1 /* 2 * Copyright (C) 2007 Oracle. All rights reserved. 3 * Copyright (C) 2014 Fujitsu. All rights reserved. 4 * 5 * This program is free software; you can redistribute it and/or 6 * modify it under the terms of the GNU General Public 7 * License v2 as published by the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it will be useful, 10 * but WITHOUT ANY WARRANTY; without even the implied warranty of 11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 12 * General Public License for more details. 13 * 14 * You should have received a copy of the GNU General Public 15 * License along with this program; if not, write to the 16 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 17 * Boston, MA 021110-1307, USA. 18 */ 19 20 #include <linux/kthread.h> 21 #include <linux/slab.h> 22 #include <linux/list.h> 23 #include <linux/spinlock.h> 24 #include <linux/freezer.h> 25 #include "async-thread.h" 26 #include "ctree.h" 27 28 #define WORK_DONE_BIT 0 29 #define WORK_ORDER_DONE_BIT 1 30 #define WORK_HIGH_PRIO_BIT 2 31 32 #define NO_THRESHOLD (-1) 33 #define DFT_THRESHOLD (32) 34 35 struct __btrfs_workqueue { 36 struct workqueue_struct *normal_wq; 37 /* List head pointing to ordered work list */ 38 struct list_head ordered_list; 39 40 /* Spinlock for ordered_list */ 41 spinlock_t list_lock; 42 43 /* Thresholding related variants */ 44 atomic_t pending; 45 int max_active; 46 int current_max; 47 int thresh; 48 unsigned int count; 49 spinlock_t thres_lock; 50 }; 51 52 struct btrfs_workqueue { 53 struct __btrfs_workqueue *normal; 54 struct __btrfs_workqueue *high; 55 }; 56 57 static void normal_work_helper(struct btrfs_work *work); 58 59 #define BTRFS_WORK_HELPER(name) \ 60 void btrfs_##name(struct work_struct *arg) \ 61 { \ 62 struct btrfs_work *work = container_of(arg, struct btrfs_work, \ 63 normal_work); \ 64 normal_work_helper(work); \ 65 } 66 67 BTRFS_WORK_HELPER(worker_helper); 68 BTRFS_WORK_HELPER(delalloc_helper); 69 BTRFS_WORK_HELPER(flush_delalloc_helper); 70 BTRFS_WORK_HELPER(cache_helper); 71 BTRFS_WORK_HELPER(submit_helper); 72 BTRFS_WORK_HELPER(fixup_helper); 73 BTRFS_WORK_HELPER(endio_helper); 74 BTRFS_WORK_HELPER(endio_meta_helper); 75 BTRFS_WORK_HELPER(endio_meta_write_helper); 76 BTRFS_WORK_HELPER(endio_raid56_helper); 77 BTRFS_WORK_HELPER(rmw_helper); 78 BTRFS_WORK_HELPER(endio_write_helper); 79 BTRFS_WORK_HELPER(freespace_write_helper); 80 BTRFS_WORK_HELPER(delayed_meta_helper); 81 BTRFS_WORK_HELPER(readahead_helper); 82 BTRFS_WORK_HELPER(qgroup_rescan_helper); 83 BTRFS_WORK_HELPER(extent_refs_helper); 84 BTRFS_WORK_HELPER(scrub_helper); 85 BTRFS_WORK_HELPER(scrubwrc_helper); 86 BTRFS_WORK_HELPER(scrubnc_helper); 87 88 static struct __btrfs_workqueue * 89 __btrfs_alloc_workqueue(const char *name, int flags, int max_active, 90 int thresh) 91 { 92 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); 93 94 if (unlikely(!ret)) 95 return NULL; 96 97 ret->max_active = max_active; 98 atomic_set(&ret->pending, 0); 99 if (thresh == 0) 100 thresh = DFT_THRESHOLD; 101 /* For low threshold, disabling threshold is a better choice */ 102 if (thresh < DFT_THRESHOLD) { 103 ret->current_max = max_active; 104 ret->thresh = NO_THRESHOLD; 105 } else { 106 ret->current_max = 1; 107 ret->thresh = thresh; 108 } 109 110 if (flags & WQ_HIGHPRI) 111 ret->normal_wq = alloc_workqueue("%s-%s-high", flags, 112 ret->max_active, 113 "btrfs", name); 114 else 115 ret->normal_wq = alloc_workqueue("%s-%s", flags, 116 ret->max_active, "btrfs", 117 name); 118 if (unlikely(!ret->normal_wq)) { 119 kfree(ret); 120 return NULL; 121 } 122 123 INIT_LIST_HEAD(&ret->ordered_list); 124 spin_lock_init(&ret->list_lock); 125 spin_lock_init(&ret->thres_lock); 126 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI); 127 return ret; 128 } 129 130 static inline void 131 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq); 132 133 struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name, 134 int flags, 135 int max_active, 136 int thresh) 137 { 138 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); 139 140 if (unlikely(!ret)) 141 return NULL; 142 143 ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI, 144 max_active, thresh); 145 if (unlikely(!ret->normal)) { 146 kfree(ret); 147 return NULL; 148 } 149 150 if (flags & WQ_HIGHPRI) { 151 ret->high = __btrfs_alloc_workqueue(name, flags, max_active, 152 thresh); 153 if (unlikely(!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_max_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_max_active = wq->current_max; 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_max_active++; 207 if (pending < wq->thresh / 2) 208 new_max_active--; 209 new_max_active = clamp_val(new_max_active, 1, wq->max_active); 210 if (new_max_active != wq->current_max) { 211 need_change = 1; 212 wq->current_max = new_max_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_max); 219 } 220 } 221 222 static void run_ordered_work(struct __btrfs_workqueue *wq) 223 { 224 struct list_head *list = &wq->ordered_list; 225 struct btrfs_work *work; 226 spinlock_t *lock = &wq->list_lock; 227 unsigned long flags; 228 229 while (1) { 230 spin_lock_irqsave(lock, flags); 231 if (list_empty(list)) 232 break; 233 work = list_entry(list->next, struct btrfs_work, 234 ordered_list); 235 if (!test_bit(WORK_DONE_BIT, &work->flags)) 236 break; 237 238 /* 239 * we are going to call the ordered done function, but 240 * we leave the work item on the list as a barrier so 241 * that later work items that are done don't have their 242 * functions called before this one returns 243 */ 244 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags)) 245 break; 246 trace_btrfs_ordered_sched(work); 247 spin_unlock_irqrestore(lock, flags); 248 work->ordered_func(work); 249 250 /* now take the lock again and drop our item from the list */ 251 spin_lock_irqsave(lock, flags); 252 list_del(&work->ordered_list); 253 spin_unlock_irqrestore(lock, flags); 254 255 /* 256 * we don't want to call the ordered free functions 257 * with the lock held though 258 */ 259 work->ordered_free(work); 260 trace_btrfs_all_work_done(work); 261 } 262 spin_unlock_irqrestore(lock, flags); 263 } 264 265 static void normal_work_helper(struct btrfs_work *work) 266 { 267 struct __btrfs_workqueue *wq; 268 int need_order = 0; 269 270 /* 271 * We should not touch things inside work in the following cases: 272 * 1) after work->func() if it has no ordered_free 273 * Since the struct is freed in work->func(). 274 * 2) after setting WORK_DONE_BIT 275 * The work may be freed in other threads almost instantly. 276 * So we save the needed things here. 277 */ 278 if (work->ordered_func) 279 need_order = 1; 280 wq = work->wq; 281 282 trace_btrfs_work_sched(work); 283 thresh_exec_hook(wq); 284 work->func(work); 285 if (need_order) { 286 set_bit(WORK_DONE_BIT, &work->flags); 287 run_ordered_work(wq); 288 } 289 if (!need_order) 290 trace_btrfs_all_work_done(work); 291 } 292 293 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func, 294 btrfs_func_t func, 295 btrfs_func_t ordered_func, 296 btrfs_func_t ordered_free) 297 { 298 work->func = func; 299 work->ordered_func = ordered_func; 300 work->ordered_free = ordered_free; 301 INIT_WORK(&work->normal_work, uniq_func); 302 INIT_LIST_HEAD(&work->ordered_list); 303 work->flags = 0; 304 } 305 306 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq, 307 struct btrfs_work *work) 308 { 309 unsigned long flags; 310 311 work->wq = wq; 312 thresh_queue_hook(wq); 313 if (work->ordered_func) { 314 spin_lock_irqsave(&wq->list_lock, flags); 315 list_add_tail(&work->ordered_list, &wq->ordered_list); 316 spin_unlock_irqrestore(&wq->list_lock, flags); 317 } 318 queue_work(wq->normal_wq, &work->normal_work); 319 trace_btrfs_work_queued(work); 320 } 321 322 void btrfs_queue_work(struct btrfs_workqueue *wq, 323 struct btrfs_work *work) 324 { 325 struct __btrfs_workqueue *dest_wq; 326 327 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high) 328 dest_wq = wq->high; 329 else 330 dest_wq = wq->normal; 331 __btrfs_queue_work(dest_wq, work); 332 } 333 334 static inline void 335 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq) 336 { 337 destroy_workqueue(wq->normal_wq); 338 trace_btrfs_workqueue_destroy(wq); 339 kfree(wq); 340 } 341 342 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq) 343 { 344 if (!wq) 345 return; 346 if (wq->high) 347 __btrfs_destroy_workqueue(wq->high); 348 __btrfs_destroy_workqueue(wq->normal); 349 kfree(wq); 350 } 351 352 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int max) 353 { 354 if (!wq) 355 return; 356 wq->normal->max_active = max; 357 if (wq->high) 358 wq->high->max_active = max; 359 } 360 361 void btrfs_set_work_high_priority(struct btrfs_work *work) 362 { 363 set_bit(WORK_HIGH_PRIO_BIT, &work->flags); 364 } 365