rb_right;
+ else if (sequence < __t->bit->sequence)
+ n = n->rb_left;
+ else if (sequence > __t->bit->sequence)
+ n = n->rb_right;
+ else
+ return __t;
+ }
+
+ /*
+ * hack - the list may not be sequence ordered because some
+ * events don't have sequence and time matched. so we end up
+ * being a little off in the rb lookup here, because we don't
+ * know the time we are looking for. compensate by browsing
+ * a little ahead from the last entry to find the match
+ */
+ if (order && prev) {
+ int max = 5;
+
+ while (((n = rb_next(prev)) != NULL) && max--) {
+ __t = rb_entry(n, struct trace, rb_node);
+
+ if (__t->bit->device == device &&
+ __t->bit->sequence == sequence)
+ return __t;
+
+ prev = n;
+ }
+ }
+
+ return NULL;
+}
+
+static inline struct trace *trace_rb_find_last(struct per_dev_info *pdi,
+ struct per_cpu_info *pci,
+ unsigned long seq)
+{
+ return trace_rb_find(pdi->dev, seq, &pci->rb_last, 0);
+}
+
+static inline int track_rb_insert(struct per_dev_info *pdi,struct io_track *iot)
+{
+ struct rb_node **p = &pdi->rb_track.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->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(pdi->dev), MINOR(pdi->dev));
+ return 1;
+ }
+ }
+
+ rb_link_node(&iot->rb_node, parent, p);
+ rb_insert_color(&iot->rb_node, &pdi->rb_track);
+ return 0;
+}
+
+static struct io_track *__find_track(struct per_dev_info *pdi, __u64 sector)
+{
+ struct rb_node *n = pdi->rb_track.rb_node;
+ struct io_track *__iot;
+
+ while (n) {
+ __iot = rb_entry(n, struct io_track, rb_node);
+
+ if (sector < __iot->sector)
+ n = n->rb_left;
+ else if (sector > __iot->sector)
+ n = n->rb_right;
+ else
+ return __iot;
+ }
+
+ return NULL;
+}
+
+static struct io_track *find_track(struct per_dev_info *pdi, pid_t pid,
+ __u64 sector)
+{
+ struct io_track *iot;
+
+ iot = __find_track(pdi, sector);
+ if (!iot) {
+ iot = malloc(sizeof(*iot));
+ iot->ppm = find_ppm(pid);
+ if (!iot->ppm)
+ iot->ppm = add_ppm_hash(pid, "unknown");
+ iot->sector = sector;
+ track_rb_insert(pdi, iot);
+ }
+
+ return iot;
+}
+
+static void log_track_frontmerge(struct per_dev_info *pdi,
+ struct blk_io_trace *t)
+{
+ struct io_track *iot;
+
+ if (!track_ios)
+ return;
+
+ iot = __find_track(pdi, t->sector + t_sec(t));
+ if (!iot) {
+ if (verbose)
+ fprintf(stderr, "merge not found for (%d,%d): %llu\n",
+ MAJOR(t->device), MINOR(t->device),
+ (unsigned long long) t->sector + t_sec(t));
+ return;
+ }
+
+ rb_erase(&iot->rb_node, &pdi->rb_track);
+ iot->sector -= t_sec(t);
+ track_rb_insert(pdi, iot);
+}
+
+static void log_track_getrq(struct per_dev_info *pdi, struct blk_io_trace *t)
+{
+ struct io_track *iot;
+
+ if (!track_ios)
+ return;
+
+ iot = find_track(pdi, t->pid, t->sector);
+ iot->allocation_time = t->time;
+}
+
+static inline int is_remapper(struct per_dev_info *pdi)
+{
+ int major = MAJOR(pdi->dev);
+
+ return (major == 253 || major == 9);
+}
+
+/*
+ * for md/dm setups, the interesting cycle is Q -> C. So track queueing
+ * time here, as dispatch time
+ */
+static void log_track_queue(struct per_dev_info *pdi, struct blk_io_trace *t)
+{
+ struct io_track *iot;
+
+ if (!track_ios)
+ return;
+ if (!is_remapper(pdi))
+ return;
+
+ iot = find_track(pdi, t->pid, t->sector);
+ iot->dispatch_time = t->time;
+}
+
+/*
+ * return time between rq allocation and insertion
+ */
+static unsigned long long log_track_insert(struct per_dev_info *pdi,
+ 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 = find_track(pdi, t->pid, t->sector);
iot->queue_time = t->time;
+
+ if (!iot->allocation_time)
+ return -1;
+
elapsed = iot->queue_time - iot->allocation_time;
if (per_process_stats) {
- struct per_process_info *ppi = find_process_by_pid(iot->pid);
+ struct per_process_info *ppi = find_ppi(iot->ppm->pid);
int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
if (ppi && elapsed > ppi->longest_allocation_wait[w])
/*
* return time between queue and issue
*/
-static unsigned long long log_track_issue(struct blk_io_trace *t)
+static unsigned long long log_track_issue(struct per_dev_info *pdi,
+ struct blk_io_trace *t)
{
unsigned long long elapsed;
struct io_track *iot;
if ((t->action & BLK_TC_ACT(BLK_TC_FS)) == 0)
return -1;
- iot = __find_track(t->device, t->sector);
+ iot = __find_track(pdi, t->sector);
if (!iot) {
- fprintf(stderr, "failed to find issue event\n");
+ if (verbose)
+ fprintf(stderr, "issue not found for (%d,%d): %llu\n",
+ MAJOR(t->device), MINOR(t->device),
+ (unsigned long long) t->sector);
return -1;
}
elapsed = iot->dispatch_time - iot->queue_time;
if (per_process_stats) {
- struct per_process_info *ppi = find_process_by_pid(iot->pid);
+ struct per_process_info *ppi = find_ppi(iot->ppm->pid);
int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
if (ppi && elapsed > ppi->longest_dispatch_wait[w])
/*
* return time between dispatch and complete
*/
-static unsigned long long log_track_complete(struct blk_io_trace *t)
+static unsigned long long log_track_complete(struct per_dev_info *pdi,
+ 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);
+ iot = __find_track(pdi, t->sector);
if (!iot) {
- fprintf(stderr, "failed to find complete event\n");
+ if (verbose)
+ fprintf(stderr,"complete not found for (%d,%d): %llu\n",
+ MAJOR(t->device), MINOR(t->device),
+ (unsigned long long) t->sector);
return -1;
}
elapsed = iot->completion_time - iot->dispatch_time;
if (per_process_stats) {
- struct per_process_info *ppi = find_process_by_pid(iot->pid);
+ struct per_process_info *ppi = find_ppi(iot->ppm->pid);
int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
if (ppi && elapsed > ppi->longest_completion_wait[w])
/*
* kill the trace, we don't need it after completion
*/
- rb_erase(&iot->rb_node, &rb_track_root);
+ rb_erase(&iot->rb_node, &pdi->rb_track);
free(iot);
return elapsed;
}
-static struct io_stats *find_process_io_stats(__u32 pid, char *name)
+static struct io_stats *find_process_io_stats(pid_t pid)
{
- struct per_process_info *ppi = find_process_by_pid(pid);
+ struct per_process_info *ppi = find_ppi(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);
+ ppi->ppm = find_ppm(pid);
+ if (!ppi->ppm)
+ ppi->ppm = add_ppm_hash(pid, "unknown");
+ add_ppi_to_hash(ppi);
+ add_ppi_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 == id)
- return &devices[i];
-
- if (resize_devices(NULL) != 0)
- return NULL;
-
- pdi = &devices[ndevices - 1];
- pdi->id = id;
- 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));
+ snprintf(buffer, size, "%d,%d",MAJOR(pdi->dev),MINOR(pdi->dev));
return buffer;
}
{
if (rw) {
ios->mwrites++;
- ios->qwrite_kb += t->bytes >> 10;
+ ios->mwrite_kb += t_kb(t);
+ ios->mwrite_b += t_b(t);
} else {
ios->mreads++;
- ios->qread_kb += t->bytes >> 10;
+ ios->mread_kb += t_kb(t);
+ ios->mread_b += t_b(t);
}
}
__account_m(&pci->io_stats, t, rw);
if (per_process_stats) {
- struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+ struct io_stats *ios = find_process_io_stats(t->pid);
__account_m(ios, t, rw);
}
}
-static inline void __account_q(struct io_stats *ios, struct blk_io_trace *t,
- int rw)
+static inline void __account_pc_queue(struct io_stats *ios,
+ struct blk_io_trace *t, int rw)
{
if (rw) {
- ios->qwrites++;
- ios->qwrite_kb += t->bytes >> 10;
+ ios->qwrites_pc++;
+ ios->qwrite_kb_pc += t_kb(t);
+ ios->qwrite_b_pc += t_b(t);
} else {
- ios->qreads++;
- ios->qread_kb += t->bytes >> 10;
+ ios->qreads_pc++;
+ ios->qread_kb += t_kb(t);
+ ios->qread_b_pc += t_b(t);
}
}
-static inline void account_q(struct blk_io_trace *t, struct per_cpu_info *pci,
- int rw)
+static inline void account_pc_queue(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int rw)
{
- __account_q(&pci->io_stats, t, rw);
+ __account_pc_queue(&pci->io_stats, t, rw);
if (per_process_stats) {
- struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+ struct io_stats *ios = find_process_io_stats(t->pid);
- __account_q(ios, t, rw);
+ __account_pc_queue(ios, t, rw);
}
}
-static inline void __account_c(struct io_stats *ios, int rw, unsigned int bytes)
+static inline void __account_pc_issue(struct io_stats *ios, int rw,
+ unsigned int bytes)
{
if (rw) {
- ios->cwrites++;
- ios->cwrite_kb += bytes >> 10;
+ ios->iwrites_pc++;
+ ios->iwrite_kb_pc += bytes >> 10;
+ ios->iwrite_b_pc += bytes & 1023;
} else {
- ios->creads++;
- ios->cread_kb += bytes >> 10;
+ ios->ireads_pc++;
+ ios->iread_kb_pc += bytes >> 10;
+ ios->iread_b_pc += bytes & 1023;
}
}
-static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
- int rw, int bytes)
+static inline void account_pc_issue(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int rw)
{
- __account_c(&pci->io_stats, rw, bytes);
+ __account_pc_issue(&pci->io_stats, rw, t->bytes);
if (per_process_stats) {
- struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+ struct io_stats *ios = find_process_io_stats(t->pid);
- __account_c(ios, rw, bytes);
+ __account_pc_issue(ios, rw, t->bytes);
}
}
-static inline void __account_i(struct io_stats *ios, int rw, unsigned int bytes)
+static inline void __account_pc_requeue(struct io_stats *ios,
+ struct blk_io_trace *t, int rw)
{
if (rw) {
- ios->iwrites++;
- ios->iwrite_kb += bytes >> 10;
+ ios->wrqueue_pc++;
+ ios->iwrite_kb_pc -= t_kb(t);
+ ios->iwrite_b_pc -= t_b(t);
} else {
- ios->ireads++;
- ios->iread_kb += bytes >> 10;
+ ios->rrqueue_pc++;
+ ios->iread_kb_pc -= t_kb(t);
+ ios->iread_b_pc -= t_b(t);
}
}
-static inline void account_i(struct blk_io_trace *t, struct per_cpu_info *pci,
- int rw)
+static inline void account_pc_requeue(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int rw)
{
- __account_i(&pci->io_stats, rw, t->bytes);
+ __account_pc_requeue(&pci->io_stats, t, rw);
if (per_process_stats) {
- struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+ struct io_stats *ios = find_process_io_stats(t->pid);
- __account_i(ios, rw, t->bytes);
+ __account_pc_requeue(ios, t, rw);
}
}
-static inline void __account_unplug(struct io_stats *ios, int timer)
+static inline void __account_pc_c(struct io_stats *ios, int rw)
{
- if (timer)
- ios->timer_unplugs++;
+ if (rw)
+ ios->cwrites_pc++;
else
- ios->io_unplugs++;
+ ios->creads_pc++;
}
-static inline void account_unplug(struct blk_io_trace *t,
- struct per_cpu_info *pci, int timer)
+static inline void account_pc_c(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int rw)
{
- __account_unplug(&pci->io_stats, timer);
+ __account_pc_c(&pci->io_stats, rw);
if (per_process_stats) {
- struct io_stats *ios = find_process_io_stats(t->pid, t->comm);
+ struct io_stats *ios = find_process_io_stats(t->pid);
- __account_unplug(ios, timer);
+ __account_pc_c(ios, rw);
}
}
-#define VALID_SPECS "BCDFGMPQRSTU"
-static char *override_format[256];
-static inline int valid_spec(int spec)
+static inline void __account_queue(struct io_stats *ios, struct blk_io_trace *t,
+ int rw)
{
- return strchr(VALID_SPECS, spec) != NULL;
+ if (rw) {
+ ios->qwrites++;
+ ios->qwrite_kb += t_kb(t);
+ ios->qwrite_b += t_b(t);
+ } else {
+ ios->qreads++;
+ ios->qread_kb += t_kb(t);
+ ios->qread_b += t_b(t);
+ }
}
-static void set_all_format_specs(char *optarg)
+static inline void account_queue(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int rw)
{
- char *p;
+ __account_queue(&pci->io_stats, t, rw);
+
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid);
- for (p = VALID_SPECS; *p; p++)
- if (override_format[(int)(*p)] == NULL)
- override_format[(int)(*p)] = strdup(optarg);
+ __account_queue(ios, t, rw);
+ }
}
-static int add_format_spec(char *optarg)
+static inline void __account_c(struct io_stats *ios, int rw, int bytes)
{
- int spec = optarg[0];
-
- if (!valid_spec(spec)) {
- fprintf(stderr,"Bad format specifier %c\n", spec);
- return 1;
- }
- if (optarg[1] != ',') {
- fprintf(stderr,"Bad format specifier - need ',' %s\n", optarg);
- return 1;
- }
- optarg += 2;
- if (*optarg == '\0') {
- fprintf(stderr,"Bad format specifier - need fmt %s\n", optarg);
- return 1;
+ if (rw) {
+ ios->cwrites++;
+ ios->cwrite_kb += bytes >> 10;
+ ios->cwrite_b += bytes & 1023;
+ } else {
+ ios->creads++;
+ ios->cread_kb += bytes >> 10;
+ ios->cread_b += bytes & 1023;
}
-
- /*
- * Set both merges (front and back)
- */
- if (spec == 'M') {
- override_format['B'] = strdup(optarg);
- override_format['M'] = strdup(optarg);
- } else
- override_format[spec] = strdup(optarg);
-
- return 0;
}
-static void print_field(char *act, struct per_cpu_info *pci,
- struct blk_io_trace *t, unsigned long long elapsed,
- int pdu_len, unsigned char *pdu_buf, char field,
- int minus, int has_w, int width)
+static inline void account_c(struct blk_io_trace *t, struct per_cpu_info *pci,
+ int rw, int bytes)
{
- char format[64];
+ __account_c(&pci->io_stats, rw, bytes);
- if (has_w) {
- if (minus)
- sprintf(format, "%%-%d", width);
- else
- sprintf(format, "%%%d", width);
- } else
- sprintf(format, "%%");
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid);
- switch (field) {
- case 'a':
- fprintf(ofp, strcat(format, "s"), act);
- break;
- case 'c':
- fprintf(ofp, strcat(format, "d"), pci->cpu);
- break;
- case 'C':
- fprintf(ofp, strcat(format, "s"), t->comm);
- break;
- case 'd': {
- char rwbs[4];
- int i = 0;
- 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);
- if (w)
- rwbs[i++] = 'W';
- else
- rwbs[i++] = 'R';
- if (b)
- rwbs[i++] = 'B';
- if (s)
- rwbs[i++] = 'S';
- rwbs[i] = '\0';
- fprintf(ofp, strcat(format, "s"), rwbs);
- break;
- }
- case 'D': /* format width ignored */
- fprintf(ofp,"%3d,%-3d", MAJOR(t->device), MINOR(t->device));
- break;
- case 'e':
- fprintf(ofp, strcat(format, "d"), t->error);
- break;
- case 'M':
- fprintf(ofp, strcat(format, "d"), MAJOR(t->device));
- break;
- case 'm':
- fprintf(ofp, strcat(format, "d"), MINOR(t->device));
- break;
- case 'n':
- fprintf(ofp, strcat(format, "u"), t->bytes >> 9);
- break;
- case 'p':
- fprintf(ofp, strcat(format, "u"), t->pid);
- break;
- case 'P': /* format width ignored */
- if ((pdu_len > 0) && (pdu_buf != NULL)) {
- int i;
- unsigned char *p = pdu_buf;
- for (i = 0; i < pdu_len; i++)
- fprintf(ofp, "%02x ", *p++);
- }
- break;
- case 's':
- fprintf(ofp, strcat(format, "ld"), t->sequence);
- break;
- case 'S':
- fprintf(ofp, strcat(format, "lu"), t->sector);
- break;
- case 't':
- sprintf(format, "%%0%dlu", has_w ? width : 9);
- fprintf(ofp, format, NANO_SECONDS(t->time));
- break;
- case 'T':
- fprintf(ofp, strcat(format, "d"), SECONDS(t->time));
- break;
- case 'u':
- if (elapsed == -1ULL) {
- fprintf(stderr, "Expecting elapsed value\n");
- exit(1);
- }
- fprintf(ofp, strcat(format, "llu"), elapsed / 1000);
- break;
- case 'U': {
- __u64 *depth = (__u64 *) ((char *) t + sizeof(*t));
- fprintf(ofp, strcat(format, "u"),
- (unsigned int) be64_to_cpu(*depth));
- break;
- }
- default:
- fprintf(ofp,strcat(format, "c"), field);
- break;
+ __account_c(ios, rw, bytes);
}
}
-static char *parse_field(char *act, struct per_cpu_info *pci,
- struct blk_io_trace *t, unsigned long long elapsed,
- int pdu_len, unsigned char *pdu_buf,
- char *master_format)
+static inline void __account_issue(struct io_stats *ios, int rw,
+ unsigned int bytes)
{
- int minus = 0;
- int has_w = 0;
- int width = 0;
- char *p = master_format;
-
- if (*p == '-') {
- minus = 1;
- p++;
- }
- if (isdigit(*p)) {
- has_w = 1;
- do {
- width = (width * 10) + (*p++ - '0');
- } while ((*p) && (isdigit(*p)));
- }
- if (*p) {
- print_field(act, pci, t, elapsed, pdu_len, pdu_buf, *p++,
- minus, has_w, width);
+ if (rw) {
+ ios->iwrites++;
+ ios->iwrite_kb += bytes >> 10;
+ ios->iwrite_b += bytes & 1023;
+ } else {
+ ios->ireads++;
+ ios->iread_kb += bytes >> 10;
+ ios->iread_b += bytes & 1023;
}
- return p;
}
-static char *fmt_select(int fmt_spec, struct blk_io_trace *t,
- unsigned long long elapsed)
+static inline void account_issue(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int rw)
{
- char *fmt;
- char scratch_format[1024];
-
- if (override_format[fmt_spec] != NULL)
- return override_format[fmt_spec];
-
- switch (fmt_spec) {
- case 'C': /* Complete */
- if (t->action & BLK_TC_ACT(BLK_TC_PC)) {
- strcpy(scratch_format, HEADER);
- strcat(scratch_format, "%P");
- } else {
- strcpy(scratch_format, HEADER "%S + %n ");
- if (elapsed != -1ULL)
- strcat(scratch_format, "(%8u) ");
- }
- strcat(scratch_format, "[%e]\n");
- fmt = scratch_format;
- break;
-
- case 'D': /* Issue */
- if (t->action & BLK_TC_ACT(BLK_TC_PC)) {
- strcpy(scratch_format, HEADER);
- strcat(scratch_format, "%P");
- } else {
- strcpy(scratch_format, HEADER "%S + %n ");
- if (elapsed != -1ULL)
- strcat(scratch_format, "(%8u) ");
- }
- strcat(scratch_format,"[%C]\n");
- fmt = scratch_format;
- break;
+ __account_issue(&pci->io_stats, rw, t->bytes);
- case 'Q': /* Queue */
- strcpy(scratch_format, HEADER "%S + %n ");
- if (elapsed != -1ULL)
- strcat(scratch_format, "(%8u) ");
- strcat(scratch_format,"[%C]\n");
- fmt = scratch_format;
- break;
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid);
- case 'B': /* Back merge */
- case 'F': /* Front merge */
- case 'M': /* Front or back merge */
- fmt = HEADER "%S + %n [%C]\n";
- break;
+ __account_issue(ios, rw, t->bytes);
+ }
+}
- case 'P': /* Plug */
- fmt = HEADER "[%C]\n";
- break;
+static inline void __account_unplug(struct io_stats *ios, int timer)
+{
+ if (timer)
+ ios->timer_unplugs++;
+ else
+ ios->io_unplugs++;
+}
- case 'G': /* Get request */
- case 'S': /* Sleep request */
- fmt = HEADER "%S + %n [%C]\n";
- break;
+static inline void account_unplug(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int timer)
+{
+ __account_unplug(&pci->io_stats, timer);
- case 'U': /* Unplug IO */
- case 'T': /* Unplug timer */
- fmt = HEADER "[%C] %U\n";
- break;
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid);
- default:
- fprintf(stderr,"FATAL: Invalid format spec %c\n", fmt_spec);
- exit(1);
- /*NOTREACHED*/
+ __account_unplug(ios, timer);
}
+}
- return fmt;
+static inline void __account_requeue(struct io_stats *ios,
+ struct blk_io_trace *t, int rw)
+{
+ if (rw) {
+ ios->wrqueue++;
+ ios->iwrite_kb -= t_kb(t);
+ ios->iwrite_b -= t_b(t);
+ } else {
+ ios->rrqueue++;
+ ios->iread_kb -= t_kb(t);
+ ios->iread_b -= t_b(t);
+ }
}
-static void process_fmt(char *act, struct per_cpu_info *pci,
- struct blk_io_trace *t, unsigned long long elapsed,
- int pdu_len, unsigned char *pdu_buf)
+static inline void account_requeue(struct blk_io_trace *t,
+ struct per_cpu_info *pci, int rw)
{
- char *p = fmt_select(act[0], t, elapsed);
+ __account_requeue(&pci->io_stats, t, rw);
- while (*p) {
- switch (*p) {
- case '%': /* Field specifier */
- p++;
- if (*p == '%')
- fprintf(ofp, "%c", *p++);
- else if (!*p)
- fprintf(ofp, "%c", '%');
- else
- p = parse_field(act, pci, t, elapsed,
- pdu_len, pdu_buf, p);
- break;
- case '\\': { /* escape */
- switch (p[1]) {
- case 'b': fprintf(ofp, "\b"); break;
- case 'n': fprintf(ofp, "\n"); break;
- case 'r': fprintf(ofp, "\r"); break;
- case 't': fprintf(ofp, "\t"); break;
- default:
- fprintf(stderr,
- "Invalid escape char in format %c\n",
- p[1]);
- exit(1);
- /*NOTREACHED*/
- }
- p += 2;
- break;
- }
- default:
- fprintf(ofp, "%c", *p++);
- break;
- }
+ if (per_process_stats) {
+ struct io_stats *ios = find_process_io_stats(t->pid);
+
+ __account_requeue(ios, t, rw);
}
}
-static void log_complete(struct per_cpu_info *pci, struct blk_io_trace *t,
- char *act)
+static void log_complete(struct per_dev_info *pdi, struct per_cpu_info *pci,
+ struct blk_io_trace *t, char *act)
{
- process_fmt(act, pci, t, log_track_complete(t), 0, NULL);
+ process_fmt(act, pci, t, log_track_complete(pdi, t), 0, NULL);
+}
+
+static void log_insert(struct per_dev_info *pdi, struct per_cpu_info *pci,
+ struct blk_io_trace *t, char *act)
+{
+ process_fmt(act, pci, t, log_track_insert(pdi, t), 0, NULL);
}
static void log_queue(struct per_cpu_info *pci, struct blk_io_trace *t,
char *act)
{
- process_fmt(act, pci, t, log_track_queue(t), 0, NULL);
+ process_fmt(act, pci, t, -1, 0, NULL);
}
-static void log_issue(struct per_cpu_info *pci, struct blk_io_trace *t,
- char *act)
+static void log_issue(struct per_dev_info *pdi, struct per_cpu_info *pci,
+ struct blk_io_trace *t, char *act)
{
- process_fmt(act, pci, t, log_track_issue(t), 0, NULL);
+ process_fmt(act, pci, t, log_track_issue(pdi, t), 0, NULL);
}
-static void log_merge(struct per_cpu_info *pci, struct blk_io_trace *t,
- char *act)
+static void log_merge(struct per_dev_info *pdi, struct per_cpu_info *pci,
+ struct blk_io_trace *t, char *act)
{
if (act[0] == 'F')
- log_track_frontmerge(t);
+ log_track_frontmerge(pdi, t);
process_fmt(act, pci, t, -1ULL, 0, NULL);
}
process_fmt(act, pci, t, -1ULL, 0, NULL);
}
+static void log_split(struct per_cpu_info *pci, struct blk_io_trace *t,
+ char *act)
+{
+ process_fmt(act, pci, t, -1ULL, 0, NULL);
+}
+
static void log_pc(struct per_cpu_info *pci, struct blk_io_trace *t, char *act)
{
unsigned char *buf = (unsigned char *) t + sizeof(*t);
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_dev_info *pdi,
+ struct per_cpu_info *pci)
{
- int ret = 0;
+ int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+ int act = t->action & 0xffff;
- switch (t->action & 0xffff) {
+ switch (act) {
case __BLK_TA_QUEUE:
log_generic(pci, t, "Q");
+ account_pc_queue(t, pci, w);
break;
case __BLK_TA_GETRQ:
log_generic(pci, t, "G");
log_generic(pci, t, "S");
break;
case __BLK_TA_REQUEUE:
+ /*
+ * can happen if we miss traces, don't let it go
+ * below zero
+ */
+ if (pdi->cur_depth[w])
+ pdi->cur_depth[w]--;
+ account_pc_requeue(t, pci, w);
log_generic(pci, t, "R");
break;
case __BLK_TA_ISSUE:
+ account_pc_issue(t, pci, w);
+ pdi->cur_depth[w]++;
+ if (pdi->cur_depth[w] > pdi->max_depth[w])
+ pdi->max_depth[w] = pdi->cur_depth[w];
log_pc(pci, t, "D");
break;
case __BLK_TA_COMPLETE:
+ if (pdi->cur_depth[w])
+ pdi->cur_depth[w]--;
log_pc(pci, t, "C");
+ account_pc_c(t, pci, w);
+ 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 0;
}
-static void dump_trace_fs(struct blk_io_trace *t, struct per_cpu_info *pci)
+static void dump_trace_fs(struct blk_io_trace *t, struct per_dev_info *pdi,
+ struct per_cpu_info *pci)
{
- int w = t->action & BLK_TC_ACT(BLK_TC_WRITE);
+ int w = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
int act = t->action & 0xffff;
switch (act) {
case __BLK_TA_QUEUE:
- account_q(t, pci, w);
+ log_track_queue(pdi, t);
+ account_queue(t, pci, w);
log_queue(pci, t, "Q");
break;
+ case __BLK_TA_INSERT:
+ log_insert(pdi, pci, t, "I");
+ break;
case __BLK_TA_BACKMERGE:
account_m(t, pci, w);
- log_merge(pci, t, "M");
+ log_merge(pdi, pci, t, "M");
break;
case __BLK_TA_FRONTMERGE:
account_m(t, pci, w);
- log_merge(pci, t, "F");
+ log_merge(pdi, pci, t, "F");
break;
case __BLK_TA_GETRQ:
- log_track_getrq(t);
+ log_track_getrq(pdi, t);
log_generic(pci, t, "G");
break;
case __BLK_TA_SLEEPRQ:
log_generic(pci, t, "S");
break;
case __BLK_TA_REQUEUE:
- account_c(t, pci, w, -t->bytes);
+ /*
+ * can happen if we miss traces, don't let it go
+ * below zero
+ */
+ if (pdi->cur_depth[w])
+ pdi->cur_depth[w]--;
+ account_requeue(t, pci, w);
log_queue(pci, t, "R");
break;
case __BLK_TA_ISSUE:
- account_i(t, pci, w);
- log_issue(pci, t, "D");
+ account_issue(t, pci, w);
+ pdi->cur_depth[w]++;
+ if (pdi->cur_depth[w] > pdi->max_depth[w])
+ pdi->max_depth[w] = pdi->cur_depth[w];
+ log_issue(pdi, pci, t, "D");
break;
case __BLK_TA_COMPLETE:
+ if (pdi->cur_depth[w])
+ pdi->cur_depth[w]--;
account_c(t, pci, w, t->bytes);
- log_complete(pci, t, "C");
+ log_complete(pdi, pci, t, "C");
break;
case __BLK_TA_PLUG:
log_action(pci, t, "P");
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;
+ case __BLK_TA_REMAP:
+ log_generic(pci, t, "A");
+ break;
+ case __BLK_TA_DRV_DATA:
+ have_drv_data = 1;
+ /* dump to binary file only */
+ 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,
- struct per_dev_info *pdi)
+static void dump_trace(struct blk_io_trace *t, struct per_cpu_info *pci,
+ struct per_dev_info *pdi)
{
- int ret = 0;
+ if (text_output) {
+ if (t->action == BLK_TN_MESSAGE)
+ handle_notify(t);
+ else if (t->action & BLK_TC_ACT(BLK_TC_PC))
+ dump_trace_pc(t, pdi, pci);
+ else
+ dump_trace_fs(t, pdi, pci);
+ }
- if (t->action & BLK_TC_ACT(BLK_TC_PC))
- ret = dump_trace_pc(t, pci);
- else
- dump_trace_fs(t, pci);
+ if (!pdi->events)
+ pdi->first_reported_time = t->time;
+
+ pdi->events++;
+
+ if (bin_output_msgs ||
+ !(t->action & BLK_TC_ACT(BLK_TC_NOTIFY) &&
+ t->action == BLK_TN_MESSAGE))
+ output_binary(t, sizeof(*t) + t->pdu_len);
+}
+
+/*
+ * print in a proper way, not too small and not too big. if more than
+ * 1000,000K, turn into M and so on
+ */
+static char *size_cnv(char *dst, unsigned long long num, int in_kb)
+{
+ char suff[] = { '\0', 'K', 'M', 'G', 'P' };
+ unsigned int i = 0;
+
+ if (in_kb)
+ i++;
+
+ while (num > 1000 * 1000ULL && (i < sizeof(suff) - 1)) {
+ i++;
+ num /= 1000;
+ }
- pdi->events++;
- return ret;
+ sprintf(dst, "%'8Lu%c", num, suff[i]);
+ return dst;
}
-static void dump_io_stats(struct io_stats *ios, char *msg)
+static void dump_io_stats(struct per_dev_info *pdi, struct io_stats *ios,
+ char *msg)
{
- fprintf(ofp, "%s\n", msg);
+ static char x[256], y[256];
- 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, "%s\n", msg);
- 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, " Reads Queued: %s, %siB\t",
+ size_cnv(x, ios->qreads, 0),
+ size_cnv(y, ios->qread_kb + (ios->qread_b>>10), 1));
+ fprintf(ofp, " Writes Queued: %s, %siB\n",
+ size_cnv(x, ios->qwrites, 0),
+ size_cnv(y, ios->qwrite_kb + (ios->qwrite_b>>10), 1));
+ fprintf(ofp, " Read Dispatches: %s, %siB\t",
+ size_cnv(x, ios->ireads, 0),
+ size_cnv(y, ios->iread_kb + (ios->iread_b>>10), 1));
+ fprintf(ofp, " Write Dispatches: %s, %siB\n",
+ size_cnv(x, ios->iwrites, 0),
+ size_cnv(y, ios->iwrite_kb + (ios->iwrite_b>>10), 1));
+ fprintf(ofp, " Reads Requeued: %s\t\t", size_cnv(x, ios->rrqueue, 0));
+ fprintf(ofp, " Writes Requeued: %s\n", size_cnv(x, ios->wrqueue, 0));
+ fprintf(ofp, " Reads Completed: %s, %siB\t",
+ size_cnv(x, ios->creads, 0),
+ size_cnv(y, ios->cread_kb + (ios->cread_b>>10), 1));
+ fprintf(ofp, " Writes Completed: %s, %siB\n",
+ size_cnv(x, ios->cwrites, 0),
+ size_cnv(y, ios->cwrite_kb + (ios->cwrite_b>>10), 1));
+ fprintf(ofp, " Read Merges: %s, %siB\t",
+ size_cnv(x, ios->mreads, 0),
+ size_cnv(y, ios->mread_kb + (ios->mread_b>>10), 1));
+ fprintf(ofp, " Write Merges: %s, %siB\n",
+ size_cnv(x, ios->mwrites, 0),
+ size_cnv(y, ios->mwrite_kb + (ios->mwrite_b>>10), 1));
+ if (pdi) {
+ fprintf(ofp, " Read depth: %'8u%8c\t", pdi->max_depth[0], ' ');
+ fprintf(ofp, " Write depth: %'8u\n", pdi->max_depth[1]);
+ }
+ if (ios->qreads_pc || ios->qwrites_pc || ios->ireads_pc || ios->iwrites_pc ||
+ ios->rrqueue_pc || ios->wrqueue_pc || ios->creads_pc || ios->cwrites_pc) {
+ fprintf(ofp, " PC Reads Queued: %s, %siB\t",
+ size_cnv(x, ios->qreads_pc, 0),
+ size_cnv(y,
+ ios->qread_kb_pc + (ios->qread_b_pc>>10), 1));
+ fprintf(ofp, " PC Writes Queued: %s, %siB\n",
+ size_cnv(x, ios->qwrites_pc, 0),
+ size_cnv(y,
+ ios->qwrite_kb_pc + (ios->qwrite_b_pc>>10), 1));
+ fprintf(ofp, " PC Read Disp.: %s, %siB\t",
+ size_cnv(x, ios->ireads_pc, 0),
+ size_cnv(y,
+ ios->iread_kb_pc + (ios->iread_b_pc>>10), 1));
+ fprintf(ofp, " PC Write Disp.: %s, %siB\n",
+ size_cnv(x, ios->iwrites_pc, 0),
+ size_cnv(y,
+ ios->iwrite_kb_pc + (ios->iwrite_b_pc>>10),
+ 1));
+ fprintf(ofp, " PC Reads Req.: %s\t\t", size_cnv(x, ios->rrqueue_pc, 0));
+ fprintf(ofp, " PC Writes Req.: %s\n", size_cnv(x, ios->wrqueue_pc, 0));
+ fprintf(ofp, " PC Reads Compl.: %s\t\t", size_cnv(x, ios->creads_pc, 0));
+ fprintf(ofp, " PC Writes Compl.: %s\n", size_cnv(x, ios->cwrites_pc, 0));
+ }
fprintf(ofp, " IO unplugs: %'8lu%8c\t", ios->io_unplugs, ' ');
fprintf(ofp, " Timer unplugs: %'8lu\n", ios->timer_unplugs);
}
struct per_process_info *ppi2 = *((struct per_process_info **) p2);
int res;
- res = strverscmp(ppi1->name, ppi2->name);
+ res = strverscmp(ppi1->ppm->comm, ppi2->ppm->comm);
if (!res)
- res = ppi1->pid < ppi2->pid;
+ res = ppi1->ppm->pid > ppi2->ppm->pid;
return res;
}
+static int ppi_event_compare(const void *p1, const void *p2)
+{
+ struct per_process_info *ppi1 = *((struct per_process_info **) p1);
+ struct per_process_info *ppi2 = *((struct per_process_info **) p2);
+ struct io_stats *ios1 = &ppi1->io_stats;
+ struct io_stats *ios2 = &ppi2->io_stats;
+ unsigned long io1, io2;
+ unsigned long long kb1,kb2;
+ int sort_by_kb = 1;
+
+ io1 = io2 = 0;
+ kb1 = kb2 = 0;
+
+ switch (per_process_stats_event) {
+ case SORT_PROG_EVENT_QKB: /* KB: Queued read and write */
+ kb1 = ios1->qwrite_kb + (ios1->qwrite_b>>10) +
+ ios1->qread_kb + (ios1->qread_b>>10);
+ kb2 = ios2->qwrite_kb + (ios2->qwrite_b>>10) +
+ ios2->qread_kb + (ios2->qread_b>>10);
+ break;
+ case SORT_PROG_EVENT_RKB: /* KB: Queued Read */
+ kb1 = ios1->qread_kb + (ios1->qread_b>>10);
+ kb2 = ios2->qread_kb + (ios2->qread_b>>10);
+ break;
+ case SORT_PROG_EVENT_WKB: /* KB: Queued Write */
+ kb1 = ios1->qwrite_kb + (ios1->qwrite_b>>10);
+ kb2 = ios2->qwrite_kb + (ios2->qwrite_b>>10);
+ break;
+ case SORT_PROG_EVENT_CKB: /* KB: Complete */
+ kb1 = ios1->cwrite_kb + (ios1->cwrite_b>>10) +
+ ios1->cread_kb + (ios1->cread_b>>10);
+ kb2 = ios2->cwrite_kb + (ios2->cwrite_b>>10) +
+ ios2->cread_kb + (ios2->cread_b>>10);
+ break;
+ case SORT_PROG_EVENT_QIO: /* IO: Queued read and write */
+ sort_by_kb = 0;
+ io1 = ios1->qreads + ios1->qwrites;
+ io2 = ios2->qreads + ios2->qwrites;
+ break;
+ case SORT_PROG_EVENT_RIO: /* IO: Queued Read */
+ sort_by_kb = 0;
+ io1 = ios1->qreads;
+ io2 = ios2->qreads;
+ break;
+ case SORT_PROG_EVENT_WIO: /* IO: Queued Write */
+ sort_by_kb = 0;
+ io1 = ios1->qwrites;
+ io2 = ios2->qwrites;
+ break;
+ case SORT_PROG_EVENT_CIO: /* IO: Complete */
+ sort_by_kb = 0;
+ io1 = ios1->creads + ios1->cwrites;
+ io2 = ios2->creads + ios2->cwrites;
+ break;
+ }
+
+
+ /* compare kb */
+ if (sort_by_kb) {
+ if (kb1 > kb2)
+ return 1;
+ else if (kb1 == kb2)
+ return 0;
+ return -1;
+ }
+
+ /* compare io */
+ if (io1 > io2)
+ return 1;
+ else if (io1 == io2)
+ return 0;
+ return -1;
+}
+
+static int ppi_compare(const void *p1, const void *p2)
+{
+ if (per_process_stats_event == SORT_PROG_EVENT_N)
+ return ppi_name_compare(p1, p2);
+
+ return ppi_event_compare(p1, p2);
+}
+
static void sort_process_list(void)
{
struct per_process_info **ppis;
ppi = ppi_list;
while (ppi) {
- ppis[i] = ppi;
+ ppis[i++] = ppi;
ppi = ppi->list_next;
}
- qsort(ppis, ppi_list_entries, sizeof(*ppi), ppi_name_compare);
+ qsort(ppis, ppi_list_entries, sizeof(ppi), ppi_compare);
i = ppi_list_entries - 1;
ppi_list = NULL;
ppi_list = ppi;
i--;
}
+
+ free(ppis);
}
static void show_process_stats(void)
ppi = ppi_list;
while (ppi) {
- dump_io_stats(&ppi->io_stats, ppi->name);
+ struct process_pid_map *ppm = ppi->ppm;
+ char name[64];
+
+ if (ppi->more_than_one)
+ sprintf(name, "%s (%u, ...)", ppm->comm, ppm->pid);
+ else
+ sprintf(name, "%s (%u)", ppm->comm, ppm->pid);
+
+ dump_io_stats(NULL, &ppi->io_stats, name);
dump_wait_stats(ppi);
ppi = ppi->list_next;
}
struct per_dev_info *pdi;
struct per_cpu_info *pci;
struct io_stats total, *ios;
+ unsigned long long rrate, wrate, msec;
int i, j, pci_events;
char line[3 + 8/*cpu*/ + 2 + 32/*dev*/ + 3];
char name[32];
+ double ratio;
for (pdi = devices, i = 0; i < ndevices; i++, pdi++) {
total.mwrites += ios->mwrites;
total.ireads += ios->ireads;
total.iwrites += ios->iwrites;
+ total.rrqueue += ios->rrqueue;
+ total.wrqueue += ios->wrqueue;
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.mread_kb += ios->mread_kb;
+ total.mwrite_kb += ios->mwrite_kb;
+ total.qread_b += ios->qread_b;
+ total.qwrite_b += ios->qwrite_b;
+ total.cread_b += ios->cread_b;
+ total.cwrite_b += ios->cwrite_b;
+ total.iread_b += ios->iread_b;
+ total.iwrite_b += ios->iwrite_b;
+ total.mread_b += ios->mread_b;
+ total.mwrite_b += ios->mwrite_b;
+
+ total.qreads_pc += ios->qreads_pc;
+ total.qwrites_pc += ios->qwrites_pc;
+ total.creads_pc += ios->creads_pc;
+ total.cwrites_pc += ios->cwrites_pc;
+ total.ireads_pc += ios->ireads_pc;
+ total.iwrites_pc += ios->iwrites_pc;
+ total.rrqueue_pc += ios->rrqueue_pc;
+ total.wrqueue_pc += ios->wrqueue_pc;
+ total.qread_kb_pc += ios->qread_kb_pc;
+ total.qwrite_kb_pc += ios->qwrite_kb_pc;
+ total.iread_kb_pc += ios->iread_kb_pc;
+ total.iwrite_kb_pc += ios->iwrite_kb_pc;
+ total.qread_b_pc += ios->qread_b_pc;
+ total.qwrite_b_pc += ios->qwrite_b_pc;
+ total.iread_b_pc += ios->iread_b_pc;
+ total.iwrite_b_pc += ios->iwrite_b_pc;
+
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);
+ dump_io_stats(pdi, ios, line);
pci_events++;
}
fprintf(ofp, "\n");
snprintf(line, sizeof(line) - 1, "Total (%s):",
get_dev_name(pdi, name, sizeof(name)));
- dump_io_stats(&total, line);
+ dump_io_stats(NULL, &total, line);
}
- fprintf(ofp, "\nEvents (%s): %'Lu entries, %'lu skips\n",
- get_dev_name(pdi, line, sizeof(line)), pdi->events,
- pdi->skips);
+ wrate = rrate = 0;
+ msec = (pdi->last_reported_time - pdi->first_reported_time) / 1000000;
+ if (msec) {
+ rrate = ((1000 * total.cread_kb) + total.cread_b) /
+ msec;
+ wrate = ((1000 * total.cwrite_kb) + total.cwrite_b) /
+ msec;
+ }
+
+ fprintf(ofp, "\nThroughput (R/W): %'LuKiB/s / %'LuKiB/s\n",
+ rrate, wrate);
+ fprintf(ofp, "Events (%s): %'Lu entries\n",
+ get_dev_name(pdi, line, sizeof(line)), pdi->events);
+
+ collect_pdi_skips(pdi);
+ if (!pdi->skips && !pdi->events)
+ ratio = 0.0;
+ else
+ ratio = 100.0 * ((double)pdi->seq_skips /
+ (double)(pdi->events + pdi->seq_skips));
+ fprintf(ofp, "Skips: %'lu forward (%'llu - %5.1lf%%)\n",
+ pdi->skips, pdi->seq_skips, ratio);
+ }
+}
+
+static void correct_abs_start_time(void)
+{
+ long delta = genesis_time - start_timestamp;
+
+ abs_start_time.tv_sec += SECONDS(delta);
+ abs_start_time.tv_nsec += NANO_SECONDS(delta);
+ if (abs_start_time.tv_nsec < 0) {
+ abs_start_time.tv_nsec += 1000000000;
+ abs_start_time.tv_sec -= 1;
+ } else
+ if (abs_start_time.tv_nsec > 1000000000) {
+ abs_start_time.tv_nsec -= 1000000000;
+ abs_start_time.tv_sec += 1;
}
}
-static struct blk_io_trace *find_trace(void *p, unsigned long offset)
+static void find_genesis(void)
{
- unsigned long max_offset = offset;
- unsigned long off;
- struct blk_io_trace *bit;
- __u32 magic;
+ struct trace *t = trace_list;
- for (off = 0; off < max_offset; off++) {
- bit = p + off;
+ genesis_time = -1ULL;
+ while (t != NULL) {
+ if (t->bit->time < genesis_time)
+ genesis_time = t->bit->time;
- magic = be32_to_cpu(bit->magic);
- if ((magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
- return bit;
+ t = t->next;
}
- return NULL;
+ /* The time stamp record will usually be the first
+ * record in the trace, but not always.
+ */
+ if (start_timestamp
+ && start_timestamp != genesis_time) {
+ correct_abs_start_time();
+ }
}
-static int sort_entries(void)
+static inline int check_stopwatch(struct blk_io_trace *bit)
{
- struct per_dev_info *pdi;
+ if (bit->time < stopwatch_end &&
+ bit->time >= stopwatch_start)
+ return 0;
+
+ return 1;
+}
+
+/*
+ * return youngest entry read
+ */
+static int sort_entries(unsigned long long *youngest)
+{
+ struct per_dev_info *pdi = NULL;
+ struct per_cpu_info *pci = NULL;
+ struct trace *t;
+
+ if (!genesis_time)
+ find_genesis();
+
+ *youngest = 0;
+ while ((t = trace_list) != NULL) {
+ struct blk_io_trace *bit = t->bit;
+
+ trace_list = t->next;
+
+ bit->time -= genesis_time;
+
+ if (bit->time < *youngest || !*youngest)
+ *youngest = bit->time;
+
+ if (!pdi || pdi->dev != bit->device) {
+ pdi = get_dev_info(bit->device);
+ pci = NULL;
+ }
+
+ if (!pci || pci->cpu != bit->cpu)
+ pci = get_cpu_info(pdi, bit->cpu);
+
+ if (bit->sequence < pci->smallest_seq_read)
+ pci->smallest_seq_read = bit->sequence;
+
+ if (check_stopwatch(bit)) {
+ bit_free(bit);
+ t_free(t);
+ continue;
+ }
+
+ if (trace_rb_insert_sort(t))
+ return -1;
+ }
+
+ return 0;
+}
+
+/*
+ * to continue, we must have traces from all online cpus in the tree
+ */
+static int check_cpu_map(struct per_dev_info *pdi)
+{
+ unsigned long *cpu_map;
+ struct rb_node *n;
+ struct trace *__t;
+ unsigned int i;
+ int ret, cpu;
+
+ /*
+ * create a map of the cpus we have traces for
+ */
+ cpu_map = malloc(pdi->cpu_map_max / sizeof(long));
+ memset(cpu_map, 0, sizeof(*cpu_map));
+ n = rb_first(&rb_sort_root);
+ while (n) {
+ __t = rb_entry(n, struct trace, rb_node);
+ cpu = __t->bit->cpu;
+
+ cpu_map[CPU_IDX(cpu)] |= (1UL << CPU_BIT(cpu));
+ n = rb_next(n);
+ }
+
+ /*
+ * we can't continue if pdi->cpu_map has entries set that we don't
+ * have in the sort rbtree. the opposite is not a problem, though
+ */
+ ret = 0;
+ for (i = 0; i < pdi->cpu_map_max / CPUS_PER_LONG; i++) {
+ if (pdi->cpu_map[i] & ~(cpu_map[i])) {
+ ret = 1;
+ break;
+ }
+ }
+
+ free(cpu_map);
+ return ret;
+}
+
+static int check_sequence(struct per_dev_info *pdi, struct trace *t, int force)
+{
+ struct blk_io_trace *bit = t->bit;
+ unsigned long expected_sequence;
struct per_cpu_info *pci;
+ struct trace *__t;
+
+ pci = get_cpu_info(pdi, bit->cpu);
+ expected_sequence = pci->last_sequence + 1;
+
+ if (!expected_sequence) {
+ /*
+ * 1 should be the first entry, just allow it
+ */
+ if (bit->sequence == 1)
+ return 0;
+ if (bit->sequence == pci->smallest_seq_read)
+ return 0;
+
+ return check_cpu_map(pdi);
+ }
+
+ if (bit->sequence == expected_sequence)
+ return 0;
+
+ /*
+ * we may not have seen that sequence yet. if we are not doing
+ * the final run, break and wait for more entries.
+ */
+ if (expected_sequence < pci->smallest_seq_read) {
+ __t = trace_rb_find_last(pdi, pci, expected_sequence);
+ if (!__t)
+ goto skip;
+
+ __put_trace_last(pdi, __t);
+ return 0;
+ } else if (!force) {
+ return 1;
+ } else {
+skip:
+ if (check_current_skips(pci, bit->sequence))
+ return 0;
+
+ if (expected_sequence < bit->sequence)
+ insert_skip(pci, expected_sequence, bit->sequence - 1);
+ return 0;
+ }
+}
+
+static void show_entries_rb(int force)
+{
+ struct per_dev_info *pdi = NULL;
+ struct per_cpu_info *pci = NULL;
struct blk_io_trace *bit;
+ struct rb_node *n;
struct trace *t;
- int nr = 0;
- while ((t = trace_list) != NULL) {
+ while ((n = rb_first(&rb_sort_root)) != NULL) {
+ if (is_done() && !force && !pipeline)
+ break;
+
+ t = rb_entry(n, struct trace, rb_node);
+ bit = t->bit;
+
+ if (read_sequence - t->read_sequence < 1 && !force)
+ break;
+
+ if (!pdi || pdi->dev != bit->device) {
+ pdi = get_dev_info(bit->device);
+ pci = NULL;
+ }
+
+ if (!pdi) {
+ fprintf(stderr, "Unknown device ID? (%d,%d)\n",
+ MAJOR(bit->device), MINOR(bit->device));
+ break;
+ }
+
+ if (!(bit->action == BLK_TN_MESSAGE) &&
+ check_sequence(pdi, t, force))
+ break;
+
+ if (!force && bit->time > last_allowed_time)
+ break;
+
+ check_time(pdi, bit);
+
+ if (!pci || pci->cpu != bit->cpu)
+ pci = get_cpu_info(pdi, bit->cpu);
+
+ if (!(bit->action == BLK_TN_MESSAGE))
+ pci->last_sequence = bit->sequence;
+
+ pci->nelems++;
+
+ if (bit->action & (act_mask << BLK_TC_SHIFT))
+ dump_trace(bit, pci, pdi);
+
+ put_trace(pdi, t);
+ }
+}
+
+static int read_data(int fd, void *buffer, int bytes, int block, int *fdblock)
+{
+ int ret, bytes_left, fl;
+ void *p;
+
+ if (block != *fdblock) {
+ fl = fcntl(fd, F_GETFL);
+
+ if (!block) {
+ *fdblock = 0;
+ fcntl(fd, F_SETFL, fl | O_NONBLOCK);
+ } else {
+ *fdblock = 1;
+ fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
+ }
+ }
+
+ bytes_left = bytes;
+ p = buffer;
+ while (bytes_left > 0) {
+ ret = read(fd, p, bytes_left);
+ if (!ret)
+ return 1;
+ else if (ret < 0) {
+ if (errno != EAGAIN) {
+ perror("read");
+ return -1;
+ }
+
+ /*
+ * never do partial reads. we can return if we
+ * didn't read anything and we should not block,
+ * otherwise wait for data
+ */
+ if ((bytes_left == bytes) && !block)
+ return 1;
+
+ usleep(10);
+ continue;
+ } else {
+ p += ret;
+ bytes_left -= ret;
+ }
+ }
+
+ return 0;
+}
+
+static inline __u16 get_pdulen(struct blk_io_trace *bit)
+{
+ if (data_is_native)
+ return bit->pdu_len;
+
+ return __bswap_16(bit->pdu_len);
+}
+
+static inline __u32 get_magic(struct blk_io_trace *bit)
+{
+ if (data_is_native)
+ return bit->magic;
+
+ return __bswap_32(bit->magic);
+}
+
+static int read_events(int fd, int always_block, int *fdblock)
+{
+ struct per_dev_info *pdi = NULL;
+ unsigned int events = 0;
+
+ while (!is_done() && events < rb_batch) {
+ struct blk_io_trace *bit;
+ struct trace *t;
+ int pdu_len, should_block, ret;
+ __u32 magic;
+
+ bit = bit_alloc();
+
+ should_block = !events || always_block;
+
+ ret = read_data(fd, bit, sizeof(*bit), should_block, fdblock);
+ if (ret) {
+ bit_free(bit);
+ if (!events && ret < 0)
+ events = ret;
+ break;
+ }
+
+ /*
+ * look at first trace to check whether we need to convert
+ * data in the future
+ */
+ if (data_is_native == -1 && check_data_endianness(bit->magic))
+ break;
+
+ magic = get_magic(bit);
+ if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
+ fprintf(stderr, "Bad magic %x\n", magic);
+ break;
+ }
+
+ pdu_len = get_pdulen(bit);
+ if (pdu_len) {
+ void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
+
+ if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1, fdblock)) {
+ bit_free(ptr);
+ break;
+ }
+
+ bit = ptr;
+ }
+
+ trace_to_cpu(bit);
+
+ if (verify_trace(bit)) {
+ bit_free(bit);
+ continue;
+ }
+
+ /*
+ * not a real trace, so grab and handle it here
+ */
+ if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY) && bit->action != BLK_TN_MESSAGE) {
+ handle_notify(bit);
+ output_binary(bit, sizeof(*bit) + bit->pdu_len);
+ continue;
+ }
+
+ t = t_alloc();
+ memset(t, 0, sizeof(*t));
+ t->bit = bit;
+ t->read_sequence = read_sequence;
+
+ t->next = trace_list;
+ trace_list = t;
+
+ if (!pdi || pdi->dev != bit->device)
+ pdi = get_dev_info(bit->device);
+
+ if (bit->time > pdi->last_read_time)
+ pdi->last_read_time = bit->time;
+
+ events++;
+ }
+
+ return events;
+}
+
+/*
+ * Managing input streams
+ */
+
+struct ms_stream {
+ struct ms_stream *next;
+ struct trace *first, *last;
+ struct per_dev_info *pdi;
+ unsigned int cpu;
+};
+
+#define MS_HASH(d, c) ((MAJOR(d) & 0xff) ^ (MINOR(d) & 0xff) ^ (cpu & 0xff))
+
+struct ms_stream *ms_head;
+struct ms_stream *ms_hash[256];
+
+static void ms_sort(struct ms_stream *msp);
+static int ms_prime(struct ms_stream *msp);
+
+static inline struct trace *ms_peek(struct ms_stream *msp)
+{
+ return (msp == NULL) ? NULL : msp->first;
+}
+
+static inline __u64 ms_peek_time(struct ms_stream *msp)
+{
+ return ms_peek(msp)->bit->time;
+}
- trace_list = t->next;
- bit = t->bit;
+static inline void ms_resort(struct ms_stream *msp)
+{
+ if (msp->next && ms_peek_time(msp) > ms_peek_time(msp->next)) {
+ ms_head = msp->next;
+ msp->next = NULL;
+ ms_sort(msp);
+ }
+}
- memset(&t->rb_node, 0, sizeof(t->rb_node));
+static inline void ms_deq(struct ms_stream *msp)
+{
+ msp->first = msp->first->next;
+ if (!msp->first) {
+ msp->last = NULL;
+ if (!ms_prime(msp)) {
+ ms_head = msp->next;
+ msp->next = NULL;
+ return;
+ }
+ }
- if (verify_trace(bit))
- break;
- if (trace_rb_insert(t))
- return -1;
+ ms_resort(msp);
+}
- pdi = get_dev_info(bit->device);
- pci = get_cpu_info(pdi, bit->cpu);
- pci->nelems++;
+static void ms_sort(struct ms_stream *msp)
+{
+ __u64 msp_t = ms_peek_time(msp);
+ struct ms_stream *this_msp = ms_head;
- nr++;
+ if (this_msp == NULL)
+ ms_head = msp;
+ else if (msp_t < ms_peek_time(this_msp)) {
+ msp->next = this_msp;
+ ms_head = msp;
}
+ else {
+ while (this_msp->next && ms_peek_time(this_msp->next) < msp_t)
+ this_msp = this_msp->next;
- return nr;
+ msp->next = this_msp->next;
+ this_msp->next = msp;
+ }
}
-static void show_entries_rb(int piped)
+static int ms_prime(struct ms_stream *msp)
{
- struct per_dev_info *pdi;
- struct blk_io_trace *bit;
- struct rb_node *n;
+ __u32 magic;
+ unsigned int i;
struct trace *t;
- int cpu;
+ struct per_dev_info *pdi = msp->pdi;
+ struct per_cpu_info *pci = get_cpu_info(pdi, msp->cpu);
+ struct blk_io_trace *bit = NULL;
+ int ret, pdu_len, ndone = 0;
- while ((n = rb_first(&rb_sort_root)) != NULL) {
+ for (i = 0; !is_done() && pci->fd >= 0 && i < rb_batch; i++) {
+ bit = bit_alloc();
+ ret = read_data(pci->fd, bit, sizeof(*bit), 1, &pci->fdblock);
+ if (ret)
+ goto err;
- t = rb_entry(n, struct trace, rb_node);
- bit = t->bit;
+ if (data_is_native == -1 && check_data_endianness(bit->magic))
+ goto err;
+
+ magic = get_magic(bit);
+ if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
+ fprintf(stderr, "Bad magic %x\n", magic);
+ goto err;
- pdi = get_dev_info(bit->device);
- if (!pdi) {
- fprintf(stderr, "Unknown device ID? (%d,%d)\n",
- MAJOR(bit->device), MINOR(bit->device));
- break;
- }
- cpu = bit->cpu;
- if (cpu > pdi->ncpus) {
- fprintf(stderr, "Unknown CPU ID? (%d, device %d,%d)\n",
- cpu, MAJOR(bit->device), MINOR(bit->device));
- break;
}
- /*
- * 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 us a few times
- */
- if (bit->sequence != (pdi->last_sequence + 1)) {
- if (piped && t->skipped < 5) {
- t->skipped++;
- break;
- } else {
- fprintf(stderr, "skipping from %lu to %u\n", pdi->last_sequence, bit->sequence);
- pdi->skips++;
+ pdu_len = get_pdulen(bit);
+ if (pdu_len) {
+ void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
+ ret = read_data(pci->fd, ptr + sizeof(*bit), pdu_len,
+ 1, &pci->fdblock);
+ if (ret) {
+ free(ptr);
+ bit = NULL;
+ goto err;
}
+
+ bit = ptr;
}
- pdi->last_sequence = bit->sequence;
+ trace_to_cpu(bit);
+ if (verify_trace(bit))
+ goto err;
- bit->time -= genesis_time;
- if (bit->time < stopwatch_start)
+ if (bit->cpu != pci->cpu) {
+ fprintf(stderr, "cpu %d trace info has error cpu %d\n",
+ pci->cpu, bit->cpu);
continue;
- if (bit->time >= stopwatch_end)
- break;
+ }
- check_time(pdi, bit);
+ if (bit->action & BLK_TC_ACT(BLK_TC_NOTIFY) && bit->action != BLK_TN_MESSAGE) {
+ handle_notify(bit);
+ output_binary(bit, sizeof(*bit) + bit->pdu_len);
+ bit_free(bit);
- if (dump_trace(bit, &pdi->cpus[cpu], pdi))
- break;
+ i -= 1;
+ continue;
+ }
- rb_erase(&t->rb_node, &rb_sort_root);
+ if (bit->time > pdi->last_read_time)
+ pdi->last_read_time = bit->time;
+
+ t = t_alloc();
+ memset(t, 0, sizeof(*t));
+ t->bit = bit;
- if (piped) {
- free(bit);
- free(t);
+ if (msp->first == NULL)
+ msp->first = msp->last = t;
+ else {
+ msp->last->next = t;
+ msp->last = t;
}
+
+ ndone++;
}
+
+ return ndone;
+
+err:
+ if (bit) bit_free(bit);
+
+ cpu_mark_offline(pdi, pci->cpu);
+ close(pci->fd);
+ pci->fd = -1;
+
+ return ndone;
}
-static int read_data(int fd, void *buffer, int bytes, int block)
+static struct ms_stream *ms_alloc(struct per_dev_info *pdi, int cpu)
{
- int ret, bytes_left, fl;
- void *p;
-
- fl = fcntl(fd, F_GETFL);
+ struct ms_stream *msp = malloc(sizeof(*msp));
- if (!block)
- fcntl(fd, F_SETFL, fl | O_NONBLOCK);
- else
- fcntl(fd, F_SETFL, fl & ~O_NONBLOCK);
+ msp->next = NULL;
+ msp->first = msp->last = NULL;
+ msp->pdi = pdi;
+ msp->cpu = cpu;
- bytes_left = bytes;
- p = buffer;
- while (bytes_left > 0) {
- ret = read(fd, p, bytes_left);
- if (!ret)
- return 1;
- else if (ret < 0) {
- if (errno != EAGAIN)
- perror("read");
- return -1;
- } else {
- p += ret;
- bytes_left -= ret;
- }
- }
+ if (ms_prime(msp))
+ ms_sort(msp);
- return 0;
+ return msp;
}
-/*
- * Find the traces in 'tb' and add them to the list for sorting and
- * displaying
- */
-static int find_entries(void *tb, unsigned long size)
+static int setup_file(struct per_dev_info *pdi, int cpu)
{
- struct blk_io_trace *bit;
- struct trace *t;
- void *start = tb;
+ int len = 0;
+ struct stat st;
+ char *p, *dname;
+ struct per_cpu_info *pci = get_cpu_info(pdi, cpu);
- while (tb - start <= size - sizeof(*bit)) {
- bit = find_trace(tb, size - (tb - start));
- if (!bit)
- break;
+ pci->cpu = cpu;
+ pci->fdblock = -1;
+
+ p = strdup(pdi->name);
+ dname = dirname(p);
+ if (strcmp(dname, ".")) {
+ input_dir = dname;
+ p = strdup(pdi->name);
+ strcpy(pdi->name, basename(p));
+ }
+ free(p);
- t = malloc(sizeof(*t));
- memset(t, 0, sizeof(*t));
- t->bit = bit;
+ if (input_dir)
+ len = sprintf(pci->fname, "%s/", input_dir);
- trace_to_cpu(bit);
- t->next = trace_list;
- trace_list = t;
+ snprintf(pci->fname + len, sizeof(pci->fname)-1-len,
+ "%s.blktrace.%d", pdi->name, pci->cpu);
+ if (stat(pci->fname, &st) < 0)
+ return 0;
+ if (!st.st_size)
+ return 1;
- tb += sizeof(*bit) + bit->pdu_len;
+ pci->fd = open(pci->fname, O_RDONLY);
+ if (pci->fd < 0) {
+ perror(pci->fname);
+ return 0;
}
- return 0;
-}
-
-static int do_file(void)
-{
- int i, j, nfiles = 0, nelems;
+ printf("Input file %s added\n", pci->fname);
+ cpu_mark_online(pdi, pci->cpu);
- for (i = 0; i < ndevices; i++) {
- for (j = 0;; j++, nfiles++) {
- struct per_dev_info *pdi;
- struct per_cpu_info *pci;
- struct stat st;
- void *tb;
-
- pdi = &devices[i];
- pci = get_cpu_info(pdi, j);
- pci->cpu = j;
-
- snprintf(pci->fname, sizeof(pci->fname)-1,
- "%s.%d", pdi->name, j);
- if (stat(pci->fname, &st) < 0)
- break;
- if (!st.st_size)
- continue;
+ pdi->nfiles++;
+ ms_alloc(pdi, pci->cpu);
- printf("Processing %s\n", pci->fname);
+ return 1;
+}
- tb = malloc(st.st_size);
- if (!tb) {
- fprintf(stderr, "Out of memory, skip file %s\n",
- pci->fname);
- continue;
- }
+static int handle(struct ms_stream *msp)
+{
+ struct trace *t;
+ struct per_dev_info *pdi;
+ struct per_cpu_info *pci;
+ struct blk_io_trace *bit;
- pci->fd = open(pci->fname, O_RDONLY);
- if (pci->fd < 0) {
- perror(pci->fname);
- free(tb);
- continue;
- }
+ t = ms_peek(msp);
- if (read_data(pci->fd, tb, st.st_size, 1)) {
- close(pci->fd);
- free(tb);
- continue;
- }
+ bit = t->bit;
+ pdi = msp->pdi;
+ pci = get_cpu_info(pdi, msp->cpu);
+ pci->nelems++;
+ bit->time -= genesis_time;
- if (find_entries(tb, st.st_size)) {
- close(pci->fd);
- free(tb);
- }
+ if (t->bit->time > stopwatch_end)
+ return 0;
- nelems = sort_entries();
- if (nelems == -1) {
- close(pci->fd);
- free(tb);
- continue;
- }
+ pdi->last_reported_time = bit->time;
+ if ((bit->action & (act_mask << BLK_TC_SHIFT))&&
+ t->bit->time >= stopwatch_start)
+ dump_trace(bit, pci, pdi);
- printf("Completed %s (CPU%d %d, entries)\n",
- pci->fname, j, nelems);
- close(pci->fd);
- }
- }
+ ms_deq(msp);
- if (!nfiles) {
- fprintf(stderr, "No files found\n");
- return 1;
+ if (text_output)
+ trace_rb_insert_last(pdi, t);
+ else {
+ bit_free(t->bit);
+ t_free(t);
}
- show_entries_rb(0);
- return 0;
+ return 1;
}
-static int read_sort_events(int fd)
+/*
+ * Check if we need to sanitize the name. We allow 'foo', or if foo.blktrace.X
+ * is given, then strip back down to 'foo' to avoid missing files.
+ */
+static int name_fixup(char *name)
{
- int events = 0;
+ char *b;
- do {
- struct blk_io_trace *bit;
- struct trace *t;
- int pdu_len;
- __u32 magic;
+ if (!name)
+ return 1;
- bit = malloc(sizeof(*bit));
+ b = strstr(name, ".blktrace.");
+ if (b)
+ *b = '\0';
- if (read_data(fd, bit, sizeof(*bit), !events))
- break;
+ return 0;
+}
- magic = be32_to_cpu(bit->magic);
- if ((magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
- fprintf(stderr, "Bad magic %x\n", magic);
- break;
- }
+static int do_file(void)
+{
+ int i, cpu, ret;
+ struct per_dev_info *pdi;
- pdu_len = be16_to_cpu(bit->pdu_len);
- if (pdu_len) {
- void *ptr = realloc(bit, sizeof(*bit) + pdu_len);
+ /*
+ * first prepare all files for reading
+ */
+ for (i = 0; i < ndevices; i++) {
+ pdi = &devices[i];
+ ret = name_fixup(pdi->name);
+ if (ret)
+ return ret;
- if (read_data(fd, ptr + sizeof(*bit), pdu_len, 1))
- break;
+ for (cpu = 0; setup_file(pdi, cpu); cpu++)
+ ;
- bit = ptr;
+ if (!cpu) {
+ fprintf(stderr,"No input files found for %s\n",
+ pdi->name);
+ return 1;
}
+ }
- t = malloc(sizeof(*t));
- memset(t, 0, sizeof(*t));
- t->bit = bit;
+ /*
+ * Get the initial time stamp
+ */
+ if (ms_head)
+ genesis_time = ms_peek_time(ms_head);
- trace_to_cpu(bit);
- t->next = trace_list;
- trace_list = t;
+ /*
+ * Correct abs_start_time if necessary
+ */
+ if (start_timestamp
+ && start_timestamp != genesis_time) {
+ correct_abs_start_time();
+ }
- events++;
- } while (!is_done() && events < rb_batch);
+ /*
+ * Keep processing traces while any are left
+ */
+ while (!is_done() && ms_head && handle(ms_head))
+ ;
- return events;
+ return 0;
}
-static int do_stdin(void)
+static void do_pipe(int fd)
{
- int fd;
+ unsigned long long youngest;
+ int events, fdblock;
- fd = dup(STDIN_FILENO);
- do {
- int events;
+ last_allowed_time = -1ULL;
+ fdblock = -1;
+ while ((events = read_events(fd, 0, &fdblock)) > 0) {
+ read_sequence++;
+
+#if 0
+ smallest_seq_read = -1U;
+#endif
- events = read_sort_events(fd);
- if (!events)
+ if (sort_entries(&youngest))
break;
-
- if (sort_entries() == -1)
+
+ if (youngest > stopwatch_end)
break;
+ show_entries_rb(0);
+ }
+
+ if (rb_sort_entries)
show_entries_rb(1);
- } while (1);
+}
+
+static int do_fifo(void)
+{
+ int fd;
+
+ if (!strcmp(pipename, "-"))
+ fd = dup(STDIN_FILENO);
+ else
+ fd = open(pipename, O_RDONLY);
+
+ if (fd == -1) {
+ perror("dup stdin");
+ return -1;
+ }
+ do_pipe(fd);
close(fd);
return 0;
}
-static void flush_output(void)
+static void show_stats(void)
{
+ if (!ofp)
+ return;
+ if (stats_printed)
+ return;
+
+ stats_printed = 1;
+
+ if (per_process_stats)
+ show_process_stats();
+
+ if (per_device_and_cpu_stats)
+ show_device_and_cpu_stats();
+
fflush(ofp);
}
-static void handle_sigint(int sig)
+static void handle_sigint(__attribute__((__unused__)) int sig)
{
done = 1;
- flush_output();
}
/*
fprintf(stderr,"Invalid stopwatch start timer: %s\n", string);
return 1;
}
- stopwatch_end = stopwatch_start + DOUBLE_TO_NANO_ULL(value);
+ 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;
+}
+
+static int is_pipe(const char *str)
+{
+ struct stat st;
+
+ if (!strcmp(str, "-"))
+ return 1;
+ if (!stat(str, &st) && S_ISFIFO(st.st_mode))
+ return 1;
+
+ return 0;
+}
+
+static int get_program_sort_event(const char *str)
+{
+ char evt = str[0];
+
+ switch (evt) {
+ case 'N':
+ per_process_stats_event = SORT_PROG_EVENT_N;
+ break;
+ case 'Q':
+ per_process_stats_event = SORT_PROG_EVENT_QKB;
+ break;
+ case 'q':
+ per_process_stats_event = SORT_PROG_EVENT_QIO;
+ break;
+ case 'R':
+ per_process_stats_event = SORT_PROG_EVENT_RKB;
+ break;
+ case 'r':
+ per_process_stats_event = SORT_PROG_EVENT_RIO;
+ break;
+ case 'W':
+ per_process_stats_event = SORT_PROG_EVENT_WKB;
+ break;
+ case 'w':
+ per_process_stats_event = SORT_PROG_EVENT_WIO;
+ break;
+ case 'C':
+ per_process_stats_event = SORT_PROG_EVENT_CKB;
+ break;
+ case 'c':
+ per_process_stats_event = SORT_PROG_EVENT_CIO;
+ break;
+ default:
+ return 1;
+ }
+
return 0;
}
+#define S_OPTS "a:A:b:D:d:f:F:hi:o:OqsS:tw:vVM"
+static char usage_str[] = "\n\n" \
+ "-i <file> | --input=<file>\n" \
+ "[ -a <action field> | --act-mask=<action field> ]\n" \
+ "[ -A <action mask> | --set-mask=<action mask> ]\n" \
+ "[ -b <traces> | --batch=<traces> ]\n" \
+ "[ -d <file> | --dump-binary=<file> ]\n" \
+ "[ -D <dir> | --input-directory=<dir> ]\n" \
+ "[ -f <format> | --format=<format> ]\n" \
+ "[ -F <spec> | --format-spec=<spec> ]\n" \
+ "[ -h | --hash-by-name ]\n" \
+ "[ -o <file> | --output=<file> ]\n" \
+ "[ -O | --no-text-output ]\n" \
+ "[ -q | --quiet ]\n" \
+ "[ -s | --per-program-stats ]\n" \
+ "[ -S <event> | --sort-program-stats=<event> ]\n" \
+ "[ -t | --track-ios ]\n" \
+ "[ -w <time> | --stopwatch=<time> ]\n" \
+ "[ -M | --no-msgs\n" \
+ "[ -v | --verbose ]\n" \
+ "[ -V | --version ]\n\n" \
+ "\t-a Only trace specified actions. See documentation\n" \
+ "\t-A Give trace mask as a single value. See documentation\n" \
+ "\t-b stdin read batching\n" \
+ "\t-d Output file. If specified, binary data is written to file\n" \
+ "\t-D Directory to prepend to input file names\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-h Hash processes by name, not pid\n" \
+ "\t-i Input file containing trace data, or '-' for stdin\n" \
+ "\t-o Output file. If not given, output is stdout\n" \
+ "\t-O Do NOT output text data\n" \
+ "\t-q Quiet. Don't display any stats at the end of the trace\n" \
+ "\t-s Show per-program io statistics\n" \
+ "\t-S Show per-program io statistics sorted by N/Q/q/R/r/W/w/C/c\n" \
+ "\t N:Name, Q/q:Queued(read & write), R/r:Queued Read, W/w:Queued Write, C/c:Complete.\n" \
+ "\t Sort programs by how much data(KB): Q,R,W,C.\n" \
+ "\t Sort programs by how many IO operations: q,r,w,c.\n" \
+ "\t if -S was used, the -s parameter will be ignored.\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-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-M Do not output messages to binary file\n" \
+ "\t-v More verbose for marginal errors\n" \
+ "\t-V Print program version info\n\n";
+
static void usage(char *prog)
{
- fprintf(stderr, "Usage: %s "
- "[-i <name>] [-o <output>] [-s] [-w N[:n]] <name>...\n",
- prog);
+ fprintf(stderr, "Usage: %s %s", prog, usage_str);
}
int main(int argc, char *argv[])
{
- char *ofp_buffer;
- int c, ret, mode;
- int per_device_and_cpu_stats = 1;
+ int i, c, ret, mode;
+ int act_mask_tmp = 0;
+ char *ofp_buffer = NULL;
+ char *bin_ofp_buffer = NULL;
while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {
switch (c) {
+ case 'a':
+ i = find_mask_map(optarg);
+ if (i < 0) {
+ fprintf(stderr,"Invalid action mask %s\n",
+ optarg);
+ return 1;
+ }
+ act_mask_tmp |= i;
+ break;
+
+ case 'A':
+ if ((sscanf(optarg, "%x", &i) != 1) ||
+ !valid_act_opt(i)) {
+ fprintf(stderr,
+ "Invalid set action mask %s/0x%x\n",
+ optarg, i);
+ return 1;
+ }
+ act_mask_tmp = i;
+ break;
case 'i':
- if (!strcmp(optarg, "-") && !pipeline)
+ if (is_pipe(optarg) && !pipeline) {
pipeline = 1;
- else if (resize_devices(optarg) != 0)
+ pipename = strdup(optarg);
+ } else if (resize_devices(optarg) != 0)
return 1;
break;
+ case 'D':
+ input_dir = optarg;
+ break;
case 'o':
output_name = optarg;
break;
+ case 'O':
+ text_output = 0;
+ break;
case 'b':
rb_batch = atoi(optarg);
if (rb_batch <= 0)
case 's':
per_process_stats = 1;
break;
+ case 'S':
+ per_process_stats = 1;
+ if (get_program_sort_event(optarg))
+ return 1;
+ break;
case 't':
track_ios = 1;
break;
if (add_format_spec(optarg) != 0)
return 1;
break;
+ case 'h':
+ ppi_hash_by_pid = 0;
+ break;
+ case 'v':
+ verbose++;
+ break;
+ case 'V':
+ printf("%s version %s\n", argv[0], blkparse_version);
+ return 0;
+ case 'd':
+ dump_binary = optarg;
+ break;
+ case 'M':
+ bin_output_msgs = 0;
+ break;
default:
usage(argv[0]);
return 1;
}
while (optind < argc) {
- if (!strcmp(argv[optind], "-") && !pipeline)
+ if (is_pipe(argv[optind]) && !pipeline) {
pipeline = 1;
- else if (resize_devices(argv[optind]) != 0)
+ pipename = strdup(argv[optind]);
+ } else if (resize_devices(argv[optind]) != 0)
return 1;
optind++;
}
return 1;
}
+ if (act_mask_tmp != 0)
+ act_mask = act_mask_tmp;
+
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 (!output_name) {
- ofp = fdopen(STDOUT_FILENO, "w");
- mode = _IOLBF;
- } else {
- char ofname[128];
+ if (text_output) {
+ if (!output_name) {
+ ofp = fdopen(STDOUT_FILENO, "w");
+ mode = _IOLBF;
+ } else {
+ char ofname[PATH_MAX];
- snprintf(ofname, sizeof(ofname) - 1, "%s.log", output_name);
- ofp = fopen(ofname, "w");
- mode = _IOFBF;
- }
+ snprintf(ofname, sizeof(ofname) - 1, "%s", output_name);
+ ofp = fopen(ofname, "w");
+ mode = _IOFBF;
+ }
- if (!ofp) {
- perror("fopen");
- return 1;
+ if (!ofp) {
+ perror("fopen");
+ return 1;
+ }
+
+ ofp_buffer = malloc(4096);
+ if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
+ perror("setvbuf");
+ return 1;
+ }
}
- ofp_buffer = malloc(4096);
- if (setvbuf(ofp, ofp_buffer, mode, 4096)) {
- perror("setvbuf");
- return 1;
+ if (dump_binary) {
+ if (!strcmp(dump_binary, "-"))
+ dump_fp = stdout;
+ else {
+ dump_fp = fopen(dump_binary, "w");
+ if (!dump_fp) {
+ perror(dump_binary);
+ dump_binary = NULL;
+ return 1;
+ }
+ }
+ bin_ofp_buffer = malloc(128 * 1024);
+ if (setvbuf(dump_fp, bin_ofp_buffer, _IOFBF, 128 * 1024)) {
+ perror("setvbuf binary");
+ return 1;
+ }
}
if (pipeline)
- ret = do_stdin();
+ ret = do_fifo();
else
ret = do_file();
- if (per_process_stats)
- show_process_stats();
+ if (!ret)
+ show_stats();
- if (per_device_and_cpu_stats)
- show_device_and_cpu_stats();
+ if (have_drv_data && !dump_binary)
+ printf("\ndiscarded traces containing low-level device driver "
+ "specific data (only available in binary output)\n");
- flush_output();
+ if (ofp_buffer) {
+ fflush(ofp);
+ free(ofp_buffer);
+ }
+ if (bin_ofp_buffer) {
+ fflush(dump_fp);
+ free(bin_ofp_buffer);
+ }
return ret;
}
projects / blktrace.git / blobdiff