fs/btrfs/async-thread.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  * Copyright (C) 2014 Fujitsu.  All rights reserved.
   5  */
   6
   7 #include <linux/kthread.h>
   8 #include <linux/slab.h>
   9 #include <linux/list.h>
  10 #include <linux/spinlock.h>
  11 #include <linux/freezer.h>
  12 #include "async-thread.h"
  13 #include "ctree.h"
  14
  15 enum {
  16         WORK_DONE_BIT,
  17         WORK_ORDER_DONE_BIT,
  18         WORK_HIGH_PRIO_BIT,
  19 };
  20
  21 #define NO_THRESHOLD (-1)
  22 #define DFT_THRESHOLD (32)
  23
  24 struct __btrfs_workqueue {
  25         struct workqueue_struct *normal_wq;
  26
  27         /* File system this workqueue services */
  28         struct btrfs_fs_info *fs_info;
  29
  30         /* List head pointing to ordered work list */
  31         struct list_head ordered_list;
  32
  33         /* Spinlock for ordered_list */
  34         spinlock_t list_lock;
  35
  36         /* Thresholding related variants */
  37         atomic_t pending;
  38
  39         /* Up limit of concurrency workers */
  40         int limit_active;
  41
  42         /* Current number of concurrency workers */
  43         int current_active;
  44
  45         /* Threshold to change current_active */
  46         int thresh;
  47         unsigned int count;
  48         spinlock_t thres_lock;
  49 };
  50
  51 struct btrfs_workqueue {
  52         struct __btrfs_workqueue *normal;
  53         struct __btrfs_workqueue *high;
  54 };
  55
  56 struct btrfs_fs_info * __pure btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
  57 {
  58         return wq->fs_info;
  59 }
  60
  61 struct btrfs_fs_info * __pure btrfs_work_owner(const struct btrfs_work *work)
  62 {
  63         return work->wq->fs_info;
  64 }
  65
  66 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
  67 {
  68         /*
  69          * We could compare wq->normal->pending with num_online_cpus()
  70          * to support "thresh == NO_THRESHOLD" case, but it requires
  71          * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
  72          * postpone it until someone needs the support of that case.
  73          */
  74         if (wq->normal->thresh == NO_THRESHOLD)
  75                 return false;
  76
  77         return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
  78 }
  79
  80 static struct __btrfs_workqueue *
  81 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
  82                         unsigned int flags, int limit_active, int thresh)
  83 {
  84         struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
  85
  86         if (!ret)
  87                 return NULL;
  88
  89         ret->fs_info = fs_info;
  90         ret->limit_active = limit_active;
  91         atomic_set(&ret->pending, 0);
  92         if (thresh == 0)
  93                 thresh = DFT_THRESHOLD;
  94         /* For low threshold, disabling threshold is a better choice */
  95         if (thresh < DFT_THRESHOLD) {
  96                 ret->current_active = limit_active;
  97                 ret->thresh = NO_THRESHOLD;
  98         } else {
  99                 /*
 100                  * For threshold-able wq, let its concurrency grow on demand.
 101                  * Use minimal max_active at alloc time to reduce resource
 102                  * usage.
 103                  */
 104                 ret->current_active = 1;
 105                 ret->thresh = thresh;
 106         }
 107
 108         if (flags & WQ_HIGHPRI)
 109                 ret->normal_wq = alloc_workqueue("btrfs-%s-high", flags,
 110                                                  ret->current_active, name);
 111         else
 112                 ret->normal_wq = alloc_workqueue("btrfs-%s", flags,
 113                                                  ret->current_active, name);
 114         if (!ret->normal_wq) {
 115                 kfree(ret);
 116                 return NULL;
 117         }
 118
 119         INIT_LIST_HEAD(&ret->ordered_list);
 120         spin_lock_init(&ret->list_lock);
 121         spin_lock_init(&ret->thres_lock);
 122         trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
 123         return ret;
 124 }
 125
 126 static inline void
 127 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
 128
 129 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
 130                                               const char *name,
 131                                               unsigned int flags,
 132                                               int limit_active,
 133                                               int thresh)
 134 {
 135         struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
 136
 137         if (!ret)
 138                 return NULL;
 139
 140         ret->normal = __btrfs_alloc_workqueue(fs_info, name,
 141                                               flags & ~WQ_HIGHPRI,
 142                                               limit_active, thresh);
 143         if (!ret->normal) {
 144                 kfree(ret);
 145                 return NULL;
 146         }
 147
 148         if (flags & WQ_HIGHPRI) {
 149                 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
 150                                                     limit_active, thresh);
 151                 if (!ret->high) {
 152                         __btrfs_destroy_workqueue(ret->normal);
 153                         kfree(ret);
 154                         return NULL;
 155                 }
 156         }
 157         return ret;
 158 }
 159
 160 /*
 161  * Hook for threshold which will be called in btrfs_queue_work.
 162  * This hook WILL be called in IRQ handler context,
 163  * so workqueue_set_max_active MUST NOT be called in this hook
 164  */
 165 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
 166 {
 167         if (wq->thresh == NO_THRESHOLD)
 168                 return;
 169         atomic_inc(&wq->pending);
 170 }
 171
 172 /*
 173  * Hook for threshold which will be called before executing the work,
 174  * This hook is called in kthread content.
 175  * So workqueue_set_max_active is called here.
 176  */
 177 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
 178 {
 179         int new_current_active;
 180         long pending;
 181         int need_change = 0;
 182
 183         if (wq->thresh == NO_THRESHOLD)
 184                 return;
 185
 186         atomic_dec(&wq->pending);
 187         spin_lock(&wq->thres_lock);
 188         /*
 189          * Use wq->count to limit the calling frequency of
 190          * workqueue_set_max_active.
 191          */
 192         wq->count++;
 193         wq->count %= (wq->thresh / 4);
 194         if (!wq->count)
 195                 goto  out;
 196         new_current_active = wq->current_active;
 197
 198         /*
 199          * pending may be changed later, but it's OK since we really
 200          * don't need it so accurate to calculate new_max_active.
 201          */
 202         pending = atomic_read(&wq->pending);
 203         if (pending > wq->thresh)
 204                 new_current_active++;
 205         if (pending < wq->thresh / 2)
 206                 new_current_active--;
 207         new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
 208         if (new_current_active != wq->current_active)  {
 209                 need_change = 1;
 210                 wq->current_active = new_current_active;
 211         }
 212 out:
 213         spin_unlock(&wq->thres_lock);
 214
 215         if (need_change) {
 216                 workqueue_set_max_active(wq->normal_wq, wq->current_active);
 217         }
 218 }
 219
 220 static void run_ordered_work(struct __btrfs_workqueue *wq,
 221                              struct btrfs_work *self)
 222 {
 223         struct list_head *list = &wq->ordered_list;
 224         struct btrfs_work *work;
 225         spinlock_t *lock = &wq->list_lock;
 226         unsigned long flags;
 227         bool free_self = false;
 228
 229         while (1) {
 230                 spin_lock_irqsave(lock, flags);
 231                 if (list_empty(list))
 232                         break;
 233                 work = list_entry(list->next, struct btrfs_work,
 234                                   ordered_list);
 235                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 236                         break;
 237                 /*
 238                  * Orders all subsequent loads after reading WORK_DONE_BIT,
 239                  * paired with the smp_mb__before_atomic in btrfs_work_helper
 240                  * this guarantees that the ordered function will see all
 241                  * updates from ordinary work function.
 242                  */
 243                 smp_rmb();
 244
 245                 /*
 246                  * we are going to call the ordered done function, but
 247                  * we leave the work item on the list as a barrier so
 248                  * that later work items that are done don't have their
 249                  * functions called before this one returns
 250                  */
 251                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 252                         break;
 253                 trace_btrfs_ordered_sched(work);
 254                 spin_unlock_irqrestore(lock, flags);
 255                 work->ordered_func(work);
 256
 257                 /* now take the lock again and drop our item from the list */
 258                 spin_lock_irqsave(lock, flags);
 259                 list_del(&work->ordered_list);
 260                 spin_unlock_irqrestore(lock, flags);
 261
 262                 if (work == self) {
 263                         /*
 264                          * This is the work item that the worker is currently
 265                          * executing.
 266                          *
 267                          * The kernel workqueue code guarantees non-reentrancy
 268                          * of work items. I.e., if a work item with the same
 269                          * address and work function is queued twice, the second
 270                          * execution is blocked until the first one finishes. A
 271                          * work item may be freed and recycled with the same
 272                          * work function; the workqueue code assumes that the
 273                          * original work item cannot depend on the recycled work
 274                          * item in that case (see find_worker_executing_work()).
 275                          *
 276                          * Note that different types of Btrfs work can depend on
 277                          * each other, and one type of work on one Btrfs
 278                          * filesystem may even depend on the same type of work
 279                          * on another Btrfs filesystem via, e.g., a loop device.
 280                          * Therefore, we must not allow the current work item to
 281                          * be recycled until we are really done, otherwise we
 282                          * break the above assumption and can deadlock.
 283                          */
 284                         free_self = true;
 285                 } else {
 286                         /*
 287                          * We don't want to call the ordered free functions with
 288                          * the lock held.
 289                          */
 290                         work->ordered_free(work);
 291                         /* NB: work must not be dereferenced past this point. */
 292                         trace_btrfs_all_work_done(wq->fs_info, work);
 293                 }
 294         }
 295         spin_unlock_irqrestore(lock, flags);
 296
 297         if (free_self) {
 298                 self->ordered_free(self);
 299                 /* NB: self must not be dereferenced past this point. */
 300                 trace_btrfs_all_work_done(wq->fs_info, self);
 301         }
 302 }
 303
 304 static void btrfs_work_helper(struct work_struct *normal_work)
 305 {
 306         struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
 307                                                normal_work);
 308         struct __btrfs_workqueue *wq;
 309         int need_order = 0;
 310
 311         /*
 312          * We should not touch things inside work in the following cases:
 313          * 1) after work->func() if it has no ordered_free
 314          *    Since the struct is freed in work->func().
 315          * 2) after setting WORK_DONE_BIT
 316          *    The work may be freed in other threads almost instantly.
 317          * So we save the needed things here.
 318          */
 319         if (work->ordered_func)
 320                 need_order = 1;
 321         wq = work->wq;
 322
 323         trace_btrfs_work_sched(work);
 324         thresh_exec_hook(wq);
 325         work->func(work);
 326         if (need_order) {
 327                 /*
 328                  * Ensures all memory accesses done in the work function are
 329                  * ordered before setting the WORK_DONE_BIT. Ensuring the thread
 330                  * which is going to executed the ordered work sees them.
 331                  * Pairs with the smp_rmb in run_ordered_work.
 332                  */
 333                 smp_mb__before_atomic();
 334                 set_bit(WORK_DONE_BIT, &work->flags);
 335                 run_ordered_work(wq, work);
 336         } else {
 337                 /* NB: work must not be dereferenced past this point. */
 338                 trace_btrfs_all_work_done(wq->fs_info, work);
 339         }
 340 }
 341
 342 void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
 343                      btrfs_func_t ordered_func, btrfs_func_t ordered_free)
 344 {
 345         work->func = func;
 346         work->ordered_func = ordered_func;
 347         work->ordered_free = ordered_free;
 348         INIT_WORK(&work->normal_work, btrfs_work_helper);
 349         INIT_LIST_HEAD(&work->ordered_list);
 350         work->flags = 0;
 351 }
 352
 353 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
 354                                       struct btrfs_work *work)
 355 {
 356         unsigned long flags;
 357
 358         work->wq = wq;
 359         thresh_queue_hook(wq);
 360         if (work->ordered_func) {
 361                 spin_lock_irqsave(&wq->list_lock, flags);
 362                 list_add_tail(&work->ordered_list, &wq->ordered_list);
 363                 spin_unlock_irqrestore(&wq->list_lock, flags);
 364         }
 365         trace_btrfs_work_queued(work);
 366         queue_work(wq->normal_wq, &work->normal_work);
 367 }
 368
 369 void btrfs_queue_work(struct btrfs_workqueue *wq,
 370                       struct btrfs_work *work)
 371 {
 372         struct __btrfs_workqueue *dest_wq;
 373
 374         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
 375                 dest_wq = wq->high;
 376         else
 377                 dest_wq = wq->normal;
 378         __btrfs_queue_work(dest_wq, work);
 379 }
 380
 381 static inline void
 382 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
 383 {
 384         destroy_workqueue(wq->normal_wq);
 385         trace_btrfs_workqueue_destroy(wq);
 386         kfree(wq);
 387 }
 388
 389 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
 390 {
 391         if (!wq)
 392                 return;
 393         if (wq->high)
 394                 __btrfs_destroy_workqueue(wq->high);
 395         __btrfs_destroy_workqueue(wq->normal);
 396         kfree(wq);
 397 }
 398
 399 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
 400 {
 401         if (!wq)
 402                 return;
 403         wq->normal->limit_active = limit_active;
 404         if (wq->high)
 405                 wq->high->limit_active = limit_active;
 406 }
 407
 408 void btrfs_set_work_high_priority(struct btrfs_work *work)
 409 {
 410         set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
 411 }
 412
 413 void btrfs_flush_workqueue(struct btrfs_workqueue *wq)
 414 {
 415         if (wq->high)
 416                 flush_workqueue(wq->high->normal_wq);
 417
 418         flush_workqueue(wq->normal->normal_wq);
 419 }