#include <time.h>
#include <math.h>
#include <assert.h>
-#include <pthread.h>
+#include <dirent.h>
+#include <libgen.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/wait.h>
#include <sys/ipc.h>
#include <sys/shm.h>
#include <sys/ioctl.h>
-#include <asm/unistd.h>
+#include <sys/mman.h>
#include "fio.h"
+#include "os.h"
#define MASK (4095)
int thread_number = 0;
char run_str[MAX_JOBS + 1];
int shm_id = 0;
+static LIST_HEAD(disk_list);
+
+static void update_io_ticks(void);
+static void disk_util_timer_arm(void);
/*
* thread life cycle
TD_REAPED,
};
-/*
- * The io unit
- */
-struct io_u {
- struct iocb iocb;
- struct timeval start_time;
- struct timeval issue_time;
-
- char *buf;
- unsigned int buflen;
- unsigned long long offset;
-
- struct list_head list;
-};
-
#define should_fsync(td) (td_write(td) && !(td)->odirect)
static sem_t startup_sem;
#define TERMINATE_ALL (-1)
-static void terminate_threads(int groupid)
+static void terminate_threads(int group_id)
{
int i;
for (i = 0; i < thread_number; i++) {
struct thread_data *td = &threads[i];
- if (groupid == TERMINATE_ALL || groupid == td->groupid) {
+ if (group_id == TERMINATE_ALL || groupid == td->groupid) {
td->terminate = 1;
td->start_delay = 0;
}
static void sig_handler(int sig)
{
- terminate_threads(TERMINATE_ALL);
+ switch (sig) {
+ case SIGALRM:
+ update_io_ticks();
+ disk_util_timer_arm();
+ break;
+ default:
+ printf("\nfio: terminating on signal\n");
+ fflush(stdout);
+ terminate_threads(TERMINATE_ALL);
+ break;
+ }
}
static unsigned long utime_since(struct timeval *s, struct timeval *e)
unsigned int blocks = 0;
while (blocks < (io_u->buflen / td->min_bs)) {
- int idx, bit;
+ unsigned int idx, bit;
if (!random_map_free(td, block))
break;
return buflen;
}
-static inline void add_stat_sample(struct thread_data *td, struct io_stat *is,
- unsigned long val)
+static inline void add_stat_sample(struct io_stat *is, unsigned long val)
{
if (val > is->max_val)
is->max_val = val;
is->samples++;
}
-static void add_log_sample(struct thread_data *td, struct io_log *log,
+static void add_log_sample(struct thread_data *td, struct io_log *iolog,
unsigned long val)
{
- if (log->nr_samples == log->max_samples) {
- int new_size = sizeof(struct io_sample) * log->max_samples * 2;
+ if (iolog->nr_samples == iolog->max_samples) {
+ int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
- log->log = realloc(log->log, new_size);
- log->max_samples <<= 1;
+ iolog->log = realloc(iolog->log, new_size);
+ iolog->max_samples <<= 1;
}
- log->log[log->nr_samples].val = val;
- log->log[log->nr_samples].time = mtime_since_now(&td->start);
- log->nr_samples++;
+ iolog->log[iolog->nr_samples].val = val;
+ iolog->log[iolog->nr_samples].time = mtime_since_now(&td->start);
+ iolog->nr_samples++;
}
static void add_clat_sample(struct thread_data *td, unsigned long msec)
{
- add_stat_sample(td, &td->clat_stat, msec);
+ add_stat_sample(&td->clat_stat, msec);
if (td->lat_log)
add_log_sample(td, td->lat_log, msec);
static void add_slat_sample(struct thread_data *td, unsigned long msec)
{
- add_stat_sample(td, &td->slat_stat, msec);
+ add_stat_sample(&td->slat_stat, msec);
}
static void add_bw_sample(struct thread_data *td)
return;
rate = (td->this_io_bytes - td->stat_io_bytes) / spent;
- add_stat_sample(td, &td->bw_stat, rate);
+ add_stat_sample(&td->bw_stat, rate);
if (td->bw_log)
add_log_sample(td, td->bw_log, rate);
/*
* busy looping version for the last few usec
*/
-static void __usec_sleep(int usec)
+static void __usec_sleep(unsigned int usec)
{
struct timeval start;
nop;
}
-static void usec_sleep(int usec)
+static void usec_sleep(struct thread_data *td, unsigned long usec)
{
- struct timespec req = { .tv_sec = 0, .tv_nsec = usec * 1000 };
- struct timespec rem;
+ struct timespec req, rem;
+
+ req.tv_sec = usec / 1000000;
+ req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
do {
if (usec < 5000) {
__usec_sleep(usec);
break;
}
+
rem.tv_sec = rem.tv_nsec = 0;
- nanosleep(&req, &rem);
- if (!rem.tv_nsec)
+ if (nanosleep(&req, &rem) < 0)
+ break;
+
+ if ((rem.tv_sec + rem.tv_nsec) == 0)
break;
req.tv_nsec = rem.tv_nsec;
- usec = rem.tv_nsec * 1000;
- } while (1);
+ req.tv_sec = rem.tv_sec;
+
+ usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
+ } while (!td->terminate);
}
static void rate_throttle(struct thread_data *td, unsigned long time_spent,
td->rate_pending_usleep += s;
if (td->rate_pending_usleep >= 100000) {
- usec_sleep(td->rate_pending_usleep);
+ usec_sleep(td, td->rate_pending_usleep);
td->rate_pending_usleep = 0;
}
} else {
{
if (!td->timeout)
return 0;
- if (mtime_since(&td->start, t) >= td->timeout * 1000)
+ if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
return 1;
return 0;
printf("\n");
}
-static int verify_io_u(struct io_u *io_u)
+static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
+{
+ unsigned char *p = (unsigned char *) io_u->buf;
+ unsigned long c;
+
+ p += sizeof(*hdr);
+ c = crc32(p, hdr->len - sizeof(*hdr));
+
+ return c != hdr->crc32;
+}
+
+static int verify_io_u_md5(struct verify_header *hdr, 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;
-
memset(&md5_ctx, 0, sizeof(md5_ctx));
p += sizeof(*hdr);
md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
return ret;
}
+static int verify_io_u(struct io_u *io_u)
+{
+ struct verify_header *hdr = (struct verify_header *) io_u->buf;
+ int ret;
+
+ if (hdr->fio_magic != FIO_HDR_MAGIC)
+ return 1;
+
+ if (hdr->verify_type == VERIFY_MD5)
+ ret = verify_io_u_md5(hdr, io_u);
+ else if (hdr->verify_type == VERIFY_CRC32)
+ ret = verify_io_u_crc32(hdr, io_u);
+ else {
+ fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
+ ret = 1;
+ }
+
+ return ret;
+}
+
+static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
+{
+ hdr->crc32 = crc32(p, len);
+}
+
+static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
+{
+ struct md5_ctx md5_ctx;
+
+ memset(&md5_ctx, 0, sizeof(md5_ctx));
+ md5_update(&md5_ctx, p, len);
+ memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
+}
+
/*
* fill body of io_u->buf with random data and add a header with the
* (eg) sha1sum of that data.
*/
static void populate_io_u(struct thread_data *td, struct io_u *io_u)
{
- struct md5_ctx md5_ctx;
- struct verify_header hdr;
unsigned char *p = (unsigned char *) io_u->buf;
+ struct verify_header hdr;
hdr.fio_magic = FIO_HDR_MAGIC;
hdr.len = io_u->buflen;
p += sizeof(hdr);
fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
- memset(&md5_ctx, 0, sizeof(md5_ctx));
- md5_update(&md5_ctx, p, io_u->buflen - sizeof(hdr));
- memcpy(hdr.md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
+ if (td->verify == VERIFY_MD5) {
+ fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
+ hdr.verify_type = VERIFY_MD5;
+ } else {
+ fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
+ hdr.verify_type = VERIFY_CRC32;
+ }
+
memcpy(io_u->buf, &hdr, sizeof(hdr));
}
if (io_u->buflen + io_u->offset > td->file_size)
io_u->buflen = td->file_size - io_u->offset;
+ if (!io_u->buflen) {
+ put_io_u(td, io_u);
+ return NULL;
+ }
+
if (!td->sequential)
mark_random_map(td, io_u);
td->last_bytes += io_u->buflen;
- if (td->verify)
+ if (td->verify != VERIFY_NONE)
populate_io_u(td, io_u);
- if (td->use_aio) {
- if (td_read(td))
- io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
- else
- io_prep_pwrite(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
- }
+ if (td->io_prep)
+ td->io_prep(td, io_u, td_read(td));
gettimeofday(&io_u->start_time, NULL);
return io_u;
io_u = __get_io_u(td);
if (!td->odirect) {
- if (fadvise(td->fd, td->file_offset, td->io_size, POSIX_FADV_DONTNEED) < 0) {
- td->error = errno;
- goto out;
+ if (!td->use_mmap) {
+ if (fadvise(td->fd, td->file_offset, td->io_size, POSIX_FADV_DONTNEED) < 0) {
+ td_verror(td, errno);
+ goto out;
+ }
+ } else {
+ if (madvise(td->mmap, td->io_size, MADV_DONTNEED)) {
+ td_verror(td, errno);
+ goto out;
+ }
}
}
if (td->cur_off != io_u->offset) {
if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
- td->error = errno;
+ td_verror(td, errno);
break;
}
}
ret = read(td->fd, io_u->buf, io_u->buflen);
if (ret < (int) io_u->buflen) {
if (ret == -1) {
- td->error = errno;
+ td_verror(td, errno);
break;
} else if (!ret)
break;
put_io_u(td, io_u);
}
+static int __do_sync_mmap(struct thread_data *td, struct io_u *io_u)
+{
+ unsigned long long real_off = io_u->offset - td->file_offset;
+
+ if (td_read(td))
+ memcpy(io_u->buf, td->mmap + real_off, io_u->buflen);
+ else
+ memcpy(td->mmap + real_off, io_u->buf, io_u->buflen);
+
+ /*
+ * not really direct, but should drop the pages from the cache
+ */
+ if (td->odirect) {
+ msync(td->mmap + real_off, io_u->buflen, MS_SYNC);
+ madvise(td->mmap + real_off, io_u->buflen, MADV_DONTNEED);
+ }
+
+ return io_u->buflen;
+}
+
+static int __do_sync_rw(struct thread_data *td, struct io_u *io_u)
+{
+ if (td->cur_off != io_u->offset) {
+ if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
+ td_verror(td, errno);
+ return 1;
+ }
+ }
+
+ if (td_read(td))
+ return read(td->fd, io_u->buf, io_u->buflen);
+ else
+ return write(td->fd, io_u->buf, io_u->buflen);
+}
+
+static void sync_td(struct thread_data *td)
+{
+ if (!td->use_mmap)
+ fsync(td->fd);
+ else
+ msync(td->mmap, td->file_size, MS_SYNC);
+}
+
static void do_sync_io(struct thread_data *td)
{
unsigned long msec, usec;
if (!io_u)
break;
- if (td->cur_off != io_u->offset) {
- if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
- td->error = errno;
- break;
- }
- }
-
- if (td_read(td))
- ret = read(td->fd, io_u->buf, io_u->buflen);
+ if (!td->use_mmap)
+ ret = __do_sync_rw(td, io_u);
else
- ret = write(td->fd, io_u->buf, io_u->buflen);
+ ret = __do_sync_mmap(td, io_u);
if (ret < (int) io_u->buflen) {
if (ret == -1)
- td->error = errno;
+ td_verror(td, errno);
break;
}
rate_throttle(td, usec, io_u->buflen);
if (check_min_rate(td, &e)) {
- td->error = ENODATA;
+ td_verror(td, ENOMEM);
break;
}
io_u = NULL;
if (td->thinktime)
- usec_sleep(td->thinktime);
+ usec_sleep(td, td->thinktime);
if (should_fsync(td) && td->fsync_blocks &&
(td->io_blocks % td->fsync_blocks) == 0)
- fsync(td->fd);
+ sync_td(td);
}
if (io_u)
put_io_u(td, io_u);
if (should_fsync(td))
- fsync(td->fd);
+ sync_td(td);
}
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;
+ return td->io_getevents(td, min, max, t);
}
static int io_u_queue(struct thread_data *td, struct io_u *io_u)
{
- struct iocb *iocb = &io_u->iocb;
- int ret;
-
- do {
- ret = io_submit(td->aio_ctx, 1, &iocb);
- if (ret == 1)
- return 0;
- else if (ret == -EAGAIN)
- usleep(100);
- else if (ret == -EINTR)
- continue;
- else
- break;
- } while (1);
-
- return ret;
+ return td->io_queue(td, io_u);
}
#define iocb_time(iocb) ((unsigned long) (iocb)->data)
-#define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj)
static int ios_completed(struct thread_data *td, int nr)
{
gettimeofday(&e, NULL);
for (i = 0, bytes_done = 0; i < nr; i++) {
- io_u = ev_to_iou(td->aio_events + i);
+ io_u = td->io_event(td, i);
td->io_blocks++;
td->io_bytes += io_u->buflen;
list_for_each_safe(entry, n, &td->io_u_busylist) {
io_u = list_entry(entry, struct io_u, list);
- r = io_cancel(td->aio_ctx, &io_u->iocb, td->aio_events);
+ r = td->io_cancel(td, io_u);
if (!r)
put_io_u(td, io_u);
}
break;
}
- io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
+ td->io_prep(td, io_u, 1);
+
ret = io_u_queue(td, io_u);
if (ret) {
put_io_u(td, io_u);
- td->error = ret;
+ td_verror(td, ret);
break;
}
ret = io_u_getevents(td, 1, 1, NULL);
if (ret != 1) {
if (ret < 0)
- td->error = ret;
+ td_verror(td, ret);
break;
}
- v_io_u = ev_to_iou(td->aio_events);
+ v_io_u = td->io_event(td, 0);
td->cur_off = v_io_u->offset + v_io_u->buflen;
ret = io_u_queue(td, io_u);
if (ret) {
put_io_u(td, io_u);
- td->error = ret;
+ td_verror(td, ret);
break;
}
ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
if (ret < 0) {
- td->error = ret;
+ td_verror(td, ret);
break;
} else if (!ret)
continue;
rate_throttle(td, usec, bytes_done);
if (check_min_rate(td, &e)) {
- td->error = ENODATA;
+ td_verror(td, ENOMEM);
break;
}
break;
if (td->thinktime)
- usec_sleep(td->thinktime);
+ usec_sleep(td, td->thinktime);
if (should_fsync(td) && td->fsync_blocks &&
(td->io_blocks % td->fsync_blocks) == 0)
static void cleanup_aio(struct thread_data *td)
{
- io_destroy(td->aio_ctx);
-
- if (td->aio_events)
- free(td->aio_events);
+ if (td->io_engine == FIO_LIBAIO)
+ fio_libaio_cleanup(td);
+ else if (td->io_engine == FIO_POSIXAIO)
+ fio_posixaio_cleanup(td);
}
static int init_aio(struct thread_data *td)
{
- if (io_queue_init(td->aio_depth, &td->aio_ctx)) {
- td->error = errno;
+ if (td->io_engine == FIO_LIBAIO)
+ return fio_libaio_init(td);
+ else if (td->io_engine == FIO_POSIXAIO)
+ return fio_posixaio_init(td);
+ else {
+ fprintf(stderr, "bad io_engine %d\n", td->io_engine);
return 1;
}
-
- td->aio_events = malloc(td->aio_depth * sizeof(struct io_event));
- return 0;
}
static void cleanup_io_u(struct thread_data *td)
shmdt(td->orig_buffer);
shmctl(td->shm_id, IPC_RMID, &sbuf);
- }
+ } else if (td->mem_type == MEM_MMAP)
+ munmap(td->orig_buffer, td->orig_buffer_size);
+ else
+ fprintf(stderr, "Bad memory type %d\n", td->mem_type);
+
+ td->orig_buffer = NULL;
}
static int init_io_u(struct thread_data *td)
{
struct io_u *io_u;
- int i, max_units, mem_size;
+ int i, max_units;
char *p;
- if (!td->use_aio)
+ if (td->io_engine == FIO_SYNCIO)
max_units = 1;
else
max_units = td->aio_depth;
- mem_size = td->max_bs * max_units + MASK;
+ td->orig_buffer_size = td->max_bs * max_units + MASK;
if (td->mem_type == MEM_MALLOC)
- td->orig_buffer = malloc(mem_size);
+ td->orig_buffer = malloc(td->orig_buffer_size);
else if (td->mem_type == MEM_SHM) {
- td->shm_id = shmget(IPC_PRIVATE, mem_size, IPC_CREAT | 0600);
+ td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
if (td->shm_id < 0) {
- td->error = errno;
+ td_verror(td, errno);
perror("shmget");
return 1;
}
td->orig_buffer = shmat(td->shm_id, NULL, 0);
if (td->orig_buffer == (void *) -1) {
- td->error = errno;
+ td_verror(td, errno);
perror("shmat");
+ td->orig_buffer = NULL;
+ return 1;
+ }
+ } else if (td->mem_type == MEM_MMAP) {
+ td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
+ if (td->orig_buffer == MAP_FAILED) {
+ td_verror(td, errno);
+ perror("mmap");
+ td->orig_buffer = NULL;
return 1;
}
}
return 0;
}
-static int create_file(struct thread_data *td)
+static int create_file(struct thread_data *td, unsigned long long size,
+ int extend)
{
unsigned long long left;
unsigned int bs;
+ int r, oflags;
char *b;
- int r;
/*
* unless specifically asked for overwrite, let normal io extend it
if (td_write(td) && !td->overwrite)
return 0;
- if (!td->file_size) {
+ if (!size) {
fprintf(stderr, "Need size for create\n");
- td->error = EINVAL;
+ td_verror(td, EINVAL);
return 1;
}
- printf("Client%d: Laying out IO file\n", td->thread_number);
+ if (!extend) {
+ oflags = O_CREAT | O_TRUNC;
+ printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
+ } else {
+ oflags = O_APPEND;
+ printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
+ }
- td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
+ td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
if (td->fd < 0) {
- td->error = errno;
+ td_verror(td, errno);
return 1;
}
- if (ftruncate(td->fd, td->file_size) == -1) {
- td->error = errno;
+ if (!extend && ftruncate(td->fd, td->file_size) == -1) {
+ td_verror(td, errno);
return 1;
}
b = malloc(td->max_bs);
memset(b, 0, td->max_bs);
- left = td->file_size;
- while (left) {
+ left = size;
+ while (left && !td->terminate) {
bs = td->max_bs;
if (bs > left)
bs = left;
continue;
} else {
if (r < 0)
- td->error = errno;
+ td_verror(td, errno);
else
- td->error = EIO;
+ td_verror(td, EIO);
break;
}
}
- if (td->create_fsync)
+ if (td->terminate)
+ unlink(td->file_name);
+ else if (td->create_fsync)
fsync(td->fd);
close(td->fd);
return 0;
}
-static int file_exists(struct thread_data *td)
+static int file_size(struct thread_data *td)
{
struct stat st;
- if (stat(td->file_name, &st) != -1)
+ if (fstat(td->fd, &st) == -1) {
+ td_verror(td, errno);
return 1;
+ }
+
+ if (!td->file_size)
+ td->file_size = st.st_size;
- return errno != ENOENT;
+ return 0;
}
-static int get_file_size(struct thread_data *td)
+static int bdev_size(struct thread_data *td)
{
- size_t bytes = 0;
- struct stat st;
+ size_t bytes;
- if (fstat(td->fd, &st) == -1) {
- td->error = errno;
+ if (ioctl(td->fd, BLKGETSIZE64, &bytes) < 0) {
+ td_verror(td, 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
+ * no extend possibilities, so limit size to device size if too large
*/
- 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->file_size || td->file_size > bytes)
+ td->file_size = bytes;
- if (td_read(td)) {
- if (td->file_size > bytes)
- bytes = td->file_size;
- } else {
- if (!td->file_size)
- td->file_size = 1024 * 1024 * 1024;
+ return 0;
+}
- bytes = td->file_size;
- }
+static int get_file_size(struct thread_data *td)
+{
+ int ret;
- if (td->file_offset > bytes) {
- fprintf(stderr, "Client%d: offset larger than length\n", td->thread_number);
+ if (td->filetype == FIO_TYPE_FILE)
+ ret = file_size(td);
+ else
+ ret = bdev_size(td);
+
+ if (ret)
+ return ret;
+
+ if (td->file_offset > td->file_size) {
+ fprintf(stderr, "Client%d: offset larger than length (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->file_size);
return 1;
}
- td->io_size = bytes - td->file_offset;
+ td->io_size = td->file_size - td->file_offset;
if (td->io_size == 0) {
fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
- td->error = EINVAL;
+ td_verror(td, EINVAL);
return 1;
}
return 0;
}
+static int setup_file_mmap(struct thread_data *td)
+{
+ int flags;
+
+ if (td_read(td))
+ flags = PROT_READ;
+ else {
+ flags = PROT_WRITE;
+
+ if (td->verify != VERIFY_NONE)
+ flags |= PROT_READ;
+ }
+
+ td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
+ if (td->mmap == MAP_FAILED) {
+ td->mmap = NULL;
+ td_verror(td, errno);
+ return 1;
+ }
+
+ if (td->invalidate_cache) {
+ if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ }
+
+ if (td->sequential) {
+ if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ } else {
+ if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
+static int setup_file_plain(struct thread_data *td)
+{
+ if (td->invalidate_cache) {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ }
+
+ if (td->sequential) {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ } else {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
static int setup_file(struct thread_data *td)
{
+ struct stat st;
int flags = 0;
- if (!file_exists(td)) {
+ if (stat(td->file_name, &st) == -1) {
+ if (errno != ENOENT) {
+ td_verror(td, errno);
+ return 1;
+ }
if (!td->create_file) {
- td->error = ENOENT;
+ td_verror(td, ENOENT);
return 1;
}
- if (create_file(td))
+ if (create_file(td, td->file_size, 0))
return 1;
+ } else if (td->filetype == FIO_TYPE_FILE) {
+ if (st.st_size < td->file_size) {
+ if (create_file(td, td->file_size - st.st_size, 1))
+ return 1;
+ }
}
if (td->odirect)
if (td_read(td))
td->fd = open(td->file_name, flags | O_RDONLY);
else {
- if (!td->overwrite)
- flags |= O_TRUNC;
+ if (td->filetype == FIO_TYPE_FILE) {
+ if (!td->overwrite)
+ flags |= O_TRUNC;
+
+ flags |= O_CREAT;
+ }
if (td->sync_io)
flags |= O_SYNC;
- if (td->verify)
- flags |= O_RDWR;
- else
- flags |= O_WRONLY;
- td->fd = open(td->file_name, flags | O_CREAT, 0600);
+ flags |= O_RDWR;
+
+ td->fd = open(td->file_name, flags, 0600);
}
if (td->fd == -1) {
- td->error = errno;
+ td_verror(td, errno);
return 1;
}
if (get_file_size(td))
return 1;
- if (td_write(td) && ftruncate(td->fd, td->file_size) == -1) {
- td->error = errno;
+ if (!td->use_mmap)
+ return setup_file_plain(td);
+ else
+ return setup_file_mmap(td);
+}
+
+static int check_dev_match(dev_t dev, char *path)
+{
+ unsigned int major, minor;
+ char line[256], *p;
+ FILE *f;
+
+ f = fopen(path, "r");
+ if (!f) {
+ perror("open path");
return 1;
}
- if (td->invalidate_cache) {
- if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
- td->error = errno;
- return 1;
+ p = fgets(line, sizeof(line), f);
+ if (!p) {
+ fclose(f);
+ return 1;
+ }
+
+ if (sscanf(p, "%u:%u", &major, &minor) != 2) {
+ fclose(f);
+ return 1;
+ }
+
+ if (((major << 8) | minor) == dev) {
+ fclose(f);
+ return 0;
+ }
+
+ fclose(f);
+ return 1;
+}
+
+static char *find_block_dir(dev_t dev, char *path)
+{
+ struct dirent *dir;
+ char *found = NULL;
+ struct stat st;
+ DIR *D;
+
+ D = opendir(path);
+ if (!D)
+ return NULL;
+
+ while ((dir = readdir(D)) != NULL) {
+ char full_path[256];
+
+ if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
+ continue;
+ if (!strcmp(dir->d_name, "device"))
+ continue;
+
+ sprintf(full_path, "%s/%s", path, dir->d_name);
+
+ if (!strcmp(dir->d_name, "dev")) {
+ if (!check_dev_match(dev, full_path)) {
+ found = path;
+ break;
+ }
}
+
+ if (stat(full_path, &st) == -1) {
+ perror("stat");
+ break;
+ }
+
+ if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
+ continue;
+
+ if ((found = find_block_dir(dev, full_path)) != NULL)
+ break;
}
+ closedir(D);
+ return found;
+}
+
+static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
+{
+ unsigned in_flight;
+ char line[256];
+ FILE *f;
+ char *p;
+
+ f = fopen(du->path, "r");
+ if (!f)
+ return 1;
+
+ p = fgets(line, sizeof(line), f);
+ if (!p) {
+ fclose(f);
+ return 1;
+ }
+
+ if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
+ fclose(f);
+ return 1;
+ }
+
+ fclose(f);
return 0;
}
+static void update_io_tick_disk(struct disk_util *du)
+{
+ struct disk_util_stat __dus, *dus, *ldus;
+ struct timeval t;
+
+ if (get_io_ticks(du, &__dus))
+ return;
+
+ dus = &du->dus;
+ ldus = &du->last_dus;
+
+ dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
+ dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
+ dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
+ dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
+ dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
+ dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
+ dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
+ dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
+ dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
+ dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
+
+ gettimeofday(&t, NULL);
+ du->msec += mtime_since(&du->time, &t);
+ memcpy(&du->time, &t, sizeof(t));
+ memcpy(ldus, &__dus, sizeof(__dus));
+}
+
+static void update_io_ticks(void)
+{
+ struct list_head *entry;
+ struct disk_util *du;
+
+ list_for_each(entry, &disk_list) {
+ du = list_entry(entry, struct disk_util, list);
+ update_io_tick_disk(du);
+ }
+}
+
+static int disk_util_exists(dev_t dev)
+{
+ struct list_head *entry;
+ struct disk_util *du;
+
+ list_for_each(entry, &disk_list) {
+ du = list_entry(entry, struct disk_util, list);
+
+ if (du->dev == dev)
+ return 1;
+ }
+
+ return 0;
+}
+
+static void disk_util_add(dev_t dev, char *path)
+{
+ struct disk_util *du = malloc(sizeof(*du));
+
+ memset(du, 0, sizeof(*du));
+ INIT_LIST_HEAD(&du->list);
+ sprintf(du->path, "%s/stat", path);
+ du->name = strdup(basename(path));
+ du->dev = dev;
+
+ gettimeofday(&du->time, NULL);
+ get_io_ticks(du, &du->last_dus);
+
+ list_add_tail(&du->list, &disk_list);
+}
+
+static void init_disk_util(struct thread_data *td)
+{
+ struct stat st;
+ char foo[256], tmp[256];
+ dev_t dev;
+ char *p, *dir;
+
+ if (!td->do_disk_util)
+ return;
+
+ if (!stat(td->file_name, &st)) {
+ if (S_ISBLK(st.st_mode))
+ dev = st.st_rdev;
+ else
+ dev = st.st_dev;
+ } else {
+ /*
+ * must be a file, open "." in that path
+ */
+ strcpy(foo, td->file_name);
+ p = dirname(foo);
+ if (stat(p, &st)) {
+ perror("disk util stat");
+ return;
+ }
+
+ dev = st.st_dev;
+ }
+
+ if (disk_util_exists(dev))
+ return;
+
+ sprintf(foo, "/sys/block");
+ dir = find_block_dir(dev, foo);
+ if (!dir)
+ return;
+
+ /*
+ * if this is inside a partition dir, jump back to parent
+ */
+ sprintf(tmp, "%s/queue", dir);
+ if (stat(tmp, &st)) {
+ p = dirname(dir);
+ sprintf(tmp, "%s/queue", p);
+ if (stat(tmp, &st)) {
+ fprintf(stderr, "unknown sysfs layout\n");
+ return;
+ }
+ sprintf(foo, "%s", p);
+ }
+
+ disk_util_add(dev, foo);
+}
+
+static void disk_util_timer_arm(void)
+{
+ struct itimerval itimer;
+
+ itimer.it_value.tv_sec = 0;
+ itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
+ setitimer(ITIMER_REAL, &itimer, NULL);
+}
+
static void clear_io_state(struct thread_data *td)
{
- if (!td->use_aio)
+ if (td->io_engine == FIO_SYNCIO)
lseek(td->fd, SEEK_SET, 0);
td->cur_off = 0;
if (init_io_u(td))
goto err;
- if (sched_setaffinity(td->pid, sizeof(td->cpumask), &td->cpumask) == -1) {
- td->error = errno;
+ if (fio_setaffinity(td) == -1) {
+ td_verror(td, errno);
goto err;
}
- if (td->use_aio && init_aio(td))
+ if ((td->io_engine != FIO_SYNCIO) && init_aio(td))
goto err;
if (td->ioprio) {
if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
- td->error = errno;
+ td_verror(td, errno);
goto err;
}
}
if (init_random_state(td))
goto err;
+ gettimeofday(&td->epoch, NULL);
+
while (td->loops--) {
getrusage(RUSAGE_SELF, &td->ru_start);
gettimeofday(&td->start, NULL);
clear_io_state(td);
prune_io_piece_log(td);
- if (!td->use_aio)
+ if (td->io_engine == FIO_SYNCIO)
do_sync_io(td);
else
do_async_io(td);
- if (td->error)
+ if (td->error || td->terminate)
break;
td->runtime += mtime_since_now(&td->start);
update_rusage_stat(td);
- if (!td->verify)
+ if (td->verify == VERIFY_NONE)
continue;
clear_io_state(td);
- if (!td->use_aio)
+ if (td->io_engine == FIO_SYNCIO)
do_sync_verify(td);
else
do_async_verify(td);
- if (td->error)
+ if (td->error || td->terminate)
break;
}
close(td->fd);
td->fd = -1;
}
- if (td->use_aio)
+ if (td->mmap)
+ munmap(td->mmap, td->file_size);
+ if (td->io_engine != FIO_SYNCIO)
cleanup_aio(td);
cleanup_io_u(td);
if (ret) {
}
-static void *fork_main(int shm_id, int offset)
+static void *fork_main(int shmid, int offset)
{
struct thread_data *td;
void *data;
- data = shmat(shm_id, NULL, 0);
+ data = shmat(shmid, NULL, 0);
if (data == (void *) -1) {
perror("shmat");
return NULL;
return 1;
}
-static void show_thread_status(struct thread_data *td)
+static void show_thread_status(struct thread_data *td,
+ struct group_run_stats *rs)
{
int prio, prio_class;
unsigned long min, max, bw = 0;
printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
if (calc_lat(&td->clat_stat, &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 (calc_lat(&td->bw_stat, &min, &max, &mean, &dev)) {
+ double p_of_agg;
+
+ p_of_agg = mean * 100 / (double) rs->agg[td->ddir];
+ printf(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
+ }
if (td->runtime) {
usr_cpu = (double) td->usr_time * 100 / (double) td->runtime;
printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
}
-static void print_thread_status(struct thread_data *td, int nr_running,
- int t_rate, int m_rate)
+static void print_thread_status(int nr_running, int t_rate, int m_rate)
{
printf("Threads now running: %d", nr_running);
if (m_rate || t_rate)
}
run_str[td->thread_number - 1] = c;
- print_thread_status(td, n, t, m);
+ print_thread_status(n, t, m);
td->old_runstate = td->runstate;
}
}
}
-static void run_threads(char *argv[])
+static void run_threads(void)
{
struct timeval genesis;
struct thread_data *td;
fflush(stdout);
signal(SIGINT, sig_handler);
+ signal(SIGALRM, sig_handler);
todo = thread_number;
nr_running = 0;
for (i = 0; i < thread_number; i++) {
td = &threads[i];
+ init_disk_util(td);
+
if (!td->create_serialize)
continue;
* start created threads (TD_CREATED -> TD_RUNNING)
*/
for (i = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
+ td = &threads[i];
if (td->runstate != TD_CREATED)
continue;
}
for (i = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
+ td = &threads[i];
if (td->runstate != TD_RUNNING &&
td->runstate != TD_VERIFYING)
reap_threads(&nr_running, &t_rate, &m_rate);
usleep(10000);
}
-}
-struct group_run_stats {
- unsigned long max_run[2], min_run[2];
- unsigned long max_bw[2], min_bw[2];
- unsigned long io_mb[2];
- unsigned long agg[2];
-};
+ update_io_ticks();
+}
static void show_group_stats(struct group_run_stats *rs, int id)
{
printf(" WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
}
+static void show_disk_util(void)
+{
+ struct disk_util_stat *dus;
+ struct list_head *entry;
+ struct disk_util *du;
+ double util;
+
+ printf("\nDisk stats (read/write):\n");
+
+ list_for_each(entry, &disk_list) {
+ du = list_entry(entry, struct disk_util, list);
+ dus = &du->dus;
+
+ util = (double) 100 * du->dus.io_ticks / (double) du->msec;
+ if (util > 100.0)
+ util = 100.0;
+
+ printf(" %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
+ }
+}
+
static void show_run_stats(void)
{
struct group_run_stats *runstats, *rs;
+ struct thread_data *td;
int i;
runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
}
for (i = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
unsigned long bw = 0;
- rs = &runstats[td->groupid];
+ td = &threads[i];
- if (!td->error) {
- if (td->runtime < rs->min_run[td->ddir])
- rs->min_run[td->ddir] = td->runtime;
- if (td->runtime > rs->max_run[td->ddir])
- rs->max_run[td->ddir] = td->runtime;
+ if (td->error) {
+ printf("Client%d: %s\n", td->thread_number, td->verror);
+ continue;
+ }
- if (td->runtime)
- bw = td->io_bytes / td->runtime;
- if (bw < rs->min_bw[td->ddir])
- rs->min_bw[td->ddir] = bw;
- if (bw > rs->max_bw[td->ddir])
- rs->max_bw[td->ddir] = bw;
+ rs = &runstats[td->groupid];
- rs->io_mb[td->ddir] += td->io_bytes >> 20;
- }
+ if (td->runtime < rs->min_run[td->ddir])
+ rs->min_run[td->ddir] = td->runtime;
+ if (td->runtime > rs->max_run[td->ddir])
+ rs->max_run[td->ddir] = td->runtime;
+
+ if (td->runtime)
+ bw = td->io_bytes / td->runtime;
+ if (bw < rs->min_bw[td->ddir])
+ rs->min_bw[td->ddir] = bw;
+ if (bw > rs->max_bw[td->ddir])
+ rs->max_bw[td->ddir] = bw;
- show_thread_status(td);
+ rs->io_mb[td->ddir] += td->io_bytes >> 20;
}
-
+
for (i = 0; i < groupid + 1; i++) {
rs = &runstats[i];
rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
if (rs->max_run[1])
rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
+ }
+
+ for (i = 0; i < thread_number; i++) {
+ td = &threads[i];
+ rs = &runstats[td->groupid];
- show_group_stats(rs, i);
+ show_thread_status(td, rs);
}
+
+ for (i = 0; i < groupid + 1; i++)
+ show_group_stats(&runstats[i], i);
+
+ show_disk_util();
}
int main(int argc, char *argv[])
return 1;
}
- run_threads(argv);
+ disk_util_timer_arm();
+
+ run_threads();
show_run_stats();
return 0;
projects / disktools.git / blobdiff