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 bool btrfs_workqueue_normal_congested(struct btrfs_workqueue *wq)
  90 {
  91         /*
  92          * We could compare wq->normal->pending with num_online_cpus()
  93          * to support "thresh == NO_THRESHOLD" case, but it requires
  94          * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
  95          * postpone it until someone needs the support of that case.
  96          */
  97         if (wq->normal->thresh == NO_THRESHOLD)
  98                 return false;
  99
 100         return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
 101 }
 102
 103 BTRFS_WORK_HELPER(worker_helper);
 104 BTRFS_WORK_HELPER(delalloc_helper);
 105 BTRFS_WORK_HELPER(flush_delalloc_helper);
 106 BTRFS_WORK_HELPER(cache_helper);
 107 BTRFS_WORK_HELPER(submit_helper);
 108 BTRFS_WORK_HELPER(fixup_helper);
 109 BTRFS_WORK_HELPER(endio_helper);
 110 BTRFS_WORK_HELPER(endio_meta_helper);
 111 BTRFS_WORK_HELPER(endio_meta_write_helper);
 112 BTRFS_WORK_HELPER(endio_raid56_helper);
 113 BTRFS_WORK_HELPER(endio_repair_helper);
 114 BTRFS_WORK_HELPER(rmw_helper);
 115 BTRFS_WORK_HELPER(endio_write_helper);
 116 BTRFS_WORK_HELPER(freespace_write_helper);
 117 BTRFS_WORK_HELPER(delayed_meta_helper);
 118 BTRFS_WORK_HELPER(readahead_helper);
 119 BTRFS_WORK_HELPER(qgroup_rescan_helper);
 120 BTRFS_WORK_HELPER(extent_refs_helper);
 121 BTRFS_WORK_HELPER(scrub_helper);
 122 BTRFS_WORK_HELPER(scrubwrc_helper);
 123 BTRFS_WORK_HELPER(scrubnc_helper);
 124 BTRFS_WORK_HELPER(scrubparity_helper);
 125
 126 static struct __btrfs_workqueue *
 127 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
 128                         unsigned int flags, int limit_active, int thresh)
 129 {
 130         struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
 131
 132         if (!ret)
 133                 return NULL;
 134
 135         ret->fs_info = fs_info;
 136         ret->limit_active = limit_active;
 137         atomic_set(&ret->pending, 0);
 138         if (thresh == 0)
 139                 thresh = DFT_THRESHOLD;
 140         /* For low threshold, disabling threshold is a better choice */
 141         if (thresh < DFT_THRESHOLD) {
 142                 ret->current_active = limit_active;
 143                 ret->thresh = NO_THRESHOLD;
 144         } else {
 145                 /*
 146                  * For threshold-able wq, let its concurrency grow on demand.
 147                  * Use minimal max_active at alloc time to reduce resource
 148                  * usage.
 149                  */
 150                 ret->current_active = 1;
 151                 ret->thresh = thresh;
 152         }
 153
 154         if (flags & WQ_HIGHPRI)
 155                 ret->normal_wq = alloc_workqueue("%s-%s-high", flags,
 156                                                  ret->current_active, "btrfs",
 157                                                  name);
 158         else
 159                 ret->normal_wq = alloc_workqueue("%s-%s", flags,
 160                                                  ret->current_active, "btrfs",
 161                                                  name);
 162         if (!ret->normal_wq) {
 163                 kfree(ret);
 164                 return NULL;
 165         }
 166
 167         INIT_LIST_HEAD(&ret->ordered_list);
 168         spin_lock_init(&ret->list_lock);
 169         spin_lock_init(&ret->thres_lock);
 170         trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
 171         return ret;
 172 }
 173
 174 static inline void
 175 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
 176
 177 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
 178                                               const char *name,
 179                                               unsigned int flags,
 180                                               int limit_active,
 181                                               int thresh)
 182 {
 183         struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
 184
 185         if (!ret)
 186                 return NULL;
 187
 188         ret->normal = __btrfs_alloc_workqueue(fs_info, name,
 189                                               flags & ~WQ_HIGHPRI,
 190                                               limit_active, thresh);
 191         if (!ret->normal) {
 192                 kfree(ret);
 193                 return NULL;
 194         }
 195
 196         if (flags & WQ_HIGHPRI) {
 197                 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
 198                                                     limit_active, thresh);
 199                 if (!ret->high) {
 200                         __btrfs_destroy_workqueue(ret->normal);
 201                         kfree(ret);
 202                         return NULL;
 203                 }
 204         }
 205         return ret;
 206 }
 207
 208 /*
 209  * Hook for threshold which will be called in btrfs_queue_work.
 210  * This hook WILL be called in IRQ handler context,
 211  * so workqueue_set_max_active MUST NOT be called in this hook
 212  */
 213 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
 214 {
 215         if (wq->thresh == NO_THRESHOLD)
 216                 return;
 217         atomic_inc(&wq->pending);
 218 }
 219
 220 /*
 221  * Hook for threshold which will be called before executing the work,
 222  * This hook is called in kthread content.
 223  * So workqueue_set_max_active is called here.
 224  */
 225 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
 226 {
 227         int new_current_active;
 228         long pending;
 229         int need_change = 0;
 230
 231         if (wq->thresh == NO_THRESHOLD)
 232                 return;
 233
 234         atomic_dec(&wq->pending);
 235         spin_lock(&wq->thres_lock);
 236         /*
 237          * Use wq->count to limit the calling frequency of
 238          * workqueue_set_max_active.
 239          */
 240         wq->count++;
 241         wq->count %= (wq->thresh / 4);
 242         if (!wq->count)
 243                 goto  out;
 244         new_current_active = wq->current_active;
 245
 246         /*
 247          * pending may be changed later, but it's OK since we really
 248          * don't need it so accurate to calculate new_max_active.
 249          */
 250         pending = atomic_read(&wq->pending);
 251         if (pending > wq->thresh)
 252                 new_current_active++;
 253         if (pending < wq->thresh / 2)
 254                 new_current_active--;
 255         new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
 256         if (new_current_active != wq->current_active)  {
 257                 need_change = 1;
 258                 wq->current_active = new_current_active;
 259         }
 260 out:
 261         spin_unlock(&wq->thres_lock);
 262
 263         if (need_change) {
 264                 workqueue_set_max_active(wq->normal_wq, wq->current_active);
 265         }
 266 }
 267
 268 static void run_ordered_work(struct __btrfs_workqueue *wq,
 269                              struct btrfs_work *self)
 270 {
 271         struct list_head *list = &wq->ordered_list;
 272         struct btrfs_work *work;
 273         spinlock_t *lock = &wq->list_lock;
 274         unsigned long flags;
 275         void *wtag;
 276         bool free_self = false;
 277
 278         while (1) {
 279                 spin_lock_irqsave(lock, flags);
 280                 if (list_empty(list))
 281                         break;
 282                 work = list_entry(list->next, struct btrfs_work,
 283                                   ordered_list);
 284                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 285                         break;
 286                 /*
 287                  * Orders all subsequent loads after reading WORK_DONE_BIT,
 288                  * paired with the smp_mb__before_atomic in btrfs_work_helper
 289                  * this guarantees that the ordered function will see all
 290                  * updates from ordinary work function.
 291                  */
 292                 smp_rmb();
 293
 294                 /*
 295                  * we are going to call the ordered done function, but
 296                  * we leave the work item on the list as a barrier so
 297                  * that later work items that are done don't have their
 298                  * functions called before this one returns
 299                  */
 300                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 301                         break;
 302                 trace_btrfs_ordered_sched(work);
 303                 spin_unlock_irqrestore(lock, flags);
 304                 work->ordered_func(work);
 305
 306                 /* now take the lock again and drop our item from the list */
 307                 spin_lock_irqsave(lock, flags);
 308                 list_del(&work->ordered_list);
 309                 spin_unlock_irqrestore(lock, flags);
 310
 311                 if (work == self) {
 312                         /*
 313                          * This is the work item that the worker is currently
 314                          * executing.
 315                          *
 316                          * The kernel workqueue code guarantees non-reentrancy
 317                          * of work items. I.e., if a work item with the same
 318                          * address and work function is queued twice, the second
 319                          * execution is blocked until the first one finishes. A
 320                          * work item may be freed and recycled with the same
 321                          * work function; the workqueue code assumes that the
 322                          * original work item cannot depend on the recycled work
 323                          * item in that case (see find_worker_executing_work()).
 324                          *
 325                          * Note that the work of one Btrfs filesystem may depend
 326                          * on the work of another Btrfs filesystem via, e.g., a
 327                          * loop device. Therefore, we must not allow the current
 328                          * work item to be recycled until we are really done,
 329                          * otherwise we break the above assumption and can
 330                          * deadlock.
 331                          */
 332                         free_self = true;
 333                 } else {
 334                         /*
 335                          * We don't want to call the ordered free functions with
 336                          * the lock held though. Save the work as tag for the
 337                          * trace event, because the callback could free the
 338                          * structure.
 339                          */
 340                         wtag = work;
 341                         work->ordered_free(work);
 342                         trace_btrfs_all_work_done(wq->fs_info, wtag);
 343                 }
 344         }
 345         spin_unlock_irqrestore(lock, flags);
 346
 347         if (free_self) {
 348                 wtag = self;
 349                 self->ordered_free(self);
 350                 trace_btrfs_all_work_done(wq->fs_info, wtag);
 351         }
 352 }
 353
 354 static void normal_work_helper(struct btrfs_work *work)
 355 {
 356         struct __btrfs_workqueue *wq;
 357         void *wtag;
 358         int need_order = 0;
 359
 360         /*
 361          * We should not touch things inside work in the following cases:
 362          * 1) after work->func() if it has no ordered_free
 363          *    Since the struct is freed in work->func().
 364          * 2) after setting WORK_DONE_BIT
 365          *    The work may be freed in other threads almost instantly.
 366          * So we save the needed things here.
 367          */
 368         if (work->ordered_func)
 369                 need_order = 1;
 370         wq = work->wq;
 371         /* Safe for tracepoints in case work gets freed by the callback */
 372         wtag = work;
 373
 374         trace_btrfs_work_sched(work);
 375         thresh_exec_hook(wq);
 376         work->func(work);
 377         if (need_order) {
 378                 /*
 379                  * Ensures all memory accesses done in the work function are
 380                  * ordered before setting the WORK_DONE_BIT. Ensuring the thread
 381                  * which is going to executed the ordered work sees them.
 382                  * Pairs with the smp_rmb in run_ordered_work.
 383                  */
 384                 smp_mb__before_atomic();
 385                 set_bit(WORK_DONE_BIT, &work->flags);
 386                 run_ordered_work(wq, work);
 387         }
 388         if (!need_order)
 389                 trace_btrfs_all_work_done(wq->fs_info, wtag);
 390 }
 391
 392 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
 393                      btrfs_func_t func,
 394                      btrfs_func_t ordered_func,
 395                      btrfs_func_t ordered_free)
 396 {
 397         work->func = func;
 398         work->ordered_func = ordered_func;
 399         work->ordered_free = ordered_free;
 400         INIT_WORK(&work->normal_work, uniq_func);
 401         INIT_LIST_HEAD(&work->ordered_list);
 402         work->flags = 0;
 403 }
 404
 405 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
 406                                       struct btrfs_work *work)
 407 {
 408         unsigned long flags;
 409
 410         work->wq = wq;
 411         thresh_queue_hook(wq);
 412         if (work->ordered_func) {
 413                 spin_lock_irqsave(&wq->list_lock, flags);
 414                 list_add_tail(&work->ordered_list, &wq->ordered_list);
 415                 spin_unlock_irqrestore(&wq->list_lock, flags);
 416         }
 417         trace_btrfs_work_queued(work);
 418         queue_work(wq->normal_wq, &work->normal_work);
 419 }
 420
 421 void btrfs_queue_work(struct btrfs_workqueue *wq,
 422                       struct btrfs_work *work)
 423 {
 424         struct __btrfs_workqueue *dest_wq;
 425
 426         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
 427                 dest_wq = wq->high;
 428         else
 429                 dest_wq = wq->normal;
 430         __btrfs_queue_work(dest_wq, work);
 431 }
 432
 433 static inline void
 434 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
 435 {
 436         destroy_workqueue(wq->normal_wq);
 437         trace_btrfs_workqueue_destroy(wq);
 438         kfree(wq);
 439 }
 440
 441 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
 442 {
 443         if (!wq)
 444                 return;
 445         if (wq->high)
 446                 __btrfs_destroy_workqueue(wq->high);
 447         __btrfs_destroy_workqueue(wq->normal);
 448         kfree(wq);
 449 }
 450
 451 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
 452 {
 453         if (!wq)
 454                 return;
 455         wq->normal->limit_active = limit_active;
 456         if (wq->high)
 457                 wq->high->limit_active = limit_active;
 458 }
 459
 460 void btrfs_set_work_high_priority(struct btrfs_work *work)
 461 {
 462         set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
 463 }
 464
 465 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
 466 {
 467         if (wq->high)
 468                 flush_workqueue(wq->high->normal_wq);
 469
 470         flush_workqueue(wq->normal->normal_wq);
 471 }