aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--fs/btrfs/extent-tree.c3
-rw-r--r--fs/buffer.c2
-rw-r--r--fs/ext4/inode.c2
-rw-r--r--fs/fs-writeback.c84
-rw-r--r--fs/quota/quota.c2
-rw-r--r--fs/sync.c4
-rw-r--r--fs/ubifs/budget.c2
-rw-r--r--include/linux/backing-dev.h14
-rw-r--r--include/linux/sched.h7
-rw-r--r--include/linux/writeback.h33
-rw-r--r--include/trace/events/writeback.h161
-rw-r--r--kernel/fork.c3
-rw-r--r--mm/backing-dev.c7
-rw-r--r--mm/page-writeback.c704
-rw-r--r--mm/vmscan.c3
15 files changed, 806 insertions, 225 deletions
diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
index f5be06a2462f..c9ee0e18bbdc 100644
--- a/fs/btrfs/extent-tree.c
+++ b/fs/btrfs/extent-tree.c
@@ -3340,7 +3340,8 @@ static int shrink_delalloc(struct btrfs_trans_handle *trans,
3340 smp_mb(); 3340 smp_mb();
3341 nr_pages = min_t(unsigned long, nr_pages, 3341 nr_pages = min_t(unsigned long, nr_pages,
3342 root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT); 3342 root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3343 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages); 3343 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3344 WB_REASON_FS_FREE_SPACE);
3344 3345
3345 spin_lock(&space_info->lock); 3346 spin_lock(&space_info->lock);
3346 if (reserved > space_info->bytes_reserved) 3347 if (reserved > space_info->bytes_reserved)
diff --git a/fs/buffer.c b/fs/buffer.c
index 70a19745cb61..19d8eb7fdc81 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -288,7 +288,7 @@ static void free_more_memory(void)
288 struct zone *zone; 288 struct zone *zone;
289 int nid; 289 int nid;
290 290
291 wakeup_flusher_threads(1024); 291 wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
292 yield(); 292 yield();
293 293
294 for_each_online_node(nid) { 294 for_each_online_node(nid) {
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index cc5a6da030a1..240f6e2dc7ee 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -2372,7 +2372,7 @@ static int ext4_nonda_switch(struct super_block *sb)
2372 * start pushing delalloc when 1/2 of free blocks are dirty. 2372 * start pushing delalloc when 1/2 of free blocks are dirty.
2373 */ 2373 */
2374 if (free_blocks < 2 * dirty_blocks) 2374 if (free_blocks < 2 * dirty_blocks)
2375 writeback_inodes_sb_if_idle(sb); 2375 writeback_inodes_sb_if_idle(sb, WB_REASON_FS_FREE_SPACE);
2376 2376
2377 return 0; 2377 return 0;
2378} 2378}
diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index 04cf3b91e501..73c3992b2bb4 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -41,11 +41,23 @@ struct wb_writeback_work {
41 unsigned int for_kupdate:1; 41 unsigned int for_kupdate:1;
42 unsigned int range_cyclic:1; 42 unsigned int range_cyclic:1;
43 unsigned int for_background:1; 43 unsigned int for_background:1;
44 enum wb_reason reason; /* why was writeback initiated? */
44 45
45 struct list_head list; /* pending work list */ 46 struct list_head list; /* pending work list */
46 struct completion *done; /* set if the caller waits */ 47 struct completion *done; /* set if the caller waits */
47}; 48};
48 49
50const char *wb_reason_name[] = {
51 [WB_REASON_BACKGROUND] = "background",
52 [WB_REASON_TRY_TO_FREE_PAGES] = "try_to_free_pages",
53 [WB_REASON_SYNC] = "sync",
54 [WB_REASON_PERIODIC] = "periodic",
55 [WB_REASON_LAPTOP_TIMER] = "laptop_timer",
56 [WB_REASON_FREE_MORE_MEM] = "free_more_memory",
57 [WB_REASON_FS_FREE_SPACE] = "fs_free_space",
58 [WB_REASON_FORKER_THREAD] = "forker_thread"
59};
60
49/* 61/*
50 * Include the creation of the trace points after defining the 62 * Include the creation of the trace points after defining the
51 * wb_writeback_work structure so that the definition remains local to this 63 * wb_writeback_work structure so that the definition remains local to this
@@ -115,7 +127,7 @@ static void bdi_queue_work(struct backing_dev_info *bdi,
115 127
116static void 128static void
117__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages, 129__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
118 bool range_cyclic) 130 bool range_cyclic, enum wb_reason reason)
119{ 131{
120 struct wb_writeback_work *work; 132 struct wb_writeback_work *work;
121 133
@@ -135,6 +147,7 @@ __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
135 work->sync_mode = WB_SYNC_NONE; 147 work->sync_mode = WB_SYNC_NONE;
136 work->nr_pages = nr_pages; 148 work->nr_pages = nr_pages;
137 work->range_cyclic = range_cyclic; 149 work->range_cyclic = range_cyclic;
150 work->reason = reason;
138 151
139 bdi_queue_work(bdi, work); 152 bdi_queue_work(bdi, work);
140} 153}
@@ -150,9 +163,10 @@ __bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
150 * completion. Caller need not hold sb s_umount semaphore. 163 * completion. Caller need not hold sb s_umount semaphore.
151 * 164 *
152 */ 165 */
153void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages) 166void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
167 enum wb_reason reason)
154{ 168{
155 __bdi_start_writeback(bdi, nr_pages, true); 169 __bdi_start_writeback(bdi, nr_pages, true, reason);
156} 170}
157 171
158/** 172/**
@@ -251,7 +265,7 @@ static bool inode_dirtied_after(struct inode *inode, unsigned long t)
251 */ 265 */
252static int move_expired_inodes(struct list_head *delaying_queue, 266static int move_expired_inodes(struct list_head *delaying_queue,
253 struct list_head *dispatch_queue, 267 struct list_head *dispatch_queue,
254 unsigned long *older_than_this) 268 struct wb_writeback_work *work)
255{ 269{
256 LIST_HEAD(tmp); 270 LIST_HEAD(tmp);
257 struct list_head *pos, *node; 271 struct list_head *pos, *node;
@@ -262,8 +276,8 @@ static int move_expired_inodes(struct list_head *delaying_queue,
262 276
263 while (!list_empty(delaying_queue)) { 277 while (!list_empty(delaying_queue)) {
264 inode = wb_inode(delaying_queue->prev); 278 inode = wb_inode(delaying_queue->prev);
265 if (older_than_this && 279 if (work->older_than_this &&
266 inode_dirtied_after(inode, *older_than_this)) 280 inode_dirtied_after(inode, *work->older_than_this))
267 break; 281 break;
268 if (sb && sb != inode->i_sb) 282 if (sb && sb != inode->i_sb)
269 do_sb_sort = 1; 283 do_sb_sort = 1;
@@ -302,13 +316,13 @@ out:
302 * | 316 * |
303 * +--> dequeue for IO 317 * +--> dequeue for IO
304 */ 318 */
305static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this) 319static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work)
306{ 320{
307 int moved; 321 int moved;
308 assert_spin_locked(&wb->list_lock); 322 assert_spin_locked(&wb->list_lock);
309 list_splice_init(&wb->b_more_io, &wb->b_io); 323 list_splice_init(&wb->b_more_io, &wb->b_io);
310 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this); 324 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, work);
311 trace_writeback_queue_io(wb, older_than_this, moved); 325 trace_writeback_queue_io(wb, work, moved);
312} 326}
313 327
314static int write_inode(struct inode *inode, struct writeback_control *wbc) 328static int write_inode(struct inode *inode, struct writeback_control *wbc)
@@ -641,31 +655,40 @@ static long __writeback_inodes_wb(struct bdi_writeback *wb,
641 return wrote; 655 return wrote;
642} 656}
643 657
644long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages) 658long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
659 enum wb_reason reason)
645{ 660{
646 struct wb_writeback_work work = { 661 struct wb_writeback_work work = {
647 .nr_pages = nr_pages, 662 .nr_pages = nr_pages,
648 .sync_mode = WB_SYNC_NONE, 663 .sync_mode = WB_SYNC_NONE,
649 .range_cyclic = 1, 664 .range_cyclic = 1,
665 .reason = reason,
650 }; 666 };
651 667
652 spin_lock(&wb->list_lock); 668 spin_lock(&wb->list_lock);
653 if (list_empty(&wb->b_io)) 669 if (list_empty(&wb->b_io))
654 queue_io(wb, NULL); 670 queue_io(wb, &work);
655 __writeback_inodes_wb(wb, &work); 671 __writeback_inodes_wb(wb, &work);
656 spin_unlock(&wb->list_lock); 672 spin_unlock(&wb->list_lock);
657 673
658 return nr_pages - work.nr_pages; 674 return nr_pages - work.nr_pages;
659} 675}
660 676
661static inline bool over_bground_thresh(void) 677static bool over_bground_thresh(struct backing_dev_info *bdi)
662{ 678{
663 unsigned long background_thresh, dirty_thresh; 679 unsigned long background_thresh, dirty_thresh;
664 680
665 global_dirty_limits(&background_thresh, &dirty_thresh); 681 global_dirty_limits(&background_thresh, &dirty_thresh);
666 682
667 return (global_page_state(NR_FILE_DIRTY) + 683 if (global_page_state(NR_FILE_DIRTY) +
668 global_page_state(NR_UNSTABLE_NFS) > background_thresh); 684 global_page_state(NR_UNSTABLE_NFS) > background_thresh)
685 return true;
686
687 if (bdi_stat(bdi, BDI_RECLAIMABLE) >
688 bdi_dirty_limit(bdi, background_thresh))
689 return true;
690
691 return false;
669} 692}
670 693
671/* 694/*
@@ -675,7 +698,7 @@ static inline bool over_bground_thresh(void)
675static void wb_update_bandwidth(struct bdi_writeback *wb, 698static void wb_update_bandwidth(struct bdi_writeback *wb,
676 unsigned long start_time) 699 unsigned long start_time)
677{ 700{
678 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, start_time); 701 __bdi_update_bandwidth(wb->bdi, 0, 0, 0, 0, 0, start_time);
679} 702}
680 703
681/* 704/*
@@ -727,7 +750,7 @@ static long wb_writeback(struct bdi_writeback *wb,
727 * For background writeout, stop when we are below the 750 * For background writeout, stop when we are below the
728 * background dirty threshold 751 * background dirty threshold
729 */ 752 */
730 if (work->for_background && !over_bground_thresh()) 753 if (work->for_background && !over_bground_thresh(wb->bdi))
731 break; 754 break;
732 755
733 if (work->for_kupdate) { 756 if (work->for_kupdate) {
@@ -738,7 +761,7 @@ static long wb_writeback(struct bdi_writeback *wb,
738 761
739 trace_writeback_start(wb->bdi, work); 762 trace_writeback_start(wb->bdi, work);
740 if (list_empty(&wb->b_io)) 763 if (list_empty(&wb->b_io))
741 queue_io(wb, work->older_than_this); 764 queue_io(wb, work);
742 if (work->sb) 765 if (work->sb)
743 progress = writeback_sb_inodes(work->sb, wb, work); 766 progress = writeback_sb_inodes(work->sb, wb, work);
744 else 767 else
@@ -811,13 +834,14 @@ static unsigned long get_nr_dirty_pages(void)
811 834
812static long wb_check_background_flush(struct bdi_writeback *wb) 835static long wb_check_background_flush(struct bdi_writeback *wb)
813{ 836{
814 if (over_bground_thresh()) { 837 if (over_bground_thresh(wb->bdi)) {
815 838
816 struct wb_writeback_work work = { 839 struct wb_writeback_work work = {
817 .nr_pages = LONG_MAX, 840 .nr_pages = LONG_MAX,
818 .sync_mode = WB_SYNC_NONE, 841 .sync_mode = WB_SYNC_NONE,
819 .for_background = 1, 842 .for_background = 1,
820 .range_cyclic = 1, 843 .range_cyclic = 1,
844 .reason = WB_REASON_BACKGROUND,
821 }; 845 };
822 846
823 return wb_writeback(wb, &work); 847 return wb_writeback(wb, &work);
@@ -851,6 +875,7 @@ static long wb_check_old_data_flush(struct bdi_writeback *wb)
851 .sync_mode = WB_SYNC_NONE, 875 .sync_mode = WB_SYNC_NONE,
852 .for_kupdate = 1, 876 .for_kupdate = 1,
853 .range_cyclic = 1, 877 .range_cyclic = 1,
878 .reason = WB_REASON_PERIODIC,
854 }; 879 };
855 880
856 return wb_writeback(wb, &work); 881 return wb_writeback(wb, &work);
@@ -969,7 +994,7 @@ int bdi_writeback_thread(void *data)
969 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back 994 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
970 * the whole world. 995 * the whole world.
971 */ 996 */
972void wakeup_flusher_threads(long nr_pages) 997void wakeup_flusher_threads(long nr_pages, enum wb_reason reason)
973{ 998{
974 struct backing_dev_info *bdi; 999 struct backing_dev_info *bdi;
975 1000
@@ -982,7 +1007,7 @@ void wakeup_flusher_threads(long nr_pages)
982 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 1007 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
983 if (!bdi_has_dirty_io(bdi)) 1008 if (!bdi_has_dirty_io(bdi))
984 continue; 1009 continue;
985 __bdi_start_writeback(bdi, nr_pages, false); 1010 __bdi_start_writeback(bdi, nr_pages, false, reason);
986 } 1011 }
987 rcu_read_unlock(); 1012 rcu_read_unlock();
988} 1013}
@@ -1203,7 +1228,9 @@ static void wait_sb_inodes(struct super_block *sb)
1203 * on how many (if any) will be written, and this function does not wait 1228 * on how many (if any) will be written, and this function does not wait
1204 * for IO completion of submitted IO. 1229 * for IO completion of submitted IO.
1205 */ 1230 */
1206void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr) 1231void writeback_inodes_sb_nr(struct super_block *sb,
1232 unsigned long nr,
1233 enum wb_reason reason)
1207{ 1234{
1208 DECLARE_COMPLETION_ONSTACK(done); 1235 DECLARE_COMPLETION_ONSTACK(done);
1209 struct wb_writeback_work work = { 1236 struct wb_writeback_work work = {
@@ -1212,6 +1239,7 @@ void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1212 .tagged_writepages = 1, 1239 .tagged_writepages = 1,
1213 .done = &done, 1240 .done = &done,
1214 .nr_pages = nr, 1241 .nr_pages = nr,
1242 .reason = reason,
1215 }; 1243 };
1216 1244
1217 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 1245 WARN_ON(!rwsem_is_locked(&sb->s_umount));
@@ -1228,9 +1256,9 @@ EXPORT_SYMBOL(writeback_inodes_sb_nr);
1228 * on how many (if any) will be written, and this function does not wait 1256 * on how many (if any) will be written, and this function does not wait
1229 * for IO completion of submitted IO. 1257 * for IO completion of submitted IO.
1230 */ 1258 */
1231void writeback_inodes_sb(struct super_block *sb) 1259void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
1232{ 1260{
1233 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages()); 1261 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
1234} 1262}
1235EXPORT_SYMBOL(writeback_inodes_sb); 1263EXPORT_SYMBOL(writeback_inodes_sb);
1236 1264
@@ -1241,11 +1269,11 @@ EXPORT_SYMBOL(writeback_inodes_sb);
1241 * Invoke writeback_inodes_sb if no writeback is currently underway. 1269 * Invoke writeback_inodes_sb if no writeback is currently underway.
1242 * Returns 1 if writeback was started, 0 if not. 1270 * Returns 1 if writeback was started, 0 if not.
1243 */ 1271 */
1244int writeback_inodes_sb_if_idle(struct super_block *sb) 1272int writeback_inodes_sb_if_idle(struct super_block *sb, enum wb_reason reason)
1245{ 1273{
1246 if (!writeback_in_progress(sb->s_bdi)) { 1274 if (!writeback_in_progress(sb->s_bdi)) {
1247 down_read(&sb->s_umount); 1275 down_read(&sb->s_umount);
1248 writeback_inodes_sb(sb); 1276 writeback_inodes_sb(sb, reason);
1249 up_read(&sb->s_umount); 1277 up_read(&sb->s_umount);
1250 return 1; 1278 return 1;
1251 } else 1279 } else
@@ -1262,11 +1290,12 @@ EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1262 * Returns 1 if writeback was started, 0 if not. 1290 * Returns 1 if writeback was started, 0 if not.
1263 */ 1291 */
1264int writeback_inodes_sb_nr_if_idle(struct super_block *sb, 1292int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1265 unsigned long nr) 1293 unsigned long nr,
1294 enum wb_reason reason)
1266{ 1295{
1267 if (!writeback_in_progress(sb->s_bdi)) { 1296 if (!writeback_in_progress(sb->s_bdi)) {
1268 down_read(&sb->s_umount); 1297 down_read(&sb->s_umount);
1269 writeback_inodes_sb_nr(sb, nr); 1298 writeback_inodes_sb_nr(sb, nr, reason);
1270 up_read(&sb->s_umount); 1299 up_read(&sb->s_umount);
1271 return 1; 1300 return 1;
1272 } else 1301 } else
@@ -1290,6 +1319,7 @@ void sync_inodes_sb(struct super_block *sb)
1290 .nr_pages = LONG_MAX, 1319 .nr_pages = LONG_MAX,
1291 .range_cyclic = 0, 1320 .range_cyclic = 0,
1292 .done = &done, 1321 .done = &done,
1322 .reason = WB_REASON_SYNC,
1293 }; 1323 };
1294 1324
1295 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 1325 WARN_ON(!rwsem_is_locked(&sb->s_umount));
diff --git a/fs/quota/quota.c b/fs/quota/quota.c
index aae0edb95c6c..35f4b0ecdeb3 100644
--- a/fs/quota/quota.c
+++ b/fs/quota/quota.c
@@ -286,7 +286,7 @@ static int do_quotactl(struct super_block *sb, int type, int cmd, qid_t id,
286 /* caller already holds s_umount */ 286 /* caller already holds s_umount */
287 if (sb->s_flags & MS_RDONLY) 287 if (sb->s_flags & MS_RDONLY)
288 return -EROFS; 288 return -EROFS;
289 writeback_inodes_sb(sb); 289 writeback_inodes_sb(sb, WB_REASON_SYNC);
290 return 0; 290 return 0;
291 default: 291 default:
292 return -EINVAL; 292 return -EINVAL;
diff --git a/fs/sync.c b/fs/sync.c
index c98a7477edfd..101b8ef901d7 100644
--- a/fs/sync.c
+++ b/fs/sync.c
@@ -43,7 +43,7 @@ static int __sync_filesystem(struct super_block *sb, int wait)
43 if (wait) 43 if (wait)
44 sync_inodes_sb(sb); 44 sync_inodes_sb(sb);
45 else 45 else
46 writeback_inodes_sb(sb); 46 writeback_inodes_sb(sb, WB_REASON_SYNC);
47 47
48 if (sb->s_op->sync_fs) 48 if (sb->s_op->sync_fs)
49 sb->s_op->sync_fs(sb, wait); 49 sb->s_op->sync_fs(sb, wait);
@@ -98,7 +98,7 @@ static void sync_filesystems(int wait)
98 */ 98 */
99SYSCALL_DEFINE0(sync) 99SYSCALL_DEFINE0(sync)
100{ 100{
101 wakeup_flusher_threads(0); 101 wakeup_flusher_threads(0, WB_REASON_SYNC);
102 sync_filesystems(0); 102 sync_filesystems(0);
103 sync_filesystems(1); 103 sync_filesystems(1);
104 if (unlikely(laptop_mode)) 104 if (unlikely(laptop_mode))
diff --git a/fs/ubifs/budget.c b/fs/ubifs/budget.c
index 315de66e52b2..bc4f94b28706 100644
--- a/fs/ubifs/budget.c
+++ b/fs/ubifs/budget.c
@@ -63,7 +63,7 @@
63static void shrink_liability(struct ubifs_info *c, int nr_to_write) 63static void shrink_liability(struct ubifs_info *c, int nr_to_write)
64{ 64{
65 down_read(&c->vfs_sb->s_umount); 65 down_read(&c->vfs_sb->s_umount);
66 writeback_inodes_sb(c->vfs_sb); 66 writeback_inodes_sb(c->vfs_sb, WB_REASON_FS_FREE_SPACE);
67 up_read(&c->vfs_sb->s_umount); 67 up_read(&c->vfs_sb->s_umount);
68} 68}
69 69
diff --git a/include/linux/backing-dev.h b/include/linux/backing-dev.h
index 3b2f9cb82986..b1038bd686ac 100644
--- a/include/linux/backing-dev.h
+++ b/include/linux/backing-dev.h
@@ -40,6 +40,7 @@ typedef int (congested_fn)(void *, int);
40enum bdi_stat_item { 40enum bdi_stat_item {
41 BDI_RECLAIMABLE, 41 BDI_RECLAIMABLE,
42 BDI_WRITEBACK, 42 BDI_WRITEBACK,
43 BDI_DIRTIED,
43 BDI_WRITTEN, 44 BDI_WRITTEN,
44 NR_BDI_STAT_ITEMS 45 NR_BDI_STAT_ITEMS
45}; 46};
@@ -74,10 +75,20 @@ struct backing_dev_info {
74 struct percpu_counter bdi_stat[NR_BDI_STAT_ITEMS]; 75 struct percpu_counter bdi_stat[NR_BDI_STAT_ITEMS];
75 76
76 unsigned long bw_time_stamp; /* last time write bw is updated */ 77 unsigned long bw_time_stamp; /* last time write bw is updated */
78 unsigned long dirtied_stamp;
77 unsigned long written_stamp; /* pages written at bw_time_stamp */ 79 unsigned long written_stamp; /* pages written at bw_time_stamp */
78 unsigned long write_bandwidth; /* the estimated write bandwidth */ 80 unsigned long write_bandwidth; /* the estimated write bandwidth */
79 unsigned long avg_write_bandwidth; /* further smoothed write bw */ 81 unsigned long avg_write_bandwidth; /* further smoothed write bw */
80 82
83 /*
84 * The base dirty throttle rate, re-calculated on every 200ms.
85 * All the bdi tasks' dirty rate will be curbed under it.
86 * @dirty_ratelimit tracks the estimated @balanced_dirty_ratelimit
87 * in small steps and is much more smooth/stable than the latter.
88 */
89 unsigned long dirty_ratelimit;
90 unsigned long balanced_dirty_ratelimit;
91
81 struct prop_local_percpu completions; 92 struct prop_local_percpu completions;
82 int dirty_exceeded; 93 int dirty_exceeded;
83 94
@@ -107,7 +118,8 @@ int bdi_register(struct backing_dev_info *bdi, struct device *parent,
107int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev); 118int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev);
108void bdi_unregister(struct backing_dev_info *bdi); 119void bdi_unregister(struct backing_dev_info *bdi);
109int bdi_setup_and_register(struct backing_dev_info *, char *, unsigned int); 120int bdi_setup_and_register(struct backing_dev_info *, char *, unsigned int);
110void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages); 121void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
122 enum wb_reason reason);
111void bdi_start_background_writeback(struct backing_dev_info *bdi); 123void bdi_start_background_writeback(struct backing_dev_info *bdi);
112int bdi_writeback_thread(void *data); 124int bdi_writeback_thread(void *data);
113int bdi_has_dirty_io(struct backing_dev_info *bdi); 125int bdi_has_dirty_io(struct backing_dev_info *bdi);
diff --git a/include/linux/sched.h b/include/linux/sched.h
index e8acce717d2a..68daf4f27e2c 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1522,6 +1522,13 @@ struct task_struct {
1522 int make_it_fail; 1522 int make_it_fail;
1523#endif 1523#endif
1524 struct prop_local_single dirties; 1524 struct prop_local_single dirties;
1525 /*
1526 * when (nr_dirtied >= nr_dirtied_pause), it's time to call
1527 * balance_dirty_pages() for some dirty throttling pause
1528 */
1529 int nr_dirtied;
1530 int nr_dirtied_pause;
1531
1525#ifdef CONFIG_LATENCYTOP 1532#ifdef CONFIG_LATENCYTOP
1526 int latency_record_count; 1533 int latency_record_count;
1527 struct latency_record latency_record[LT_SAVECOUNT]; 1534 struct latency_record latency_record[LT_SAVECOUNT];
diff --git a/include/linux/writeback.h b/include/linux/writeback.h
index 2b8963ff0f35..a378c295851f 100644
--- a/include/linux/writeback.h
+++ b/include/linux/writeback.h
@@ -39,6 +39,23 @@ enum writeback_sync_modes {
39}; 39};
40 40
41/* 41/*
42 * why some writeback work was initiated
43 */
44enum wb_reason {
45 WB_REASON_BACKGROUND,
46 WB_REASON_TRY_TO_FREE_PAGES,
47 WB_REASON_SYNC,
48 WB_REASON_PERIODIC,
49 WB_REASON_LAPTOP_TIMER,
50 WB_REASON_FREE_MORE_MEM,
51 WB_REASON_FS_FREE_SPACE,
52 WB_REASON_FORKER_THREAD,
53
54 WB_REASON_MAX,
55};
56extern const char *wb_reason_name[];
57
58/*
42 * A control structure which tells the writeback code what to do. These are 59 * A control structure which tells the writeback code what to do. These are
43 * always on the stack, and hence need no locking. They are always initialised 60 * always on the stack, and hence need no locking. They are always initialised
44 * in a manner such that unspecified fields are set to zero. 61 * in a manner such that unspecified fields are set to zero.
@@ -69,14 +86,17 @@ struct writeback_control {
69 */ 86 */
70struct bdi_writeback; 87struct bdi_writeback;
71int inode_wait(void *); 88int inode_wait(void *);
72void writeback_inodes_sb(struct super_block *); 89void writeback_inodes_sb(struct super_block *, enum wb_reason reason);
73void writeback_inodes_sb_nr(struct super_block *, unsigned long nr); 90void writeback_inodes_sb_nr(struct super_block *, unsigned long nr,
74int writeback_inodes_sb_if_idle(struct super_block *); 91 enum wb_reason reason);
75int writeback_inodes_sb_nr_if_idle(struct super_block *, unsigned long nr); 92int writeback_inodes_sb_if_idle(struct super_block *, enum wb_reason reason);
93int writeback_inodes_sb_nr_if_idle(struct super_block *, unsigned long nr,
94 enum wb_reason reason);
76void sync_inodes_sb(struct super_block *); 95void sync_inodes_sb(struct super_block *);
77long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages); 96long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
97 enum wb_reason reason);
78long wb_do_writeback(struct bdi_writeback *wb, int force_wait); 98long wb_do_writeback(struct bdi_writeback *wb, int force_wait);
79void wakeup_flusher_threads(long nr_pages); 99void wakeup_flusher_threads(long nr_pages, enum wb_reason reason);
80 100
81/* writeback.h requires fs.h; it, too, is not included from here. */ 101/* writeback.h requires fs.h; it, too, is not included from here. */
82static inline void wait_on_inode(struct inode *inode) 102static inline void wait_on_inode(struct inode *inode)
@@ -143,6 +163,7 @@ unsigned long bdi_dirty_limit(struct backing_dev_info *bdi,
143 163
144void __bdi_update_bandwidth(struct backing_dev_info *bdi, 164void __bdi_update_bandwidth(struct backing_dev_info *bdi,
145 unsigned long thresh, 165 unsigned long thresh,
166 unsigned long bg_thresh,
146 unsigned long dirty, 167 unsigned long dirty,
147 unsigned long bdi_thresh, 168 unsigned long bdi_thresh,
148 unsigned long bdi_dirty, 169 unsigned long bdi_dirty,
diff --git a/include/trace/events/writeback.h b/include/trace/events/writeback.h
index 5f172703eb4f..b99caa8b780c 100644
--- a/include/trace/events/writeback.h
+++ b/include/trace/events/writeback.h
@@ -34,6 +34,7 @@ DECLARE_EVENT_CLASS(writeback_work_class,
34 __field(int, for_kupdate) 34 __field(int, for_kupdate)
35 __field(int, range_cyclic) 35 __field(int, range_cyclic)
36 __field(int, for_background) 36 __field(int, for_background)
37 __field(int, reason)
37 ), 38 ),
38 TP_fast_assign( 39 TP_fast_assign(
39 strncpy(__entry->name, dev_name(bdi->dev), 32); 40 strncpy(__entry->name, dev_name(bdi->dev), 32);
@@ -43,16 +44,18 @@ DECLARE_EVENT_CLASS(writeback_work_class,
43 __entry->for_kupdate = work->for_kupdate; 44 __entry->for_kupdate = work->for_kupdate;
44 __entry->range_cyclic = work->range_cyclic; 45 __entry->range_cyclic = work->range_cyclic;
45 __entry->for_background = work->for_background; 46 __entry->for_background = work->for_background;
47 __entry->reason = work->reason;
46 ), 48 ),
47 TP_printk("bdi %s: sb_dev %d:%d nr_pages=%ld sync_mode=%d " 49 TP_printk("bdi %s: sb_dev %d:%d nr_pages=%ld sync_mode=%d "
48 "kupdate=%d range_cyclic=%d background=%d", 50 "kupdate=%d range_cyclic=%d background=%d reason=%s",
49 __entry->name, 51 __entry->name,
50 MAJOR(__entry->sb_dev), MINOR(__entry->sb_dev), 52 MAJOR(__entry->sb_dev), MINOR(__entry->sb_dev),
51 __entry->nr_pages, 53 __entry->nr_pages,
52 __entry->sync_mode, 54 __entry->sync_mode,
53 __entry->for_kupdate, 55 __entry->for_kupdate,
54 __entry->range_cyclic, 56 __entry->range_cyclic,
55 __entry->for_background 57 __entry->for_background,
58 wb_reason_name[__entry->reason]
56 ) 59 )
57); 60);
58#define DEFINE_WRITEBACK_WORK_EVENT(name) \ 61#define DEFINE_WRITEBACK_WORK_EVENT(name) \
@@ -104,30 +107,6 @@ DEFINE_WRITEBACK_EVENT(writeback_bdi_register);
104DEFINE_WRITEBACK_EVENT(writeback_bdi_unregister); 107DEFINE_WRITEBACK_EVENT(writeback_bdi_unregister);
105DEFINE_WRITEBACK_EVENT(writeback_thread_start); 108DEFINE_WRITEBACK_EVENT(writeback_thread_start);
106DEFINE_WRITEBACK_EVENT(writeback_thread_stop); 109DEFINE_WRITEBACK_EVENT(writeback_thread_stop);
107DEFINE_WRITEBACK_EVENT(balance_dirty_start);
108DEFINE_WRITEBACK_EVENT(balance_dirty_wait);
109
110TRACE_EVENT(balance_dirty_written,
111
112 TP_PROTO(struct backing_dev_info *bdi, int written),
113
114 TP_ARGS(bdi, written),
115
116 TP_STRUCT__entry(
117 __array(char, name, 32)
118 __field(int, written)
119 ),
120
121 TP_fast_assign(
122 strncpy(__entry->name, dev_name(bdi->dev), 32);
123 __entry->written = written;
124 ),
125
126 TP_printk("bdi %s written %d",
127 __entry->name,
128 __entry->written
129 )
130);
131 110
132DECLARE_EVENT_CLASS(wbc_class, 111DECLARE_EVENT_CLASS(wbc_class,
133 TP_PROTO(struct writeback_control *wbc, struct backing_dev_info *bdi), 112 TP_PROTO(struct writeback_control *wbc, struct backing_dev_info *bdi),
@@ -181,27 +160,31 @@ DEFINE_WBC_EVENT(wbc_writepage);
181 160
182TRACE_EVENT(writeback_queue_io, 161TRACE_EVENT(writeback_queue_io,
183 TP_PROTO(struct bdi_writeback *wb, 162 TP_PROTO(struct bdi_writeback *wb,
184 unsigned long *older_than_this, 163 struct wb_writeback_work *work,
185 int moved), 164 int moved),
186 TP_ARGS(wb, older_than_this, moved), 165 TP_ARGS(wb, work, moved),
187 TP_STRUCT__entry( 166 TP_STRUCT__entry(
188 __array(char, name, 32) 167 __array(char, name, 32)
189 __field(unsigned long, older) 168 __field(unsigned long, older)
190 __field(long, age) 169 __field(long, age)
191 __field(int, moved) 170 __field(int, moved)
171 __field(int, reason)
192 ), 172 ),
193 TP_fast_assign( 173 TP_fast_assign(
174 unsigned long *older_than_this = work->older_than_this;
194 strncpy(__entry->name, dev_name(wb->bdi->dev), 32); 175 strncpy(__entry->name, dev_name(wb->bdi->dev), 32);
195 __entry->older = older_than_this ? *older_than_this : 0; 176 __entry->older = older_than_this ? *older_than_this : 0;
196 __entry->age = older_than_this ? 177 __entry->age = older_than_this ?
197 (jiffies - *older_than_this) * 1000 / HZ : -1; 178 (jiffies - *older_than_this) * 1000 / HZ : -1;
198 __entry->moved = moved; 179 __entry->moved = moved;
180 __entry->reason = work->reason;
199 ), 181 ),
200 TP_printk("bdi %s: older=%lu age=%ld enqueue=%d", 182 TP_printk("bdi %s: older=%lu age=%ld enqueue=%d reason=%s",
201 __entry->name, 183 __entry->name,
202 __entry->older, /* older_than_this in jiffies */ 184 __entry->older, /* older_than_this in jiffies */
203 __entry->age, /* older_than_this in relative milliseconds */ 185 __entry->age, /* older_than_this in relative milliseconds */
204 __entry->moved) 186 __entry->moved,
187 wb_reason_name[__entry->reason])
205); 188);
206 189
207TRACE_EVENT(global_dirty_state, 190TRACE_EVENT(global_dirty_state,
@@ -250,6 +233,124 @@ TRACE_EVENT(global_dirty_state,
250 ) 233 )
251); 234);
252 235
236#define KBps(x) ((x) << (PAGE_SHIFT - 10))
237
238TRACE_EVENT(bdi_dirty_ratelimit,
239
240 TP_PROTO(struct backing_dev_info *bdi,
241 unsigned long dirty_rate,
242 unsigned long task_ratelimit),
243
244 TP_ARGS(bdi, dirty_rate, task_ratelimit),
245
246 TP_STRUCT__entry(
247 __array(char, bdi, 32)
248 __field(unsigned long, write_bw)
249 __field(unsigned long, avg_write_bw)
250 __field(unsigned long, dirty_rate)
251 __field(unsigned long, dirty_ratelimit)
252 __field(unsigned long, task_ratelimit)
253 __field(unsigned long, balanced_dirty_ratelimit)
254 ),
255
256 TP_fast_assign(
257 strlcpy(__entry->bdi, dev_name(bdi->dev), 32);
258 __entry->write_bw = KBps(bdi->write_bandwidth);
259 __entry->avg_write_bw = KBps(bdi->avg_write_bandwidth);
260 __entry->dirty_rate = KBps(dirty_rate);
261 __entry->dirty_ratelimit = KBps(bdi->dirty_ratelimit);
262 __entry->task_ratelimit = KBps(task_ratelimit);
263 __entry->balanced_dirty_ratelimit =
264 KBps(bdi->balanced_dirty_ratelimit);
265 ),
266
267 TP_printk("bdi %s: "
268 "write_bw=%lu awrite_bw=%lu dirty_rate=%lu "
269 "dirty_ratelimit=%lu task_ratelimit=%lu "
270 "balanced_dirty_ratelimit=%lu",
271 __entry->bdi,
272 __entry->write_bw, /* write bandwidth */
273 __entry->avg_write_bw, /* avg write bandwidth */
274 __entry->dirty_rate, /* bdi dirty rate */
275 __entry->dirty_ratelimit, /* base ratelimit */
276 __entry->task_ratelimit, /* ratelimit with position control */
277 __entry->balanced_dirty_ratelimit /* the balanced ratelimit */
278 )
279);
280
281TRACE_EVENT(balance_dirty_pages,
282
283 TP_PROTO(struct backing_dev_info *bdi,
284 unsigned long thresh,
285 unsigned long bg_thresh,
286 unsigned long dirty,
287 unsigned long bdi_thresh,
288 unsigned long bdi_dirty,
289 unsigned long dirty_ratelimit,
290 unsigned long task_ratelimit,
291 unsigned long dirtied,
292 long pause,
293 unsigned long start_time),
294
295 TP_ARGS(bdi, thresh, bg_thresh, dirty, bdi_thresh, bdi_dirty,
296 dirty_ratelimit, task_ratelimit,
297 dirtied, pause, start_time),
298
299 TP_STRUCT__entry(
300 __array( char, bdi, 32)
301 __field(unsigned long, limit)
302 __field(unsigned long, setpoint)
303 __field(unsigned long, dirty)
304 __field(unsigned long, bdi_setpoint)
305 __field(unsigned long, bdi_dirty)
306 __field(unsigned long, dirty_ratelimit)
307 __field(unsigned long, task_ratelimit)
308 __field(unsigned int, dirtied)
309 __field(unsigned int, dirtied_pause)
310 __field(unsigned long, paused)
311 __field( long, pause)
312 ),
313
314 TP_fast_assign(
315 unsigned long freerun = (thresh + bg_thresh) / 2;
316 strlcpy(__entry->bdi, dev_name(bdi->dev), 32);
317
318 __entry->limit = global_dirty_limit;
319 __entry->setpoint = (global_dirty_limit + freerun) / 2;
320 __entry->dirty = dirty;
321 __entry->bdi_setpoint = __entry->setpoint *
322 bdi_thresh / (thresh + 1);
323 __entry->bdi_dirty = bdi_dirty;
324 __entry->dirty_ratelimit = KBps(dirty_ratelimit);
325 __entry->task_ratelimit = KBps(task_ratelimit);
326 __entry->dirtied = dirtied;
327 __entry->dirtied_pause = current->nr_dirtied_pause;
328 __entry->pause = pause * 1000 / HZ;
329 __entry->paused = (jiffies - start_time) * 1000 / HZ;
330 ),
331
332
333 TP_printk("bdi %s: "
334 "limit=%lu setpoint=%lu dirty=%lu "
335 "bdi_setpoint=%lu bdi_dirty=%lu "
336 "dirty_ratelimit=%lu task_ratelimit=%lu "
337 "dirtied=%u dirtied_pause=%u "
338 "paused=%lu pause=%ld",
339 __entry->bdi,
340 __entry->limit,
341 __entry->setpoint,
342 __entry->dirty,
343 __entry->bdi_setpoint,
344 __entry->bdi_dirty,
345 __entry->dirty_ratelimit,
346 __entry->task_ratelimit,
347 __entry->dirtied,
348 __entry->dirtied_pause,
349 __entry->paused, /* ms */
350 __entry->pause /* ms */
351 )
352);
353
253DECLARE_EVENT_CLASS(writeback_congest_waited_template, 354DECLARE_EVENT_CLASS(writeback_congest_waited_template,
254 355
255 TP_PROTO(unsigned int usec_timeout, unsigned int usec_delayed), 356 TP_PROTO(unsigned int usec_timeout, unsigned int usec_delayed),
diff --git a/kernel/fork.c b/kernel/fork.c
index 70d76191afb9..ba0d17261329 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -1299,6 +1299,9 @@ static struct task_struct *copy_process(unsigned long clone_flags,
1299 p->pdeath_signal = 0; 1299 p->pdeath_signal = 0;
1300 p->exit_state = 0; 1300 p->exit_state = 0;
1301 1301
1302 p->nr_dirtied = 0;
1303 p->nr_dirtied_pause = 128 >> (PAGE_SHIFT - 10);
1304
1302 /* 1305 /*
1303 * Ok, make it visible to the rest of the system. 1306 * Ok, make it visible to the rest of the system.
1304 * We dont wake it up yet. 1307 * We dont wake it up yet.
diff --git a/mm/backing-dev.c b/mm/backing-dev.c
index 7520ef0bfd47..a0860640378d 100644
--- a/mm/backing-dev.c
+++ b/mm/backing-dev.c
@@ -97,6 +97,7 @@ static int bdi_debug_stats_show(struct seq_file *m, void *v)
97 "BdiDirtyThresh: %10lu kB\n" 97 "BdiDirtyThresh: %10lu kB\n"
98 "DirtyThresh: %10lu kB\n" 98 "DirtyThresh: %10lu kB\n"
99 "BackgroundThresh: %10lu kB\n" 99 "BackgroundThresh: %10lu kB\n"
100 "BdiDirtied: %10lu kB\n"
100 "BdiWritten: %10lu kB\n" 101 "BdiWritten: %10lu kB\n"
101 "BdiWriteBandwidth: %10lu kBps\n" 102 "BdiWriteBandwidth: %10lu kBps\n"
102 "b_dirty: %10lu\n" 103 "b_dirty: %10lu\n"
@@ -109,6 +110,7 @@ static int bdi_debug_stats_show(struct seq_file *m, void *v)
109 K(bdi_thresh), 110 K(bdi_thresh),
110 K(dirty_thresh), 111 K(dirty_thresh),
111 K(background_thresh), 112 K(background_thresh),
113 (unsigned long) K(bdi_stat(bdi, BDI_DIRTIED)),
112 (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)), 114 (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
113 (unsigned long) K(bdi->write_bandwidth), 115 (unsigned long) K(bdi->write_bandwidth),
114 nr_dirty, 116 nr_dirty,
@@ -473,7 +475,8 @@ static int bdi_forker_thread(void *ptr)
473 * the bdi from the thread. Hopefully 1024 is 475 * the bdi from the thread. Hopefully 1024 is
474 * large enough for efficient IO. 476 * large enough for efficient IO.
475 */ 477 */
476 writeback_inodes_wb(&bdi->wb, 1024); 478 writeback_inodes_wb(&bdi->wb, 1024,
479 WB_REASON_FORKER_THREAD);
477 } else { 480 } else {
478 /* 481 /*
479 * The spinlock makes sure we do not lose 482 * The spinlock makes sure we do not lose
@@ -683,6 +686,8 @@ int bdi_init(struct backing_dev_info *bdi)
683 bdi->bw_time_stamp = jiffies; 686 bdi->bw_time_stamp = jiffies;
684 bdi->written_stamp = 0; 687 bdi->written_stamp = 0;
685 688
689 bdi->balanced_dirty_ratelimit = INIT_BW;
690 bdi->dirty_ratelimit = INIT_BW;
686 bdi->write_bandwidth = INIT_BW; 691 bdi->write_bandwidth = INIT_BW;
687 bdi->avg_write_bandwidth = INIT_BW; 692 bdi->avg_write_bandwidth = INIT_BW;
688 693
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 793e9874de51..9c31199dae9a 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -46,26 +46,14 @@
46 */ 46 */
47#define BANDWIDTH_INTERVAL max(HZ/5, 1) 47#define BANDWIDTH_INTERVAL max(HZ/5, 1)
48 48
49#define RATELIMIT_CALC_SHIFT 10
50
49/* 51/*
50 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited 52 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
51 * will look to see if it needs to force writeback or throttling. 53 * will look to see if it needs to force writeback or throttling.
52 */ 54 */
53static long ratelimit_pages = 32; 55static long ratelimit_pages = 32;
54 56
55/*
56 * When balance_dirty_pages decides that the caller needs to perform some
57 * non-background writeback, this is how many pages it will attempt to write.
58 * It should be somewhat larger than dirtied pages to ensure that reasonably
59 * large amounts of I/O are submitted.
60 */
61static inline long sync_writeback_pages(unsigned long dirtied)
62{
63 if (dirtied < ratelimit_pages)
64 dirtied = ratelimit_pages;
65
66 return dirtied + dirtied / 2;
67}
68
69/* The following parameters are exported via /proc/sys/vm */ 57/* The following parameters are exported via /proc/sys/vm */
70 58
71/* 59/*
@@ -167,6 +155,8 @@ static void update_completion_period(void)
167 int shift = calc_period_shift(); 155 int shift = calc_period_shift();
168 prop_change_shift(&vm_completions, shift); 156 prop_change_shift(&vm_completions, shift);
169 prop_change_shift(&vm_dirties, shift); 157 prop_change_shift(&vm_dirties, shift);
158
159 writeback_set_ratelimit();
170} 160}
171 161
172int dirty_background_ratio_handler(struct ctl_table *table, int write, 162int dirty_background_ratio_handler(struct ctl_table *table, int write,
@@ -260,50 +250,6 @@ static void bdi_writeout_fraction(struct backing_dev_info *bdi,
260 numerator, denominator); 250 numerator, denominator);
261} 251}
262 252
263static inline void task_dirties_fraction(struct task_struct *tsk,
264 long *numerator, long *denominator)
265{
266 prop_fraction_single(&vm_dirties, &tsk->dirties,
267 numerator, denominator);
268}
269
270/*
271 * task_dirty_limit - scale down dirty throttling threshold for one task
272 *
273 * task specific dirty limit:
274 *
275 * dirty -= (dirty/8) * p_{t}
276 *
277 * To protect light/slow dirtying tasks from heavier/fast ones, we start
278 * throttling individual tasks before reaching the bdi dirty limit.
279 * Relatively low thresholds will be allocated to heavy dirtiers. So when
280 * dirty pages grow large, heavy dirtiers will be throttled first, which will
281 * effectively curb the growth of dirty pages. Light dirtiers with high enough
282 * dirty threshold may never get throttled.
283 */
284#define TASK_LIMIT_FRACTION 8
285static unsigned long task_dirty_limit(struct task_struct *tsk,
286 unsigned long bdi_dirty)
287{
288 long numerator, denominator;
289 unsigned long dirty = bdi_dirty;
290 u64 inv = dirty / TASK_LIMIT_FRACTION;
291
292 task_dirties_fraction(tsk, &numerator, &denominator);
293 inv *= numerator;
294 do_div(inv, denominator);
295
296 dirty -= inv;
297
298 return max(dirty, bdi_dirty/2);
299}
300
301/* Minimum limit for any task */
302static unsigned long task_min_dirty_limit(unsigned long bdi_dirty)
303{
304 return bdi_dirty - bdi_dirty / TASK_LIMIT_FRACTION;
305}
306
307/* 253/*
308 * bdi_min_ratio keeps the sum of the minimum dirty shares of all 254 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
309 * registered backing devices, which, for obvious reasons, can not 255 * registered backing devices, which, for obvious reasons, can not
@@ -413,6 +359,12 @@ unsigned long determine_dirtyable_memory(void)
413 return x + 1; /* Ensure that we never return 0 */ 359 return x + 1; /* Ensure that we never return 0 */
414} 360}
415 361
362static unsigned long dirty_freerun_ceiling(unsigned long thresh,
363 unsigned long bg_thresh)
364{
365 return (thresh + bg_thresh) / 2;
366}
367
416static unsigned long hard_dirty_limit(unsigned long thresh) 368static unsigned long hard_dirty_limit(unsigned long thresh)
417{ 369{
418 return max(thresh, global_dirty_limit); 370 return max(thresh, global_dirty_limit);
@@ -497,6 +449,198 @@ unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
497 return bdi_dirty; 449 return bdi_dirty;
498} 450}
499 451
452/*
453 * Dirty position control.
454 *
455 * (o) global/bdi setpoints
456 *
457 * We want the dirty pages be balanced around the global/bdi setpoints.
458 * When the number of dirty pages is higher/lower than the setpoint, the
459 * dirty position control ratio (and hence task dirty ratelimit) will be
460 * decreased/increased to bring the dirty pages back to the setpoint.
461 *
462 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
463 *
464 * if (dirty < setpoint) scale up pos_ratio
465 * if (dirty > setpoint) scale down pos_ratio
466 *
467 * if (bdi_dirty < bdi_setpoint) scale up pos_ratio
468 * if (bdi_dirty > bdi_setpoint) scale down pos_ratio
469 *
470 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
471 *
472 * (o) global control line
473 *
474 * ^ pos_ratio
475 * |
476 * | |<===== global dirty control scope ======>|
477 * 2.0 .............*
478 * | .*
479 * | . *
480 * | . *
481 * | . *
482 * | . *
483 * | . *
484 * 1.0 ................................*
485 * | . . *
486 * | . . *
487 * | . . *
488 * | . . *
489 * | . . *
490 * 0 +------------.------------------.----------------------*------------->
491 * freerun^ setpoint^ limit^ dirty pages
492 *
493 * (o) bdi control line
494 *
495 * ^ pos_ratio
496 * |
497 * | *
498 * | *
499 * | *
500 * | *
501 * | * |<=========== span ============>|
502 * 1.0 .......................*
503 * | . *
504 * | . *
505 * | . *
506 * | . *
507 * | . *
508 * | . *
509 * | . *
510 * | . *
511 * | . *
512 * | . *
513 * | . *
514 * 1/4 ...............................................* * * * * * * * * * * *
515 * | . .
516 * | . .
517 * | . .
518 * 0 +----------------------.-------------------------------.------------->
519 * bdi_setpoint^ x_intercept^
520 *
521 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
522 * be smoothly throttled down to normal if it starts high in situations like
523 * - start writing to a slow SD card and a fast disk at the same time. The SD
524 * card's bdi_dirty may rush to many times higher than bdi_setpoint.
525 * - the bdi dirty thresh drops quickly due to change of JBOD workload
526 */
527static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
528 unsigned long thresh,
529 unsigned long bg_thresh,
530 unsigned long dirty,
531 unsigned long bdi_thresh,
532 unsigned long bdi_dirty)
533{
534 unsigned long write_bw = bdi->avg_write_bandwidth;
535 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
536 unsigned long limit = hard_dirty_limit(thresh);
537 unsigned long x_intercept;
538 unsigned long setpoint; /* dirty pages' target balance point */
539 unsigned long bdi_setpoint;
540 unsigned long span;
541 long long pos_ratio; /* for scaling up/down the rate limit */
542 long x;
543
544 if (unlikely(dirty >= limit))
545 return 0;
546
547 /*
548 * global setpoint
549 *
550 * setpoint - dirty 3
551 * f(dirty) := 1.0 + (----------------)
552 * limit - setpoint
553 *
554 * it's a 3rd order polynomial that subjects to
555 *
556 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
557 * (2) f(setpoint) = 1.0 => the balance point
558 * (3) f(limit) = 0 => the hard limit
559 * (4) df/dx <= 0 => negative feedback control
560 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
561 * => fast response on large errors; small oscillation near setpoint
562 */
563 setpoint = (freerun + limit) / 2;
564 x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
565 limit - setpoint + 1);
566 pos_ratio = x;
567 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
568 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
569 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
570
571 /*
572 * We have computed basic pos_ratio above based on global situation. If
573 * the bdi is over/under its share of dirty pages, we want to scale
574 * pos_ratio further down/up. That is done by the following mechanism.
575 */
576
577 /*
578 * bdi setpoint
579 *
580 * f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
581 *
582 * x_intercept - bdi_dirty
583 * := --------------------------
584 * x_intercept - bdi_setpoint
585 *
586 * The main bdi control line is a linear function that subjects to
587 *
588 * (1) f(bdi_setpoint) = 1.0
589 * (2) k = - 1 / (8 * write_bw) (in single bdi case)
590 * or equally: x_intercept = bdi_setpoint + 8 * write_bw
591 *
592 * For single bdi case, the dirty pages are observed to fluctuate
593 * regularly within range
594 * [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
595 * for various filesystems, where (2) can yield in a reasonable 12.5%
596 * fluctuation range for pos_ratio.
597 *
598 * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
599 * own size, so move the slope over accordingly and choose a slope that
600 * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
601 */
602 if (unlikely(bdi_thresh > thresh))
603 bdi_thresh = thresh;
604 bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
605 /*
606 * scale global setpoint to bdi's:
607 * bdi_setpoint = setpoint * bdi_thresh / thresh
608 */
609 x = div_u64((u64)bdi_thresh << 16, thresh + 1);
610 bdi_setpoint = setpoint * (u64)x >> 16;
611 /*
612 * Use span=(8*write_bw) in single bdi case as indicated by
613 * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
614 *
615 * bdi_thresh thresh - bdi_thresh
616 * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
617 * thresh thresh
618 */
619 span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
620 x_intercept = bdi_setpoint + span;
621
622 if (bdi_dirty < x_intercept - span / 4) {
623 pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
624 x_intercept - bdi_setpoint + 1);
625 } else
626 pos_ratio /= 4;
627
628 /*
629 * bdi reserve area, safeguard against dirty pool underrun and disk idle
630 * It may push the desired control point of global dirty pages higher
631 * than setpoint.
632 */
633 x_intercept = bdi_thresh / 2;
634 if (bdi_dirty < x_intercept) {
635 if (bdi_dirty > x_intercept / 8)
636 pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
637 else
638 pos_ratio *= 8;
639 }
640
641 return pos_ratio;
642}
643
500static void bdi_update_write_bandwidth(struct backing_dev_info *bdi, 644static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
501 unsigned long elapsed, 645 unsigned long elapsed,
502 unsigned long written) 646 unsigned long written)
@@ -593,8 +737,153 @@ static void global_update_bandwidth(unsigned long thresh,
593 spin_unlock(&dirty_lock); 737 spin_unlock(&dirty_lock);
594} 738}
595 739
740/*
741 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
742 *
743 * Normal bdi tasks will be curbed at or below it in long term.
744 * Obviously it should be around (write_bw / N) when there are N dd tasks.
745 */
746static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
747 unsigned long thresh,
748 unsigned long bg_thresh,
749 unsigned long dirty,
750 unsigned long bdi_thresh,
751 unsigned long bdi_dirty,
752 unsigned long dirtied,
753 unsigned long elapsed)
754{
755 unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
756 unsigned long limit = hard_dirty_limit(thresh);
757 unsigned long setpoint = (freerun + limit) / 2;
758 unsigned long write_bw = bdi->avg_write_bandwidth;
759 unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
760 unsigned long dirty_rate;
761 unsigned long task_ratelimit;
762 unsigned long balanced_dirty_ratelimit;
763 unsigned long pos_ratio;
764 unsigned long step;
765 unsigned long x;
766
767 /*
768 * The dirty rate will match the writeout rate in long term, except
769 * when dirty pages are truncated by userspace or re-dirtied by FS.
770 */
771 dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
772
773 pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
774 bdi_thresh, bdi_dirty);
775 /*
776 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
777 */
778 task_ratelimit = (u64)dirty_ratelimit *
779 pos_ratio >> RATELIMIT_CALC_SHIFT;
780 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
781
782 /*
783 * A linear estimation of the "balanced" throttle rate. The theory is,
784 * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
785 * dirty_rate will be measured to be (N * task_ratelimit). So the below
786 * formula will yield the balanced rate limit (write_bw / N).
787 *
788 * Note that the expanded form is not a pure rate feedback:
789 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
790 * but also takes pos_ratio into account:
791 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
792 *
793 * (1) is not realistic because pos_ratio also takes part in balancing
794 * the dirty rate. Consider the state
795 * pos_ratio = 0.5 (3)
796 * rate = 2 * (write_bw / N) (4)
797 * If (1) is used, it will stuck in that state! Because each dd will
798 * be throttled at
799 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
800 * yielding
801 * dirty_rate = N * task_ratelimit = write_bw (6)
802 * put (6) into (1) we get
803 * rate_(i+1) = rate_(i) (7)
804 *
805 * So we end up using (2) to always keep
806 * rate_(i+1) ~= (write_bw / N) (8)
807 * regardless of the value of pos_ratio. As long as (8) is satisfied,
808 * pos_ratio is able to drive itself to 1.0, which is not only where
809 * the dirty count meet the setpoint, but also where the slope of
810 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
811 */
812 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
813 dirty_rate | 1);
814
815 /*
816 * We could safely do this and return immediately:
817 *
818 * bdi->dirty_ratelimit = balanced_dirty_ratelimit;
819 *
820 * However to get a more stable dirty_ratelimit, the below elaborated
821 * code makes use of task_ratelimit to filter out sigular points and
822 * limit the step size.
823 *
824 * The below code essentially only uses the relative value of
825 *
826 * task_ratelimit - dirty_ratelimit
827 * = (pos_ratio - 1) * dirty_ratelimit
828 *
829 * which reflects the direction and size of dirty position error.
830 */
831
832 /*
833 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
834 * task_ratelimit is on the same side of dirty_ratelimit, too.
835 * For example, when
836 * - dirty_ratelimit > balanced_dirty_ratelimit
837 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
838 * lowering dirty_ratelimit will help meet both the position and rate
839 * control targets. Otherwise, don't update dirty_ratelimit if it will
840 * only help meet the rate target. After all, what the users ultimately
841 * feel and care are stable dirty rate and small position error.
842 *
843 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
844 * and filter out the sigular points of balanced_dirty_ratelimit. Which
845 * keeps jumping around randomly and can even leap far away at times
846 * due to the small 200ms estimation period of dirty_rate (we want to
847 * keep that period small to reduce time lags).
848 */
849 step = 0;
850 if (dirty < setpoint) {
851 x = min(bdi->balanced_dirty_ratelimit,
852 min(balanced_dirty_ratelimit, task_ratelimit));
853 if (dirty_ratelimit < x)
854 step = x - dirty_ratelimit;
855 } else {
856 x = max(bdi->balanced_dirty_ratelimit,
857 max(balanced_dirty_ratelimit, task_ratelimit));
858 if (dirty_ratelimit > x)
859 step = dirty_ratelimit - x;
860 }
861
862 /*
863 * Don't pursue 100% rate matching. It's impossible since the balanced
864 * rate itself is constantly fluctuating. So decrease the track speed
865 * when it gets close to the target. Helps eliminate pointless tremors.
866 */
867 step >>= dirty_ratelimit / (2 * step + 1);
868 /*
869 * Limit the tracking speed to avoid overshooting.
870 */
871 step = (step + 7) / 8;
872
873 if (dirty_ratelimit < balanced_dirty_ratelimit)
874 dirty_ratelimit += step;
875 else
876 dirty_ratelimit -= step;
877
878 bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
879 bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
880
881 trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
882}
883
596void __bdi_update_bandwidth(struct backing_dev_info *bdi, 884void __bdi_update_bandwidth(struct backing_dev_info *bdi,
597 unsigned long thresh, 885 unsigned long thresh,
886 unsigned long bg_thresh,
598 unsigned long dirty, 887 unsigned long dirty,
599 unsigned long bdi_thresh, 888 unsigned long bdi_thresh,
600 unsigned long bdi_dirty, 889 unsigned long bdi_dirty,
@@ -602,6 +891,7 @@ void __bdi_update_bandwidth(struct backing_dev_info *bdi,
602{ 891{
603 unsigned long now = jiffies; 892 unsigned long now = jiffies;
604 unsigned long elapsed = now - bdi->bw_time_stamp; 893 unsigned long elapsed = now - bdi->bw_time_stamp;
894 unsigned long dirtied;
605 unsigned long written; 895 unsigned long written;
606 896
607 /* 897 /*
@@ -610,6 +900,7 @@ void __bdi_update_bandwidth(struct backing_dev_info *bdi,
610 if (elapsed < BANDWIDTH_INTERVAL) 900 if (elapsed < BANDWIDTH_INTERVAL)
611 return; 901 return;
612 902
903 dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
613 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]); 904 written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
614 905
615 /* 906 /*
@@ -619,18 +910,23 @@ void __bdi_update_bandwidth(struct backing_dev_info *bdi,
619 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time)) 910 if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
620 goto snapshot; 911 goto snapshot;
621 912
622 if (thresh) 913 if (thresh) {
623 global_update_bandwidth(thresh, dirty, now); 914 global_update_bandwidth(thresh, dirty, now);
624 915 bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
916 bdi_thresh, bdi_dirty,
917 dirtied, elapsed);
918 }
625 bdi_update_write_bandwidth(bdi, elapsed, written); 919 bdi_update_write_bandwidth(bdi, elapsed, written);
626 920
627snapshot: 921snapshot:
922 bdi->dirtied_stamp = dirtied;
628 bdi->written_stamp = written; 923 bdi->written_stamp = written;
629 bdi->bw_time_stamp = now; 924 bdi->bw_time_stamp = now;
630} 925}
631 926
632static void bdi_update_bandwidth(struct backing_dev_info *bdi, 927static void bdi_update_bandwidth(struct backing_dev_info *bdi,
633 unsigned long thresh, 928 unsigned long thresh,
929 unsigned long bg_thresh,
634 unsigned long dirty, 930 unsigned long dirty,
635 unsigned long bdi_thresh, 931 unsigned long bdi_thresh,
636 unsigned long bdi_dirty, 932 unsigned long bdi_dirty,
@@ -639,37 +935,99 @@ static void bdi_update_bandwidth(struct backing_dev_info *bdi,
639 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL)) 935 if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
640 return; 936 return;
641 spin_lock(&bdi->wb.list_lock); 937 spin_lock(&bdi->wb.list_lock);
642 __bdi_update_bandwidth(bdi, thresh, dirty, bdi_thresh, bdi_dirty, 938 __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
643 start_time); 939 bdi_thresh, bdi_dirty, start_time);
644 spin_unlock(&bdi->wb.list_lock); 940 spin_unlock(&bdi->wb.list_lock);
645} 941}
646 942
647/* 943/*
944 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
945 * will look to see if it needs to start dirty throttling.
946 *
947 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
948 * global_page_state() too often. So scale it near-sqrt to the safety margin
949 * (the number of pages we may dirty without exceeding the dirty limits).
950 */
951static unsigned long dirty_poll_interval(unsigned long dirty,
952 unsigned long thresh)
953{
954 if (thresh > dirty)
955 return 1UL << (ilog2(thresh - dirty) >> 1);
956
957 return 1;
958}
959
960static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
961 unsigned long bdi_dirty)
962{
963 unsigned long bw = bdi->avg_write_bandwidth;
964 unsigned long hi = ilog2(bw);
965 unsigned long lo = ilog2(bdi->dirty_ratelimit);
966 unsigned long t;
967
968 /* target for 20ms max pause on 1-dd case */
969 t = HZ / 50;
970
971 /*
972 * Scale up pause time for concurrent dirtiers in order to reduce CPU
973 * overheads.
974 *
975 * (N * 20ms) on 2^N concurrent tasks.
976 */
977 if (hi > lo)
978 t += (hi - lo) * (20 * HZ) / 1024;
979
980 /*
981 * Limit pause time for small memory systems. If sleeping for too long
982 * time, a small pool of dirty/writeback pages may go empty and disk go
983 * idle.
984 *
985 * 8 serves as the safety ratio.
986 */
987 if (bdi_dirty)
988 t = min(t, bdi_dirty * HZ / (8 * bw + 1));
989
990 /*
991 * The pause time will be settled within range (max_pause/4, max_pause).
992 * Apply a minimal value of 4 to get a non-zero max_pause/4.
993 */
994 return clamp_val(t, 4, MAX_PAUSE);
995}
996
997/*
648 * balance_dirty_pages() must be called by processes which are generating dirty 998 * balance_dirty_pages() must be called by processes which are generating dirty
649 * data. It looks at the number of dirty pages in the machine and will force 999 * data. It looks at the number of dirty pages in the machine and will force
650 * the caller to perform writeback if the system is over `vm_dirty_ratio'. 1000 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
651 * If we're over `background_thresh' then the writeback threads are woken to 1001 * If we're over `background_thresh' then the writeback threads are woken to
652 * perform some writeout. 1002 * perform some writeout.
653 */ 1003 */
654static void balance_dirty_pages(struct address_space *mapping, 1004static void balance_dirty_pages(struct address_space *mapping,
655 unsigned long write_chunk) 1005 unsigned long pages_dirtied)
656{ 1006{
657 unsigned long nr_reclaimable, bdi_nr_reclaimable; 1007 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
1008 unsigned long bdi_reclaimable;
658 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */ 1009 unsigned long nr_dirty; /* = file_dirty + writeback + unstable_nfs */
659 unsigned long bdi_dirty; 1010 unsigned long bdi_dirty;
1011 unsigned long freerun;
660 unsigned long background_thresh; 1012 unsigned long background_thresh;
661 unsigned long dirty_thresh; 1013 unsigned long dirty_thresh;
662 unsigned long bdi_thresh; 1014 unsigned long bdi_thresh;
663 unsigned long task_bdi_thresh; 1015 long pause = 0;
664 unsigned long min_task_bdi_thresh; 1016 long uninitialized_var(max_pause);
665 unsigned long pages_written = 0;
666 unsigned long pause = 1;
667 bool dirty_exceeded = false; 1017 bool dirty_exceeded = false;
668 bool clear_dirty_exceeded = true; 1018 unsigned long task_ratelimit;
1019 unsigned long uninitialized_var(dirty_ratelimit);
1020 unsigned long pos_ratio;
669 struct backing_dev_info *bdi = mapping->backing_dev_info; 1021 struct backing_dev_info *bdi = mapping->backing_dev_info;
670 unsigned long start_time = jiffies; 1022 unsigned long start_time = jiffies;
671 1023
672 for (;;) { 1024 for (;;) {
1025 /*
1026 * Unstable writes are a feature of certain networked
1027 * filesystems (i.e. NFS) in which data may have been
1028 * written to the server's write cache, but has not yet
1029 * been flushed to permanent storage.
1030 */
673 nr_reclaimable = global_page_state(NR_FILE_DIRTY) + 1031 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
674 global_page_state(NR_UNSTABLE_NFS); 1032 global_page_state(NR_UNSTABLE_NFS);
675 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK); 1033 nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
@@ -681,12 +1039,28 @@ static void balance_dirty_pages(struct address_space *mapping,
681 * catch-up. This avoids (excessively) small writeouts 1039 * catch-up. This avoids (excessively) small writeouts
682 * when the bdi limits are ramping up. 1040 * when the bdi limits are ramping up.
683 */ 1041 */
684 if (nr_dirty <= (background_thresh + dirty_thresh) / 2) 1042 freerun = dirty_freerun_ceiling(dirty_thresh,
1043 background_thresh);
1044 if (nr_dirty <= freerun)
685 break; 1045 break;
686 1046
1047 if (unlikely(!writeback_in_progress(bdi)))
1048 bdi_start_background_writeback(bdi);
1049
1050 /*
1051 * bdi_thresh is not treated as some limiting factor as
1052 * dirty_thresh, due to reasons
1053 * - in JBOD setup, bdi_thresh can fluctuate a lot
1054 * - in a system with HDD and USB key, the USB key may somehow
1055 * go into state (bdi_dirty >> bdi_thresh) either because
1056 * bdi_dirty starts high, or because bdi_thresh drops low.
1057 * In this case we don't want to hard throttle the USB key
1058 * dirtiers for 100 seconds until bdi_dirty drops under
1059 * bdi_thresh. Instead the auxiliary bdi control line in
1060 * bdi_position_ratio() will let the dirtier task progress
1061 * at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1062 */
687 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh); 1063 bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
688 min_task_bdi_thresh = task_min_dirty_limit(bdi_thresh);
689 task_bdi_thresh = task_dirty_limit(current, bdi_thresh);
690 1064
691 /* 1065 /*
692 * In order to avoid the stacked BDI deadlock we need 1066 * In order to avoid the stacked BDI deadlock we need
@@ -698,56 +1072,69 @@ static void balance_dirty_pages(struct address_space *mapping,
698 * actually dirty; with m+n sitting in the percpu 1072 * actually dirty; with m+n sitting in the percpu
699 * deltas. 1073 * deltas.
700 */ 1074 */
701 if (task_bdi_thresh < 2 * bdi_stat_error(bdi)) { 1075 if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
702 bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE); 1076 bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
703 bdi_dirty = bdi_nr_reclaimable + 1077 bdi_dirty = bdi_reclaimable +
704 bdi_stat_sum(bdi, BDI_WRITEBACK); 1078 bdi_stat_sum(bdi, BDI_WRITEBACK);
705 } else { 1079 } else {
706 bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE); 1080 bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
707 bdi_dirty = bdi_nr_reclaimable + 1081 bdi_dirty = bdi_reclaimable +
708 bdi_stat(bdi, BDI_WRITEBACK); 1082 bdi_stat(bdi, BDI_WRITEBACK);
709 } 1083 }
710 1084
711 /* 1085 dirty_exceeded = (bdi_dirty > bdi_thresh) ||
712 * The bdi thresh is somehow "soft" limit derived from the
713 * global "hard" limit. The former helps to prevent heavy IO
714 * bdi or process from holding back light ones; The latter is
715 * the last resort safeguard.
716 */
717 dirty_exceeded = (bdi_dirty > task_bdi_thresh) ||
718 (nr_dirty > dirty_thresh); 1086 (nr_dirty > dirty_thresh);
719 clear_dirty_exceeded = (bdi_dirty <= min_task_bdi_thresh) && 1087 if (dirty_exceeded && !bdi->dirty_exceeded)
720 (nr_dirty <= dirty_thresh);
721
722 if (!dirty_exceeded)
723 break;
724
725 if (!bdi->dirty_exceeded)
726 bdi->dirty_exceeded = 1; 1088 bdi->dirty_exceeded = 1;
727 1089
728 bdi_update_bandwidth(bdi, dirty_thresh, nr_dirty, 1090 bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
729 bdi_thresh, bdi_dirty, start_time); 1091 nr_dirty, bdi_thresh, bdi_dirty,
1092 start_time);
730 1093
731 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable. 1094 max_pause = bdi_max_pause(bdi, bdi_dirty);
732 * Unstable writes are a feature of certain networked 1095
733 * filesystems (i.e. NFS) in which data may have been 1096 dirty_ratelimit = bdi->dirty_ratelimit;
734 * written to the server's write cache, but has not yet 1097 pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
735 * been flushed to permanent storage. 1098 background_thresh, nr_dirty,
736 * Only move pages to writeback if this bdi is over its 1099 bdi_thresh, bdi_dirty);
737 * threshold otherwise wait until the disk writes catch 1100 if (unlikely(pos_ratio == 0)) {
738 * up. 1101 pause = max_pause;
739 */ 1102 goto pause;
740 trace_balance_dirty_start(bdi); 1103 }
741 if (bdi_nr_reclaimable > task_bdi_thresh) { 1104 task_ratelimit = (u64)dirty_ratelimit *
742 pages_written += writeback_inodes_wb(&bdi->wb, 1105 pos_ratio >> RATELIMIT_CALC_SHIFT;
743 write_chunk); 1106 pause = (HZ * pages_dirtied) / (task_ratelimit | 1);
744 trace_balance_dirty_written(bdi, pages_written); 1107 if (unlikely(pause <= 0)) {
745 if (pages_written >= write_chunk) 1108 trace_balance_dirty_pages(bdi,
746 break; /* We've done our duty */ 1109 dirty_thresh,
1110 background_thresh,
1111 nr_dirty,
1112 bdi_thresh,
1113 bdi_dirty,
1114 dirty_ratelimit,
1115 task_ratelimit,
1116 pages_dirtied,
1117 pause,
1118 start_time);
1119 pause = 1; /* avoid resetting nr_dirtied_pause below */
1120 break;
747 } 1121 }
1122 pause = min(pause, max_pause);
1123
1124pause:
1125 trace_balance_dirty_pages(bdi,
1126 dirty_thresh,
1127 background_thresh,
1128 nr_dirty,
1129 bdi_thresh,
1130 bdi_dirty,
1131 dirty_ratelimit,
1132 task_ratelimit,
1133 pages_dirtied,
1134 pause,
1135 start_time);
748 __set_current_state(TASK_UNINTERRUPTIBLE); 1136 __set_current_state(TASK_UNINTERRUPTIBLE);
749 io_schedule_timeout(pause); 1137 io_schedule_timeout(pause);
750 trace_balance_dirty_wait(bdi);
751 1138
752 dirty_thresh = hard_dirty_limit(dirty_thresh); 1139 dirty_thresh = hard_dirty_limit(dirty_thresh);
753 /* 1140 /*
@@ -756,24 +1143,30 @@ static void balance_dirty_pages(struct address_space *mapping,
756 * 200ms is typically more than enough to curb heavy dirtiers; 1143 * 200ms is typically more than enough to curb heavy dirtiers;
757 * (b) the pause time limit makes the dirtiers more responsive. 1144 * (b) the pause time limit makes the dirtiers more responsive.
758 */ 1145 */
759 if (nr_dirty < dirty_thresh && 1146 if (nr_dirty < dirty_thresh)
760 bdi_dirty < (task_bdi_thresh + bdi_thresh) / 2 &&
761 time_after(jiffies, start_time + MAX_PAUSE))
762 break; 1147 break;
763
764 /*
765 * Increase the delay for each loop, up to our previous
766 * default of taking a 100ms nap.
767 */
768 pause <<= 1;
769 if (pause > HZ / 10)
770 pause = HZ / 10;
771 } 1148 }
772 1149
773 /* Clear dirty_exceeded flag only when no task can exceed the limit */ 1150 if (!dirty_exceeded && bdi->dirty_exceeded)
774 if (clear_dirty_exceeded && bdi->dirty_exceeded)
775 bdi->dirty_exceeded = 0; 1151 bdi->dirty_exceeded = 0;
776 1152
1153 current->nr_dirtied = 0;
1154 if (pause == 0) { /* in freerun area */
1155 current->nr_dirtied_pause =
1156 dirty_poll_interval(nr_dirty, dirty_thresh);
1157 } else if (pause <= max_pause / 4 &&
1158 pages_dirtied >= current->nr_dirtied_pause) {
1159 current->nr_dirtied_pause = clamp_val(
1160 dirty_ratelimit * (max_pause / 2) / HZ,
1161 pages_dirtied + pages_dirtied / 8,
1162 pages_dirtied * 4);
1163 } else if (pause >= max_pause) {
1164 current->nr_dirtied_pause = 1 | clamp_val(
1165 dirty_ratelimit * (max_pause / 2) / HZ,
1166 pages_dirtied / 4,
1167 pages_dirtied - pages_dirtied / 8);
1168 }
1169
777 if (writeback_in_progress(bdi)) 1170 if (writeback_in_progress(bdi))
778 return; 1171 return;
779 1172
@@ -785,8 +1178,10 @@ static void balance_dirty_pages(struct address_space *mapping,
785 * In normal mode, we start background writeout at the lower 1178 * In normal mode, we start background writeout at the lower
786 * background_thresh, to keep the amount of dirty memory low. 1179 * background_thresh, to keep the amount of dirty memory low.
787 */ 1180 */
788 if ((laptop_mode && pages_written) || 1181 if (laptop_mode)
789 (!laptop_mode && (nr_reclaimable > background_thresh))) 1182 return;
1183
1184 if (nr_reclaimable > background_thresh)
790 bdi_start_background_writeback(bdi); 1185 bdi_start_background_writeback(bdi);
791} 1186}
792 1187
@@ -800,7 +1195,7 @@ void set_page_dirty_balance(struct page *page, int page_mkwrite)
800 } 1195 }
801} 1196}
802 1197
803static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0; 1198static DEFINE_PER_CPU(int, bdp_ratelimits);
804 1199
805/** 1200/**
806 * balance_dirty_pages_ratelimited_nr - balance dirty memory state 1201 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
@@ -820,31 +1215,39 @@ void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
820 unsigned long nr_pages_dirtied) 1215 unsigned long nr_pages_dirtied)
821{ 1216{
822 struct backing_dev_info *bdi = mapping->backing_dev_info; 1217 struct backing_dev_info *bdi = mapping->backing_dev_info;
823 unsigned long ratelimit; 1218 int ratelimit;
824 unsigned long *p; 1219 int *p;
825 1220
826 if (!bdi_cap_account_dirty(bdi)) 1221 if (!bdi_cap_account_dirty(bdi))
827 return; 1222 return;
828 1223
829 ratelimit = ratelimit_pages; 1224 ratelimit = current->nr_dirtied_pause;
830 if (mapping->backing_dev_info->dirty_exceeded) 1225 if (bdi->dirty_exceeded)
831 ratelimit = 8; 1226 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1227
1228 current->nr_dirtied += nr_pages_dirtied;
832 1229
1230 preempt_disable();
833 /* 1231 /*
834 * Check the rate limiting. Also, we do not want to throttle real-time 1232 * This prevents one CPU to accumulate too many dirtied pages without
835 * tasks in balance_dirty_pages(). Period. 1233 * calling into balance_dirty_pages(), which can happen when there are
1234 * 1000+ tasks, all of them start dirtying pages at exactly the same
1235 * time, hence all honoured too large initial task->nr_dirtied_pause.
836 */ 1236 */
837 preempt_disable();
838 p = &__get_cpu_var(bdp_ratelimits); 1237 p = &__get_cpu_var(bdp_ratelimits);
839 *p += nr_pages_dirtied; 1238 if (unlikely(current->nr_dirtied >= ratelimit))
840 if (unlikely(*p >= ratelimit)) {
841 ratelimit = sync_writeback_pages(*p);
842 *p = 0; 1239 *p = 0;
843 preempt_enable(); 1240 else {
844 balance_dirty_pages(mapping, ratelimit); 1241 *p += nr_pages_dirtied;
845 return; 1242 if (unlikely(*p >= ratelimit_pages)) {
1243 *p = 0;
1244 ratelimit = 0;
1245 }
846 } 1246 }
847 preempt_enable(); 1247 preempt_enable();
1248
1249 if (unlikely(current->nr_dirtied >= ratelimit))
1250 balance_dirty_pages(mapping, current->nr_dirtied);
848} 1251}
849EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr); 1252EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
850 1253
@@ -900,7 +1303,8 @@ void laptop_mode_timer_fn(unsigned long data)
900 * threshold 1303 * threshold
901 */ 1304 */
902 if (bdi_has_dirty_io(&q->backing_dev_info)) 1305 if (bdi_has_dirty_io(&q->backing_dev_info))
903 bdi_start_writeback(&q->backing_dev_info, nr_pages); 1306 bdi_start_writeback(&q->backing_dev_info, nr_pages,
1307 WB_REASON_LAPTOP_TIMER);
904} 1308}
905 1309
906/* 1310/*
@@ -939,22 +1343,17 @@ void laptop_sync_completion(void)
939 * 1343 *
940 * Here we set ratelimit_pages to a level which ensures that when all CPUs are 1344 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
941 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory 1345 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
942 * thresholds before writeback cuts in. 1346 * thresholds.
943 *
944 * But the limit should not be set too high. Because it also controls the
945 * amount of memory which the balance_dirty_pages() caller has to write back.
946 * If this is too large then the caller will block on the IO queue all the
947 * time. So limit it to four megabytes - the balance_dirty_pages() caller
948 * will write six megabyte chunks, max.
949 */ 1347 */
950 1348
951void writeback_set_ratelimit(void) 1349void writeback_set_ratelimit(void)
952{ 1350{
953 ratelimit_pages = vm_total_pages / (num_online_cpus() * 32); 1351 unsigned long background_thresh;
1352 unsigned long dirty_thresh;
1353 global_dirty_limits(&background_thresh, &dirty_thresh);
1354 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
954 if (ratelimit_pages < 16) 1355 if (ratelimit_pages < 16)
955 ratelimit_pages = 16; 1356 ratelimit_pages = 16;
956 if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
957 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
958} 1357}
959 1358
960static int __cpuinit 1359static int __cpuinit
@@ -1324,6 +1723,7 @@ void account_page_dirtied(struct page *page, struct address_space *mapping)
1324 __inc_zone_page_state(page, NR_FILE_DIRTY); 1723 __inc_zone_page_state(page, NR_FILE_DIRTY);
1325 __inc_zone_page_state(page, NR_DIRTIED); 1724 __inc_zone_page_state(page, NR_DIRTIED);
1326 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE); 1725 __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1726 __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1327 task_dirty_inc(current); 1727 task_dirty_inc(current);
1328 task_io_account_write(PAGE_CACHE_SIZE); 1728 task_io_account_write(PAGE_CACHE_SIZE);
1329 } 1729 }
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 132d1ddb2238..a1893c050795 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2266,7 +2266,8 @@ static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
2266 */ 2266 */
2267 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2; 2267 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2268 if (total_scanned > writeback_threshold) { 2268 if (total_scanned > writeback_threshold) {
2269 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned); 2269 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2270 WB_REASON_TRY_TO_FREE_PAGES);
2270 sc->may_writepage = 1; 2271 sc->may_writepage = 1;
2271 } 2272 }
2272 2273
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-02 17:04:01 -0500