#include <sys/mman.h>
#include "fio.h"
-#include "os.h"
+#include "hash.h"
+#include "smalloc.h"
unsigned long page_mask;
unsigned long page_size;
int nr_thread = 0;
int shm_id = 0;
int temp_stall_ts;
+unsigned long done_secs = 0;
-static struct fio_sem *startup_sem;
+static struct fio_mutex *startup_mutex;
static volatile int fio_abort;
static int exit_value;
static inline void td_set_runstate(struct thread_data *td, int runstate)
{
+ if (td->runstate == runstate)
+ return;
+
+ dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
+ td->runstate, runstate);
td->runstate = runstate;
}
struct thread_data *td;
int i;
+ dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
+
for_each_td(td, i) {
if (group_id == TERMINATE_ALL || groupid == td->groupid) {
+ dprint(FD_PROCESS, "setting terminate on %s/%d\n",
+ td->o.name, (int) td->pid);
+ td->terminate = 1;
+ td->o.start_delay = 0;
+
/*
* if the thread is running, just let it exit
*/
if (td->runstate < TD_RUNNING)
kill(td->pid, SIGQUIT);
- td->terminate = 1;
- td->o.start_delay = 0;
+ else {
+ struct ioengine_ops *ops = td->io_ops;
+
+ if (ops && (ops->flags & FIO_SIGQUIT))
+ kill(td->pid, SIGQUIT);
+ }
}
}
}
static void sig_handler(int sig)
{
+ if (!threads)
+ return;
+
switch (sig) {
- case SIGALRM:
- update_io_ticks();
- disk_util_timer_arm();
- print_thread_status();
- break;
- default:
- printf("\nfio: terminating on signal %d\n", sig);
- fflush(stdout);
- terminate_threads(TERMINATE_ALL);
- break;
+ case SIGALRM:
+ update_io_ticks();
+ disk_util_timer_arm();
+ print_thread_status();
+ break;
+ default:
+ printf("\nfio: terminating on signal %d\n", sig);
+ fflush(stdout);
+ terminate_threads(TERMINATE_ALL);
+ break;
}
}
* check bandwidth specified rate
*/
if (bytes < td->rate_bytes) {
- log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
+ log_err("%s: min rate %u not met\n", td->o.name,
+ td->o.ratemin);
return 1;
} else {
rate = (bytes - td->rate_bytes) / spent;
- if (rate < td->o.ratemin || bytes < td->rate_bytes) {
- log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
+ if (rate < td->o.ratemin ||
+ bytes < td->rate_bytes) {
+ log_err("%s: min rate %u not met, got"
+ " %luKiB/sec\n", td->o.name,
+ td->o.ratemin, rate);
return 1;
}
}
* checks iops specified rate
*/
if (iops < td->o.rate_iops) {
- log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
+ log_err("%s: min iops rate %u not met\n",
+ td->o.name, td->o.rate_iops);
return 1;
} else {
rate = (iops - td->rate_blocks) / spent;
- if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
- log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
+ if (rate < td->o.rate_iops_min ||
+ iops < td->rate_blocks) {
+ log_err("%s: min iops rate %u not met,"
+ " got %lu\n", td->o.name,
+ td->o.rate_iops_min,
+ rate);
}
}
}
if (runtime_exceeded(td, &io_u->start_time)) {
put_io_u(td, io_u);
+ td->terminate = 1;
break;
}
ret = -io_u->error;
else if (io_u->resid) {
int bytes = io_u->xfer_buflen - io_u->resid;
+ struct fio_file *f = io_u->file;
/*
* zero read, fail
put_io_u(td, io_u);
break;
}
+
io_u->xfer_buflen = io_u->resid;
io_u->xfer_buf += bytes;
+ io_u->offset += bytes;
+
+ td->ts.short_io_u[io_u->ddir]++;
+
+ if (io_u->offset == f->real_file_size)
+ goto sync_done;
+
requeue_io_u(td, &io_u);
} else {
+sync_done:
ret = io_u_sync_complete(td, io_u);
if (ret < 0)
break;
*/
min_events = 0;
if (queue_full(td) || ret == FIO_Q_BUSY) {
- min_events = 1;
-
- if (td->cur_depth > td->o.iodepth_low)
+ if (td->cur_depth >= td->o.iodepth_low)
min_events = td->cur_depth - td->o.iodepth_low;
+ if (!min_events)
+ min_events = 1;
}
/*
td_set_runstate(td, TD_RUNNING);
- while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
+ while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
struct timeval comp_time;
long bytes_done = 0;
int min_evts = 0;
if (runtime_exceeded(td, &s)) {
put_io_u(td, io_u);
+ td->terminate = 1;
break;
}
+ /*
+ * Add verification end_io handler, if asked to verify
+ * a previously written file.
+ */
+ if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
+ io_u->end_io = verify_io_u;
+ td_set_runstate(td, TD_VERIFYING);
+ } else
+ td_set_runstate(td, TD_RUNNING);
+
ret = td_io_queue(td, io_u);
switch (ret) {
case FIO_Q_COMPLETED:
ret = -io_u->error;
else if (io_u->resid) {
int bytes = io_u->xfer_buflen - io_u->resid;
+ struct fio_file *f = io_u->file;
/*
* zero read, fail
io_u->xfer_buflen = io_u->resid;
io_u->xfer_buf += bytes;
+ io_u->offset += bytes;
+
+ td->ts.short_io_u[io_u->ddir]++;
+
+ if (io_u->offset == f->real_file_size)
+ goto sync_done;
+
requeue_io_u(td, &io_u);
} else {
+sync_done:
fio_gettime(&comp_time, NULL);
bytes_done = io_u_sync_complete(td, io_u);
if (bytes_done < 0)
/*
* See if we need to complete some commands
*/
- if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
+ if (queue_full(td) || ret == FIO_Q_BUSY) {
min_evts = 0;
- if (queue_full(td) || ret == FIO_Q_BUSY) {
+ if (td->cur_depth >= td->o.iodepth_low)
+ min_evts = td->cur_depth - td->o.iodepth_low;
+ if (!min_evts)
min_evts = 1;
-
- if (td->cur_depth > td->o.iodepth_low)
- min_evts = td->cur_depth - td->o.iodepth_low;
- }
-
fio_gettime(&comp_time, NULL);
bytes_done = io_u_queued_complete(td, min_evts);
if (bytes_done < 0)
}
}
+ if (td->o.fill_device && td->error == ENOSPC) {
+ td->error = 0;
+ td->terminate = 1;
+ }
if (!td->error) {
struct fio_file *f;
}
} else
cleanup_pending_aio(td);
+
+ /*
+ * stop job if we failed doing any IO
+ */
+ if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
+ td->done = 1;
}
static void cleanup_io_u(struct thread_data *td)
free_io_mem(td);
}
-/*
- * "randomly" fill the buffer contents
- */
-static void fill_rand_buf(struct io_u *io_u, int max_bs)
-{
- int *ptr = io_u->buf;
-
- while ((void *) ptr - io_u->buf < max_bs) {
- *ptr = rand() * 0x9e370001;
- ptr++;
- }
-}
-
static int init_io_u(struct thread_data *td)
{
- unsigned long long buf_size;
struct io_u *io_u;
unsigned int max_bs;
int i, max_units;
char *p;
- if (td->io_ops->flags & FIO_SYNCIO)
- max_units = 1;
- else
- max_units = td->o.iodepth;
-
+ max_units = td->o.iodepth;
max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
- buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
- buf_size += page_mask;
- if (buf_size != (size_t) buf_size) {
+ td->orig_buffer_size = (unsigned long long) max_bs
+ * (unsigned long long) max_units;
+
+ if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
+ unsigned long bs;
+
+ bs = td->orig_buffer_size + td->o.hugepage_size - 1;
+ td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
+ }
+
+ if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
return 1;
}
- td->orig_buffer_size = buf_size;
-
- if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
- td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
- else if (td->orig_buffer_size & page_mask)
- td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
-
if (allocate_io_mem(td))
return 1;
- p = ALIGN(td->orig_buffer);
+ if (td->o.odirect)
+ p = ALIGN(td->orig_buffer);
+ else
+ p = td->orig_buffer;
+
for (i = 0; i < max_units; i++) {
+ if (td->terminate)
+ return 1;
io_u = malloc(sizeof(*io_u));
memset(io_u, 0, sizeof(*io_u));
INIT_LIST_HEAD(&io_u->list);
- io_u->buf = p + max_bs * i;
- if (td_write(td) || td_rw(td))
- fill_rand_buf(io_u, max_bs);
+ if (!(td->io_ops->flags & FIO_NOIO)) {
+ io_u->buf = p + max_bs * i;
+
+ if (td_write(td) && !td->o.refill_buffers)
+ io_u_fill_buffer(td, io_u, max_bs);
+ }
io_u->index = i;
io_u->flags = IO_U_F_FREE;
f = fopen(tmp, "r+");
if (!f) {
- td_verror(td, errno, "fopen");
+ if (errno == ENOENT) {
+ log_err("fio: os or kernel doesn't support IO scheduler"
+ " switching\n");
+ return 0;
+ }
+ td_verror(td, errno, "fopen iosched");
return 1;
}
return 0;
}
+static int keep_running(struct thread_data *td)
+{
+ unsigned long long io_done;
+
+ if (td->done)
+ return 0;
+ if (td->o.time_based)
+ return 1;
+ if (td->o.loops) {
+ td->o.loops--;
+ return 1;
+ }
+
+ io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
+ + td->io_skip_bytes;
+ if (io_done < td->o.size)
+ return 1;
+
+ return 0;
+}
+
static int clear_io_state(struct thread_data *td)
{
struct fio_file *f;
td->last_was_sync = 0;
- for_each_file(td, f, i)
- td_io_close_file(td, f);
+ /*
+ * reset file done count if we are to start over
+ */
+ if (td->o.time_based || td->o.loops)
+ td->nr_done_files = 0;
+
+ close_files(td);
ret = 0;
for_each_file(td, f, i) {
+ f->flags &= ~FIO_FILE_DONE;
ret = td_io_open_file(td, f);
if (ret)
break;
*/
static void *thread_main(void *data)
{
- unsigned long long runtime[2];
+ unsigned long long runtime[2], elapsed;
struct thread_data *td = data;
- unsigned long elapsed;
int clear_state;
if (!td->o.use_thread)
td->pid = getpid();
+ dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
+
INIT_LIST_HEAD(&td->io_u_freelist);
INIT_LIST_HEAD(&td->io_u_busylist);
INIT_LIST_HEAD(&td->io_u_requeues);
INIT_LIST_HEAD(&td->io_log_list);
+ INIT_LIST_HEAD(&td->io_hist_list);
+ td->io_hist_tree = RB_ROOT;
+
+ td_set_runstate(td, TD_INITIALIZED);
+ fio_mutex_up(startup_mutex);
+ fio_mutex_down(td->mutex);
+
+ /*
+ * the ->mutex mutex is now no longer used, close it to avoid
+ * eating a file descriptor
+ */
+ fio_mutex_remove(td->mutex);
+
+ /*
+ * May alter parameters that init_io_u() will use, so we need to
+ * do this first.
+ */
+ if (init_iolog(td))
+ goto err;
if (init_io_u(td))
- goto err_sem;
+ goto err;
- if (fio_setaffinity(td) == -1) {
+ if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
td_verror(td, errno, "cpu_set_affinity");
- goto err_sem;
+ goto err;
}
- if (init_iolog(td))
- goto err_sem;
-
- if (td->ioprio) {
+ if (td->ioprio_set) {
if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
td_verror(td, errno, "ioprio_set");
- goto err_sem;
+ goto err;
}
}
if (nice(td->o.nice) == -1) {
td_verror(td, errno, "nice");
- goto err_sem;
+ goto err;
}
if (td->o.ioscheduler && switch_ioscheduler(td))
- goto err_sem;
-
- td_set_runstate(td, TD_INITIALIZED);
- fio_sem_up(startup_sem);
- fio_sem_down(td->mutex);
-
- /*
- * the ->mutex semaphore is now no longer used, close it to avoid
- * eating a file descriptor
- */
- fio_sem_remove(td->mutex);
+ goto err;
if (!td->o.create_serialize && setup_files(td))
goto err;
runtime[0] = runtime[1] = 0;
clear_state = 0;
- while (td->o.loops--) {
+ while (keep_running(td)) {
fio_gettime(&td->start, NULL);
memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
if (td->o.ratemin)
- memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
+ memcpy(&td->lastrate, &td->ts.stat_sample_time,
+ sizeof(td->lastrate));
if (clear_state && clear_io_state(td))
break;
if (td_read(td) && td->io_bytes[DDIR_READ]) {
if (td->rw_end_set[DDIR_READ])
- elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
+ elapsed = utime_since(&td->start,
+ &td->rw_end[DDIR_READ]);
else
elapsed = utime_since_now(&td->start);
}
if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
if (td->rw_end_set[DDIR_WRITE])
- elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
+ elapsed = utime_since(&td->start,
+ &td->rw_end[DDIR_WRITE]);
else
elapsed = utime_since_now(&td->start);
runtime[DDIR_WRITE] += elapsed;
}
-
+
if (td->error || td->terminate)
break;
- if (td->o.verify == VERIFY_NONE)
+ if (!td->o.do_verify ||
+ td->o.verify == VERIFY_NONE ||
+ (td->io_ops->flags & FIO_UNIDIR))
continue;
if (clear_io_state(td))
}
update_rusage_stat(td);
- td->ts.runtime[0] = runtime[0] / 1000;
- td->ts.runtime[1] = runtime[1] / 1000;
+ td->ts.runtime[0] = (runtime[0] + 999) / 1000;
+ td->ts.runtime[1] = (runtime[1] + 999) / 1000;
td->ts.total_run_time = mtime_since_now(&td->epoch);
td->ts.io_bytes[0] = td->io_bytes[0];
td->ts.io_bytes[1] = td->io_bytes[1];
finish_log(td, td->ts.slat_log, "slat");
if (td->ts.clat_log)
finish_log(td, td->ts.clat_log, "clat");
- if (td->o.write_iolog_file)
- write_iolog_close(td);
if (td->o.exec_postrun) {
if (system(td->o.exec_postrun) < 0)
log_err("fio: postrun %s failed\n", td->o.exec_postrun);
err:
if (td->error)
- printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
- close_files(td);
+ printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
+ td->verror);
+ close_and_free_files(td);
close_ioengine(td);
cleanup_io_u(td);
+
+ /*
+ * do this very late, it will log file closing as well
+ */
+ if (td->o.write_iolog_file)
+ write_iolog_close(td);
+
options_mem_free(td);
td_set_runstate(td, TD_EXITED);
return (void *) (unsigned long) td->error;
-err_sem:
- fio_sem_up(startup_sem);
- goto err;
}
/*
static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
{
struct thread_data *td;
- int i, cputhreads, pending, status, ret;
+ int i, cputhreads, realthreads, pending, status, ret;
/*
* reap exited threads (TD_EXITED -> TD_REAPED)
*/
- pending = cputhreads = 0;
+ realthreads = pending = cputhreads = 0;
for_each_td(td, i) {
int flags = 0;
*/
if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
cputhreads++;
+ else
+ realthreads++;
- if (!td->pid || td->runstate == TD_REAPED)
+ if (!td->pid) {
+ pending++;
+ continue;
+ }
+ if (td->runstate == TD_REAPED)
continue;
if (td->o.use_thread) {
if (td->runstate == TD_EXITED) {
ret = waitpid(td->pid, &status, flags);
if (ret < 0) {
if (errno == ECHILD) {
- log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
+ log_err("fio: pid=%d disappeared %d\n",
+ (int) td->pid, td->runstate);
td_set_runstate(td, TD_REAPED);
goto reaped;
}
int sig = WTERMSIG(status);
if (sig != SIGQUIT)
- log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+ log_err("fio: pid=%d, got signal=%d\n",
+ (int) td->pid, sig);
td_set_runstate(td, TD_REAPED);
goto reaped;
}
/*
* thread is not dead, continue
*/
+ pending++;
continue;
reaped:
- if (td->o.use_thread) {
- long ret;
-
- if (pthread_join(td->thread, (void *) &ret))
- perror("pthread_join");
- }
-
(*nr_running)--;
(*m_rate) -= td->o.ratemin;
(*t_rate) -= td->o.rate;
+ if (!td->pid)
+ pending--;
if (td->error)
exit_value++;
+
+ done_secs += mtime_since_now(&td->epoch) / 1000;
}
- if (*nr_running == cputhreads && !pending)
+ if (*nr_running == cputhreads && !pending && realthreads)
terminate_threads(TERMINATE_ALL);
}
if (!terse_output) {
printf("Starting ");
if (nr_thread)
- printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
+ printf("%d thread%s", nr_thread,
+ nr_thread > 1 ? "s" : "");
if (nr_process) {
if (nr_thread)
printf(" and ");
- printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
+ printf("%d process%s", nr_process,
+ nr_process > 1 ? "es" : "");
}
printf("\n");
fflush(stdout);
if (setup_files(td)) {
exit_value++;
if (td->error)
- log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
+ log_err("fio: pid=%d, err=%d/%s\n",
+ (int) td->pid, td->error, td->verror);
td_set_runstate(td, TD_REAPED);
todo--;
+ } else {
+ struct fio_file *f;
+ unsigned int i;
+
+ /*
+ * for sharing to work, each job must always open
+ * its own files. so close them, if we opened them
+ * for creation
+ */
+ for_each_file(td, f, i)
+ td_io_close_file(td, f);
}
init_disk_util(td);
continue;
}
- if (td->o.stonewall && (nr_started || nr_running))
+ if (td->o.stonewall && (nr_started || nr_running)) {
+ dprint(FD_PROCESS, "%s: stonewall wait\n",
+ td->o.name);
break;
+ }
/*
* Set state to created. Thread will transition
nr_started++;
if (td->o.use_thread) {
- if (pthread_create(&td->thread, NULL, thread_main, td)) {
- perror("thread_create");
+ dprint(FD_PROCESS, "will pthread_create\n");
+ if (pthread_create(&td->thread, NULL,
+ thread_main, td)) {
+ perror("pthread_create");
nr_started--;
break;
}
+ if (pthread_detach(td->thread) < 0)
+ perror("pthread_detach");
} else {
- if (!fork()) {
+ pid_t pid;
+ dprint(FD_PROCESS, "will fork\n");
+ pid = fork();
+ if (!pid) {
int ret = fork_main(shm_id, i);
- exit(ret);
- }
+ _exit(ret);
+ } else if (i == fio_debug_jobno)
+ *fio_debug_jobp = pid;
}
- fio_sem_down(startup_sem);
+ fio_mutex_down(startup_mutex);
}
/*
m_rate += td->o.ratemin;
t_rate += td->o.rate;
todo--;
- fio_sem_up(td->mutex);
+ fio_mutex_up(td->mutex);
}
reap_threads(&nr_running, &t_rate, &m_rate);
{
long ps;
+ sinit();
+
/*
* We need locale for number printing, if it isn't set then just
* go with the US format.
if (parse_options(argc, argv))
return 1;
- if (!thread_number) {
- log_err("Nothing to do\n");
- return 1;
- }
+ if (!thread_number)
+ return 0;
ps = sysconf(_SC_PAGESIZE);
if (ps < 0) {
setup_log(&agg_io_log[DDIR_WRITE]);
}
- startup_sem = fio_sem_init(0);
+ startup_mutex = fio_mutex_init(0);
set_genesis_time();
if (!fio_abort) {
show_run_stats();
if (write_bw_log) {
- __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
- __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
+ __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
+ __finish_log(agg_io_log[DDIR_WRITE],
+ "agg-write_bw.log");
}
}
- fio_sem_remove(startup_sem);
+ fio_mutex_remove(startup_mutex);
return exit_value;
}