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