1 /*
2 * Rated submission helpers
3 *
4 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
5 *
6 */
7 #include "fio.h"
8 #include "ioengines.h"
9 #include "lib/getrusage.h"
10 #include "rate-submit.h"
11
io_workqueue_fn(struct submit_worker * sw,struct workqueue_work * work)12 static int io_workqueue_fn(struct submit_worker *sw,
13 struct workqueue_work *work)
14 {
15 struct io_u *io_u = container_of(work, struct io_u, work);
16 const enum fio_ddir ddir = io_u->ddir;
17 struct thread_data *td = sw->priv;
18 int ret;
19
20 dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
21
22 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
23
24 td->cur_depth++;
25
26 do {
27 ret = td_io_queue(td, io_u);
28 if (ret != FIO_Q_BUSY)
29 break;
30 ret = io_u_queued_complete(td, 1);
31 if (ret > 0)
32 td->cur_depth -= ret;
33 io_u_clear(td, io_u, IO_U_F_FLIGHT);
34 } while (1);
35
36 dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
37
38 io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
39
40 if (ret == FIO_Q_COMPLETED)
41 td->cur_depth--;
42 else if (ret == FIO_Q_QUEUED) {
43 unsigned int min_evts;
44
45 if (td->o.iodepth == 1)
46 min_evts = 1;
47 else
48 min_evts = 0;
49
50 ret = io_u_queued_complete(td, min_evts);
51 if (ret > 0)
52 td->cur_depth -= ret;
53 } else if (ret == FIO_Q_BUSY) {
54 ret = io_u_queued_complete(td, td->cur_depth);
55 if (ret > 0)
56 td->cur_depth -= ret;
57 }
58
59 return 0;
60 }
61
io_workqueue_pre_sleep_flush_fn(struct submit_worker * sw)62 static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
63 {
64 struct thread_data *td = sw->priv;
65
66 if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
67 return true;
68
69 return false;
70 }
71
io_workqueue_pre_sleep_fn(struct submit_worker * sw)72 static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
73 {
74 struct thread_data *td = sw->priv;
75 int ret;
76
77 ret = io_u_quiesce(td);
78 if (ret > 0)
79 td->cur_depth -= ret;
80 }
81
io_workqueue_alloc_fn(struct submit_worker * sw)82 static int io_workqueue_alloc_fn(struct submit_worker *sw)
83 {
84 struct thread_data *td;
85
86 td = calloc(1, sizeof(*td));
87 sw->priv = td;
88 return 0;
89 }
90
io_workqueue_free_fn(struct submit_worker * sw)91 static void io_workqueue_free_fn(struct submit_worker *sw)
92 {
93 free(sw->priv);
94 sw->priv = NULL;
95 }
96
io_workqueue_init_worker_fn(struct submit_worker * sw)97 static int io_workqueue_init_worker_fn(struct submit_worker *sw)
98 {
99 struct thread_data *parent = sw->wq->td;
100 struct thread_data *td = sw->priv;
101
102 memcpy(&td->o, &parent->o, sizeof(td->o));
103 memcpy(&td->ts, &parent->ts, sizeof(td->ts));
104 td->o.uid = td->o.gid = -1U;
105 dup_files(td, parent);
106 td->eo = parent->eo;
107 fio_options_mem_dupe(td);
108
109 if (ioengine_load(td))
110 goto err;
111
112 td->pid = gettid();
113
114 INIT_FLIST_HEAD(&td->io_log_list);
115 INIT_FLIST_HEAD(&td->io_hist_list);
116 INIT_FLIST_HEAD(&td->verify_list);
117 INIT_FLIST_HEAD(&td->trim_list);
118 INIT_FLIST_HEAD(&td->next_rand_list);
119 td->io_hist_tree = RB_ROOT;
120
121 td->o.iodepth = 1;
122 if (td_io_init(td))
123 goto err_io_init;
124
125 set_epoch_time(td, td->o.log_unix_epoch);
126 fio_getrusage(&td->ru_start);
127 clear_io_state(td, 1);
128
129 td_set_runstate(td, TD_RUNNING);
130 td->flags |= TD_F_CHILD;
131 td->parent = parent;
132 return 0;
133
134 err_io_init:
135 close_ioengine(td);
136 err:
137 return 1;
138
139 }
140
io_workqueue_exit_worker_fn(struct submit_worker * sw,unsigned int * sum_cnt)141 static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
142 unsigned int *sum_cnt)
143 {
144 struct thread_data *td = sw->priv;
145
146 (*sum_cnt)++;
147 sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
148
149 fio_options_free(td);
150 close_and_free_files(td);
151 if (td->io_ops)
152 close_ioengine(td);
153 td_set_runstate(td, TD_EXITED);
154 }
155
156 #ifdef CONFIG_SFAA
sum_val(uint64_t * dst,uint64_t * src)157 static void sum_val(uint64_t *dst, uint64_t *src)
158 {
159 if (*src) {
160 __sync_fetch_and_add(dst, *src);
161 *src = 0;
162 }
163 }
164 #else
sum_val(uint64_t * dst,uint64_t * src)165 static void sum_val(uint64_t *dst, uint64_t *src)
166 {
167 if (*src) {
168 *dst += *src;
169 *src = 0;
170 }
171 }
172 #endif
173
pthread_double_unlock(pthread_mutex_t * lock1,pthread_mutex_t * lock2)174 static void pthread_double_unlock(pthread_mutex_t *lock1,
175 pthread_mutex_t *lock2)
176 {
177 #ifndef CONFIG_SFAA
178 pthread_mutex_unlock(lock1);
179 pthread_mutex_unlock(lock2);
180 #endif
181 }
182
pthread_double_lock(pthread_mutex_t * lock1,pthread_mutex_t * lock2)183 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
184 {
185 #ifndef CONFIG_SFAA
186 if (lock1 < lock2) {
187 pthread_mutex_lock(lock1);
188 pthread_mutex_lock(lock2);
189 } else {
190 pthread_mutex_lock(lock2);
191 pthread_mutex_lock(lock1);
192 }
193 #endif
194 }
195
sum_ddir(struct thread_data * dst,struct thread_data * src,enum fio_ddir ddir)196 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
197 enum fio_ddir ddir)
198 {
199 pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
200
201 sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
202 sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
203 sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
204 sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
205 sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
206
207 pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
208 }
209
io_workqueue_update_acct_fn(struct submit_worker * sw)210 static void io_workqueue_update_acct_fn(struct submit_worker *sw)
211 {
212 struct thread_data *src = sw->priv;
213 struct thread_data *dst = sw->wq->td;
214
215 if (td_read(src))
216 sum_ddir(dst, src, DDIR_READ);
217 if (td_write(src))
218 sum_ddir(dst, src, DDIR_WRITE);
219 if (td_trim(src))
220 sum_ddir(dst, src, DDIR_TRIM);
221
222 }
223
224 static struct workqueue_ops rated_wq_ops = {
225 .fn = io_workqueue_fn,
226 .pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
227 .pre_sleep_fn = io_workqueue_pre_sleep_fn,
228 .update_acct_fn = io_workqueue_update_acct_fn,
229 .alloc_worker_fn = io_workqueue_alloc_fn,
230 .free_worker_fn = io_workqueue_free_fn,
231 .init_worker_fn = io_workqueue_init_worker_fn,
232 .exit_worker_fn = io_workqueue_exit_worker_fn,
233 };
234
rate_submit_init(struct thread_data * td,struct sk_out * sk_out)235 int rate_submit_init(struct thread_data *td, struct sk_out *sk_out)
236 {
237 if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
238 return 0;
239
240 return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
241 }
242
rate_submit_exit(struct thread_data * td)243 void rate_submit_exit(struct thread_data *td)
244 {
245 if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
246 return;
247
248 workqueue_exit(&td->io_wq);
249 }
250