fs/btrfs/async-thread.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  * Copyright (C) 2014 Fujitsu.  All rights reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public
   7  * License v2 as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public
  15  * License along with this program; if not, write to the
  16  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  17  * Boston, MA 021110-1307, USA.
  18  */
  19
  20 #include <linux/kthread.h>
  21 #include <linux/slab.h>
  22 #include <linux/list.h>
  23 #include <linux/spinlock.h>
  24 #include <linux/freezer.h>
  25 #include "async-thread.h"
  26 #include "ctree.h"
  27
  28 #define WORK_DONE_BIT 0
  29 #define WORK_ORDER_DONE_BIT 1
  30 #define WORK_HIGH_PRIO_BIT 2
  31
  32 #define NO_THRESHOLD (-1)
  33 #define DFT_THRESHOLD (32)
  34
  35 struct __btrfs_workqueue {
  36         struct workqueue_struct *normal_wq;
  37         /* List head pointing to ordered work list */
  38         struct list_head ordered_list;
  39
  40         /* Spinlock for ordered_list */
  41         spinlock_t list_lock;
  42
  43         /* Thresholding related variants */
  44         atomic_t pending;
  45
  46         /* Up limit of concurrency workers */
  47         int limit_active;
  48
  49         /* Current number of concurrency workers */
  50         int current_active;
  51
  52         /* Threshold to change current_active */
  53         int thresh;
  54         unsigned int count;
  55         spinlock_t thres_lock;
  56 };
  57
  58 struct btrfs_workqueue {
  59         struct __btrfs_workqueue *normal;
  60         struct __btrfs_workqueue *high;
  61 };
  62
  63 static void normal_work_helper(struct btrfs_work *work);
  64
  65 #define BTRFS_WORK_HELPER(name)                                 \
  66 void btrfs_##name(struct work_struct *arg)                              \
  67 {                                                                       \
  68         struct btrfs_work *work = container_of(arg, struct btrfs_work,  \
  69                                                normal_work);            \
  70         normal_work_helper(work);                                       \
  71 }
  72
  73 bool btrfs_workqueue_normal_congested(struct btrfs_workqueue *wq)
  74 {
  75         /*
  76          * We could compare wq->normal->pending with num_online_cpus()
  77          * to support "thresh == NO_THRESHOLD" case, but it requires
  78          * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
  79          * postpone it until someone needs the support of that case.
  80          */
  81         if (wq->normal->thresh == NO_THRESHOLD)
  82                 return false;
  83
  84         return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
  85 }
  86
  87 BTRFS_WORK_HELPER(worker_helper);
  88 BTRFS_WORK_HELPER(delalloc_helper);
  89 BTRFS_WORK_HELPER(flush_delalloc_helper);
  90 BTRFS_WORK_HELPER(cache_helper);
  91 BTRFS_WORK_HELPER(submit_helper);
  92 BTRFS_WORK_HELPER(fixup_helper);
  93 BTRFS_WORK_HELPER(endio_helper);
  94 BTRFS_WORK_HELPER(endio_meta_helper);
  95 BTRFS_WORK_HELPER(endio_meta_write_helper);
  96 BTRFS_WORK_HELPER(endio_raid56_helper);
  97 BTRFS_WORK_HELPER(endio_repair_helper);
  98 BTRFS_WORK_HELPER(rmw_helper);
  99 BTRFS_WORK_HELPER(endio_write_helper);
 100 BTRFS_WORK_HELPER(freespace_write_helper);
 101 BTRFS_WORK_HELPER(delayed_meta_helper);
 102 BTRFS_WORK_HELPER(readahead_helper);
 103 BTRFS_WORK_HELPER(qgroup_rescan_helper);
 104 BTRFS_WORK_HELPER(extent_refs_helper);
 105 BTRFS_WORK_HELPER(scrub_helper);
 106 BTRFS_WORK_HELPER(scrubwrc_helper);
 107 BTRFS_WORK_HELPER(scrubnc_helper);
 108 BTRFS_WORK_HELPER(scrubparity_helper);
 109
 110 static struct __btrfs_workqueue *
 111 __btrfs_alloc_workqueue(const char *name, unsigned int flags, int limit_active,
 112                          int thresh)
 113 {
 114         struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
 115
 116         if (!ret)
 117                 return NULL;
 118
 119         ret->limit_active = limit_active;
 120         atomic_set(&ret->pending, 0);
 121         if (thresh == 0)
 122                 thresh = DFT_THRESHOLD;
 123         /* For low threshold, disabling threshold is a better choice */
 124         if (thresh < DFT_THRESHOLD) {
 125                 ret->current_active = limit_active;
 126                 ret->thresh = NO_THRESHOLD;
 127         } else {
 128                 /*
 129                  * For threshold-able wq, let its concurrency grow on demand.
 130                  * Use minimal max_active at alloc time to reduce resource
 131                  * usage.
 132                  */
 133                 ret->current_active = 1;
 134                 ret->thresh = thresh;
 135         }
 136
 137         if (flags & WQ_HIGHPRI)
 138                 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
 139                                                  ret->current_active, "btrfs",
 140                                                  name);
 141         else
 142                 ret->normal_wq = alloc_workqueue("%s-%s", flags,
 143                                                  ret->current_active, "btrfs",
 144                                                  name);
 145         if (!ret->normal_wq) {
 146                 kfree(ret);
 147                 return NULL;
 148         }
 149
 150         INIT_LIST_HEAD(&ret->ordered_list);
 151         spin_lock_init(&ret->list_lock);
 152         spin_lock_init(&ret->thres_lock);
 153         trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
 154         return ret;
 155 }
 156
 157 static inline void
 158 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
 159
 160 struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name,
 161                                               unsigned int flags,
 162                                               int limit_active,
 163                                               int thresh)
 164 {
 165         struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
 166
 167         if (!ret)
 168                 return NULL;
 169
 170         ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI,
 171                                               limit_active, thresh);
 172         if (!ret->normal) {
 173                 kfree(ret);
 174                 return NULL;
 175         }
 176
 177         if (flags & WQ_HIGHPRI) {
 178                 ret->high = __btrfs_alloc_workqueue(name, flags, limit_active,
 179                                                     thresh);
 180                 if (!ret->high) {
 181                         __btrfs_destroy_workqueue(ret->normal);
 182                         kfree(ret);
 183                         return NULL;
 184                 }
 185         }
 186         return ret;
 187 }
 188
 189 /*
 190  * Hook for threshold which will be called in btrfs_queue_work.
 191  * This hook WILL be called in IRQ handler context,
 192  * so workqueue_set_max_active MUST NOT be called in this hook
 193  */
 194 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
 195 {
 196         if (wq->thresh == NO_THRESHOLD)
 197                 return;
 198         atomic_inc(&wq->pending);
 199 }
 200
 201 /*
 202  * Hook for threshold which will be called before executing the work,
 203  * This hook is called in kthread content.
 204  * So workqueue_set_max_active is called here.
 205  */
 206 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
 207 {
 208         int new_current_active;
 209         long pending;
 210         int need_change = 0;
 211
 212         if (wq->thresh == NO_THRESHOLD)
 213                 return;
 214
 215         atomic_dec(&wq->pending);
 216         spin_lock(&wq->thres_lock);
 217         /*
 218          * Use wq->count to limit the calling frequency of
 219          * workqueue_set_max_active.
 220          */
 221         wq->count++;
 222         wq->count %= (wq->thresh / 4);
 223         if (!wq->count)
 224                 goto  out;
 225         new_current_active = wq->current_active;
 226
 227         /*
 228          * pending may be changed later, but it's OK since we really
 229          * don't need it so accurate to calculate new_max_active.
 230          */
 231         pending = atomic_read(&wq->pending);
 232         if (pending > wq->thresh)
 233                 new_current_active++;
 234         if (pending < wq->thresh / 2)
 235                 new_current_active--;
 236         new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
 237         if (new_current_active != wq->current_active)  {
 238                 need_change = 1;
 239                 wq->current_active = new_current_active;
 240         }
 241 out:
 242         spin_unlock(&wq->thres_lock);
 243
 244         if (need_change) {
 245                 workqueue_set_max_active(wq->normal_wq, wq->current_active);
 246         }
 247 }
 248
 249 static void run_ordered_work(struct __btrfs_workqueue *wq)
 250 {
 251         struct list_head *list = &wq->ordered_list;
 252         struct btrfs_work *work;
 253         spinlock_t *lock = &wq->list_lock;
 254         unsigned long flags;
 255
 256         while (1) {
 257                 spin_lock_irqsave(lock, flags);
 258                 if (list_empty(list))
 259                         break;
 260                 work = list_entry(list->next, struct btrfs_work,
 261                                   ordered_list);
 262                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 263                         break;
 264
 265                 /*
 266                  * we are going to call the ordered done function, but
 267                  * we leave the work item on the list as a barrier so
 268                  * that later work items that are done don't have their
 269                  * functions called before this one returns
 270                  */
 271                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 272                         break;
 273                 trace_btrfs_ordered_sched(work);
 274                 spin_unlock_irqrestore(lock, flags);
 275                 work->ordered_func(work);
 276
 277                 /* now take the lock again and drop our item from the list */
 278                 spin_lock_irqsave(lock, flags);
 279                 list_del(&work->ordered_list);
 280                 spin_unlock_irqrestore(lock, flags);
 281
 282                 /*
 283                  * we don't want to call the ordered free functions
 284                  * with the lock held though
 285                  */
 286                 work->ordered_free(work);
 287                 trace_btrfs_all_work_done(work);
 288         }
 289         spin_unlock_irqrestore(lock, flags);
 290 }
 291
 292 static void normal_work_helper(struct btrfs_work *work)
 293 {
 294         struct __btrfs_workqueue *wq;
 295         int need_order = 0;
 296
 297         /*
 298          * We should not touch things inside work in the following cases:
 299          * 1) after work->func() if it has no ordered_free
 300          *    Since the struct is freed in work->func().
 301          * 2) after setting WORK_DONE_BIT
 302          *    The work may be freed in other threads almost instantly.
 303          * So we save the needed things here.
 304          */
 305         if (work->ordered_func)
 306                 need_order = 1;
 307         wq = work->wq;
 308
 309         trace_btrfs_work_sched(work);
 310         thresh_exec_hook(wq);
 311         work->func(work);
 312         if (need_order) {
 313                 set_bit(WORK_DONE_BIT, &work->flags);
 314                 run_ordered_work(wq);
 315         }
 316         if (!need_order)
 317                 trace_btrfs_all_work_done(work);
 318 }
 319
 320 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
 321                      btrfs_func_t func,
 322                      btrfs_func_t ordered_func,
 323                      btrfs_func_t ordered_free)
 324 {
 325         work->func = func;
 326         work->ordered_func = ordered_func;
 327         work->ordered_free = ordered_free;
 328         INIT_WORK(&work->normal_work, uniq_func);
 329         INIT_LIST_HEAD(&work->ordered_list);
 330         work->flags = 0;
 331 }
 332
 333 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
 334                                       struct btrfs_work *work)
 335 {
 336         unsigned long flags;
 337
 338         work->wq = wq;
 339         thresh_queue_hook(wq);
 340         if (work->ordered_func) {
 341                 spin_lock_irqsave(&wq->list_lock, flags);
 342                 list_add_tail(&work->ordered_list, &wq->ordered_list);
 343                 spin_unlock_irqrestore(&wq->list_lock, flags);
 344         }
 345         trace_btrfs_work_queued(work);
 346         queue_work(wq->normal_wq, &work->normal_work);
 347 }
 348
 349 void btrfs_queue_work(struct btrfs_workqueue *wq,
 350                       struct btrfs_work *work)
 351 {
 352         struct __btrfs_workqueue *dest_wq;
 353
 354         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
 355                 dest_wq = wq->high;
 356         else
 357                 dest_wq = wq->normal;
 358         __btrfs_queue_work(dest_wq, work);
 359 }
 360
 361 static inline void
 362 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
 363 {
 364         destroy_workqueue(wq->normal_wq);
 365         trace_btrfs_workqueue_destroy(wq);
 366         kfree(wq);
 367 }
 368
 369 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
 370 {
 371         if (!wq)
 372                 return;
 373         if (wq->high)
 374                 __btrfs_destroy_workqueue(wq->high);
 375         __btrfs_destroy_workqueue(wq->normal);
 376         kfree(wq);
 377 }
 378
 379 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
 380 {
 381         if (!wq)
 382                 return;
 383         wq->normal->limit_active = limit_active;
 384         if (wq->high)
 385                 wq->high->limit_active = limit_active;
 386 }
 387
 388 void btrfs_set_work_high_priority(struct btrfs_work *work)
 389 {
 390         set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
 391 }
 392
 393 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
 394 {
 395         if (wq->high)
 396                 flush_workqueue(wq->high->normal_wq);
 397
 398         flush_workqueue(wq->normal->normal_wq);
 399 }