X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=8f10bfdc0ce3552d1a1f7ccd6f1cdc613bf4ad73;hp=b48c99dad3ed4a3eb25650de53f59203463af850;hb=4ba66134744885e676167dede2f5900512834a31;hpb=1020a139beff50faf1bb18f761b024b664a09e14

diff --git a/io_u.c b/io_u.c
index b48c99da..8f10bfdc 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
@@ -56,7 +55,7 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 		 * 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))
+		if (!td->o.ddir_nr == 1 && !random_map_free(td, f, block))
 			break;
 
 		idx = RAND_MAP_IDX(td, f, block);
@@ -73,11 +72,24 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 		io_u->buflen = blocks * min_bs;
 }
 
+static inline unsigned long long last_block(struct thread_data *td,
+					    struct fio_file *f,
+					    enum fio_ddir ddir)
+{
+	unsigned long long max_blocks;
+
+	max_blocks = f->io_size / td->o.min_bs[ddir];
+	if (!max_blocks)
+		return 0;
+
+	return max_blocks - 1;
+}
+
 /*
  * Return the next free block in the map.
  */
 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
-			       unsigned long long *b)
+			       enum fio_ddir ddir, unsigned long long *b)
 {
 	int i;
 
@@ -85,7 +97,9 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 	*b = (i * BLOCKS_PER_MAP);
 	while ((*b) * td->o.rw_min_bs < f->real_file_size) {
 		if (f->file_map[i] != -1UL) {
-			*b += ffz(f->file_map[i]);
+			*b += fio_ffz(f->file_map[i]);
+			if (*b > last_block(td, f, ddir))
+				break;
 			f->last_free_lookup = i;
 			return 0;
 		}
@@ -94,36 +108,54 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 		i++;
 	}
 
+	dprint(FD_IO, "failed finding a free block\n");
 	return 1;
 }
 
 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
-				int ddir, unsigned long long *b)
+				enum fio_ddir 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));
+		*b = last_block(td, f, ddir);
+
+		/*
+		 * if we are not maintaining a random map, we are done.
+		 */
 		if (td->o.norandommap)
-			break;
-		rb = *b + (f->file_offset / td->o.min_bs[ddir]);
-		loops--;
-	} while (!random_map_free(td, f, rb) && loops);
+			return 0;
+
+		/*
+		 * calculate map offset and chec if it's free
+		 */
+		rb = *b;
+		if (random_map_free(td, f, rb))
+			return 0;
+
+	} while (--loops);
 
 	/*
-	 * if we failed to retrieve a truly random offset within
-	 * the loops assigned, see if there are free ones left at all
+	 * we get here, if we didn't suceed in looking up a block. generate
+	 * a random start offset into the filemap, and find the first free
+	 * block from there.
 	 */
-	if (!loops && get_next_free_block(td, f, b))
-		return 1;
+	loops = 10;
+	do {
+		f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
+		if (!get_next_free_block(td, f, ddir, b))
+			return 0;
 
-	return 0;
+		r = os_random_long(&td->random_state);
+	} while (--loops);
+
+	/*
+	 * that didn't work either, try exhaustive search from the start
+	 */
+	f->last_free_lookup = 0;
+	return get_next_free_block(td, f, ddir, b);
 }
 
 /*
@@ -134,8 +166,8 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 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;
+	enum fio_ddir ddir = io_u->ddir;
 
 	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
 		td->ddir_nr = td->o.ddir_nr;
@@ -143,15 +175,20 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 		if (get_next_rand_offset(td, f, ddir, &b))
 			return 1;
 	} else {
-		if (f->last_pos >= f->real_file_size)
-			return 1;
-
-		b = f->last_pos / td->o.min_bs[ddir];
+		if (f->last_pos >= f->real_file_size) {
+			if (!td_random(td) ||
+			     get_next_rand_offset(td, f, ddir, &b))
+				return 1;
+		} else
+			b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
 	}
 
 	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
-	if (io_u->offset >= f->real_file_size)
+	if (io_u->offset >= f->real_file_size) {
+		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
+					io_u->offset, f->real_file_size);
 		return 1;
+	}
 
 	return 0;
 }
@@ -166,11 +203,31 @@ static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 		buflen = td->o.min_bs[ddir];
 	else {
 		r = os_random_long(&td->bsrange_state);
-		buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
+		if (!td->o.bssplit_nr)
+			buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
+		else {
+			long perc = 0;
+			unsigned int i;
+
+			for (i = 0; i < td->o.bssplit_nr; i++) {
+				struct bssplit *bsp = &td->o.bssplit[i];
+
+				buflen = bsp->bs;
+				perc += bsp->perc;
+				if (r <= ((LONG_MAX / 100L) * perc))
+					break;
+			}
+		}
 		if (!td->o.bs_unaligned)
 			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
 	}
 
+	if (io_u->offset + buflen > io_u->file->real_file_size) {
+		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
+						td->o.min_bs[ddir], ddir);
+		buflen = td->o.min_bs[ddir];
+	}
+
 	return buflen;
 }
 
@@ -266,6 +323,9 @@ void put_io_u(struct thread_data *td, struct io_u *io_u)
 	assert((io_u->flags & IO_U_F_FREE) == 0);
 	io_u->flags |= IO_U_F_FREE;
 
+	if (io_u->file)
+		put_file(td, io_u->file);
+
 	io_u->file = NULL;
 	list_del(&io_u->list);
 	list_add(&io_u->list, &td->io_u_freelist);
@@ -277,6 +337,9 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 	struct io_u *__io_u = *io_u;
 
 	__io_u->flags |= IO_U_F_FREE;
+	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
+		td->io_issues[__io_u->ddir]--;
+		
 	__io_u->flags &= ~IO_U_F_FLIGHT;
 
 	list_del(&__io_u->list);
@@ -287,11 +350,8 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 
 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->o.read_iolog)
-		return read_iolog_get(td, io_u);
+	if (td->io_ops->flags & FIO_NOIO)
+		goto out;
 
 	/*
 	 * see if it's time to sync
@@ -305,16 +365,36 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 
 	io_u->ddir = get_rw_ddir(td);
 
+	/*
+	 * See if it's time to switch to a new zone
+	 */
+	if (td->zone_bytes >= td->o.zone_size) {
+		td->zone_bytes = 0;
+		io_u->file->last_pos += td->o.zone_skip;
+		td->io_skip_bytes += td->o.zone_skip;
+	}
+
 	/*
 	 * No log, let the seq/rand engine retrieve the next buflen and
 	 * position.
 	 */
-	if (get_next_offset(td, io_u))
+	if (get_next_offset(td, io_u)) {
+		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
 		return 1;
+	}
 
 	io_u->buflen = get_next_buflen(td, io_u);
-	if (!io_u->buflen)
+	if (!io_u->buflen) {
+		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
+		return 1;
+	}
+
+	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
+		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
+		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
+				io_u->buflen, io_u->file->real_file_size);
 		return 1;
+	}
 
 	/*
 	 * mark entry before potentially trimming io_u
@@ -326,9 +406,9 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	 * If using a write iolog, store this entry.
 	 */
 out:
-	if (td->o.write_iolog_file)
-		write_iolog_put(td, io_u);
-
+	dprint_io_u(io_u, "fill_io_u");
+	td->zone_bytes += io_u->buflen;
+	log_io_u(td, io_u);
 	return 0;
 }
 
@@ -341,17 +421,22 @@ void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
 
 	switch (td->cur_depth) {
 	default:
-		index++;
+		index = 6;
+		break;
 	case 32 ... 63:
-		index++;
+		index = 5;
+		break;
 	case 16 ... 31:
-		index++;
+		index = 4;
+		break;
 	case 8 ... 15:
-		index++;
+		index = 3;
+		break;
 	case 4 ... 7:
-		index++;
+		index = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		index = 1;
 	case 1:
 		break;
 	}
@@ -360,38 +445,98 @@ void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
 	td->ts.total_io_u[io_u->ddir]++;
 }
 
-static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
+static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
+{
+	int index = 0;
+
+	assert(usec < 1000);
+
+	switch (usec) {
+	case 750 ... 999:
+		index = 9;
+		break;
+	case 500 ... 749:
+		index = 8;
+		break;
+	case 250 ... 499:
+		index = 7;
+		break;
+	case 100 ... 249:
+		index = 6;
+		break;
+	case 50 ... 99:
+		index = 5;
+		break;
+	case 20 ... 49:
+		index = 4;
+		break;
+	case 10 ... 19:
+		index = 3;
+		break;
+	case 4 ... 9:
+		index = 2;
+		break;
+	case 2 ... 3:
+		index = 1;
+	case 0 ... 1:
+		break;
+	}
+
+	assert(index < FIO_IO_U_LAT_U_NR);
+	td->ts.io_u_lat_u[index]++;
+}
+
+static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
 {
 	int index = 0;
 
 	switch (msec) {
 	default:
-		index++;
+		index = 11;
+		break;
 	case 1000 ... 1999:
-		index++;
+		index = 10;
+		break;
 	case 750 ... 999:
-		index++;
+		index = 9;
+		break;
 	case 500 ... 749:
-		index++;
+		index = 8;
+		break;
 	case 250 ... 499:
-		index++;
+		index = 7;
+		break;
 	case 100 ... 249:
-		index++;
+		index = 6;
+		break;
 	case 50 ... 99:
-		index++;
+		index = 5;
+		break;
 	case 20 ... 49:
-		index++;
+		index = 4;
+		break;
 	case 10 ... 19:
-		index++;
+		index = 3;
+		break;
 	case 4 ... 9:
-		index++;
+		index = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		index = 1;
 	case 0 ... 1:
 		break;
 	}
 
-	td->ts.io_u_lat[index]++;
+	assert(index < FIO_IO_U_LAT_M_NR);
+	td->ts.io_u_lat_m[index]++;
+}
+
+static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
+{
+	if (usec < 1000)
+		io_u_mark_lat_usec(td, usec);
+	else
+		io_u_mark_lat_msec(td, usec / 1000);
 }
 
 /*
@@ -411,8 +556,10 @@ static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
 		if (f->flags & FIO_FILE_DONE)
 			continue;
 
-		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
+		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
+			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
 			return f;
+		}
 	} while (1);
 }
 
@@ -443,6 +590,7 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 		f = NULL;
 	} while (td->next_file != old_next_file);
 
+	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
 	return f;
 }
 
@@ -452,12 +600,14 @@ static struct fio_file *get_next_file(struct thread_data *td)
 
 	assert(td->o.nr_files <= td->files_index);
 
-	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
+	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
+		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d, nr_files=%d\n", 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;
+		goto out;
 
 	if (td->o.file_service_type == FIO_FSERVICE_RR)
 		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
@@ -466,6 +616,8 @@ static struct fio_file *get_next_file(struct thread_data *td)
 
 	td->file_service_file = f;
 	td->file_service_left = td->file_service_nr - 1;
+out:
+	dprint(FD_FILE, "get_next_file: %p\n", f);
 	return f;
 }
 
@@ -484,6 +636,50 @@ static struct fio_file *find_next_new_file(struct thread_data *td)
 	return f;
 }
 
+static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
+{
+	struct fio_file *f;
+
+	do {
+		f = get_next_file(td);
+		if (!f)
+			return 1;
+
+set_file:
+		io_u->file = f;
+		get_file(f);
+
+		if (!fill_io_u(td, io_u))
+			break;
+
+		/*
+		 * td_io_close() does a put_file() as well, so no need to
+		 * do that here.
+		 */
+		io_u->file = NULL;
+		td_io_close_file(td, f);
+		f->flags |= FIO_FILE_DONE;
+		td->nr_done_files++;
+
+		/*
+		 * 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 || td_io_open_file(td, f))
+				return 1;
+
+			goto set_file;
+		}
+	} while (1);
+
+	return 0;
+}
+
+
 struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
@@ -520,11 +716,12 @@ 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)
+	if (!io_u) {
+		dprint(FD_IO, "__get_io_u failed\n");
 		return NULL;
+	}
 
 	/*
 	 * from a requeue, io_u already setup
@@ -532,52 +729,24 @@ struct io_u *get_io_u(struct thread_data *td)
 	if (io_u->file)
 		goto out;
 
-	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))
-			break;
-
-		/*
-		 * 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++;
-
-		/*
-		 * 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);
-
-	if (td->zone_bytes >= td->o.zone_size) {
-		td->zone_bytes = 0;
-		f->last_pos += td->o.zone_skip;
+	/*
+	 * If using an iolog, grab next piece if any available.
+	 */
+	if (td->o.read_iolog_file) {
+		if (read_iolog_get(td, io_u))
+			goto err_put;
+	} else if (set_io_u_file(td, io_u)) {
+		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
+		goto err_put;
 	}
+	
+	f = io_u->file;
+	assert(f->flags & FIO_FILE_OPEN);
 
 	if (io_u->ddir != DDIR_SYNC) {
-		if (!io_u->buflen) {
-			put_io_u(td, io_u);
-			return NULL;
+		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;
 		}
 
 		f->last_pos = io_u->offset + io_u->buflen;
@@ -589,17 +758,19 @@ set_file:
 	/*
 	 * 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;
 
-	if (td_io_prep(td, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
+	if (!td_io_prep(td, io_u)) {
+		fio_gettime(&io_u->start_time, NULL);
+		return io_u;
 	}
-
-	fio_gettime(&io_u->start_time, NULL);
-	return io_u;
+err_put:
+	dprint(FD_IO, "get_io_u failed\n");
+	put_io_u(td, io_u);
+	return NULL;
 }
 
 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
@@ -622,13 +793,13 @@ void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 static void io_completed(struct thread_data *td, struct io_u *io_u,
 			 struct io_completion_data *icd)
 {
-	unsigned long msec;
+	unsigned long usec;
+
+	dprint_io_u(io_u, "io complete");
 
 	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;
@@ -643,18 +814,18 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
-		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);
+		usec = utime_since(&io_u->issue_time, &icd->time);
 
-		add_clat_sample(td, idx, msec);
+		add_clat_sample(td, idx, usec);
 		add_bw_sample(td, idx, &icd->time);
-		io_u_mark_latency(td, msec);
+		io_u_mark_latency(td, usec);
 
-		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
+		if (td_write(td) && idx == DDIR_WRITE &&
+		    td->o.do_verify &&
 		    td->o.verify != VERIFY_NONE)
 			log_io_piece(td, io_u);
 
@@ -723,6 +894,8 @@ long io_u_queued_complete(struct thread_data *td, int min_events)
 	int ret;
 	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
+	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
+
 	if (!min_events)
 		tvp = &ts;
 
@@ -749,7 +922,7 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 {
 	unsigned long slat_time;
 
-	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
+	slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
 	add_slat_sample(td, io_u->ddir, slat_time);
 }