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
 158         if (ioengine_load(td))
 159                 goto err;
 160
 161         td->pid = gettid();
 162
 163         INIT_FLIST_HEAD(&td->io_log_list);
 164         INIT_FLIST_HEAD(&td->io_hist_list);
 165         INIT_FLIST_HEAD(&td->verify_list);
 166         INIT_FLIST_HEAD(&td->trim_list);
 167         td->io_hist_tree = RB_ROOT;
 168
 169         td->o.iodepth = 1;
 170         if (td_io_init(td))
 171                 goto err_io_init;
 172
 173         if (td->io_ops->post_init && td->io_ops->post_init(td))
 174                 goto err_io_init;
 175
 176         set_epoch_time(td, td->o.log_unix_epoch);
 177         fio_getrusage(&td->ru_start);
 178         clear_io_state(td, 1);
 179
 180         td_set_runstate(td, TD_RUNNING);
 181         td->flags |= TD_F_CHILD | TD_F_NEED_LOCK;
 182         td->parent = parent;
 183         return 0;
 184
 185 err_io_init:
 186         close_ioengine(td);
 187 err:
 188         return 1;
 189
 190 }
 191
 192 static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
 193                                         unsigned int *sum_cnt)
 194 {
 195         struct thread_data *td = sw->priv;
 196
 197         (*sum_cnt)++;
 198         sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
 199
 200         fio_options_free(td);
 201         close_and_free_files(td);
 202         if (td->io_ops)
 203                 close_ioengine(td);
 204         td_set_runstate(td, TD_EXITED);
 205 }
 206
 207 #ifdef CONFIG_SFAA
 208 static void sum_val(uint64_t *dst, uint64_t *src)
 209 {
 210         if (*src) {
 211                 __sync_fetch_and_add(dst, *src);
 212                 *src = 0;
 213         }
 214 }
 215 #else
 216 static void sum_val(uint64_t *dst, uint64_t *src)
 217 {
 218         if (*src) {
 219                 *dst += *src;
 220                 *src = 0;
 221         }
 222 }
 223 #endif
 224
 225 static void pthread_double_unlock(pthread_mutex_t *lock1,
 226                                   pthread_mutex_t *lock2)
 227 {
 228 #ifndef CONFIG_SFAA
 229         pthread_mutex_unlock(lock1);
 230         pthread_mutex_unlock(lock2);
 231 #endif
 232 }
 233
 234 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
 235 {
 236 #ifndef CONFIG_SFAA
 237         if (lock1 < lock2) {
 238                 pthread_mutex_lock(lock1);
 239                 pthread_mutex_lock(lock2);
 240         } else {
 241                 pthread_mutex_lock(lock2);
 242                 pthread_mutex_lock(lock1);
 243         }
 244 #endif
 245 }
 246
 247 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 248                      enum fio_ddir ddir)
 249 {
 250         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 251
 252         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 253         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 254         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 255         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 256         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 257
 258         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 259 }
 260
 261 static void io_workqueue_update_acct_fn(struct submit_worker *sw)
 262 {
 263         struct thread_data *src = sw->priv;
 264         struct thread_data *dst = sw->wq->td;
 265
 266         if (td_read(src))
 267                 sum_ddir(dst, src, DDIR_READ);
 268         if (td_write(src))
 269                 sum_ddir(dst, src, DDIR_WRITE);
 270         if (td_trim(src))
 271                 sum_ddir(dst, src, DDIR_TRIM);
 272
 273 }
 274
 275 static struct workqueue_ops rated_wq_ops = {
 276         .fn                     = io_workqueue_fn,
 277         .pre_sleep_flush_fn     = io_workqueue_pre_sleep_flush_fn,
 278         .pre_sleep_fn           = io_workqueue_pre_sleep_fn,
 279         .update_acct_fn         = io_workqueue_update_acct_fn,
 280         .alloc_worker_fn        = io_workqueue_alloc_fn,
 281         .free_worker_fn         = io_workqueue_free_fn,
 282         .init_worker_fn         = io_workqueue_init_worker_fn,
 283         .exit_worker_fn         = io_workqueue_exit_worker_fn,
 284 };
 285
 286 int rate_submit_init(struct thread_data *td, struct sk_out *sk_out)
 287 {
 288         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
 289                 return 0;
 290
 291         return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
 292 }
 293
 294 void rate_submit_exit(struct thread_data *td)
 295 {
 296         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
 297                 return;
 298
 299         workqueue_exit(&td->io_wq);
 300 }