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