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 35struct __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 52struct btrfs_workqueue { 53 struct __btrfs_workqueue *normal; 54 struct __btrfs_workqueue *high; 55}; 56 57static void normal_work_helper(struct btrfs_work *work); 58 59#define BTRFS_WORK_HELPER(name) \ 60void 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 67BTRFS_WORK_HELPER(worker_helper); 68BTRFS_WORK_HELPER(delalloc_helper); 69BTRFS_WORK_HELPER(flush_delalloc_helper); 70BTRFS_WORK_HELPER(cache_helper); 71BTRFS_WORK_HELPER(submit_helper); 72BTRFS_WORK_HELPER(fixup_helper); 73BTRFS_WORK_HELPER(endio_helper); 74BTRFS_WORK_HELPER(endio_meta_helper); 75BTRFS_WORK_HELPER(endio_meta_write_helper); 76BTRFS_WORK_HELPER(endio_raid56_helper); 77BTRFS_WORK_HELPER(endio_repair_helper); 78BTRFS_WORK_HELPER(rmw_helper); 79BTRFS_WORK_HELPER(endio_write_helper); 80BTRFS_WORK_HELPER(freespace_write_helper); 81BTRFS_WORK_HELPER(delayed_meta_helper); 82BTRFS_WORK_HELPER(readahead_helper); 83BTRFS_WORK_HELPER(qgroup_rescan_helper); 84BTRFS_WORK_HELPER(extent_refs_helper); 85BTRFS_WORK_HELPER(scrub_helper); 86BTRFS_WORK_HELPER(scrubwrc_helper); 87BTRFS_WORK_HELPER(scrubnc_helper); 88 89static struct __btrfs_workqueue * 90__btrfs_alloc_workqueue(const char *name, unsigned int flags, int max_active, 91 int thresh) 92{ 93 struct __btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); 94 95 if (!ret) 96 return NULL; 97 98 ret->max_active = max_active; 99 atomic_set(&ret->pending, 0); 100 if (thresh == 0) 101 thresh = DFT_THRESHOLD; 102 /* For low threshold, disabling threshold is a better choice */ 103 if (thresh < DFT_THRESHOLD) { 104 ret->current_max = max_active; 105 ret->thresh = NO_THRESHOLD; 106 } else { 107 ret->current_max = 1; 108 ret->thresh = thresh; 109 } 110 111 if (flags & WQ_HIGHPRI) 112 ret->normal_wq = alloc_workqueue("%s-%s-high", flags, 113 ret->max_active, 114 "btrfs", name); 115 else 116 ret->normal_wq = alloc_workqueue("%s-%s", flags, 117 ret->max_active, "btrfs", 118 name); 119 if (!ret->normal_wq) { 120 kfree(ret); 121 return NULL; 122 } 123 124 INIT_LIST_HEAD(&ret->ordered_list); 125 spin_lock_init(&ret->list_lock); 126 spin_lock_init(&ret->thres_lock); 127 trace_btrfs_workqueue_alloc(ret, name, flags & WQ_HIGHPRI); 128 return ret; 129} 130 131static inline void 132__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq); 133 134struct btrfs_workqueue *btrfs_alloc_workqueue(const char *name, 135 unsigned int flags, 136 int max_active, 137 int thresh) 138{ 139 struct btrfs_workqueue *ret = kzalloc(sizeof(*ret), GFP_NOFS); 140 141 if (!ret) 142 return NULL; 143 144 ret->normal = __btrfs_alloc_workqueue(name, flags & ~WQ_HIGHPRI, 145 max_active, thresh); 146 if (!ret->normal) { 147 kfree(ret); 148 return NULL; 149 } 150 151 if (flags & WQ_HIGHPRI) { 152 ret->high = __btrfs_alloc_workqueue(name, flags, max_active, 153 thresh); 154 if (!ret->high) { 155 __btrfs_destroy_workqueue(ret->normal); 156 kfree(ret); 157 return NULL; 158 } 159 } 160 return ret; 161} 162 163/* 164 * Hook for threshold which will be called in btrfs_queue_work. 165 * This hook WILL be called in IRQ handler context, 166 * so workqueue_set_max_active MUST NOT be called in this hook 167 */ 168static inline void thresh_queue_hook(struct __btrfs_workqueue *wq) 169{ 170 if (wq->thresh == NO_THRESHOLD) 171 return; 172 atomic_inc(&wq->pending); 173} 174 175/* 176 * Hook for threshold which will be called before executing the work, 177 * This hook is called in kthread content. 178 * So workqueue_set_max_active is called here. 179 */ 180static inline void thresh_exec_hook(struct __btrfs_workqueue *wq) 181{ 182 int new_max_active; 183 long pending; 184 int need_change = 0; 185 186 if (wq->thresh == NO_THRESHOLD) 187 return; 188 189 atomic_dec(&wq->pending); 190 spin_lock(&wq->thres_lock); 191 /* 192 * Use wq->count to limit the calling frequency of 193 * workqueue_set_max_active. 194 */ 195 wq->count++; 196 wq->count %= (wq->thresh / 4); 197 if (!wq->count) 198 goto out; 199 new_max_active = wq->current_max; 200 201 /* 202 * pending may be changed later, but it's OK since we really 203 * don't need it so accurate to calculate new_max_active. 204 */ 205 pending = atomic_read(&wq->pending); 206 if (pending > wq->thresh) 207 new_max_active++; 208 if (pending < wq->thresh / 2) 209 new_max_active--; 210 new_max_active = clamp_val(new_max_active, 1, wq->max_active); 211 if (new_max_active != wq->current_max) { 212 need_change = 1; 213 wq->current_max = new_max_active; 214 } 215out: 216 spin_unlock(&wq->thres_lock); 217 218 if (need_change) { 219 workqueue_set_max_active(wq->normal_wq, wq->current_max); 220 } 221} 222 223static void run_ordered_work(struct __btrfs_workqueue *wq) 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; 229 230 while (1) { 231 spin_lock_irqsave(lock, flags); 232 if (list_empty(list)) 233 break; 234 work = list_entry(list->next, struct btrfs_work, 235 ordered_list); 236 if (!test_bit(WORK_DONE_BIT, &work->flags)) 237 break; 238 239 /* 240 * we are going to call the ordered done function, but 241 * we leave the work item on the list as a barrier so 242 * that later work items that are done don't have their 243 * functions called before this one returns 244 */ 245 if (test_and_set_bit(WORK_ORDER_DONE_BIT, &work->flags)) 246 break; 247 trace_btrfs_ordered_sched(work); 248 spin_unlock_irqrestore(lock, flags); 249 work->ordered_func(work); 250 251 /* now take the lock again and drop our item from the list */ 252 spin_lock_irqsave(lock, flags); 253 list_del(&work->ordered_list); 254 spin_unlock_irqrestore(lock, flags); 255 256 /* 257 * we don't want to call the ordered free functions 258 * with the lock held though 259 */ 260 work->ordered_free(work); 261 trace_btrfs_all_work_done(work); 262 } 263 spin_unlock_irqrestore(lock, flags); 264} 265 266static void normal_work_helper(struct btrfs_work *work) 267{ 268 struct __btrfs_workqueue *wq; 269 int need_order = 0; 270 271 /* 272 * We should not touch things inside work in the following cases: 273 * 1) after work->func() if it has no ordered_free 274 * Since the struct is freed in work->func(). 275 * 2) after setting WORK_DONE_BIT 276 * The work may be freed in other threads almost instantly. 277 * So we save the needed things here. 278 */ 279 if (work->ordered_func) 280 need_order = 1; 281 wq = work->wq; 282 283 trace_btrfs_work_sched(work); 284 thresh_exec_hook(wq); 285 work->func(work); 286 if (need_order) { 287 set_bit(WORK_DONE_BIT, &work->flags); 288 run_ordered_work(wq); 289 } 290 if (!need_order) 291 trace_btrfs_all_work_done(work); 292} 293 294void btrfs_init_work(struct btrfs_work *work, btrfs_work_func_t uniq_func, 295 btrfs_func_t func, 296 btrfs_func_t ordered_func, 297 btrfs_func_t ordered_free) 298{ 299 work->func = func; 300 work->ordered_func = ordered_func; 301 work->ordered_free = ordered_free; 302 INIT_WORK(&work->normal_work, uniq_func); 303 INIT_LIST_HEAD(&work->ordered_list); 304 work->flags = 0; 305} 306 307static inline void __btrfs_queue_work(struct __btrfs_workqueue *wq, 308 struct btrfs_work *work) 309{ 310 unsigned long flags; 311 312 work->wq = wq; 313 thresh_queue_hook(wq); 314 if (work->ordered_func) { 315 spin_lock_irqsave(&wq->list_lock, flags); 316 list_add_tail(&work->ordered_list, &wq->ordered_list); 317 spin_unlock_irqrestore(&wq->list_lock, flags); 318 } 319 trace_btrfs_work_queued(work); 320 queue_work(wq->normal_wq, &work->normal_work); 321} 322 323void btrfs_queue_work(struct btrfs_workqueue *wq, 324 struct btrfs_work *work) 325{ 326 struct __btrfs_workqueue *dest_wq; 327 328 if (test_bit(WORK_HIGH_PRIO_BIT, &work->flags) && wq->high) 329 dest_wq = wq->high; 330 else 331 dest_wq = wq->normal; 332 __btrfs_queue_work(dest_wq, work); 333} 334 335static inline void 336__btrfs_destroy_workqueue(struct __btrfs_workqueue *wq) 337{ 338 destroy_workqueue(wq->normal_wq); 339 trace_btrfs_workqueue_destroy(wq); 340 kfree(wq); 341} 342 343void btrfs_destroy_workqueue(struct btrfs_workqueue *wq) 344{ 345 if (!wq) 346 return; 347 if (wq->high) 348 __btrfs_destroy_workqueue(wq->high); 349 __btrfs_destroy_workqueue(wq->normal); 350 kfree(wq); 351} 352 353void btrfs_workqueue_set_max(struct btrfs_workqueue *wq, int max) 354{ 355 if (!wq) 356 return; 357 wq->normal->max_active = max; 358 if (wq->high) 359 wq->high->max_active = max; 360} 361 362void btrfs_set_work_high_priority(struct btrfs_work *work) 363{ 364 set_bit(WORK_HIGH_PRIO_BIT, &work->flags); 365} 366