X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=4d3116b7fd4fc0901f4419e8690270c2c18a2d43;hp=d92d940522191451a830e2b1c44428d385e4ccbd;hb=1fbbf72e16c27a6fda636db3891a41cd37dc6666;hpb=720e84ad8292ab7b3a8e264fb00db71d796600d1

diff --git a/io_u.c b/io_u.c
index d92d9405..4d3116b7 100644
--- a/io_u.c
+++ b/io_u.c
@@ -7,6 +7,8 @@
 
 #include "fio.h"
 #include "hash.h"
+#include "verify.h"
+#include "lib/rand.h"
 
 struct io_completion_data {
 	int nr;				/* input */
@@ -113,7 +115,8 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 
 	i = f->last_free_lookup;
 	*b = (i * BLOCKS_PER_MAP);
-	while ((*b) * min_bs < f->real_file_size) {
+	while ((*b) * min_bs < f->real_file_size &&
+		(*b) * min_bs < f->io_size) {
 		if (f->file_map[i] != (unsigned int) -1) {
 			*b += ffz(f->file_map[i]);
 			if (*b > last_block(td, f, ddir))
@@ -185,7 +188,7 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
  * 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, struct io_u *io_u)
+static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f = io_u->file;
 	unsigned long long b;
@@ -229,12 +232,17 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 	return 0;
 }
 
-static inline int is_power_of_2(unsigned int val)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
-	return (val != 0 && ((val & (val - 1)) == 0));
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_off)
+		return ops->fill_io_u_off(td, io_u);
+
+	return __get_next_offset(td, io_u);
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
 	const int ddir = io_u->ddir;
 	unsigned int uninitialized_var(buflen);
@@ -279,6 +287,16 @@ static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 	return buflen;
 }
 
+static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_size)
+		return ops->fill_io_u_size(td, io_u);
+
+	return __get_next_buflen(td, io_u);
+}
+
 static void set_rwmix_bytes(struct thread_data *td)
 {
 	unsigned int diff;
@@ -305,6 +323,53 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 	return DDIR_WRITE;
 }
 
+static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
+{
+	enum fio_ddir odir = ddir ^ 1;
+	struct timeval t;
+	long usec;
+
+	if (td->rate_pending_usleep[ddir] <= 0)
+		return ddir;
+
+	/*
+	 * We have too much pending sleep in this direction. See if we
+	 * should switch.
+	 */
+	if (td_rw(td)) {
+		/*
+		 * Other direction does not have too much pending, switch
+		 */
+		if (td->rate_pending_usleep[odir] < 100000)
+			return odir;
+
+		/*
+		 * Both directions have pending sleep. Sleep the minimum time
+		 * and deduct from both.
+		 */
+		if (td->rate_pending_usleep[ddir] <=
+			td->rate_pending_usleep[odir]) {
+			usec = td->rate_pending_usleep[ddir];
+		} else {
+			usec = td->rate_pending_usleep[odir];
+			ddir = odir;
+		}
+	} else
+		usec = td->rate_pending_usleep[ddir];
+
+	fio_gettime(&t, NULL);
+	usec_sleep(td, usec);
+	usec = utime_since_now(&t);
+
+	td->rate_pending_usleep[ddir] -= usec;
+
+	odir = ddir ^ 1;
+	if (td_rw(td) && __should_check_rate(td, odir))
+		td->rate_pending_usleep[odir] -= usec;
+
+	return ddir;
+}
+
 /*
  * 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
@@ -312,44 +377,60 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
  */
 static enum fio_ddir get_rw_ddir(struct thread_data *td)
 {
+	enum fio_ddir ddir;
+
+	/*
+	 * see if it's time to fsync
+	 */
+	if (td->o.fsync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC;
+
+	/*
+	 * see if it's time to fdatasync
+	 */
+	if (td->o.fdatasync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_DATASYNC;
+
+	/*
+	 * see if it's time to sync_file_range
+	 */
+	if (td->sync_file_range_nr &&
+	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC_FILE_RANGE;
+
 	if (td_rw(td)) {
 		/*
 		 * Check if it's time to seed a new data direction.
 		 */
 		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
-			unsigned long long max_bytes;
-			enum fio_ddir ddir;
-
 			/*
 			 * Put a top limit on how many bytes we do for
 			 * one data direction, to avoid overflowing the
 			 * ranges too much
 			 */
 			ddir = get_rand_ddir(td);
-			max_bytes = td->this_io_bytes[ddir];
-			if (max_bytes >=
-			    (td->o.size * td->o.rwmix[ddir] / 100)) {
-				if (!td->rw_end_set[ddir]) {
-					td->rw_end_set[ddir] = 1;
-					fio_gettime(&td->rw_end[ddir], NULL);
-				}
-
-				ddir ^= 1;
-			}
 
 			if (ddir != td->rwmix_ddir)
 				set_rwmix_bytes(td);
 
 			td->rwmix_ddir = ddir;
 		}
-		return td->rwmix_ddir;
+		ddir = td->rwmix_ddir;
 	} else if (td_read(td))
-		return DDIR_READ;
+		ddir = DDIR_READ;
 	else
-		return DDIR_WRITE;
+		ddir = DDIR_WRITE;
+
+	td->rwmix_ddir = rate_ddir(td, ddir);
+	return td->rwmix_ddir;
 }
 
-static void put_file_log(struct thread_data *td, struct fio_file *f)
+void put_file_log(struct thread_data *td, struct fio_file *f)
 {
 	int ret = put_file(td, f);
 
@@ -359,16 +440,27 @@ static void put_file_log(struct thread_data *td, struct fio_file *f)
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
-	assert((io_u->flags & IO_U_F_FREE) == 0);
+	td_io_u_lock(td);
+
 	io_u->flags |= IO_U_F_FREE;
+	io_u->flags &= ~IO_U_F_FREE_DEF;
 
 	if (io_u->file)
 		put_file_log(td, io_u->file);
 
 	io_u->file = NULL;
-	flist_del(&io_u->list);
+	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
+	flist_del_init(&io_u->list);
 	flist_add(&io_u->list, &td->io_u_freelist);
-	td->cur_depth--;
+	td_io_u_unlock(td);
+	td_io_u_free_notify(td);
+}
+
+void clear_io_u(struct thread_data *td, struct io_u *io_u)
+{
+	io_u->flags &= ~IO_U_F_FLIGHT;
+	put_io_u(td, io_u);
 }
 
 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
@@ -377,15 +469,18 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 
 	dprint(FD_IO, "requeue %p\n", __io_u);
 
+	td_io_u_lock(td);
+
 	__io_u->flags |= IO_U_F_FREE;
-	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
+	if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir))
 		td->io_issues[__io_u->ddir]--;
 
 	__io_u->flags &= ~IO_U_F_FLIGHT;
-
+	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
 	flist_del(&__io_u->list);
 	flist_add_tail(&__io_u->list, &td->io_u_requeues);
-	td->cur_depth--;
+	td_io_u_unlock(td);
 	*io_u = NULL;
 }
 
@@ -394,17 +489,13 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	if (td->io_ops->flags & FIO_NOIO)
 		goto out;
 
+	io_u->ddir = get_rw_ddir(td);
+
 	/*
-	 * see if it's time to sync
+	 * fsync() or fdatasync(), we are done
 	 */
-	if (td->o.fsync_blocks &&
-	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
-	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
-		io_u->ddir = DDIR_SYNC;
+	if (ddir_sync(io_u->ddir))
 		goto out;
-	}
-
-	io_u->ddir = get_rw_ddir(td);
 
 	/*
 	 * See if it's time to switch to a new zone
@@ -620,8 +711,9 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
 /*
  * Get next file to service by choosing one at random
  */
-static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
-					   int badf)
+static struct fio_file *get_next_file_rand(struct thread_data *td,
+					   enum fio_file_flags goodf,
+					   enum fio_file_flags badf)
 {
 	struct fio_file *f;
 	int fno;
@@ -633,10 +725,10 @@ static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
 		fno = (unsigned int) ((double) td->o.nr_files
 			* (r / (OS_RAND_MAX + 1.0)));
 		f = td->files[fno];
-		if (f->flags & FIO_FILE_DONE)
+		if (fio_file_done(f))
 			continue;
 
-		if (!(f->flags & FIO_FILE_OPEN)) {
+		if (!fio_file_open(f)) {
 			int err;
 
 			err = td_io_open_file(td, f);
@@ -673,24 +765,26 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 			td->next_file = 0;
 
 		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
-		if (f->flags & FIO_FILE_DONE) {
+		if (fio_file_done(f)) {
 			f = NULL;
 			continue;
 		}
 
-		if (!(f->flags & FIO_FILE_OPEN)) {
+		if (!fio_file_open(f)) {
 			int err;
 
 			err = td_io_open_file(td, f);
 			if (err) {
 				dprint(FD_FILE, "error %d on open of %s\n",
 					err, f->file_name);
+				f = NULL;
 				continue;
 			}
 			opened = 1;
 		}
 
-		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
+		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
+								f->flags);
 		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			break;
 
@@ -704,7 +798,7 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 	return f;
 }
 
-static struct fio_file *get_next_file(struct thread_data *td)
+static struct fio_file *__get_next_file(struct thread_data *td)
 {
 	struct fio_file *f;
 
@@ -719,7 +813,7 @@ static struct fio_file *get_next_file(struct thread_data *td)
 	}
 
 	f = td->file_service_file;
-	if (f && (f->flags & FIO_FILE_OPEN) && !(f->flags & FIO_FILE_CLOSING)) {
+	if (f && fio_file_open(f) && !fio_file_closing(f)) {
 		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
 			goto out;
 		if (td->file_service_left--)
@@ -728,9 +822,9 @@ static struct fio_file *get_next_file(struct thread_data *td)
 
 	if (td->o.file_service_type == FIO_FSERVICE_RR ||
 	    td->o.file_service_type == FIO_FSERVICE_SEQ)
-		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
 	else
-		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
 
 	td->file_service_file = f;
 	td->file_service_left = td->file_service_nr - 1;
@@ -739,6 +833,16 @@ out:
 	return f;
 }
 
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->get_next_file)
+		return ops->get_next_file(td);
+
+	return __get_next_file(td);
+}
+
 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f;
@@ -757,9 +861,10 @@ static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 		put_file_log(td, f);
 		td_io_close_file(td, f);
 		io_u->file = NULL;
-		f->flags |= FIO_FILE_DONE;
+		fio_file_set_done(f);
 		td->nr_done_files++;
-		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
+		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
+					td->nr_done_files, td->o.nr_files);
 	} while (1);
 
 	return 0;
@@ -770,6 +875,9 @@ struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
 
+	td_io_u_lock(td);
+
+again:
 	if (!flist_empty(&td->io_u_requeues))
 		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
 	else if (!queue_full(td)) {
@@ -783,14 +891,23 @@ struct io_u *__get_io_u(struct thread_data *td)
 
 	if (io_u) {
 		assert(io_u->flags & IO_U_F_FREE);
-		io_u->flags &= ~IO_U_F_FREE;
+		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
 
 		io_u->error = 0;
 		flist_del(&io_u->list);
 		flist_add(&io_u->list, &td->io_u_busylist);
 		td->cur_depth++;
+		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+	} else if (td->o.verify_async) {
+		/*
+		 * We ran out, wait for async verify threads to finish and
+		 * return one
+		 */
+		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
+		goto again;
 	}
 
+	td_io_u_unlock(td);
 	return io_u;
 }
 
@@ -827,9 +944,9 @@ struct io_u *get_io_u(struct thread_data *td)
 	}
 
 	f = io_u->file;
-	assert(f->flags & FIO_FILE_OPEN);
+	assert(fio_file_open(f));
 
-	if (io_u->ddir != DDIR_SYNC) {
+	if (!ddir_sync(io_u->ddir)) {
 		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
 			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
 			goto err_put;
@@ -837,7 +954,7 @@ struct io_u *get_io_u(struct thread_data *td)
 
 		f->last_pos = io_u->offset + io_u->buflen;
 
-		if (td->o.verify != VERIFY_NONE)
+		if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
 			populate_verify_io_u(td, io_u);
 		else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
 			io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
@@ -846,7 +963,6 @@ struct io_u *get_io_u(struct thread_data *td)
 	/*
 	 * Set io data pointers.
 	 */
-	io_u->endpos = io_u->offset + io_u->buflen;
 	io_u->xfer_buf = io_u->buf;
 	io_u->xfer_buflen = io_u->buflen;
 
@@ -888,14 +1004,22 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	 * initialized, silence that warning.
 	 */
 	unsigned long uninitialized_var(usec);
+	struct fio_file *f;
 
 	dprint_io_u(io_u, "io complete");
 
+	td_io_u_lock(td);
 	assert(io_u->flags & IO_U_F_FLIGHT);
 	io_u->flags &= ~IO_U_F_FLIGHT;
+	td_io_u_unlock(td);
 
-	if (io_u->ddir == DDIR_SYNC) {
+	if (ddir_sync(io_u->ddir)) {
 		td->last_was_sync = 1;
+		f = io_u->file;
+		if (f) {
+			f->first_write = -1ULL;
+			f->last_write = -1ULL;
+		}
 		return;
 	}
 
@@ -904,23 +1028,49 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	if (!io_u->error) {
 		unsigned int bytes = io_u->buflen - io_u->resid;
 		const enum fio_ddir idx = io_u->ddir;
+		const enum fio_ddir odx = io_u->ddir ^ 1;
 		int ret;
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
 		td->this_io_bytes[idx] += bytes;
 
+		if (idx == DDIR_WRITE) {
+			f = io_u->file;
+			if (f) {
+				if (f->first_write == -1ULL ||
+				    io_u->offset < f->first_write)
+					f->first_write = io_u->offset;
+				if (f->last_write == -1ULL ||
+				    ((io_u->offset + bytes) > f->last_write))
+					f->last_write = io_u->offset + bytes;
+			}
+		}
+
 		if (ramp_time_over(td)) {
+			unsigned long uninitialized_var(lusec);
+
 			if (!td->o.disable_clat || !td->o.disable_bw)
-				usec = utime_since(&io_u->issue_time,
+				lusec = utime_since(&io_u->issue_time,
 							&icd->time);
 
 			if (!td->o.disable_clat) {
-				add_clat_sample(td, idx, usec);
-				io_u_mark_latency(td, usec);
+				add_clat_sample(td, idx, lusec, bytes);
+				io_u_mark_latency(td, lusec);
 			}
 			if (!td->o.disable_bw)
-				add_bw_sample(td, idx, &icd->time);
+				add_bw_sample(td, idx, bytes, &icd->time);
+			if (__should_check_rate(td, idx)) {
+				td->rate_pending_usleep[idx] =
+					((td->this_io_bytes[idx] *
+					  td->rate_nsec_cycle[idx]) / 1000 -
+					 utime_since_now(&td->start));
+			}
+			if (__should_check_rate(td, idx ^ 1))
+				td->rate_pending_usleep[odx] =
+					((td->this_io_bytes[odx] *
+					  td->rate_nsec_cycle[odx]) / 1000 -
+					 utime_since_now(&td->start));
 		}
 
 		if (td_write(td) && idx == DDIR_WRITE &&
@@ -939,6 +1089,17 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 		icd->error = io_u->error;
 		io_u_log_error(td, io_u);
 	}
+	if (td->o.continue_on_error && icd->error &&
+	    td_non_fatal_error(icd->error)) {
+		/*
+		 * If there is a non_fatal error, then add to the error count
+		 * and clear all the errors.
+		 */
+		update_error_count(td, icd->error);
+		td_clear_error(td);
+		icd->error = 0;
+		io_u->error = 0;
+	}
 }
 
 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
@@ -963,32 +1124,44 @@ static void ios_completed(struct thread_data *td,
 		io_u = td->io_ops->event(td, i);
 
 		io_completed(td, io_u, icd);
-		put_io_u(td, io_u);
+
+		if (!(io_u->flags & IO_U_F_FREE_DEF))
+			put_io_u(td, io_u);
 	}
 }
 
 /*
  * Complete a single io_u for the sync engines.
  */
-long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
+int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
+		       unsigned long *bytes)
 {
 	struct io_completion_data icd;
 
 	init_icd(td, &icd, 1);
 	io_completed(td, io_u, &icd);
-	put_io_u(td, io_u);
 
-	if (!icd.error)
-		return icd.bytes_done[0] + icd.bytes_done[1];
+	if (!(io_u->flags & IO_U_F_FREE_DEF))
+		put_io_u(td, io_u);
+
+	if (icd.error) {
+		td_verror(td, icd.error, "io_u_sync_complete");
+		return -1;
+	}
+
+	if (bytes) {
+		bytes[0] += icd.bytes_done[0];
+		bytes[1] += icd.bytes_done[1];
+	}
 
-	td_verror(td, icd.error, "io_u_sync_complete");
-	return -1;
+	return 0;
 }
 
 /*
  * Called to complete min_events number of io for the async engines.
  */
-long io_u_queued_complete(struct thread_data *td, int min_evts)
+int io_u_queued_complete(struct thread_data *td, int min_evts,
+			 unsigned long *bytes)
 {
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
@@ -1009,11 +1182,17 @@ long io_u_queued_complete(struct thread_data *td, int min_evts)
 
 	init_icd(td, &icd, ret);
 	ios_completed(td, &icd);
-	if (!icd.error)
-		return icd.bytes_done[0] + icd.bytes_done[1];
+	if (icd.error) {
+		td_verror(td, icd.error, "io_u_queued_complete");
+		return -1;
+	}
 
-	td_verror(td, icd.error, "io_u_queued_complete");
-	return -1;
+	if (bytes) {
+		bytes[0] += icd.bytes_done[0];
+		bytes[1] += icd.bytes_done[1];
+	}
+
+	return 0;
 }
 
 /*
@@ -1025,7 +1204,7 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 		unsigned long slat_time;
 
 		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
-		add_slat_sample(td, io_u->ddir, slat_time);
+		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
 	}
 }
 
@@ -1039,7 +1218,7 @@ void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
 
 	if (!td->o.zero_buffers) {
 		while ((void *) ptr - io_u->buf < max_bs) {
-			*ptr = rand() * GOLDEN_RATIO_PRIME;
+			*ptr = __rand(&__fio_rand_state);
 			ptr++;
 		}
 	} else