#include <libaio.h>
#include <math.h>
#include <limits.h>
+#include <assert.h>
#include <sys/time.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <semaphore.h>
#include <sys/ipc.h>
#include <sys/shm.h>
+#include <sys/ioctl.h>
#include <asm/unistd.h>
#include <asm/types.h>
+#include <asm/bitops.h>
+#include "arch.h"
#include "list.h"
#include "md5.h"
-#define MAX_JOBS (1024)
-
-/*
- * assume we don't have _get either, if _set isn't defined
- */
-#ifndef __NR_ioprio_set
-#if defined(__i386__)
-#define __NR_ioprio_set 289
-#define __NR_ioprio_get 290
-#elif defined(__powerpc__) || defined(__powerpc64__)
-#define __NR_ioprio_set 273
-#define __NR_ioprio_get 274
-#elif defined(__x86_64__)
-#define __NR_ioprio_set 251
-#define __NR_ioprio_get 252
-#elif defined(__ia64__)
-#define __NR_ioprio_set 1274
-#define __NR_ioprio_get 1275
-#elif defined(__alpha__)
-#define __NR_ioprio_set 442
-#define __NR_ioprio_get 443
-#elif defined(__s390x__) || defined(__s390__)
-#define __NR_ioprio_set 282
-#define __NR_ioprio_get 283
-#else
-#error "Unsupported arch"
-#endif
+#ifndef BLKGETSIZE64
+#define BLKGETSIZE64 _IOR(0x12,114,size_t)
#endif
-#ifndef __NR_fadvise64
-#if defined(__i386__)
-#define __NR_fadvise64 250
-#elif defined(__powerpc__) || defined(__powerpc64__)
-#define __NR_fadvise64 233
-#elif defined(__x86_64__)
-#define __NR_fadvise64 221
-#elif defined(__ia64__)
-#define __NR_fadvise64 1234
-#elif defined(__alpha__)
-#define __NR_fadvise64 413
-#elif defined(__s390x__) || defined(__s390__)
-#define __NR_fadvise64 253
-#else
-#error "Unsupported arch"
-#endif
-#endif
+#define MAX_JOBS (1024)
static int ioprio_set(int which, int who, int ioprio)
{
#define MASK (4095)
#define DEF_BS (4096)
-#define DEF_TIMEOUT (30)
+#define DEF_TIMEOUT (0)
#define DEF_RATE_CYCLE (1000)
#define DEF_ODIRECT (1)
#define DEF_SEQUENTIAL (1)
#define DEF_CREATE_FSYNC (1)
#define DEF_LOOPS (1)
#define DEF_VERIFY (0)
+#define DEF_STONEWALL (0)
+#define DEF_NUMJOBS (1)
#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
struct io_sample *log;
};
+struct io_piece {
+ struct list_head list;
+ unsigned long long offset;
+ unsigned int len;
+};
+
#define FIO_HDR_MAGIC 0xf00baaef
struct verify_header {
};
#define td_read(td) ((td)->ddir == DDIR_READ)
-#define should_fsync(td) (!td_read(td) && !(td)->odirect)
+#define td_write(td) ((td)->ddir == DDIR_WRITE)
+#define should_fsync(td) (td_write(td) && !(td)->odirect)
+
+#define BLOCKS_PER_MAP (8 * sizeof(long))
+#define TO_MAP_BLOCK(td, b) ((b) - ((td)->file_offset / (td)->min_bs))
+#define RAND_MAP_IDX(td, b) (TO_MAP_BLOCK(td, b) / BLOCKS_PER_MAP)
+#define RAND_MAP_BIT(td, b) (TO_MAP_BLOCK(td, b) & (BLOCKS_PER_MAP - 1))
struct thread_data {
char file_name[256];
+ char directory[256];
int thread_number;
int error;
int fd;
unsigned int sync_io;
unsigned int mem_type;
unsigned int verify;
+ unsigned int stonewall;
+ unsigned int numjobs;
cpu_set_t cpumask;
struct drand48_data bsrange_state;
unsigned int ratecycle;
unsigned long rate_usec_cycle;
long rate_pending_usleep;
- unsigned long rate_kb;
+ unsigned long rate_bytes;
struct timeval lastrate;
unsigned long runtime; /* sec */
- unsigned long kb;
+ unsigned long long io_size;
+
unsigned long io_blocks;
- unsigned long io_kb;
- unsigned long this_io_kb;
- unsigned long last_kb;
+ unsigned long io_bytes;
+ unsigned long this_io_bytes;
+ unsigned long last_bytes;
sem_t mutex;
+
struct drand48_data random_state;
+ unsigned long *file_map;
+ unsigned int num_maps;
/*
* bandwidth and latency stats
struct io_stat slat_stat; /* submission latency */
struct io_stat bw_stat; /* bandwidth stats */
- unsigned long stat_io_kb;
+ unsigned long stat_io_bytes;
struct timeval stat_sample_time;
struct io_log *lat_log;
struct io_log *bw_log;
struct timeval start;
+ struct rusage ru_start;
+ struct rusage ru_end;
+
+ struct list_head io_hist_list;
};
static struct thread_data *threads;
static int init_random_state(struct thread_data *td)
{
unsigned long seed;
- int fd;
+ int fd, num_maps, blocks;
fd = open("/dev/random", O_RDONLY);
if (fd == -1) {
if (repeatable)
seed = DEF_RANDSEED;
+ blocks = (td->io_size + td->min_bs - 1) / td->min_bs;
+ num_maps = blocks / BLOCKS_PER_MAP;
+ td->file_map = malloc(num_maps * sizeof(long));
+ td->num_maps = num_maps;
+ memset(td->file_map, 0, num_maps * sizeof(long));
+
srand48_r(seed, &td->random_state);
return 0;
}
return sec + usec;
}
+static unsigned long utime_since_now(struct timeval *s)
+{
+ struct timeval t;
+
+ gettimeofday(&t, NULL);
+ return utime_since(s, &t);
+}
+
static unsigned long mtime_since(struct timeval *s, struct timeval *e)
{
double sec, usec;
return s->tv_sec * 1000 + s->tv_usec / 1000;
}
-static unsigned long long get_next_offset(struct thread_data *td)
+static int random_map_free(struct thread_data *td, unsigned long long block)
{
- unsigned long long kb;
- double r;
+ unsigned int idx = RAND_MAP_IDX(td, block);
+ unsigned int bit = RAND_MAP_BIT(td, block);
+
+ return (td->file_map[idx] & (1UL << bit)) == 0;
+}
+
+static int get_next_free_block(struct thread_data *td, unsigned long long *b)
+{
+ int i;
+
+ *b = 0;
+ i = 0;
+ while ((*b) * td->min_bs < td->io_size) {
+ if (td->file_map[i] != -1UL) {
+ *b += ffz(td->file_map[i]);
+ return 0;
+ }
+
+ *b += BLOCKS_PER_MAP;
+ i++;
+ }
+
+ return 1;
+}
+
+static void mark_random_map(struct thread_data *td, struct io_u *io_u)
+{
+ unsigned long block = io_u->offset / td->min_bs;
+ unsigned int blocks = 0;
+
+ while (blocks < (io_u->buflen / td->min_bs)) {
+ int idx, bit;
+
+ if (!random_map_free(td, block))
+ break;
+
+ idx = RAND_MAP_IDX(td, block);
+ bit = RAND_MAP_BIT(td, block);
+
+ assert(idx < td->num_maps);
+
+ td->file_map[idx] |= (1UL << bit);
+ block++;
+ blocks++;
+ }
+
+ if ((blocks * td->min_bs) < io_u->buflen)
+ io_u->buflen = blocks * td->min_bs;
+}
+
+static int get_next_offset(struct thread_data *td, unsigned long long *offset)
+{
+ unsigned long long b, rb;
+ long r;
if (!td->sequential) {
- drand48_r(&td->random_state, &r);
- kb = (1+(double) (td->kb-1) * r / (RAND_MAX+1.0));
+ unsigned long max_blocks = td->io_size / td->min_bs;
+ int loops = 50;
+
+ do {
+ lrand48_r(&td->random_state, &r);
+ b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
+ rb = b + (td->file_offset / td->min_bs);
+ loops--;
+ } while (!random_map_free(td, rb) && loops);
+
+ if (!loops) {
+ if (get_next_free_block(td, &b))
+ return 1;
+ }
} else
- kb = td->last_kb;
+ b = td->last_bytes / td->min_bs;
+
+ *offset = (b * td->min_bs) + td->file_offset;
+ if (*offset > td->file_size)
+ return 1;
- return (kb << 10) + td->file_offset;
+ return 0;
}
static unsigned int get_next_buflen(struct thread_data *td)
{
unsigned int buflen;
- double r;
+ long r;
if (td->min_bs == td->max_bs)
buflen = td->min_bs;
else {
- drand48_r(&td->bsrange_state, &r);
+ lrand48_r(&td->bsrange_state, &r);
buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
}
- if (buflen > ((td->kb - td->this_io_kb) << 10))
- buflen = (td->kb - td->this_io_kb) << 10;
+ if (buflen > td->io_size - td->this_io_bytes)
+ buflen = td->io_size - td->this_io_bytes;
- td->last_kb += buflen >> 10;
return buflen;
}
add_stat_sample(td, &td->slat_stat, msec);
}
-static void add_bw_sample(struct thread_data *td, unsigned long msec)
+static void add_bw_sample(struct thread_data *td)
{
unsigned long spent = mtime_since_now(&td->stat_sample_time);
unsigned long rate;
if (spent < td->bw_avg_time)
return;
- rate = ((td->this_io_kb - td->stat_io_kb) * 1024) / spent;
+ rate = (td->this_io_bytes - td->stat_io_bytes) / spent;
add_stat_sample(td, &td->bw_stat, rate);
if (td->bw_log)
add_log_sample(td, td->bw_log, rate);
gettimeofday(&td->stat_sample_time, NULL);
- td->stat_io_kb = td->this_io_kb;
+ td->stat_io_bytes = td->this_io_bytes;
+}
+
+/*
+ * busy looping version for the last few usec
+ */
+static void __usec_sleep(int usec)
+{
+ struct timeval start;
+
+ gettimeofday(&start, NULL);
+ while (utime_since_now(&start) < usec)
+ __asm__ __volatile__("rep;nop": : :"memory");
}
static void usec_sleep(int usec)
struct timespec rem;
do {
+ if (usec < 5000) {
+ __usec_sleep(usec);
+ break;
+ }
rem.tv_sec = rem.tv_nsec = 0;
nanosleep(&req, &rem);
if (!rem.tv_nsec)
break;
req.tv_nsec = rem.tv_nsec;
+ usec = rem.tv_nsec * 1000;
} while (1);
}
/*
* if rate blocks is set, sample is running
*/
- if (td->rate_kb) {
+ if (td->rate_bytes) {
spent = mtime_since(&td->lastrate, now);
if (spent < td->ratecycle)
return 0;
- rate = ((td->this_io_kb - td->rate_kb) * 1024) / spent;
+ rate = (td->this_io_bytes - td->rate_bytes) / spent;
if (rate < td->ratemin) {
printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
if (rate_quit)
}
}
- td->rate_kb = td->this_io_kb;
+ td->rate_bytes = td->this_io_bytes;
memcpy(&td->lastrate, now, sizeof(*now));
return 0;
}
static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
{
+ if (!td->timeout)
+ return 0;
if (mtime_since(&td->start, t) >= td->timeout * 1000)
return 1;
static void fill_random_bytes(struct thread_data *td,
unsigned char *p, unsigned int len)
{
+ unsigned int todo;
double r;
- int todo;
while (len) {
drand48_r(&td->verify_state, &r);
}
}
-static int verify_io_u(struct thread_data *td, struct io_u *io_u)
+static void hexdump(void *buffer, int len)
+{
+ unsigned char *p = buffer;
+ int i;
+
+ for (i = 0; i < len; i++)
+ printf("%02x", p[i]);
+ printf("\n");
+}
+
+static int verify_io_u(struct io_u *io_u)
{
struct verify_header *hdr = (struct verify_header *) io_u->buf;
unsigned char *p = (unsigned char *) io_u->buf;
struct md5_ctx md5_ctx;
+ int ret;
if (hdr->fio_magic != FIO_HDR_MAGIC)
return 1;
p += sizeof(*hdr);
md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
- return memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
-}
-
-static int verify_io_us(struct thread_data *td, struct io_u *io_u, int *back)
-{
- struct verify_header *hdr;
- unsigned int left;
- struct io_u i;
- char *buf;
- off_t off;
-
- if (back)
- *back = 0;
-
- left = io_u->buflen;
- buf = io_u->buf;
- off = io_u->offset;
- while (left) {
- hdr = (struct verify_header *) buf;
- i.buf = buf;
- i.buflen = hdr->len;
-
- if (hdr->len > left) {
- if (back)
- *back = left;
- return 0;
- }
-
- if (verify_io_u(td, &i)) {
- printf("failed verify at offset %lu\n", (unsigned long) off);
- td->error = EBADMSG;
- return 1;
- }
-
- buf += hdr->len;
- left -= hdr->len;
- off += hdr->len;
+ ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
+ if (ret) {
+ hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
+ hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
}
- return 0;
+ return ret;
}
/*
if (!io_u)
return NULL;
- io_u->offset = get_next_offset(td);
+ if (get_next_offset(td, &io_u->offset)) {
+ put_io_u(td, io_u);
+ return NULL;
+ }
+
io_u->buflen = get_next_buflen(td);
if (!io_u->buflen) {
put_io_u(td, io_u);
return NULL;
}
- if (!td_read(td) && td->verify)
+ if (io_u->buflen + io_u->offset > td->file_size)
+ io_u->buflen = td->file_size - io_u->offset;
+
+ if (!td->sequential)
+ mark_random_map(td, io_u);
+
+ td->last_bytes += io_u->buflen;
+
+ if (td->verify)
populate_io_u(td, io_u);
if (td->use_aio) {
td->runstate = runstate;
}
-static int do_sync_verify(struct thread_data *td)
+static int get_next_verify(struct thread_data *td,
+ unsigned long long *offset, unsigned int *len)
+{
+ struct io_piece *ipo;
+
+ if (list_empty(&td->io_hist_list))
+ return 1;
+
+ ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
+ list_del(&ipo->list);
+
+ *offset = ipo->offset;
+ *len = ipo->len;
+ free(ipo);
+ return 0;
+}
+
+static void prune_io_piece_log(struct thread_data *td)
+{
+ struct io_piece *ipo;
+
+ while (!list_empty(&td->io_hist_list)) {
+ ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
+
+ list_del(&ipo->list);
+ free(ipo);
+ }
+}
+
+/*
+ * log a succesful write, so we can unwind the log for verify
+ */
+static void log_io_piece(struct thread_data *td, struct io_u *io_u)
+{
+ struct io_piece *ipo = malloc(sizeof(struct io_piece));
+ struct list_head *entry;
+
+ INIT_LIST_HEAD(&ipo->list);
+ ipo->offset = io_u->offset;
+ ipo->len = io_u->buflen;
+
+ /*
+ * for random io where the writes extend the file, it will typically
+ * be laid out with the block scattered as written. it's faster to
+ * read them in in that order again, so don't sort
+ */
+ if (td->sequential || !td->overwrite) {
+ list_add_tail(&ipo->list, &td->io_hist_list);
+ return;
+ }
+
+ /*
+ * for random io, sort the list so verify will run faster
+ */
+ entry = &td->io_hist_list;
+ while ((entry = entry->prev) != &td->io_hist_list) {
+ struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
+
+ if (__ipo->offset < ipo->offset)
+ break;
+ }
+
+ list_add(&ipo->list, entry);
+}
+
+static void do_sync_verify(struct thread_data *td)
{
struct timeval t;
struct io_u *io_u = NULL;
- loff_t off = 0;
- int back, ret;
+ int ret;
td_set_runstate(td, TD_VERIFYING);
io_u = __get_io_u(td);
- if (lseek(td->fd, 0, SEEK_SET) < 0) {
- td->error = errno;
- goto out;
- }
-
if (!td->odirect) {
- unsigned long size = td->kb << 10;
-
- if (fadvise(td->fd, 0, size, POSIX_FADV_DONTNEED) < 0) {
+ if (fadvise(td->fd, td->file_offset, td->io_size, POSIX_FADV_DONTNEED) < 0) {
td->error = errno;
goto out;
}
if (runtime_exceeded(td, &t))
break;
- io_u->offset = off;
- io_u->buflen = td->max_bs;
+ if (get_next_verify(td, &io_u->offset, &io_u->buflen))
+ break;
+
+ if (td->cur_off != io_u->offset) {
+ if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
+ td->error = errno;
+ break;
+ }
+ }
ret = read(td->fd, io_u->buf, io_u->buflen);
if (ret < (int) io_u->buflen) {
io_u->buflen = ret;
}
- if (verify_io_us(td, io_u, &back))
+ if (verify_io_u(io_u))
break;
- if (back) {
- ret -= back;
- if (lseek(td->fd, -back, SEEK_CUR) < 0) {
- td->error = errno;
- break;
- }
- }
-
- off += ret;
+ td->cur_off = io_u->offset + io_u->buflen;
} while (1);
out:
td_set_runstate(td, TD_RUNNING);
put_io_u(td, io_u);
-
- return td->error == 0;
}
static void do_sync_io(struct thread_data *td)
struct io_u *io_u = NULL;
struct timeval e;
- for (td->this_io_kb = 0; td->this_io_kb < td->kb;) {
+ while (td->this_io_bytes < td->io_size) {
int ret;
if (td->terminate)
break;
}
+ if (td_write(td))
+ log_io_piece(td, io_u);
+
td->io_blocks++;
- td->io_kb += io_u->buflen >> 10;
- td->this_io_kb += io_u->buflen >> 10;
+ td->io_bytes += io_u->buflen;
+ td->this_io_bytes += io_u->buflen;
td->cur_off = io_u->offset + io_u->buflen;
gettimeofday(&e, NULL);
msec = usec / 1000;
add_clat_sample(td, msec);
- add_bw_sample(td, msec);
+ add_bw_sample(td);
if (runtime_exceeded(td, &e))
break;
fsync(td->fd);
}
+static int io_u_getevents(struct thread_data *td, int min, int max,
+ struct timespec *t)
+{
+ int r;
+
+ do {
+ r = io_getevents(td->aio_ctx, min, max, td->aio_events, t);
+ if (r != -EAGAIN && r != -EINTR)
+ break;
+ } while (1);
+
+ return r;
+}
+
static int io_u_queue(struct thread_data *td, struct io_u *io_u)
{
struct iocb *iocb = &io_u->iocb;
ret = io_submit(td->aio_ctx, 1, &iocb);
if (ret == 1)
return 0;
- else if (ret == EAGAIN)
+ else if (ret == -EAGAIN)
usleep(100);
- else if (ret == EINTR)
+ else if (ret == -EINTR)
continue;
else
break;
io_u = ev_to_iou(td->aio_events + i);
td->io_blocks++;
- td->io_kb += io_u->buflen >> 10;
- td->this_io_kb += io_u->buflen >> 10;
+ td->io_bytes += io_u->buflen;
+ td->this_io_bytes += io_u->buflen;
msec = mtime_since(&io_u->issue_time, &e);
add_clat_sample(td, msec);
- add_bw_sample(td, msec);
+ add_bw_sample(td);
+
+ if (td_write(td))
+ log_io_piece(td, io_u);
bytes_done += io_u->buflen;
put_io_u(td, io_u);
/*
* get immediately available events, if any
*/
- r = io_getevents(td->aio_ctx, 0, td->cur_depth, td->aio_events, &ts);
+ r = io_u_getevents(td, 0, td->cur_depth, &ts);
if (r > 0)
ios_completed(td, r);
}
if (td->cur_depth) {
- r = io_getevents(td->aio_ctx, td->cur_depth, td->cur_depth, td->aio_events, NULL);
+ r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
if (r > 0)
ios_completed(td, r);
}
}
-static int do_async_verify(struct thread_data *td)
+static int async_do_verify(struct thread_data *td, struct io_u **io_u)
+{
+ struct io_u *v_io_u = *io_u;
+ int ret = 0;
+
+ if (v_io_u) {
+ ret = verify_io_u(v_io_u);
+ put_io_u(td, v_io_u);
+ *io_u = NULL;
+ }
+
+ return ret;
+}
+
+static void do_async_verify(struct thread_data *td)
{
struct timeval t;
- struct io_u *io_u, *v_io_u;
- struct verify_header *hdr;
- int ret, back;
- char *p;
+ struct io_u *io_u, *v_io_u = NULL;
+ int ret;
td_set_runstate(td, TD_VERIFYING);
- td->cur_off = 0;
- td->last_kb = 0;
- v_io_u = NULL;
-
do {
if (td->terminate)
break;
if (!io_u)
break;
- io_u->offset = td->cur_off;
- io_u->buflen = td->max_bs;
-
- if (io_u->offset + io_u->buflen > (td->kb << 10)) {
- io_u->buflen = (td->kb << 10) - io_u->offset;
- if (!io_u->buflen) {
- put_io_u(td, io_u);
- break;
- }
+ if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
+ put_io_u(td, io_u);
+ break;
}
io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
* we have one pending to verify, do that while the next
* we are doing io on the next one
*/
- if (v_io_u) {
- ret = verify_io_us(td, v_io_u, NULL);
- put_io_u(td, v_io_u);
- v_io_u = NULL;
- if (ret)
- break;
- }
+ if (async_do_verify(td, &v_io_u))
+ break;
- ret = io_getevents(td->aio_ctx, 1, 1, td->aio_events, NULL);
+ ret = io_u_getevents(td, 1, 1, NULL);
if (ret != 1) {
if (ret < 0)
td->error = ret;
break;
}
- /*
- * got our io_u to verify, find back offset so we can
- * submit the next one before verifying this one
- */
v_io_u = ev_to_iou(td->aio_events);
- p = v_io_u->buf;
- back = v_io_u->buflen;
- do {
- hdr = (struct verify_header *) p;
- if (hdr->len > back)
- break;
-
- back -= hdr->len;
- p += hdr->len;
- } while (back);
-
- td->cur_off += (v_io_u->buflen - back);
+ td->cur_off = v_io_u->offset + v_io_u->buflen;
/*
- * if max depth is 1, we need to verify now
+ * if we can't submit more io, we need to verify now
*/
- if (queue_full(td)) {
- ret = verify_io_us(td, v_io_u, NULL);
- put_io_u(td, v_io_u);
- v_io_u = NULL;
- if (ret)
- break;
- }
+ if (queue_full(td) && async_do_verify(td, &v_io_u))
+ break;
+
} while (1);
- if (v_io_u) {
- verify_io_us(td, v_io_u, NULL);
- put_io_u(td, v_io_u);
- }
+ async_do_verify(td, &v_io_u);
if (td->cur_depth)
cleanup_pending_aio(td);
td_set_runstate(td, TD_RUNNING);
- return td->error == 0;
}
static void do_async_io(struct thread_data *td)
struct timeval s, e;
unsigned long usec;
- for (td->this_io_kb = 0; td->this_io_kb < td->kb;) {
+ while (td->this_io_bytes < td->io_size) {
struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
struct timespec *timeout;
int ret, min_evts = 0;
min_evts = 1;
}
- ret = io_getevents(td->aio_ctx, min_evts, td->cur_depth, td->aio_events, timeout);
+ ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
if (ret < 0) {
- td->error = errno;
+ td->error = ret;
break;
} else if (!ret)
continue;
INIT_LIST_HEAD(&td->io_u_freelist);
INIT_LIST_HEAD(&td->io_u_busylist);
+ INIT_LIST_HEAD(&td->io_hist_list);
p = ALIGN(td->orig_buffer);
for (i = 0; i < max_units; i++) {
{
char file_name[128];
FILE *f;
- int i;
+ unsigned int i;
sprintf(file_name, "client%d_%s.log", td->thread_number, name);
f = fopen(file_name, "w");
static int create_file(struct thread_data *td)
{
unsigned long long left;
+ unsigned int bs;
char *b;
- int r, bs;
+ int r;
/*
* unless specifically asked for overwrite, let normal io extend it
*/
- if (!td_read(td) && !td->overwrite)
+ if (td_write(td) && !td->overwrite)
return 0;
if (!td->file_size) {
return 1;
}
- td->kb = td->file_size >> 10;
+ if (ftruncate(td->fd, td->file_size) == -1) {
+ td->error = errno;
+ return 1;
+ }
+
+ td->io_size = td->file_size;
b = malloc(td->max_bs);
memset(b, 0, td->max_bs);
r = write(td->fd, b, bs);
- if (r == bs) {
+ if (r == (int) bs) {
left -= bs;
continue;
} else {
return errno != ENOENT;
}
-static int setup_file(struct thread_data *td)
+static int get_file_size(struct thread_data *td)
{
+ size_t bytes = 0;
struct stat st;
+
+ if (fstat(td->fd, &st) == -1) {
+ td->error = errno;
+ return 1;
+ }
+
+ /*
+ * if block device, get size via BLKGETSIZE64 ioctl. try that as well
+ * if this is a link, fall back to st.st_size if it fails
+ */
+ if (S_ISBLK(st.st_mode) || S_ISLNK(st.st_mode)) {
+ if (ioctl(td->fd, BLKGETSIZE64, &bytes)) {
+ if (S_ISBLK(st.st_mode)) {
+ td->error = errno;
+ return 1;
+ } else
+ bytes = st.st_size;
+ }
+ } else
+ bytes = st.st_size;
+
+ if (td_read(td)) {
+ if (td->file_size > bytes)
+ bytes = td->file_size;
+ } else {
+ if (!td->file_size)
+ td->file_size = 1024 * 1024 * 1024;
+
+ bytes = td->file_size;
+ }
+
+ if (td->file_offset > bytes) {
+ fprintf(stderr, "Client%d: offset larger than length\n", td->thread_number);
+ return 1;
+ }
+
+ td->io_size = bytes - td->file_offset;
+ if (td->io_size == 0) {
+ fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
+ td->error = EINVAL;
+ return 1;
+ }
+
+ return 0;
+}
+
+static int setup_file(struct thread_data *td)
+{
int flags = 0;
if (!file_exists(td)) {
return 1;
}
- if (td_read(td)) {
- if (fstat(td->fd, &st) == -1) {
- td->error = errno;
- return 1;
- }
-
- if (td->file_size > st.st_size)
- st.st_size = td->file_size;
- } else {
- if (!td->file_size)
- td->file_size = 1024 * 1024 * 1024;
-
- st.st_size = td->file_size;
- }
+ if (get_file_size(td))
+ return 1;
- td->kb = (st.st_size - td->file_offset) / 1024;
- if (!td->kb) {
- fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
- td->error = EINVAL;
+ if (td_write(td) && ftruncate(td->fd, td->file_size) == -1) {
+ td->error = errno;
return 1;
}
if (td->invalidate_cache) {
- if (fadvise(td->fd, 0, st.st_size, POSIX_FADV_DONTNEED) < 0) {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
td->error = errno;
return 1;
}
return 0;
}
+static void clear_io_state(struct thread_data *td)
+{
+ if (!td->use_aio)
+ lseek(td->fd, SEEK_SET, 0);
+
+ td->cur_off = 0;
+ td->last_bytes = 0;
+ td->stat_io_bytes = 0;
+ td->this_io_bytes = 0;
+
+ if (td->file_map)
+ memset(td->file_map, 0, td->num_maps * sizeof(long));
+}
+
static void *thread_main(int shm_id, int offset, char *argv[])
{
struct thread_data *td;
if (td->use_aio && init_aio(td))
goto err;
- if (init_random_state(td))
- goto err;
-
if (td->ioprio) {
if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
td->error = errno;
if (!td->create_serialize && setup_file(td))
goto err;
+ if (init_random_state(td))
+ goto err;
+
gettimeofday(&td->start, NULL);
+ getrusage(RUSAGE_SELF, &td->ru_start);
+
while (td->loops--) {
gettimeofday(&td->stat_sample_time, NULL);
if (td->ratemin)
memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
- td->cur_off = 0;
- td->last_kb = 0;
- td->stat_io_kb = 0;
+ clear_io_state(td);
+ prune_io_piece_log(td);
- if (!td->use_aio) {
+ if (!td->use_aio)
do_sync_io(td);
-
- if (td->verify) {
- if (!do_sync_verify(td))
- break;
- }
- } else {
+ else
do_async_io(td);
- if (td->verify) {
- if (!do_async_verify(td))
- break;
- }
- }
+ if (td->error)
+ break;
+
+ if (!td->verify)
+ continue;
+
+ clear_io_state(td);
+
+ if (!td->use_aio)
+ do_sync_verify(td);
+ else
+ do_async_verify(td);
+
+ if (td->error)
+ break;
}
td->runtime = mtime_since_now(&td->start);
+ getrusage(RUSAGE_SELF, &td->ru_end);
ret = 0;
if (td->bw_log)
static void show_thread_status(struct thread_data *td)
{
int prio, prio_class;
- unsigned long min, max, bw = 0;
- double mean, dev;
+ unsigned long min, max, bw = 0, ctx;
+ double mean, dev, usr_cpu, sys_cpu;
- if (!td->io_kb && !td->error)
+ if (!td->io_bytes && !td->error)
return;
if (td->runtime)
- bw = td->io_kb * 1024 / td->runtime;
+ bw = td->io_bytes / td->runtime;
prio = td->ioprio & 0xff;
prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;
- printf("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->error, td->io_kb >> 10, bw, td->runtime);
+ printf("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->error, td->io_bytes >> 20, bw, td->runtime);
if (calc_lat(&td->slat_stat, &min, &max, &mean, &dev))
printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev))
printf(" bw (KiB/s) : min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
+
+ if (td->runtime) {
+ unsigned long t;
+
+ t = mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
+ usr_cpu = (double) t * 100 / (double) td->runtime;
+
+ t = mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
+ sys_cpu = (double) t * 100 / (double) td->runtime;
+ } else {
+ usr_cpu = 0;
+ sys_cpu = 0;
+ }
+
+ ctx = td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
+
+ printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ctx);
}
static int setup_rate(struct thread_data *td)
return -1;
}
- nr_reads_per_sec = td->rate * 1024 / td->min_bs;
+ nr_reads_per_sec = (td->rate * 1024) / td->min_bs;
td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
td->rate_pending_usleep = 0;
return 0;
}
-static struct thread_data *get_new_job(int global)
+static struct thread_data *get_new_job(int global, struct thread_data *parent)
{
struct thread_data *td;
td = &threads[thread_number++];
memset(td, 0, sizeof(*td));
+ sprintf(td->directory, ".");
+
td->fd = -1;
td->thread_number = thread_number;
- td->ddir = def_thread.ddir;
- td->ioprio = def_thread.ioprio;
- td->sequential = def_thread.sequential;
- td->bs = def_thread.bs;
- td->min_bs = def_thread.min_bs;
- td->max_bs = def_thread.max_bs;
- td->odirect = def_thread.odirect;
- td->thinktime = def_thread.thinktime;
- td->fsync_blocks = def_thread.fsync_blocks;
- td->start_delay = def_thread.start_delay;
- td->timeout = def_thread.timeout;
- td->use_aio = def_thread.use_aio;
- td->create_file = def_thread.create_file;
- td->overwrite = def_thread.overwrite;
- td->invalidate_cache = def_thread.invalidate_cache;
- td->file_size = def_thread.file_size;
- td->file_offset = def_thread.file_offset;
- td->rate = def_thread.rate;
- td->ratemin = def_thread.ratemin;
- td->ratecycle = def_thread.ratecycle;
- td->aio_depth = def_thread.aio_depth;
- td->sync_io = def_thread.sync_io;
- td->mem_type = def_thread.mem_type;
- td->bw_avg_time = def_thread.bw_avg_time;
- td->create_serialize = def_thread.create_serialize;
- td->create_fsync = def_thread.create_fsync;
- td->loops = def_thread.loops;
- td->verify = def_thread.verify;
- memcpy(&td->cpumask, &def_thread.cpumask, sizeof(td->cpumask));
+ td->ddir = parent->ddir;
+ td->ioprio = parent->ioprio;
+ td->sequential = parent->sequential;
+ td->bs = parent->bs;
+ td->min_bs = parent->min_bs;
+ td->max_bs = parent->max_bs;
+ td->odirect = parent->odirect;
+ td->thinktime = parent->thinktime;
+ td->fsync_blocks = parent->fsync_blocks;
+ td->start_delay = parent->start_delay;
+ td->timeout = parent->timeout;
+ td->use_aio = parent->use_aio;
+ td->create_file = parent->create_file;
+ td->overwrite = parent->overwrite;
+ td->invalidate_cache = parent->invalidate_cache;
+ td->file_size = parent->file_size;
+ td->file_offset = parent->file_offset;
+ td->rate = parent->rate;
+ td->ratemin = parent->ratemin;
+ td->ratecycle = parent->ratecycle;
+ td->aio_depth = parent->aio_depth;
+ td->sync_io = parent->sync_io;
+ td->mem_type = parent->mem_type;
+ td->bw_avg_time = parent->bw_avg_time;
+ td->create_serialize = parent->create_serialize;
+ td->create_fsync = parent->create_fsync;
+ td->loops = parent->loops;
+ td->verify = parent->verify;
+ td->stonewall = parent->stonewall;
+ td->numjobs = parent->numjobs;
+ memcpy(&td->cpumask, &parent->cpumask, sizeof(td->cpumask));
return td;
}
thread_number--;
}
-static int add_job(struct thread_data *td, const char *filename, int prioclass,
+static int add_job(struct thread_data *td, const char *jobname, int prioclass,
int prio)
{
+ int numjobs;
+
if (td == &def_thread)
return 0;
- strcpy(td->file_name, filename);
+ sprintf(td->file_name, "%s/%s.%d", td->directory, jobname, td->thread_number);
sem_init(&td->mutex, 1, 0);
td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;
if (td->use_aio && !td->aio_depth)
td->aio_depth = 1;
- if (td->min_bs == -1)
+ if (td->min_bs == -1U)
td->min_bs = td->bs;
- if (td->max_bs == -1)
+ if (td->max_bs == -1U)
td->max_bs = td->bs;
+ if (td_read(td))
+ td->verify = 0;
if (setup_rate(td))
- return -1;
+ goto err;
if (write_lat_log)
setup_log(&td->lat_log);
if (write_bw_log)
setup_log(&td->bw_log);
- printf("Client%d: file=%s, rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d-%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, filename, td->ddir, prioclass, prio, td->sequential, td->odirect, td->min_bs, td->max_bs, td->rate, td->use_aio, td->aio_depth);
+ printf("Client%d: rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d-%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, td->ddir, prioclass, prio, td->sequential, td->odirect, td->min_bs, td->max_bs, td->rate, td->use_aio, td->aio_depth);
+
+ /*
+ * recurse add identical jobs, clear numjobs and stonewall options
+ * as they don't apply to sub-jobs
+ */
+ numjobs = td->numjobs;
+ while (--numjobs) {
+ struct thread_data *td_new = get_new_job(0, td);
+
+ if (!td_new)
+ break;
+
+ td_new->numjobs = 1;
+ td_new->stonewall = 0;
+
+ if (add_job(td_new, jobname, prioclass, prio))
+ break;
+ }
return 0;
+err:
+ put_job(td);
+ return -1;
}
static void fill_cpu_mask(cpu_set_t cpumask, int cpu)
return 1;
}
+static int check_strstore(char *p, char *name, char *dest)
+{
+ char *s = strstr(p, name);
+
+ if (!s)
+ return 1;
+
+ s = strstr(p, "=");
+ if (!s)
+ return 1;
+
+ s++;
+ while (isblank(*s))
+ s++;
+
+ strcpy(dest, s);
+ return 0;
+}
+
static int check_range(char *p, char *name, unsigned long *s, unsigned long *e)
{
char str[128];
name[strlen(name) - 1] = '\0';
- td = get_new_job(global);
+ td = get_new_job(global, &def_thread);
if (!td)
break;
fgetpos(f, &off);
continue;
}
+ if (!check_int(p, "numjobs", &td->numjobs)) {
+ fgetpos(f, &off);
+ continue;
+ }
if (!check_range(p, "bsrange", &ul1, &ul2)) {
- td->min_bs = ul1;
- td->max_bs = ul2;
+ if (ul1 & 511)
+ printf("bad min block size, must be a multiple of 512\n");
+ else
+ td->min_bs = ul1;
+ if (ul2 & 511)
+ printf("bad max block size, must be a multiple of 512\n");
+ else
+ td->max_bs = ul2;
fgetpos(f, &off);
continue;
}
if (!check_strcnv(p, "bs", &ull)) {
- td->bs = ull;
+ if (ull & 511)
+ printf("bad block size, must be a multiple of 512\n");
+ else
+ td->bs = ull;
fgetpos(f, &off);
continue;
}
fgetpos(f, &off);
continue;
}
+ if (!check_strstore(p, "directory", td->directory)) {
+ printf("got dir %s\n", td->directory);
+ fgetpos(f, &off);
+ continue;
+ }
if (!check_str(p, "mem", "malloc")) {
td->mem_type = MEM_MALLOC;
fgetpos(f, &off);
fgetpos(f, &off);
continue;
}
+ if (!strncmp(p, "stonewall", 9)) {
+ td->stonewall = 1;
+ fgetpos(f, &off);
+ continue;
+ }
printf("Client%d: bad option %s\n",td->thread_number,p);
}
fsetpos(f, &off);
if (add_job(td, name, prioclass, prio))
- put_job(td);
+ break;
}
free(string);
c = 'E';
break;
case TD_RUNNING:
- c = '+';
+ if (td_read(td)) {
+ if (td->sequential)
+ c = 'R';
+ else
+ c = 'r';
+ } else {
+ if (td->sequential)
+ c = 'W';
+ else
+ c = 'w';
+ }
break;
case TD_VERIFYING:
c = 'V';
struct timeval genesis;
struct thread_data *td;
unsigned long spent;
- int i, todo, nr_running, m_rate, t_rate;
+ int i, todo, nr_running, m_rate, t_rate, nr_started;
printf("Starting %d threads\n", thread_number);
fflush(stdout);
todo = thread_number;
nr_running = 0;
+ nr_started = 0;
m_rate = t_rate = 0;
for (i = 0; i < thread_number; i++) {
continue;
}
+ if (td->stonewall && (nr_started || nr_running))
+ break;
+
td_set_runstate(td, TD_CREATED);
check_str_update(td, nr_running, t_rate, m_rate);
sem_init(&startup_sem, 1, 1);
todo--;
+ nr_started++;
if (fork())
sem_wait(&startup_sem);
td_set_runstate(td, TD_RUNNING);
nr_running++;
+ nr_started--;
m_rate += td->ratemin;
t_rate += td->rate;
check_str_update(td, nr_running, t_rate, m_rate);
for (i = 0; i < thread_number; i++) {
struct thread_data *td = &threads[i];
- if (td->runstate == TD_RUNNING)
- run_str[td->thread_number - 1] = '+';
- else if (td->runstate == TD_VERIFYING)
- run_str[td->thread_number - 1] = 'V';
- else
+ if (td->runstate != TD_RUNNING &&
+ td->runstate != TD_VERIFYING)
continue;
check_str_update(td, nr_running, t_rate, m_rate);
def_thread.create_fsync = DEF_CREATE_FSYNC;
def_thread.loops = DEF_LOOPS;
def_thread.verify = DEF_VERIFY;
+ def_thread.stonewall = DEF_STONEWALL;
+ def_thread.numjobs = DEF_NUMJOBS;
i = parse_options(argc, argv);
max_run[td->ddir] = td->runtime;
if (td->runtime)
- bw = td->io_kb * 1024 / td->runtime;
+ bw = td->io_bytes / td->runtime;
if (bw < min_bw[td->ddir])
min_bw[td->ddir] = bw;
if (bw > max_bw[td->ddir])
max_bw[td->ddir] = bw;
- io_mb[td->ddir] += td->io_kb >> 10;
+ io_mb[td->ddir] += td->io_bytes >> 20;
}
show_thread_status(td);
}
if (max_run[0])
- agg[0] = io_mb[0] * 1024 * 1000 / max_run[0];
+ agg[0] = (io_mb[0] * 1024 * 1000) / max_run[0];
if (max_run[1])
- agg[1] = io_mb[1] * 1024 * 1000 / max_run[1];
+ agg[1] = (io_mb[1] * 1024 * 1000) / max_run[1];
printf("\nRun status:\n");
if (max_run[DDIR_READ])