NVML references renamed to PMDK
[fio.git] / rate-submit.c
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
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
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
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
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
91 static void io_workqueue_free_fn(struct submit_worker *sw)
92 {
93         free(sw->priv);
94         sw->priv = NULL;
95 }
96
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
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
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
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
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
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
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
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
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
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 }