blob: 1845d3b81e09437d3fc7bc95ba3d366da9d73006 [file] [log] [blame]
Jens Axboe10ba5352006-10-20 11:39:27 +02001#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
Jens Axboe0c6e7512007-02-22 11:19:39 +01006#include <assert.h>
Jens Axboe10ba5352006-10-20 11:39:27 +02007
8#include "fio.h"
Jens Axboe5973caf2008-05-21 19:52:35 +02009#include "hash.h"
Jens Axboe4f5af7b2009-06-03 08:45:40 +020010#include "verify.h"
Jens Axboe10ba5352006-10-20 11:39:27 +020011
Jens Axboe97601022007-02-18 12:47:29 +010012struct io_completion_data {
13 int nr; /* input */
Jens Axboe97601022007-02-18 12:47:29 +010014
15 int error; /* output */
16 unsigned long bytes_done[2]; /* output */
17 struct timeval time; /* output */
18};
19
Jens Axboe10ba5352006-10-20 11:39:27 +020020/*
21 * The ->file_map[] contains a map of blocks we have or have not done io
22 * to yet. Used to make sure we cover the entire range in a fair fashion.
23 */
Jens Axboeaec2de22008-04-24 12:44:42 +020024static int random_map_free(struct fio_file *f, const unsigned long long block)
Jens Axboe10ba5352006-10-20 11:39:27 +020025{
Jens Axboeaec2de22008-04-24 12:44:42 +020026 unsigned int idx = RAND_MAP_IDX(f, block);
27 unsigned int bit = RAND_MAP_BIT(f, block);
Jens Axboe10ba5352006-10-20 11:39:27 +020028
Jens Axboe84422ac2008-02-19 20:10:26 +010029 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
30
Jens Axboede605662008-05-31 00:21:12 +020031 return (f->file_map[idx] & (1 << bit)) == 0;
Jens Axboe10ba5352006-10-20 11:39:27 +020032}
33
34/*
Jens Axboedf415582006-10-20 11:41:03 +020035 * Mark a given offset as used in the map.
36 */
Jens Axboe9bf20612007-03-01 09:33:57 +010037static void mark_random_map(struct thread_data *td, struct io_u *io_u)
Jens Axboedf415582006-10-20 11:41:03 +020038{
Jens Axboe2dc1bbe2007-03-15 15:01:33 +010039 unsigned int min_bs = td->o.rw_min_bs;
Jens Axboe9bf20612007-03-01 09:33:57 +010040 struct fio_file *f = io_u->file;
Jens Axboea00735e2006-11-03 08:58:08 +010041 unsigned long long block;
Jens Axboe3e3357b2008-06-02 09:53:05 +020042 unsigned int blocks, nr_blocks;
Jens Axboedf415582006-10-20 11:41:03 +020043
Jens Axboeb9c5b642008-02-18 14:13:08 +010044 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
Jens Axboec685b5b2007-02-10 20:02:28 +010045 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
Jens Axboe3e3357b2008-06-02 09:53:05 +020046 blocks = 0;
Jens Axboec685b5b2007-02-10 20:02:28 +010047
Jens Axboe3e3357b2008-06-02 09:53:05 +020048 while (nr_blocks) {
49 unsigned int this_blocks, mask;
Jens Axboedf415582006-10-20 11:41:03 +020050 unsigned int idx, bit;
51
Jens Axboe1e3d53a2007-03-22 19:01:48 +010052 /*
53 * If we have a mixed random workload, we may
54 * encounter blocks we already did IO to.
55 */
Jens Axboe8b113b42008-06-02 10:15:48 +020056 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
57 if (!blocks)
58 blocks = 1;
Jens Axboedf415582006-10-20 11:41:03 +020059 break;
Jens Axboe8b113b42008-06-02 10:15:48 +020060 }
Jens Axboedf415582006-10-20 11:41:03 +020061
Jens Axboeaec2de22008-04-24 12:44:42 +020062 idx = RAND_MAP_IDX(f, block);
63 bit = RAND_MAP_BIT(f, block);
Jens Axboedf415582006-10-20 11:41:03 +020064
Jens Axboe0032bf92007-02-13 17:39:56 +010065 fio_assert(td, idx < f->num_maps);
Jens Axboedf415582006-10-20 11:41:03 +020066
Jens Axboe3e3357b2008-06-02 09:53:05 +020067 this_blocks = nr_blocks;
68 if (this_blocks + bit > BLOCKS_PER_MAP)
69 this_blocks = BLOCKS_PER_MAP - bit;
70
71 if (this_blocks == BLOCKS_PER_MAP)
72 mask = -1U;
73 else
74 mask = ((1U << this_blocks) - 1) << bit;
75
Jens Axboe3e3357b2008-06-02 09:53:05 +020076 f->file_map[idx] |= mask;
77 nr_blocks -= this_blocks;
78 blocks += this_blocks;
Jens Axboec9dd34b2008-06-02 10:09:43 +020079 block += this_blocks;
Jens Axboedf415582006-10-20 11:41:03 +020080 }
81
Jens Axboea00735e2006-11-03 08:58:08 +010082 if ((blocks * min_bs) < io_u->buflen)
83 io_u->buflen = blocks * min_bs;
Jens Axboedf415582006-10-20 11:41:03 +020084}
85
Jens Axboe3e3357b2008-06-02 09:53:05 +020086static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir)
Jens Axboe2ba1c292008-02-01 13:16:38 +010088{
89 unsigned long long max_blocks;
Jens Axboed9dd70f2008-05-23 12:37:23 +020090 unsigned long long max_size;
Jens Axboe2ba1c292008-02-01 13:16:38 +010091
Jens Axboed9dd70f2008-05-23 12:37:23 +020092 /*
93 * Hmm, should we make sure that ->io_size <= ->real_file_size?
94 */
95 max_size = f->io_size;
96 if (max_size > f->real_file_size)
97 max_size = f->real_file_size;
98
Jens Axboe2b7a01d2009-03-11 11:00:13 +010099 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
Jens Axboe2ba1c292008-02-01 13:16:38 +0100100 if (!max_blocks)
101 return 0;
102
Jens Axboe67778e82008-05-15 09:20:08 +0200103 return max_blocks;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100104}
105
Jens Axboedf415582006-10-20 11:41:03 +0200106/*
Jens Axboe10ba5352006-10-20 11:39:27 +0200107 * Return the next free block in the map.
108 */
109static int get_next_free_block(struct thread_data *td, struct fio_file *f,
Jens Axboe4ba66132008-02-05 10:05:50 +0100110 enum fio_ddir ddir, unsigned long long *b)
Jens Axboe10ba5352006-10-20 11:39:27 +0200111{
Jens Axboe84422ac2008-02-19 20:10:26 +0100112 unsigned long long min_bs = td->o.rw_min_bs;
Jens Axboe10ba5352006-10-20 11:39:27 +0200113 int i;
114
Jens Axboec685b5b2007-02-10 20:02:28 +0100115 i = f->last_free_lookup;
116 *b = (i * BLOCKS_PER_MAP);
Radha Ramachandran899c29f2009-04-22 08:17:42 +0200117 while ((*b) * min_bs < f->real_file_size &&
118 (*b) * min_bs < f->io_size) {
Jens Axboede605662008-05-31 00:21:12 +0200119 if (f->file_map[i] != (unsigned int) -1) {
Jens Axboe697a6062008-05-31 00:04:45 +0200120 *b += ffz(f->file_map[i]);
Jens Axboe4ba66132008-02-05 10:05:50 +0100121 if (*b > last_block(td, f, ddir))
Jens Axboe2ba1c292008-02-01 13:16:38 +0100122 break;
Jens Axboec685b5b2007-02-10 20:02:28 +0100123 f->last_free_lookup = i;
Jens Axboe10ba5352006-10-20 11:39:27 +0200124 return 0;
125 }
126
127 *b += BLOCKS_PER_MAP;
128 i++;
129 }
130
Jens Axboe2ba1c292008-02-01 13:16:38 +0100131 dprint(FD_IO, "failed finding a free block\n");
Jens Axboe10ba5352006-10-20 11:39:27 +0200132 return 1;
133}
134
Jens Axboeec4015d2007-03-23 08:04:27 +0100135static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
Jens Axboe4ba66132008-02-05 10:05:50 +0100136 enum fio_ddir ddir, unsigned long long *b)
Jens Axboeec4015d2007-03-23 08:04:27 +0100137{
Jens Axboe84422ac2008-02-19 20:10:26 +0100138 unsigned long long r;
Jens Axboeec4015d2007-03-23 08:04:27 +0100139 int loops = 5;
140
141 do {
142 r = os_random_long(&td->random_state);
Jens Axboe84422ac2008-02-19 20:10:26 +0100143 dprint(FD_RANDOM, "off rand %llu\n", r);
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100144 *b = (last_block(td, f, ddir) - 1)
Jens Axboedc873b62008-06-04 20:13:04 +0200145 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
Jens Axboe2ba1c292008-02-01 13:16:38 +0100146
Jens Axboe43c63a72007-05-21 13:23:30 +0200147 /*
148 * if we are not maintaining a random map, we are done.
149 */
Jens Axboe303032a2008-03-26 10:11:10 +0100150 if (!file_randommap(td, f))
Jens Axboe43c63a72007-05-21 13:23:30 +0200151 return 0;
152
153 /*
Jens Axboe84422ac2008-02-19 20:10:26 +0100154 * calculate map offset and check if it's free
Jens Axboe43c63a72007-05-21 13:23:30 +0200155 */
Jens Axboeaec2de22008-04-24 12:44:42 +0200156 if (random_map_free(f, *b))
Jens Axboe43c63a72007-05-21 13:23:30 +0200157 return 0;
158
Jens Axboe84422ac2008-02-19 20:10:26 +0100159 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
160 *b);
Jens Axboe43c63a72007-05-21 13:23:30 +0200161 } while (--loops);
Jens Axboeec4015d2007-03-23 08:04:27 +0100162
163 /*
Jens Axboe43c63a72007-05-21 13:23:30 +0200164 * we get here, if we didn't suceed in looking up a block. generate
165 * a random start offset into the filemap, and find the first free
166 * block from there.
Jens Axboeec4015d2007-03-23 08:04:27 +0100167 */
Jens Axboe43c63a72007-05-21 13:23:30 +0200168 loops = 10;
169 do {
Jens Axboedc873b62008-06-04 20:13:04 +0200170 f->last_free_lookup = (f->num_maps - 1) *
171 (r / (OS_RAND_MAX + 1.0));
Jens Axboe4ba66132008-02-05 10:05:50 +0100172 if (!get_next_free_block(td, f, ddir, b))
Jens Axboe43c63a72007-05-21 13:23:30 +0200173 return 0;
Jens Axboeec4015d2007-03-23 08:04:27 +0100174
Jens Axboe43c63a72007-05-21 13:23:30 +0200175 r = os_random_long(&td->random_state);
176 } while (--loops);
177
178 /*
179 * that didn't work either, try exhaustive search from the start
180 */
181 f->last_free_lookup = 0;
Jens Axboe4ba66132008-02-05 10:05:50 +0100182 return get_next_free_block(td, f, ddir, b);
Jens Axboeec4015d2007-03-23 08:04:27 +0100183}
184
Jens Axboe10ba5352006-10-20 11:39:27 +0200185/*
186 * For random io, generate a random new block and see if it's used. Repeat
187 * until we find a free one. For sequential io, just return the end of
188 * the last io issued.
189 */
Jens Axboe9bf20612007-03-01 09:33:57 +0100190static int get_next_offset(struct thread_data *td, struct io_u *io_u)
Jens Axboe10ba5352006-10-20 11:39:27 +0200191{
Jens Axboe9bf20612007-03-01 09:33:57 +0100192 struct fio_file *f = io_u->file;
Jens Axboeec4015d2007-03-23 08:04:27 +0100193 unsigned long long b;
Jens Axboe4ba66132008-02-05 10:05:50 +0100194 enum fio_ddir ddir = io_u->ddir;
Jens Axboe10ba5352006-10-20 11:39:27 +0200195
Jens Axboeec4015d2007-03-23 08:04:27 +0100196 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
197 td->ddir_nr = td->o.ddir_nr;
Jens Axboe10ba5352006-10-20 11:39:27 +0200198
Jens Axboe683023e2009-03-04 08:41:31 +0100199 if (get_next_rand_offset(td, f, ddir, &b)) {
200 dprint(FD_IO, "%s: getting rand offset failed\n",
201 f->file_name);
Jens Axboebca4ed42007-02-12 05:13:23 +0100202 return 1;
Jens Axboe683023e2009-03-04 08:41:31 +0100203 }
Jens Axboe43063a12007-03-27 20:21:24 +0200204 } else {
ljzhang,Yaxin Hu,Jianchao Tang0c3d7682007-07-26 10:59:25 +0200205 if (f->last_pos >= f->real_file_size) {
Jens Axboe4ba66132008-02-05 10:05:50 +0100206 if (!td_random(td) ||
Jens Axboe683023e2009-03-04 08:41:31 +0100207 get_next_rand_offset(td, f, ddir, &b)) {
208 dprint(FD_IO, "%s: pos %llu > size %llu\n",
209 f->file_name, f->last_pos,
210 f->real_file_size);
ljzhang,Yaxin Hu,Jianchao Tang0c3d7682007-07-26 10:59:25 +0200211 return 1;
Jens Axboe683023e2009-03-04 08:41:31 +0100212 }
ljzhang,Yaxin Hu,Jianchao Tangbcdedd02007-07-27 13:28:26 +0200213 } else
Jens Axboe4ba66132008-02-05 10:05:50 +0100214 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
Jens Axboe43063a12007-03-27 20:21:24 +0200215 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200216
Jens Axboe2b7a01d2009-03-11 11:00:13 +0100217 io_u->offset = b * td->o.ba[ddir];
Jens Axboe009bd842008-05-15 10:19:46 +0200218 if (io_u->offset >= f->io_size) {
219 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
220 io_u->offset, f->io_size);
221 return 1;
222 }
223
224 io_u->offset += f->file_offset;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100225 if (io_u->offset >= f->real_file_size) {
226 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
227 io_u->offset, f->real_file_size);
Jens Axboe10ba5352006-10-20 11:39:27 +0200228 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100229 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200230
231 return 0;
232}
233
Jens Axboe9bf20612007-03-01 09:33:57 +0100234static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
Jens Axboe10ba5352006-10-20 11:39:27 +0200235{
Jens Axboebca4ed42007-02-12 05:13:23 +0100236 const int ddir = io_u->ddir;
Jens Axboef3f552b2008-06-13 10:27:05 +0200237 unsigned int uninitialized_var(buflen);
Jens Axboef3059de2008-06-11 15:37:32 +0200238 unsigned int minbs, maxbs;
Jens Axboe10ba5352006-10-20 11:39:27 +0200239 long r;
240
Jens Axboef3059de2008-06-11 15:37:32 +0200241 minbs = td->o.min_bs[ddir];
242 maxbs = td->o.max_bs[ddir];
243
244 if (minbs == maxbs)
245 buflen = minbs;
Jens Axboe10ba5352006-10-20 11:39:27 +0200246 else {
247 r = os_random_long(&td->bsrange_state);
Jens Axboe720e84a2009-04-21 08:29:55 +0200248 if (!td->o.bssplit_nr[ddir]) {
Jens Axboef3059de2008-06-11 15:37:32 +0200249 buflen = 1 + (unsigned int) ((double) maxbs *
250 (r / (OS_RAND_MAX + 1.0)));
251 if (buflen < minbs)
252 buflen = minbs;
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100253 } else {
Jens Axboe564ca972007-12-14 12:21:19 +0100254 long perc = 0;
255 unsigned int i;
256
Jens Axboe720e84a2009-04-21 08:29:55 +0200257 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
258 struct bssplit *bsp = &td->o.bssplit[ddir][i];
Jens Axboe564ca972007-12-14 12:21:19 +0100259
260 buflen = bsp->bs;
261 perc += bsp->perc;
Jens Axboef3059de2008-06-11 15:37:32 +0200262 if (r <= ((OS_RAND_MAX / 100L) * perc))
Jens Axboe564ca972007-12-14 12:21:19 +0100263 break;
264 }
265 }
Jens Axboef3059de2008-06-11 15:37:32 +0200266 if (!td->o.bs_unaligned && is_power_of_2(minbs))
267 buflen = (buflen + minbs - 1) & ~(minbs - 1);
Jens Axboe10ba5352006-10-20 11:39:27 +0200268 }
269
Jens Axboe4ba66132008-02-05 10:05:50 +0100270 if (io_u->offset + buflen > io_u->file->real_file_size) {
271 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
Jens Axboef3059de2008-06-11 15:37:32 +0200272 minbs, ddir);
273 buflen = minbs;
Jens Axboe4ba66132008-02-05 10:05:50 +0100274 }
Jens Axboe6a5e6882007-07-26 10:47:51 +0200275
Jens Axboe10ba5352006-10-20 11:39:27 +0200276 return buflen;
277}
278
Jens Axboeafe24a52007-03-16 20:27:27 +0100279static void set_rwmix_bytes(struct thread_data *td)
280{
Jens Axboeafe24a52007-03-16 20:27:27 +0100281 unsigned int diff;
282
283 /*
284 * we do time or byte based switch. this is needed because
285 * buffered writes may issue a lot quicker than they complete,
286 * whereas reads do not.
287 */
Jens Axboee47f7992007-03-21 14:05:39 +0100288 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
Jens Axboe04c540d2008-05-28 10:35:26 +0200289 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
Jens Axboee47f7992007-03-21 14:05:39 +0100290}
291
292static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
293{
294 unsigned int v;
295 long r;
296
297 r = os_random_long(&td->rwmix_state);
Jens Axboedc873b62008-06-04 20:13:04 +0200298 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
Jens Axboe04c540d2008-05-28 10:35:26 +0200299 if (v <= td->o.rwmix[DDIR_READ])
Jens Axboee47f7992007-03-21 14:05:39 +0100300 return DDIR_READ;
301
302 return DDIR_WRITE;
Jens Axboeafe24a52007-03-16 20:27:27 +0100303}
304
Jens Axboe581e7142009-06-09 12:47:16 +0200305static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
306{
307 enum fio_ddir odir = ddir ^ 1;
308 struct timeval t;
309 long usec;
310
311 if (td->rate_pending_usleep[ddir] <= 0)
312 return ddir;
313
314 /*
315 * We have too much pending sleep in this direction. See if we
316 * should switch.
317 */
318 if (td_rw(td)) {
319 /*
320 * Other direction does not have too much pending, switch
321 */
322 if (td->rate_pending_usleep[odir] < 100000)
323 return odir;
324
325 /*
326 * Both directions have pending sleep. Sleep the minimum time
327 * and deduct from both.
328 */
329 if (td->rate_pending_usleep[ddir] <=
330 td->rate_pending_usleep[odir]) {
331 usec = td->rate_pending_usleep[ddir];
332 } else {
333 usec = td->rate_pending_usleep[odir];
334 ddir = odir;
335 }
336 } else
337 usec = td->rate_pending_usleep[ddir];
338
339 fio_gettime(&t, NULL);
340 usec_sleep(td, usec);
341 usec = utime_since_now(&t);
342
343 td->rate_pending_usleep[ddir] -= usec;
344
345 odir = ddir ^ 1;
346 if (td_rw(td) && __should_check_rate(td, odir))
347 td->rate_pending_usleep[odir] -= usec;
Jens Axboe0b9d69e2009-09-11 22:29:54 +0200348
Jens Axboe581e7142009-06-09 12:47:16 +0200349 return ddir;
350}
351
Jens Axboe10ba5352006-10-20 11:39:27 +0200352/*
353 * Return the data direction for the next io_u. If the job is a
354 * mixed read/write workload, check the rwmix cycle and switch if
355 * necessary.
356 */
Jens Axboe1e97cce2006-12-05 11:44:16 +0100357static enum fio_ddir get_rw_ddir(struct thread_data *td)
Jens Axboe10ba5352006-10-20 11:39:27 +0200358{
Jens Axboe581e7142009-06-09 12:47:16 +0200359 enum fio_ddir ddir;
360
Jens Axboe5f9099e2009-06-16 22:40:26 +0200361 /*
362 * see if it's time to fsync
363 */
364 if (td->o.fsync_blocks &&
365 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
366 td->io_issues[DDIR_WRITE] && should_fsync(td))
367 return DDIR_SYNC;
368
369 /*
370 * see if it's time to fdatasync
371 */
372 if (td->o.fdatasync_blocks &&
373 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
374 td->io_issues[DDIR_WRITE] && should_fsync(td))
375 return DDIR_DATASYNC;
376
Jens Axboe10ba5352006-10-20 11:39:27 +0200377 if (td_rw(td)) {
Jens Axboe10ba5352006-10-20 11:39:27 +0200378 /*
379 * Check if it's time to seed a new data direction.
380 */
Jens Axboee4928662008-04-07 09:19:46 +0200381 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
Jens Axboee47f7992007-03-21 14:05:39 +0100382 /*
383 * Put a top limit on how many bytes we do for
384 * one data direction, to avoid overflowing the
385 * ranges too much
386 */
387 ddir = get_rand_ddir(td);
Jens Axboee47f7992007-03-21 14:05:39 +0100388
389 if (ddir != td->rwmix_ddir)
390 set_rwmix_bytes(td);
391
392 td->rwmix_ddir = ddir;
Jens Axboe10ba5352006-10-20 11:39:27 +0200393 }
Jens Axboe581e7142009-06-09 12:47:16 +0200394 ddir = td->rwmix_ddir;
Jens Axboe10ba5352006-10-20 11:39:27 +0200395 } else if (td_read(td))
Jens Axboe581e7142009-06-09 12:47:16 +0200396 ddir = DDIR_READ;
Jens Axboe10ba5352006-10-20 11:39:27 +0200397 else
Jens Axboe581e7142009-06-09 12:47:16 +0200398 ddir = DDIR_WRITE;
399
400 td->rwmix_ddir = rate_ddir(td, ddir);
401 return td->rwmix_ddir;
Jens Axboe10ba5352006-10-20 11:39:27 +0200402}
403
Jens Axboee8462bd2009-07-06 12:59:04 +0200404void put_file_log(struct thread_data *td, struct fio_file *f)
Jens Axboe60f2c652008-05-16 12:31:36 +0200405{
406 int ret = put_file(td, f);
407
408 if (ret)
409 td_verror(td, ret, "file close");
410}
411
Jens Axboe10ba5352006-10-20 11:39:27 +0200412void put_io_u(struct thread_data *td, struct io_u *io_u)
413{
Jens Axboee8462bd2009-07-06 12:59:04 +0200414 td_io_u_lock(td);
415
Jens Axboe0c6e7512007-02-22 11:19:39 +0100416 io_u->flags |= IO_U_F_FREE;
Jens Axboee8462bd2009-07-06 12:59:04 +0200417 io_u->flags &= ~IO_U_F_FREE_DEF;
Jens Axboe0c6e7512007-02-22 11:19:39 +0100418
Jens Axboe60f2c652008-05-16 12:31:36 +0200419 if (io_u->file)
420 put_file_log(td, io_u->file);
Jens Axboe2dbdab72007-04-26 10:21:15 +0200421
Jens Axboe10ba5352006-10-20 11:39:27 +0200422 io_u->file = NULL;
Radha Ramachandran0c412142009-11-03 21:45:31 +0100423 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
424 td->cur_depth--;
Jens Axboee8462bd2009-07-06 12:59:04 +0200425 flist_del_init(&io_u->list);
Jens Axboe01743ee2008-06-02 12:19:19 +0200426 flist_add(&io_u->list, &td->io_u_freelist);
Jens Axboee8462bd2009-07-06 12:59:04 +0200427 td_io_u_unlock(td);
428 td_io_u_free_notify(td);
Jens Axboe10ba5352006-10-20 11:39:27 +0200429}
430
Radha Ramachandranf2bba182009-06-15 08:40:16 +0200431void clear_io_u(struct thread_data *td, struct io_u *io_u)
432{
433 io_u->flags &= ~IO_U_F_FLIGHT;
434 put_io_u(td, io_u);
435}
436
Jens Axboe755200a2007-02-19 13:08:12 +0100437void requeue_io_u(struct thread_data *td, struct io_u **io_u)
438{
439 struct io_u *__io_u = *io_u;
440
Jens Axboe465221b2008-05-30 22:07:49 +0200441 dprint(FD_IO, "requeue %p\n", __io_u);
442
Jens Axboee8462bd2009-07-06 12:59:04 +0200443 td_io_u_lock(td);
444
Jens Axboe4d2e0f42007-02-24 13:31:57 +0100445 __io_u->flags |= IO_U_F_FREE;
Jens Axboe5f9099e2009-06-16 22:40:26 +0200446 if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir))
Jens Axboee4f54ad2008-02-04 10:56:07 +0100447 td->io_issues[__io_u->ddir]--;
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100448
Jens Axboe4d2e0f42007-02-24 13:31:57 +0100449 __io_u->flags &= ~IO_U_F_FLIGHT;
Radha Ramachandran0c412142009-11-03 21:45:31 +0100450 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
451 td->cur_depth--;
Jens Axboe01743ee2008-06-02 12:19:19 +0200452 flist_del(&__io_u->list);
453 flist_add_tail(&__io_u->list, &td->io_u_requeues);
Jens Axboee8462bd2009-07-06 12:59:04 +0200454 td_io_u_unlock(td);
Jens Axboe755200a2007-02-19 13:08:12 +0100455 *io_u = NULL;
456}
457
Jens Axboe9bf20612007-03-01 09:33:57 +0100458static int fill_io_u(struct thread_data *td, struct io_u *io_u)
Jens Axboe10ba5352006-10-20 11:39:27 +0200459{
Jens Axboeb4c5e1a2007-10-25 18:44:45 +0200460 if (td->io_ops->flags & FIO_NOIO)
461 goto out;
462
Jens Axboea00735e2006-11-03 08:58:08 +0100463 io_u->ddir = get_rw_ddir(td);
464
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200465 /*
Jens Axboe5f9099e2009-06-16 22:40:26 +0200466 * fsync() or fdatasync(), we are done
467 */
468 if (ddir_sync(io_u->ddir))
469 goto out;
470
471 /*
Jens Axboe48f5abd2007-07-20 13:25:04 +0200472 * See if it's time to switch to a new zone
473 */
474 if (td->zone_bytes >= td->o.zone_size) {
475 td->zone_bytes = 0;
476 io_u->file->last_pos += td->o.zone_skip;
477 td->io_skip_bytes += td->o.zone_skip;
478 }
479
480 /*
Jens Axboec685b5b2007-02-10 20:02:28 +0100481 * No log, let the seq/rand engine retrieve the next buflen and
482 * position.
Jens Axboe10ba5352006-10-20 11:39:27 +0200483 */
Jens Axboe2ba1c292008-02-01 13:16:38 +0100484 if (get_next_offset(td, io_u)) {
485 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
Jens Axboebca4ed42007-02-12 05:13:23 +0100486 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100487 }
Jens Axboec685b5b2007-02-10 20:02:28 +0100488
Jens Axboe9bf20612007-03-01 09:33:57 +0100489 io_u->buflen = get_next_buflen(td, io_u);
Jens Axboe2ba1c292008-02-01 13:16:38 +0100490 if (!io_u->buflen) {
491 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
Jens Axboebca4ed42007-02-12 05:13:23 +0100492 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100493 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200494
Jens Axboe2ba1c292008-02-01 13:16:38 +0100495 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
496 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
Jens Axboe4ba66132008-02-05 10:05:50 +0100497 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
498 io_u->buflen, io_u->file->real_file_size);
Jens Axboe6a5e6882007-07-26 10:47:51 +0200499 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100500 }
Jens Axboe6a5e6882007-07-26 10:47:51 +0200501
Jens Axboebca4ed42007-02-12 05:13:23 +0100502 /*
503 * mark entry before potentially trimming io_u
504 */
Jens Axboe303032a2008-03-26 10:11:10 +0100505 if (td_random(td) && file_randommap(td, io_u->file))
Jens Axboe9bf20612007-03-01 09:33:57 +0100506 mark_random_map(td, io_u);
Jens Axboe10ba5352006-10-20 11:39:27 +0200507
Jens Axboebca4ed42007-02-12 05:13:23 +0100508 /*
509 * If using a write iolog, store this entry.
510 */
Jens Axboec38e9462007-03-27 08:48:48 +0200511out:
Jens Axboe2ba1c292008-02-01 13:16:38 +0100512 dprint_io_u(io_u, "fill_io_u");
Jens Axboed9d91e32008-01-31 13:25:42 +0100513 td->zone_bytes += io_u->buflen;
Jens Axboef29b25a2007-07-23 08:56:43 +0200514 log_io_u(td, io_u);
Jens Axboebca4ed42007-02-12 05:13:23 +0100515 return 0;
Jens Axboe10ba5352006-10-20 11:39:27 +0200516}
517
Jens Axboe838bc702008-05-22 13:08:23 +0200518static void __io_u_mark_map(unsigned int *map, unsigned int nr)
519{
520 int index = 0;
521
522 switch (nr) {
523 default:
524 index = 6;
525 break;
526 case 33 ... 64:
527 index = 5;
528 break;
529 case 17 ... 32:
530 index = 4;
531 break;
532 case 9 ... 16:
533 index = 3;
534 break;
535 case 5 ... 8:
536 index = 2;
537 break;
538 case 1 ... 4:
539 index = 1;
540 case 0:
541 break;
542 }
543
544 map[index]++;
545}
546
547void io_u_mark_submit(struct thread_data *td, unsigned int nr)
548{
549 __io_u_mark_map(td->ts.io_u_submit, nr);
550 td->ts.total_submit++;
551}
552
553void io_u_mark_complete(struct thread_data *td, unsigned int nr)
554{
555 __io_u_mark_map(td->ts.io_u_complete, nr);
556 td->ts.total_complete++;
557}
558
Jens Axboed8005752008-05-15 09:49:09 +0200559void io_u_mark_depth(struct thread_data *td, unsigned int nr)
Jens Axboe71619dc2007-01-13 23:56:33 +0100560{
561 int index = 0;
562
563 switch (td->cur_depth) {
564 default:
Jens Axboea783e612007-06-19 09:50:28 +0200565 index = 6;
566 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100567 case 32 ... 63:
Jens Axboea783e612007-06-19 09:50:28 +0200568 index = 5;
569 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100570 case 16 ... 31:
Jens Axboea783e612007-06-19 09:50:28 +0200571 index = 4;
572 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100573 case 8 ... 15:
Jens Axboea783e612007-06-19 09:50:28 +0200574 index = 3;
575 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100576 case 4 ... 7:
Jens Axboea783e612007-06-19 09:50:28 +0200577 index = 2;
578 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100579 case 2 ... 3:
Jens Axboea783e612007-06-19 09:50:28 +0200580 index = 1;
Jens Axboe71619dc2007-01-13 23:56:33 +0100581 case 1:
582 break;
583 }
584
Jens Axboe3bec7ae2008-05-14 21:08:37 +0200585 td->ts.io_u_map[index] += nr;
Jens Axboe71619dc2007-01-13 23:56:33 +0100586}
587
Jens Axboe04a0fea2007-06-19 12:48:41 +0200588static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
589{
590 int index = 0;
591
592 assert(usec < 1000);
593
594 switch (usec) {
595 case 750 ... 999:
596 index = 9;
597 break;
598 case 500 ... 749:
599 index = 8;
600 break;
601 case 250 ... 499:
602 index = 7;
603 break;
604 case 100 ... 249:
605 index = 6;
606 break;
607 case 50 ... 99:
608 index = 5;
609 break;
610 case 20 ... 49:
611 index = 4;
612 break;
613 case 10 ... 19:
614 index = 3;
615 break;
616 case 4 ... 9:
617 index = 2;
618 break;
619 case 2 ... 3:
620 index = 1;
621 case 0 ... 1:
622 break;
623 }
624
625 assert(index < FIO_IO_U_LAT_U_NR);
626 td->ts.io_u_lat_u[index]++;
627}
628
629static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
Jens Axboeec118302007-02-17 04:38:20 +0100630{
631 int index = 0;
632
633 switch (msec) {
634 default:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200635 index = 11;
636 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100637 case 1000 ... 1999:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200638 index = 10;
639 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100640 case 750 ... 999:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200641 index = 9;
642 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100643 case 500 ... 749:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200644 index = 8;
645 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100646 case 250 ... 499:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200647 index = 7;
648 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100649 case 100 ... 249:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200650 index = 6;
651 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100652 case 50 ... 99:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200653 index = 5;
654 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100655 case 20 ... 49:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200656 index = 4;
657 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100658 case 10 ... 19:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200659 index = 3;
660 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100661 case 4 ... 9:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200662 index = 2;
663 break;
Jens Axboeec118302007-02-17 04:38:20 +0100664 case 2 ... 3:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200665 index = 1;
Jens Axboeec118302007-02-17 04:38:20 +0100666 case 0 ... 1:
667 break;
668 }
669
Jens Axboe04a0fea2007-06-19 12:48:41 +0200670 assert(index < FIO_IO_U_LAT_M_NR);
671 td->ts.io_u_lat_m[index]++;
672}
673
674static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
675{
676 if (usec < 1000)
677 io_u_mark_lat_usec(td, usec);
678 else
679 io_u_mark_lat_msec(td, usec / 1000);
Jens Axboeec118302007-02-17 04:38:20 +0100680}
681
Jens Axboe0aabe162007-02-23 08:45:55 +0100682/*
683 * Get next file to service by choosing one at random
684 */
Jens Axboe2cc52932009-06-09 14:14:20 +0200685static struct fio_file *get_next_file_rand(struct thread_data *td,
686 enum fio_file_flags goodf,
Jens Axboed6aed792009-06-03 08:41:15 +0200687 enum fio_file_flags badf)
Jens Axboe0aabe162007-02-23 08:45:55 +0100688{
Jens Axboe0aabe162007-02-23 08:45:55 +0100689 struct fio_file *f;
Jens Axboe1c178182007-03-13 13:25:18 +0100690 int fno;
Jens Axboe0aabe162007-02-23 08:45:55 +0100691
692 do {
Jens Axboe7c83c082007-03-01 10:04:15 +0100693 long r = os_random_long(&td->next_file_state);
Jens Axboe87b10672009-03-04 09:39:47 +0100694 int opened = 0;
Jens Axboe7c83c082007-03-01 10:04:15 +0100695
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100696 fno = (unsigned int) ((double) td->o.nr_files
Jens Axboedc873b62008-06-04 20:13:04 +0200697 * (r / (OS_RAND_MAX + 1.0)));
Jens Axboe126d65c2008-03-01 18:04:31 +0100698 f = td->files[fno];
Jens Axboed6aed792009-06-03 08:41:15 +0200699 if (fio_file_done(f))
Jens Axboe059e63c2007-03-27 20:34:47 +0200700 continue;
Jens Axboe1c178182007-03-13 13:25:18 +0100701
Jens Axboed6aed792009-06-03 08:41:15 +0200702 if (!fio_file_open(f)) {
Jens Axboe87b10672009-03-04 09:39:47 +0100703 int err;
704
705 err = td_io_open_file(td, f);
706 if (err)
707 continue;
708 opened = 1;
709 }
710
Jens Axboe2ba1c292008-02-01 13:16:38 +0100711 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
712 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
Jens Axboe0aabe162007-02-23 08:45:55 +0100713 return f;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100714 }
Jens Axboe87b10672009-03-04 09:39:47 +0100715 if (opened)
716 td_io_close_file(td, f);
Jens Axboe0aabe162007-02-23 08:45:55 +0100717 } while (1);
718}
719
720/*
721 * Get next file to service by doing round robin between all available ones
722 */
Jens Axboe1c178182007-03-13 13:25:18 +0100723static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
724 int badf)
Jens Axboe3d7c3912007-02-19 13:16:12 +0100725{
726 unsigned int old_next_file = td->next_file;
727 struct fio_file *f;
728
729 do {
Jens Axboe87b10672009-03-04 09:39:47 +0100730 int opened = 0;
731
Jens Axboe126d65c2008-03-01 18:04:31 +0100732 f = td->files[td->next_file];
Jens Axboe3d7c3912007-02-19 13:16:12 +0100733
734 td->next_file++;
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100735 if (td->next_file >= td->o.nr_files)
Jens Axboe3d7c3912007-02-19 13:16:12 +0100736 td->next_file = 0;
737
Jens Axboe87b10672009-03-04 09:39:47 +0100738 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
Jens Axboed6aed792009-06-03 08:41:15 +0200739 if (fio_file_done(f)) {
Jens Axboed5ed68e2007-03-28 09:33:43 +0200740 f = NULL;
Jens Axboe059e63c2007-03-27 20:34:47 +0200741 continue;
Jens Axboed5ed68e2007-03-28 09:33:43 +0200742 }
Jens Axboe059e63c2007-03-27 20:34:47 +0200743
Jens Axboed6aed792009-06-03 08:41:15 +0200744 if (!fio_file_open(f)) {
Jens Axboe87b10672009-03-04 09:39:47 +0100745 int err;
746
747 err = td_io_open_file(td, f);
Jens Axboeb5696bf2009-03-04 16:03:49 +0100748 if (err) {
749 dprint(FD_FILE, "error %d on open of %s\n",
750 err, f->file_name);
Jens Axboe87c27b42009-05-20 10:45:12 +0200751 f = NULL;
Jens Axboe87b10672009-03-04 09:39:47 +0100752 continue;
Jens Axboeb5696bf2009-03-04 16:03:49 +0100753 }
Jens Axboe87b10672009-03-04 09:39:47 +0100754 opened = 1;
755 }
756
Jens Axboe0b9d69e2009-09-11 22:29:54 +0200757 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
758 f->flags);
Jens Axboe1c178182007-03-13 13:25:18 +0100759 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
Jens Axboe3d7c3912007-02-19 13:16:12 +0100760 break;
761
Jens Axboe87b10672009-03-04 09:39:47 +0100762 if (opened)
763 td_io_close_file(td, f);
764
Jens Axboe3d7c3912007-02-19 13:16:12 +0100765 f = NULL;
766 } while (td->next_file != old_next_file);
767
Jens Axboe2ba1c292008-02-01 13:16:38 +0100768 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
Jens Axboe3d7c3912007-02-19 13:16:12 +0100769 return f;
770}
771
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100772static struct fio_file *get_next_file(struct thread_data *td)
773{
Jens Axboe1907dbc2007-03-12 11:44:28 +0100774 struct fio_file *f;
775
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100776 assert(td->o.nr_files <= td->files_index);
Jens Axboe1c178182007-03-13 13:25:18 +0100777
Jens Axboeb5696bf2009-03-04 16:03:49 +0100778 if (td->nr_done_files >= td->o.nr_files) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100779 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
780 " nr_files=%d\n", td->nr_open_files,
781 td->nr_done_files,
782 td->o.nr_files);
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100783 return NULL;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100784 }
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100785
Jens Axboe1907dbc2007-03-12 11:44:28 +0100786 f = td->file_service_file;
Jens Axboed6aed792009-06-03 08:41:15 +0200787 if (f && fio_file_open(f) && !fio_file_closing(f)) {
Jens Axboea086c252009-03-04 08:27:37 +0100788 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
789 goto out;
790 if (td->file_service_left--)
791 goto out;
792 }
Jens Axboe1907dbc2007-03-12 11:44:28 +0100793
Jens Axboea086c252009-03-04 08:27:37 +0100794 if (td->o.file_service_type == FIO_FSERVICE_RR ||
795 td->o.file_service_type == FIO_FSERVICE_SEQ)
Jens Axboed6aed792009-06-03 08:41:15 +0200796 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100797 else
Jens Axboed6aed792009-06-03 08:41:15 +0200798 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
Jens Axboe1907dbc2007-03-12 11:44:28 +0100799
800 td->file_service_file = f;
801 td->file_service_left = td->file_service_nr - 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100802out:
Jens Axboe683023e2009-03-04 08:41:31 +0100803 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
Jens Axboe1907dbc2007-03-12 11:44:28 +0100804 return f;
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100805}
806
Jens Axboe429f6672007-07-23 10:38:43 +0200807static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
808{
809 struct fio_file *f;
810
811 do {
812 f = get_next_file(td);
813 if (!f)
814 return 1;
815
Jens Axboe429f6672007-07-23 10:38:43 +0200816 io_u->file = f;
817 get_file(f);
818
819 if (!fill_io_u(td, io_u))
820 break;
821
Jens Axboeb5696bf2009-03-04 16:03:49 +0100822 put_file_log(td, f);
Jens Axboe429f6672007-07-23 10:38:43 +0200823 td_io_close_file(td, f);
Jens Axboeb5696bf2009-03-04 16:03:49 +0100824 io_u->file = NULL;
Jens Axboed6aed792009-06-03 08:41:15 +0200825 fio_file_set_done(f);
Jens Axboe429f6672007-07-23 10:38:43 +0200826 td->nr_done_files++;
Jens Axboe0b9d69e2009-09-11 22:29:54 +0200827 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
828 td->nr_done_files, td->o.nr_files);
Jens Axboe429f6672007-07-23 10:38:43 +0200829 } while (1);
830
831 return 0;
832}
833
834
Jens Axboe10ba5352006-10-20 11:39:27 +0200835struct io_u *__get_io_u(struct thread_data *td)
836{
837 struct io_u *io_u = NULL;
838
Jens Axboee8462bd2009-07-06 12:59:04 +0200839 td_io_u_lock(td);
840
841again:
Jens Axboe01743ee2008-06-02 12:19:19 +0200842 if (!flist_empty(&td->io_u_requeues))
843 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
Jens Axboe755200a2007-02-19 13:08:12 +0100844 else if (!queue_full(td)) {
Jens Axboe01743ee2008-06-02 12:19:19 +0200845 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
Jens Axboe10ba5352006-10-20 11:39:27 +0200846
Jens Axboe6040dab2006-10-24 19:38:15 +0200847 io_u->buflen = 0;
Jens Axboe10ba5352006-10-20 11:39:27 +0200848 io_u->resid = 0;
Jens Axboe755200a2007-02-19 13:08:12 +0100849 io_u->file = NULL;
Jens Axboed7762cf2007-02-23 12:34:57 +0100850 io_u->end_io = NULL;
Jens Axboe755200a2007-02-19 13:08:12 +0100851 }
852
Jens Axboee8462bd2009-07-06 12:59:04 +0200853 /*
854 * We ran out, wait for async verify threads to finish and return one
855 */
856 if (!io_u && td->o.verify_async) {
857 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
858 goto again;
859 }
860
Jens Axboe755200a2007-02-19 13:08:12 +0100861 if (io_u) {
Jens Axboe0c6e7512007-02-22 11:19:39 +0100862 assert(io_u->flags & IO_U_F_FREE);
Jens Axboe2ecc1b52009-11-04 20:58:09 +0100863 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
Jens Axboe0c6e7512007-02-22 11:19:39 +0100864
Jens Axboe755200a2007-02-19 13:08:12 +0100865 io_u->error = 0;
Jens Axboe01743ee2008-06-02 12:19:19 +0200866 flist_del(&io_u->list);
867 flist_add(&io_u->list, &td->io_u_busylist);
Jens Axboe10ba5352006-10-20 11:39:27 +0200868 td->cur_depth++;
Radha Ramachandran0c412142009-11-03 21:45:31 +0100869 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
Jens Axboe10ba5352006-10-20 11:39:27 +0200870 }
871
Jens Axboee8462bd2009-07-06 12:59:04 +0200872 td_io_u_unlock(td);
Jens Axboe10ba5352006-10-20 11:39:27 +0200873 return io_u;
874}
875
876/*
877 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
878 * etc. The returned io_u is fully ready to be prepped and submitted.
879 */
Jens Axboe3d7c3912007-02-19 13:16:12 +0100880struct io_u *get_io_u(struct thread_data *td)
Jens Axboe10ba5352006-10-20 11:39:27 +0200881{
Jens Axboe3d7c3912007-02-19 13:16:12 +0100882 struct fio_file *f;
Jens Axboe10ba5352006-10-20 11:39:27 +0200883 struct io_u *io_u;
884
885 io_u = __get_io_u(td);
Jens Axboe2ba1c292008-02-01 13:16:38 +0100886 if (!io_u) {
887 dprint(FD_IO, "__get_io_u failed\n");
Jens Axboe10ba5352006-10-20 11:39:27 +0200888 return NULL;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100889 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200890
Jens Axboe755200a2007-02-19 13:08:12 +0100891 /*
892 * from a requeue, io_u already setup
893 */
894 if (io_u->file)
Jens Axboe77f392b2007-02-19 20:13:09 +0100895 goto out;
Jens Axboe755200a2007-02-19 13:08:12 +0100896
Jens Axboe429f6672007-07-23 10:38:43 +0200897 /*
898 * If using an iolog, grab next piece if any available.
899 */
900 if (td->o.read_iolog_file) {
901 if (read_iolog_get(td, io_u))
902 goto err_put;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100903 } else if (set_io_u_file(td, io_u)) {
904 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
Jens Axboe429f6672007-07-23 10:38:43 +0200905 goto err_put;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100906 }
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100907
Jens Axboe429f6672007-07-23 10:38:43 +0200908 f = io_u->file;
Jens Axboed6aed792009-06-03 08:41:15 +0200909 assert(fio_file_open(f));
Jens Axboe97af62c2007-05-22 11:12:13 +0200910
Jens Axboe5f9099e2009-06-16 22:40:26 +0200911 if (!ddir_sync(io_u->ddir)) {
Jens Axboed0656a92008-02-01 18:33:23 +0100912 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
Jens Axboe2ba1c292008-02-01 13:16:38 +0100913 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
Jens Axboe429f6672007-07-23 10:38:43 +0200914 goto err_put;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100915 }
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200916
Jens Axboe36167d82007-02-18 05:41:31 +0100917 f->last_pos = io_u->offset + io_u->buflen;
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200918
Radha Ramachandranc311cd22009-04-28 08:12:56 +0200919 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200920 populate_verify_io_u(td, io_u);
Jens Axboee4dad9c2008-05-28 10:53:44 +0200921 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
922 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
Jens Axboe10ba5352006-10-20 11:39:27 +0200923 }
924
Jens Axboe165faf12007-02-07 11:30:37 +0100925 /*
926 * Set io data pointers.
927 */
Jens Axboecec6b552007-02-06 20:15:38 +0100928 io_u->xfer_buf = io_u->buf;
929 io_u->xfer_buflen = io_u->buflen;
Jens Axboe5973caf2008-05-21 19:52:35 +0200930
Jens Axboe6ac7a332008-03-01 15:22:32 +0100931out:
Jens Axboe429f6672007-07-23 10:38:43 +0200932 if (!td_io_prep(td, io_u)) {
Jens Axboe993bf482008-11-14 13:04:53 +0100933 if (!td->o.disable_slat)
934 fio_gettime(&io_u->start_time, NULL);
Jens Axboe429f6672007-07-23 10:38:43 +0200935 return io_u;
Jens Axboe36167d82007-02-18 05:41:31 +0100936 }
Jens Axboe429f6672007-07-23 10:38:43 +0200937err_put:
Jens Axboe2ba1c292008-02-01 13:16:38 +0100938 dprint(FD_IO, "get_io_u failed\n");
Jens Axboe429f6672007-07-23 10:38:43 +0200939 put_io_u(td, io_u);
940 return NULL;
Jens Axboe10ba5352006-10-20 11:39:27 +0200941}
942
Jens Axboe54517922007-03-05 10:06:06 +0100943void io_u_log_error(struct thread_data *td, struct io_u *io_u)
944{
945 const char *msg[] = { "read", "write", "sync" };
946
947 log_err("fio: io_u error");
948
949 if (io_u->file)
950 log_err(" on file %s", io_u->file->file_name);
951
952 log_err(": %s\n", strerror(io_u->error));
953
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100954 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
955 io_u->offset, io_u->xfer_buflen);
Jens Axboe54517922007-03-05 10:06:06 +0100956
957 if (!td->error)
958 td_verror(td, io_u->error, "io_u error");
959}
960
Jens Axboe97601022007-02-18 12:47:29 +0100961static void io_completed(struct thread_data *td, struct io_u *io_u,
962 struct io_completion_data *icd)
Jens Axboe10ba5352006-10-20 11:39:27 +0200963{
Jens Axboedbad30b2008-12-05 13:17:16 +0100964 /*
965 * Older gcc's are too dumb to realize that usec is always used
966 * initialized, silence that warning.
967 */
968 unsigned long uninitialized_var(usec);
Jens Axboe10ba5352006-10-20 11:39:27 +0200969
Jens Axboe2ba1c292008-02-01 13:16:38 +0100970 dprint_io_u(io_u, "io complete");
971
Jens Axboe2ecc1b52009-11-04 20:58:09 +0100972 td_io_u_lock(td);
Jens Axboe0c6e7512007-02-22 11:19:39 +0100973 assert(io_u->flags & IO_U_F_FLIGHT);
974 io_u->flags &= ~IO_U_F_FLIGHT;
Jens Axboe2ecc1b52009-11-04 20:58:09 +0100975 td_io_u_unlock(td);
Jens Axboe0c6e7512007-02-22 11:19:39 +0100976
Jens Axboe5f9099e2009-06-16 22:40:26 +0200977 if (ddir_sync(io_u->ddir)) {
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200978 td->last_was_sync = 1;
979 return;
980 }
981
982 td->last_was_sync = 0;
983
Jens Axboe10ba5352006-10-20 11:39:27 +0200984 if (!io_u->error) {
985 unsigned int bytes = io_u->buflen - io_u->resid;
Jens Axboe1e97cce2006-12-05 11:44:16 +0100986 const enum fio_ddir idx = io_u->ddir;
Radha Ramachandranba3e4e02009-12-09 22:31:44 +0100987 const enum fio_ddir odx = io_u->ddir ^ 1;
Jens Axboeb29ee5b2008-09-11 10:17:26 +0200988 int ret;
Jens Axboe10ba5352006-10-20 11:39:27 +0200989
Jens Axboeb29ee5b2008-09-11 10:17:26 +0200990 td->io_blocks[idx]++;
991 td->io_bytes[idx] += bytes;
992 td->this_io_bytes[idx] += bytes;
Jens Axboe10ba5352006-10-20 11:39:27 +0200993
Jens Axboeb29ee5b2008-09-11 10:17:26 +0200994 if (ramp_time_over(td)) {
Jens Axboe40e1a6f2009-06-11 10:55:39 +0200995 unsigned long uninitialized_var(lusec);
Jens Axboe40e1a6f2009-06-11 10:55:39 +0200996
997 if (!td->o.disable_clat || !td->o.disable_bw)
998 lusec = utime_since(&io_u->issue_time,
999 &icd->time);
Jens Axboe721938a2008-09-10 09:46:16 +02001000
Jens Axboe9520ebb2008-10-16 21:03:27 +02001001 if (!td->o.disable_clat) {
Jens Axboe01a1caa2009-07-10 08:47:57 +02001002 add_clat_sample(td, idx, lusec, bytes);
Jens Axboe40e1a6f2009-06-11 10:55:39 +02001003 io_u_mark_latency(td, lusec);
Jens Axboe9520ebb2008-10-16 21:03:27 +02001004 }
1005 if (!td->o.disable_bw)
Jens Axboe306ddc92009-05-18 13:08:12 +02001006 add_bw_sample(td, idx, bytes, &icd->time);
Jens Axboeb23b6a22009-06-11 22:06:23 +02001007 if (__should_check_rate(td, idx)) {
Radha Ramachandranba3e4e02009-12-09 22:31:44 +01001008 td->rate_pending_usleep[idx] =
1009 ((td->this_io_bytes[idx] *
1010 td->rate_nsec_cycle[idx]) / 1000 -
1011 utime_since_now(&td->start));
Jens Axboeb23b6a22009-06-11 22:06:23 +02001012 }
Jens Axboe581e7142009-06-09 12:47:16 +02001013 if (__should_check_rate(td, idx ^ 1))
Radha Ramachandranba3e4e02009-12-09 22:31:44 +01001014 td->rate_pending_usleep[odx] =
1015 ((td->this_io_bytes[odx] *
1016 td->rate_nsec_cycle[odx]) / 1000 -
1017 utime_since_now(&td->start));
Jens Axboe721938a2008-09-10 09:46:16 +02001018 }
Jens Axboe10ba5352006-10-20 11:39:27 +02001019
Jens Axboe660a1cb2007-04-17 15:37:47 +02001020 if (td_write(td) && idx == DDIR_WRITE &&
Shawn Lewis86705792007-11-21 09:35:41 +01001021 td->o.do_verify &&
Jens Axboe41128402007-03-27 08:32:48 +02001022 td->o.verify != VERIFY_NONE)
Jens Axboe10ba5352006-10-20 11:39:27 +02001023 log_io_piece(td, io_u);
1024
Jens Axboeb29ee5b2008-09-11 10:17:26 +02001025 icd->bytes_done[idx] += bytes;
Jens Axboe3af6ef32007-02-18 06:57:43 +01001026
Jens Axboed7762cf2007-02-23 12:34:57 +01001027 if (io_u->end_io) {
Jens Axboe36690c92007-03-26 10:23:34 +02001028 ret = io_u->end_io(td, io_u);
Jens Axboe3af6ef32007-02-18 06:57:43 +01001029 if (ret && !icd->error)
1030 icd->error = ret;
1031 }
Jens Axboe54517922007-03-05 10:06:06 +01001032 } else {
Jens Axboe10ba5352006-10-20 11:39:27 +02001033 icd->error = io_u->error;
Jens Axboe54517922007-03-05 10:06:06 +01001034 io_u_log_error(td, io_u);
1035 }
Radha Ramachandranf2bba182009-06-15 08:40:16 +02001036 if (td->o.continue_on_error && icd->error &&
1037 td_non_fatal_error(icd->error)) {
1038 /*
1039 * If there is a non_fatal error, then add to the error count
1040 * and clear all the errors.
1041 */
1042 update_error_count(td, icd->error);
1043 td_clear_error(td);
1044 icd->error = 0;
1045 io_u->error = 0;
1046 }
Jens Axboe10ba5352006-10-20 11:39:27 +02001047}
1048
Jens Axboe9520ebb2008-10-16 21:03:27 +02001049static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1050 int nr)
Jens Axboe36167d82007-02-18 05:41:31 +01001051{
Jens Axboe9520ebb2008-10-16 21:03:27 +02001052 if (!td->o.disable_clat || !td->o.disable_bw)
1053 fio_gettime(&icd->time, NULL);
Jens Axboe36167d82007-02-18 05:41:31 +01001054
Jens Axboe3af6ef32007-02-18 06:57:43 +01001055 icd->nr = nr;
1056
Jens Axboe36167d82007-02-18 05:41:31 +01001057 icd->error = 0;
1058 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1059}
1060
Jens Axboe97601022007-02-18 12:47:29 +01001061static void ios_completed(struct thread_data *td,
1062 struct io_completion_data *icd)
Jens Axboe10ba5352006-10-20 11:39:27 +02001063{
1064 struct io_u *io_u;
1065 int i;
1066
Jens Axboe10ba5352006-10-20 11:39:27 +02001067 for (i = 0; i < icd->nr; i++) {
1068 io_u = td->io_ops->event(td, i);
1069
1070 io_completed(td, io_u, icd);
Jens Axboee8462bd2009-07-06 12:59:04 +02001071
1072 if (!(io_u->flags & IO_U_F_FREE_DEF))
1073 put_io_u(td, io_u);
Jens Axboe10ba5352006-10-20 11:39:27 +02001074 }
1075}
Jens Axboe97601022007-02-18 12:47:29 +01001076
Jens Axboee7e6cfb2007-02-20 10:58:34 +01001077/*
1078 * Complete a single io_u for the sync engines.
1079 */
Jens Axboe581e7142009-06-09 12:47:16 +02001080int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1081 unsigned long *bytes)
Jens Axboe97601022007-02-18 12:47:29 +01001082{
1083 struct io_completion_data icd;
1084
Jens Axboe9520ebb2008-10-16 21:03:27 +02001085 init_icd(td, &icd, 1);
Jens Axboe97601022007-02-18 12:47:29 +01001086 io_completed(td, io_u, &icd);
Jens Axboee8462bd2009-07-06 12:59:04 +02001087
1088 if (!(io_u->flags & IO_U_F_FREE_DEF))
1089 put_io_u(td, io_u);
Jens Axboe97601022007-02-18 12:47:29 +01001090
Jens Axboe581e7142009-06-09 12:47:16 +02001091 if (icd.error) {
1092 td_verror(td, icd.error, "io_u_sync_complete");
1093 return -1;
1094 }
Jens Axboe97601022007-02-18 12:47:29 +01001095
Jens Axboe581e7142009-06-09 12:47:16 +02001096 if (bytes) {
1097 bytes[0] += icd.bytes_done[0];
1098 bytes[1] += icd.bytes_done[1];
1099 }
1100
1101 return 0;
Jens Axboe97601022007-02-18 12:47:29 +01001102}
1103
Jens Axboee7e6cfb2007-02-20 10:58:34 +01001104/*
1105 * Called to complete min_events number of io for the async engines.
1106 */
Jens Axboe581e7142009-06-09 12:47:16 +02001107int io_u_queued_complete(struct thread_data *td, int min_evts,
1108 unsigned long *bytes)
Jens Axboe97601022007-02-18 12:47:29 +01001109{
Jens Axboe97601022007-02-18 12:47:29 +01001110 struct io_completion_data icd;
Jens Axboe00de55e2007-02-20 10:45:57 +01001111 struct timespec *tvp = NULL;
Jens Axboe97601022007-02-18 12:47:29 +01001112 int ret;
Davide Libenzi4d06a332007-03-22 07:43:50 +01001113 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
Jens Axboe97601022007-02-18 12:47:29 +01001114
Jens Axboe49504212008-06-05 09:03:30 +02001115 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
Jens Axboeb271fe62008-02-04 10:49:41 +01001116
Jens Axboe49504212008-06-05 09:03:30 +02001117 if (!min_evts)
Jens Axboe00de55e2007-02-20 10:45:57 +01001118 tvp = &ts;
Jens Axboe97601022007-02-18 12:47:29 +01001119
Jens Axboe49504212008-06-05 09:03:30 +02001120 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
Jens Axboe97601022007-02-18 12:47:29 +01001121 if (ret < 0) {
Jens Axboee1161c32007-02-22 19:36:48 +01001122 td_verror(td, -ret, "td_io_getevents");
Jens Axboe97601022007-02-18 12:47:29 +01001123 return ret;
1124 } else if (!ret)
1125 return ret;
1126
Jens Axboe9520ebb2008-10-16 21:03:27 +02001127 init_icd(td, &icd, ret);
Jens Axboe97601022007-02-18 12:47:29 +01001128 ios_completed(td, &icd);
Jens Axboe581e7142009-06-09 12:47:16 +02001129 if (icd.error) {
1130 td_verror(td, icd.error, "io_u_queued_complete");
1131 return -1;
1132 }
Jens Axboe97601022007-02-18 12:47:29 +01001133
Jens Axboe581e7142009-06-09 12:47:16 +02001134 if (bytes) {
1135 bytes[0] += icd.bytes_done[0];
1136 bytes[1] += icd.bytes_done[1];
1137 }
1138
1139 return 0;
Jens Axboe97601022007-02-18 12:47:29 +01001140}
Jens Axboe7e77dd02007-02-20 10:57:34 +01001141
1142/*
1143 * Call when io_u is really queued, to update the submission latency.
1144 */
1145void io_u_queued(struct thread_data *td, struct io_u *io_u)
1146{
Jens Axboe9520ebb2008-10-16 21:03:27 +02001147 if (!td->o.disable_slat) {
1148 unsigned long slat_time;
Jens Axboe7e77dd02007-02-20 10:57:34 +01001149
Jens Axboe9520ebb2008-10-16 21:03:27 +02001150 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
Jens Axboe29a90dd2009-06-10 06:57:47 +02001151 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
Jens Axboe9520ebb2008-10-16 21:03:27 +02001152 }
Jens Axboe7e77dd02007-02-20 10:57:34 +01001153}
Jens Axboe433afcb2007-02-22 10:39:01 +01001154
Jens Axboe5973caf2008-05-21 19:52:35 +02001155/*
1156 * "randomly" fill the buffer contents
1157 */
1158void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1159 unsigned int max_bs)
1160{
1161 long *ptr = io_u->buf;
1162
1163 if (!td->o.zero_buffers) {
1164 while ((void *) ptr - io_u->buf < max_bs) {
1165 *ptr = rand() * GOLDEN_RATIO_PRIME;
1166 ptr++;
1167 }
1168 } else
1169 memset(ptr, 0, max_bs);
1170}