Revert "ioengines.c:346: td_io_queue: Assertion `res == 0' failed"
[fio.git] / rate-submit.c
1 /*
2  * Rated submission helpers
3  *
4  * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
5  *
6  */
7 #include <assert.h>
8 #include "fio.h"
9 #include "ioengines.h"
10 #include "lib/getrusage.h"
11 #include "rate-submit.h"
12
13 static void check_overlap(struct io_u *io_u)
14 {
15         int i, res;
16         struct thread_data *td;
17
18         /*
19          * Allow only one thread to check for overlap at a time to prevent two
20          * threads from thinking the coast is clear and then submitting IOs
21          * that overlap with each other.
22          *
23          * If an overlap is found, release the lock and re-acquire it before
24          * checking again to give other threads a chance to make progress.
25          *
26          * If no overlap is found, release the lock when the io_u's
27          * IO_U_F_FLIGHT flag is set so that this io_u can be checked by other
28          * threads as they assess overlap.
29          */
30         res = pthread_mutex_lock(&overlap_check);
31         assert(res == 0);
32
33 retry:
34         for_each_td(td, i) {
35                 if (td->runstate <= TD_SETTING_UP ||
36                     td->runstate >= TD_FINISHING ||
37                     !td->o.serialize_overlap ||
38                     td->o.io_submit_mode != IO_MODE_OFFLOAD)
39                         continue;
40
41                 if (!in_flight_overlap(&td->io_u_all, io_u))
42                         continue;
43
44                 res = pthread_mutex_unlock(&overlap_check);
45                 assert(res == 0);
46                 res = pthread_mutex_lock(&overlap_check);
47                 assert(res == 0);
48                 goto retry;
49         }
50 }
51
52 static int io_workqueue_fn(struct submit_worker *sw,
53                            struct workqueue_work *work)
54 {
55         struct io_u *io_u = container_of(work, struct io_u, work);
56         const enum fio_ddir ddir = io_u->ddir;
57         struct thread_data *td = sw->priv;
58         int ret, error;
59
60         if (td->o.serialize_overlap)
61                 check_overlap(io_u);
62
63         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
64
65         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
66
67         td->cur_depth++;
68
69         do {
70                 ret = td_io_queue(td, io_u);
71                 if (ret != FIO_Q_BUSY)
72                         break;
73                 ret = io_u_queued_complete(td, 1);
74                 if (ret > 0)
75                         td->cur_depth -= ret;
76                 else if (ret < 0)
77                         break;
78                 io_u_clear(td, io_u, IO_U_F_FLIGHT);
79         } while (1);
80
81         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
82
83         error = io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
84
85         if (ret == FIO_Q_COMPLETED)
86                 td->cur_depth--;
87         else if (ret == FIO_Q_QUEUED) {
88                 unsigned int min_evts;
89
90                 if (td->o.iodepth == 1)
91                         min_evts = 1;
92                 else
93                         min_evts = 0;
94
95                 ret = io_u_queued_complete(td, min_evts);
96                 if (ret > 0)
97                         td->cur_depth -= ret;
98         }
99
100         if (error || td->error) {
101                 pthread_mutex_lock(&td->io_u_lock);
102                 pthread_cond_signal(&td->parent->free_cond);
103                 pthread_mutex_unlock(&td->io_u_lock);
104         }
105
106         return 0;
107 }
108
109 static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
110 {
111         struct thread_data *td = sw->priv;
112
113         if (td->error)
114                 return false;
115         if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
116                 return true;
117
118         return false;
119 }
120
121 static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
122 {
123         struct thread_data *td = sw->priv;
124         int ret;
125
126         ret = io_u_quiesce(td);
127         if (ret > 0)
128                 td->cur_depth -= ret;
129 }
130
131 static int io_workqueue_alloc_fn(struct submit_worker *sw)
132 {
133         struct thread_data *td;
134
135         td = calloc(1, sizeof(*td));
136         sw->priv = td;
137         return 0;
138 }
139
140 static void io_workqueue_free_fn(struct submit_worker *sw)
141 {
142         free(sw->priv);
143         sw->priv = NULL;
144 }
145
146 static int io_workqueue_init_worker_fn(struct submit_worker *sw)
147 {
148         struct thread_data *parent = sw->wq->td;
149         struct thread_data *td = sw->priv;
150
151         memcpy(&td->o, &parent->o, sizeof(td->o));
152         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
153         td->o.uid = td->o.gid = -1U;
154         dup_files(td, parent);
155         td->eo = parent->eo;
156         fio_options_mem_dupe(td);
157         td->iolog_f = parent->iolog_f;
158
159         if (ioengine_load(td))
160                 goto err;
161
162         td->pid = gettid();
163
164         INIT_FLIST_HEAD(&td->io_log_list);
165         INIT_FLIST_HEAD(&td->io_hist_list);
166         INIT_FLIST_HEAD(&td->verify_list);
167         INIT_FLIST_HEAD(&td->trim_list);
168         td->io_hist_tree = RB_ROOT;
169
170         td->o.iodepth = 1;
171         if (td_io_init(td))
172                 goto err_io_init;
173
174         if (td->io_ops->post_init && td->io_ops->post_init(td))
175                 goto err_io_init;
176
177         set_epoch_time(td, td->o.log_unix_epoch | td->o.log_alternate_epoch, td->o.log_alternate_epoch_clock_id);
178         fio_getrusage(&td->ru_start);
179         clear_io_state(td, 1);
180
181         td_set_runstate(td, TD_RUNNING);
182         td->flags |= TD_F_CHILD | TD_F_NEED_LOCK;
183         td->parent = parent;
184         return 0;
185
186 err_io_init:
187         close_ioengine(td);
188 err:
189         return 1;
190
191 }
192
193 static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
194                                         unsigned int *sum_cnt)
195 {
196         struct thread_data *td = sw->priv;
197
198         (*sum_cnt)++;
199
200         /*
201          * io_workqueue_update_acct_fn() doesn't support per prio stats, and
202          * even if it did, offload can't be used with all async IO engines.
203          * If group reporting is set in the parent td, the group result
204          * generated by __show_run_stats() can still contain multiple prios
205          * from different offloaded jobs.
206          */
207         sw->wq->td->ts.disable_prio_stat = 1;
208         sum_thread_stats(&sw->wq->td->ts, &td->ts);
209
210         fio_options_free(td);
211         close_and_free_files(td);
212         if (td->io_ops)
213                 close_ioengine(td);
214         td_set_runstate(td, TD_EXITED);
215 }
216
217 #ifdef CONFIG_SFAA
218 static void sum_val(uint64_t *dst, uint64_t *src)
219 {
220         if (*src) {
221                 __sync_fetch_and_add(dst, *src);
222                 *src = 0;
223         }
224 }
225 #else
226 static void sum_val(uint64_t *dst, uint64_t *src)
227 {
228         if (*src) {
229                 *dst += *src;
230                 *src = 0;
231         }
232 }
233 #endif
234
235 static void pthread_double_unlock(pthread_mutex_t *lock1,
236                                   pthread_mutex_t *lock2)
237 {
238 #ifndef CONFIG_SFAA
239         pthread_mutex_unlock(lock1);
240         pthread_mutex_unlock(lock2);
241 #endif
242 }
243
244 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
245 {
246 #ifndef CONFIG_SFAA
247         if (lock1 < lock2) {
248                 pthread_mutex_lock(lock1);
249                 pthread_mutex_lock(lock2);
250         } else {
251                 pthread_mutex_lock(lock2);
252                 pthread_mutex_lock(lock1);
253         }
254 #endif
255 }
256
257 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
258                      enum fio_ddir ddir)
259 {
260         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
261
262         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
263         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
264         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
265         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
266         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
267         if (ddir == DDIR_READ)
268                 sum_val(&dst->bytes_verified, &src->bytes_verified);
269
270         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
271 }
272
273 static void io_workqueue_update_acct_fn(struct submit_worker *sw)
274 {
275         struct thread_data *src = sw->priv;
276         struct thread_data *dst = sw->wq->td;
277
278         if (td_read(src))
279                 sum_ddir(dst, src, DDIR_READ);
280         if (td_write(src))
281                 sum_ddir(dst, src, DDIR_WRITE);
282         if (td_trim(src))
283                 sum_ddir(dst, src, DDIR_TRIM);
284
285 }
286
287 static struct workqueue_ops rated_wq_ops = {
288         .fn                     = io_workqueue_fn,
289         .pre_sleep_flush_fn     = io_workqueue_pre_sleep_flush_fn,
290         .pre_sleep_fn           = io_workqueue_pre_sleep_fn,
291         .update_acct_fn         = io_workqueue_update_acct_fn,
292         .alloc_worker_fn        = io_workqueue_alloc_fn,
293         .free_worker_fn         = io_workqueue_free_fn,
294         .init_worker_fn         = io_workqueue_init_worker_fn,
295         .exit_worker_fn         = io_workqueue_exit_worker_fn,
296 };
297
298 int rate_submit_init(struct thread_data *td, struct sk_out *sk_out)
299 {
300         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
301                 return 0;
302
303         return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
304 }
305
306 void rate_submit_exit(struct thread_data *td)
307 {
308         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
309                 return;
310
311         workqueue_exit(&td->io_wq);
312 }