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