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 #define WORK_DONE_BIT 0
  16 #define WORK_ORDER_DONE_BIT 1
  17 #define WORK_HIGH_PRIO_BIT 2
  18
  19 #define NO_THRESHOLD (-1)
  20 #define DFT_THRESHOLD (32)
  21
  22 struct __btrfs_workqueue {
  23         struct workqueue_struct *normal_wq;
  24
  25         /* File system this workqueue services */
  26         struct btrfs_fs_info *fs_info;
  27
  28         /* List head pointing to ordered work list */
  29         struct list_head ordered_list;
  30
  31         /* Spinlock for ordered_list */
  32         spinlock_t list_lock;
  33
  34         /* Thresholding related variants */
  35         atomic_t pending;
  36
  37         /* Up limit of concurrency workers */
  38         int limit_active;
  39
  40         /* Current number of concurrency workers */
  41         int current_active;
  42
  43         /* Threshold to change current_active */
  44         int thresh;
  45         unsigned int count;
  46         spinlock_t thres_lock;
  47 };
  48
  49 struct btrfs_workqueue {
  50         struct __btrfs_workqueue *normal;
  51         struct __btrfs_workqueue *high;
  52 };
  53
  54 static void normal_work_helper(struct btrfs_work *work);
  55
  56 #define BTRFS_WORK_HELPER(name)                                 \
  57 noinline_for_stack void btrfs_##name(struct work_struct *arg)           \
  58 {                                                                       \
  59         struct btrfs_work *work = container_of(arg, struct btrfs_work,  \
  60                                                normal_work);            \
  61         normal_work_helper(work);                                       \
  62 }
  63
  64 struct btrfs_fs_info *
  65 btrfs_workqueue_owner(const struct __btrfs_workqueue *wq)
  66 {
  67         return wq->fs_info;
  68 }
  69
  70 struct btrfs_fs_info *
  71 btrfs_work_owner(const struct btrfs_work *work)
  72 {
  73         return work->wq->fs_info;
  74 }
  75
  76 bool btrfs_workqueue_normal_congested(const struct btrfs_workqueue *wq)
  77 {
  78         /*
  79          * We could compare wq->normal->pending with num_online_cpus()
  80          * to support "thresh == NO_THRESHOLD" case, but it requires
  81          * moving up atomic_inc/dec in thresh_queue/exec_hook. Let's
  82          * postpone it until someone needs the support of that case.
  83          */
  84         if (wq->normal->thresh == NO_THRESHOLD)
  85                 return false;
  86
  87         return atomic_read(&wq->normal->pending) > wq->normal->thresh * 2;
  88 }
  89
  90 BTRFS_WORK_HELPER(worker_helper);
  91 BTRFS_WORK_HELPER(delalloc_helper);
  92 BTRFS_WORK_HELPER(flush_delalloc_helper);
  93 BTRFS_WORK_HELPER(cache_helper);
  94 BTRFS_WORK_HELPER(submit_helper);
  95 BTRFS_WORK_HELPER(fixup_helper);
  96 BTRFS_WORK_HELPER(endio_helper);
  97 BTRFS_WORK_HELPER(endio_meta_helper);
  98 BTRFS_WORK_HELPER(endio_meta_write_helper);
  99 BTRFS_WORK_HELPER(endio_raid56_helper);
 100 BTRFS_WORK_HELPER(endio_repair_helper);
 101 BTRFS_WORK_HELPER(rmw_helper);
 102 BTRFS_WORK_HELPER(endio_write_helper);
 103 BTRFS_WORK_HELPER(freespace_write_helper);
 104 BTRFS_WORK_HELPER(delayed_meta_helper);
 105 BTRFS_WORK_HELPER(readahead_helper);
 106 BTRFS_WORK_HELPER(qgroup_rescan_helper);
 107 BTRFS_WORK_HELPER(extent_refs_helper);
 108 BTRFS_WORK_HELPER(scrub_helper);
 109 BTRFS_WORK_HELPER(scrubwrc_helper);
 110 BTRFS_WORK_HELPER(scrubnc_helper);
 111 BTRFS_WORK_HELPER(scrubparity_helper);
 112
 113 static struct __btrfs_workqueue *
 114 __btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info, const char *name,
 115                         unsigned int flags, int limit_active, int thresh)
 116 {
 117         struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
 118
 119         if (!ret)
 120                 return NULL;
 121
 122         ret->fs_info = fs_info;
 123         ret->limit_active = limit_active;
 124         atomic_set(&ret->pending, 0);
 125         if (thresh == 0)
 126                 thresh = DFT_THRESHOLD;
 127         /* For low threshold, disabling threshold is a better choice */
 128         if (thresh < DFT_THRESHOLD) {
 129                 ret->current_active = limit_active;
 130                 ret->thresh = NO_THRESHOLD;
 131         } else {
 132                 /*
 133                  * For threshold-able wq, let its concurrency grow on demand.
 134                  * Use minimal max_active at alloc time to reduce resource
 135                  * usage.
 136                  */
 137                 ret->current_active = 1;
 138                 ret->thresh = thresh;
 139         }
 140
 141         if (flags & WQ_HIGHPRI)
 142                 ret->normal_wq = alloc_workqueue("btrfs-%s-high", flags,
 143                                                  ret->current_active, name);
 144         else
 145                 ret->normal_wq = alloc_workqueue("btrfs-%s", flags,
 146                                                  ret->current_active, name);
 147         if (!ret->normal_wq) {
 148                 kfree(ret);
 149                 return NULL;
 150         }
 151
 152         INIT_LIST_HEAD(&ret->ordered_list);
 153         spin_lock_init(&ret->list_lock);
 154         spin_lock_init(&ret->thres_lock);
 155         trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI);
 156         return ret;
 157 }
 158
 159 static inline void
 160 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq);
 161
 162 struct btrfs_workqueue *btrfs_alloc_workqueue(struct btrfs_fs_info *fs_info,
 163                                               const char *name,
 164                                               unsigned int flags,
 165                                               int limit_active,
 166                                               int thresh)
 167 {
 168         struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_KERNEL);
 169
 170         if (!ret)
 171                 return NULL;
 172
 173         ret->normal = __btrfs_alloc_workqueue(fs_info, name,
 174                                               flags & ~WQ_HIGHPRI,
 175                                               limit_active, thresh);
 176         if (!ret->normal) {
 177                 kfree(ret);
 178                 return NULL;
 179         }
 180
 181         if (flags & WQ_HIGHPRI) {
 182                 ret->high = __btrfs_alloc_workqueue(fs_info, name, flags,
 183                                                     limit_active, thresh);
 184                 if (!ret->high) {
 185                         __btrfs_destroy_workqueue(ret->normal);
 186                         kfree(ret);
 187                         return NULL;
 188                 }
 189         }
 190         return ret;
 191 }
 192
 193 /*
 194  * Hook for threshold which will be called in btrfs_queue_work.
 195  * This hook WILL be called in IRQ handler context,
 196  * so workqueue_set_max_active MUST NOT be called in this hook
 197  */
 198 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq)
 199 {
 200         if (wq->thresh == NO_THRESHOLD)
 201                 return;
 202         atomic_inc(&wq->pending);
 203 }
 204
 205 /*
 206  * Hook for threshold which will be called before executing the work,
 207  * This hook is called in kthread content.
 208  * So workqueue_set_max_active is called here.
 209  */
 210 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq)
 211 {
 212         int new_current_active;
 213         long pending;
 214         int need_change = 0;
 215
 216         if (wq->thresh == NO_THRESHOLD)
 217                 return;
 218
 219         atomic_dec(&wq->pending);
 220         spin_lock(&wq->thres_lock);
 221         /*
 222          * Use wq->count to limit the calling frequency of
 223          * workqueue_set_max_active.
 224          */
 225         wq->count++;
 226         wq->count %= (wq->thresh / 4);
 227         if (!wq->count)
 228                 goto  out;
 229         new_current_active = wq->current_active;
 230
 231         /*
 232          * pending may be changed later, but it's OK since we really
 233          * don't need it so accurate to calculate new_max_active.
 234          */
 235         pending = atomic_read(&wq->pending);
 236         if (pending > wq->thresh)
 237                 new_current_active++;
 238         if (pending < wq->thresh / 2)
 239                 new_current_active--;
 240         new_current_active = clamp_val(new_current_active, 1, wq->limit_active);
 241         if (new_current_active != wq->current_active)  {
 242                 need_change = 1;
 243                 wq->current_active = new_current_active;
 244         }
 245 out:
 246         spin_unlock(&wq->thres_lock);
 247
 248         if (need_change) {
 249                 workqueue_set_max_active(wq->normal_wq, wq->current_active);
 250         }
 251 }
 252
 253 static void run_ordered_work(struct __btrfs_workqueue *wq)
 254 {
 255         struct list_head *list = &wq->ordered_list;
 256         struct btrfs_work *work;
 257         spinlock_t *lock = &wq->list_lock;
 258         unsigned long flags;
 259
 260         while (1) {
 261                 void *wtag;
 262
 263                 spin_lock_irqsave(lock, flags);
 264                 if (list_empty(list))
 265                         break;
 266                 work = list_entry(list->next, struct btrfs_work,
 267                                   ordered_list);
 268                 if (!test_bit(WORK_DONE_BIT, &work->flags))
 269                         break;
 270
 271                 /*
 272                  * we are going to call the ordered done function, but
 273                  * we leave the work item on the list as a barrier so
 274                  * that later work items that are done don't have their
 275                  * functions called before this one returns
 276                  */
 277                 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags))
 278                         break;
 279                 trace_btrfs_ordered_sched(work);
 280                 spin_unlock_irqrestore(lock, flags);
 281                 work->ordered_func(work);
 282
 283                 /* now take the lock again and drop our item from the list */
 284                 spin_lock_irqsave(lock, flags);
 285                 list_del(&work->ordered_list);
 286                 spin_unlock_irqrestore(lock, flags);
 287
 288                 /*
 289                  * We don't want to call the ordered free functions with the
 290                  * lock held though. Save the work as tag for the trace event,
 291                  * because the callback could free the structure.
 292                  */
 293                 wtag = work;
 294                 work->ordered_free(work);
 295                 trace_btrfs_all_work_done(wq->fs_info, wtag);
 296         }
 297         spin_unlock_irqrestore(lock, flags);
 298 }
 299
 300 static void normal_work_helper(struct btrfs_work *work)
 301 {
 302         struct __btrfs_workqueue *wq;
 303         void *wtag;
 304         int need_order = 0;
 305
 306         /*
 307          * We should not touch things inside work in the following cases:
 308          * 1) after work->func() if it has no ordered_free
 309          *    Since the struct is freed in work->func().
 310          * 2) after setting WORK_DONE_BIT
 311          *    The work may be freed in other threads almost instantly.
 312          * So we save the needed things here.
 313          */
 314         if (work->ordered_func)
 315                 need_order = 1;
 316         wq = work->wq;
 317         /* Safe for tracepoints in case work gets freed by the callback */
 318         wtag = work;
 319
 320         trace_btrfs_work_sched(work);
 321         thresh_exec_hook(wq);
 322         work->func(work);
 323         if (need_order) {
 324                 set_bit(WORK_DONE_BIT, &work->flags);
 325                 run_ordered_work(wq);
 326         }
 327         if (!need_order)
 328                 trace_btrfs_all_work_done(wq->fs_info, wtag);
 329 }
 330
 331 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func,
 332                      btrfs_func_t func,
 333                      btrfs_func_t ordered_func,
 334                      btrfs_func_t ordered_free)
 335 {
 336         work->func = func;
 337         work->ordered_func = ordered_func;
 338         work->ordered_free = ordered_free;
 339         INIT_WORK(&work->normal_work, uniq_func);
 340         INIT_LIST_HEAD(&work->ordered_list);
 341         work->flags = 0;
 342 }
 343
 344 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq,
 345                                       struct btrfs_work *work)
 346 {
 347         unsigned long flags;
 348
 349         work->wq = wq;
 350         thresh_queue_hook(wq);
 351         if (work->ordered_func) {
 352                 spin_lock_irqsave(&wq->list_lock, flags);
 353                 list_add_tail(&work->ordered_list, &wq->ordered_list);
 354                 spin_unlock_irqrestore(&wq->list_lock, flags);
 355         }
 356         trace_btrfs_work_queued(work);
 357         queue_work(wq->normal_wq, &work->normal_work);
 358 }
 359
 360 void btrfs_queue_work(struct btrfs_workqueue *wq,
 361                       struct btrfs_work *work)
 362 {
 363         struct __btrfs_workqueue *dest_wq;
 364
 365         if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high)
 366                 dest_wq = wq->high;
 367         else
 368                 dest_wq = wq->normal;
 369         __btrfs_queue_work(dest_wq, work);
 370 }
 371
 372 static inline void
 373 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq)
 374 {
 375         destroy_workqueue(wq->normal_wq);
 376         trace_btrfs_workqueue_destroy(wq);
 377         kfree(wq);
 378 }
 379
 380 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq)
 381 {
 382         if (!wq)
 383                 return;
 384         if (wq->high)
 385                 __btrfs_destroy_workqueue(wq->high);
 386         __btrfs_destroy_workqueue(wq->normal);
 387         kfree(wq);
 388 }
 389
 390 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int limit_active)
 391 {
 392         if (!wq)
 393                 return;
 394         wq->normal->limit_active = limit_active;
 395         if (wq->high)
 396                 wq->high->limit_active = limit_active;
 397 }
 398
 399 void btrfs_set_work_high_priority(struct btrfs_work *work)
 400 {
 401         set_bit(WORK_HIGH_PRIO_BIT, &work->flags);
 402 }