X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=b636f576eb808caa286eef42c3e3e7bc68bc8d8a;hp=f71856fc8b943b08fcc0835f3c91eeacd790b6bb;hb=48f5abd393cc614e1984391b28ba491ef9b26f5e;hpb=43063a1c8288d987eb9c14e2736b3b857836def6

diff --git a/io_u.c b/io_u.c
index f71856fc..b636f576 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
@@ -85,7 +84,7 @@ 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]);
 			f->last_free_lookup = i;
 			return 0;
 		}
@@ -110,20 +109,40 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 			*b = 0;
 		else
 			*b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
+		/*
+		 * if we are not maintaining a random map, we are done.
+		 */
 		if (td->o.norandommap)
-			break;
+			return 0;
+
+		/*
+		 * calculate map offset and chec if it's free
+		 */
 		rb = *b + (f->file_offset / td->o.min_bs[ddir]);
-		loops--;
-	} while (!random_map_free(td, f, rb) && loops);
+		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, 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, b);
 }
 
 /*
@@ -158,7 +177,6 @@ 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;
@@ -172,19 +190,6 @@ static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 			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->o.min_bs[ddir]) {
-			if (!td->o.odirect) {
-				assert(io_u->offset <= f->real_file_size);
-				buflen = f->real_file_size - io_u->offset;
-				return buflen;
-			}
-			return 0;
-		}
-
-		buflen = td->o.min_bs[ddir];
-	}
-
 	return buflen;
 }
 
@@ -280,6 +285,9 @@ 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);
@@ -304,7 +312,7 @@ 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)
+	if (td->o.read_iolog_file)
 		return read_iolog_get(td, io_u);
 
 	/*
@@ -319,6 +327,15 @@ 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.
@@ -355,17 +372,22 @@ 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;
 	}
@@ -374,38 +396,98 @@ 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);
 }
 
 /*
@@ -422,6 +504,8 @@ static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
 
 		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;
@@ -444,6 +528,11 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 		if (td->next_file >= td->o.nr_files)
 			td->next_file = 0;
 
+		if (f->flags & FIO_FILE_DONE) {
+			f = NULL;
+			continue;
+		}
+
 		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			break;
 
@@ -459,7 +548,7 @@ static struct fio_file *get_next_file(struct thread_data *td)
 
 	assert(td->o.nr_files <= td->files_index);
 
-	if (!td->nr_open_files)
+	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
 		return NULL;
 
 	f = td->file_service_file;
@@ -480,6 +569,9 @@ 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
@@ -545,15 +637,19 @@ struct io_u *get_io_u(struct thread_data *td)
 
 set_file:
 		io_u->file = f;
+		get_file(f);
 
 		if (!fill_io_u(td, io_u))
 			break;
 
 		/*
-		 * No more to do for this file, close it
+		 * 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
@@ -571,19 +667,7 @@ set_file:
 		}
 	} while (1);
 
-	if (td->zone_bytes >= td->o.zone_size) {
-		td->zone_bytes = 0;
-		f->last_pos += td->o.zone_skip;
-	}
-
-	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;
-		}
-
-		io_u->buflen = f->real_file_size - io_u->offset;
-	}
+	assert(io_u->file->flags & FIO_FILE_OPEN);
 
 	if (io_u->ddir != DDIR_SYNC) {
 		if (!io_u->buflen) {
@@ -600,6 +684,7 @@ 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;
@@ -633,13 +718,11 @@ 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;
 
 	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;
@@ -657,15 +740,15 @@ 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);
+		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.verify != VERIFY_NONE)
 			log_io_piece(td, io_u);
 
@@ -760,7 +843,7 @@ 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);
 }