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(endio_repair_helper); 78 BTRFS_WORK_HELPER(rmw_helper); 79 BTRFS_WORK_HELPER(endio_write_helper); 80 BTRFS_WORK_HELPER(freespace_write_helper); 81 BTRFS_WORK_HELPER(delayed_meta_helper); 82 BTRFS_WORK_HELPER(readahead_helper); 83 BTRFS_WORK_HELPER(qgroup_rescan_helper); 84 BTRFS_WORK_HELPER(extent_refs_helper); 85 BTRFS_WORK_HELPER(scrub_helper); 86 BTRFS_WORK_HELPER(scrubwrc_helper); 87 BTRFS_WORK_HELPER(scrubnc_helper); 88 BTRFS_WORK_HELPER(scrubparity_helper); 89 90 static struct __btrfs_workqueue * 91 __btrfs_alloc_workqueue(const char *name, unsigned int flags, int max_active, 92 int thresh) 93 { 94 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); 95 96 if (!ret) 97 return NULL; 98 99 ret->max_active = max_active; 100 atomic_set(&ret->pending, 0); 101 if (thresh == 0) 102 thresh = DFT_THRESHOLD; 103 /* For low threshold, disabling threshold is a better choice */ 104 if (thresh < DFT_THRESHOLD) { 105 ret->current_max = max_active; 106 ret->thresh = NO_THRESHOLD; 107 } else { 108 ret->current_max = 1; 109 ret->thresh = thresh; 110 } 111 112 if (flags & WQ_HIGHPRI) 113 ret->normal_wq = alloc_workqueue("%s-%s-high", flags, 114 ret->max_active, 115 "btrfs", name); 116 else 117 ret->normal_wq = alloc_workqueue("%s-%s", flags, 118 ret->max_active, "btrfs", 119 name); 120 if (!ret->normal_wq) { 121 kfree(ret); 122 return NULL; 123 } 124 125 INIT_LIST_HEAD(&ret->ordered_list); 126 spin_lock_init(&ret->list_lock); 127 spin_lock_init(&ret->thres_lock); 128 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI); 129 return ret; 130 } 131 132 static inline void 133 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq); 134 135 struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name, 136 unsigned int flags, 137 int max_active, 138 int thresh) 139 { 140 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); 141 142 if (!ret) 143 return NULL; 144 145 ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI, 146 max_active, thresh); 147 if (!ret->normal) { 148 kfree(ret); 149 return NULL; 150 } 151 152 if (flags & WQ_HIGHPRI) { 153 ret->high = __btrfs_alloc_workqueue(name, flags, max_active, 154 thresh); 155 if (!ret->high) { 156 __btrfs_destroy_workqueue(ret->normal); 157 kfree(ret); 158 return NULL; 159 } 160 } 161 return ret; 162 } 163 164 /* 165 * Hook for threshold which will be called in btrfs_queue_work. 166 * This hook WILL be called in IRQ handler context, 167 * so workqueue_set_max_active MUST NOT be called in this hook 168 */ 169 static inline void thresh_queue_hook(struct __btrfs_workqueue *wq) 170 { 171 if (wq->thresh == NO_THRESHOLD) 172 return; 173 atomic_inc(&wq->pending); 174 } 175 176 /* 177 * Hook for threshold which will be called before executing the work, 178 * This hook is called in kthread content. 179 * So workqueue_set_max_active is called here. 180 */ 181 static inline void thresh_exec_hook(struct __btrfs_workqueue *wq) 182 { 183 int new_max_active; 184 long pending; 185 int need_change = 0; 186 187 if (wq->thresh == NO_THRESHOLD) 188 return; 189 190 atomic_dec(&wq->pending); 191 spin_lock(&wq->thres_lock); 192 /* 193 * Use wq->count to limit the calling frequency of 194 * workqueue_set_max_active. 195 */ 196 wq->count++; 197 wq->count %= (wq->thresh / 4); 198 if (!wq->count) 199 goto out; 200 new_max_active = wq->current_max; 201 202 /* 203 * pending may be changed later, but it's OK since we really 204 * don't need it so accurate to calculate new_max_active. 205 */ 206 pending = atomic_read(&wq->pending); 207 if (pending > wq->thresh) 208 new_max_active++; 209 if (pending < wq->thresh / 2) 210 new_max_active--; 211 new_max_active = clamp_val(new_max_active, 1, wq->max_active); 212 if (new_max_active != wq->current_max) { 213 need_change = 1; 214 wq->current_max = new_max_active; 215 } 216 out: 217 spin_unlock(&wq->thres_lock); 218 219 if (need_change) { 220 workqueue_set_max_active(wq->normal_wq, wq->current_max); 221 } 222 } 223 224 static void run_ordered_work(struct __btrfs_workqueue *wq) 225 { 226 struct list_head *list = &wq->ordered_list; 227 struct btrfs_work *work; 228 spinlock_t *lock = &wq->list_lock; 229 unsigned long flags; 230 231 while (1) { 232 spin_lock_irqsave(lock, flags); 233 if (list_empty(list)) 234 break; 235 work = list_entry(list->next, struct btrfs_work, 236 ordered_list); 237 if (!test_bit(WORK_DONE_BIT, &work->flags)) 238 break; 239 240 /* 241 * we are going to call the ordered done function, but 242 * we leave the work item on the list as a barrier so 243 * that later work items that are done don't have their 244 * functions called before this one returns 245 */ 246 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags)) 247 break; 248 trace_btrfs_ordered_sched(work); 249 spin_unlock_irqrestore(lock, flags); 250 work->ordered_func(work); 251 252 /* now take the lock again and drop our item from the list */ 253 spin_lock_irqsave(lock, flags); 254 list_del(&work->ordered_list); 255 spin_unlock_irqrestore(lock, flags); 256 257 /* 258 * we don't want to call the ordered free functions 259 * with the lock held though 260 */ 261 work->ordered_free(work); 262 trace_btrfs_all_work_done(work); 263 } 264 spin_unlock_irqrestore(lock, flags); 265 } 266 267 static void normal_work_helper(struct btrfs_work *work) 268 { 269 struct __btrfs_workqueue *wq; 270 int need_order = 0; 271 272 /* 273 * We should not touch things inside work in the following cases: 274 * 1) after work->func() if it has no ordered_free 275 * Since the struct is freed in work->func(). 276 * 2) after setting WORK_DONE_BIT 277 * The work may be freed in other threads almost instantly. 278 * So we save the needed things here. 279 */ 280 if (work->ordered_func) 281 need_order = 1; 282 wq = work->wq; 283 284 trace_btrfs_work_sched(work); 285 thresh_exec_hook(wq); 286 work->func(work); 287 if (need_order) { 288 set_bit(WORK_DONE_BIT, &work->flags); 289 run_ordered_work(wq); 290 } 291 if (!need_order) 292 trace_btrfs_all_work_done(work); 293 } 294 295 void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func, 296 btrfs_func_t func, 297 btrfs_func_t ordered_func, 298 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, uniq_func); 304 INIT_LIST_HEAD(&work->ordered_list); 305 work->flags = 0; 306 } 307 308 static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq, 309 struct btrfs_work *work) 310 { 311 unsigned long flags; 312 313 work->wq = wq; 314 thresh_queue_hook(wq); 315 if (work->ordered_func) { 316 spin_lock_irqsave(&wq->list_lock, flags); 317 list_add_tail(&work->ordered_list, &wq->ordered_list); 318 spin_unlock_irqrestore(&wq->list_lock, flags); 319 } 320 queue_work(wq->normal_wq, &work->normal_work); 321 trace_btrfs_work_queued(work); 322 } 323 324 void btrfs_queue_work(struct btrfs_workqueue *wq, 325 struct btrfs_work *work) 326 { 327 struct __btrfs_workqueue *dest_wq; 328 329 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high) 330 dest_wq = wq->high; 331 else 332 dest_wq = wq->normal; 333 __btrfs_queue_work(dest_wq, work); 334 } 335 336 static inline void 337 __btrfs_destroy_workqueue(struct __btrfs_workqueue *wq) 338 { 339 destroy_workqueue(wq->normal_wq); 340 trace_btrfs_workqueue_destroy(wq); 341 kfree(wq); 342 } 343 344 void btrfs_destroy_workqueue(struct btrfs_workqueue *wq) 345 { 346 if (!wq) 347 return; 348 if (wq->high) 349 __btrfs_destroy_workqueue(wq->high); 350 __btrfs_destroy_workqueue(wq->normal); 351 kfree(wq); 352 } 353 354 void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int max) 355 { 356 if (!wq) 357 return; 358 wq->normal->max_active = max; 359 if (wq->high) 360 wq->high->max_active = max; 361 } 362 363 void btrfs_set_work_high_priority(struct btrfs_work *work) 364 { 365 set_bit(WORK_HIGH_PRIO_BIT, &work->flags); 366 } 367