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 <trace/events/btrfs.h>
  13 #include "async-thread.h"
  14 #include "ctree.h"
  15
  16 enum {
  17         WORK_DONE_BIT,
  18         WORK_ORDER_DONE_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_fs_info * __pure btrfs_workqueue_owner(const struct btrfs_workqueue *wq)
  52 {
  53         return wq->fs_info;
  54 }
  55
  56 struct btrfs_fs_info * __pure btrfs_work_owner(const struct btrfs_work *work)
  57 {
  58         return work->wq->fs_info;
  59 }
  60
  61 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
  62 {
  63         /*
  64          * We could compare wq->pending with num_online_cpus()
  65          * to support "thresh == NO_THRESHOLD" case, but it requires
  66          * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
  67          * postpone it until someone needs the support of that case.
  68          */
  69         if (wq->thresh == NO_THRESHOLD)
  70                 return false;
  71
  72         return atomic_read(&wq->pending) > wq->thresh * 2;
  73 }
  74
  75 static void btrfs_init_workqueue(struct btrfs_workqueue *wq,
  76                                  struct btrfs_fs_info *fs_info)
  77 {
  78         wq->fs_info = fs_info;
  79         atomic_set(&wq->pending, 0);
  80         INIT_LIST_HEAD(&wq->ordered_list);
  81         spin_lock_init(&wq->list_lock);
  82         spin_lock_init(&wq->thres_lock);
  83 }
  84
  85 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
  86                                               const char *name, unsigned int flags,
  87                                               int limit_active, int thresh)
  88 {
  89         struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
  90
  91         if (!ret)
  92                 return NULL;
  93
  94         btrfs_init_workqueue(ret, fs_info);
  95
  96         ret->limit_active = limit_active;
  97         if (thresh == 0)
  98                 thresh = DFT_THRESHOLD;
  99         /* For low threshold, disabling threshold is a better choice */
 100         if (thresh < DFT_THRESHOLD) {
 101                 ret->current_active = limit_active;
 102                 ret->thresh = NO_THRESHOLD;
 103         } else {
 104                 /*
 105                  * For threshold-able wq, let its concurrency grow on demand.
 106                  * Use minimal max_active at alloc time to reduce resource
 107                  * usage.
 108                  */
 109                 ret->current_active = 1;
 110                 ret->thresh = thresh;
 111         }
 112
 113         ret->normal_wq = alloc_workqueue("btrfs-%s", flags, ret->current_active,
 114                                          name);
 115         if (!ret->normal_wq) {
 116                 kfree(ret);
 117                 return NULL;
 118         }
 119
 120         trace_btrfs_workqueue_alloc(ret, name);
 121         return ret;
 122 }
 123
 124 struct btrfs_workqueue *btrfs_alloc_ordered_workqueue(
 125                                 struct btrfs_fs_info *fs_info, const char *name,
 126                                 unsigned int flags)
 127 {
 128         struct btrfs_workqueue *ret;
 129
 130         ret = kzalloc(sizeof(*ret), GFP_KERNEL);
 131         if (!ret)
 132                 return NULL;
 133
 134         btrfs_init_workqueue(ret, fs_info);
 135
 136         /* Ordered workqueues don't allow @max_active adjustments. */
 137         ret->limit_active = 1;
 138         ret->current_active = 1;
 139         ret->thresh = NO_THRESHOLD;
 140
 141         ret->normal_wq = alloc_ordered_workqueue("btrfs-%s", flags, name);
 142         if (!ret->normal_wq) {
 143                 kfree(ret);
 144                 return NULL;
 145         }
 146
 147         trace_btrfs_workqueue_alloc(ret, name);
 148         return ret;
 149 }
 150
 151 /*
 152  * Hook for threshold which will be called in btrfs_queue_work.
 153  * This hook WILL be called in IRQ handler context,
 154  * so workqueue_set_max_active MUST NOT be called in this hook
 155  */
 156 static inline void thresh_queue_hook(struct btrfs_workqueue *wq)
 157 {
 158         if (wq->thresh == NO_THRESHOLD)
 159                 return;
 160         atomic_inc(&wq->pending);
 161 }
 162
 163 /*
 164  * Hook for threshold which will be called before executing the work,
 165  * This hook is called in kthread content.
 166  * So workqueue_set_max_active is called here.
 167  */
 168 static inline void thresh_exec_hook(struct btrfs_workqueue *wq)
 169 {
 170         int new_current_active;
 171         long pending;
 172         int need_change = 0;
 173
 174         if (wq->thresh == NO_THRESHOLD)
 175                 return;
 176
 177         atomic_dec(&wq->pending);
 178         spin_lock(&wq->thres_lock);
 179         /*
 180          * Use wq->count to limit the calling frequency of
 181          * workqueue_set_max_active.
 182          */
 183         wq->count++;
 184         wq->count %= (wq->thresh / 4);
 185         if (!wq->count)
 186                 goto  out;
 187         new_current_active = wq->current_active;
 188
 189         /*
 190          * pending may be changed later, but it's OK since we really
 191          * don't need it so accurate to calculate new_max_active.
 192          */
 193         pending = atomic_read(&wq->pending);
 194         if (pending > wq->thresh)
 195                 new_current_active++;
 196         if (pending < wq->thresh / 2)
 197                 new_current_active--;
 198         new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
 199         if (new_current_active != wq->current_active)  {
 200                 need_change = 1;
 201                 wq->current_active = new_current_active;
 202         }
 203 out:
 204         spin_unlock(&wq->thres_lock);
 205
 206         if (need_change) {
 207                 workqueue_set_max_active(wq->normal_wq, wq->current_active);
 208         }
 209 }
 210
 211 static void run_ordered_work(struct btrfs_workqueue *wq,
 212                              struct btrfs_work *self)
 213 {
 214         struct list_head *list = &wq->ordered_list;
 215         struct btrfs_work *work;
 216         spinlock_t *lock = &wq->list_lock;
 217         unsigned long flags;
 218         bool free_self = false;
 219
 220         while (1) {
 221                 spin_lock_irqsave(lock, flags);
 222                 if (list_empty(list))
 223                         break;
 224                 work = list_entry(list->next, struct btrfs_work,
 225                                   ordered_list);
 226                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 227                         break;
 228                 /*
 229                  * Orders all subsequent loads after reading WORK_DONE_BIT,
 230                  * paired with the smp_mb__before_atomic in btrfs_work_helper
 231                  * this guarantees that the ordered function will see all
 232                  * updates from ordinary work function.
 233                  */
 234                 smp_rmb();
 235
 236                 /*
 237                  * we are going to call the ordered done function, but
 238                  * we leave the work item on the list as a barrier so
 239                  * that later work items that are done don't have their
 240                  * functions called before this one returns
 241                  */
 242                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 243                         break;
 244                 trace_btrfs_ordered_sched(work);
 245                 spin_unlock_irqrestore(lock, flags);
 246                 work->ordered_func(work, false);
 247
 248                 /* now take the lock again and drop our item from the list */
 249                 spin_lock_irqsave(lock, flags);
 250                 list_del(&work->ordered_list);
 251                 spin_unlock_irqrestore(lock, flags);
 252
 253                 if (work == self) {
 254                         /*
 255                          * This is the work item that the worker is currently
 256                          * executing.
 257                          *
 258                          * The kernel workqueue code guarantees non-reentrancy
 259                          * of work items. I.e., if a work item with the same
 260                          * address and work function is queued twice, the second
 261                          * execution is blocked until the first one finishes. A
 262                          * work item may be freed and recycled with the same
 263                          * work function; the workqueue code assumes that the
 264                          * original work item cannot depend on the recycled work
 265                          * item in that case (see find_worker_executing_work()).
 266                          *
 267                          * Note that different types of Btrfs work can depend on
 268                          * each other, and one type of work on one Btrfs
 269                          * filesystem may even depend on the same type of work
 270                          * on another Btrfs filesystem via, e.g., a loop device.
 271                          * Therefore, we must not allow the current work item to
 272                          * be recycled until we are really done, otherwise we
 273                          * break the above assumption and can deadlock.
 274                          */
 275                         free_self = true;
 276                 } else {
 277                         /*
 278                          * We don't want to call the ordered free functions with
 279                          * the lock held.
 280                          */
 281                         work->ordered_func(work, true);
 282                         /* NB: work must not be dereferenced past this point. */
 283                         trace_btrfs_all_work_done(wq->fs_info, work);
 284                 }
 285         }
 286         spin_unlock_irqrestore(lock, flags);
 287
 288         if (free_self) {
 289                 self->ordered_func(self, true);
 290                 /* NB: self must not be dereferenced past this point. */
 291                 trace_btrfs_all_work_done(wq->fs_info, self);
 292         }
 293 }
 294
 295 static void btrfs_work_helper(struct work_struct *normal_work)
 296 {
 297         struct btrfs_work *work = container_of(normal_work, struct btrfs_work,
 298                                                normal_work);
 299         struct btrfs_workqueue *wq = work->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_func(..., true) to 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
 313         trace_btrfs_work_sched(work);
 314         thresh_exec_hook(wq);
 315         work->func(work);
 316         if (need_order) {
 317                 /*
 318                  * Ensures all memory accesses done in the work function are
 319                  * ordered before setting the WORK_DONE_BIT. Ensuring the thread
 320                  * which is going to executed the ordered work sees them.
 321                  * Pairs with the smp_rmb in run_ordered_work.
 322                  */
 323                 smp_mb__before_atomic();
 324                 set_bit(WORK_DONE_BIT, &work->flags);
 325                 run_ordered_work(wq, work);
 326         } else {
 327                 /* NB: work must not be dereferenced past this point. */
 328                 trace_btrfs_all_work_done(wq->fs_info, work);
 329         }
 330 }
 331
 332 void btrfs_init_work(struct btrfs_work *work, btrfs_func_t func,
 333                      btrfs_ordered_func_t ordered_func)
 334 {
 335         work->func = func;
 336         work->ordered_func = ordered_func;
 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 }