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