| 1 | /* |
| 2 | * Rated submission helpers |
| 3 | * |
| 4 | * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk> |
| 5 | * |
| 6 | */ |
| 7 | #include "fio.h" |
| 8 | #include "ioengine.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->private; |
| 18 | int ret; |
| 19 | |
| 20 | dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid()); |
| 21 | |
| 22 | io_u_set(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(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->private; |
| 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->private; |
| 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->private = td; |
| 88 | return 0; |
| 89 | } |
| 90 | |
| 91 | static void io_workqueue_free_fn(struct submit_worker *sw) |
| 92 | { |
| 93 | free(sw->private); |
| 94 | sw->private = 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->private; |
| 101 | int fio_unused ret; |
| 102 | |
| 103 | memcpy(&td->o, &parent->o, sizeof(td->o)); |
| 104 | memcpy(&td->ts, &parent->ts, sizeof(td->ts)); |
| 105 | td->o.uid = td->o.gid = -1U; |
| 106 | dup_files(td, parent); |
| 107 | td->eo = parent->eo; |
| 108 | fio_options_mem_dupe(td); |
| 109 | |
| 110 | if (ioengine_load(td)) |
| 111 | goto err; |
| 112 | |
| 113 | if (td->o.odirect) |
| 114 | td->io_ops->flags |= FIO_RAWIO; |
| 115 | |
| 116 | td->pid = gettid(); |
| 117 | |
| 118 | INIT_FLIST_HEAD(&td->io_log_list); |
| 119 | INIT_FLIST_HEAD(&td->io_hist_list); |
| 120 | INIT_FLIST_HEAD(&td->verify_list); |
| 121 | INIT_FLIST_HEAD(&td->trim_list); |
| 122 | INIT_FLIST_HEAD(&td->next_rand_list); |
| 123 | td->io_hist_tree = RB_ROOT; |
| 124 | |
| 125 | td->o.iodepth = 1; |
| 126 | if (td_io_init(td)) |
| 127 | goto err_io_init; |
| 128 | |
| 129 | fio_gettime(&td->epoch, NULL); |
| 130 | fio_getrusage(&td->ru_start); |
| 131 | clear_io_state(td, 1); |
| 132 | |
| 133 | td_set_runstate(td, TD_RUNNING); |
| 134 | td->flags |= TD_F_CHILD; |
| 135 | td->parent = parent; |
| 136 | return 0; |
| 137 | |
| 138 | err_io_init: |
| 139 | close_ioengine(td); |
| 140 | err: |
| 141 | return 1; |
| 142 | |
| 143 | } |
| 144 | |
| 145 | static void io_workqueue_exit_worker_fn(struct submit_worker *sw, |
| 146 | unsigned int *sum_cnt) |
| 147 | { |
| 148 | struct thread_data *td = sw->private; |
| 149 | |
| 150 | (*sum_cnt)++; |
| 151 | sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1); |
| 152 | |
| 153 | fio_options_free(td); |
| 154 | close_and_free_files(td); |
| 155 | if (td->io_ops) |
| 156 | close_ioengine(td); |
| 157 | td_set_runstate(td, TD_EXITED); |
| 158 | } |
| 159 | |
| 160 | #ifdef CONFIG_SFAA |
| 161 | static void sum_val(uint64_t *dst, uint64_t *src) |
| 162 | { |
| 163 | if (*src) { |
| 164 | __sync_fetch_and_add(dst, *src); |
| 165 | *src = 0; |
| 166 | } |
| 167 | } |
| 168 | #else |
| 169 | static void sum_val(uint64_t *dst, uint64_t *src) |
| 170 | { |
| 171 | if (*src) { |
| 172 | *dst += *src; |
| 173 | *src = 0; |
| 174 | } |
| 175 | } |
| 176 | #endif |
| 177 | |
| 178 | static void pthread_double_unlock(pthread_mutex_t *lock1, |
| 179 | pthread_mutex_t *lock2) |
| 180 | { |
| 181 | #ifndef CONFIG_SFAA |
| 182 | pthread_mutex_unlock(lock1); |
| 183 | pthread_mutex_unlock(lock2); |
| 184 | #endif |
| 185 | } |
| 186 | |
| 187 | static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2) |
| 188 | { |
| 189 | #ifndef CONFIG_SFAA |
| 190 | if (lock1 < lock2) { |
| 191 | pthread_mutex_lock(lock1); |
| 192 | pthread_mutex_lock(lock2); |
| 193 | } else { |
| 194 | pthread_mutex_lock(lock2); |
| 195 | pthread_mutex_lock(lock1); |
| 196 | } |
| 197 | #endif |
| 198 | } |
| 199 | |
| 200 | static void sum_ddir(struct thread_data *dst, struct thread_data *src, |
| 201 | enum fio_ddir ddir) |
| 202 | { |
| 203 | pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock); |
| 204 | |
| 205 | sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]); |
| 206 | sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]); |
| 207 | sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]); |
| 208 | sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]); |
| 209 | sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]); |
| 210 | |
| 211 | pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock); |
| 212 | } |
| 213 | |
| 214 | static void io_workqueue_update_acct_fn(struct submit_worker *sw) |
| 215 | { |
| 216 | struct thread_data *src = sw->private; |
| 217 | struct thread_data *dst = sw->wq->td; |
| 218 | |
| 219 | if (td_read(src)) |
| 220 | sum_ddir(dst, src, DDIR_READ); |
| 221 | if (td_write(src)) |
| 222 | sum_ddir(dst, src, DDIR_WRITE); |
| 223 | if (td_trim(src)) |
| 224 | sum_ddir(dst, src, DDIR_TRIM); |
| 225 | |
| 226 | } |
| 227 | |
| 228 | static struct workqueue_ops rated_wq_ops = { |
| 229 | .fn = io_workqueue_fn, |
| 230 | .pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn, |
| 231 | .pre_sleep_fn = io_workqueue_pre_sleep_fn, |
| 232 | .update_acct_fn = io_workqueue_update_acct_fn, |
| 233 | .alloc_worker_fn = io_workqueue_alloc_fn, |
| 234 | .free_worker_fn = io_workqueue_free_fn, |
| 235 | .init_worker_fn = io_workqueue_init_worker_fn, |
| 236 | .exit_worker_fn = io_workqueue_exit_worker_fn, |
| 237 | }; |
| 238 | |
| 239 | int rate_submit_init(struct thread_data *td, struct sk_out *sk_out) |
| 240 | { |
| 241 | if (td->o.io_submit_mode != IO_MODE_OFFLOAD) |
| 242 | return 0; |
| 243 | |
| 244 | return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out); |
| 245 | } |
| 246 | |
| 247 | void rate_submit_exit(struct thread_data *td) |
| 248 | { |
| 249 | if (td->o.io_submit_mode != IO_MODE_OFFLOAD) |
| 250 | return; |
| 251 | |
| 252 | workqueue_exit(&td->io_wq); |
| 253 | } |