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         int max_active;
  46         int current_max;
  47         int thresh;
  48         unsigned int count;
  49         spinlock_t thres_lock;
  50 };
  51
  52 struct btrfs_workqueue {
  53         struct __btrfs_workqueue *normal;
  54         struct __btrfs_workqueue *high;
  55 };
  56
  57 static void normal_work_helper(struct btrfs_work *work);
  58
  59 #define BTRFS_WORK_HELPER(name)                                 \
  60 void btrfs_##name(struct work_struct *arg)                              \
  61 {                                                                       \
  62         struct btrfs_work *work = container_of(arg, struct btrfs_work,  \
  63                                                normal_work);            \
  64         normal_work_helper(work);                                       \
  65 }
  66
  67 bool btrfs_workqueue_normal_congested(struct btrfs_workqueue *wq)
  68 {
  69         /*
  70          * We could compare wq->normal->pending with num_online_cpus()
  71          * to support "thresh == NO_THRESHOLD" case, but it requires
  72          * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
  73          * postpone it until someone needs the support of that case.
  74          */
  75         if (wq->normal->thresh == NO_THRESHOLD)
  76                 return false;
  77
  78         return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
  79 }
  80
  81 BTRFS_WORK_HELPER(worker_helper);
  82 BTRFS_WORK_HELPER(delalloc_helper);
  83 BTRFS_WORK_HELPER(flush_delalloc_helper);
  84 BTRFS_WORK_HELPER(cache_helper);
  85 BTRFS_WORK_HELPER(submit_helper);
  86 BTRFS_WORK_HELPER(fixup_helper);
  87 BTRFS_WORK_HELPER(endio_helper);
  88 BTRFS_WORK_HELPER(endio_meta_helper);
  89 BTRFS_WORK_HELPER(endio_meta_write_helper);
  90 BTRFS_WORK_HELPER(endio_raid56_helper);
  91 BTRFS_WORK_HELPER(endio_repair_helper);
  92 BTRFS_WORK_HELPER(rmw_helper);
  93 BTRFS_WORK_HELPER(endio_write_helper);
  94 BTRFS_WORK_HELPER(freespace_write_helper);
  95 BTRFS_WORK_HELPER(delayed_meta_helper);
  96 BTRFS_WORK_HELPER(readahead_helper);
  97 BTRFS_WORK_HELPER(qgroup_rescan_helper);
  98 BTRFS_WORK_HELPER(extent_refs_helper);
  99 BTRFS_WORK_HELPER(scrub_helper);
 100 BTRFS_WORK_HELPER(scrubwrc_helper);
 101 BTRFS_WORK_HELPER(scrubnc_helper);
 102
 103 static struct __btrfs_workqueue *
 104 __btrfs_alloc_workqueue(const char *name, unsigned int flags, int max_active,
 105                          int thresh)
 106 {
 107         struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
 108
 109         if (!ret)
 110                 return NULL;
 111
 112         ret->max_active = max_active;
 113         atomic_set(&ret->pending, 0);
 114         if (thresh == 0)
 115                 thresh = DFT_THRESHOLD;
 116         /* For low threshold, disabling threshold is a better choice */
 117         if (thresh < DFT_THRESHOLD) {
 118                 ret->current_max = max_active;
 119                 ret->thresh = NO_THRESHOLD;
 120         } else {
 121                 ret->current_max = 1;
 122                 ret->thresh = thresh;
 123         }
 124
 125         if (flags & WQ_HIGHPRI)
 126                 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
 127                                                  ret->max_active,
 128                                                  "btrfs", name);
 129         else
 130                 ret->normal_wq = alloc_workqueue("%s-%s", flags,
 131                                                  ret->max_active, "btrfs",
 132                                                  name);
 133         if (!ret->normal_wq) {
 134                 kfree(ret);
 135                 return NULL;
 136         }
 137
 138         INIT_LIST_HEAD(&ret->ordered_list);
 139         spin_lock_init(&ret->list_lock);
 140         spin_lock_init(&ret->thres_lock);
 141         trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
 142         return ret;
 143 }
 144
 145 static inline void
 146 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
 147
 148 struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name,
 149                                               unsigned int flags,
 150                                               int max_active,
 151                                               int thresh)
 152 {
 153         struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS);
 154
 155         if (!ret)
 156                 return NULL;
 157
 158         ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI,
 159                                               max_active, thresh);
 160         if (!ret->normal) {
 161                 kfree(ret);
 162                 return NULL;
 163         }
 164
 165         if (flags & WQ_HIGHPRI) {
 166                 ret->high = __btrfs_alloc_workqueue(name, flags, max_active,
 167                                                     thresh);
 168                 if (!ret->high) {
 169                         __btrfs_destroy_workqueue(ret->normal);
 170                         kfree(ret);
 171                         return NULL;
 172                 }
 173         }
 174         return ret;
 175 }
 176
 177 /*
 178  * Hook for threshold which will be called in btrfs_queue_work.
 179  * This hook WILL be called in IRQ handler context,
 180  * so workqueue_set_max_active MUST NOT be called in this hook
 181  */
 182 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
 183 {
 184         if (wq->thresh == NO_THRESHOLD)
 185                 return;
 186         atomic_inc(&wq->pending);
 187 }
 188
 189 /*
 190  * Hook for threshold which will be called before executing the work,
 191  * This hook is called in kthread content.
 192  * So workqueue_set_max_active is called here.
 193  */
 194 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
 195 {
 196         int new_max_active;
 197         long pending;
 198         int need_change = 0;
 199
 200         if (wq->thresh == NO_THRESHOLD)
 201                 return;
 202
 203         atomic_dec(&wq->pending);
 204         spin_lock(&wq->thres_lock);
 205         /*
 206          * Use wq->count to limit the calling frequency of
 207          * workqueue_set_max_active.
 208          */
 209         wq->count++;
 210         wq->count %= (wq->thresh / 4);
 211         if (!wq->count)
 212                 goto  out;
 213         new_max_active = wq->current_max;
 214
 215         /*
 216          * pending may be changed later, but it's OK since we really
 217          * don't need it so accurate to calculate new_max_active.
 218          */
 219         pending = atomic_read(&wq->pending);
 220         if (pending > wq->thresh)
 221                 new_max_active++;
 222         if (pending < wq->thresh / 2)
 223                 new_max_active--;
 224         new_max_active = clamp_val(new_max_active, 1, wq->max_active);
 225         if (new_max_active != wq->current_max)  {
 226                 need_change = 1;
 227                 wq->current_max = new_max_active;
 228         }
 229 out:
 230         spin_unlock(&wq->thres_lock);
 231
 232         if (need_change) {
 233                 workqueue_set_max_active(wq->normal_wq, wq->current_max);
 234         }
 235 }
 236
 237 static void run_ordered_work(struct __btrfs_workqueue *wq)
 238 {
 239         struct list_head *list = &wq->ordered_list;
 240         struct btrfs_work *work;
 241         spinlock_t *lock = &wq->list_lock;
 242         unsigned long flags;
 243
 244         while (1) {
 245                 spin_lock_irqsave(lock, flags);
 246                 if (list_empty(list))
 247                         break;
 248                 work = list_entry(list->next, struct btrfs_work,
 249                                   ordered_list);
 250                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 251                         break;
 252
 253                 /*
 254                  * we are going to call the ordered done function, but
 255                  * we leave the work item on the list as a barrier so
 256                  * that later work items that are done don't have their
 257                  * functions called before this one returns
 258                  */
 259                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 260                         break;
 261                 trace_btrfs_ordered_sched(work);
 262                 spin_unlock_irqrestore(lock, flags);
 263                 work->ordered_func(work);
 264
 265                 /* now take the lock again and drop our item from the list */
 266                 spin_lock_irqsave(lock, flags);
 267                 list_del(&work->ordered_list);
 268                 spin_unlock_irqrestore(lock, flags);
 269
 270                 /*
 271                  * we don't want to call the ordered free functions
 272                  * with the lock held though
 273                  */
 274                 work->ordered_free(work);
 275                 trace_btrfs_all_work_done(work);
 276         }
 277         spin_unlock_irqrestore(lock, flags);
 278 }
 279
 280 static void normal_work_helper(struct btrfs_work *work)
 281 {
 282         struct __btrfs_workqueue *wq;
 283         int need_order = 0;
 284
 285         /*
 286          * We should not touch things inside work in the following cases:
 287          * 1) after work->func() if it has no ordered_free
 288          *    Since the struct is freed in work->func().
 289          * 2) after setting WORK_DONE_BIT
 290          *    The work may be freed in other threads almost instantly.
 291          * So we save the needed things here.
 292          */
 293         if (work->ordered_func)
 294                 need_order = 1;
 295         wq = work->wq;
 296
 297         trace_btrfs_work_sched(work);
 298         thresh_exec_hook(wq);
 299         work->func(work);
 300         if (need_order) {
 301                 set_bit(WORK_DONE_BIT, &work->flags);
 302                 run_ordered_work(wq);
 303         }
 304         if (!need_order)
 305                 trace_btrfs_all_work_done(work);
 306 }
 307
 308 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
 309                      btrfs_func_t func,
 310                      btrfs_func_t ordered_func,
 311                      btrfs_func_t ordered_free)
 312 {
 313         work->func = func;
 314         work->ordered_func = ordered_func;
 315         work->ordered_free = ordered_free;
 316         INIT_WORK(&work->normal_work, uniq_func);
 317         INIT_LIST_HEAD(&work->ordered_list);
 318         work->flags = 0;
 319 }
 320
 321 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
 322                                       struct btrfs_work *work)
 323 {
 324         unsigned long flags;
 325
 326         work->wq = wq;
 327         thresh_queue_hook(wq);
 328         if (work->ordered_func) {
 329                 spin_lock_irqsave(&wq->list_lock, flags);
 330                 list_add_tail(&work->ordered_list, &wq->ordered_list);
 331                 spin_unlock_irqrestore(&wq->list_lock, flags);
 332         }
 333         trace_btrfs_work_queued(work);
 334         queue_work(wq->normal_wq, &work->normal_work);
 335 }
 336
 337 void btrfs_queue_work(struct btrfs_workqueue *wq,
 338                       struct btrfs_work *work)
 339 {
 340         struct __btrfs_workqueue *dest_wq;
 341
 342         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
 343                 dest_wq = wq->high;
 344         else
 345                 dest_wq = wq->normal;
 346         __btrfs_queue_work(dest_wq, work);
 347 }
 348
 349 static inline void
 350 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
 351 {
 352         destroy_workqueue(wq->normal_wq);
 353         trace_btrfs_workqueue_destroy(wq);
 354         kfree(wq);
 355 }
 356
 357 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
 358 {
 359         if (!wq)
 360                 return;
 361         if (wq->high)
 362                 __btrfs_destroy_workqueue(wq->high);
 363         __btrfs_destroy_workqueue(wq->normal);
 364         kfree(wq);
 365 }
 366
 367 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int max)
 368 {
 369         if (!wq)
 370                 return;
 371         wq->normal->max_active = max;
 372         if (wq->high)
 373                 wq->high->max_active = max;
 374 }
 375
 376 void btrfs_set_work_high_priority(struct btrfs_work *work)
 377 {
 378         set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
 379 }