X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=99f07288376f5cd7f668f169cdfb24f130091299;hp=426a5f8a38b49b1f2e7a1b18de199ddccf53a536;hb=733ed597a79e118bc2480974a4dc4676f2edede1;hpb=9bf2061e2b76a396936c5fe70060574a6c96b8f8

diff --git a/io_u.c b/io_u.c
index 426a5f8a..99f07288 100644
--- a/io_u.c
+++ b/io_u.c
@@ -6,7 +6,6 @@
 #include <assert.h>
 
 #include "fio.h"
-#include "os.h"
 
 /*
  * Change this define to play with the timeout handling
@@ -39,7 +38,7 @@ static int random_map_free(struct thread_data *td, struct fio_file *f,
  */
 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 {
-	unsigned int min_bs = td->rw_min_bs;
+	unsigned int min_bs = td->o.rw_min_bs;
 	struct fio_file *f = io_u->file;
 	unsigned long long block;
 	unsigned int blocks;
@@ -52,7 +51,11 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 	while (blocks < nr_blocks) {
 		unsigned int idx, bit;
 
-		if (!random_map_free(td, f, block))
+		/*
+		 * If we have a mixed random workload, we may
+		 * encounter blocks we already did IO to.
+		 */
+		if (!td->o.ddir_nr && !random_map_free(td, f, block))
 			break;
 
 		idx = RAND_MAP_IDX(td, f, block);
@@ -79,7 +82,7 @@ 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) * td->rw_min_bs < f->real_file_size) {
+	while ((*b) * td->o.rw_min_bs < f->real_file_size) {
 		if (f->file_map[i] != -1UL) {
 			*b += ffz(f->file_map[i]);
 			f->last_free_lookup = i;
@@ -93,6 +96,35 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 	return 1;
 }
 
+static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
+				int ddir, unsigned long long *b)
+{
+	unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
+	unsigned long long r, rb;
+	int loops = 5;
+
+	do {
+		r = os_random_long(&td->random_state);
+		if (!max_blocks)
+			*b = 0;
+		else
+			*b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
+		if (td->o.norandommap)
+			break;
+		rb = *b + (f->file_offset / td->o.min_bs[ddir]);
+		loops--;
+	} while (!random_map_free(td, f, rb) && loops);
+
+	/*
+	 * if we failed to retrieve a truly random offset within
+	 * the loops assigned, see if there are free ones left at all
+	 */
+	if (!loops && get_next_free_block(td, f, b))
+		return 1;
+
+	return 0;
+}
+
 /*
  * 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
@@ -102,32 +134,21 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f = io_u->file;
 	const int ddir = io_u->ddir;
-	unsigned long long b, rb;
-	long r;
-
-	if (td_random(td)) {
-		unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
-		int loops = 5;
+	unsigned long long b;
 
-		do {
-			r = os_random_long(&td->random_state);
-			b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-			if (td->norandommap)
-				break;
-			rb = b + (f->file_offset / td->min_bs[ddir]);
-			loops--;
-		} while (!random_map_free(td, f, rb) && loops);
+	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
+		td->ddir_nr = td->o.ddir_nr;
 
-		/*
-		 * if we failed to retrieve a truly random offset within
-		 * the loops assigned, see if there are free ones left at all
-		 */
-		if (!loops && get_next_free_block(td, f, &b))
+		if (get_next_rand_offset(td, f, ddir, &b))
 			return 1;
-	} else
-		b = f->last_pos / td->min_bs[ddir];
+	} else {
+		if (f->last_pos >= f->real_file_size)
+			return 1;
+
+		b = f->last_pos / td->o.min_bs[ddir];
+	}
 
-	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
+	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
 	if (io_u->offset >= f->real_file_size)
 		return 1;
 
@@ -136,28 +157,49 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 
 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
-	struct fio_file *f = io_u->file;
 	const int ddir = io_u->ddir;
 	unsigned int buflen;
 	long r;
 
-	if (td->min_bs[ddir] == td->max_bs[ddir])
-		buflen = td->min_bs[ddir];
+	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
+		buflen = td->o.min_bs[ddir];
 	else {
 		r = os_random_long(&td->bsrange_state);
-		buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
-		if (!td->bs_unaligned)
-			buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
+		buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
+		if (!td->o.bs_unaligned)
+			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
 	}
 
-	while (buflen + io_u->offset > f->real_file_size) {
-		if (buflen == td->min_bs[ddir])
-			return 0;
+	return buflen;
+}
 
-		buflen = td->min_bs[ddir];
-	}
+static void set_rwmix_bytes(struct thread_data *td)
+{
+	unsigned long long rbytes;
+	unsigned int diff;
 
-	return buflen;
+	/*
+	 * we do time or byte based switch. this is needed because
+	 * buffered writes may issue a lot quicker than they complete,
+	 * whereas reads do not.
+	 */
+	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
+	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
+
+	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
+}
+
+static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
+{
+	unsigned int v;
+	long r;
+
+	r = os_random_long(&td->rwmix_state);
+	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
+	if (v < td->o.rwmix[DDIR_READ])
+		return DDIR_READ;
+
+	return DDIR_WRITE;
 }
 
 /*
@@ -170,23 +212,45 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 	if (td_rw(td)) {
 		struct timeval now;
 		unsigned long elapsed;
+		unsigned int cycle;
 
 		fio_gettime(&now, NULL);
 	 	elapsed = mtime_since_now(&td->rwmix_switch);
 
+		/*
+		 * if this is the first cycle, make it shorter
+		 */
+		cycle = td->o.rwmixcycle;
+		if (!td->rwmix_bytes)
+			cycle /= 10;
+
 		/*
 		 * Check if it's time to seed a new data direction.
 		 */
-		if (elapsed >= td->rwmixcycle) {
-			unsigned int v;
-			long r;
-
-			r = os_random_long(&td->rwmix_state);
-			v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
-			if (v < td->rwmixread)
-				td->rwmix_ddir = DDIR_READ;
-			else
-				td->rwmix_ddir = DDIR_WRITE;
+		if (elapsed >= cycle ||
+		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
+			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;
+					memcpy(&td->rw_end[ddir], &now, sizeof(now));
+				}
+				ddir ^= 1;
+			}
+
+			if (ddir != td->rwmix_ddir)
+				set_rwmix_bytes(td);
+
+			td->rwmix_ddir = ddir;
 			memcpy(&td->rwmix_switch, &now, sizeof(now));
 		}
 		return td->rwmix_ddir;
@@ -225,16 +289,17 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	/*
 	 * If using an iolog, grab next piece if any available.
 	 */
-	if (td->read_iolog)
+	if (td->o.read_iolog_file)
 		return read_iolog_get(td, io_u);
 
 	/*
 	 * see if it's time to sync
 	 */
-	if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
-	    && td->io_issues[DDIR_WRITE] && should_fsync(td)) {
+	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;
-		return 0;
+		goto out;
 	}
 
 	io_u->ddir = get_rw_ddir(td);
@@ -253,22 +318,26 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	/*
 	 * mark entry before potentially trimming io_u
 	 */
-	if (!td->read_iolog && td_random(td) && !td->norandommap)
+	if (td_random(td) && !td->o.norandommap)
 		mark_random_map(td, io_u);
 
 	/*
 	 * If using a write iolog, store this entry.
 	 */
-	if (td->write_iolog_file)
+out:
+	if (td->o.write_iolog_file)
 		write_iolog_put(td, io_u);
 
 	return 0;
 }
 
-static void io_u_mark_depth(struct thread_data *td)
+void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
 {
 	int index = 0;
 
+	if (io_u->ddir == DDIR_SYNC)
+		return;
+
 	switch (td->cur_depth) {
 	default:
 		index++;
@@ -286,8 +355,8 @@ static void io_u_mark_depth(struct thread_data *td)
 		break;
 	}
 
-	td->io_u_map[index]++;
-	td->total_io_u++;
+	td->ts.io_u_map[index]++;
+	td->ts.total_io_u[io_u->ddir]++;
 }
 
 static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
@@ -321,22 +390,27 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
 		break;
 	}
 
-	td->io_u_lat[index]++;
+	td->ts.io_u_lat[index]++;
 }
 
 /*
  * Get next file to service by choosing one at random
  */
-static struct fio_file *get_next_file_rand(struct thread_data *td)
+static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
+					   int badf)
 {
-	long r = os_random_long(&td->next_file_state);
-	unsigned int fileno;
 	struct fio_file *f;
+	int fno;
 
 	do {
-		fileno = (unsigned int) ((double) (td->nr_files - 1) * r / (RAND_MAX + 1.0));
-		f = &td->files[fileno];
-		if (f->fd != -1)
+		long r = os_random_long(&td->next_file_state);
+
+		fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
+		f = &td->files[fno];
+		if (f->flags & FIO_FILE_DONE)
+			continue;
+
+		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			return f;
 	} while (1);
 }
@@ -344,7 +418,8 @@ static struct fio_file *get_next_file_rand(struct thread_data *td)
 /*
  * Get next file to service by doing round robin between all available ones
  */
-static struct fio_file *get_next_file_rr(struct thread_data *td)
+static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
+					 int badf)
 {
 	unsigned int old_next_file = td->next_file;
 	struct fio_file *f;
@@ -353,10 +428,15 @@ static struct fio_file *get_next_file_rr(struct thread_data *td)
 		f = &td->files[td->next_file];
 
 		td->next_file++;
-		if (td->next_file >= td->nr_files)
+		if (td->next_file >= td->o.nr_files)
 			td->next_file = 0;
 
-		if (f->fd != -1)
+		if (f->flags & FIO_FILE_DONE) {
+			f = NULL;
+			continue;
+		}
+
+		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			break;
 
 		f = NULL;
@@ -365,6 +445,44 @@ static struct fio_file *get_next_file_rr(struct thread_data *td)
 	return f;
 }
 
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	struct fio_file *f;
+
+	assert(td->o.nr_files <= td->files_index);
+
+	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
+		return NULL;
+
+	f = td->file_service_file;
+	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
+		return f;
+
+	if (td->o.file_service_type == FIO_FSERVICE_RR)
+		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);
+
+	td->file_service_file = f;
+	td->file_service_left = td->file_service_nr - 1;
+	return f;
+}
+
+static struct fio_file *find_next_new_file(struct thread_data *td)
+{
+	struct fio_file *f;
+
+	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
+		return NULL;
+
+	if (td->o.file_service_type == FIO_FSERVICE_RR)
+		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
+	else
+		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
+
+	return f;
+}
+
 struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
@@ -388,7 +506,6 @@ struct io_u *__get_io_u(struct thread_data *td)
 		list_del(&io_u->list);
 		list_add(&io_u->list, &td->io_u_busylist);
 		td->cur_depth++;
-		io_u_mark_depth(td);
 	}
 
 	return io_u;
@@ -402,6 +519,7 @@ struct io_u *get_io_u(struct thread_data *td)
 {
 	struct fio_file *f;
 	struct io_u *io_u;
+	int ret;
 
 	io_u = __get_io_u(td);
 	if (!io_u)
@@ -413,36 +531,46 @@ struct io_u *get_io_u(struct thread_data *td)
 	if (io_u->file)
 		goto out;
 
-	if (td->file_service_type == FIO_FSERVICE_RR)
-		f = get_next_file_rr(td);
-	else
-		f = get_next_file_rand(td);
-
-	if (!f) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
-
-	io_u->file = f;
+	do {
+		f = get_next_file(td);
+		if (!f) {
+			put_io_u(td, io_u);
+			return NULL;
+		}
 
+set_file:
+		io_u->file = f;
 
-	if (fill_io_u(td, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
+		if (!fill_io_u(td, io_u))
+			break;
 
-	if (td->zone_bytes >= td->zone_size) {
-		td->zone_bytes = 0;
-		f->last_pos += td->zone_skip;
-	}
+		/*
+		 * No more to do for this file, close it
+		 */
+		io_u->file = NULL;
+		td_io_close_file(td, f);
+		f->flags |= FIO_FILE_DONE;
+		td->nr_done_files++;
 
-	if (io_u->buflen + io_u->offset > f->real_file_size) {
-		if (td->io_ops->flags & FIO_RAWIO) {
-			put_io_u(td, io_u);
-			return NULL;
+		/*
+		 * probably not the right place to do this, but see
+		 * if we need to open a new file
+		 */
+		if (td->nr_open_files < td->o.open_files &&
+		    td->o.open_files != td->o.nr_files) {
+			f = find_next_new_file(td);
+
+			if (!f || (ret = td_io_open_file(td, f))) {
+				put_io_u(td, io_u);
+				return NULL;
+			}
+			goto set_file;
 		}
+	} while (1);
 
-		io_u->buflen = f->real_file_size - io_u->offset;
+	if (td->zone_bytes >= td->o.zone_size) {
+		td->zone_bytes = 0;
+		f->last_pos += td->o.zone_skip;
 	}
 
 	if (io_u->ddir != DDIR_SYNC) {
@@ -453,13 +581,14 @@ struct io_u *get_io_u(struct thread_data *td)
 
 		f->last_pos = io_u->offset + io_u->buflen;
 
-		if (td->verify != VERIFY_NONE)
+		if (td->o.verify != VERIFY_NONE)
 			populate_verify_io_u(td, io_u);
 	}
 
 	/*
 	 * Set io data pointers.
 	 */
+	io_u->endpos = io_u->offset + io_u->buflen;
 out:
 	io_u->xfer_buf = io_u->buf;
 	io_u->xfer_buflen = io_u->buflen;
@@ -473,6 +602,23 @@ out:
 	return io_u;
 }
 
+void io_u_log_error(struct thread_data *td, struct io_u *io_u)
+{
+	const char *msg[] = { "read", "write", "sync" };
+
+	log_err("fio: io_u error");
+
+	if (io_u->file)
+		log_err(" on file %s", io_u->file->file_name);
+
+	log_err(": %s\n", strerror(io_u->error));
+
+	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
+
+	if (!td->error)
+		td_verror(td, io_u->error, "io_u error");
+}
+
 static void io_completed(struct thread_data *td, struct io_u *io_u,
 			 struct io_completion_data *icd)
 {
@@ -481,6 +627,8 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	assert(io_u->flags & IO_U_F_FLIGHT);
 	io_u->flags &= ~IO_U_F_FLIGHT;
 
+	put_file(td, io_u->file);
+
 	if (io_u->ddir == DDIR_SYNC) {
 		td->last_was_sync = 1;
 		return;
@@ -498,7 +646,7 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 		td->zone_bytes += bytes;
 		td->this_io_bytes[idx] += bytes;
 
-		io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
+		io_u->file->last_completed_pos = io_u->endpos;
 
 		msec = mtime_since(&io_u->issue_time, &icd->time);
 
@@ -506,18 +654,21 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 		add_bw_sample(td, idx, &icd->time);
 		io_u_mark_latency(td, msec);
 
-		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
+		if (td_write(td) && idx == DDIR_WRITE &&
+		    td->o.verify != VERIFY_NONE)
 			log_io_piece(td, io_u);
 
 		icd->bytes_done[idx] += bytes;
 
 		if (io_u->end_io) {
-			ret = io_u->end_io(io_u);
+			ret = io_u->end_io(td, io_u);
 			if (ret && !icd->error)
 				icd->error = ret;
 		}
-	} else
+	} else {
 		icd->error = io_u->error;
+		io_u_log_error(td, io_u);
+	}
 }
 
 static void init_icd(struct io_completion_data *icd, int nr)
@@ -558,6 +709,7 @@ long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 	if (!icd.error)
 		return icd.bytes_done[0] + icd.bytes_done[1];
 
+	td_verror(td, icd.error, "io_u_sync_complete");
 	return -1;
 }
 
@@ -569,12 +721,10 @@ long io_u_queued_complete(struct thread_data *td, int min_events)
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
 	int ret;
+	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-	if (!min_events) {
-		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
-
+	if (!min_events)
 		tvp = &ts;
-	}
 
 	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
 	if (ret < 0) {
@@ -588,6 +738,7 @@ long io_u_queued_complete(struct thread_data *td, int min_events)
 	if (!icd.error)
 		return icd.bytes_done[0] + icd.bytes_done[1];
 
+	td_verror(td, icd.error, "io_u_queued_complete");
 	return -1;
 }
 
@@ -662,7 +813,7 @@ static void io_u_timeout_handler(int fio_unused sig)
 		return;
 	}
 
-	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
+	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
 
 	list_for_each(entry, &td->io_u_busylist) {
 		io_u = list_entry(entry, struct io_u, list);