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