* fio - the flexible io tester
*
* 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
static char run_str[MAX_JOBS + 1];
int shm_id = 0;
static struct timeval genesis;
+static int temp_stall_ts;
static void print_thread_status(void);
extern unsigned long long mlock_size;
/*
- * thread life cycle
+ * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
+ * will never back again. It may cycle between running/verififying/fsyncing.
+ * Once the thread reaches TD_EXITED, it is just waiting for the core to
+ * reap it.
*/
enum {
TD_NOT_CREATED = 0,
}
}
+/*
+ * The ->file_map[] contains a map of blocks we have or have not done io
+ * to yet. Used to make sure we cover the entire range in a fair fashion.
+ */
static int random_map_free(struct thread_data *td, unsigned long long block)
{
unsigned int idx = RAND_MAP_IDX(td, block);
return (td->file_map[idx] & (1UL << bit)) == 0;
}
+/*
+ * Return the next free block in the map.
+ */
static int get_next_free_block(struct thread_data *td, unsigned long long *b)
{
int i;
return 1;
}
+/*
+ * Mark a given offset as used in the map.
+ */
static void mark_random_map(struct thread_data *td, struct io_u *io_u)
{
unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
io_u->buflen = blocks * td->min_bs;
}
+/*
+ * For random io, generate a random new block and see if it's used. Repeat
+ * until we find a free one. For sequential io, just return the end of
+ * the last io issued.
+ */
static int get_next_offset(struct thread_data *td, unsigned long long *offset)
{
unsigned long long b, rb;
return buflen;
}
+/*
+ * Check if we are above the minimum rate given.
+ */
static int check_min_rate(struct thread_data *td, struct timeval *now)
{
unsigned long spent;
memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
}
+/*
+ * Return the data direction for the next io_u. If the job is a
+ * mixed read/write workload, check the rwmix cycle and switch if
+ * necessary.
+ */
static int get_rw_ddir(struct thread_data *td)
{
if (td_rw(td)) {
static int td_io_prep(struct thread_data *td, struct io_u *io_u)
{
- if (td->io_prep && td->io_prep(td, io_u))
+ if (td->io_ops->prep && td->io_ops->prep(td, io_u))
return 1;
return 0;
return io_u;
}
+/*
+ * Return an io_u to be processed. Gets a buflen and offset, sets direction,
+ * etc. The returned io_u is fully ready to be prepped and submitted.
+ */
static struct io_u *get_io_u(struct thread_data *td)
{
struct io_u *io_u;
static int sync_td(struct thread_data *td)
{
- if (td->io_sync)
- return td->io_sync(td);
+ if (td->io_ops->sync)
+ return td->io_ops->sync(td);
return 0;
}
static int io_u_getevents(struct thread_data *td, int min, int max,
struct timespec *t)
{
- return td->io_getevents(td, min, max, t);
+ return td->io_ops->getevents(td, min, max, t);
}
static int io_u_queue(struct thread_data *td, struct io_u *io_u)
{
gettimeofday(&io_u->issue_time, NULL);
- return td->io_queue(td, io_u);
+ return td->io_ops->queue(td, io_u);
}
#define iocb_time(iocb) ((unsigned long) (iocb)->data)
icd->bytes_done[0] = icd->bytes_done[1] = 0;
for (i = 0; i < icd->nr; i++) {
- io_u = td->io_event(td, i);
+ io_u = td->io_ops->event(td, i);
io_completed(td, io_u, icd);
put_io_u(td, io_u);
}
}
+/*
+ * When job exits, we can cancel the in-flight IO if we are using async
+ * io. Attempt to do so.
+ */
static void cleanup_pending_aio(struct thread_data *td)
{
struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
/*
* now cancel remaining active events
*/
- if (td->io_cancel) {
+ if (td->io_ops->cancel) {
list_for_each_safe(entry, n, &td->io_u_busylist) {
io_u = list_entry(entry, struct io_u, list);
- r = td->io_cancel(td, io_u);
+ r = td->io_ops->cancel(td, io_u);
if (!r)
put_io_u(td, io_u);
}
return ret;
}
+/*
+ * The main verify engine. Runs over the writes we previusly submitted,
+ * reads the blocks back in, and checks the crc/md5 of the data.
+ */
static void do_verify(struct thread_data *td)
{
struct timeval t;
break;
}
- v_io_u = td->io_event(td, 0);
+ v_io_u = td->io_ops->event(td, 0);
icd.nr = 1;
icd.error = 0;
io_completed(td, v_io_u, &icd);
}
/*
- * Main IO worker functions. It retrieves io_u's to process and queues
+ * Not really an io thread, all it does is burn CPU cycles in the specified
+ * manner.
+ */
+static void do_cpuio(struct thread_data *td)
+{
+ struct timeval e;
+ int split = 100 / td->cpuload;
+ int i = 0;
+
+ while (!td->terminate) {
+ gettimeofday(&e, NULL);
+
+ if (runtime_exceeded(td, &e))
+ break;
+
+ if (!(i % split))
+ __usec_sleep(10000);
+ else
+ usec_sleep(td, 10000);
+
+ i++;
+ }
+}
+
+/*
+ * Main IO worker function. It retrieves io_u's to process and queues
* and reaps them, checking for rate and errors along the way.
*/
static void do_io(struct thread_data *td)
}
}
-static void cleanup_io(struct thread_data *td)
-{
- if (td->io_cleanup)
- td->io_cleanup(td);
-}
-
static int init_io(struct thread_data *td)
{
- if (td->io_engine == FIO_SYNCIO)
- return fio_syncio_init(td);
- else if (td->io_engine == FIO_MMAPIO)
- return fio_mmapio_init(td);
- else if (td->io_engine == FIO_LIBAIO)
- return fio_libaio_init(td);
- else if (td->io_engine == FIO_POSIXAIO)
- return fio_posixaio_init(td);
- else if (td->io_engine == FIO_SGIO)
- return fio_sgio_init(td);
- else if (td->io_engine == FIO_SPLICEIO)
- return fio_spliceio_init(td);
- else {
- log_err("bad io_engine %d\n", td->io_engine);
- return 1;
- }
+ if (td->io_ops->init)
+ return td->io_ops->init(td);
+
+ return 0;
}
static void cleanup_io_u(struct thread_data *td)
int i, max_units;
char *p;
- if (td->io_engine & FIO_SYNCIO)
+ if (td->io_ops->flags & FIO_CPUIO)
+ return 0;
+
+ if (td->io_ops->flags & FIO_SYNCIO)
max_units = 1;
else
max_units = td->iodepth;
return 0;
}
-static int create_file(struct thread_data *td, unsigned long long size,
- int extend)
+static int create_file(struct thread_data *td, unsigned long long size)
{
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)
+ if (!td->overwrite) {
+ td->real_file_size = size;
return 0;
+ }
if (!size) {
log_err("Need size for create\n");
return 1;
}
- if (!extend) {
- oflags = O_CREAT | O_TRUNC;
- fprintf(f_out, "%s: Laying out IO file (%LuMiB)\n", td->name, size >> 20);
- } else {
- oflags = O_APPEND;
- fprintf(f_out, "%s: Extending IO file (%Lu -> %LuMiB)\n", td->name, (td->file_size - size) >> 20, td->file_size >> 20);
- }
+ temp_stall_ts = 1;
+ fprintf(f_out, "%s: Laying out IO file (%LuMiB)\n",td->name,size >> 20);
- td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
+ td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
if (td->fd < 0) {
td_verror(td, errno);
- return 1;
+ goto done_noclose;
}
- if (!extend && ftruncate(td->fd, td->file_size) == -1) {
+ if (ftruncate(td->fd, td->file_size) == -1) {
td_verror(td, errno);
- return 1;
+ goto done;
}
td->io_size = td->file_size;
else if (td->create_fsync)
fsync(td->fd);
+ free(b);
+done:
close(td->fd);
td->fd = -1;
- free(b);
+done_noclose:
+ temp_stall_ts = 0;
return 0;
}
{
struct stat st;
- if (fstat(td->fd, &st) == -1) {
- td_verror(td, errno);
- return 1;
- }
+ if (td->overwrite) {
+ if (fstat(td->fd, &st) == -1) {
+ td_verror(td, errno);
+ return 1;
+ }
- td->real_file_size = st.st_size;
+ td->real_file_size = st.st_size;
- if (!td->file_size || td->file_size > td->real_file_size)
- td->file_size = td->real_file_size;
+ if (!td->file_size || td->file_size > td->real_file_size)
+ td->file_size = td->real_file_size;
+ }
td->file_size -= td->file_offset;
return 0;
struct stat st;
int flags = 0;
+ if (td->io_ops->flags & FIO_CPUIO)
+ return 0;
+
if (stat(td->file_name, &st) == -1) {
if (errno != ENOENT) {
td_verror(td, errno);
td_verror(td, ENOENT);
return 1;
}
- if (create_file(td, td->file_size, 0))
+ if (create_file(td, td->file_size))
+ return 1;
+ } else if (td->filetype == FIO_TYPE_FILE &&
+ st.st_size < (off_t) td->file_size) {
+ if (create_file(td, td->file_size))
return 1;
- } else if (td->filetype == FIO_TYPE_FILE) {
- if (st.st_size < (off_t) td->file_size) {
- if (create_file(td, td->file_size - st.st_size, 1))
- return 1;
- }
}
if (td->odirect)
if (get_file_size(td))
return 1;
- if (td->io_engine != FIO_MMAPIO)
- return setup_file_plain(td);
- else
+ if (td->io_ops->flags & FIO_MMAPIO)
return setup_file_mmap(td);
+ else
+ return setup_file_plain(td);
}
static int switch_ioscheduler(struct thread_data *td)
static void clear_io_state(struct thread_data *td)
{
- if (td->io_engine == FIO_SYNCIO)
+ if (td->io_ops->flags & FIO_SYNCIO)
lseek(td->fd, SEEK_SET, 0);
td->last_pos = 0;
memset(td->file_map, 0, td->num_maps * sizeof(long));
}
+/*
+ * Entry point for the thread based jobs. The process based jobs end up
+ * here as well, after a little setup.
+ */
static void *thread_main(void *data)
{
struct thread_data *td = data;
}
}
- if (nice(td->nice) < 0) {
+ if (nice(td->nice) == -1) {
td_verror(td, errno);
goto err;
}
clear_io_state(td);
prune_io_piece_log(td);
- do_io(td);
+ if (td->io_ops->flags & FIO_CPUIO)
+ do_cpuio(td);
+ else
+ do_io(td);
td->runtime[td->ddir] += mtime_since_now(&td->start);
if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
}
if (td->mmap)
munmap(td->mmap, td->file_size);
- cleanup_io(td);
+ close_ioengine(td);
cleanup_io_u(td);
td_set_runstate(td, TD_EXITED);
return NULL;
}
+/*
+ * We cannot pass the td data into a forked process, so attach the td and
+ * pass it to the thread worker.
+ */
static void *fork_main(int shmid, int offset)
{
struct thread_data *td;
return NULL;
}
+/*
+ * Sets the status of the 'td' in the printed status map.
+ */
static void check_str_update(struct thread_data *td)
{
char c = run_str[td->thread_number - 1];
run_str[td->thread_number - 1] = c;
}
+/*
+ * Convert seconds to a printable string.
+ */
static void eta_to_str(char *str, int eta_sec)
{
unsigned int d, h, m, s;
str += sprintf(str, "%02ds", s);
}
+/*
+ * Best effort calculation of the estimated pending runtime of a job.
+ */
static int thread_eta(struct thread_data *td, unsigned long elapsed)
{
unsigned long long bytes_total, bytes_done;
else if (t_eta)
eta_sec = t_eta;
else
- eta_sec = INT_MAX;
+ eta_sec = 0;
} else {
/*
* thread is already done or waiting for fsync
return eta_sec;
}
+/*
+ * Print status of the jobs we know about. This includes rate estimates,
+ * ETA, thread state, etc.
+ */
static void print_thread_status(void)
{
unsigned long elapsed = time_since_now(&genesis);
char eta_str[32];
double perc = 0.0;
+ if (temp_stall_ts || terse_output)
+ return;
+
eta_secs = malloc(thread_number * sizeof(int));
memset(eta_secs, 0, thread_number * sizeof(int));
printf("Threads running: %d", nr_running);
if (m_rate || t_rate)
printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
- if (eta_sec != INT_MAX) {
+ if (eta_sec != INT_MAX && nr_running) {
perc *= 100.0;
printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
}
free(eta_secs);
}
+/*
+ * Run over the job map and reap the threads that have exited, if any.
+ */
static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
{
- int i;
+ int i, cputhreads;
/*
* reap exited threads (TD_EXITED -> TD_REAPED)
*/
- for (i = 0; i < thread_number; i++) {
+ for (i = 0, cputhreads = 0; i < thread_number; i++) {
struct thread_data *td = &threads[i];
+ if (td->io_ops->flags & FIO_CPUIO)
+ cputhreads++;
+
if (td->runstate != TD_EXITED)
continue;
(*m_rate) -= td->ratemin;
(*t_rate) -= td->rate;
}
+
+ if (*nr_running == cputhreads)
+ terminate_threads(TERMINATE_ALL);
}
static void fio_unpin_memory(void *pinned)
return ptr;
}
+/*
+ * Main function for kicking off and reaping jobs, as needed.
+ */
static void run_threads(void)
{
struct thread_data *td;
mlocked_mem = fio_pin_memory();
- printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
- fflush(stdout);
+ if (!terse_output) {
+ printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
+ fflush(stdout);
+ }
signal(SIGINT, sig_handler);
signal(SIGALRM, sig_handler);
* Wait for the started threads to transition to
* TD_INITIALIZED.
*/
- printf("fio: Waiting for threads to initialize...\n");
gettimeofday(&this_start, NULL);
left = this_jobs;
while (left) {
/*
* start created threads (TD_INITIALIZED -> TD_RUNNING).
*/
- printf("fio: Go for launch\n");
for (i = 0; i < thread_number; i++) {
td = &threads[i];