blob: 1dff88f1edf64a5a10d9ea296ee1f5ae51e72288 [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 Axboe10ba5352006-10-20 11:39:27 +020010
Jens Axboe5945b9b2007-02-22 20:33:01 +010011/*
12 * Change this define to play with the timeout handling
13 */
14#undef FIO_USE_TIMEOUT
15
Jens Axboe97601022007-02-18 12:47:29 +010016struct io_completion_data {
17 int nr; /* input */
Jens Axboe97601022007-02-18 12:47:29 +010018
19 int error; /* output */
20 unsigned long bytes_done[2]; /* output */
21 struct timeval time; /* output */
22};
23
Jens Axboe10ba5352006-10-20 11:39:27 +020024/*
25 * The ->file_map[] contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 */
Jens Axboeaec2de22008-04-24 12:44:42 +020028static int random_map_free(struct fio_file *f, const unsigned long long block)
Jens Axboe10ba5352006-10-20 11:39:27 +020029{
Jens Axboeaec2de22008-04-24 12:44:42 +020030 unsigned int idx = RAND_MAP_IDX(f, block);
31 unsigned int bit = RAND_MAP_BIT(f, block);
Jens Axboe10ba5352006-10-20 11:39:27 +020032
Jens Axboe84422ac2008-02-19 20:10:26 +010033 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
34
Jens Axboede605662008-05-31 00:21:12 +020035 return (f->file_map[idx] & (1 << bit)) == 0;
Jens Axboe10ba5352006-10-20 11:39:27 +020036}
37
38/*
Jens Axboedf415582006-10-20 11:41:03 +020039 * Mark a given offset as used in the map.
40 */
Jens Axboe9bf20612007-03-01 09:33:57 +010041static void mark_random_map(struct thread_data *td, struct io_u *io_u)
Jens Axboedf415582006-10-20 11:41:03 +020042{
Jens Axboe2dc1bbe2007-03-15 15:01:33 +010043 unsigned int min_bs = td->o.rw_min_bs;
Jens Axboe9bf20612007-03-01 09:33:57 +010044 struct fio_file *f = io_u->file;
Jens Axboea00735e2006-11-03 08:58:08 +010045 unsigned long long block;
Jens Axboe3e3357b2008-06-02 09:53:05 +020046 unsigned int blocks, nr_blocks;
Jens Axboedf415582006-10-20 11:41:03 +020047
Jens Axboeb9c5b642008-02-18 14:13:08 +010048 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
Jens Axboec685b5b2007-02-10 20:02:28 +010049 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
Jens Axboe3e3357b2008-06-02 09:53:05 +020050 blocks = 0;
Jens Axboec685b5b2007-02-10 20:02:28 +010051
Jens Axboe3e3357b2008-06-02 09:53:05 +020052 while (nr_blocks) {
53 unsigned int this_blocks, mask;
Jens Axboedf415582006-10-20 11:41:03 +020054 unsigned int idx, bit;
55
Jens Axboe1e3d53a2007-03-22 19:01:48 +010056 /*
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
59 */
Jens Axboe8b113b42008-06-02 10:15:48 +020060 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
61 if (!blocks)
62 blocks = 1;
Jens Axboedf415582006-10-20 11:41:03 +020063 break;
Jens Axboe8b113b42008-06-02 10:15:48 +020064 }
Jens Axboedf415582006-10-20 11:41:03 +020065
Jens Axboeaec2de22008-04-24 12:44:42 +020066 idx = RAND_MAP_IDX(f, block);
67 bit = RAND_MAP_BIT(f, block);
Jens Axboedf415582006-10-20 11:41:03 +020068
Jens Axboe0032bf92007-02-13 17:39:56 +010069 fio_assert(td, idx < f->num_maps);
Jens Axboedf415582006-10-20 11:41:03 +020070
Jens Axboe3e3357b2008-06-02 09:53:05 +020071 this_blocks = nr_blocks;
72 if (this_blocks + bit > BLOCKS_PER_MAP)
73 this_blocks = BLOCKS_PER_MAP - bit;
74
75 if (this_blocks == BLOCKS_PER_MAP)
76 mask = -1U;
77 else
78 mask = ((1U << this_blocks) - 1) << bit;
79
Jens Axboe3e3357b2008-06-02 09:53:05 +020080 f->file_map[idx] |= mask;
81 nr_blocks -= this_blocks;
82 blocks += this_blocks;
Jens Axboec9dd34b2008-06-02 10:09:43 +020083 block += this_blocks;
Jens Axboedf415582006-10-20 11:41:03 +020084 }
85
Jens Axboea00735e2006-11-03 08:58:08 +010086 if ((blocks * min_bs) < io_u->buflen)
87 io_u->buflen = blocks * min_bs;
Jens Axboedf415582006-10-20 11:41:03 +020088}
89
Jens Axboe3e3357b2008-06-02 09:53:05 +020090static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
91 enum fio_ddir ddir)
Jens Axboe2ba1c292008-02-01 13:16:38 +010092{
93 unsigned long long max_blocks;
Jens Axboed9dd70f2008-05-23 12:37:23 +020094 unsigned long long max_size;
Jens Axboe2ba1c292008-02-01 13:16:38 +010095
Jens Axboed9dd70f2008-05-23 12:37:23 +020096 /*
97 * Hmm, should we make sure that ->io_size <= ->real_file_size?
98 */
99 max_size = f->io_size;
100 if (max_size > f->real_file_size)
101 max_size = f->real_file_size;
102
103 max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir];
Jens Axboe2ba1c292008-02-01 13:16:38 +0100104 if (!max_blocks)
105 return 0;
106
Jens Axboe67778e82008-05-15 09:20:08 +0200107 return max_blocks;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100108}
109
Jens Axboedf415582006-10-20 11:41:03 +0200110/*
Jens Axboe10ba5352006-10-20 11:39:27 +0200111 * Return the next free block in the map.
112 */
113static int get_next_free_block(struct thread_data *td, struct fio_file *f,
Jens Axboe4ba66132008-02-05 10:05:50 +0100114 enum fio_ddir ddir, unsigned long long *b)
Jens Axboe10ba5352006-10-20 11:39:27 +0200115{
Jens Axboe84422ac2008-02-19 20:10:26 +0100116 unsigned long long min_bs = td->o.rw_min_bs;
Jens Axboe10ba5352006-10-20 11:39:27 +0200117 int i;
118
Jens Axboec685b5b2007-02-10 20:02:28 +0100119 i = f->last_free_lookup;
120 *b = (i * BLOCKS_PER_MAP);
Jens Axboe84422ac2008-02-19 20:10:26 +0100121 while ((*b) * min_bs < f->real_file_size) {
Jens Axboede605662008-05-31 00:21:12 +0200122 if (f->file_map[i] != (unsigned int) -1) {
Jens Axboe697a6062008-05-31 00:04:45 +0200123 *b += ffz(f->file_map[i]);
Jens Axboe4ba66132008-02-05 10:05:50 +0100124 if (*b > last_block(td, f, ddir))
Jens Axboe2ba1c292008-02-01 13:16:38 +0100125 break;
Jens Axboec685b5b2007-02-10 20:02:28 +0100126 f->last_free_lookup = i;
Jens Axboe10ba5352006-10-20 11:39:27 +0200127 return 0;
128 }
129
130 *b += BLOCKS_PER_MAP;
131 i++;
132 }
133
Jens Axboe2ba1c292008-02-01 13:16:38 +0100134 dprint(FD_IO, "failed finding a free block\n");
Jens Axboe10ba5352006-10-20 11:39:27 +0200135 return 1;
136}
137
Jens Axboeec4015d2007-03-23 08:04:27 +0100138static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
Jens Axboe4ba66132008-02-05 10:05:50 +0100139 enum fio_ddir ddir, unsigned long long *b)
Jens Axboeec4015d2007-03-23 08:04:27 +0100140{
Jens Axboe84422ac2008-02-19 20:10:26 +0100141 unsigned long long r;
Jens Axboeec4015d2007-03-23 08:04:27 +0100142 int loops = 5;
143
144 do {
145 r = os_random_long(&td->random_state);
Jens Axboe84422ac2008-02-19 20:10:26 +0100146 dprint(FD_RANDOM, "off rand %llu\n", r);
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100147 *b = (last_block(td, f, ddir) - 1)
Jens Axboedc873b62008-06-04 20:13:04 +0200148 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
Jens Axboe2ba1c292008-02-01 13:16:38 +0100149
Jens Axboe43c63a72007-05-21 13:23:30 +0200150 /*
151 * if we are not maintaining a random map, we are done.
152 */
Jens Axboe303032a2008-03-26 10:11:10 +0100153 if (!file_randommap(td, f))
Jens Axboe43c63a72007-05-21 13:23:30 +0200154 return 0;
155
156 /*
Jens Axboe84422ac2008-02-19 20:10:26 +0100157 * calculate map offset and check if it's free
Jens Axboe43c63a72007-05-21 13:23:30 +0200158 */
Jens Axboeaec2de22008-04-24 12:44:42 +0200159 if (random_map_free(f, *b))
Jens Axboe43c63a72007-05-21 13:23:30 +0200160 return 0;
161
Jens Axboe84422ac2008-02-19 20:10:26 +0100162 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
163 *b);
Jens Axboe43c63a72007-05-21 13:23:30 +0200164 } while (--loops);
Jens Axboeec4015d2007-03-23 08:04:27 +0100165
166 /*
Jens Axboe43c63a72007-05-21 13:23:30 +0200167 * we get here, if we didn't suceed in looking up a block. generate
168 * a random start offset into the filemap, and find the first free
169 * block from there.
Jens Axboeec4015d2007-03-23 08:04:27 +0100170 */
Jens Axboe43c63a72007-05-21 13:23:30 +0200171 loops = 10;
172 do {
Jens Axboedc873b62008-06-04 20:13:04 +0200173 f->last_free_lookup = (f->num_maps - 1) *
174 (r / (OS_RAND_MAX + 1.0));
Jens Axboe4ba66132008-02-05 10:05:50 +0100175 if (!get_next_free_block(td, f, ddir, b))
Jens Axboe43c63a72007-05-21 13:23:30 +0200176 return 0;
Jens Axboeec4015d2007-03-23 08:04:27 +0100177
Jens Axboe43c63a72007-05-21 13:23:30 +0200178 r = os_random_long(&td->random_state);
179 } while (--loops);
180
181 /*
182 * that didn't work either, try exhaustive search from the start
183 */
184 f->last_free_lookup = 0;
Jens Axboe4ba66132008-02-05 10:05:50 +0100185 return get_next_free_block(td, f, ddir, b);
Jens Axboeec4015d2007-03-23 08:04:27 +0100186}
187
Jens Axboe10ba5352006-10-20 11:39:27 +0200188/*
189 * For random io, generate a random new block and see if it's used. Repeat
190 * until we find a free one. For sequential io, just return the end of
191 * the last io issued.
192 */
Jens Axboe9bf20612007-03-01 09:33:57 +0100193static int get_next_offset(struct thread_data *td, struct io_u *io_u)
Jens Axboe10ba5352006-10-20 11:39:27 +0200194{
Jens Axboe9bf20612007-03-01 09:33:57 +0100195 struct fio_file *f = io_u->file;
Jens Axboeec4015d2007-03-23 08:04:27 +0100196 unsigned long long b;
Jens Axboe4ba66132008-02-05 10:05:50 +0100197 enum fio_ddir ddir = io_u->ddir;
Jens Axboe10ba5352006-10-20 11:39:27 +0200198
Jens Axboeec4015d2007-03-23 08:04:27 +0100199 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
200 td->ddir_nr = td->o.ddir_nr;
Jens Axboe10ba5352006-10-20 11:39:27 +0200201
Jens Axboe4ba66132008-02-05 10:05:50 +0100202 if (get_next_rand_offset(td, f, ddir, &b))
Jens Axboebca4ed42007-02-12 05:13:23 +0100203 return 1;
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) ||
207 get_next_rand_offset(td, f, ddir, &b))
ljzhang,Yaxin Hu,Jianchao Tang0c3d7682007-07-26 10:59:25 +0200208 return 1;
ljzhang,Yaxin Hu,Jianchao Tangbcdedd02007-07-27 13:28:26 +0200209 } else
Jens Axboe4ba66132008-02-05 10:05:50 +0100210 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
Jens Axboe43063a12007-03-27 20:21:24 +0200211 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200212
Jens Axboe009bd842008-05-15 10:19:46 +0200213 io_u->offset = b * td->o.min_bs[ddir];
214 if (io_u->offset >= f->io_size) {
215 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
216 io_u->offset, f->io_size);
217 return 1;
218 }
219
220 io_u->offset += f->file_offset;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100221 if (io_u->offset >= f->real_file_size) {
222 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
223 io_u->offset, f->real_file_size);
Jens Axboe10ba5352006-10-20 11:39:27 +0200224 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100225 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200226
227 return 0;
228}
229
Jens Axboef3059de2008-06-11 15:37:32 +0200230static inline int is_power_of_2(unsigned int val)
231{
232 return (val != 0 && ((val & (val - 1)) == 0));
233}
234
Jens Axboe9bf20612007-03-01 09:33:57 +0100235static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
Jens Axboe10ba5352006-10-20 11:39:27 +0200236{
Jens Axboebca4ed42007-02-12 05:13:23 +0100237 const int ddir = io_u->ddir;
Jens Axboef3f552b2008-06-13 10:27:05 +0200238 unsigned int uninitialized_var(buflen);
Jens Axboef3059de2008-06-11 15:37:32 +0200239 unsigned int minbs, maxbs;
Jens Axboe10ba5352006-10-20 11:39:27 +0200240 long r;
241
Jens Axboef3059de2008-06-11 15:37:32 +0200242 minbs = td->o.min_bs[ddir];
243 maxbs = td->o.max_bs[ddir];
244
245 if (minbs == maxbs)
246 buflen = minbs;
Jens Axboe10ba5352006-10-20 11:39:27 +0200247 else {
248 r = os_random_long(&td->bsrange_state);
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100249 if (!td->o.bssplit_nr) {
Jens Axboef3059de2008-06-11 15:37:32 +0200250 buflen = 1 + (unsigned int) ((double) maxbs *
251 (r / (OS_RAND_MAX + 1.0)));
252 if (buflen < minbs)
253 buflen = minbs;
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100254 } else {
Jens Axboe564ca972007-12-14 12:21:19 +0100255 long perc = 0;
256 unsigned int i;
257
258 for (i = 0; i < td->o.bssplit_nr; i++) {
259 struct bssplit *bsp = &td->o.bssplit[i];
260
261 buflen = bsp->bs;
262 perc += bsp->perc;
Jens Axboef3059de2008-06-11 15:37:32 +0200263 if (r <= ((OS_RAND_MAX / 100L) * perc))
Jens Axboe564ca972007-12-14 12:21:19 +0100264 break;
265 }
266 }
Jens Axboef3059de2008-06-11 15:37:32 +0200267 if (!td->o.bs_unaligned && is_power_of_2(minbs))
268 buflen = (buflen + minbs - 1) & ~(minbs - 1);
Jens Axboe10ba5352006-10-20 11:39:27 +0200269 }
270
Jens Axboe4ba66132008-02-05 10:05:50 +0100271 if (io_u->offset + buflen > io_u->file->real_file_size) {
272 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
Jens Axboef3059de2008-06-11 15:37:32 +0200273 minbs, ddir);
274 buflen = minbs;
Jens Axboe4ba66132008-02-05 10:05:50 +0100275 }
Jens Axboe6a5e6882007-07-26 10:47:51 +0200276
Jens Axboe10ba5352006-10-20 11:39:27 +0200277 return buflen;
278}
279
Jens Axboeafe24a52007-03-16 20:27:27 +0100280static void set_rwmix_bytes(struct thread_data *td)
281{
Jens Axboeafe24a52007-03-16 20:27:27 +0100282 unsigned int diff;
283
284 /*
285 * we do time or byte based switch. this is needed because
286 * buffered writes may issue a lot quicker than they complete,
287 * whereas reads do not.
288 */
Jens Axboee47f7992007-03-21 14:05:39 +0100289 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
Jens Axboe04c540d2008-05-28 10:35:26 +0200290 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
Jens Axboee47f7992007-03-21 14:05:39 +0100291}
292
293static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
294{
295 unsigned int v;
296 long r;
297
298 r = os_random_long(&td->rwmix_state);
Jens Axboedc873b62008-06-04 20:13:04 +0200299 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
Jens Axboe04c540d2008-05-28 10:35:26 +0200300 if (v <= td->o.rwmix[DDIR_READ])
Jens Axboee47f7992007-03-21 14:05:39 +0100301 return DDIR_READ;
302
303 return DDIR_WRITE;
Jens Axboeafe24a52007-03-16 20:27:27 +0100304}
305
Jens Axboe10ba5352006-10-20 11:39:27 +0200306/*
307 * Return the data direction for the next io_u. If the job is a
308 * mixed read/write workload, check the rwmix cycle and switch if
309 * necessary.
310 */
Jens Axboe1e97cce2006-12-05 11:44:16 +0100311static enum fio_ddir get_rw_ddir(struct thread_data *td)
Jens Axboe10ba5352006-10-20 11:39:27 +0200312{
313 if (td_rw(td)) {
Jens Axboe10ba5352006-10-20 11:39:27 +0200314 /*
315 * Check if it's time to seed a new data direction.
316 */
Jens Axboee4928662008-04-07 09:19:46 +0200317 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
Jens Axboee47f7992007-03-21 14:05:39 +0100318 unsigned long long max_bytes;
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100319 enum fio_ddir ddir;
Jens Axboe10ba5352006-10-20 11:39:27 +0200320
Jens Axboee47f7992007-03-21 14:05:39 +0100321 /*
322 * Put a top limit on how many bytes we do for
323 * one data direction, to avoid overflowing the
324 * ranges too much
325 */
326 ddir = get_rand_ddir(td);
327 max_bytes = td->this_io_bytes[ddir];
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100328 if (max_bytes >=
329 (td->o.size * td->o.rwmix[ddir] / 100)) {
Jens Axboe6162f522008-06-04 10:35:25 +0200330 if (!td->rw_end_set[ddir]) {
Jens Axboe38d77ca2007-03-21 14:20:20 +0100331 td->rw_end_set[ddir] = 1;
Jens Axboe6162f522008-06-04 10:35:25 +0200332 fio_gettime(&td->rw_end[ddir], NULL);
333 }
Jens Axboee4928662008-04-07 09:19:46 +0200334
Jens Axboee47f7992007-03-21 14:05:39 +0100335 ddir ^= 1;
Jens Axboe38d77ca2007-03-21 14:20:20 +0100336 }
Jens Axboee47f7992007-03-21 14:05:39 +0100337
338 if (ddir != td->rwmix_ddir)
339 set_rwmix_bytes(td);
340
341 td->rwmix_ddir = ddir;
Jens Axboe10ba5352006-10-20 11:39:27 +0200342 }
343 return td->rwmix_ddir;
344 } else if (td_read(td))
345 return DDIR_READ;
346 else
347 return DDIR_WRITE;
348}
349
Jens Axboe60f2c652008-05-16 12:31:36 +0200350static void put_file_log(struct thread_data *td, struct fio_file *f)
351{
352 int ret = put_file(td, f);
353
354 if (ret)
355 td_verror(td, ret, "file close");
356}
357
Jens Axboe10ba5352006-10-20 11:39:27 +0200358void put_io_u(struct thread_data *td, struct io_u *io_u)
359{
Jens Axboe0c6e7512007-02-22 11:19:39 +0100360 assert((io_u->flags & IO_U_F_FREE) == 0);
361 io_u->flags |= IO_U_F_FREE;
362
Jens Axboe60f2c652008-05-16 12:31:36 +0200363 if (io_u->file)
364 put_file_log(td, io_u->file);
Jens Axboe2dbdab72007-04-26 10:21:15 +0200365
Jens Axboe10ba5352006-10-20 11:39:27 +0200366 io_u->file = NULL;
Jens Axboe01743ee2008-06-02 12:19:19 +0200367 flist_del(&io_u->list);
368 flist_add(&io_u->list, &td->io_u_freelist);
Jens Axboe10ba5352006-10-20 11:39:27 +0200369 td->cur_depth--;
370}
371
Jens Axboe755200a2007-02-19 13:08:12 +0100372void requeue_io_u(struct thread_data *td, struct io_u **io_u)
373{
374 struct io_u *__io_u = *io_u;
375
Jens Axboe465221b2008-05-30 22:07:49 +0200376 dprint(FD_IO, "requeue %p\n", __io_u);
377
Jens Axboe4d2e0f42007-02-24 13:31:57 +0100378 __io_u->flags |= IO_U_F_FREE;
Jens Axboee4f54ad2008-02-04 10:56:07 +0100379 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
380 td->io_issues[__io_u->ddir]--;
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100381
Jens Axboe4d2e0f42007-02-24 13:31:57 +0100382 __io_u->flags &= ~IO_U_F_FLIGHT;
383
Jens Axboe01743ee2008-06-02 12:19:19 +0200384 flist_del(&__io_u->list);
385 flist_add_tail(&__io_u->list, &td->io_u_requeues);
Jens Axboe755200a2007-02-19 13:08:12 +0100386 td->cur_depth--;
387 *io_u = NULL;
388}
389
Jens Axboe9bf20612007-03-01 09:33:57 +0100390static int fill_io_u(struct thread_data *td, struct io_u *io_u)
Jens Axboe10ba5352006-10-20 11:39:27 +0200391{
Jens Axboeb4c5e1a2007-10-25 18:44:45 +0200392 if (td->io_ops->flags & FIO_NOIO)
393 goto out;
394
Jens Axboe10ba5352006-10-20 11:39:27 +0200395 /*
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200396 * see if it's time to sync
397 */
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100398 if (td->o.fsync_blocks &&
399 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
400 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200401 io_u->ddir = DDIR_SYNC;
Jens Axboec38e9462007-03-27 08:48:48 +0200402 goto out;
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200403 }
404
Jens Axboea00735e2006-11-03 08:58:08 +0100405 io_u->ddir = get_rw_ddir(td);
406
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200407 /*
Jens Axboe48f5abd2007-07-20 13:25:04 +0200408 * See if it's time to switch to a new zone
409 */
410 if (td->zone_bytes >= td->o.zone_size) {
411 td->zone_bytes = 0;
412 io_u->file->last_pos += td->o.zone_skip;
413 td->io_skip_bytes += td->o.zone_skip;
414 }
415
416 /*
Jens Axboec685b5b2007-02-10 20:02:28 +0100417 * No log, let the seq/rand engine retrieve the next buflen and
418 * position.
Jens Axboe10ba5352006-10-20 11:39:27 +0200419 */
Jens Axboe2ba1c292008-02-01 13:16:38 +0100420 if (get_next_offset(td, io_u)) {
421 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
Jens Axboebca4ed42007-02-12 05:13:23 +0100422 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100423 }
Jens Axboec685b5b2007-02-10 20:02:28 +0100424
Jens Axboe9bf20612007-03-01 09:33:57 +0100425 io_u->buflen = get_next_buflen(td, io_u);
Jens Axboe2ba1c292008-02-01 13:16:38 +0100426 if (!io_u->buflen) {
427 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
Jens Axboebca4ed42007-02-12 05:13:23 +0100428 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100429 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200430
Jens Axboe2ba1c292008-02-01 13:16:38 +0100431 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
432 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
Jens Axboe4ba66132008-02-05 10:05:50 +0100433 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
434 io_u->buflen, io_u->file->real_file_size);
Jens Axboe6a5e6882007-07-26 10:47:51 +0200435 return 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100436 }
Jens Axboe6a5e6882007-07-26 10:47:51 +0200437
Jens Axboebca4ed42007-02-12 05:13:23 +0100438 /*
439 * mark entry before potentially trimming io_u
440 */
Jens Axboe303032a2008-03-26 10:11:10 +0100441 if (td_random(td) && file_randommap(td, io_u->file))
Jens Axboe9bf20612007-03-01 09:33:57 +0100442 mark_random_map(td, io_u);
Jens Axboe10ba5352006-10-20 11:39:27 +0200443
Jens Axboebca4ed42007-02-12 05:13:23 +0100444 /*
445 * If using a write iolog, store this entry.
446 */
Jens Axboec38e9462007-03-27 08:48:48 +0200447out:
Jens Axboe2ba1c292008-02-01 13:16:38 +0100448 dprint_io_u(io_u, "fill_io_u");
Jens Axboed9d91e32008-01-31 13:25:42 +0100449 td->zone_bytes += io_u->buflen;
Jens Axboef29b25a2007-07-23 08:56:43 +0200450 log_io_u(td, io_u);
Jens Axboebca4ed42007-02-12 05:13:23 +0100451 return 0;
Jens Axboe10ba5352006-10-20 11:39:27 +0200452}
453
Jens Axboe838bc702008-05-22 13:08:23 +0200454static void __io_u_mark_map(unsigned int *map, unsigned int nr)
455{
456 int index = 0;
457
458 switch (nr) {
459 default:
460 index = 6;
461 break;
462 case 33 ... 64:
463 index = 5;
464 break;
465 case 17 ... 32:
466 index = 4;
467 break;
468 case 9 ... 16:
469 index = 3;
470 break;
471 case 5 ... 8:
472 index = 2;
473 break;
474 case 1 ... 4:
475 index = 1;
476 case 0:
477 break;
478 }
479
480 map[index]++;
481}
482
483void io_u_mark_submit(struct thread_data *td, unsigned int nr)
484{
485 __io_u_mark_map(td->ts.io_u_submit, nr);
486 td->ts.total_submit++;
487}
488
489void io_u_mark_complete(struct thread_data *td, unsigned int nr)
490{
491 __io_u_mark_map(td->ts.io_u_complete, nr);
492 td->ts.total_complete++;
493}
494
Jens Axboed8005752008-05-15 09:49:09 +0200495void io_u_mark_depth(struct thread_data *td, unsigned int nr)
Jens Axboe71619dc2007-01-13 23:56:33 +0100496{
497 int index = 0;
498
499 switch (td->cur_depth) {
500 default:
Jens Axboea783e612007-06-19 09:50:28 +0200501 index = 6;
502 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100503 case 32 ... 63:
Jens Axboea783e612007-06-19 09:50:28 +0200504 index = 5;
505 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100506 case 16 ... 31:
Jens Axboea783e612007-06-19 09:50:28 +0200507 index = 4;
508 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100509 case 8 ... 15:
Jens Axboea783e612007-06-19 09:50:28 +0200510 index = 3;
511 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100512 case 4 ... 7:
Jens Axboea783e612007-06-19 09:50:28 +0200513 index = 2;
514 break;
Jens Axboe71619dc2007-01-13 23:56:33 +0100515 case 2 ... 3:
Jens Axboea783e612007-06-19 09:50:28 +0200516 index = 1;
Jens Axboe71619dc2007-01-13 23:56:33 +0100517 case 1:
518 break;
519 }
520
Jens Axboe3bec7ae2008-05-14 21:08:37 +0200521 td->ts.io_u_map[index] += nr;
Jens Axboe71619dc2007-01-13 23:56:33 +0100522}
523
Jens Axboe04a0fea2007-06-19 12:48:41 +0200524static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
525{
526 int index = 0;
527
528 assert(usec < 1000);
529
530 switch (usec) {
531 case 750 ... 999:
532 index = 9;
533 break;
534 case 500 ... 749:
535 index = 8;
536 break;
537 case 250 ... 499:
538 index = 7;
539 break;
540 case 100 ... 249:
541 index = 6;
542 break;
543 case 50 ... 99:
544 index = 5;
545 break;
546 case 20 ... 49:
547 index = 4;
548 break;
549 case 10 ... 19:
550 index = 3;
551 break;
552 case 4 ... 9:
553 index = 2;
554 break;
555 case 2 ... 3:
556 index = 1;
557 case 0 ... 1:
558 break;
559 }
560
561 assert(index < FIO_IO_U_LAT_U_NR);
562 td->ts.io_u_lat_u[index]++;
563}
564
565static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
Jens Axboeec118302007-02-17 04:38:20 +0100566{
567 int index = 0;
568
569 switch (msec) {
570 default:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200571 index = 11;
572 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100573 case 1000 ... 1999:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200574 index = 10;
575 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100576 case 750 ... 999:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200577 index = 9;
578 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100579 case 500 ... 749:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200580 index = 8;
581 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100582 case 250 ... 499:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200583 index = 7;
584 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100585 case 100 ... 249:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200586 index = 6;
587 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100588 case 50 ... 99:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200589 index = 5;
590 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100591 case 20 ... 49:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200592 index = 4;
593 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100594 case 10 ... 19:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200595 index = 3;
596 break;
Jens Axboe8abdce62007-02-21 10:22:55 +0100597 case 4 ... 9:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200598 index = 2;
599 break;
Jens Axboeec118302007-02-17 04:38:20 +0100600 case 2 ... 3:
Jens Axboe04a0fea2007-06-19 12:48:41 +0200601 index = 1;
Jens Axboeec118302007-02-17 04:38:20 +0100602 case 0 ... 1:
603 break;
604 }
605
Jens Axboe04a0fea2007-06-19 12:48:41 +0200606 assert(index < FIO_IO_U_LAT_M_NR);
607 td->ts.io_u_lat_m[index]++;
608}
609
610static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
611{
612 if (usec < 1000)
613 io_u_mark_lat_usec(td, usec);
614 else
615 io_u_mark_lat_msec(td, usec / 1000);
Jens Axboeec118302007-02-17 04:38:20 +0100616}
617
Jens Axboe0aabe162007-02-23 08:45:55 +0100618/*
619 * Get next file to service by choosing one at random
620 */
Jens Axboe1c178182007-03-13 13:25:18 +0100621static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
622 int badf)
Jens Axboe0aabe162007-02-23 08:45:55 +0100623{
Jens Axboe0aabe162007-02-23 08:45:55 +0100624 struct fio_file *f;
Jens Axboe1c178182007-03-13 13:25:18 +0100625 int fno;
Jens Axboe0aabe162007-02-23 08:45:55 +0100626
627 do {
Jens Axboe7c83c082007-03-01 10:04:15 +0100628 long r = os_random_long(&td->next_file_state);
629
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100630 fno = (unsigned int) ((double) td->o.nr_files
Jens Axboedc873b62008-06-04 20:13:04 +0200631 * (r / (OS_RAND_MAX + 1.0)));
Jens Axboe126d65c2008-03-01 18:04:31 +0100632 f = td->files[fno];
Jens Axboe059e63c2007-03-27 20:34:47 +0200633 if (f->flags & FIO_FILE_DONE)
634 continue;
Jens Axboe1c178182007-03-13 13:25:18 +0100635
Jens Axboe2ba1c292008-02-01 13:16:38 +0100636 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
637 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
Jens Axboe0aabe162007-02-23 08:45:55 +0100638 return f;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100639 }
Jens Axboe0aabe162007-02-23 08:45:55 +0100640 } while (1);
641}
642
643/*
644 * Get next file to service by doing round robin between all available ones
645 */
Jens Axboe1c178182007-03-13 13:25:18 +0100646static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
647 int badf)
Jens Axboe3d7c3912007-02-19 13:16:12 +0100648{
649 unsigned int old_next_file = td->next_file;
650 struct fio_file *f;
651
652 do {
Jens Axboe126d65c2008-03-01 18:04:31 +0100653 f = td->files[td->next_file];
Jens Axboe3d7c3912007-02-19 13:16:12 +0100654
655 td->next_file++;
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100656 if (td->next_file >= td->o.nr_files)
Jens Axboe3d7c3912007-02-19 13:16:12 +0100657 td->next_file = 0;
658
Jens Axboed5ed68e2007-03-28 09:33:43 +0200659 if (f->flags & FIO_FILE_DONE) {
660 f = NULL;
Jens Axboe059e63c2007-03-27 20:34:47 +0200661 continue;
Jens Axboed5ed68e2007-03-28 09:33:43 +0200662 }
Jens Axboe059e63c2007-03-27 20:34:47 +0200663
Jens Axboe1c178182007-03-13 13:25:18 +0100664 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
Jens Axboe3d7c3912007-02-19 13:16:12 +0100665 break;
666
667 f = NULL;
668 } while (td->next_file != old_next_file);
669
Jens Axboe2ba1c292008-02-01 13:16:38 +0100670 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
Jens Axboe3d7c3912007-02-19 13:16:12 +0100671 return f;
672}
673
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100674static struct fio_file *get_next_file(struct thread_data *td)
675{
Jens Axboe1907dbc2007-03-12 11:44:28 +0100676 struct fio_file *f;
677
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100678 assert(td->o.nr_files <= td->files_index);
Jens Axboe1c178182007-03-13 13:25:18 +0100679
Jens Axboe2ba1c292008-02-01 13:16:38 +0100680 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100681 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
682 " nr_files=%d\n", td->nr_open_files,
683 td->nr_done_files,
684 td->o.nr_files);
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100685 return NULL;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100686 }
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100687
Jens Axboe1907dbc2007-03-12 11:44:28 +0100688 f = td->file_service_file;
Jens Axboef11bd942007-03-13 10:16:34 +0100689 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
Jens Axboe2ba1c292008-02-01 13:16:38 +0100690 goto out;
Jens Axboe1907dbc2007-03-12 11:44:28 +0100691
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100692 if (td->o.file_service_type == FIO_FSERVICE_RR)
Jens Axboe1c178182007-03-13 13:25:18 +0100693 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100694 else
Jens Axboe1c178182007-03-13 13:25:18 +0100695 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
Jens Axboe1907dbc2007-03-12 11:44:28 +0100696
697 td->file_service_file = f;
698 td->file_service_left = td->file_service_nr - 1;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100699out:
700 dprint(FD_FILE, "get_next_file: %p\n", f);
Jens Axboe1907dbc2007-03-12 11:44:28 +0100701 return f;
Jens Axboebdb4e2e2007-03-01 09:54:57 +0100702}
703
Jens Axboe1c178182007-03-13 13:25:18 +0100704static struct fio_file *find_next_new_file(struct thread_data *td)
705{
706 struct fio_file *f;
707
Jens Axboe1020a132007-03-29 11:15:06 +0200708 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
709 return NULL;
710
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100711 if (td->o.file_service_type == FIO_FSERVICE_RR)
Jens Axboe1c178182007-03-13 13:25:18 +0100712 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
713 else
714 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
715
716 return f;
717}
718
Jens Axboe429f6672007-07-23 10:38:43 +0200719static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
720{
721 struct fio_file *f;
722
723 do {
724 f = get_next_file(td);
725 if (!f)
726 return 1;
727
728set_file:
729 io_u->file = f;
730 get_file(f);
731
732 if (!fill_io_u(td, io_u))
733 break;
734
735 /*
Jens Axboe009bd842008-05-15 10:19:46 +0200736 * optimization to prevent close/open of the same file. This
737 * way we preserve queueing etc.
738 */
739 if (td->o.nr_files == 1 && td->o.time_based) {
Jens Axboe60f2c652008-05-16 12:31:36 +0200740 put_file_log(td, f);
Jens Axboe009bd842008-05-15 10:19:46 +0200741 fio_file_reset(f);
742 goto set_file;
743 }
744
745 /*
Jens Axboe429f6672007-07-23 10:38:43 +0200746 * td_io_close() does a put_file() as well, so no need to
747 * do that here.
748 */
749 io_u->file = NULL;
750 td_io_close_file(td, f);
751 f->flags |= FIO_FILE_DONE;
752 td->nr_done_files++;
753
754 /*
755 * probably not the right place to do this, but see
756 * if we need to open a new file
757 */
758 if (td->nr_open_files < td->o.open_files &&
759 td->o.open_files != td->o.nr_files) {
760 f = find_next_new_file(td);
761
762 if (!f || td_io_open_file(td, f))
763 return 1;
764
765 goto set_file;
766 }
767 } while (1);
768
769 return 0;
770}
771
772
Jens Axboe10ba5352006-10-20 11:39:27 +0200773struct io_u *__get_io_u(struct thread_data *td)
774{
775 struct io_u *io_u = NULL;
776
Jens Axboe01743ee2008-06-02 12:19:19 +0200777 if (!flist_empty(&td->io_u_requeues))
778 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
Jens Axboe755200a2007-02-19 13:08:12 +0100779 else if (!queue_full(td)) {
Jens Axboe01743ee2008-06-02 12:19:19 +0200780 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
Jens Axboe10ba5352006-10-20 11:39:27 +0200781
Jens Axboe6040dab2006-10-24 19:38:15 +0200782 io_u->buflen = 0;
Jens Axboe10ba5352006-10-20 11:39:27 +0200783 io_u->resid = 0;
Jens Axboe755200a2007-02-19 13:08:12 +0100784 io_u->file = NULL;
Jens Axboed7762cf2007-02-23 12:34:57 +0100785 io_u->end_io = NULL;
Jens Axboe755200a2007-02-19 13:08:12 +0100786 }
787
788 if (io_u) {
Jens Axboe0c6e7512007-02-22 11:19:39 +0100789 assert(io_u->flags & IO_U_F_FREE);
790 io_u->flags &= ~IO_U_F_FREE;
791
Jens Axboe755200a2007-02-19 13:08:12 +0100792 io_u->error = 0;
Jens Axboe01743ee2008-06-02 12:19:19 +0200793 flist_del(&io_u->list);
794 flist_add(&io_u->list, &td->io_u_busylist);
Jens Axboe10ba5352006-10-20 11:39:27 +0200795 td->cur_depth++;
796 }
797
798 return io_u;
799}
800
801/*
802 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
803 * etc. The returned io_u is fully ready to be prepped and submitted.
804 */
Jens Axboe3d7c3912007-02-19 13:16:12 +0100805struct io_u *get_io_u(struct thread_data *td)
Jens Axboe10ba5352006-10-20 11:39:27 +0200806{
Jens Axboe3d7c3912007-02-19 13:16:12 +0100807 struct fio_file *f;
Jens Axboe10ba5352006-10-20 11:39:27 +0200808 struct io_u *io_u;
809
810 io_u = __get_io_u(td);
Jens Axboe2ba1c292008-02-01 13:16:38 +0100811 if (!io_u) {
812 dprint(FD_IO, "__get_io_u failed\n");
Jens Axboe10ba5352006-10-20 11:39:27 +0200813 return NULL;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100814 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200815
Jens Axboe755200a2007-02-19 13:08:12 +0100816 /*
817 * from a requeue, io_u already setup
818 */
819 if (io_u->file)
Jens Axboe77f392b2007-02-19 20:13:09 +0100820 goto out;
Jens Axboe755200a2007-02-19 13:08:12 +0100821
Jens Axboe429f6672007-07-23 10:38:43 +0200822 /*
823 * If using an iolog, grab next piece if any available.
824 */
825 if (td->o.read_iolog_file) {
826 if (read_iolog_get(td, io_u))
827 goto err_put;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100828 } else if (set_io_u_file(td, io_u)) {
829 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
Jens Axboe429f6672007-07-23 10:38:43 +0200830 goto err_put;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100831 }
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100832
Jens Axboe429f6672007-07-23 10:38:43 +0200833 f = io_u->file;
834 assert(f->flags & FIO_FILE_OPEN);
Jens Axboe97af62c2007-05-22 11:12:13 +0200835
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200836 if (io_u->ddir != DDIR_SYNC) {
Jens Axboed0656a92008-02-01 18:33:23 +0100837 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
Jens Axboe2ba1c292008-02-01 13:16:38 +0100838 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
Jens Axboe429f6672007-07-23 10:38:43 +0200839 goto err_put;
Jens Axboe2ba1c292008-02-01 13:16:38 +0100840 }
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200841
Jens Axboe36167d82007-02-18 05:41:31 +0100842 f->last_pos = io_u->offset + io_u->buflen;
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200843
Jens Axboe2dc1bbe2007-03-15 15:01:33 +0100844 if (td->o.verify != VERIFY_NONE)
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200845 populate_verify_io_u(td, io_u);
Jens Axboee4dad9c2008-05-28 10:53:44 +0200846 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
847 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
Jens Axboe10ba5352006-10-20 11:39:27 +0200848 }
849
Jens Axboe165faf12007-02-07 11:30:37 +0100850 /*
851 * Set io data pointers.
852 */
Jens Axboed460eb32007-04-16 21:39:33 +0200853 io_u->endpos = io_u->offset + io_u->buflen;
Jens Axboecec6b552007-02-06 20:15:38 +0100854 io_u->xfer_buf = io_u->buf;
855 io_u->xfer_buflen = io_u->buflen;
Jens Axboe5973caf2008-05-21 19:52:35 +0200856
Jens Axboe6ac7a332008-03-01 15:22:32 +0100857out:
Jens Axboe429f6672007-07-23 10:38:43 +0200858 if (!td_io_prep(td, io_u)) {
859 fio_gettime(&io_u->start_time, NULL);
860 return io_u;
Jens Axboe36167d82007-02-18 05:41:31 +0100861 }
Jens Axboe429f6672007-07-23 10:38:43 +0200862err_put:
Jens Axboe2ba1c292008-02-01 13:16:38 +0100863 dprint(FD_IO, "get_io_u failed\n");
Jens Axboe429f6672007-07-23 10:38:43 +0200864 put_io_u(td, io_u);
865 return NULL;
Jens Axboe10ba5352006-10-20 11:39:27 +0200866}
867
Jens Axboe54517922007-03-05 10:06:06 +0100868void io_u_log_error(struct thread_data *td, struct io_u *io_u)
869{
870 const char *msg[] = { "read", "write", "sync" };
871
872 log_err("fio: io_u error");
873
874 if (io_u->file)
875 log_err(" on file %s", io_u->file->file_name);
876
877 log_err(": %s\n", strerror(io_u->error));
878
Jens Axboe5ec10ea2008-03-06 15:42:00 +0100879 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
880 io_u->offset, io_u->xfer_buflen);
Jens Axboe54517922007-03-05 10:06:06 +0100881
882 if (!td->error)
883 td_verror(td, io_u->error, "io_u error");
884}
885
Jens Axboe97601022007-02-18 12:47:29 +0100886static void io_completed(struct thread_data *td, struct io_u *io_u,
887 struct io_completion_data *icd)
Jens Axboe10ba5352006-10-20 11:39:27 +0200888{
Jens Axboed85f5112007-06-18 09:41:23 +0200889 unsigned long usec;
Jens Axboe10ba5352006-10-20 11:39:27 +0200890
Jens Axboe2ba1c292008-02-01 13:16:38 +0100891 dprint_io_u(io_u, "io complete");
892
Jens Axboe0c6e7512007-02-22 11:19:39 +0100893 assert(io_u->flags & IO_U_F_FLIGHT);
894 io_u->flags &= ~IO_U_F_FLIGHT;
895
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200896 if (io_u->ddir == DDIR_SYNC) {
897 td->last_was_sync = 1;
898 return;
899 }
900
901 td->last_was_sync = 0;
902
Jens Axboe10ba5352006-10-20 11:39:27 +0200903 if (!io_u->error) {
904 unsigned int bytes = io_u->buflen - io_u->resid;
Jens Axboe1e97cce2006-12-05 11:44:16 +0100905 const enum fio_ddir idx = io_u->ddir;
Jens Axboe3af6ef32007-02-18 06:57:43 +0100906 int ret;
Jens Axboe10ba5352006-10-20 11:39:27 +0200907
908 td->io_blocks[idx]++;
909 td->io_bytes[idx] += bytes;
Jens Axboe10ba5352006-10-20 11:39:27 +0200910 td->this_io_bytes[idx] += bytes;
911
Jens Axboed85f5112007-06-18 09:41:23 +0200912 usec = utime_since(&io_u->issue_time, &icd->time);
Jens Axboe10ba5352006-10-20 11:39:27 +0200913
Jens Axboed85f5112007-06-18 09:41:23 +0200914 add_clat_sample(td, idx, usec);
Jens Axboe02bcaa82006-11-24 10:42:00 +0100915 add_bw_sample(td, idx, &icd->time);
Jens Axboe04a0fea2007-06-19 12:48:41 +0200916 io_u_mark_latency(td, usec);
Jens Axboe10ba5352006-10-20 11:39:27 +0200917
Jens Axboe660a1cb2007-04-17 15:37:47 +0200918 if (td_write(td) && idx == DDIR_WRITE &&
Shawn Lewis86705792007-11-21 09:35:41 +0100919 td->o.do_verify &&
Jens Axboe41128402007-03-27 08:32:48 +0200920 td->o.verify != VERIFY_NONE)
Jens Axboe10ba5352006-10-20 11:39:27 +0200921 log_io_piece(td, io_u);
922
923 icd->bytes_done[idx] += bytes;
Jens Axboe3af6ef32007-02-18 06:57:43 +0100924
Jens Axboed7762cf2007-02-23 12:34:57 +0100925 if (io_u->end_io) {
Jens Axboe36690c92007-03-26 10:23:34 +0200926 ret = io_u->end_io(td, io_u);
Jens Axboe3af6ef32007-02-18 06:57:43 +0100927 if (ret && !icd->error)
928 icd->error = ret;
929 }
Jens Axboe54517922007-03-05 10:06:06 +0100930 } else {
Jens Axboe10ba5352006-10-20 11:39:27 +0200931 icd->error = io_u->error;
Jens Axboe54517922007-03-05 10:06:06 +0100932 io_u_log_error(td, io_u);
933 }
Jens Axboe10ba5352006-10-20 11:39:27 +0200934}
935
Jens Axboed7762cf2007-02-23 12:34:57 +0100936static void init_icd(struct io_completion_data *icd, int nr)
Jens Axboe36167d82007-02-18 05:41:31 +0100937{
938 fio_gettime(&icd->time, NULL);
939
Jens Axboe3af6ef32007-02-18 06:57:43 +0100940 icd->nr = nr;
941
Jens Axboe36167d82007-02-18 05:41:31 +0100942 icd->error = 0;
943 icd->bytes_done[0] = icd->bytes_done[1] = 0;
944}
945
Jens Axboe97601022007-02-18 12:47:29 +0100946static void ios_completed(struct thread_data *td,
947 struct io_completion_data *icd)
Jens Axboe10ba5352006-10-20 11:39:27 +0200948{
949 struct io_u *io_u;
950 int i;
951
Jens Axboe10ba5352006-10-20 11:39:27 +0200952 for (i = 0; i < icd->nr; i++) {
953 io_u = td->io_ops->event(td, i);
954
955 io_completed(td, io_u, icd);
956 put_io_u(td, io_u);
957 }
958}
Jens Axboe97601022007-02-18 12:47:29 +0100959
Jens Axboee7e6cfb2007-02-20 10:58:34 +0100960/*
961 * Complete a single io_u for the sync engines.
962 */
Jens Axboed7762cf2007-02-23 12:34:57 +0100963long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
Jens Axboe97601022007-02-18 12:47:29 +0100964{
965 struct io_completion_data icd;
966
Jens Axboed7762cf2007-02-23 12:34:57 +0100967 init_icd(&icd, 1);
Jens Axboe97601022007-02-18 12:47:29 +0100968 io_completed(td, io_u, &icd);
969 put_io_u(td, io_u);
970
971 if (!icd.error)
972 return icd.bytes_done[0] + icd.bytes_done[1];
973
Jens Axboe37e974a2007-03-02 15:16:45 +0100974 td_verror(td, icd.error, "io_u_sync_complete");
Jens Axboe97601022007-02-18 12:47:29 +0100975 return -1;
976}
977
Jens Axboee7e6cfb2007-02-20 10:58:34 +0100978/*
979 * Called to complete min_events number of io for the async engines.
980 */
Jens Axboe49504212008-06-05 09:03:30 +0200981long io_u_queued_complete(struct thread_data *td, int min_evts)
Jens Axboe97601022007-02-18 12:47:29 +0100982{
Jens Axboe97601022007-02-18 12:47:29 +0100983 struct io_completion_data icd;
Jens Axboe00de55e2007-02-20 10:45:57 +0100984 struct timespec *tvp = NULL;
Jens Axboe97601022007-02-18 12:47:29 +0100985 int ret;
Davide Libenzi4d06a332007-03-22 07:43:50 +0100986 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
Jens Axboe97601022007-02-18 12:47:29 +0100987
Jens Axboe49504212008-06-05 09:03:30 +0200988 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
Jens Axboeb271fe62008-02-04 10:49:41 +0100989
Jens Axboe49504212008-06-05 09:03:30 +0200990 if (!min_evts)
Jens Axboe00de55e2007-02-20 10:45:57 +0100991 tvp = &ts;
Jens Axboe97601022007-02-18 12:47:29 +0100992
Jens Axboe49504212008-06-05 09:03:30 +0200993 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
Jens Axboe97601022007-02-18 12:47:29 +0100994 if (ret < 0) {
Jens Axboee1161c32007-02-22 19:36:48 +0100995 td_verror(td, -ret, "td_io_getevents");
Jens Axboe97601022007-02-18 12:47:29 +0100996 return ret;
997 } else if (!ret)
998 return ret;
999
Jens Axboed7762cf2007-02-23 12:34:57 +01001000 init_icd(&icd, ret);
Jens Axboe97601022007-02-18 12:47:29 +01001001 ios_completed(td, &icd);
1002 if (!icd.error)
1003 return icd.bytes_done[0] + icd.bytes_done[1];
1004
Jens Axboe37e974a2007-03-02 15:16:45 +01001005 td_verror(td, icd.error, "io_u_queued_complete");
Jens Axboe97601022007-02-18 12:47:29 +01001006 return -1;
1007}
Jens Axboe7e77dd02007-02-20 10:57:34 +01001008
1009/*
1010 * Call when io_u is really queued, to update the submission latency.
1011 */
1012void io_u_queued(struct thread_data *td, struct io_u *io_u)
1013{
1014 unsigned long slat_time;
1015
Jens Axboed85f5112007-06-18 09:41:23 +02001016 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
Jens Axboe7e77dd02007-02-20 10:57:34 +01001017 add_slat_sample(td, io_u->ddir, slat_time);
1018}
Jens Axboe433afcb2007-02-22 10:39:01 +01001019
Jens Axboe5973caf2008-05-21 19:52:35 +02001020/*
1021 * "randomly" fill the buffer contents
1022 */
1023void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1024 unsigned int max_bs)
1025{
1026 long *ptr = io_u->buf;
1027
1028 if (!td->o.zero_buffers) {
1029 while ((void *) ptr - io_u->buf < max_bs) {
1030 *ptr = rand() * GOLDEN_RATIO_PRIME;
1031 ptr++;
1032 }
1033 } else
1034 memset(ptr, 0, max_bs);
1035}
1036
Jens Axboe55bc9722007-02-22 11:30:05 +01001037#ifdef FIO_USE_TIMEOUT
Jens Axboe433afcb2007-02-22 10:39:01 +01001038void io_u_set_timeout(struct thread_data *td)
1039{
1040 assert(td->cur_depth);
1041
1042 td->timer.it_interval.tv_sec = 0;
1043 td->timer.it_interval.tv_usec = 0;
1044 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
1045 td->timer.it_value.tv_usec = 0;
1046 setitimer(ITIMER_REAL, &td->timer, NULL);
1047 fio_gettime(&td->timeout_end, NULL);
1048}
Jens Axboe5945b9b2007-02-22 20:33:01 +01001049
1050static void io_u_dump(struct io_u *io_u)
1051{
1052 unsigned long t_start = mtime_since_now(&io_u->start_time);
1053 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
1054
1055 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
Jens Axboe5ec10ea2008-03-06 15:42:00 +01001056 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
1057 io_u->xfer_buf, io_u->buflen,
1058 io_u->xfer_buflen,
1059 io_u->offset);
Jens Axboe5945b9b2007-02-22 20:33:01 +01001060 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
1061}
Jens Axboe55bc9722007-02-22 11:30:05 +01001062#else
1063void io_u_set_timeout(struct thread_data fio_unused *td)
1064{
1065}
1066#endif
Jens Axboe433afcb2007-02-22 10:39:01 +01001067
Jens Axboe55bc9722007-02-22 11:30:05 +01001068#ifdef FIO_USE_TIMEOUT
Jens Axboe433afcb2007-02-22 10:39:01 +01001069static void io_u_timeout_handler(int fio_unused sig)
1070{
1071 struct thread_data *td, *__td;
1072 pid_t pid = getpid();
Jens Axboe01743ee2008-06-02 12:19:19 +02001073 struct flist_head *entry;
Jens Axboe5945b9b2007-02-22 20:33:01 +01001074 struct io_u *io_u;
Jens Axboe433afcb2007-02-22 10:39:01 +01001075 int i;
1076
1077 log_err("fio: io_u timeout\n");
1078
1079 /*
1080 * TLS would be nice...
1081 */
1082 td = NULL;
1083 for_each_td(__td, i) {
1084 if (__td->pid == pid) {
1085 td = __td;
1086 break;
1087 }
1088 }
1089
1090 if (!td) {
1091 log_err("fio: io_u timeout, can't find job\n");
1092 exit(1);
1093 }
1094
1095 if (!td->cur_depth) {
1096 log_err("fio: timeout without pending work?\n");
1097 return;
1098 }
1099
Jens Axboe15506d02007-03-15 15:04:43 +01001100 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
Jens Axboe5945b9b2007-02-22 20:33:01 +01001101
Jens Axboe01743ee2008-06-02 12:19:19 +02001102 flist_for_each(entry, &td->io_u_busylist) {
1103 io_u = flist_entry(entry, struct io_u, list);
Jens Axboe5945b9b2007-02-22 20:33:01 +01001104
1105 io_u_dump(io_u);
1106 }
1107
1108 td_verror(td, ETIMEDOUT, "io_u timeout");
Jens Axboe433afcb2007-02-22 10:39:01 +01001109 exit(1);
1110}
Jens Axboe55bc9722007-02-22 11:30:05 +01001111#endif
Jens Axboe433afcb2007-02-22 10:39:01 +01001112
1113void io_u_init_timeout(void)
1114{
Jens Axboe55bc9722007-02-22 11:30:05 +01001115#ifdef FIO_USE_TIMEOUT
Jens Axboe433afcb2007-02-22 10:39:01 +01001116 signal(SIGALRM, io_u_timeout_handler);
Jens Axboe55bc9722007-02-22 11:30:05 +01001117#endif
Jens Axboe433afcb2007-02-22 10:39:01 +01001118}