#include <stdio.h>
-#include <stdio.h>
#include <unistd.h>
#include <inttypes.h>
+#include <math.h>
#include <assert.h>
#include "../io_ddir.h"
#include "../blktrace_api.h"
#include "../os/os.h"
#include "../log.h"
-#include "../lib/linux-dev-lookup.h"
+#include "../minmax.h"
+#include "../oslib/linux-dev-lookup.h"
#define TRACE_FIFO_SIZE 8192
static int output_ascii = 1;
static char *filename;
+static char **add_opts;
+static int n_add_opts;
+
+/*
+ * Collapse defaults
+ */
+static unsigned int collapse_entries = 0;
+static unsigned int depth_diff = 1;
+static unsigned int random_diff = 5;
+
struct bs {
unsigned int bs;
unsigned int nr;
uint64_t first_ttime[DDIR_RWDIR_CNT];
uint64_t last_ttime[DDIR_RWDIR_CNT];
- uint64_t kb[DDIR_RWDIR_CNT];
+ uint64_t kib[DDIR_RWDIR_CNT];
uint64_t start_delay;
};
struct flist_head pid_list;
pid_t pid;
+ pid_t *merge_pids;
+ unsigned int nr_merge_pids;
+
struct trace_file *files;
int nr_files;
unsigned int last_major, last_minor;
+ int numjobs;
+ int ignore;
struct btrace_out o;
};
if (act == __BLK_TA_QUEUE) {
inflight_add(p, t->sector, t->bytes);
ret = handle_queue_trace(t, p);
- } else if (act == __BLK_TA_REQUEUE) {
- p->o.inflight--;
} else if (act == __BLK_TA_BACKMERGE) {
struct inflight *i;
i = inflight_find(t->sector + (t->bytes >> 9));
if (i) {
- i->p->o.kb[t_to_rwdir(t)] += (t->bytes >> 10);
+ i->p->o.kib[t_to_rwdir(t)] += (t->bytes >> 10);
i->p->o.complete_seen = 1;
inflight_remove(i);
}
}
p->pid = pid;
+ p->numjobs = 1;
flist_add_tail(&p->hash_list, hash_list);
flist_add_tail(&p->pid_list, &pid_list);
}
return bsb->nr - bsa->nr;
}
-static unsigned long o_to_kb_rate(struct btrace_out *o, int rw)
+static unsigned long o_to_kib_rate(struct btrace_out *o, int rw)
{
uint64_t usec = (o->last_ttime[rw] - o->first_ttime[rw]) / 1000ULL;
uint64_t val;
if (!usec)
return 0;
- val = o->kb[rw] * 1000ULL;
+ val = o->kib[rw] * 1000ULL;
return val / usec;
}
unsigned long total, usec;
int i, j;
- printf("[pid:\t%u]\n", p->pid);
+ printf("[pid:\t%u", p->pid);
+ if (p->nr_merge_pids)
+ for (i = 0; i < p->nr_merge_pids; i++)
+ printf(", %u", p->merge_pids[i]);
+ printf("]\n");
total = ddir_rw_sum(o->ios);
for (i = 0; i < DDIR_RWDIR_CNT; i++) {
printf("\tmerges: %lu (perc=%3.2f%%)\n", o->merges[i], perc);
perc = ((float) o->seq[i] * 100.0) / (float) o->ios[i];
printf("\tseq: %lu (perc=%3.2f%%)\n", (unsigned long) o->seq[i], perc);
- printf("\trate: %lu KB/sec\n", o_to_kb_rate(o, i));
+ printf("\trate: %lu KiB/sec\n", o_to_kib_rate(o, i));
for (j = 0; j < o->nr_bs[i]; j++) {
struct bs *bs = &o->bs[i][j];
printf("\n");
}
-static int __output_p_fio(struct btrace_pid *p, unsigned long *ios)
+static int __output_p_fio(struct btrace_pid *p, unsigned long *ios,
+ const char *name_postfix)
{
struct btrace_out *o = &p->o;
unsigned long total;
int i, j;
if ((o->ios[0] + o->ios[1]) && o->ios[2]) {
- log_err("fio: trace has both read/write and trim\n");
- return 1;
+ unsigned long ios_bak[DDIR_RWDIR_CNT];
+
+ memcpy(ios_bak, o->ios, DDIR_RWDIR_CNT * sizeof(unsigned long));
+
+ /* create job for read/write */
+ o->ios[2] = 0;
+ __output_p_fio(p, ios, "");
+ o->ios[2] = ios_bak[2];
+
+ /* create job for trim */
+ o->ios[0] = 0;
+ o->ios[1] = 0;
+ __output_p_fio(p, ios, "_trim");
+ o->ios[0] = ios_bak[0];
+ o->ios[1] = ios_bak[1];
+
+ return 0;
}
if (!p->nr_files) {
log_err("fio: no devices found\n");
return 1;
}
- printf("[pid%u]\n", p->pid);
+ printf("[pid%u%s", p->pid, name_postfix);
+ if (p->nr_merge_pids)
+ for (i = 0; i < p->nr_merge_pids; i++)
+ printf(",pid%u", p->merge_pids[i]);
+ printf("]\n");
+
+ printf("numjobs=%u\n", p->numjobs);
printf("direct=1\n");
if (o->depth == 1)
printf("ioengine=sync\n");
printf("rw=randrw\n");
total = ddir_rw_sum(o->ios);
perc = ((float) o->ios[0] * 100.0) / (float) total;
- printf("rwmixread=%u\n", (int) (perc + 0.99));
+ printf("rwmixread=%u\n", (int) floor(perc + 0.50));
}
printf("percentage_random=");
if (i)
printf(",");
perc = 100.0 - perc;
- printf("%u", (int) perc);
+ printf("%u", (int) floor(perc + 0.5));
}
printf("\n");
}
printf("\n");
- printf("startdelay=%llus\n", o->start_delay / 1000000ULL);
+ if (o->start_delay / 1000000ULL)
+ printf("startdelay=%llus\n", o->start_delay / 1000000ULL);
time = o_longest_ttime(o);
time = (time + 1000000000ULL - 1) / 1000000000ULL;
if (j + 1 == o->nr_bs[i])
printf("%u/", bs->bs);
else
- printf("%u/%u", bs->bs, (int) perc);
+ printf("%u/%u", bs->bs, (int) floor(perc + 0.5));
}
}
printf("\n");
for (i = 0; i < DDIR_RWDIR_CNT; i++) {
unsigned long rate;
- rate = o_to_kb_rate(o, i);
+ rate = o_to_kib_rate(o, i);
if (i)
printf(",");
if (rate)
printf("\n");
}
+ if (n_add_opts)
+ for (i = 0; i < n_add_opts; i++)
+ printf("%s\n", add_opts[i]);
+
printf("\n");
return 0;
}
if (output_ascii)
__output_p_ascii(p, ios);
else
- ret = __output_p_fio(p, ios);
+ ret = __output_p_fio(p, ios, "");
return ret;
}
for (i = 0; i < DDIR_RWDIR_CNT; i++) {
unsigned long this_rate;
- this_rate = o_to_kb_rate(o, i);
+ this_rate = o_to_kib_rate(o, i);
if (this_rate < rate_threshold) {
remove_ddir(o, i);
this_rate = 0;
free(p);
}
+static int entries_close(struct btrace_pid *pida, struct btrace_pid *pidb)
+{
+ float perca, percb, fdiff;
+ int i, idiff;
+
+ for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+ if ((pida->o.ios[i] && !pidb->o.ios[i]) ||
+ (pidb->o.ios[i] && !pida->o.ios[i]))
+ return 0;
+ if (pida->o.ios[i] && pidb->o.ios[i]) {
+ perca = ((float) pida->o.seq[i] * 100.0) / (float) pida->o.ios[i];
+ percb = ((float) pidb->o.seq[i] * 100.0) / (float) pidb->o.ios[i];
+ fdiff = perca - percb;
+ if (fabs(fdiff) > random_diff)
+ return 0;
+ }
+
+ idiff = pida->o.depth - pidb->o.depth;
+ if (abs(idiff) > depth_diff)
+ return 0;
+ }
+
+ return 1;
+}
+
+static void merge_bs(struct bs **bsap, unsigned int *nr_bsap,
+ struct bs *bsb, unsigned int nr_bsb)
+{
+ struct bs *bsa = *bsap;
+ unsigned int nr_bsa = *nr_bsap;
+ int a, b;
+
+ for (b = 0; b < nr_bsb; b++) {
+ int next, found = 0;
+
+ for (a = 0; a < nr_bsa; a++) {
+ if (bsb[b].bs != bsa[a].bs)
+ continue;
+
+ bsa[a].nr += bsb[b].nr;
+ bsa[a].merges += bsb[b].merges;
+ found = 1;
+ break;
+ }
+
+ if (found)
+ continue;
+
+ next = *nr_bsap;
+ bsa = realloc(bsa, (next + 1) * sizeof(struct bs));
+ bsa[next].bs = bsb[b].bs;
+ bsa[next].nr = bsb[b].nr;
+ (*nr_bsap)++;
+ *bsap = bsa;
+ }
+}
+
+static int merge_entries(struct btrace_pid *pida, struct btrace_pid *pidb)
+{
+ int i;
+
+ if (!entries_close(pida, pidb))
+ return 0;
+
+ pida->nr_merge_pids++;
+ pida->merge_pids = realloc(pida->merge_pids, pida->nr_merge_pids * sizeof(pid_t));
+ pida->merge_pids[pida->nr_merge_pids - 1] = pidb->pid;
+
+ for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+ struct btrace_out *oa = &pida->o;
+ struct btrace_out *ob = &pidb->o;
+
+ oa->ios[i] += ob->ios[i];
+ oa->merges[i] += ob->merges[i];
+ oa->seq[i] += ob->seq[i];
+ oa->kib[i] += ob->kib[i];
+ oa->first_ttime[i] = min(oa->first_ttime[i], ob->first_ttime[i]);
+ oa->last_ttime[i] = max(oa->last_ttime[i], ob->last_ttime[i]);
+ merge_bs(&oa->bs[i], &oa->nr_bs[i], ob->bs[i], ob->nr_bs[i]);
+ }
+
+ pida->o.start_delay = min(pida->o.start_delay, pidb->o.start_delay);
+ pida->o.depth = (pida->o.depth + pidb->o.depth) / 2;
+ return 1;
+}
+
+static void check_merges(struct btrace_pid *p, struct flist_head *pidlist)
+{
+ struct flist_head *e, *tmp;
+
+ if (p->ignore)
+ return;
+
+ flist_for_each_safe(e, tmp, pidlist) {
+ struct btrace_pid *pidb;
+
+ pidb = flist_entry(e, struct btrace_pid, pid_list);
+ if (pidb == p)
+ continue;
+
+ if (merge_entries(p, pidb)) {
+ pidb->ignore = 1;
+ p->numjobs++;
+ }
+ }
+}
+
static int output_p(void)
{
unsigned long ios[DDIR_RWDIR_CNT];
depth_disabled += p->o.depth_disabled;
}
+ if (collapse_entries) {
+ struct btrace_pid *p;
+
+ flist_for_each_safe(e, tmp, &pid_list) {
+ p = flist_entry(e, struct btrace_pid, pid_list);
+ check_merges(p, &pid_list);
+ }
+
+ flist_for_each_safe(e, tmp, &pid_list) {
+ p = flist_entry(e, struct btrace_pid, pid_list);
+ if (p->ignore)
+ free_p(p);
+ }
+ }
+
if (depth_disabled)
log_err("fio: missing completion traces, depths capped at %u\n", max_depth);
static int usage(char *argv[])
{
- log_err("%s: <blktrace bin file>\n", argv[0]);
+ log_err("%s: [options] <blktrace bin file>\n", argv[0]);
log_err("\t-t\tUsec threshold to ignore task\n");
log_err("\t-n\tNumber IOS threshold to ignore task\n");
log_err("\t-f\tFio job file output\n");
log_err("\t-d\tUse this file/device for replay\n");
- log_err("\t-r\tIgnore jobs with less than this KB/sec rate\n");
- log_err("\t-R\tSet rate in fio job\n");
+ log_err("\t-r\tIgnore jobs with less than this KiB/sec rate\n");
+ log_err("\t-R\tSet rate in fio job (def=%u)\n", set_rate);
log_err("\t-D\tCap queue depth at this value (def=%u)\n", max_depth);
+ log_err("\t-c\tCollapse \"identical\" jobs (def=%u)\n", collapse_entries);
+ log_err("\t-u\tDepth difference for collapse (def=%u)\n", depth_diff);
+ log_err("\t-x\tRandom difference for collapse (def=%u)\n", random_diff);
+ log_err("\t-a\tAdditional fio option to add to job file\n");
return 1;
}
int fd, ret;
*swap = -1;
-
+
fd = open(trace_file, O_RDONLY);
if (fd < 0) {
perror("open");
if (argc < 2)
return usage(argv);
- while ((c = getopt(argc, argv, "t:n:fd:r:RD:")) != -1) {
+ while ((c = getopt(argc, argv, "t:n:fd:r:RD:c:u:x:a:")) != -1) {
switch (c) {
case 'R':
set_rate = 1;
case 'D':
max_depth = atoi(optarg);
break;
+ case 'c':
+ collapse_entries = atoi(optarg);
+ break;
+ case 'u':
+ depth_diff = atoi(optarg);
+ break;
+ case 'x':
+ random_diff = atoi(optarg);
+ break;
+ case 'a':
+ add_opts = realloc(add_opts, (n_add_opts + 1) * sizeof(char *));
+ add_opts[n_add_opts] = strdup(optarg);
+ n_add_opts++;
+ break;
case '?':
default:
return usage(argv);