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         }
  87
  88         return 0;
  89 }
  90
  91 static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
  92 {
  93         struct thread_data *td = sw->priv;
  94
  95         if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
  96                 return true;
  97
  98         return false;
  99 }
 100
 101 static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
 102 {
 103         struct thread_data *td = sw->priv;
 104         int ret;
 105
 106         ret = io_u_quiesce(td);
 107         if (ret > 0)
 108                 td->cur_depth -= ret;
 109 }
 110
 111 static int io_workqueue_alloc_fn(struct submit_worker *sw)
 112 {
 113         struct thread_data *td;
 114
 115         td = calloc(1, sizeof(*td));
 116         sw->priv = td;
 117         return 0;
 118 }
 119
 120 static void io_workqueue_free_fn(struct submit_worker *sw)
 121 {
 122         free(sw->priv);
 123         sw->priv = NULL;
 124 }
 125
 126 static int io_workqueue_init_worker_fn(struct submit_worker *sw)
 127 {
 128         struct thread_data *parent = sw->wq->td;
 129         struct thread_data *td = sw->priv;
 130
 131         memcpy(&td->o, &parent->o, sizeof(td->o));
 132         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
 133         td->o.uid = td->o.gid = -1U;
 134         dup_files(td, parent);
 135         td->eo = parent->eo;
 136         fio_options_mem_dupe(td);
 137
 138         if (ioengine_load(td))
 139                 goto err;
 140
 141         td->pid = gettid();
 142
 143         INIT_FLIST_HEAD(&td->io_log_list);
 144         INIT_FLIST_HEAD(&td->io_hist_list);
 145         INIT_FLIST_HEAD(&td->verify_list);
 146         INIT_FLIST_HEAD(&td->trim_list);
 147         td->io_hist_tree = RB_ROOT;
 148
 149         td->o.iodepth = 1;
 150         if (td_io_init(td))
 151                 goto err_io_init;
 152
 153         set_epoch_time(td, td->o.log_unix_epoch);
 154         fio_getrusage(&td->ru_start);
 155         clear_io_state(td, 1);
 156
 157         td_set_runstate(td, TD_RUNNING);
 158         td->flags |= TD_F_CHILD | TD_F_NEED_LOCK;
 159         td->parent = parent;
 160         return 0;
 161
 162 err_io_init:
 163         close_ioengine(td);
 164 err:
 165         return 1;
 166
 167 }
 168
 169 static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
 170                                         unsigned int *sum_cnt)
 171 {
 172         struct thread_data *td = sw->priv;
 173
 174         (*sum_cnt)++;
 175         sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
 176
 177         fio_options_free(td);
 178         close_and_free_files(td);
 179         if (td->io_ops)
 180                 close_ioengine(td);
 181         td_set_runstate(td, TD_EXITED);
 182 }
 183
 184 #ifdef CONFIG_SFAA
 185 static void sum_val(uint64_t *dst, uint64_t *src)
 186 {
 187         if (*src) {
 188                 __sync_fetch_and_add(dst, *src);
 189                 *src = 0;
 190         }
 191 }
 192 #else
 193 static void sum_val(uint64_t *dst, uint64_t *src)
 194 {
 195         if (*src) {
 196                 *dst += *src;
 197                 *src = 0;
 198         }
 199 }
 200 #endif
 201
 202 static void pthread_double_unlock(pthread_mutex_t *lock1,
 203                                   pthread_mutex_t *lock2)
 204 {
 205 #ifndef CONFIG_SFAA
 206         pthread_mutex_unlock(lock1);
 207         pthread_mutex_unlock(lock2);
 208 #endif
 209 }
 210
 211 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
 212 {
 213 #ifndef CONFIG_SFAA
 214         if (lock1 < lock2) {
 215                 pthread_mutex_lock(lock1);
 216                 pthread_mutex_lock(lock2);
 217         } else {
 218                 pthread_mutex_lock(lock2);
 219                 pthread_mutex_lock(lock1);
 220         }
 221 #endif
 222 }
 223
 224 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 225                      enum fio_ddir ddir)
 226 {
 227         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 228
 229         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 230         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 231         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 232         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 233         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 234
 235         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 236 }
 237
 238 static void io_workqueue_update_acct_fn(struct submit_worker *sw)
 239 {
 240         struct thread_data *src = sw->priv;
 241         struct thread_data *dst = sw->wq->td;
 242
 243         if (td_read(src))
 244                 sum_ddir(dst, src, DDIR_READ);
 245         if (td_write(src))
 246                 sum_ddir(dst, src, DDIR_WRITE);
 247         if (td_trim(src))
 248                 sum_ddir(dst, src, DDIR_TRIM);
 249
 250 }
 251
 252 static struct workqueue_ops rated_wq_ops = {
 253         .fn                     = io_workqueue_fn,
 254         .pre_sleep_flush_fn     = io_workqueue_pre_sleep_flush_fn,
 255         .pre_sleep_fn           = io_workqueue_pre_sleep_fn,
 256         .update_acct_fn         = io_workqueue_update_acct_fn,
 257         .alloc_worker_fn        = io_workqueue_alloc_fn,
 258         .free_worker_fn         = io_workqueue_free_fn,
 259         .init_worker_fn         = io_workqueue_init_worker_fn,
 260         .exit_worker_fn         = io_workqueue_exit_worker_fn,
 261 };
 262
 263 int rate_submit_init(struct thread_data *td, struct sk_out *sk_out)
 264 {
 265         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
 266                 return 0;
 267
 268         return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
 269 }
 270
 271 void rate_submit_exit(struct thread_data *td)
 272 {
 273         if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
 274                 return;
 275
 276         workqueue_exit(&td->io_wq);
 277 }