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 void check_overlap(struct io_u *io_u)
  13 {
  14         int i;
  15         struct thread_data *td;
  16         bool overlap = false;
  17
  18         do {
  19                 /*
  20                  * Allow only one thread to check for overlap at a
  21                  * time to prevent two threads from thinking the coast
  22                  * is clear and then submitting IOs that overlap with
  23                  * each other
  24                  */
  25                 pthread_mutex_lock(&overlap_check);
  26                 for_each_td(td, i) {
  27                         if (td->runstate <= TD_SETTING_UP ||
  28                                 td->runstate >= TD_FINISHING ||
  29                                 !td->o.serialize_overlap ||
  30                                 td->o.io_submit_mode != IO_MODE_OFFLOAD)
  31                                 continue;
  32
  33                         overlap = in_flight_overlap(&td->io_u_all, io_u);
  34                         if (overlap) {
  35                                 pthread_mutex_unlock(&overlap_check);
  36                                 break;
  37                         }
  38                 }
  39         } while (overlap);
  40 }
  41
  42 static int io_workqueue_fn(struct submit_worker *sw,
  43                            struct workqueue_work *work)
  44 {
  45         struct io_u *io_u = container_of(work, struct io_u, work);
  46         const enum fio_ddir ddir = io_u->ddir;
  47         struct thread_data *td = sw->priv;
  48         int ret;
  49
  50         if (td->o.serialize_overlap)
  51                 check_overlap(io_u);
  52
  53         dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
  54
  55         io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
  56
  57         td->cur_depth++;
  58
  59         do {
  60                 ret = td_io_queue(td, io_u);
  61                 if (ret != FIO_Q_BUSY)
  62                         break;
  63                 ret = io_u_queued_complete(td, 1);
  64                 if (ret > 0)
  65                         td->cur_depth -= ret;
  66                 io_u_clear(td, io_u, IO_U_F_FLIGHT);
  67         } while (1);
  68
  69         dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
  70
  71         io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
  72
  73         if (ret == FIO_Q_COMPLETED)
  74                 td->cur_depth--;
  75         else if (ret == FIO_Q_QUEUED) {
  76                 unsigned int min_evts;
  77
  78                 if (td->o.iodepth == 1)
  79                         min_evts = 1;
  80                 else
  81                         min_evts = 0;
  82
  83                 ret = io_u_queued_complete(td, min_evts);
  84                 if (ret > 0)
  85                         td->cur_depth -= ret;
  86         } else if (ret == FIO_Q_BUSY) {
  87                 ret = io_u_queued_complete(td, td->cur_depth);
  88                 if (ret > 0)
  89                         td->cur_depth -= ret;
  90         }
  91
  92         return 0;
  93 }
  94
  95 static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
  96 {
  97         struct thread_data *td = sw->priv;
  98
  99         if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
 100                 return true;
 101
 102         return false;
 103 }
 104
 105 static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
 106 {
 107         struct thread_data *td = sw->priv;
 108         int ret;
 109
 110         ret = io_u_quiesce(td);
 111         if (ret > 0)
 112                 td->cur_depth -= ret;
 113 }
 114
 115 static int io_workqueue_alloc_fn(struct submit_worker *sw)
 116 {
 117         struct thread_data *td;
 118
 119         td = calloc(1, sizeof(*td));
 120         sw->priv = td;
 121         return 0;
 122 }
 123
 124 static void io_workqueue_free_fn(struct submit_worker *sw)
 125 {
 126         free(sw->priv);
 127         sw->priv = NULL;
 128 }
 129
 130 static int io_workqueue_init_worker_fn(struct submit_worker *sw)
 131 {
 132         struct thread_data *parent = sw->wq->td;
 133         struct thread_data *td = sw->priv;
 134
 135         memcpy(&td->o, &parent->o, sizeof(td->o));
 136         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
 137         td->o.uid = td->o.gid = -1U;
 138         dup_files(td, parent);
 139         td->eo = parent->eo;
 140         fio_options_mem_dupe(td);
 141
 142         if (ioengine_load(td))
 143                 goto err;
 144
 145         td->pid = gettid();
 146
 147         INIT_FLIST_HEAD(&td->io_log_list);
 148         INIT_FLIST_HEAD(&td->io_hist_list);
 149         INIT_FLIST_HEAD(&td->verify_list);
 150         INIT_FLIST_HEAD(&td->trim_list);
 151         td->io_hist_tree = RB_ROOT;
 152
 153         td->o.iodepth = 1;
 154         if (td_io_init(td))
 155                 goto err_io_init;
 156
 157         set_epoch_time(td, td->o.log_unix_epoch);
 158         fio_getrusage(&td->ru_start);
 159         clear_io_state(td, 1);
 160
 161         td_set_runstate(td, TD_RUNNING);
 162         td->flags |= TD_F_CHILD | TD_F_NEED_LOCK;
 163         td->parent = parent;
 164         return 0;
 165
 166 err_io_init:
 167         close_ioengine(td);
 168 err:
 169         return 1;
 170
 171 }
 172
 173 static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
 174                                         unsigned int *sum_cnt)
 175 {
 176         struct thread_data *td = sw->priv;
 177
 178         (*sum_cnt)++;
 179         sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
 180
 181         fio_options_free(td);
 182         close_and_free_files(td);
 183         if (td->io_ops)
 184                 close_ioengine(td);
 185         td_set_runstate(td, TD_EXITED);
 186 }
 187
 188 #ifdef CONFIG_SFAA
 189 static void sum_val(uint64_t *dst, uint64_t *src)
 190 {
 191         if (*src) {
 192                 __sync_fetch_and_add(dst, *src);
 193                 *src = 0;
 194         }
 195 }
 196 #else
 197 static void sum_val(uint64_t *dst, uint64_t *src)
 198 {
 199         if (*src) {
 200                 *dst += *src;
 201                 *src = 0;
 202         }
 203 }
 204 #endif
 205
 206 static void pthread_double_unlock(pthread_mutex_t *lock1,
 207                                   pthread_mutex_t *lock2)
 208 {
 209 #ifndef CONFIG_SFAA
 210         pthread_mutex_unlock(lock1);
 211         pthread_mutex_unlock(lock2);
 212 #endif
 213 }
 214
 215 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
 216 {
 217 #ifndef CONFIG_SFAA
 218         if (lock1 < lock2) {
 219                 pthread_mutex_lock(lock1);
 220                 pthread_mutex_lock(lock2);
 221         } else {
 222                 pthread_mutex_lock(lock2);
 223                 pthread_mutex_lock(lock1);
 224         }
 225 #endif
 226 }
 227
 228 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 229                      enum fio_ddir ddir)
 230 {
 231         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 232
 233         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 234         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 235         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 236         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 237         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 238
 239         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 240 }
 241
 242 static void io_workqueue_update_acct_fn(struct submit_worker *sw)
 243 {
 244         struct thread_data *src = sw->priv;
 245         struct thread_data *dst = sw->wq->td;
 246
 247         if (td_read(src))
 248                 sum_ddir(dst, src, DDIR_READ);
 249         if (td_write(src))
 250                 sum_ddir(dst, src, DDIR_WRITE);
 251         if (td_trim(src))
 252                 sum_ddir(dst, src, DDIR_TRIM);
 253
 254 }
 255
 256 static struct workqueue_ops rated_wq_ops = {
 257         .fn                     = io_workqueue_fn,
 258         .pre_sleep_flush_fn     = io_workqueue_pre_sleep_flush_fn,
 259         .pre_sleep_fn           = io_workqueue_pre_sleep_fn,
 260         .update_acct_fn         = io_workqueue_update_acct_fn,
 261         .alloc_worker_fn        = io_workqueue_alloc_fn,
 262         .free_worker_fn         = io_workqueue_free_fn,
 263         .init_worker_fn         = io_workqueue_init_worker_fn,
 264         .exit_worker_fn         = io_workqueue_exit_worker_fn,
 265 };
 266
 267 int rate_submit_init(struct thread_data *td, struct sk_out *sk_out)
 268 {
 269         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
 270                 return 0;
 271
 272         return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
 273 }
 274
 275 void rate_submit_exit(struct thread_data *td)
 276 {
 277         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
 278                 return;
 279
 280         workqueue_exit(&td->io_wq);
 281 }