io_u.c revision f697125ab9003d358a37186d610e09799afd190f 
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9#include "os.h"
10
11/*
12 * The ->file_map[] contains a map of blocks we have or have not done io
13 * to yet. Used to make sure we cover the entire range in a fair fashion.
14 */
15static int random_map_free(struct thread_data *td, struct fio_file *f,
16			   unsigned long long block)
17{
18	unsigned int idx = RAND_MAP_IDX(td, f, block);
19	unsigned int bit = RAND_MAP_BIT(td, f, block);
20
21	return (f->file_map[idx] & (1UL << bit)) == 0;
22}
23
24/*
25 * Mark a given offset as used in the map.
26 */
27static void mark_random_map(struct thread_data *td, struct fio_file *f,
28			    struct io_u *io_u)
29{
30	unsigned int min_bs = td->min_bs[io_u->ddir];
31	unsigned long long block;
32	unsigned int blocks;
33
34	block = io_u->offset / (unsigned long long) min_bs;
35	blocks = 0;
36	while (blocks < (io_u->buflen / min_bs)) {
37		unsigned int idx, bit;
38
39		if (!random_map_free(td, f, block))
40			break;
41
42		idx = RAND_MAP_IDX(td, f, block);
43		bit = RAND_MAP_BIT(td, f, block);
44
45		assert(idx < f->num_maps);
46
47		f->file_map[idx] |= (1UL << bit);
48		block++;
49		blocks++;
50	}
51
52	if ((blocks * min_bs) < io_u->buflen)
53		io_u->buflen = blocks * min_bs;
54}
55
56/*
57 * Return the next free block in the map.
58 */
59static int get_next_free_block(struct thread_data *td, struct fio_file *f,
60			       unsigned long long *b)
61{
62	int i;
63
64	*b = 0;
65	i = 0;
66	while ((*b) * td->rw_min_bs < f->file_size) {
67		if (f->file_map[i] != -1UL) {
68			*b += ffz(f->file_map[i]);
69			return 0;
70		}
71
72		*b += BLOCKS_PER_MAP;
73		i++;
74	}
75
76	return 1;
77}
78
79/*
80 * For random io, generate a random new block and see if it's used. Repeat
81 * until we find a free one. For sequential io, just return the end of
82 * the last io issued.
83 */
84static int get_next_offset(struct thread_data *td, struct fio_file *f,
85			   unsigned long long *offset, int ddir)
86{
87	unsigned long long b, rb;
88	long r;
89
90	if (!td->sequential) {
91		unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
92		int loops = 50;
93
94		do {
95			r = os_random_long(&td->random_state);
96			b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
97			if (td->norandommap)
98				break;
99			rb = b + (f->file_offset / td->min_bs[ddir]);
100			loops--;
101		} while (!random_map_free(td, f, rb) && loops);
102
103		if (!loops) {
104			if (get_next_free_block(td, f, &b))
105				return 1;
106		}
107	} else
108		b = f->last_pos / td->min_bs[ddir];
109
110	*offset = (b * td->min_bs[ddir]) + f->file_offset;
111	if (*offset > f->file_size)
112		return 1;
113
114	return 0;
115}
116
117static unsigned int get_next_buflen(struct thread_data *td, int ddir)
118{
119	unsigned int buflen;
120	long r;
121
122	if (td->min_bs[ddir] == td->max_bs[ddir])
123		buflen = td->min_bs[ddir];
124	else {
125		r = os_random_long(&td->bsrange_state);
126		buflen = (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
127		if (!td->bs_unaligned)
128			buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
129	}
130
131	if (buflen > td->io_size - td->this_io_bytes[ddir]) {
132		/*
133		 * if using direct/raw io, we may not be able to
134		 * shrink the size. so just fail it.
135		 */
136		if (td->io_ops->flags & FIO_RAWIO)
137			return 0;
138
139		buflen = td->io_size - td->this_io_bytes[ddir];
140	}
141
142	return buflen;
143}
144
145/*
146 * Return the data direction for the next io_u. If the job is a
147 * mixed read/write workload, check the rwmix cycle and switch if
148 * necessary.
149 */
150static int get_rw_ddir(struct thread_data *td)
151{
152	if (td_rw(td)) {
153		struct timeval now;
154		unsigned long elapsed;
155
156		gettimeofday(&now, NULL);
157	 	elapsed = mtime_since_now(&td->rwmix_switch);
158
159		/*
160		 * Check if it's time to seed a new data direction.
161		 */
162		if (elapsed >= td->rwmixcycle) {
163			unsigned int v;
164			long r;
165
166			r = os_random_long(&td->rwmix_state);
167			v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
168			if (v < td->rwmixread)
169				td->rwmix_ddir = DDIR_READ;
170			else
171				td->rwmix_ddir = DDIR_WRITE;
172			memcpy(&td->rwmix_switch, &now, sizeof(now));
173		}
174		return td->rwmix_ddir;
175	} else if (td_read(td))
176		return DDIR_READ;
177	else
178		return DDIR_WRITE;
179}
180
181void put_io_u(struct thread_data *td, struct io_u *io_u)
182{
183	io_u->file = NULL;
184	list_del(&io_u->list);
185	list_add(&io_u->list, &td->io_u_freelist);
186	td->cur_depth--;
187}
188
189static int fill_io_u(struct thread_data *td, struct fio_file *f,
190		     struct io_u *io_u)
191{
192	/*
193	 * If using an iolog, grab next piece if any available.
194	 */
195	if (td->read_iolog)
196		return read_iolog_get(td, io_u);
197
198	/*
199	 * see if it's time to sync
200	 */
201	if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
202	    && should_fsync(td)) {
203		io_u->ddir = DDIR_SYNC;
204		io_u->file = f;
205		return 0;
206	}
207
208	io_u->ddir = get_rw_ddir(td);
209
210	/*
211	 * No log, let the seq/rand engine retrieve the next position.
212	 */
213	if (!get_next_offset(td, f, &io_u->offset, io_u->ddir)) {
214		io_u->buflen = get_next_buflen(td, io_u->ddir);
215		if (io_u->buflen) {
216			/*
217			 * If using a write iolog, store this entry.
218			 */
219			if (td->write_iolog_file)
220				write_iolog_put(td, io_u);
221
222			io_u->file = f;
223			return 0;
224		}
225	}
226
227	return 1;
228}
229
230struct io_u *__get_io_u(struct thread_data *td)
231{
232	struct io_u *io_u = NULL;
233
234	if (!queue_full(td)) {
235		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
236
237		io_u->buflen = 0;
238		io_u->error = 0;
239		io_u->resid = 0;
240		list_del(&io_u->list);
241		list_add(&io_u->list, &td->io_u_busylist);
242		td->cur_depth++;
243	}
244
245	return io_u;
246}
247
248/*
249 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
250 * etc. The returned io_u is fully ready to be prepped and submitted.
251 */
252struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
253{
254	struct io_u *io_u;
255
256	io_u = __get_io_u(td);
257	if (!io_u)
258		return NULL;
259
260	if (td->zone_bytes >= td->zone_size) {
261		td->zone_bytes = 0;
262		f->last_pos += td->zone_skip;
263	}
264
265	if (fill_io_u(td, f, io_u)) {
266		put_io_u(td, io_u);
267		return NULL;
268	}
269
270	if (io_u->buflen + io_u->offset > f->file_size) {
271		if (td->io_ops->flags & FIO_RAWIO) {
272			put_io_u(td, io_u);
273			return NULL;
274		}
275
276		io_u->buflen = f->file_size - io_u->offset;
277	}
278
279	if (io_u->ddir != DDIR_SYNC) {
280		if (!io_u->buflen) {
281			put_io_u(td, io_u);
282			return NULL;
283		}
284
285		if (!td->read_iolog && !td->sequential && !td->norandommap)
286			mark_random_map(td, f, io_u);
287
288		f->last_pos += io_u->buflen;
289
290		if (td->verify != VERIFY_NONE)
291			populate_verify_io_u(td, io_u);
292	}
293
294	if (td_io_prep(td, io_u)) {
295		put_io_u(td, io_u);
296		return NULL;
297	}
298
299	gettimeofday(&io_u->start_time, NULL);
300	return io_u;
301}
302
303void io_completed(struct thread_data *td, struct io_u *io_u,
304		  struct io_completion_data *icd)
305{
306	struct timeval e;
307	unsigned long msec;
308
309	if (io_u->ddir == DDIR_SYNC) {
310		td->last_was_sync = 1;
311		return;
312	}
313
314	td->last_was_sync = 0;
315
316	gettimeofday(&e, NULL);
317
318	if (!io_u->error) {
319		unsigned int bytes = io_u->buflen - io_u->resid;
320		const int idx = io_u->ddir;
321
322		td->io_blocks[idx]++;
323		td->io_bytes[idx] += bytes;
324		td->zone_bytes += bytes;
325		td->this_io_bytes[idx] += bytes;
326
327		msec = mtime_since(&io_u->issue_time, &e);
328
329		add_clat_sample(td, idx, msec);
330		add_bw_sample(td, idx);
331
332		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
333			log_io_piece(td, io_u);
334
335		icd->bytes_done[idx] += bytes;
336	} else
337		icd->error = io_u->error;
338}
339
340void ios_completed(struct thread_data *td, struct io_completion_data *icd)
341{
342	struct io_u *io_u;
343	int i;
344
345	icd->error = 0;
346	icd->bytes_done[0] = icd->bytes_done[1] = 0;
347
348	for (i = 0; i < icd->nr; i++) {
349		io_u = td->io_ops->event(td, i);
350
351		io_completed(td, io_u, icd);
352		put_io_u(td, io_u);
353	}
354}
355
356
357