Thomas Gleixner | 457c899 | 2019-05-19 13:08:55 +0100 | [diff] [blame] | 1 | // SPDX-License-Identifier: GPL-2.0-only |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2 | /* |
Uwe Zeisberger | f30c226 | 2006-10-03 23:01:26 +0200 | [diff] [blame] | 3 | * mm/page-writeback.c |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 4 | * |
| 5 | * Copyright (C) 2002, Linus Torvalds. |
Peter Zijlstra | 90eec10 | 2015-11-16 11:08:45 +0100 | [diff] [blame] | 6 | * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 7 | * |
| 8 | * Contains functions related to writing back dirty pages at the |
| 9 | * address_space level. |
| 10 | * |
Francois Cami | e1f8e87 | 2008-10-15 22:01:59 -0700 | [diff] [blame] | 11 | * 10Apr2002 Andrew Morton |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 12 | * Initial version |
| 13 | */ |
| 14 | |
| 15 | #include <linux/kernel.h> |
Stefan Roesch | 1bf27e9 | 2022-11-18 16:52:02 -0800 | [diff] [blame] | 16 | #include <linux/math64.h> |
Paul Gortmaker | b95f1b31 | 2011-10-16 02:01:52 -0400 | [diff] [blame] | 17 | #include <linux/export.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 18 | #include <linux/spinlock.h> |
| 19 | #include <linux/fs.h> |
| 20 | #include <linux/mm.h> |
| 21 | #include <linux/swap.h> |
| 22 | #include <linux/slab.h> |
| 23 | #include <linux/pagemap.h> |
| 24 | #include <linux/writeback.h> |
| 25 | #include <linux/init.h> |
| 26 | #include <linux/backing-dev.h> |
Andrew Morton | 55e829a | 2006-12-10 02:19:27 -0800 | [diff] [blame] | 27 | #include <linux/task_io_accounting_ops.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 28 | #include <linux/blkdev.h> |
| 29 | #include <linux/mpage.h> |
Peter Zijlstra | d08b385 | 2006-09-25 23:30:57 -0700 | [diff] [blame] | 30 | #include <linux/rmap.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 31 | #include <linux/percpu.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 32 | #include <linux/smp.h> |
| 33 | #include <linux/sysctl.h> |
| 34 | #include <linux/cpu.h> |
| 35 | #include <linux/syscalls.h> |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 36 | #include <linux/pagevec.h> |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 37 | #include <linux/timer.h> |
Clark Williams | 8bd75c7 | 2013-02-07 09:47:07 -0600 | [diff] [blame] | 38 | #include <linux/sched/rt.h> |
Ingo Molnar | f361bf4 | 2017-02-03 23:47:37 +0100 | [diff] [blame] | 39 | #include <linux/sched/signal.h> |
Lisa Du | 6e543d5 | 2013-09-11 14:22:36 -0700 | [diff] [blame] | 40 | #include <linux/mm_inline.h> |
Dave Chinner | 028c2dd | 2010-07-07 13:24:07 +1000 | [diff] [blame] | 41 | #include <trace/events/writeback.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 42 | |
Lisa Du | 6e543d5 | 2013-09-11 14:22:36 -0700 | [diff] [blame] | 43 | #include "internal.h" |
| 44 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 45 | /* |
Wu Fengguang | ffd1f60 | 2011-06-19 22:18:42 -0600 | [diff] [blame] | 46 | * Sleep at most 200ms at a time in balance_dirty_pages(). |
| 47 | */ |
| 48 | #define MAX_PAUSE max(HZ/5, 1) |
| 49 | |
| 50 | /* |
Wu Fengguang | 5b9b357 | 2011-12-06 13:17:17 -0600 | [diff] [blame] | 51 | * Try to keep balance_dirty_pages() call intervals higher than this many pages |
| 52 | * by raising pause time to max_pause when falls below it. |
| 53 | */ |
| 54 | #define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10)) |
| 55 | |
| 56 | /* |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 57 | * Estimate write bandwidth at 200ms intervals. |
| 58 | */ |
| 59 | #define BANDWIDTH_INTERVAL max(HZ/5, 1) |
| 60 | |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 61 | #define RATELIMIT_CALC_SHIFT 10 |
| 62 | |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 63 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 64 | * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited |
| 65 | * will look to see if it needs to force writeback or throttling. |
| 66 | */ |
| 67 | static long ratelimit_pages = 32; |
| 68 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 69 | /* The following parameters are exported via /proc/sys/vm */ |
| 70 | |
| 71 | /* |
Jens Axboe | 5b0830c | 2009-09-23 19:37:09 +0200 | [diff] [blame] | 72 | * Start background writeback (via writeback threads) at this percentage |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 73 | */ |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 74 | static int dirty_background_ratio = 10; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 75 | |
| 76 | /* |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 77 | * dirty_background_bytes starts at 0 (disabled) so that it is a function of |
| 78 | * dirty_background_ratio * the amount of dirtyable memory |
| 79 | */ |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 80 | static unsigned long dirty_background_bytes; |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 81 | |
| 82 | /* |
Bron Gondwana | 195cf453 | 2008-02-04 22:29:20 -0800 | [diff] [blame] | 83 | * free highmem will not be subtracted from the total free memory |
| 84 | * for calculating free ratios if vm_highmem_is_dirtyable is true |
| 85 | */ |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 86 | static int vm_highmem_is_dirtyable; |
Bron Gondwana | 195cf453 | 2008-02-04 22:29:20 -0800 | [diff] [blame] | 87 | |
| 88 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 89 | * The generator of dirty data starts writeback at this percentage |
| 90 | */ |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 91 | static int vm_dirty_ratio = 20; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 92 | |
| 93 | /* |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 94 | * vm_dirty_bytes starts at 0 (disabled) so that it is a function of |
| 95 | * vm_dirty_ratio * the amount of dirtyable memory |
| 96 | */ |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 97 | static unsigned long vm_dirty_bytes; |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 98 | |
| 99 | /* |
Alexey Dobriyan | 704503d | 2009-03-31 15:23:18 -0700 | [diff] [blame] | 100 | * The interval between `kupdate'-style writebacks |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 101 | */ |
Toshiyuki Okajima | 22ef37e | 2009-05-16 22:56:28 -0700 | [diff] [blame] | 102 | unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 103 | |
Artem Bityutskiy | 91913a2 | 2012-03-21 22:33:00 -0400 | [diff] [blame] | 104 | EXPORT_SYMBOL_GPL(dirty_writeback_interval); |
| 105 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 106 | /* |
Alexey Dobriyan | 704503d | 2009-03-31 15:23:18 -0700 | [diff] [blame] | 107 | * The longest time for which data is allowed to remain dirty |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 108 | */ |
Toshiyuki Okajima | 22ef37e | 2009-05-16 22:56:28 -0700 | [diff] [blame] | 109 | unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 110 | |
| 111 | /* |
Bart Samwel | ed5b43f | 2006-03-24 03:15:49 -0800 | [diff] [blame] | 112 | * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies: |
| 113 | * a full sync is triggered after this time elapses without any disk activity. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 114 | */ |
| 115 | int laptop_mode; |
| 116 | |
| 117 | EXPORT_SYMBOL(laptop_mode); |
| 118 | |
| 119 | /* End of sysctl-exported parameters */ |
| 120 | |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 121 | struct wb_domain global_wb_domain; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 122 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 123 | /* consolidated parameters for balance_dirty_pages() and its subroutines */ |
| 124 | struct dirty_throttle_control { |
Tejun Heo | e9f07df | 2015-05-22 18:23:28 -0400 | [diff] [blame] | 125 | #ifdef CONFIG_CGROUP_WRITEBACK |
| 126 | struct wb_domain *dom; |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 127 | struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */ |
Tejun Heo | e9f07df | 2015-05-22 18:23:28 -0400 | [diff] [blame] | 128 | #endif |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 129 | struct bdi_writeback *wb; |
Tejun Heo | e9770b3 | 2015-05-22 18:23:27 -0400 | [diff] [blame] | 130 | struct fprop_local_percpu *wb_completions; |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 131 | |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 132 | unsigned long avail; /* dirtyable */ |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 133 | unsigned long dirty; /* file_dirty + write + nfs */ |
| 134 | unsigned long thresh; /* dirty threshold */ |
| 135 | unsigned long bg_thresh; /* dirty background threshold */ |
| 136 | |
| 137 | unsigned long wb_dirty; /* per-wb counterparts */ |
| 138 | unsigned long wb_thresh; |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 139 | unsigned long wb_bg_thresh; |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 140 | |
| 141 | unsigned long pos_ratio; |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 142 | }; |
| 143 | |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 144 | /* |
| 145 | * Length of period for aging writeout fractions of bdis. This is an |
| 146 | * arbitrarily chosen number. The longer the period, the slower fractions will |
| 147 | * reflect changes in current writeout rate. |
| 148 | */ |
| 149 | #define VM_COMPLETIONS_PERIOD_LEN (3*HZ) |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 150 | |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 151 | #ifdef CONFIG_CGROUP_WRITEBACK |
| 152 | |
Tejun Heo | d60d1bd | 2015-09-29 12:47:53 -0400 | [diff] [blame] | 153 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
| 154 | .dom = &global_wb_domain, \ |
| 155 | .wb_completions = &(__wb)->completions |
| 156 | |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 157 | #define GDTC_INIT_NO_WB .dom = &global_wb_domain |
Tejun Heo | d60d1bd | 2015-09-29 12:47:53 -0400 | [diff] [blame] | 158 | |
| 159 | #define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \ |
| 160 | .dom = mem_cgroup_wb_domain(__wb), \ |
| 161 | .wb_completions = &(__wb)->memcg_completions, \ |
| 162 | .gdtc = __gdtc |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 163 | |
| 164 | static bool mdtc_valid(struct dirty_throttle_control *dtc) |
| 165 | { |
| 166 | return dtc->dom; |
| 167 | } |
Tejun Heo | e9f07df | 2015-05-22 18:23:28 -0400 | [diff] [blame] | 168 | |
| 169 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) |
| 170 | { |
| 171 | return dtc->dom; |
| 172 | } |
| 173 | |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 174 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
| 175 | { |
| 176 | return mdtc->gdtc; |
| 177 | } |
| 178 | |
Tejun Heo | 841710a | 2015-05-22 18:23:33 -0400 | [diff] [blame] | 179 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
| 180 | { |
| 181 | return &wb->memcg_completions; |
| 182 | } |
| 183 | |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 184 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
| 185 | unsigned long *minp, unsigned long *maxp) |
| 186 | { |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 187 | unsigned long this_bw = READ_ONCE(wb->avg_write_bandwidth); |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 188 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); |
| 189 | unsigned long long min = wb->bdi->min_ratio; |
| 190 | unsigned long long max = wb->bdi->max_ratio; |
| 191 | |
| 192 | /* |
| 193 | * @wb may already be clean by the time control reaches here and |
| 194 | * the total may not include its bw. |
| 195 | */ |
| 196 | if (this_bw < tot_bw) { |
| 197 | if (min) { |
| 198 | min *= this_bw; |
Wen Yang | 6d9e8c6 | 2020-01-13 16:29:23 -0800 | [diff] [blame] | 199 | min = div64_ul(min, tot_bw); |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 200 | } |
Stefan Roesch | ae82291 | 2022-11-18 16:51:59 -0800 | [diff] [blame] | 201 | if (max < 100 * BDI_RATIO_SCALE) { |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 202 | max *= this_bw; |
Wen Yang | 6d9e8c6 | 2020-01-13 16:29:23 -0800 | [diff] [blame] | 203 | max = div64_ul(max, tot_bw); |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 204 | } |
| 205 | } |
| 206 | |
| 207 | *minp = min; |
| 208 | *maxp = max; |
| 209 | } |
| 210 | |
| 211 | #else /* CONFIG_CGROUP_WRITEBACK */ |
| 212 | |
Tejun Heo | d60d1bd | 2015-09-29 12:47:53 -0400 | [diff] [blame] | 213 | #define GDTC_INIT(__wb) .wb = (__wb), \ |
| 214 | .wb_completions = &(__wb)->completions |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 215 | #define GDTC_INIT_NO_WB |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 216 | #define MDTC_INIT(__wb, __gdtc) |
| 217 | |
| 218 | static bool mdtc_valid(struct dirty_throttle_control *dtc) |
| 219 | { |
| 220 | return false; |
| 221 | } |
Tejun Heo | e9f07df | 2015-05-22 18:23:28 -0400 | [diff] [blame] | 222 | |
| 223 | static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc) |
| 224 | { |
| 225 | return &global_wb_domain; |
| 226 | } |
| 227 | |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 228 | static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc) |
| 229 | { |
| 230 | return NULL; |
| 231 | } |
| 232 | |
Tejun Heo | 841710a | 2015-05-22 18:23:33 -0400 | [diff] [blame] | 233 | static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb) |
| 234 | { |
| 235 | return NULL; |
| 236 | } |
| 237 | |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 238 | static void wb_min_max_ratio(struct bdi_writeback *wb, |
| 239 | unsigned long *minp, unsigned long *maxp) |
| 240 | { |
| 241 | *minp = wb->bdi->min_ratio; |
| 242 | *maxp = wb->bdi->max_ratio; |
| 243 | } |
| 244 | |
| 245 | #endif /* CONFIG_CGROUP_WRITEBACK */ |
| 246 | |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 247 | /* |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 248 | * In a memory zone, there is a certain amount of pages we consider |
| 249 | * available for the page cache, which is essentially the number of |
| 250 | * free and reclaimable pages, minus some zone reserves to protect |
| 251 | * lowmem and the ability to uphold the zone's watermarks without |
| 252 | * requiring writeback. |
| 253 | * |
| 254 | * This number of dirtyable pages is the base value of which the |
Ethon Paul | e0857cf | 2020-06-04 16:49:40 -0700 | [diff] [blame] | 255 | * user-configurable dirty ratio is the effective number of pages that |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 256 | * are allowed to be actually dirtied. Per individual zone, or |
| 257 | * globally by using the sum of dirtyable pages over all zones. |
| 258 | * |
| 259 | * Because the user is allowed to specify the dirty limit globally as |
| 260 | * absolute number of bytes, calculating the per-zone dirty limit can |
| 261 | * require translating the configured limit into a percentage of |
| 262 | * global dirtyable memory first. |
| 263 | */ |
| 264 | |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 265 | /** |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 266 | * node_dirtyable_memory - number of dirtyable pages in a node |
| 267 | * @pgdat: the node |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 268 | * |
Mike Rapoport | a862f68 | 2019-03-05 15:48:42 -0800 | [diff] [blame] | 269 | * Return: the node's number of pages potentially available for dirty |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 270 | * page cache. This is the base value for the per-node dirty limits. |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 271 | */ |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 272 | static unsigned long node_dirtyable_memory(struct pglist_data *pgdat) |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 273 | { |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 274 | unsigned long nr_pages = 0; |
| 275 | int z; |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 276 | |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 277 | for (z = 0; z < MAX_NR_ZONES; z++) { |
| 278 | struct zone *zone = pgdat->node_zones + z; |
| 279 | |
| 280 | if (!populated_zone(zone)) |
| 281 | continue; |
| 282 | |
| 283 | nr_pages += zone_page_state(zone, NR_FREE_PAGES); |
| 284 | } |
| 285 | |
Johannes Weiner | a8d0143 | 2016-01-14 15:20:15 -0800 | [diff] [blame] | 286 | /* |
| 287 | * Pages reserved for the kernel should not be considered |
| 288 | * dirtyable, to prevent a situation where reclaim has to |
| 289 | * clean pages in order to balance the zones. |
| 290 | */ |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 291 | nr_pages -= min(nr_pages, pgdat->totalreserve_pages); |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 292 | |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 293 | nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE); |
| 294 | nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE); |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 295 | |
| 296 | return nr_pages; |
| 297 | } |
| 298 | |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 299 | static unsigned long highmem_dirtyable_memory(unsigned long total) |
| 300 | { |
| 301 | #ifdef CONFIG_HIGHMEM |
| 302 | int node; |
Mel Gorman | bb4cc2b | 2016-07-28 15:47:29 -0700 | [diff] [blame] | 303 | unsigned long x = 0; |
Joonsoo Kim | 09b4ab3 | 2016-05-19 17:12:20 -0700 | [diff] [blame] | 304 | int i; |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 305 | |
| 306 | for_each_node_state(node, N_HIGH_MEMORY) { |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 307 | for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) { |
| 308 | struct zone *z; |
Minchan Kim | 9cb937e | 2016-07-28 15:47:08 -0700 | [diff] [blame] | 309 | unsigned long nr_pages; |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 310 | |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 311 | if (!is_highmem_idx(i)) |
| 312 | continue; |
| 313 | |
| 314 | z = &NODE_DATA(node)->node_zones[i]; |
Minchan Kim | 9cb937e | 2016-07-28 15:47:08 -0700 | [diff] [blame] | 315 | if (!populated_zone(z)) |
| 316 | continue; |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 317 | |
Minchan Kim | 9cb937e | 2016-07-28 15:47:08 -0700 | [diff] [blame] | 318 | nr_pages = zone_page_state(z, NR_FREE_PAGES); |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 319 | /* watch for underflows */ |
Minchan Kim | 9cb937e | 2016-07-28 15:47:08 -0700 | [diff] [blame] | 320 | nr_pages -= min(nr_pages, high_wmark_pages(z)); |
Mel Gorman | bb4cc2b | 2016-07-28 15:47:29 -0700 | [diff] [blame] | 321 | nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE); |
| 322 | nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE); |
| 323 | x += nr_pages; |
Joonsoo Kim | 09b4ab3 | 2016-05-19 17:12:20 -0700 | [diff] [blame] | 324 | } |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 325 | } |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 326 | |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 327 | /* |
| 328 | * Make sure that the number of highmem pages is never larger |
| 329 | * than the number of the total dirtyable memory. This can only |
| 330 | * occur in very strange VM situations but we want to make sure |
| 331 | * that this does not occur. |
| 332 | */ |
| 333 | return min(x, total); |
| 334 | #else |
| 335 | return 0; |
| 336 | #endif |
| 337 | } |
| 338 | |
| 339 | /** |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 340 | * global_dirtyable_memory - number of globally dirtyable pages |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 341 | * |
Mike Rapoport | a862f68 | 2019-03-05 15:48:42 -0800 | [diff] [blame] | 342 | * Return: the global number of pages potentially available for dirty |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 343 | * page cache. This is the base value for the global dirty limits. |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 344 | */ |
H Hartley Sweeten | 18cf8cf | 2012-04-12 13:44:20 -0700 | [diff] [blame] | 345 | static unsigned long global_dirtyable_memory(void) |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 346 | { |
| 347 | unsigned long x; |
| 348 | |
Michal Hocko | c41f012 | 2017-09-06 16:23:36 -0700 | [diff] [blame] | 349 | x = global_zone_page_state(NR_FREE_PAGES); |
Johannes Weiner | a8d0143 | 2016-01-14 15:20:15 -0800 | [diff] [blame] | 350 | /* |
| 351 | * Pages reserved for the kernel should not be considered |
| 352 | * dirtyable, to prevent a situation where reclaim has to |
| 353 | * clean pages in order to balance the zones. |
| 354 | */ |
| 355 | x -= min(x, totalreserve_pages); |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 356 | |
Mel Gorman | 599d0c9 | 2016-07-28 15:45:31 -0700 | [diff] [blame] | 357 | x += global_node_page_state(NR_INACTIVE_FILE); |
| 358 | x += global_node_page_state(NR_ACTIVE_FILE); |
Johannes Weiner | a804552b | 2014-01-29 14:05:39 -0800 | [diff] [blame] | 359 | |
Johannes Weiner | 1edf223 | 2012-01-10 15:06:57 -0800 | [diff] [blame] | 360 | if (!vm_highmem_is_dirtyable) |
| 361 | x -= highmem_dirtyable_memory(x); |
| 362 | |
| 363 | return x + 1; /* Ensure that we never return 0 */ |
| 364 | } |
| 365 | |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 366 | /** |
| 367 | * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain |
| 368 | * @dtc: dirty_throttle_control of interest |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 369 | * |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 370 | * Calculate @dtc->thresh and ->bg_thresh considering |
| 371 | * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller |
| 372 | * must ensure that @dtc->avail is set before calling this function. The |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 373 | * dirty limits will be lifted by 1/4 for real-time tasks. |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 374 | */ |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 375 | static void domain_dirty_limits(struct dirty_throttle_control *dtc) |
| 376 | { |
| 377 | const unsigned long available_memory = dtc->avail; |
| 378 | struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc); |
| 379 | unsigned long bytes = vm_dirty_bytes; |
| 380 | unsigned long bg_bytes = dirty_background_bytes; |
Tejun Heo | 62a584f | 2016-05-27 14:34:46 -0400 | [diff] [blame] | 381 | /* convert ratios to per-PAGE_SIZE for higher precision */ |
| 382 | unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100; |
| 383 | unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100; |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 384 | unsigned long thresh; |
| 385 | unsigned long bg_thresh; |
| 386 | struct task_struct *tsk; |
| 387 | |
| 388 | /* gdtc is !NULL iff @dtc is for memcg domain */ |
| 389 | if (gdtc) { |
| 390 | unsigned long global_avail = gdtc->avail; |
| 391 | |
| 392 | /* |
| 393 | * The byte settings can't be applied directly to memcg |
| 394 | * domains. Convert them to ratios by scaling against |
Tejun Heo | 62a584f | 2016-05-27 14:34:46 -0400 | [diff] [blame] | 395 | * globally available memory. As the ratios are in |
| 396 | * per-PAGE_SIZE, they can be obtained by dividing bytes by |
| 397 | * number of pages. |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 398 | */ |
| 399 | if (bytes) |
Tejun Heo | 62a584f | 2016-05-27 14:34:46 -0400 | [diff] [blame] | 400 | ratio = min(DIV_ROUND_UP(bytes, global_avail), |
| 401 | PAGE_SIZE); |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 402 | if (bg_bytes) |
Tejun Heo | 62a584f | 2016-05-27 14:34:46 -0400 | [diff] [blame] | 403 | bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail), |
| 404 | PAGE_SIZE); |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 405 | bytes = bg_bytes = 0; |
| 406 | } |
| 407 | |
| 408 | if (bytes) |
| 409 | thresh = DIV_ROUND_UP(bytes, PAGE_SIZE); |
| 410 | else |
Tejun Heo | 62a584f | 2016-05-27 14:34:46 -0400 | [diff] [blame] | 411 | thresh = (ratio * available_memory) / PAGE_SIZE; |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 412 | |
| 413 | if (bg_bytes) |
| 414 | bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE); |
| 415 | else |
Tejun Heo | 62a584f | 2016-05-27 14:34:46 -0400 | [diff] [blame] | 416 | bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE; |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 417 | |
Michal Hocko | 90daf30 | 2017-11-29 16:10:58 -0800 | [diff] [blame] | 418 | if (bg_thresh >= thresh) |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 419 | bg_thresh = thresh / 2; |
| 420 | tsk = current; |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 421 | if (rt_task(tsk)) { |
NeilBrown | a53eaff | 2016-05-20 16:58:53 -0700 | [diff] [blame] | 422 | bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32; |
| 423 | thresh += thresh / 4 + global_wb_domain.dirty_limit / 32; |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 424 | } |
| 425 | dtc->thresh = thresh; |
| 426 | dtc->bg_thresh = bg_thresh; |
| 427 | |
| 428 | /* we should eventually report the domain in the TP */ |
| 429 | if (!gdtc) |
| 430 | trace_global_dirty_state(bg_thresh, thresh); |
| 431 | } |
| 432 | |
| 433 | /** |
| 434 | * global_dirty_limits - background-writeback and dirty-throttling thresholds |
| 435 | * @pbackground: out parameter for bg_thresh |
| 436 | * @pdirty: out parameter for thresh |
| 437 | * |
| 438 | * Calculate bg_thresh and thresh for global_wb_domain. See |
| 439 | * domain_dirty_limits() for details. |
| 440 | */ |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 441 | void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty) |
| 442 | { |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 443 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB }; |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 444 | |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 445 | gdtc.avail = global_dirtyable_memory(); |
| 446 | domain_dirty_limits(&gdtc); |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 447 | |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 448 | *pbackground = gdtc.bg_thresh; |
| 449 | *pdirty = gdtc.thresh; |
Johannes Weiner | ccafa28 | 2012-01-10 15:07:44 -0800 | [diff] [blame] | 450 | } |
| 451 | |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 452 | /** |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 453 | * node_dirty_limit - maximum number of dirty pages allowed in a node |
| 454 | * @pgdat: the node |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 455 | * |
Mike Rapoport | a862f68 | 2019-03-05 15:48:42 -0800 | [diff] [blame] | 456 | * Return: the maximum number of dirty pages allowed in a node, based |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 457 | * on the node's dirtyable memory. |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 458 | */ |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 459 | static unsigned long node_dirty_limit(struct pglist_data *pgdat) |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 460 | { |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 461 | unsigned long node_memory = node_dirtyable_memory(pgdat); |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 462 | struct task_struct *tsk = current; |
| 463 | unsigned long dirty; |
| 464 | |
| 465 | if (vm_dirty_bytes) |
| 466 | dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) * |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 467 | node_memory / global_dirtyable_memory(); |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 468 | else |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 469 | dirty = vm_dirty_ratio * node_memory / 100; |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 470 | |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 471 | if (rt_task(tsk)) |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 472 | dirty += dirty / 4; |
| 473 | |
| 474 | return dirty; |
| 475 | } |
| 476 | |
| 477 | /** |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 478 | * node_dirty_ok - tells whether a node is within its dirty limits |
| 479 | * @pgdat: the node to check |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 480 | * |
Mike Rapoport | a862f68 | 2019-03-05 15:48:42 -0800 | [diff] [blame] | 481 | * Return: %true when the dirty pages in @pgdat are within the node's |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 482 | * dirty limit, %false if the limit is exceeded. |
| 483 | */ |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 484 | bool node_dirty_ok(struct pglist_data *pgdat) |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 485 | { |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 486 | unsigned long limit = node_dirty_limit(pgdat); |
| 487 | unsigned long nr_pages = 0; |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 488 | |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 489 | nr_pages += node_page_state(pgdat, NR_FILE_DIRTY); |
Mel Gorman | 11fb998 | 2016-07-28 15:46:20 -0700 | [diff] [blame] | 490 | nr_pages += node_page_state(pgdat, NR_WRITEBACK); |
Mel Gorman | 281e372 | 2016-07-28 15:46:11 -0700 | [diff] [blame] | 491 | |
| 492 | return nr_pages <= limit; |
Johannes Weiner | a756cf5 | 2012-01-10 15:07:49 -0800 | [diff] [blame] | 493 | } |
| 494 | |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 495 | #ifdef CONFIG_SYSCTL |
| 496 | static int dirty_background_ratio_handler(struct ctl_table *table, int write, |
Christoph Hellwig | 3292739 | 2020-04-24 08:43:38 +0200 | [diff] [blame] | 497 | void *buffer, size_t *lenp, loff_t *ppos) |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 498 | { |
| 499 | int ret; |
| 500 | |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 501 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 502 | if (ret == 0 && write) |
| 503 | dirty_background_bytes = 0; |
| 504 | return ret; |
| 505 | } |
| 506 | |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 507 | static int dirty_background_bytes_handler(struct ctl_table *table, int write, |
Christoph Hellwig | 3292739 | 2020-04-24 08:43:38 +0200 | [diff] [blame] | 508 | void *buffer, size_t *lenp, loff_t *ppos) |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 509 | { |
| 510 | int ret; |
| 511 | |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 512 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 513 | if (ret == 0 && write) |
| 514 | dirty_background_ratio = 0; |
| 515 | return ret; |
| 516 | } |
| 517 | |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 518 | static int dirty_ratio_handler(struct ctl_table *table, int write, void *buffer, |
Christoph Hellwig | 3292739 | 2020-04-24 08:43:38 +0200 | [diff] [blame] | 519 | size_t *lenp, loff_t *ppos) |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 520 | { |
| 521 | int old_ratio = vm_dirty_ratio; |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 522 | int ret; |
| 523 | |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 524 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 525 | if (ret == 0 && write && vm_dirty_ratio != old_ratio) { |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 526 | writeback_set_ratelimit(); |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 527 | vm_dirty_bytes = 0; |
| 528 | } |
| 529 | return ret; |
| 530 | } |
| 531 | |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 532 | static int dirty_bytes_handler(struct ctl_table *table, int write, |
Christoph Hellwig | 3292739 | 2020-04-24 08:43:38 +0200 | [diff] [blame] | 533 | void *buffer, size_t *lenp, loff_t *ppos) |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 534 | { |
Sven Wegener | fc3501d | 2009-02-11 13:04:23 -0800 | [diff] [blame] | 535 | unsigned long old_bytes = vm_dirty_bytes; |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 536 | int ret; |
| 537 | |
Alexey Dobriyan | 8d65af7 | 2009-09-23 15:57:19 -0700 | [diff] [blame] | 538 | ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos); |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 539 | if (ret == 0 && write && vm_dirty_bytes != old_bytes) { |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 540 | writeback_set_ratelimit(); |
David Rientjes | 2da0299 | 2009-01-06 14:39:31 -0800 | [diff] [blame] | 541 | vm_dirty_ratio = 0; |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 542 | } |
| 543 | return ret; |
| 544 | } |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 545 | #endif |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 546 | |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 547 | static unsigned long wp_next_time(unsigned long cur_time) |
| 548 | { |
| 549 | cur_time += VM_COMPLETIONS_PERIOD_LEN; |
| 550 | /* 0 has a special meaning... */ |
| 551 | if (!cur_time) |
| 552 | return 1; |
| 553 | return cur_time; |
| 554 | } |
| 555 | |
Matthew Wilcox (Oracle) | cc24df4 | 2021-03-20 16:45:15 -0400 | [diff] [blame] | 556 | static void wb_domain_writeout_add(struct wb_domain *dom, |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 557 | struct fprop_local_percpu *completions, |
Matthew Wilcox (Oracle) | cc24df4 | 2021-03-20 16:45:15 -0400 | [diff] [blame] | 558 | unsigned int max_prop_frac, long nr) |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 559 | { |
Matthew Wilcox (Oracle) | be5f179 | 2021-03-20 16:34:54 -0400 | [diff] [blame] | 560 | __fprop_add_percpu_max(&dom->completions, completions, |
Matthew Wilcox (Oracle) | cc24df4 | 2021-03-20 16:45:15 -0400 | [diff] [blame] | 561 | max_prop_frac, nr); |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 562 | /* First event after period switching was turned off? */ |
Steven Rostedt (VMware) | 517663e | 2017-02-24 14:59:24 -0800 | [diff] [blame] | 563 | if (unlikely(!dom->period_time)) { |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 564 | /* |
| 565 | * We can race with other __bdi_writeout_inc calls here but |
| 566 | * it does not cause any harm since the resulting time when |
| 567 | * timer will fire and what is in writeout_period_time will be |
| 568 | * roughly the same. |
| 569 | */ |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 570 | dom->period_time = wp_next_time(jiffies); |
| 571 | mod_timer(&dom->period_timer, dom->period_time); |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 572 | } |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 573 | } |
| 574 | |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 575 | /* |
| 576 | * Increment @wb's writeout completion count and the global writeout |
Matthew Wilcox (Oracle) | 269ccca3 | 2021-01-15 23:34:16 -0500 | [diff] [blame] | 577 | * completion count. Called from __folio_end_writeback(). |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 578 | */ |
Matthew Wilcox (Oracle) | cc24df4 | 2021-03-20 16:45:15 -0400 | [diff] [blame] | 579 | static inline void __wb_writeout_add(struct bdi_writeback *wb, long nr) |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 580 | { |
Tejun Heo | 841710a | 2015-05-22 18:23:33 -0400 | [diff] [blame] | 581 | struct wb_domain *cgdom; |
| 582 | |
Matthew Wilcox (Oracle) | cc24df4 | 2021-03-20 16:45:15 -0400 | [diff] [blame] | 583 | wb_stat_mod(wb, WB_WRITTEN, nr); |
| 584 | wb_domain_writeout_add(&global_wb_domain, &wb->completions, |
| 585 | wb->bdi->max_prop_frac, nr); |
Tejun Heo | 841710a | 2015-05-22 18:23:33 -0400 | [diff] [blame] | 586 | |
| 587 | cgdom = mem_cgroup_wb_domain(wb); |
| 588 | if (cgdom) |
Matthew Wilcox (Oracle) | cc24df4 | 2021-03-20 16:45:15 -0400 | [diff] [blame] | 589 | wb_domain_writeout_add(cgdom, wb_memcg_completions(wb), |
| 590 | wb->bdi->max_prop_frac, nr); |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 591 | } |
| 592 | |
Tejun Heo | 93f78d8 | 2015-05-22 17:13:27 -0400 | [diff] [blame] | 593 | void wb_writeout_inc(struct bdi_writeback *wb) |
Miklos Szeredi | dd5656e | 2008-04-30 00:54:37 -0700 | [diff] [blame] | 594 | { |
| 595 | unsigned long flags; |
| 596 | |
| 597 | local_irq_save(flags); |
Matthew Wilcox (Oracle) | cc24df4 | 2021-03-20 16:45:15 -0400 | [diff] [blame] | 598 | __wb_writeout_add(wb, 1); |
Miklos Szeredi | dd5656e | 2008-04-30 00:54:37 -0700 | [diff] [blame] | 599 | local_irq_restore(flags); |
| 600 | } |
Tejun Heo | 93f78d8 | 2015-05-22 17:13:27 -0400 | [diff] [blame] | 601 | EXPORT_SYMBOL_GPL(wb_writeout_inc); |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 602 | |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 603 | /* |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 604 | * On idle system, we can be called long after we scheduled because we use |
| 605 | * deferred timers so count with missed periods. |
| 606 | */ |
Kees Cook | 9823e51 | 2017-11-15 17:37:04 -0800 | [diff] [blame] | 607 | static void writeout_period(struct timer_list *t) |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 608 | { |
Kees Cook | 9823e51 | 2017-11-15 17:37:04 -0800 | [diff] [blame] | 609 | struct wb_domain *dom = from_timer(dom, t, period_timer); |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 610 | int miss_periods = (jiffies - dom->period_time) / |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 611 | VM_COMPLETIONS_PERIOD_LEN; |
| 612 | |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 613 | if (fprop_new_period(&dom->completions, miss_periods + 1)) { |
| 614 | dom->period_time = wp_next_time(dom->period_time + |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 615 | miss_periods * VM_COMPLETIONS_PERIOD_LEN); |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 616 | mod_timer(&dom->period_timer, dom->period_time); |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 617 | } else { |
| 618 | /* |
| 619 | * Aging has zeroed all fractions. Stop wasting CPU on period |
| 620 | * updates. |
| 621 | */ |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 622 | dom->period_time = 0; |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 623 | } |
| 624 | } |
| 625 | |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 626 | int wb_domain_init(struct wb_domain *dom, gfp_t gfp) |
| 627 | { |
| 628 | memset(dom, 0, sizeof(*dom)); |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 629 | |
| 630 | spin_lock_init(&dom->lock); |
| 631 | |
Kees Cook | 9823e51 | 2017-11-15 17:37:04 -0800 | [diff] [blame] | 632 | timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE); |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 633 | |
| 634 | dom->dirty_limit_tstamp = jiffies; |
| 635 | |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 636 | return fprop_global_init(&dom->completions, gfp); |
| 637 | } |
| 638 | |
Tejun Heo | 841710a | 2015-05-22 18:23:33 -0400 | [diff] [blame] | 639 | #ifdef CONFIG_CGROUP_WRITEBACK |
| 640 | void wb_domain_exit(struct wb_domain *dom) |
| 641 | { |
| 642 | del_timer_sync(&dom->period_timer); |
| 643 | fprop_global_destroy(&dom->completions); |
| 644 | } |
| 645 | #endif |
| 646 | |
Jan Kara | eb608e3 | 2012-05-24 18:59:11 +0200 | [diff] [blame] | 647 | /* |
Johannes Weiner | d08c429 | 2011-10-31 17:07:05 -0700 | [diff] [blame] | 648 | * bdi_min_ratio keeps the sum of the minimum dirty shares of all |
| 649 | * registered backing devices, which, for obvious reasons, can not |
| 650 | * exceed 100%. |
Peter Zijlstra | 189d3c4 | 2008-04-30 00:54:35 -0700 | [diff] [blame] | 651 | */ |
Peter Zijlstra | 189d3c4 | 2008-04-30 00:54:35 -0700 | [diff] [blame] | 652 | static unsigned int bdi_min_ratio; |
| 653 | |
Stefan Roesch | 1bf27e9 | 2022-11-18 16:52:02 -0800 | [diff] [blame] | 654 | static int bdi_check_pages_limit(unsigned long pages) |
| 655 | { |
| 656 | unsigned long max_dirty_pages = global_dirtyable_memory(); |
| 657 | |
| 658 | if (pages > max_dirty_pages) |
| 659 | return -EINVAL; |
| 660 | |
| 661 | return 0; |
| 662 | } |
| 663 | |
| 664 | static unsigned long bdi_ratio_from_pages(unsigned long pages) |
| 665 | { |
| 666 | unsigned long background_thresh; |
| 667 | unsigned long dirty_thresh; |
| 668 | unsigned long ratio; |
| 669 | |
| 670 | global_dirty_limits(&background_thresh, &dirty_thresh); |
| 671 | ratio = div64_u64(pages * 100ULL * BDI_RATIO_SCALE, dirty_thresh); |
| 672 | |
| 673 | return ratio; |
| 674 | } |
| 675 | |
Stefan Roesch | 00df7d5 | 2022-11-18 16:52:00 -0800 | [diff] [blame] | 676 | static u64 bdi_get_bytes(unsigned int ratio) |
| 677 | { |
| 678 | unsigned long background_thresh; |
| 679 | unsigned long dirty_thresh; |
| 680 | u64 bytes; |
| 681 | |
| 682 | global_dirty_limits(&background_thresh, &dirty_thresh); |
| 683 | bytes = (dirty_thresh * PAGE_SIZE * ratio) / BDI_RATIO_SCALE / 100; |
| 684 | |
| 685 | return bytes; |
| 686 | } |
| 687 | |
Stefan Roesch | 8021fb3 | 2022-11-18 16:52:06 -0800 | [diff] [blame] | 688 | static int __bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) |
Peter Zijlstra | 189d3c4 | 2008-04-30 00:54:35 -0700 | [diff] [blame] | 689 | { |
Chen Wandun | 21f0dd8 | 2022-04-28 23:15:57 -0700 | [diff] [blame] | 690 | unsigned int delta; |
Peter Zijlstra | 189d3c4 | 2008-04-30 00:54:35 -0700 | [diff] [blame] | 691 | int ret = 0; |
Peter Zijlstra | 189d3c4 | 2008-04-30 00:54:35 -0700 | [diff] [blame] | 692 | |
Stefan Roesch | 2c44af4f | 2022-11-18 16:52:13 -0800 | [diff] [blame] | 693 | if (min_ratio > 100 * BDI_RATIO_SCALE) |
| 694 | return -EINVAL; |
Stefan Roesch | ae82291 | 2022-11-18 16:51:59 -0800 | [diff] [blame] | 695 | |
Jens Axboe | cfc4ba5 | 2009-09-14 13:12:40 +0200 | [diff] [blame] | 696 | spin_lock_bh(&bdi_lock); |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 697 | if (min_ratio > bdi->max_ratio) { |
Peter Zijlstra | 189d3c4 | 2008-04-30 00:54:35 -0700 | [diff] [blame] | 698 | ret = -EINVAL; |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 699 | } else { |
Chen Wandun | 21f0dd8 | 2022-04-28 23:15:57 -0700 | [diff] [blame] | 700 | if (min_ratio < bdi->min_ratio) { |
| 701 | delta = bdi->min_ratio - min_ratio; |
| 702 | bdi_min_ratio -= delta; |
| 703 | bdi->min_ratio = min_ratio; |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 704 | } else { |
Chen Wandun | 21f0dd8 | 2022-04-28 23:15:57 -0700 | [diff] [blame] | 705 | delta = min_ratio - bdi->min_ratio; |
Stefan Roesch | ae82291 | 2022-11-18 16:51:59 -0800 | [diff] [blame] | 706 | if (bdi_min_ratio + delta < 100 * BDI_RATIO_SCALE) { |
Chen Wandun | 21f0dd8 | 2022-04-28 23:15:57 -0700 | [diff] [blame] | 707 | bdi_min_ratio += delta; |
| 708 | bdi->min_ratio = min_ratio; |
| 709 | } else { |
| 710 | ret = -EINVAL; |
| 711 | } |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 712 | } |
| 713 | } |
Jens Axboe | cfc4ba5 | 2009-09-14 13:12:40 +0200 | [diff] [blame] | 714 | spin_unlock_bh(&bdi_lock); |
Peter Zijlstra | 189d3c4 | 2008-04-30 00:54:35 -0700 | [diff] [blame] | 715 | |
| 716 | return ret; |
| 717 | } |
| 718 | |
Stefan Roesch | efc3e6a | 2022-11-18 16:52:01 -0800 | [diff] [blame] | 719 | static int __bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio) |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 720 | { |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 721 | int ret = 0; |
| 722 | |
Stefan Roesch | 4e230b4 | 2022-11-18 16:52:10 -0800 | [diff] [blame] | 723 | if (max_ratio > 100 * BDI_RATIO_SCALE) |
| 724 | return -EINVAL; |
| 725 | |
Jens Axboe | cfc4ba5 | 2009-09-14 13:12:40 +0200 | [diff] [blame] | 726 | spin_lock_bh(&bdi_lock); |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 727 | if (bdi->min_ratio > max_ratio) { |
| 728 | ret = -EINVAL; |
| 729 | } else { |
| 730 | bdi->max_ratio = max_ratio; |
Jingbo Xu | fa151a3 | 2023-12-19 22:25:08 +0800 | [diff] [blame] | 731 | bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / |
| 732 | (100 * BDI_RATIO_SCALE); |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 733 | } |
Jens Axboe | cfc4ba5 | 2009-09-14 13:12:40 +0200 | [diff] [blame] | 734 | spin_unlock_bh(&bdi_lock); |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 735 | |
| 736 | return ret; |
| 737 | } |
Stefan Roesch | efc3e6a | 2022-11-18 16:52:01 -0800 | [diff] [blame] | 738 | |
Stefan Roesch | 2c44af4f | 2022-11-18 16:52:13 -0800 | [diff] [blame] | 739 | int bdi_set_min_ratio_no_scale(struct backing_dev_info *bdi, unsigned int min_ratio) |
| 740 | { |
| 741 | return __bdi_set_min_ratio(bdi, min_ratio); |
| 742 | } |
| 743 | |
Stefan Roesch | 4e230b4 | 2022-11-18 16:52:10 -0800 | [diff] [blame] | 744 | int bdi_set_max_ratio_no_scale(struct backing_dev_info *bdi, unsigned int max_ratio) |
| 745 | { |
| 746 | return __bdi_set_max_ratio(bdi, max_ratio); |
| 747 | } |
| 748 | |
Stefan Roesch | 8021fb3 | 2022-11-18 16:52:06 -0800 | [diff] [blame] | 749 | int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio) |
| 750 | { |
| 751 | return __bdi_set_min_ratio(bdi, min_ratio * BDI_RATIO_SCALE); |
| 752 | } |
| 753 | |
Stefan Roesch | efc3e6a | 2022-11-18 16:52:01 -0800 | [diff] [blame] | 754 | int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio) |
| 755 | { |
Stefan Roesch | efc3e6a | 2022-11-18 16:52:01 -0800 | [diff] [blame] | 756 | return __bdi_set_max_ratio(bdi, max_ratio * BDI_RATIO_SCALE); |
| 757 | } |
Peter Zijlstra | a42dde0 | 2008-04-30 00:54:36 -0700 | [diff] [blame] | 758 | EXPORT_SYMBOL(bdi_set_max_ratio); |
| 759 | |
Stefan Roesch | 712c00d | 2022-11-18 16:52:05 -0800 | [diff] [blame] | 760 | u64 bdi_get_min_bytes(struct backing_dev_info *bdi) |
| 761 | { |
| 762 | return bdi_get_bytes(bdi->min_ratio); |
| 763 | } |
| 764 | |
Stefan Roesch | 803c980 | 2022-11-18 16:52:07 -0800 | [diff] [blame] | 765 | int bdi_set_min_bytes(struct backing_dev_info *bdi, u64 min_bytes) |
| 766 | { |
| 767 | int ret; |
| 768 | unsigned long pages = min_bytes >> PAGE_SHIFT; |
| 769 | unsigned long min_ratio; |
| 770 | |
| 771 | ret = bdi_check_pages_limit(pages); |
| 772 | if (ret) |
| 773 | return ret; |
| 774 | |
| 775 | min_ratio = bdi_ratio_from_pages(pages); |
| 776 | return __bdi_set_min_ratio(bdi, min_ratio); |
| 777 | } |
| 778 | |
Stefan Roesch | 00df7d5 | 2022-11-18 16:52:00 -0800 | [diff] [blame] | 779 | u64 bdi_get_max_bytes(struct backing_dev_info *bdi) |
| 780 | { |
| 781 | return bdi_get_bytes(bdi->max_ratio); |
| 782 | } |
| 783 | |
Stefan Roesch | 1bf27e9 | 2022-11-18 16:52:02 -0800 | [diff] [blame] | 784 | int bdi_set_max_bytes(struct backing_dev_info *bdi, u64 max_bytes) |
| 785 | { |
| 786 | int ret; |
| 787 | unsigned long pages = max_bytes >> PAGE_SHIFT; |
| 788 | unsigned long max_ratio; |
| 789 | |
| 790 | ret = bdi_check_pages_limit(pages); |
| 791 | if (ret) |
| 792 | return ret; |
| 793 | |
| 794 | max_ratio = bdi_ratio_from_pages(pages); |
| 795 | return __bdi_set_max_ratio(bdi, max_ratio); |
| 796 | } |
| 797 | |
Stefan Roesch | 8e9d5ea | 2022-11-18 16:51:56 -0800 | [diff] [blame] | 798 | int bdi_set_strict_limit(struct backing_dev_info *bdi, unsigned int strict_limit) |
| 799 | { |
| 800 | if (strict_limit > 1) |
| 801 | return -EINVAL; |
| 802 | |
| 803 | spin_lock_bh(&bdi_lock); |
| 804 | if (strict_limit) |
| 805 | bdi->capabilities |= BDI_CAP_STRICTLIMIT; |
| 806 | else |
| 807 | bdi->capabilities &= ~BDI_CAP_STRICTLIMIT; |
| 808 | spin_unlock_bh(&bdi_lock); |
| 809 | |
| 810 | return 0; |
| 811 | } |
| 812 | |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 813 | static unsigned long dirty_freerun_ceiling(unsigned long thresh, |
| 814 | unsigned long bg_thresh) |
| 815 | { |
| 816 | return (thresh + bg_thresh) / 2; |
| 817 | } |
| 818 | |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 819 | static unsigned long hard_dirty_limit(struct wb_domain *dom, |
| 820 | unsigned long thresh) |
Wu Fengguang | ffd1f60 | 2011-06-19 22:18:42 -0600 | [diff] [blame] | 821 | { |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 822 | return max(thresh, dom->dirty_limit); |
Wu Fengguang | ffd1f60 | 2011-06-19 22:18:42 -0600 | [diff] [blame] | 823 | } |
| 824 | |
Tejun Heo | c5edf9c | 2015-09-29 13:04:26 -0400 | [diff] [blame] | 825 | /* |
| 826 | * Memory which can be further allocated to a memcg domain is capped by |
| 827 | * system-wide clean memory excluding the amount being used in the domain. |
| 828 | */ |
| 829 | static void mdtc_calc_avail(struct dirty_throttle_control *mdtc, |
| 830 | unsigned long filepages, unsigned long headroom) |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 831 | { |
| 832 | struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc); |
Tejun Heo | c5edf9c | 2015-09-29 13:04:26 -0400 | [diff] [blame] | 833 | unsigned long clean = filepages - min(filepages, mdtc->dirty); |
| 834 | unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty); |
| 835 | unsigned long other_clean = global_clean - min(global_clean, clean); |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 836 | |
Tejun Heo | c5edf9c | 2015-09-29 13:04:26 -0400 | [diff] [blame] | 837 | mdtc->avail = filepages + min(headroom, other_clean); |
Christoph Lameter | 1b42446 | 2007-05-06 14:48:59 -0700 | [diff] [blame] | 838 | } |
| 839 | |
Kemeng Shi | ba62d5c | 2024-05-14 20:52:48 +0800 | [diff] [blame] | 840 | static inline bool dtc_is_global(struct dirty_throttle_control *dtc) |
| 841 | { |
| 842 | return mdtc_gdtc(dtc) == NULL; |
| 843 | } |
| 844 | |
| 845 | /* |
| 846 | * Dirty background will ignore pages being written as we're trying to |
| 847 | * decide whether to put more under writeback. |
| 848 | */ |
| 849 | static void domain_dirty_avail(struct dirty_throttle_control *dtc, |
| 850 | bool include_writeback) |
| 851 | { |
| 852 | if (dtc_is_global(dtc)) { |
| 853 | dtc->avail = global_dirtyable_memory(); |
| 854 | dtc->dirty = global_node_page_state(NR_FILE_DIRTY); |
| 855 | if (include_writeback) |
| 856 | dtc->dirty += global_node_page_state(NR_WRITEBACK); |
| 857 | } else { |
| 858 | unsigned long filepages = 0, headroom = 0, writeback = 0; |
| 859 | |
| 860 | mem_cgroup_wb_stats(dtc->wb, &filepages, &headroom, &dtc->dirty, |
| 861 | &writeback); |
| 862 | if (include_writeback) |
| 863 | dtc->dirty += writeback; |
| 864 | mdtc_calc_avail(dtc, filepages, headroom); |
| 865 | } |
| 866 | } |
| 867 | |
Wu Fengguang | 6f71865 | 2011-03-02 17:14:34 -0600 | [diff] [blame] | 868 | /** |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 869 | * __wb_calc_thresh - @wb's share of dirty threshold |
Tejun Heo | b1cbc6d | 2015-05-22 18:23:25 -0400 | [diff] [blame] | 870 | * @dtc: dirty_throttle_context of interest |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 871 | * @thresh: dirty throttling or dirty background threshold of wb_domain in @dtc |
Wu Fengguang | 1babe18 | 2010-08-11 14:17:40 -0700 | [diff] [blame] | 872 | * |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 873 | * Note that balance_dirty_pages() will only seriously take dirty throttling |
| 874 | * threshold as a hard limit when sleeping max_pause per page is not enough |
| 875 | * to keep the dirty pages under control. For example, when the device is |
| 876 | * completely stalled due to some error conditions, or when there are 1000 |
| 877 | * dd tasks writing to a slow 10MB/s USB key. |
Wu Fengguang | aed21ad | 2011-11-23 11:44:41 -0600 | [diff] [blame] | 878 | * In the other normal situations, it acts more gently by throttling the tasks |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 879 | * more (rather than completely block them) when the wb dirty pages go high. |
Wu Fengguang | 6f71865 | 2011-03-02 17:14:34 -0600 | [diff] [blame] | 880 | * |
| 881 | * It allocates high/low dirty limits to fast/slow devices, in order to prevent |
Wu Fengguang | 1babe18 | 2010-08-11 14:17:40 -0700 | [diff] [blame] | 882 | * - starving fast devices |
| 883 | * - piling up dirty pages (that will take long time to sync) on slow devices |
| 884 | * |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 885 | * The wb's share of dirty limit will be adapting to its throughput and |
Wu Fengguang | 1babe18 | 2010-08-11 14:17:40 -0700 | [diff] [blame] | 886 | * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set. |
Mike Rapoport | a862f68 | 2019-03-05 15:48:42 -0800 | [diff] [blame] | 887 | * |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 888 | * Return: @wb's dirty limit in pages. For dirty throttling limit, the term |
| 889 | * "dirty" in the context of dirty balancing includes all PG_dirty and |
| 890 | * PG_writeback pages. |
Wu Fengguang | 1babe18 | 2010-08-11 14:17:40 -0700 | [diff] [blame] | 891 | */ |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 892 | static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc, |
| 893 | unsigned long thresh) |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 894 | { |
Tejun Heo | e9f07df | 2015-05-22 18:23:28 -0400 | [diff] [blame] | 895 | struct wb_domain *dom = dtc_dom(dtc); |
Tejun Heo | 0d960a3 | 2015-05-22 18:23:19 -0400 | [diff] [blame] | 896 | u64 wb_thresh; |
Wen Yang | d3ac946 | 2020-01-13 16:29:26 -0800 | [diff] [blame] | 897 | unsigned long numerator, denominator; |
Tejun Heo | 693108a | 2015-05-22 17:13:49 -0400 | [diff] [blame] | 898 | unsigned long wb_min_ratio, wb_max_ratio; |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 899 | |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 900 | /* |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 901 | * Calculate this wb's share of the thresh ratio. |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 902 | */ |
Tejun Heo | e9770b3 | 2015-05-22 18:23:27 -0400 | [diff] [blame] | 903 | fprop_fraction_percpu(&dom->completions, dtc->wb_completions, |
Tejun Heo | 380c27c | 2015-05-22 18:23:21 -0400 | [diff] [blame] | 904 | &numerator, &denominator); |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 905 | |
Stefan Roesch | ae82291 | 2022-11-18 16:51:59 -0800 | [diff] [blame] | 906 | wb_thresh = (thresh * (100 * BDI_RATIO_SCALE - bdi_min_ratio)) / (100 * BDI_RATIO_SCALE); |
Tejun Heo | 0d960a3 | 2015-05-22 18:23:19 -0400 | [diff] [blame] | 907 | wb_thresh *= numerator; |
Wen Yang | d3ac946 | 2020-01-13 16:29:26 -0800 | [diff] [blame] | 908 | wb_thresh = div64_ul(wb_thresh, denominator); |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 909 | |
Tejun Heo | b1cbc6d | 2015-05-22 18:23:25 -0400 | [diff] [blame] | 910 | wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio); |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 911 | |
Stefan Roesch | ae82291 | 2022-11-18 16:51:59 -0800 | [diff] [blame] | 912 | wb_thresh += (thresh * wb_min_ratio) / (100 * BDI_RATIO_SCALE); |
| 913 | if (wb_thresh > (thresh * wb_max_ratio) / (100 * BDI_RATIO_SCALE)) |
| 914 | wb_thresh = thresh * wb_max_ratio / (100 * BDI_RATIO_SCALE); |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 915 | |
Tejun Heo | 0d960a3 | 2015-05-22 18:23:19 -0400 | [diff] [blame] | 916 | return wb_thresh; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 917 | } |
| 918 | |
Tejun Heo | b1cbc6d | 2015-05-22 18:23:25 -0400 | [diff] [blame] | 919 | unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh) |
| 920 | { |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 921 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
| 922 | |
| 923 | return __wb_calc_thresh(&gdtc, thresh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 924 | } |
| 925 | |
Kemeng Shi | 4b5bbc3 | 2024-04-23 11:46:40 +0800 | [diff] [blame] | 926 | unsigned long cgwb_calc_thresh(struct bdi_writeback *wb) |
| 927 | { |
| 928 | struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB }; |
| 929 | struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) }; |
Kemeng Shi | 4b5bbc3 | 2024-04-23 11:46:40 +0800 | [diff] [blame] | 930 | |
Kemeng Shi | ba62d5c | 2024-05-14 20:52:48 +0800 | [diff] [blame] | 931 | domain_dirty_avail(&gdtc, true); |
| 932 | domain_dirty_avail(&mdtc, true); |
Kemeng Shi | 4b5bbc3 | 2024-04-23 11:46:40 +0800 | [diff] [blame] | 933 | domain_dirty_limits(&mdtc); |
| 934 | |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 935 | return __wb_calc_thresh(&mdtc, mdtc.thresh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 936 | } |
| 937 | |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 938 | /* |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 939 | * setpoint - dirty 3 |
| 940 | * f(dirty) := 1.0 + (----------------) |
| 941 | * limit - setpoint |
| 942 | * |
| 943 | * it's a 3rd order polynomial that subjects to |
| 944 | * |
| 945 | * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast |
| 946 | * (2) f(setpoint) = 1.0 => the balance point |
| 947 | * (3) f(limit) = 0 => the hard limit |
| 948 | * (4) df/dx <= 0 => negative feedback control |
| 949 | * (5) the closer to setpoint, the smaller |df/dx| (and the reverse) |
| 950 | * => fast response on large errors; small oscillation near setpoint |
| 951 | */ |
Rik van Riel | d5c9fde | 2014-05-06 12:50:01 -0700 | [diff] [blame] | 952 | static long long pos_ratio_polynom(unsigned long setpoint, |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 953 | unsigned long dirty, |
| 954 | unsigned long limit) |
| 955 | { |
| 956 | long long pos_ratio; |
| 957 | long x; |
| 958 | |
Rik van Riel | d5c9fde | 2014-05-06 12:50:01 -0700 | [diff] [blame] | 959 | x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT, |
Tejun Heo | 464d138 | 2015-04-21 16:49:13 -0400 | [diff] [blame] | 960 | (limit - setpoint) | 1); |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 961 | pos_ratio = x; |
| 962 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; |
| 963 | pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT; |
| 964 | pos_ratio += 1 << RATELIMIT_CALC_SHIFT; |
| 965 | |
| 966 | return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT); |
| 967 | } |
| 968 | |
| 969 | /* |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 970 | * Dirty position control. |
| 971 | * |
| 972 | * (o) global/bdi setpoints |
| 973 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 974 | * We want the dirty pages be balanced around the global/wb setpoints. |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 975 | * When the number of dirty pages is higher/lower than the setpoint, the |
| 976 | * dirty position control ratio (and hence task dirty ratelimit) will be |
| 977 | * decreased/increased to bring the dirty pages back to the setpoint. |
| 978 | * |
| 979 | * pos_ratio = 1 << RATELIMIT_CALC_SHIFT |
| 980 | * |
| 981 | * if (dirty < setpoint) scale up pos_ratio |
| 982 | * if (dirty > setpoint) scale down pos_ratio |
| 983 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 984 | * if (wb_dirty < wb_setpoint) scale up pos_ratio |
| 985 | * if (wb_dirty > wb_setpoint) scale down pos_ratio |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 986 | * |
| 987 | * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT |
| 988 | * |
| 989 | * (o) global control line |
| 990 | * |
| 991 | * ^ pos_ratio |
| 992 | * | |
| 993 | * | |<===== global dirty control scope ======>| |
Chi Wu | 0323155 | 2021-06-28 19:35:34 -0700 | [diff] [blame] | 994 | * 2.0 * * * * * * * |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 995 | * | .* |
| 996 | * | . * |
| 997 | * | . * |
| 998 | * | . * |
| 999 | * | . * |
| 1000 | * | . * |
| 1001 | * 1.0 ................................* |
| 1002 | * | . . * |
| 1003 | * | . . * |
| 1004 | * | . . * |
| 1005 | * | . . * |
| 1006 | * | . . * |
| 1007 | * 0 +------------.------------------.----------------------*-------------> |
| 1008 | * freerun^ setpoint^ limit^ dirty pages |
| 1009 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1010 | * (o) wb control line |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1011 | * |
| 1012 | * ^ pos_ratio |
| 1013 | * | |
| 1014 | * | * |
| 1015 | * | * |
| 1016 | * | * |
| 1017 | * | * |
| 1018 | * | * |<=========== span ============>| |
| 1019 | * 1.0 .......................* |
| 1020 | * | . * |
| 1021 | * | . * |
| 1022 | * | . * |
| 1023 | * | . * |
| 1024 | * | . * |
| 1025 | * | . * |
| 1026 | * | . * |
| 1027 | * | . * |
| 1028 | * | . * |
| 1029 | * | . * |
| 1030 | * | . * |
| 1031 | * 1/4 ...............................................* * * * * * * * * * * * |
| 1032 | * | . . |
| 1033 | * | . . |
| 1034 | * | . . |
| 1035 | * 0 +----------------------.-------------------------------.-------------> |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1036 | * wb_setpoint^ x_intercept^ |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1037 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1038 | * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1039 | * be smoothly throttled down to normal if it starts high in situations like |
| 1040 | * - start writing to a slow SD card and a fast disk at the same time. The SD |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1041 | * card's wb_dirty may rush to many times higher than wb_setpoint. |
| 1042 | * - the wb dirty thresh drops quickly due to change of JBOD workload |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1043 | */ |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1044 | static void wb_position_ratio(struct dirty_throttle_control *dtc) |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1045 | { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1046 | struct bdi_writeback *wb = dtc->wb; |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1047 | unsigned long write_bw = READ_ONCE(wb->avg_write_bandwidth); |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1048 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 1049 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1050 | unsigned long wb_thresh = dtc->wb_thresh; |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1051 | unsigned long x_intercept; |
| 1052 | unsigned long setpoint; /* dirty pages' target balance point */ |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1053 | unsigned long wb_setpoint; |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1054 | unsigned long span; |
| 1055 | long long pos_ratio; /* for scaling up/down the rate limit */ |
| 1056 | long x; |
| 1057 | |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1058 | dtc->pos_ratio = 0; |
| 1059 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1060 | if (unlikely(dtc->dirty >= limit)) |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1061 | return; |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1062 | |
| 1063 | /* |
| 1064 | * global setpoint |
| 1065 | * |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1066 | * See comment for pos_ratio_polynom(). |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1067 | */ |
| 1068 | setpoint = (freerun + limit) / 2; |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1069 | pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit); |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1070 | |
| 1071 | /* |
| 1072 | * The strictlimit feature is a tool preventing mistrusted filesystems |
| 1073 | * from growing a large number of dirty pages before throttling. For |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1074 | * such filesystems balance_dirty_pages always checks wb counters |
| 1075 | * against wb limits. Even if global "nr_dirty" is under "freerun". |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1076 | * This is especially important for fuse which sets bdi->max_ratio to |
| 1077 | * 1% by default. Without strictlimit feature, fuse writeback may |
| 1078 | * consume arbitrary amount of RAM because it is accounted in |
| 1079 | * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty". |
| 1080 | * |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1081 | * Here, in wb_position_ratio(), we calculate pos_ratio based on |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1082 | * two values: wb_dirty and wb_thresh. Let's consider an example: |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1083 | * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global |
| 1084 | * limits are set by default to 10% and 20% (background and throttle). |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1085 | * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages. |
Tejun Heo | 0d960a3 | 2015-05-22 18:23:19 -0400 | [diff] [blame] | 1086 | * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1087 | * about ~6K pages (as the average of background and throttle wb |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1088 | * limits). The 3rd order polynomial will provide positive feedback if |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1089 | * wb_dirty is under wb_setpoint and vice versa. |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1090 | * |
| 1091 | * Note, that we cannot use global counters in these calculations |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1092 | * because we want to throttle process writing to a strictlimit wb |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1093 | * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB |
| 1094 | * in the example above). |
| 1095 | */ |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1096 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1097 | long long wb_pos_ratio; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1098 | |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1099 | if (dtc->wb_dirty < 8) { |
| 1100 | dtc->pos_ratio = min_t(long long, pos_ratio * 2, |
| 1101 | 2 << RATELIMIT_CALC_SHIFT); |
| 1102 | return; |
| 1103 | } |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1104 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1105 | if (dtc->wb_dirty >= wb_thresh) |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1106 | return; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1107 | |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1108 | wb_setpoint = dirty_freerun_ceiling(wb_thresh, |
| 1109 | dtc->wb_bg_thresh); |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1110 | |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1111 | if (wb_setpoint == 0 || wb_setpoint == wb_thresh) |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1112 | return; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1113 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1114 | wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty, |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1115 | wb_thresh); |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1116 | |
| 1117 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1118 | * Typically, for strictlimit case, wb_setpoint << setpoint |
| 1119 | * and pos_ratio >> wb_pos_ratio. In the other words global |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1120 | * state ("dirty") is not limiting factor and we have to |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1121 | * make decision based on wb counters. But there is an |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1122 | * important case when global pos_ratio should get precedence: |
| 1123 | * global limits are exceeded (e.g. due to activities on other |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1124 | * wb's) while given strictlimit wb is below limit. |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1125 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1126 | * "pos_ratio * wb_pos_ratio" would work for the case above, |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1127 | * but it would look too non-natural for the case of all |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1128 | * activity in the system coming from a single strictlimit wb |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1129 | * with bdi->max_ratio == 100%. |
| 1130 | * |
| 1131 | * Note that min() below somewhat changes the dynamics of the |
| 1132 | * control system. Normally, pos_ratio value can be well over 3 |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1133 | * (when globally we are at freerun and wb is well below wb |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1134 | * setpoint). Now the maximum pos_ratio in the same situation |
| 1135 | * is 2. We might want to tweak this if we observe the control |
| 1136 | * system is too slow to adapt. |
| 1137 | */ |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1138 | dtc->pos_ratio = min(pos_ratio, wb_pos_ratio); |
| 1139 | return; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1140 | } |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1141 | |
| 1142 | /* |
| 1143 | * We have computed basic pos_ratio above based on global situation. If |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1144 | * the wb is over/under its share of dirty pages, we want to scale |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1145 | * pos_ratio further down/up. That is done by the following mechanism. |
| 1146 | */ |
| 1147 | |
| 1148 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1149 | * wb setpoint |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1150 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1151 | * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint) |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1152 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1153 | * x_intercept - wb_dirty |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1154 | * := -------------------------- |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1155 | * x_intercept - wb_setpoint |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1156 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1157 | * The main wb control line is a linear function that subjects to |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1158 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1159 | * (1) f(wb_setpoint) = 1.0 |
| 1160 | * (2) k = - 1 / (8 * write_bw) (in single wb case) |
| 1161 | * or equally: x_intercept = wb_setpoint + 8 * write_bw |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1162 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1163 | * For single wb case, the dirty pages are observed to fluctuate |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1164 | * regularly within range |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1165 | * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2] |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1166 | * for various filesystems, where (2) can yield in a reasonable 12.5% |
| 1167 | * fluctuation range for pos_ratio. |
| 1168 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1169 | * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1170 | * own size, so move the slope over accordingly and choose a slope that |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1171 | * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh. |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1172 | */ |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1173 | if (unlikely(wb_thresh > dtc->thresh)) |
| 1174 | wb_thresh = dtc->thresh; |
Wu Fengguang | aed21ad | 2011-11-23 11:44:41 -0600 | [diff] [blame] | 1175 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1176 | * It's very possible that wb_thresh is close to 0 not because the |
Wu Fengguang | aed21ad | 2011-11-23 11:44:41 -0600 | [diff] [blame] | 1177 | * device is slow, but that it has remained inactive for long time. |
| 1178 | * Honour such devices a reasonable good (hopefully IO efficient) |
| 1179 | * threshold, so that the occasional writes won't be blocked and active |
| 1180 | * writes can rampup the threshold quickly. |
| 1181 | */ |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1182 | wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8); |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1183 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1184 | * scale global setpoint to wb's: |
| 1185 | * wb_setpoint = setpoint * wb_thresh / thresh |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1186 | */ |
Linus Torvalds | e4bc13a | 2015-06-25 16:00:17 -0700 | [diff] [blame] | 1187 | x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1); |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1188 | wb_setpoint = setpoint * (u64)x >> 16; |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1189 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1190 | * Use span=(8*write_bw) in single wb case as indicated by |
| 1191 | * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case. |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1192 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1193 | * wb_thresh thresh - wb_thresh |
| 1194 | * span = --------- * (8 * write_bw) + ------------------ * wb_thresh |
| 1195 | * thresh thresh |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1196 | */ |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1197 | span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16; |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1198 | x_intercept = wb_setpoint + span; |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1199 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1200 | if (dtc->wb_dirty < x_intercept - span / 4) { |
| 1201 | pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty), |
Linus Torvalds | e4bc13a | 2015-06-25 16:00:17 -0700 | [diff] [blame] | 1202 | (x_intercept - wb_setpoint) | 1); |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1203 | } else |
| 1204 | pos_ratio /= 4; |
| 1205 | |
Wu Fengguang | 8927f66 | 2011-08-04 22:16:46 -0600 | [diff] [blame] | 1206 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1207 | * wb reserve area, safeguard against dirty pool underrun and disk idle |
Wu Fengguang | 8927f66 | 2011-08-04 22:16:46 -0600 | [diff] [blame] | 1208 | * It may push the desired control point of global dirty pages higher |
| 1209 | * than setpoint. |
| 1210 | */ |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1211 | x_intercept = wb_thresh / 2; |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1212 | if (dtc->wb_dirty < x_intercept) { |
| 1213 | if (dtc->wb_dirty > x_intercept / 8) |
| 1214 | pos_ratio = div_u64(pos_ratio * x_intercept, |
| 1215 | dtc->wb_dirty); |
Wu Fengguang | 50657fc | 2011-10-11 17:06:33 -0600 | [diff] [blame] | 1216 | else |
Wu Fengguang | 8927f66 | 2011-08-04 22:16:46 -0600 | [diff] [blame] | 1217 | pos_ratio *= 8; |
| 1218 | } |
| 1219 | |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1220 | dtc->pos_ratio = pos_ratio; |
Wu Fengguang | 6c14ae1 | 2011-03-02 16:04:18 -0600 | [diff] [blame] | 1221 | } |
| 1222 | |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1223 | static void wb_update_write_bandwidth(struct bdi_writeback *wb, |
| 1224 | unsigned long elapsed, |
| 1225 | unsigned long written) |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1226 | { |
| 1227 | const unsigned long period = roundup_pow_of_two(3 * HZ); |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1228 | unsigned long avg = wb->avg_write_bandwidth; |
| 1229 | unsigned long old = wb->write_bandwidth; |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1230 | u64 bw; |
| 1231 | |
| 1232 | /* |
| 1233 | * bw = written * HZ / elapsed |
| 1234 | * |
| 1235 | * bw * elapsed + write_bandwidth * (period - elapsed) |
| 1236 | * write_bandwidth = --------------------------------------------------- |
| 1237 | * period |
Tejun Heo | c72efb6 | 2015-03-23 00:18:48 -0400 | [diff] [blame] | 1238 | * |
Christoph Hellwig | ed2da92 | 2023-06-28 17:31:44 +0200 | [diff] [blame] | 1239 | * @written may have decreased due to folio_redirty_for_writepage(). |
Tejun Heo | c72efb6 | 2015-03-23 00:18:48 -0400 | [diff] [blame] | 1240 | * Avoid underflowing @bw calculation. |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1241 | */ |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1242 | bw = written - min(written, wb->written_stamp); |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1243 | bw *= HZ; |
| 1244 | if (unlikely(elapsed > period)) { |
Wen Yang | 0a5d1a7 | 2020-01-13 16:29:29 -0800 | [diff] [blame] | 1245 | bw = div64_ul(bw, elapsed); |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1246 | avg = bw; |
| 1247 | goto out; |
| 1248 | } |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1249 | bw += (u64)wb->write_bandwidth * (period - elapsed); |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1250 | bw >>= ilog2(period); |
| 1251 | |
| 1252 | /* |
| 1253 | * one more level of smoothing, for filtering out sudden spikes |
| 1254 | */ |
| 1255 | if (avg > old && old >= (unsigned long)bw) |
| 1256 | avg -= (avg - old) >> 3; |
| 1257 | |
| 1258 | if (avg < old && old <= (unsigned long)bw) |
| 1259 | avg += (old - avg) >> 3; |
| 1260 | |
| 1261 | out: |
Tejun Heo | 95a46c6 | 2015-05-22 17:13:47 -0400 | [diff] [blame] | 1262 | /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */ |
| 1263 | avg = max(avg, 1LU); |
| 1264 | if (wb_has_dirty_io(wb)) { |
| 1265 | long delta = avg - wb->avg_write_bandwidth; |
| 1266 | WARN_ON_ONCE(atomic_long_add_return(delta, |
| 1267 | &wb->bdi->tot_write_bandwidth) <= 0); |
| 1268 | } |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1269 | wb->write_bandwidth = bw; |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1270 | WRITE_ONCE(wb->avg_write_bandwidth, avg); |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1271 | } |
| 1272 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1273 | static void update_dirty_limit(struct dirty_throttle_control *dtc) |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1274 | { |
Tejun Heo | e9f07df | 2015-05-22 18:23:28 -0400 | [diff] [blame] | 1275 | struct wb_domain *dom = dtc_dom(dtc); |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1276 | unsigned long thresh = dtc->thresh; |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 1277 | unsigned long limit = dom->dirty_limit; |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1278 | |
| 1279 | /* |
| 1280 | * Follow up in one step. |
| 1281 | */ |
| 1282 | if (limit < thresh) { |
| 1283 | limit = thresh; |
| 1284 | goto update; |
| 1285 | } |
| 1286 | |
| 1287 | /* |
| 1288 | * Follow down slowly. Use the higher one as the target, because thresh |
| 1289 | * may drop below dirty. This is exactly the reason to introduce |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 1290 | * dom->dirty_limit which is guaranteed to lie above the dirty pages. |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1291 | */ |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1292 | thresh = max(thresh, dtc->dirty); |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1293 | if (limit > thresh) { |
| 1294 | limit -= (limit - thresh) >> 5; |
| 1295 | goto update; |
| 1296 | } |
| 1297 | return; |
| 1298 | update: |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 1299 | dom->dirty_limit = limit; |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1300 | } |
| 1301 | |
Jan Kara | 42dd235 | 2021-09-02 14:53:12 -0700 | [diff] [blame] | 1302 | static void domain_update_dirty_limit(struct dirty_throttle_control *dtc, |
| 1303 | unsigned long now) |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1304 | { |
Tejun Heo | e9f07df | 2015-05-22 18:23:28 -0400 | [diff] [blame] | 1305 | struct wb_domain *dom = dtc_dom(dtc); |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1306 | |
| 1307 | /* |
| 1308 | * check locklessly first to optimize away locking for the most time |
| 1309 | */ |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 1310 | if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1311 | return; |
| 1312 | |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 1313 | spin_lock(&dom->lock); |
| 1314 | if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1315 | update_dirty_limit(dtc); |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 1316 | dom->dirty_limit_tstamp = now; |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1317 | } |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 1318 | spin_unlock(&dom->lock); |
Wu Fengguang | c42843f | 2011-03-02 15:54:09 -0600 | [diff] [blame] | 1319 | } |
| 1320 | |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1321 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1322 | * Maintain wb->dirty_ratelimit, the base dirty throttle rate. |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1323 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1324 | * Normal wb tasks will be curbed at or below it in long term. |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1325 | * Obviously it should be around (write_bw / N) when there are N dd tasks. |
| 1326 | */ |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1327 | static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc, |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1328 | unsigned long dirtied, |
| 1329 | unsigned long elapsed) |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1330 | { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1331 | struct bdi_writeback *wb = dtc->wb; |
| 1332 | unsigned long dirty = dtc->dirty; |
| 1333 | unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh); |
Tejun Heo | c798143 | 2015-05-22 18:23:29 -0400 | [diff] [blame] | 1334 | unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh); |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1335 | unsigned long setpoint = (freerun + limit) / 2; |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1336 | unsigned long write_bw = wb->avg_write_bandwidth; |
| 1337 | unsigned long dirty_ratelimit = wb->dirty_ratelimit; |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1338 | unsigned long dirty_rate; |
| 1339 | unsigned long task_ratelimit; |
| 1340 | unsigned long balanced_dirty_ratelimit; |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1341 | unsigned long step; |
| 1342 | unsigned long x; |
Andrey Ryabinin | d59b108 | 2016-03-15 14:55:27 -0700 | [diff] [blame] | 1343 | unsigned long shift; |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1344 | |
| 1345 | /* |
| 1346 | * The dirty rate will match the writeout rate in long term, except |
| 1347 | * when dirty pages are truncated by userspace or re-dirtied by FS. |
| 1348 | */ |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1349 | dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed; |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1350 | |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1351 | /* |
| 1352 | * task_ratelimit reflects each dd's dirty rate for the past 200ms. |
| 1353 | */ |
| 1354 | task_ratelimit = (u64)dirty_ratelimit * |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1355 | dtc->pos_ratio >> RATELIMIT_CALC_SHIFT; |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1356 | task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */ |
| 1357 | |
| 1358 | /* |
| 1359 | * A linear estimation of the "balanced" throttle rate. The theory is, |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1360 | * if there are N dd tasks, each throttled at task_ratelimit, the wb's |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1361 | * dirty_rate will be measured to be (N * task_ratelimit). So the below |
| 1362 | * formula will yield the balanced rate limit (write_bw / N). |
| 1363 | * |
| 1364 | * Note that the expanded form is not a pure rate feedback: |
| 1365 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1) |
| 1366 | * but also takes pos_ratio into account: |
| 1367 | * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2) |
| 1368 | * |
| 1369 | * (1) is not realistic because pos_ratio also takes part in balancing |
| 1370 | * the dirty rate. Consider the state |
| 1371 | * pos_ratio = 0.5 (3) |
| 1372 | * rate = 2 * (write_bw / N) (4) |
| 1373 | * If (1) is used, it will stuck in that state! Because each dd will |
| 1374 | * be throttled at |
| 1375 | * task_ratelimit = pos_ratio * rate = (write_bw / N) (5) |
| 1376 | * yielding |
| 1377 | * dirty_rate = N * task_ratelimit = write_bw (6) |
| 1378 | * put (6) into (1) we get |
| 1379 | * rate_(i+1) = rate_(i) (7) |
| 1380 | * |
| 1381 | * So we end up using (2) to always keep |
| 1382 | * rate_(i+1) ~= (write_bw / N) (8) |
| 1383 | * regardless of the value of pos_ratio. As long as (8) is satisfied, |
| 1384 | * pos_ratio is able to drive itself to 1.0, which is not only where |
| 1385 | * the dirty count meet the setpoint, but also where the slope of |
| 1386 | * pos_ratio is most flat and hence task_ratelimit is least fluctuated. |
| 1387 | */ |
| 1388 | balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw, |
| 1389 | dirty_rate | 1); |
Wu Fengguang | bdaac49 | 2011-08-03 14:30:36 -0600 | [diff] [blame] | 1390 | /* |
| 1391 | * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw |
| 1392 | */ |
| 1393 | if (unlikely(balanced_dirty_ratelimit > write_bw)) |
| 1394 | balanced_dirty_ratelimit = write_bw; |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1395 | |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1396 | /* |
| 1397 | * We could safely do this and return immediately: |
| 1398 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1399 | * wb->dirty_ratelimit = balanced_dirty_ratelimit; |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1400 | * |
| 1401 | * However to get a more stable dirty_ratelimit, the below elaborated |
Wanpeng Li | 331cbde | 2012-06-09 11:10:55 +0800 | [diff] [blame] | 1402 | * code makes use of task_ratelimit to filter out singular points and |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1403 | * limit the step size. |
| 1404 | * |
| 1405 | * The below code essentially only uses the relative value of |
| 1406 | * |
| 1407 | * task_ratelimit - dirty_ratelimit |
| 1408 | * = (pos_ratio - 1) * dirty_ratelimit |
| 1409 | * |
| 1410 | * which reflects the direction and size of dirty position error. |
| 1411 | */ |
| 1412 | |
| 1413 | /* |
| 1414 | * dirty_ratelimit will follow balanced_dirty_ratelimit iff |
| 1415 | * task_ratelimit is on the same side of dirty_ratelimit, too. |
| 1416 | * For example, when |
| 1417 | * - dirty_ratelimit > balanced_dirty_ratelimit |
| 1418 | * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint) |
| 1419 | * lowering dirty_ratelimit will help meet both the position and rate |
| 1420 | * control targets. Otherwise, don't update dirty_ratelimit if it will |
| 1421 | * only help meet the rate target. After all, what the users ultimately |
| 1422 | * feel and care are stable dirty rate and small position error. |
| 1423 | * |
| 1424 | * |task_ratelimit - dirty_ratelimit| is used to limit the step size |
Wanpeng Li | 331cbde | 2012-06-09 11:10:55 +0800 | [diff] [blame] | 1425 | * and filter out the singular points of balanced_dirty_ratelimit. Which |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1426 | * keeps jumping around randomly and can even leap far away at times |
| 1427 | * due to the small 200ms estimation period of dirty_rate (we want to |
| 1428 | * keep that period small to reduce time lags). |
| 1429 | */ |
| 1430 | step = 0; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1431 | |
| 1432 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1433 | * For strictlimit case, calculations above were based on wb counters |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1434 | * and limits (starting from pos_ratio = wb_position_ratio() and up to |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1435 | * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate). |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1436 | * Hence, to calculate "step" properly, we have to use wb_dirty as |
| 1437 | * "dirty" and wb_setpoint as "setpoint". |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1438 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1439 | * We rampup dirty_ratelimit forcibly if wb_dirty is low because |
| 1440 | * it's possible that wb_thresh is close to zero due to inactivity |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1441 | * of backing device. |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1442 | */ |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1443 | if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1444 | dirty = dtc->wb_dirty; |
| 1445 | if (dtc->wb_dirty < 8) |
| 1446 | setpoint = dtc->wb_dirty + 1; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1447 | else |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1448 | setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1449 | } |
| 1450 | |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1451 | if (dirty < setpoint) { |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1452 | x = min3(wb->balanced_dirty_ratelimit, |
Mark Rustad | 7c80996 | 2014-10-09 15:28:15 -0700 | [diff] [blame] | 1453 | balanced_dirty_ratelimit, task_ratelimit); |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1454 | if (dirty_ratelimit < x) |
| 1455 | step = x - dirty_ratelimit; |
| 1456 | } else { |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1457 | x = max3(wb->balanced_dirty_ratelimit, |
Mark Rustad | 7c80996 | 2014-10-09 15:28:15 -0700 | [diff] [blame] | 1458 | balanced_dirty_ratelimit, task_ratelimit); |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1459 | if (dirty_ratelimit > x) |
| 1460 | step = dirty_ratelimit - x; |
| 1461 | } |
| 1462 | |
| 1463 | /* |
| 1464 | * Don't pursue 100% rate matching. It's impossible since the balanced |
| 1465 | * rate itself is constantly fluctuating. So decrease the track speed |
| 1466 | * when it gets close to the target. Helps eliminate pointless tremors. |
| 1467 | */ |
Andrey Ryabinin | d59b108 | 2016-03-15 14:55:27 -0700 | [diff] [blame] | 1468 | shift = dirty_ratelimit / (2 * step + 1); |
| 1469 | if (shift < BITS_PER_LONG) |
| 1470 | step = DIV_ROUND_UP(step >> shift, 8); |
| 1471 | else |
| 1472 | step = 0; |
Wu Fengguang | 7381131 | 2011-08-26 15:53:24 -0600 | [diff] [blame] | 1473 | |
| 1474 | if (dirty_ratelimit < balanced_dirty_ratelimit) |
| 1475 | dirty_ratelimit += step; |
| 1476 | else |
| 1477 | dirty_ratelimit -= step; |
| 1478 | |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1479 | WRITE_ONCE(wb->dirty_ratelimit, max(dirty_ratelimit, 1UL)); |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1480 | wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit; |
Wu Fengguang | b48c104 | 2011-03-02 17:22:49 -0600 | [diff] [blame] | 1481 | |
Tejun Heo | 5634cc2 | 2015-08-18 14:54:56 -0700 | [diff] [blame] | 1482 | trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit); |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1483 | } |
| 1484 | |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1485 | static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc, |
| 1486 | struct dirty_throttle_control *mdtc, |
Tejun Heo | 8a73179 | 2015-05-22 18:23:20 -0400 | [diff] [blame] | 1487 | bool update_ratelimit) |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1488 | { |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1489 | struct bdi_writeback *wb = gdtc->wb; |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1490 | unsigned long now = jiffies; |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 1491 | unsigned long elapsed; |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1492 | unsigned long dirtied; |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1493 | unsigned long written; |
| 1494 | |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 1495 | spin_lock(&wb->list_lock); |
Tejun Heo | 8a73179 | 2015-05-22 18:23:20 -0400 | [diff] [blame] | 1496 | |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1497 | /* |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 1498 | * Lockless checks for elapsed time are racy and delayed update after |
| 1499 | * IO completion doesn't do it at all (to make sure written pages are |
| 1500 | * accounted reasonably quickly). Make sure elapsed >= 1 to avoid |
| 1501 | * division errors. |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1502 | */ |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 1503 | elapsed = max(now - wb->bw_time_stamp, 1UL); |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1504 | dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]); |
| 1505 | written = percpu_counter_read(&wb->stat[WB_WRITTEN]); |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1506 | |
Tejun Heo | 8a73179 | 2015-05-22 18:23:20 -0400 | [diff] [blame] | 1507 | if (update_ratelimit) { |
Jan Kara | 42dd235 | 2021-09-02 14:53:12 -0700 | [diff] [blame] | 1508 | domain_update_dirty_limit(gdtc, now); |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1509 | wb_update_dirty_ratelimit(gdtc, dirtied, elapsed); |
| 1510 | |
| 1511 | /* |
| 1512 | * @mdtc is always NULL if !CGROUP_WRITEBACK but the |
| 1513 | * compiler has no way to figure that out. Help it. |
| 1514 | */ |
| 1515 | if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) { |
Jan Kara | 42dd235 | 2021-09-02 14:53:12 -0700 | [diff] [blame] | 1516 | domain_update_dirty_limit(mdtc, now); |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1517 | wb_update_dirty_ratelimit(mdtc, dirtied, elapsed); |
| 1518 | } |
Wu Fengguang | be3ffa2 | 2011-06-12 10:51:31 -0600 | [diff] [blame] | 1519 | } |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1520 | wb_update_write_bandwidth(wb, elapsed, written); |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1521 | |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1522 | wb->dirtied_stamp = dirtied; |
| 1523 | wb->written_stamp = written; |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1524 | WRITE_ONCE(wb->bw_time_stamp, now); |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 1525 | spin_unlock(&wb->list_lock); |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1526 | } |
| 1527 | |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 1528 | void wb_update_bandwidth(struct bdi_writeback *wb) |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1529 | { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1530 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
| 1531 | |
Jan Kara | fee468f | 2021-09-02 14:53:06 -0700 | [diff] [blame] | 1532 | __wb_update_bandwidth(&gdtc, NULL, false); |
Jan Kara | fee468f | 2021-09-02 14:53:06 -0700 | [diff] [blame] | 1533 | } |
| 1534 | |
| 1535 | /* Interval after which we consider wb idle and don't estimate bandwidth */ |
| 1536 | #define WB_BANDWIDTH_IDLE_JIF (HZ) |
| 1537 | |
| 1538 | static void wb_bandwidth_estimate_start(struct bdi_writeback *wb) |
| 1539 | { |
| 1540 | unsigned long now = jiffies; |
| 1541 | unsigned long elapsed = now - READ_ONCE(wb->bw_time_stamp); |
| 1542 | |
| 1543 | if (elapsed > WB_BANDWIDTH_IDLE_JIF && |
| 1544 | !atomic_read(&wb->writeback_inodes)) { |
| 1545 | spin_lock(&wb->list_lock); |
| 1546 | wb->dirtied_stamp = wb_stat(wb, WB_DIRTIED); |
| 1547 | wb->written_stamp = wb_stat(wb, WB_WRITTEN); |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1548 | WRITE_ONCE(wb->bw_time_stamp, now); |
Jan Kara | fee468f | 2021-09-02 14:53:06 -0700 | [diff] [blame] | 1549 | spin_unlock(&wb->list_lock); |
| 1550 | } |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1551 | } |
| 1552 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1553 | /* |
Namjae Jeon | d0e1d66 | 2012-12-11 16:00:21 -0800 | [diff] [blame] | 1554 | * After a task dirtied this many pages, balance_dirty_pages_ratelimited() |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 1555 | * will look to see if it needs to start dirty throttling. |
| 1556 | * |
| 1557 | * If dirty_poll_interval is too low, big NUMA machines will call the expensive |
Michal Hocko | c41f012 | 2017-09-06 16:23:36 -0700 | [diff] [blame] | 1558 | * global_zone_page_state() too often. So scale it near-sqrt to the safety margin |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 1559 | * (the number of pages we may dirty without exceeding the dirty limits). |
| 1560 | */ |
| 1561 | static unsigned long dirty_poll_interval(unsigned long dirty, |
| 1562 | unsigned long thresh) |
| 1563 | { |
| 1564 | if (thresh > dirty) |
| 1565 | return 1UL << (ilog2(thresh - dirty) >> 1); |
| 1566 | |
| 1567 | return 1; |
| 1568 | } |
| 1569 | |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1570 | static unsigned long wb_max_pause(struct bdi_writeback *wb, |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1571 | unsigned long wb_dirty) |
Wu Fengguang | c8462cc | 2011-06-11 19:21:43 -0600 | [diff] [blame] | 1572 | { |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1573 | unsigned long bw = READ_ONCE(wb->avg_write_bandwidth); |
Fengguang Wu | e3b6c65 | 2013-10-16 13:47:03 -0700 | [diff] [blame] | 1574 | unsigned long t; |
Wu Fengguang | c8462cc | 2011-06-11 19:21:43 -0600 | [diff] [blame] | 1575 | |
| 1576 | /* |
| 1577 | * Limit pause time for small memory systems. If sleeping for too long |
| 1578 | * time, a small pool of dirty/writeback pages may go empty and disk go |
| 1579 | * idle. |
| 1580 | * |
| 1581 | * 8 serves as the safety ratio. |
| 1582 | */ |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1583 | t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8)); |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1584 | t++; |
| 1585 | |
Fengguang Wu | e3b6c65 | 2013-10-16 13:47:03 -0700 | [diff] [blame] | 1586 | return min_t(unsigned long, t, MAX_PAUSE); |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1587 | } |
| 1588 | |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1589 | static long wb_min_pause(struct bdi_writeback *wb, |
| 1590 | long max_pause, |
| 1591 | unsigned long task_ratelimit, |
| 1592 | unsigned long dirty_ratelimit, |
| 1593 | int *nr_dirtied_pause) |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1594 | { |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1595 | long hi = ilog2(READ_ONCE(wb->avg_write_bandwidth)); |
| 1596 | long lo = ilog2(READ_ONCE(wb->dirty_ratelimit)); |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1597 | long t; /* target pause */ |
| 1598 | long pause; /* estimated next pause */ |
| 1599 | int pages; /* target nr_dirtied_pause */ |
| 1600 | |
| 1601 | /* target for 10ms pause on 1-dd case */ |
| 1602 | t = max(1, HZ / 100); |
Wu Fengguang | c8462cc | 2011-06-11 19:21:43 -0600 | [diff] [blame] | 1603 | |
| 1604 | /* |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1605 | * Scale up pause time for concurrent dirtiers in order to reduce CPU |
| 1606 | * overheads. |
| 1607 | * |
| 1608 | * (N * 10ms) on 2^N concurrent tasks. |
Wu Fengguang | c8462cc | 2011-06-11 19:21:43 -0600 | [diff] [blame] | 1609 | */ |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1610 | if (hi > lo) |
| 1611 | t += (hi - lo) * (10 * HZ) / 1024; |
| 1612 | |
| 1613 | /* |
| 1614 | * This is a bit convoluted. We try to base the next nr_dirtied_pause |
| 1615 | * on the much more stable dirty_ratelimit. However the next pause time |
| 1616 | * will be computed based on task_ratelimit and the two rate limits may |
| 1617 | * depart considerably at some time. Especially if task_ratelimit goes |
| 1618 | * below dirty_ratelimit/2 and the target pause is max_pause, the next |
| 1619 | * pause time will be max_pause*2 _trimmed down_ to max_pause. As a |
| 1620 | * result task_ratelimit won't be executed faithfully, which could |
| 1621 | * eventually bring down dirty_ratelimit. |
| 1622 | * |
| 1623 | * We apply two rules to fix it up: |
| 1624 | * 1) try to estimate the next pause time and if necessary, use a lower |
| 1625 | * nr_dirtied_pause so as not to exceed max_pause. When this happens, |
| 1626 | * nr_dirtied_pause will be "dancing" with task_ratelimit. |
| 1627 | * 2) limit the target pause time to max_pause/2, so that the normal |
| 1628 | * small fluctuations of task_ratelimit won't trigger rule (1) and |
| 1629 | * nr_dirtied_pause will remain as stable as dirty_ratelimit. |
| 1630 | */ |
| 1631 | t = min(t, 1 + max_pause / 2); |
| 1632 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); |
| 1633 | |
Wu Fengguang | 5b9b357 | 2011-12-06 13:17:17 -0600 | [diff] [blame] | 1634 | /* |
| 1635 | * Tiny nr_dirtied_pause is found to hurt I/O performance in the test |
| 1636 | * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}. |
| 1637 | * When the 16 consecutive reads are often interrupted by some dirty |
| 1638 | * throttling pause during the async writes, cfq will go into idles |
| 1639 | * (deadline is fine). So push nr_dirtied_pause as high as possible |
| 1640 | * until reaches DIRTY_POLL_THRESH=32 pages. |
| 1641 | */ |
| 1642 | if (pages < DIRTY_POLL_THRESH) { |
| 1643 | t = max_pause; |
| 1644 | pages = dirty_ratelimit * t / roundup_pow_of_two(HZ); |
| 1645 | if (pages > DIRTY_POLL_THRESH) { |
| 1646 | pages = DIRTY_POLL_THRESH; |
| 1647 | t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit; |
| 1648 | } |
| 1649 | } |
| 1650 | |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1651 | pause = HZ * pages / (task_ratelimit + 1); |
| 1652 | if (pause > max_pause) { |
| 1653 | t = max_pause; |
| 1654 | pages = task_ratelimit * t / roundup_pow_of_two(HZ); |
| 1655 | } |
| 1656 | |
| 1657 | *nr_dirtied_pause = pages; |
| 1658 | /* |
| 1659 | * The minimal pause time will normally be half the target pause time. |
| 1660 | */ |
Wu Fengguang | 5b9b357 | 2011-12-06 13:17:17 -0600 | [diff] [blame] | 1661 | return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t; |
Wu Fengguang | c8462cc | 2011-06-11 19:21:43 -0600 | [diff] [blame] | 1662 | } |
| 1663 | |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1664 | static inline void wb_dirty_limits(struct dirty_throttle_control *dtc) |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1665 | { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1666 | struct bdi_writeback *wb = dtc->wb; |
Tejun Heo | 93f78d8 | 2015-05-22 17:13:27 -0400 | [diff] [blame] | 1667 | unsigned long wb_reclaimable; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1668 | |
| 1669 | /* |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1670 | * wb_thresh is not treated as some limiting factor as |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1671 | * dirty_thresh, due to reasons |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1672 | * - in JBOD setup, wb_thresh can fluctuate a lot |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1673 | * - in a system with HDD and USB key, the USB key may somehow |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1674 | * go into state (wb_dirty >> wb_thresh) either because |
| 1675 | * wb_dirty starts high, or because wb_thresh drops low. |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1676 | * In this case we don't want to hard throttle the USB key |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1677 | * dirtiers for 100 seconds until wb_dirty drops under |
| 1678 | * wb_thresh. Instead the auxiliary wb control line in |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1679 | * wb_position_ratio() will let the dirtier task progress |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1680 | * at some rate <= (write_bw / 2) for bringing down wb_dirty. |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1681 | */ |
Kemeng Shi | 13fc441 | 2024-04-25 21:17:21 +0800 | [diff] [blame] | 1682 | dtc->wb_thresh = __wb_calc_thresh(dtc, dtc->thresh); |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1683 | dtc->wb_bg_thresh = dtc->thresh ? |
Zach O'Keefe | 9319b64 | 2024-01-18 10:19:53 -0800 | [diff] [blame] | 1684 | div64_u64(dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1685 | |
| 1686 | /* |
| 1687 | * In order to avoid the stacked BDI deadlock we need |
| 1688 | * to ensure we accurately count the 'dirty' pages when |
| 1689 | * the threshold is low. |
| 1690 | * |
| 1691 | * Otherwise it would be possible to get thresh+n pages |
| 1692 | * reported dirty, even though there are thresh-m pages |
| 1693 | * actually dirty; with m+n sitting in the percpu |
| 1694 | * deltas. |
| 1695 | */ |
Wang Long | 2bce774 | 2017-11-15 17:39:03 -0800 | [diff] [blame] | 1696 | if (dtc->wb_thresh < 2 * wb_stat_error()) { |
Tejun Heo | 93f78d8 | 2015-05-22 17:13:27 -0400 | [diff] [blame] | 1697 | wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE); |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1698 | dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK); |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1699 | } else { |
Tejun Heo | 93f78d8 | 2015-05-22 17:13:27 -0400 | [diff] [blame] | 1700 | wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE); |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1701 | dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK); |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1702 | } |
| 1703 | } |
| 1704 | |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 1705 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1706 | * balance_dirty_pages() must be called by processes which are generating dirty |
| 1707 | * data. It looks at the number of dirty pages in the machine and will force |
Wu Fengguang | 143dfe8 | 2010-08-27 18:45:12 -0600 | [diff] [blame] | 1708 | * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2. |
Jens Axboe | 5b0830c | 2009-09-23 19:37:09 +0200 | [diff] [blame] | 1709 | * If we're over `background_thresh' then the writeback threads are woken to |
| 1710 | * perform some writeout. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1711 | */ |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 1712 | static int balance_dirty_pages(struct bdi_writeback *wb, |
| 1713 | unsigned long pages_dirtied, unsigned int flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1714 | { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1715 | struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) }; |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1716 | struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) }; |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1717 | struct dirty_throttle_control * const gdtc = &gdtc_stor; |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1718 | struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ? |
| 1719 | &mdtc_stor : NULL; |
| 1720 | struct dirty_throttle_control *sdtc; |
Kemeng Shi | 826881a | 2024-04-23 11:46:43 +0800 | [diff] [blame] | 1721 | unsigned long nr_dirty; |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1722 | long period; |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1723 | long pause; |
| 1724 | long max_pause; |
| 1725 | long min_pause; |
| 1726 | int nr_dirtied_pause; |
Wu Fengguang | e50e372 | 2010-08-11 14:17:37 -0700 | [diff] [blame] | 1727 | bool dirty_exceeded = false; |
Wu Fengguang | 143dfe8 | 2010-08-27 18:45:12 -0600 | [diff] [blame] | 1728 | unsigned long task_ratelimit; |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1729 | unsigned long dirty_ratelimit; |
Tejun Heo | dfb8ae5 | 2015-05-22 17:13:40 -0400 | [diff] [blame] | 1730 | struct backing_dev_info *bdi = wb->bdi; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1731 | bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT; |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1732 | unsigned long start_time = jiffies; |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 1733 | int ret = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1734 | |
| 1735 | for (;;) { |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1736 | unsigned long now = jiffies; |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1737 | unsigned long dirty, thresh, bg_thresh; |
Yang Shi | 50e55bf | 2015-11-20 15:57:10 -0800 | [diff] [blame] | 1738 | unsigned long m_dirty = 0; /* stop bogus uninit warnings */ |
| 1739 | unsigned long m_thresh = 0; |
| 1740 | unsigned long m_bg_thresh = 0; |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1741 | |
Kemeng Shi | 826881a | 2024-04-23 11:46:43 +0800 | [diff] [blame] | 1742 | nr_dirty = global_node_page_state(NR_FILE_DIRTY); |
Peter Zijlstra | 5fce25a | 2007-11-14 16:59:15 -0800 | [diff] [blame] | 1743 | |
Kemeng Shi | ba62d5c | 2024-05-14 20:52:48 +0800 | [diff] [blame] | 1744 | domain_dirty_avail(gdtc, true); |
Tejun Heo | 9fc3a43 | 2015-05-22 18:23:30 -0400 | [diff] [blame] | 1745 | domain_dirty_limits(gdtc); |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 1746 | |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1747 | if (unlikely(strictlimit)) { |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1748 | wb_dirty_limits(gdtc); |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1749 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1750 | dirty = gdtc->wb_dirty; |
| 1751 | thresh = gdtc->wb_thresh; |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1752 | bg_thresh = gdtc->wb_bg_thresh; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1753 | } else { |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1754 | dirty = gdtc->dirty; |
| 1755 | thresh = gdtc->thresh; |
| 1756 | bg_thresh = gdtc->bg_thresh; |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1757 | } |
| 1758 | |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1759 | if (mdtc) { |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1760 | /* |
| 1761 | * If @wb belongs to !root memcg, repeat the same |
| 1762 | * basic calculations for the memcg domain. |
| 1763 | */ |
Kemeng Shi | ba62d5c | 2024-05-14 20:52:48 +0800 | [diff] [blame] | 1764 | domain_dirty_avail(mdtc, true); |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1765 | domain_dirty_limits(mdtc); |
| 1766 | |
| 1767 | if (unlikely(strictlimit)) { |
| 1768 | wb_dirty_limits(mdtc); |
| 1769 | m_dirty = mdtc->wb_dirty; |
| 1770 | m_thresh = mdtc->wb_thresh; |
| 1771 | m_bg_thresh = mdtc->wb_bg_thresh; |
| 1772 | } else { |
| 1773 | m_dirty = mdtc->dirty; |
| 1774 | m_thresh = mdtc->thresh; |
| 1775 | m_bg_thresh = mdtc->bg_thresh; |
| 1776 | } |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 1777 | } |
| 1778 | |
| 1779 | /* |
Jan Kara | ea6813b | 2022-06-23 10:51:44 -0700 | [diff] [blame] | 1780 | * In laptop mode, we wait until hitting the higher threshold |
| 1781 | * before starting background writeout, and then write out all |
| 1782 | * the way down to the lower threshold. So slow writers cause |
| 1783 | * minimal disk activity. |
| 1784 | * |
| 1785 | * In normal mode, we start background writeout at the lower |
| 1786 | * background_thresh, to keep the amount of dirty memory low. |
| 1787 | */ |
Kemeng Shi | 826881a | 2024-04-23 11:46:43 +0800 | [diff] [blame] | 1788 | if (!laptop_mode && nr_dirty > gdtc->bg_thresh && |
Jan Kara | ea6813b | 2022-06-23 10:51:44 -0700 | [diff] [blame] | 1789 | !writeback_in_progress(wb)) |
| 1790 | wb_start_background_writeback(wb); |
| 1791 | |
| 1792 | /* |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 1793 | * Throttle it only when the background writeback cannot |
| 1794 | * catch-up. This avoids (excessively) small writeouts |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1795 | * when the wb limits are ramping up in case of !strictlimit. |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1796 | * |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1797 | * In strictlimit case make decision based on the wb counters |
| 1798 | * and limits. Small writeouts when the wb limits are ramping |
Maxim Patlasov | 5a53748 | 2013-09-11 14:22:46 -0700 | [diff] [blame] | 1799 | * up are the price we consciously pay for strictlimit-ing. |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1800 | * |
| 1801 | * If memcg domain is in effect, @dirty should be under |
| 1802 | * both global and memcg freerun ceilings. |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 1803 | */ |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1804 | if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) && |
| 1805 | (!mdtc || |
| 1806 | m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) { |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 1807 | unsigned long intv; |
| 1808 | unsigned long m_intv; |
| 1809 | |
| 1810 | free_running: |
| 1811 | intv = dirty_poll_interval(dirty, thresh); |
| 1812 | m_intv = ULONG_MAX; |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1813 | |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1814 | current->dirty_paused_when = now; |
| 1815 | current->nr_dirtied = 0; |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1816 | if (mdtc) |
| 1817 | m_intv = dirty_poll_interval(m_dirty, m_thresh); |
| 1818 | current->nr_dirtied_pause = min(intv, m_intv); |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 1819 | break; |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1820 | } |
Wu Fengguang | 16c4042 | 2010-08-11 14:17:39 -0700 | [diff] [blame] | 1821 | |
Jan Kara | ea6813b | 2022-06-23 10:51:44 -0700 | [diff] [blame] | 1822 | /* Start writeback even when in laptop mode */ |
Tejun Heo | bc05873 | 2015-05-22 17:13:53 -0400 | [diff] [blame] | 1823 | if (unlikely(!writeback_in_progress(wb))) |
Tejun Heo | 9ecf4866 | 2015-05-22 17:13:54 -0400 | [diff] [blame] | 1824 | wb_start_background_writeback(wb); |
Wu Fengguang | 143dfe8 | 2010-08-27 18:45:12 -0600 | [diff] [blame] | 1825 | |
Tejun Heo | 97b2782 | 2019-08-26 09:06:56 -0700 | [diff] [blame] | 1826 | mem_cgroup_flush_foreign(wb); |
| 1827 | |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1828 | /* |
| 1829 | * Calculate global domain's pos_ratio and select the |
| 1830 | * global dtc by default. |
| 1831 | */ |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 1832 | if (!strictlimit) { |
Tejun Heo | 970fb01 | 2015-05-22 18:23:24 -0400 | [diff] [blame] | 1833 | wb_dirty_limits(gdtc); |
Peter Zijlstra | 5fce25a | 2007-11-14 16:59:15 -0800 | [diff] [blame] | 1834 | |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 1835 | if ((current->flags & PF_LOCAL_THROTTLE) && |
| 1836 | gdtc->wb_dirty < |
| 1837 | dirty_freerun_ceiling(gdtc->wb_thresh, |
| 1838 | gdtc->wb_bg_thresh)) |
| 1839 | /* |
| 1840 | * LOCAL_THROTTLE tasks must not be throttled |
| 1841 | * when below the per-wb freerun ceiling. |
| 1842 | */ |
| 1843 | goto free_running; |
| 1844 | } |
| 1845 | |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1846 | dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) && |
| 1847 | ((gdtc->dirty > gdtc->thresh) || strictlimit); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1848 | |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1849 | wb_position_ratio(gdtc); |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1850 | sdtc = gdtc; |
Wu Fengguang | e98be2d | 2010-08-29 11:22:30 -0600 | [diff] [blame] | 1851 | |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1852 | if (mdtc) { |
| 1853 | /* |
| 1854 | * If memcg domain is in effect, calculate its |
| 1855 | * pos_ratio. @wb should satisfy constraints from |
| 1856 | * both global and memcg domains. Choose the one |
| 1857 | * w/ lower pos_ratio. |
| 1858 | */ |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 1859 | if (!strictlimit) { |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1860 | wb_dirty_limits(mdtc); |
| 1861 | |
NeilBrown | a37b071 | 2020-06-01 21:48:18 -0700 | [diff] [blame] | 1862 | if ((current->flags & PF_LOCAL_THROTTLE) && |
| 1863 | mdtc->wb_dirty < |
| 1864 | dirty_freerun_ceiling(mdtc->wb_thresh, |
| 1865 | mdtc->wb_bg_thresh)) |
| 1866 | /* |
| 1867 | * LOCAL_THROTTLE tasks must not be |
| 1868 | * throttled when below the per-wb |
| 1869 | * freerun ceiling. |
| 1870 | */ |
| 1871 | goto free_running; |
| 1872 | } |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1873 | dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) && |
| 1874 | ((mdtc->dirty > mdtc->thresh) || strictlimit); |
| 1875 | |
| 1876 | wb_position_ratio(mdtc); |
| 1877 | if (mdtc->pos_ratio < gdtc->pos_ratio) |
| 1878 | sdtc = mdtc; |
| 1879 | } |
Tejun Heo | daddfa3 | 2015-05-22 18:23:26 -0400 | [diff] [blame] | 1880 | |
Jan Kara | e92eebbb | 2022-06-23 10:51:45 -0700 | [diff] [blame] | 1881 | if (dirty_exceeded != wb->dirty_exceeded) |
| 1882 | wb->dirty_exceeded = dirty_exceeded; |
Peter Zijlstra | 04fbfdc | 2007-10-16 23:25:50 -0700 | [diff] [blame] | 1883 | |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1884 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 1885 | BANDWIDTH_INTERVAL)) |
Jan Kara | fee468f | 2021-09-02 14:53:06 -0700 | [diff] [blame] | 1886 | __wb_update_bandwidth(gdtc, mdtc, true); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1887 | |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1888 | /* throttle according to the chosen dtc */ |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 1889 | dirty_ratelimit = READ_ONCE(wb->dirty_ratelimit); |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1890 | task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >> |
Wu Fengguang | 3a73dbb | 2011-11-07 19:19:28 +0800 | [diff] [blame] | 1891 | RATELIMIT_CALC_SHIFT; |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1892 | max_pause = wb_max_pause(wb, sdtc->wb_dirty); |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 1893 | min_pause = wb_min_pause(wb, max_pause, |
| 1894 | task_ratelimit, dirty_ratelimit, |
| 1895 | &nr_dirtied_pause); |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1896 | |
Wu Fengguang | 3a73dbb | 2011-11-07 19:19:28 +0800 | [diff] [blame] | 1897 | if (unlikely(task_ratelimit == 0)) { |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1898 | period = max_pause; |
Wu Fengguang | c8462cc | 2011-06-11 19:21:43 -0600 | [diff] [blame] | 1899 | pause = max_pause; |
Wu Fengguang | 143dfe8 | 2010-08-27 18:45:12 -0600 | [diff] [blame] | 1900 | goto pause; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1901 | } |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1902 | period = HZ * pages_dirtied / task_ratelimit; |
| 1903 | pause = period; |
| 1904 | if (current->dirty_paused_when) |
| 1905 | pause -= now - current->dirty_paused_when; |
| 1906 | /* |
| 1907 | * For less than 1s think time (ext3/4 may block the dirtier |
| 1908 | * for up to 800ms from time to time on 1-HDD; so does xfs, |
| 1909 | * however at much less frequency), try to compensate it in |
| 1910 | * future periods by updating the virtual time; otherwise just |
| 1911 | * do a reset, as it may be a light dirtier. |
| 1912 | */ |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1913 | if (pause < min_pause) { |
Tejun Heo | 5634cc2 | 2015-08-18 14:54:56 -0700 | [diff] [blame] | 1914 | trace_balance_dirty_pages(wb, |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1915 | sdtc->thresh, |
| 1916 | sdtc->bg_thresh, |
| 1917 | sdtc->dirty, |
| 1918 | sdtc->wb_thresh, |
| 1919 | sdtc->wb_dirty, |
Wu Fengguang | ece13ac3 | 2010-08-29 23:33:20 -0600 | [diff] [blame] | 1920 | dirty_ratelimit, |
| 1921 | task_ratelimit, |
| 1922 | pages_dirtied, |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1923 | period, |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1924 | min(pause, 0L), |
Wu Fengguang | ece13ac3 | 2010-08-29 23:33:20 -0600 | [diff] [blame] | 1925 | start_time); |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1926 | if (pause < -HZ) { |
| 1927 | current->dirty_paused_when = now; |
| 1928 | current->nr_dirtied = 0; |
| 1929 | } else if (period) { |
| 1930 | current->dirty_paused_when += period; |
| 1931 | current->nr_dirtied = 0; |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1932 | } else if (current->nr_dirtied_pause <= pages_dirtied) |
| 1933 | current->nr_dirtied_pause += pages_dirtied; |
Wu Fengguang | 57fc978 | 2011-06-11 19:32:32 -0600 | [diff] [blame] | 1934 | break; |
| 1935 | } |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1936 | if (unlikely(pause > max_pause)) { |
| 1937 | /* for occasional dropped task_ratelimit */ |
| 1938 | now += min(pause - max_pause, max_pause); |
| 1939 | pause = max_pause; |
| 1940 | } |
Wu Fengguang | 143dfe8 | 2010-08-27 18:45:12 -0600 | [diff] [blame] | 1941 | |
| 1942 | pause: |
Tejun Heo | 5634cc2 | 2015-08-18 14:54:56 -0700 | [diff] [blame] | 1943 | trace_balance_dirty_pages(wb, |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 1944 | sdtc->thresh, |
| 1945 | sdtc->bg_thresh, |
| 1946 | sdtc->dirty, |
| 1947 | sdtc->wb_thresh, |
| 1948 | sdtc->wb_dirty, |
Wu Fengguang | ece13ac3 | 2010-08-29 23:33:20 -0600 | [diff] [blame] | 1949 | dirty_ratelimit, |
| 1950 | task_ratelimit, |
| 1951 | pages_dirtied, |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1952 | period, |
Wu Fengguang | ece13ac3 | 2010-08-29 23:33:20 -0600 | [diff] [blame] | 1953 | pause, |
| 1954 | start_time); |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 1955 | if (flags & BDP_ASYNC) { |
| 1956 | ret = -EAGAIN; |
| 1957 | break; |
| 1958 | } |
Jan Kara | 499d05e | 2011-11-16 19:34:48 +0800 | [diff] [blame] | 1959 | __set_current_state(TASK_KILLABLE); |
Jan Kara | f814bdd | 2024-01-23 18:58:26 +0100 | [diff] [blame] | 1960 | bdi->last_bdp_sleep = jiffies; |
Wu Fengguang | d25105e | 2009-10-09 12:40:42 +0200 | [diff] [blame] | 1961 | io_schedule_timeout(pause); |
Jens Axboe | 87c6a9b | 2009-09-17 19:59:14 +0200 | [diff] [blame] | 1962 | |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1963 | current->dirty_paused_when = now + pause; |
| 1964 | current->nr_dirtied = 0; |
Wu Fengguang | 7ccb9ad | 2011-11-30 11:08:55 -0600 | [diff] [blame] | 1965 | current->nr_dirtied_pause = nr_dirtied_pause; |
Wu Fengguang | 8371235 | 2011-06-11 19:25:42 -0600 | [diff] [blame] | 1966 | |
Wu Fengguang | ffd1f60 | 2011-06-19 22:18:42 -0600 | [diff] [blame] | 1967 | /* |
Tejun Heo | 2bc00ae | 2015-05-22 18:23:23 -0400 | [diff] [blame] | 1968 | * This is typically equal to (dirty < thresh) and can also |
| 1969 | * keep "1000+ dd on a slow USB stick" under control. |
Wu Fengguang | ffd1f60 | 2011-06-19 22:18:42 -0600 | [diff] [blame] | 1970 | */ |
Wu Fengguang | 1df64719 | 2011-11-13 19:47:32 -0600 | [diff] [blame] | 1971 | if (task_ratelimit) |
Wu Fengguang | ffd1f60 | 2011-06-19 22:18:42 -0600 | [diff] [blame] | 1972 | break; |
Jan Kara | 499d05e | 2011-11-16 19:34:48 +0800 | [diff] [blame] | 1973 | |
Wu Fengguang | c5c6343 | 2011-12-02 10:21:33 -0600 | [diff] [blame] | 1974 | /* |
Ingo Molnar | f0953a1 | 2021-05-06 18:06:47 -0700 | [diff] [blame] | 1975 | * In the case of an unresponsive NFS server and the NFS dirty |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1976 | * pages exceeds dirty_thresh, give the other good wb's a pipe |
Wu Fengguang | c5c6343 | 2011-12-02 10:21:33 -0600 | [diff] [blame] | 1977 | * to go through, so that tasks on them still remain responsive. |
| 1978 | * |
Masahiro Yamada | 3f8b6fb | 2017-02-27 14:29:25 -0800 | [diff] [blame] | 1979 | * In theory 1 page is enough to keep the consumer-producer |
Wu Fengguang | c5c6343 | 2011-12-02 10:21:33 -0600 | [diff] [blame] | 1980 | * pipe going: the flusher cleans 1 page => the task dirties 1 |
Tejun Heo | de1fff3 | 2015-05-22 17:13:29 -0400 | [diff] [blame] | 1981 | * more page. However wb_dirty has accounting errors. So use |
Tejun Heo | 93f78d8 | 2015-05-22 17:13:27 -0400 | [diff] [blame] | 1982 | * the larger and more IO friendly wb_stat_error. |
Wu Fengguang | c5c6343 | 2011-12-02 10:21:33 -0600 | [diff] [blame] | 1983 | */ |
Wang Long | 2bce774 | 2017-11-15 17:39:03 -0800 | [diff] [blame] | 1984 | if (sdtc->wb_dirty <= wb_stat_error()) |
Wu Fengguang | c5c6343 | 2011-12-02 10:21:33 -0600 | [diff] [blame] | 1985 | break; |
| 1986 | |
Jan Kara | 499d05e | 2011-11-16 19:34:48 +0800 | [diff] [blame] | 1987 | if (fatal_signal_pending(current)) |
| 1988 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1989 | } |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 1990 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1991 | } |
| 1992 | |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 1993 | static DEFINE_PER_CPU(int, bdp_ratelimits); |
Tejun Heo | 245b2e7 | 2009-06-24 15:13:48 +0900 | [diff] [blame] | 1994 | |
Wu Fengguang | 54848d7 | 2011-04-05 13:21:19 -0600 | [diff] [blame] | 1995 | /* |
| 1996 | * Normal tasks are throttled by |
| 1997 | * loop { |
| 1998 | * dirty tsk->nr_dirtied_pause pages; |
| 1999 | * take a snap in balance_dirty_pages(); |
| 2000 | * } |
| 2001 | * However there is a worst case. If every task exit immediately when dirtied |
| 2002 | * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be |
| 2003 | * called to throttle the page dirties. The solution is to save the not yet |
| 2004 | * throttled page dirties in dirty_throttle_leaks on task exit and charge them |
| 2005 | * randomly into the running tasks. This works well for the above worst case, |
| 2006 | * as the new task will pick up and accumulate the old task's leaked dirty |
| 2007 | * count and eventually get throttled. |
| 2008 | */ |
| 2009 | DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0; |
| 2010 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2011 | /** |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2012 | * balance_dirty_pages_ratelimited_flags - Balance dirty memory state. |
| 2013 | * @mapping: address_space which was dirtied. |
| 2014 | * @flags: BDP flags. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2015 | * |
| 2016 | * Processes which are dirtying memory should call in here once for each page |
| 2017 | * which was newly dirtied. The function will periodically check the system's |
| 2018 | * dirty state and will initiate writeback if needed. |
| 2019 | * |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2020 | * See balance_dirty_pages_ratelimited() for details. |
| 2021 | * |
| 2022 | * Return: If @flags contains BDP_ASYNC, it may return -EAGAIN to |
| 2023 | * indicate that memory is out of balance and the caller must wait |
| 2024 | * for I/O to complete. Otherwise, it will return 0 to indicate |
| 2025 | * that either memory was already in balance, or it was able to sleep |
| 2026 | * until the amount of dirty memory returned to balance. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2027 | */ |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2028 | int balance_dirty_pages_ratelimited_flags(struct address_space *mapping, |
| 2029 | unsigned int flags) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2030 | { |
Tejun Heo | dfb8ae5 | 2015-05-22 17:13:40 -0400 | [diff] [blame] | 2031 | struct inode *inode = mapping->host; |
| 2032 | struct backing_dev_info *bdi = inode_to_bdi(inode); |
| 2033 | struct bdi_writeback *wb = NULL; |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2034 | int ratelimit; |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2035 | int ret = 0; |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2036 | int *p; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2037 | |
Christoph Hellwig | f56753a | 2020-09-24 08:51:40 +0200 | [diff] [blame] | 2038 | if (!(bdi->capabilities & BDI_CAP_WRITEBACK)) |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2039 | return ret; |
Wu Fengguang | 36715ce | 2011-06-11 17:53:57 -0600 | [diff] [blame] | 2040 | |
Tejun Heo | dfb8ae5 | 2015-05-22 17:13:40 -0400 | [diff] [blame] | 2041 | if (inode_cgwb_enabled(inode)) |
| 2042 | wb = wb_get_create_current(bdi, GFP_KERNEL); |
| 2043 | if (!wb) |
| 2044 | wb = &bdi->wb; |
| 2045 | |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2046 | ratelimit = current->nr_dirtied_pause; |
Tejun Heo | a88a341 | 2015-05-22 17:13:28 -0400 | [diff] [blame] | 2047 | if (wb->dirty_exceeded) |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2048 | ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2049 | |
Andrew Morton | fa5a734 | 2006-03-24 03:18:10 -0800 | [diff] [blame] | 2050 | preempt_disable(); |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2051 | /* |
| 2052 | * This prevents one CPU to accumulate too many dirtied pages without |
| 2053 | * calling into balance_dirty_pages(), which can happen when there are |
| 2054 | * 1000+ tasks, all of them start dirtying pages at exactly the same |
| 2055 | * time, hence all honoured too large initial task->nr_dirtied_pause. |
| 2056 | */ |
Christoph Lameter | 7c8e018 | 2014-06-04 16:07:56 -0700 | [diff] [blame] | 2057 | p = this_cpu_ptr(&bdp_ratelimits); |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2058 | if (unlikely(current->nr_dirtied >= ratelimit)) |
Andrew Morton | fa5a734 | 2006-03-24 03:18:10 -0800 | [diff] [blame] | 2059 | *p = 0; |
Wu Fengguang | d3bc1fe | 2011-04-14 07:52:37 -0600 | [diff] [blame] | 2060 | else if (unlikely(*p >= ratelimit_pages)) { |
| 2061 | *p = 0; |
| 2062 | ratelimit = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2063 | } |
Wu Fengguang | 54848d7 | 2011-04-05 13:21:19 -0600 | [diff] [blame] | 2064 | /* |
| 2065 | * Pick up the dirtied pages by the exited tasks. This avoids lots of |
| 2066 | * short-lived tasks (eg. gcc invocations in a kernel build) escaping |
| 2067 | * the dirty throttling and livelock other long-run dirtiers. |
| 2068 | */ |
Christoph Lameter | 7c8e018 | 2014-06-04 16:07:56 -0700 | [diff] [blame] | 2069 | p = this_cpu_ptr(&dirty_throttle_leaks); |
Wu Fengguang | 54848d7 | 2011-04-05 13:21:19 -0600 | [diff] [blame] | 2070 | if (*p > 0 && current->nr_dirtied < ratelimit) { |
Namjae Jeon | d0e1d66 | 2012-12-11 16:00:21 -0800 | [diff] [blame] | 2071 | unsigned long nr_pages_dirtied; |
Wu Fengguang | 54848d7 | 2011-04-05 13:21:19 -0600 | [diff] [blame] | 2072 | nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied); |
| 2073 | *p -= nr_pages_dirtied; |
| 2074 | current->nr_dirtied += nr_pages_dirtied; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2075 | } |
Andrew Morton | fa5a734 | 2006-03-24 03:18:10 -0800 | [diff] [blame] | 2076 | preempt_enable(); |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2077 | |
| 2078 | if (unlikely(current->nr_dirtied >= ratelimit)) |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2079 | ret = balance_dirty_pages(wb, current->nr_dirtied, flags); |
Tejun Heo | dfb8ae5 | 2015-05-22 17:13:40 -0400 | [diff] [blame] | 2080 | |
| 2081 | wb_put(wb); |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2082 | return ret; |
| 2083 | } |
Stefan Roesch | 611df5d | 2022-09-12 12:27:41 -0700 | [diff] [blame] | 2084 | EXPORT_SYMBOL_GPL(balance_dirty_pages_ratelimited_flags); |
Jan Kara | fe6c9c6 | 2022-06-23 10:51:46 -0700 | [diff] [blame] | 2085 | |
| 2086 | /** |
| 2087 | * balance_dirty_pages_ratelimited - balance dirty memory state. |
| 2088 | * @mapping: address_space which was dirtied. |
| 2089 | * |
| 2090 | * Processes which are dirtying memory should call in here once for each page |
| 2091 | * which was newly dirtied. The function will periodically check the system's |
| 2092 | * dirty state and will initiate writeback if needed. |
| 2093 | * |
| 2094 | * Once we're over the dirty memory limit we decrease the ratelimiting |
| 2095 | * by a lot, to prevent individual processes from overshooting the limit |
| 2096 | * by (ratelimit_pages) each. |
| 2097 | */ |
| 2098 | void balance_dirty_pages_ratelimited(struct address_space *mapping) |
| 2099 | { |
| 2100 | balance_dirty_pages_ratelimited_flags(mapping, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2101 | } |
Namjae Jeon | d0e1d66 | 2012-12-11 16:00:21 -0800 | [diff] [blame] | 2102 | EXPORT_SYMBOL(balance_dirty_pages_ratelimited); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2103 | |
Kemeng Shi | 7c0c629 | 2024-05-14 20:52:47 +0800 | [diff] [blame] | 2104 | /* |
| 2105 | * Similar to wb_dirty_limits, wb_bg_dirty_limits also calculates dirty |
| 2106 | * and thresh, but it's for background writeback. |
| 2107 | */ |
| 2108 | static void wb_bg_dirty_limits(struct dirty_throttle_control *dtc) |
| 2109 | { |
| 2110 | struct bdi_writeback *wb = dtc->wb; |
| 2111 | |
| 2112 | dtc->wb_bg_thresh = __wb_calc_thresh(dtc, dtc->bg_thresh); |
| 2113 | if (dtc->wb_bg_thresh < 2 * wb_stat_error()) |
| 2114 | dtc->wb_dirty = wb_stat_sum(wb, WB_RECLAIMABLE); |
| 2115 | else |
| 2116 | dtc->wb_dirty = wb_stat(wb, WB_RECLAIMABLE); |
| 2117 | } |
| 2118 | |
Kemeng Shi | 6e20832 | 2024-05-14 20:52:49 +0800 | [diff] [blame^] | 2119 | static bool domain_over_bg_thresh(struct dirty_throttle_control *dtc) |
| 2120 | { |
| 2121 | domain_dirty_avail(dtc, false); |
| 2122 | domain_dirty_limits(dtc); |
| 2123 | if (dtc->dirty > dtc->bg_thresh) |
| 2124 | return true; |
| 2125 | |
| 2126 | wb_bg_dirty_limits(dtc); |
| 2127 | if (dtc->wb_dirty > dtc->wb_bg_thresh) |
| 2128 | return true; |
| 2129 | |
| 2130 | return false; |
| 2131 | } |
| 2132 | |
Tejun Heo | aa661bb | 2015-05-22 18:23:31 -0400 | [diff] [blame] | 2133 | /** |
| 2134 | * wb_over_bg_thresh - does @wb need to be written back? |
| 2135 | * @wb: bdi_writeback of interest |
| 2136 | * |
| 2137 | * Determines whether background writeback should keep writing @wb or it's |
Mike Rapoport | a862f68 | 2019-03-05 15:48:42 -0800 | [diff] [blame] | 2138 | * clean enough. |
| 2139 | * |
| 2140 | * Return: %true if writeback should continue. |
Tejun Heo | aa661bb | 2015-05-22 18:23:31 -0400 | [diff] [blame] | 2141 | */ |
| 2142 | bool wb_over_bg_thresh(struct bdi_writeback *wb) |
| 2143 | { |
Kemeng Shi | 6e20832 | 2024-05-14 20:52:49 +0800 | [diff] [blame^] | 2144 | struct dirty_throttle_control gdtc = { GDTC_INIT(wb) }; |
| 2145 | struct dirty_throttle_control mdtc = { MDTC_INIT(wb, &gdtc) }; |
Tejun Heo | aa661bb | 2015-05-22 18:23:31 -0400 | [diff] [blame] | 2146 | |
Kemeng Shi | 6e20832 | 2024-05-14 20:52:49 +0800 | [diff] [blame^] | 2147 | if (domain_over_bg_thresh(&gdtc)) |
Tejun Heo | aa661bb | 2015-05-22 18:23:31 -0400 | [diff] [blame] | 2148 | return true; |
| 2149 | |
Kemeng Shi | 6e20832 | 2024-05-14 20:52:49 +0800 | [diff] [blame^] | 2150 | if (mdtc_valid(&mdtc)) |
| 2151 | return domain_over_bg_thresh(&mdtc); |
Tejun Heo | c2aa723 | 2015-05-22 18:23:35 -0400 | [diff] [blame] | 2152 | |
Tejun Heo | aa661bb | 2015-05-22 18:23:31 -0400 | [diff] [blame] | 2153 | return false; |
| 2154 | } |
| 2155 | |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 2156 | #ifdef CONFIG_SYSCTL |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2157 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2158 | * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs |
| 2159 | */ |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 2160 | static int dirty_writeback_centisecs_handler(struct ctl_table *table, int write, |
Christoph Hellwig | 3292739 | 2020-04-24 08:43:38 +0200 | [diff] [blame] | 2161 | void *buffer, size_t *length, loff_t *ppos) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2162 | { |
Yafang Shao | 94af584 | 2017-10-10 05:44:13 +0800 | [diff] [blame] | 2163 | unsigned int old_interval = dirty_writeback_interval; |
| 2164 | int ret; |
| 2165 | |
| 2166 | ret = proc_dointvec(table, write, buffer, length, ppos); |
Yafang Shao | 515c24c | 2017-10-14 16:38:27 +0800 | [diff] [blame] | 2167 | |
| 2168 | /* |
| 2169 | * Writing 0 to dirty_writeback_interval will disable periodic writeback |
| 2170 | * and a different non-zero value will wakeup the writeback threads. |
| 2171 | * wb_wakeup_delayed() would be more appropriate, but it's a pain to |
| 2172 | * iterate over all bdis and wbs. |
| 2173 | * The reason we do this is to make the change take effect immediately. |
| 2174 | */ |
| 2175 | if (!ret && write && dirty_writeback_interval && |
| 2176 | dirty_writeback_interval != old_interval) |
Yafang Shao | 94af584 | 2017-10-10 05:44:13 +0800 | [diff] [blame] | 2177 | wakeup_flusher_threads(WB_REASON_PERIODIC); |
| 2178 | |
| 2179 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2180 | } |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 2181 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2182 | |
Kees Cook | bca237a | 2017-08-28 15:03:41 -0700 | [diff] [blame] | 2183 | void laptop_mode_timer_fn(struct timer_list *t) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2184 | { |
Kees Cook | bca237a | 2017-08-28 15:03:41 -0700 | [diff] [blame] | 2185 | struct backing_dev_info *backing_dev_info = |
| 2186 | from_timer(backing_dev_info, t, laptop_mode_wb_timer); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2187 | |
Kees Cook | bca237a | 2017-08-28 15:03:41 -0700 | [diff] [blame] | 2188 | wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2189 | } |
| 2190 | |
| 2191 | /* |
| 2192 | * We've spun up the disk and we're in laptop mode: schedule writeback |
| 2193 | * of all dirty data a few seconds from now. If the flush is already scheduled |
| 2194 | * then push it back - the user is still using the disk. |
| 2195 | */ |
Matthew Garrett | 31373d0 | 2010-04-06 14:25:14 +0200 | [diff] [blame] | 2196 | void laptop_io_completion(struct backing_dev_info *info) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2197 | { |
Matthew Garrett | 31373d0 | 2010-04-06 14:25:14 +0200 | [diff] [blame] | 2198 | mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2199 | } |
| 2200 | |
| 2201 | /* |
| 2202 | * We're in laptop mode and we've just synced. The sync's writes will have |
| 2203 | * caused another writeback to be scheduled by laptop_io_completion. |
| 2204 | * Nothing needs to be written back anymore, so we unschedule the writeback. |
| 2205 | */ |
| 2206 | void laptop_sync_completion(void) |
| 2207 | { |
Matthew Garrett | 31373d0 | 2010-04-06 14:25:14 +0200 | [diff] [blame] | 2208 | struct backing_dev_info *bdi; |
| 2209 | |
| 2210 | rcu_read_lock(); |
| 2211 | |
| 2212 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) |
| 2213 | del_timer(&bdi->laptop_mode_wb_timer); |
| 2214 | |
| 2215 | rcu_read_unlock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2216 | } |
| 2217 | |
| 2218 | /* |
| 2219 | * If ratelimit_pages is too high then we can get into dirty-data overload |
| 2220 | * if a large number of processes all perform writes at the same time. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2221 | * |
| 2222 | * Here we set ratelimit_pages to a level which ensures that when all CPUs are |
| 2223 | * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2224 | * thresholds. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2225 | */ |
| 2226 | |
Chandra Seetharaman | 2d1d43f | 2006-09-29 02:01:25 -0700 | [diff] [blame] | 2227 | void writeback_set_ratelimit(void) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2228 | { |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 2229 | struct wb_domain *dom = &global_wb_domain; |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2230 | unsigned long background_thresh; |
| 2231 | unsigned long dirty_thresh; |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 2232 | |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2233 | global_dirty_limits(&background_thresh, &dirty_thresh); |
Tejun Heo | dcc25ae | 2015-05-22 18:23:22 -0400 | [diff] [blame] | 2234 | dom->dirty_limit = dirty_thresh; |
Wu Fengguang | 9d823e8 | 2011-06-11 18:10:12 -0600 | [diff] [blame] | 2235 | ratelimit_pages = dirty_thresh / (num_online_cpus() * 32); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2236 | if (ratelimit_pages < 16) |
| 2237 | ratelimit_pages = 16; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2238 | } |
| 2239 | |
Sebastian Andrzej Siewior | 1d7ac6a | 2016-08-18 14:57:20 +0200 | [diff] [blame] | 2240 | static int page_writeback_cpu_online(unsigned int cpu) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2241 | { |
Sebastian Andrzej Siewior | 1d7ac6a | 2016-08-18 14:57:20 +0200 | [diff] [blame] | 2242 | writeback_set_ratelimit(); |
| 2243 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2244 | } |
| 2245 | |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 2246 | #ifdef CONFIG_SYSCTL |
Luis Chamberlain | 3c6a4cb | 2022-04-15 15:08:02 -0700 | [diff] [blame] | 2247 | |
| 2248 | /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */ |
| 2249 | static const unsigned long dirty_bytes_min = 2 * PAGE_SIZE; |
| 2250 | |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 2251 | static struct ctl_table vm_page_writeback_sysctls[] = { |
| 2252 | { |
| 2253 | .procname = "dirty_background_ratio", |
| 2254 | .data = &dirty_background_ratio, |
| 2255 | .maxlen = sizeof(dirty_background_ratio), |
| 2256 | .mode = 0644, |
| 2257 | .proc_handler = dirty_background_ratio_handler, |
| 2258 | .extra1 = SYSCTL_ZERO, |
| 2259 | .extra2 = SYSCTL_ONE_HUNDRED, |
| 2260 | }, |
| 2261 | { |
| 2262 | .procname = "dirty_background_bytes", |
| 2263 | .data = &dirty_background_bytes, |
| 2264 | .maxlen = sizeof(dirty_background_bytes), |
| 2265 | .mode = 0644, |
| 2266 | .proc_handler = dirty_background_bytes_handler, |
| 2267 | .extra1 = SYSCTL_LONG_ONE, |
| 2268 | }, |
| 2269 | { |
| 2270 | .procname = "dirty_ratio", |
| 2271 | .data = &vm_dirty_ratio, |
| 2272 | .maxlen = sizeof(vm_dirty_ratio), |
| 2273 | .mode = 0644, |
| 2274 | .proc_handler = dirty_ratio_handler, |
| 2275 | .extra1 = SYSCTL_ZERO, |
| 2276 | .extra2 = SYSCTL_ONE_HUNDRED, |
| 2277 | }, |
| 2278 | { |
| 2279 | .procname = "dirty_bytes", |
| 2280 | .data = &vm_dirty_bytes, |
| 2281 | .maxlen = sizeof(vm_dirty_bytes), |
| 2282 | .mode = 0644, |
| 2283 | .proc_handler = dirty_bytes_handler, |
| 2284 | .extra1 = (void *)&dirty_bytes_min, |
| 2285 | }, |
| 2286 | { |
| 2287 | .procname = "dirty_writeback_centisecs", |
| 2288 | .data = &dirty_writeback_interval, |
| 2289 | .maxlen = sizeof(dirty_writeback_interval), |
| 2290 | .mode = 0644, |
| 2291 | .proc_handler = dirty_writeback_centisecs_handler, |
| 2292 | }, |
| 2293 | { |
| 2294 | .procname = "dirty_expire_centisecs", |
| 2295 | .data = &dirty_expire_interval, |
| 2296 | .maxlen = sizeof(dirty_expire_interval), |
| 2297 | .mode = 0644, |
| 2298 | .proc_handler = proc_dointvec_minmax, |
| 2299 | .extra1 = SYSCTL_ZERO, |
| 2300 | }, |
| 2301 | #ifdef CONFIG_HIGHMEM |
| 2302 | { |
| 2303 | .procname = "highmem_is_dirtyable", |
| 2304 | .data = &vm_highmem_is_dirtyable, |
| 2305 | .maxlen = sizeof(vm_highmem_is_dirtyable), |
| 2306 | .mode = 0644, |
| 2307 | .proc_handler = proc_dointvec_minmax, |
| 2308 | .extra1 = SYSCTL_ZERO, |
| 2309 | .extra2 = SYSCTL_ONE, |
| 2310 | }, |
| 2311 | #endif |
| 2312 | { |
| 2313 | .procname = "laptop_mode", |
| 2314 | .data = &laptop_mode, |
| 2315 | .maxlen = sizeof(laptop_mode), |
| 2316 | .mode = 0644, |
| 2317 | .proc_handler = proc_dointvec_jiffies, |
| 2318 | }, |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 2319 | }; |
| 2320 | #endif |
| 2321 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2322 | /* |
Linus Torvalds | dc6e29d | 2007-01-29 16:37:38 -0800 | [diff] [blame] | 2323 | * Called early on to tune the page writeback dirty limits. |
| 2324 | * |
| 2325 | * We used to scale dirty pages according to how total memory |
David Hildenbrand | 0a18e60 | 2020-08-06 23:25:27 -0700 | [diff] [blame] | 2326 | * related to pages that could be allocated for buffers. |
Linus Torvalds | dc6e29d | 2007-01-29 16:37:38 -0800 | [diff] [blame] | 2327 | * |
| 2328 | * However, that was when we used "dirty_ratio" to scale with |
| 2329 | * all memory, and we don't do that any more. "dirty_ratio" |
David Hildenbrand | 0a18e60 | 2020-08-06 23:25:27 -0700 | [diff] [blame] | 2330 | * is now applied to total non-HIGHPAGE memory, and as such we can't |
Linus Torvalds | dc6e29d | 2007-01-29 16:37:38 -0800 | [diff] [blame] | 2331 | * get into the old insane situation any more where we had |
| 2332 | * large amounts of dirty pages compared to a small amount of |
| 2333 | * non-HIGHMEM memory. |
| 2334 | * |
| 2335 | * But we might still want to scale the dirty_ratio by how |
| 2336 | * much memory the box has.. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2337 | */ |
| 2338 | void __init page_writeback_init(void) |
| 2339 | { |
Rabin Vincent | a50fcb5 | 2015-08-06 15:47:14 -0700 | [diff] [blame] | 2340 | BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL)); |
| 2341 | |
Sebastian Andrzej Siewior | 1d7ac6a | 2016-08-18 14:57:20 +0200 | [diff] [blame] | 2342 | cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online", |
| 2343 | page_writeback_cpu_online, NULL); |
| 2344 | cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL, |
| 2345 | page_writeback_cpu_online); |
zhanglianjie | aa779e5 | 2022-02-17 18:51:51 -0800 | [diff] [blame] | 2346 | #ifdef CONFIG_SYSCTL |
| 2347 | register_sysctl_init("vm", vm_page_writeback_sysctls); |
| 2348 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2349 | } |
| 2350 | |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2351 | /** |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2352 | * tag_pages_for_writeback - tag pages to be written by writeback |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2353 | * @mapping: address space structure to write |
| 2354 | * @start: starting page index |
| 2355 | * @end: ending page index (inclusive) |
| 2356 | * |
| 2357 | * This function scans the page range from @start to @end (inclusive) and tags |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2358 | * all pages that have DIRTY tag set with a special TOWRITE tag. The caller |
| 2359 | * can then use the TOWRITE tag to identify pages eligible for writeback. |
| 2360 | * This mechanism is used to avoid livelocking of writeback by a process |
| 2361 | * steadily creating new dirty pages in the file (thus it is important for this |
| 2362 | * function to be quick so that it can tag pages faster than a dirtying process |
| 2363 | * can create them). |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2364 | */ |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2365 | void tag_pages_for_writeback(struct address_space *mapping, |
| 2366 | pgoff_t start, pgoff_t end) |
| 2367 | { |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 2368 | XA_STATE(xas, &mapping->i_pages, start); |
| 2369 | unsigned int tagged = 0; |
| 2370 | void *page; |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2371 | |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 2372 | xas_lock_irq(&xas); |
| 2373 | xas_for_each_marked(&xas, page, end, PAGECACHE_TAG_DIRTY) { |
| 2374 | xas_set_mark(&xas, PAGECACHE_TAG_TOWRITE); |
| 2375 | if (++tagged % XA_CHECK_SCHED) |
Matthew Wilcox | 268f42d | 2016-12-14 15:08:55 -0800 | [diff] [blame] | 2376 | continue; |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 2377 | |
| 2378 | xas_pause(&xas); |
| 2379 | xas_unlock_irq(&xas); |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2380 | cond_resched(); |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 2381 | xas_lock_irq(&xas); |
Matthew Wilcox | 268f42d | 2016-12-14 15:08:55 -0800 | [diff] [blame] | 2382 | } |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 2383 | xas_unlock_irq(&xas); |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2384 | } |
| 2385 | EXPORT_SYMBOL(tag_pages_for_writeback); |
| 2386 | |
Matthew Wilcox (Oracle) | b179392 | 2024-02-15 07:36:42 +0100 | [diff] [blame] | 2387 | static bool folio_prepare_writeback(struct address_space *mapping, |
| 2388 | struct writeback_control *wbc, struct folio *folio) |
| 2389 | { |
| 2390 | /* |
| 2391 | * Folio truncated or invalidated. We can freely skip it then, |
| 2392 | * even for data integrity operations: the folio has disappeared |
| 2393 | * concurrently, so there could be no real expectation of this |
| 2394 | * data integrity operation even if there is now a new, dirty |
| 2395 | * folio at the same pagecache index. |
| 2396 | */ |
| 2397 | if (unlikely(folio->mapping != mapping)) |
| 2398 | return false; |
| 2399 | |
| 2400 | /* |
| 2401 | * Did somebody else write it for us? |
| 2402 | */ |
| 2403 | if (!folio_test_dirty(folio)) |
| 2404 | return false; |
| 2405 | |
| 2406 | if (folio_test_writeback(folio)) { |
| 2407 | if (wbc->sync_mode == WB_SYNC_NONE) |
| 2408 | return false; |
| 2409 | folio_wait_writeback(folio); |
| 2410 | } |
| 2411 | BUG_ON(folio_test_writeback(folio)); |
| 2412 | |
| 2413 | if (!folio_clear_dirty_for_io(folio)) |
| 2414 | return false; |
| 2415 | |
| 2416 | return true; |
| 2417 | } |
| 2418 | |
Matthew Wilcox (Oracle) | 751e0d5 | 2024-02-15 07:36:43 +0100 | [diff] [blame] | 2419 | static xa_mark_t wbc_to_tag(struct writeback_control *wbc) |
| 2420 | { |
| 2421 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
| 2422 | return PAGECACHE_TAG_TOWRITE; |
| 2423 | return PAGECACHE_TAG_DIRTY; |
| 2424 | } |
| 2425 | |
| 2426 | static pgoff_t wbc_end(struct writeback_control *wbc) |
| 2427 | { |
| 2428 | if (wbc->range_cyclic) |
| 2429 | return -1; |
| 2430 | return wbc->range_end >> PAGE_SHIFT; |
| 2431 | } |
| 2432 | |
Matthew Wilcox (Oracle) | e6d0ab8 | 2024-02-15 07:36:46 +0100 | [diff] [blame] | 2433 | static struct folio *writeback_get_folio(struct address_space *mapping, |
Matthew Wilcox (Oracle) | 751e0d5 | 2024-02-15 07:36:43 +0100 | [diff] [blame] | 2434 | struct writeback_control *wbc) |
| 2435 | { |
Matthew Wilcox (Oracle) | e6d0ab8 | 2024-02-15 07:36:46 +0100 | [diff] [blame] | 2436 | struct folio *folio; |
| 2437 | |
Matthew Wilcox (Oracle) | a2cbc13 | 2024-02-15 07:36:47 +0100 | [diff] [blame] | 2438 | retry: |
Matthew Wilcox (Oracle) | e6d0ab8 | 2024-02-15 07:36:46 +0100 | [diff] [blame] | 2439 | folio = folio_batch_next(&wbc->fbatch); |
| 2440 | if (!folio) { |
| 2441 | folio_batch_release(&wbc->fbatch); |
| 2442 | cond_resched(); |
| 2443 | filemap_get_folios_tag(mapping, &wbc->index, wbc_end(wbc), |
| 2444 | wbc_to_tag(wbc), &wbc->fbatch); |
| 2445 | folio = folio_batch_next(&wbc->fbatch); |
Matthew Wilcox (Oracle) | a2cbc13 | 2024-02-15 07:36:47 +0100 | [diff] [blame] | 2446 | if (!folio) |
| 2447 | return NULL; |
Matthew Wilcox (Oracle) | e6d0ab8 | 2024-02-15 07:36:46 +0100 | [diff] [blame] | 2448 | } |
| 2449 | |
Matthew Wilcox (Oracle) | a2cbc13 | 2024-02-15 07:36:47 +0100 | [diff] [blame] | 2450 | folio_lock(folio); |
| 2451 | if (unlikely(!folio_prepare_writeback(mapping, wbc, folio))) { |
| 2452 | folio_unlock(folio); |
| 2453 | goto retry; |
| 2454 | } |
| 2455 | |
| 2456 | trace_wbc_writepage(wbc, inode_to_bdi(mapping->host)); |
Matthew Wilcox (Oracle) | e6d0ab8 | 2024-02-15 07:36:46 +0100 | [diff] [blame] | 2457 | return folio; |
Matthew Wilcox (Oracle) | 751e0d5 | 2024-02-15 07:36:43 +0100 | [diff] [blame] | 2458 | } |
| 2459 | |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2460 | /** |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2461 | * writeback_iter - iterate folio of a mapping for writeback |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2462 | * @mapping: address space structure to write |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2463 | * @wbc: writeback context |
| 2464 | * @folio: previously iterated folio (%NULL to start) |
| 2465 | * @error: in-out pointer for writeback errors (see below) |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2466 | * |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2467 | * This function returns the next folio for the writeback operation described by |
| 2468 | * @wbc on @mapping and should be called in a while loop in the ->writepages |
| 2469 | * implementation. |
Jan Kara | f446daae | 2010-08-09 17:19:12 -0700 | [diff] [blame] | 2470 | * |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2471 | * To start the writeback operation, %NULL is passed in the @folio argument, and |
| 2472 | * for every subsequent iteration the folio returned previously should be passed |
| 2473 | * back in. |
Dave Chinner | 6408136 | 2018-10-26 15:09:45 -0700 | [diff] [blame] | 2474 | * |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2475 | * If there was an error in the per-folio writeback inside the writeback_iter() |
| 2476 | * loop, @error should be set to the error value. |
Mike Rapoport | a862f68 | 2019-03-05 15:48:42 -0800 | [diff] [blame] | 2477 | * |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2478 | * Once the writeback described in @wbc has finished, this function will return |
| 2479 | * %NULL and if there was an error in any iteration restore it to @error. |
| 2480 | * |
| 2481 | * Note: callers should not manually break out of the loop using break or goto |
| 2482 | * but must keep calling writeback_iter() until it returns %NULL. |
| 2483 | * |
| 2484 | * Return: the folio to write or %NULL if the loop is done. |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2485 | */ |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2486 | struct folio *writeback_iter(struct address_space *mapping, |
| 2487 | struct writeback_control *wbc, struct folio *folio, int *error) |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2488 | { |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2489 | if (!folio) { |
| 2490 | folio_batch_init(&wbc->fbatch); |
| 2491 | wbc->saved_err = *error = 0; |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2492 | |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2493 | /* |
| 2494 | * For range cyclic writeback we remember where we stopped so |
| 2495 | * that we can continue where we stopped. |
| 2496 | * |
| 2497 | * For non-cyclic writeback we always start at the beginning of |
| 2498 | * the passed in range. |
| 2499 | */ |
| 2500 | if (wbc->range_cyclic) |
| 2501 | wbc->index = mapping->writeback_index; |
| 2502 | else |
| 2503 | wbc->index = wbc->range_start >> PAGE_SHIFT; |
| 2504 | |
| 2505 | /* |
| 2506 | * To avoid livelocks when other processes dirty new pages, we |
| 2507 | * first tag pages which should be written back and only then |
| 2508 | * start writing them. |
| 2509 | * |
| 2510 | * For data-integrity writeback we have to be careful so that we |
| 2511 | * do not miss some pages (e.g., because some other process has |
| 2512 | * cleared the TOWRITE tag we set). The rule we follow is that |
| 2513 | * TOWRITE tag can be cleared only by the process clearing the |
| 2514 | * DIRTY tag (and submitting the page for I/O). |
| 2515 | */ |
| 2516 | if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages) |
| 2517 | tag_pages_for_writeback(mapping, wbc->index, |
| 2518 | wbc_end(wbc)); |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2519 | } else { |
Matthew Wilcox (Oracle) | 807d1fe | 2024-02-15 07:36:44 +0100 | [diff] [blame] | 2520 | wbc->nr_to_write -= folio_nr_pages(folio); |
Christoph Hellwig | f946e0d | 2024-02-15 07:36:41 +0100 | [diff] [blame] | 2521 | |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2522 | WARN_ON_ONCE(*error > 0); |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2523 | |
Matthew Wilcox (Oracle) | 807d1fe | 2024-02-15 07:36:44 +0100 | [diff] [blame] | 2524 | /* |
| 2525 | * For integrity writeback we have to keep going until we have |
| 2526 | * written all the folios we tagged for writeback above, even if |
| 2527 | * we run past wbc->nr_to_write or encounter errors. |
| 2528 | * We stash away the first error we encounter in wbc->saved_err |
| 2529 | * so that it can be retrieved when we're done. This is because |
| 2530 | * the file system may still have state to clear for each folio. |
| 2531 | * |
| 2532 | * For background writeback we exit as soon as we run past |
| 2533 | * wbc->nr_to_write or encounter the first error. |
| 2534 | */ |
| 2535 | if (wbc->sync_mode == WB_SYNC_ALL) { |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2536 | if (*error && !wbc->saved_err) |
| 2537 | wbc->saved_err = *error; |
Matthew Wilcox (Oracle) | 807d1fe | 2024-02-15 07:36:44 +0100 | [diff] [blame] | 2538 | } else { |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2539 | if (*error || wbc->nr_to_write <= 0) |
Matthew Wilcox (Oracle) | 807d1fe | 2024-02-15 07:36:44 +0100 | [diff] [blame] | 2540 | goto done; |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2541 | } |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2542 | } |
Dave Chinner | 6408136 | 2018-10-26 15:09:45 -0700 | [diff] [blame] | 2543 | |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2544 | folio = writeback_get_folio(mapping, wbc); |
| 2545 | if (!folio) { |
| 2546 | /* |
| 2547 | * To avoid deadlocks between range_cyclic writeback and callers |
| 2548 | * that hold pages in PageWriteback to aggregate I/O until |
| 2549 | * the writeback iteration finishes, we do not loop back to the |
| 2550 | * start of the file. Doing so causes a page lock/page |
| 2551 | * writeback access order inversion - we should only ever lock |
| 2552 | * multiple pages in ascending page->index order, and looping |
| 2553 | * back to the start of the file violates that rule and causes |
| 2554 | * deadlocks. |
| 2555 | */ |
| 2556 | if (wbc->range_cyclic) |
| 2557 | mapping->writeback_index = 0; |
| 2558 | |
| 2559 | /* |
| 2560 | * Return the first error we encountered (if there was any) to |
| 2561 | * the caller. |
| 2562 | */ |
| 2563 | *error = wbc->saved_err; |
| 2564 | } |
| 2565 | return folio; |
Christoph Hellwig | f946e0d | 2024-02-15 07:36:41 +0100 | [diff] [blame] | 2566 | |
| 2567 | done: |
| 2568 | if (wbc->range_cyclic) |
| 2569 | mapping->writeback_index = folio->index + folio_nr_pages(folio); |
Matthew Wilcox (Oracle) | 751e0d5 | 2024-02-15 07:36:43 +0100 | [diff] [blame] | 2570 | folio_batch_release(&wbc->fbatch); |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2571 | return NULL; |
| 2572 | } |
David Howells | 5a550a0 | 2024-03-18 16:18:49 +0000 | [diff] [blame] | 2573 | EXPORT_SYMBOL_GPL(writeback_iter); |
Christoph Hellwig | cdc150b | 2024-02-15 07:36:48 +0100 | [diff] [blame] | 2574 | |
| 2575 | /** |
| 2576 | * write_cache_pages - walk the list of dirty pages of the given address space and write all of them. |
| 2577 | * @mapping: address space structure to write |
| 2578 | * @wbc: subtract the number of written pages from *@wbc->nr_to_write |
| 2579 | * @writepage: function called for each page |
| 2580 | * @data: data passed to writepage function |
| 2581 | * |
| 2582 | * Return: %0 on success, negative error code otherwise |
| 2583 | * |
| 2584 | * Note: please use writeback_iter() instead. |
| 2585 | */ |
| 2586 | int write_cache_pages(struct address_space *mapping, |
| 2587 | struct writeback_control *wbc, writepage_t writepage, |
| 2588 | void *data) |
| 2589 | { |
| 2590 | struct folio *folio = NULL; |
| 2591 | int error; |
| 2592 | |
| 2593 | while ((folio = writeback_iter(mapping, wbc, folio, &error))) { |
| 2594 | error = writepage(folio, wbc, data); |
| 2595 | if (error == AOP_WRITEPAGE_ACTIVATE) { |
| 2596 | folio_unlock(folio); |
| 2597 | error = 0; |
| 2598 | } |
| 2599 | } |
| 2600 | |
Christoph Hellwig | f946e0d | 2024-02-15 07:36:41 +0100 | [diff] [blame] | 2601 | return error; |
David Howells | 811d736 | 2006-08-29 19:06:09 +0100 | [diff] [blame] | 2602 | } |
Miklos Szeredi | 0ea9718 | 2007-05-10 22:22:51 -0700 | [diff] [blame] | 2603 | EXPORT_SYMBOL(write_cache_pages); |
| 2604 | |
Matthew Wilcox (Oracle) | c44ed5b | 2024-02-15 07:36:49 +0100 | [diff] [blame] | 2605 | static int writeback_use_writepage(struct address_space *mapping, |
| 2606 | struct writeback_control *wbc) |
Miklos Szeredi | 0ea9718 | 2007-05-10 22:22:51 -0700 | [diff] [blame] | 2607 | { |
Matthew Wilcox (Oracle) | c44ed5b | 2024-02-15 07:36:49 +0100 | [diff] [blame] | 2608 | struct folio *folio = NULL; |
| 2609 | struct blk_plug plug; |
| 2610 | int err; |
Christoph Hellwig | 6768907 | 2024-02-15 07:36:36 +0100 | [diff] [blame] | 2611 | |
Matthew Wilcox (Oracle) | c44ed5b | 2024-02-15 07:36:49 +0100 | [diff] [blame] | 2612 | blk_start_plug(&plug); |
| 2613 | while ((folio = writeback_iter(mapping, wbc, folio, &err))) { |
| 2614 | err = mapping->a_ops->writepage(&folio->page, wbc); |
| 2615 | if (err == AOP_WRITEPAGE_ACTIVATE) { |
| 2616 | folio_unlock(folio); |
| 2617 | err = 0; |
| 2618 | } |
| 2619 | mapping_set_error(mapping, err); |
| 2620 | } |
| 2621 | blk_finish_plug(&plug); |
| 2622 | |
| 2623 | return err; |
Miklos Szeredi | 0ea9718 | 2007-05-10 22:22:51 -0700 | [diff] [blame] | 2624 | } |
| 2625 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2626 | int do_writepages(struct address_space *mapping, struct writeback_control *wbc) |
| 2627 | { |
Andrew Morton | 22905f7 | 2005-11-16 15:07:01 -0800 | [diff] [blame] | 2628 | int ret; |
Jan Kara | fee468f | 2021-09-02 14:53:06 -0700 | [diff] [blame] | 2629 | struct bdi_writeback *wb; |
Andrew Morton | 22905f7 | 2005-11-16 15:07:01 -0800 | [diff] [blame] | 2630 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2631 | if (wbc->nr_to_write <= 0) |
| 2632 | return 0; |
Jan Kara | fee468f | 2021-09-02 14:53:06 -0700 | [diff] [blame] | 2633 | wb = inode_to_wb_wbc(mapping->host, wbc); |
| 2634 | wb_bandwidth_estimate_start(wb); |
Theodore Ts'o | 80a2ea9 | 2017-04-28 09:51:54 -0400 | [diff] [blame] | 2635 | while (1) { |
Christoph Hellwig | c2ca7a5 | 2022-12-29 06:10:31 -1000 | [diff] [blame] | 2636 | if (mapping->a_ops->writepages) { |
Theodore Ts'o | 80a2ea9 | 2017-04-28 09:51:54 -0400 | [diff] [blame] | 2637 | ret = mapping->a_ops->writepages(mapping, wbc); |
Christoph Hellwig | c2ca7a5 | 2022-12-29 06:10:31 -1000 | [diff] [blame] | 2638 | } else if (mapping->a_ops->writepage) { |
Matthew Wilcox (Oracle) | c44ed5b | 2024-02-15 07:36:49 +0100 | [diff] [blame] | 2639 | ret = writeback_use_writepage(mapping, wbc); |
Christoph Hellwig | c2ca7a5 | 2022-12-29 06:10:31 -1000 | [diff] [blame] | 2640 | } else { |
| 2641 | /* deal with chardevs and other special files */ |
| 2642 | ret = 0; |
| 2643 | } |
| 2644 | if (ret != -ENOMEM || wbc->sync_mode != WB_SYNC_ALL) |
Theodore Ts'o | 80a2ea9 | 2017-04-28 09:51:54 -0400 | [diff] [blame] | 2645 | break; |
Mel Gorman | 8d58802 | 2021-11-05 13:42:35 -0700 | [diff] [blame] | 2646 | |
| 2647 | /* |
| 2648 | * Lacking an allocation context or the locality or writeback |
| 2649 | * state of any of the inode's pages, throttle based on |
| 2650 | * writeback activity on the local node. It's as good a |
| 2651 | * guess as any. |
| 2652 | */ |
| 2653 | reclaim_throttle(NODE_DATA(numa_node_id()), |
Mel Gorman | c3f4a9a | 2021-11-05 13:42:42 -0700 | [diff] [blame] | 2654 | VMSCAN_THROTTLE_WRITEBACK); |
Theodore Ts'o | 80a2ea9 | 2017-04-28 09:51:54 -0400 | [diff] [blame] | 2655 | } |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 2656 | /* |
| 2657 | * Usually few pages are written by now from those we've just submitted |
| 2658 | * but if there's constant writeback being submitted, this makes sure |
| 2659 | * writeback bandwidth is updated once in a while. |
| 2660 | */ |
Jan Kara | 20792eb | 2021-09-02 14:53:15 -0700 | [diff] [blame] | 2661 | if (time_is_before_jiffies(READ_ONCE(wb->bw_time_stamp) + |
| 2662 | BANDWIDTH_INTERVAL)) |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 2663 | wb_update_bandwidth(wb); |
Andrew Morton | 22905f7 | 2005-11-16 15:07:01 -0800 | [diff] [blame] | 2664 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2665 | } |
| 2666 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2667 | /* |
Ken Chen | 7671932 | 2007-02-10 01:43:15 -0800 | [diff] [blame] | 2668 | * For address_spaces which do not use buffers nor write back. |
| 2669 | */ |
Matthew Wilcox (Oracle) | 46de8b97 | 2022-02-09 20:22:13 +0000 | [diff] [blame] | 2670 | bool noop_dirty_folio(struct address_space *mapping, struct folio *folio) |
Ken Chen | 7671932 | 2007-02-10 01:43:15 -0800 | [diff] [blame] | 2671 | { |
Matthew Wilcox (Oracle) | 46de8b97 | 2022-02-09 20:22:13 +0000 | [diff] [blame] | 2672 | if (!folio_test_dirty(folio)) |
| 2673 | return !folio_test_set_dirty(folio); |
| 2674 | return false; |
Ken Chen | 7671932 | 2007-02-10 01:43:15 -0800 | [diff] [blame] | 2675 | } |
Matthew Wilcox (Oracle) | 46de8b97 | 2022-02-09 20:22:13 +0000 | [diff] [blame] | 2676 | EXPORT_SYMBOL(noop_dirty_folio); |
Ken Chen | 7671932 | 2007-02-10 01:43:15 -0800 | [diff] [blame] | 2677 | |
| 2678 | /* |
Edward Shishkin | e3a7cca | 2009-03-31 15:19:39 -0700 | [diff] [blame] | 2679 | * Helper function for set_page_dirty family. |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2680 | * |
Kefeng Wang | 6c77b60 | 2023-06-14 22:36:12 +0800 | [diff] [blame] | 2681 | * Caller must hold folio_memcg_lock(). |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2682 | * |
Edward Shishkin | e3a7cca | 2009-03-31 15:19:39 -0700 | [diff] [blame] | 2683 | * NOTE: This relies on being atomic wrt interrupts. |
| 2684 | */ |
Matthew Wilcox (Oracle) | 203a315 | 2021-05-04 11:01:10 -0400 | [diff] [blame] | 2685 | static void folio_account_dirtied(struct folio *folio, |
Matthew Wilcox (Oracle) | 6e1cae88 | 2021-06-28 19:36:15 -0700 | [diff] [blame] | 2686 | struct address_space *mapping) |
Edward Shishkin | e3a7cca | 2009-03-31 15:19:39 -0700 | [diff] [blame] | 2687 | { |
Tejun Heo | 52ebea7 | 2015-05-22 17:13:37 -0400 | [diff] [blame] | 2688 | struct inode *inode = mapping->host; |
| 2689 | |
Matthew Wilcox (Oracle) | b9b0ff6 | 2021-07-13 22:45:50 -0400 | [diff] [blame] | 2690 | trace_writeback_dirty_folio(folio, mapping); |
Tejun Heo | 9fb0a7d | 2013-01-11 13:06:37 -0800 | [diff] [blame] | 2691 | |
Christoph Hellwig | f56753a | 2020-09-24 08:51:40 +0200 | [diff] [blame] | 2692 | if (mapping_can_writeback(mapping)) { |
Tejun Heo | 52ebea7 | 2015-05-22 17:13:37 -0400 | [diff] [blame] | 2693 | struct bdi_writeback *wb; |
Matthew Wilcox (Oracle) | 203a315 | 2021-05-04 11:01:10 -0400 | [diff] [blame] | 2694 | long nr = folio_nr_pages(folio); |
Christoph Hellwig | de1414a | 2015-01-14 10:42:36 +0100 | [diff] [blame] | 2695 | |
Matthew Wilcox (Oracle) | 9cfb816 | 2023-01-16 19:25:06 +0000 | [diff] [blame] | 2696 | inode_attach_wb(inode, folio); |
Tejun Heo | 52ebea7 | 2015-05-22 17:13:37 -0400 | [diff] [blame] | 2697 | wb = inode_to_wb(inode); |
Edward Shishkin | e3a7cca | 2009-03-31 15:19:39 -0700 | [diff] [blame] | 2698 | |
Matthew Wilcox (Oracle) | 203a315 | 2021-05-04 11:01:10 -0400 | [diff] [blame] | 2699 | __lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, nr); |
| 2700 | __zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); |
| 2701 | __node_stat_mod_folio(folio, NR_DIRTIED, nr); |
| 2702 | wb_stat_mod(wb, WB_RECLAIMABLE, nr); |
| 2703 | wb_stat_mod(wb, WB_DIRTIED, nr); |
| 2704 | task_io_account_write(nr * PAGE_SIZE); |
| 2705 | current->nr_dirtied += nr; |
| 2706 | __this_cpu_add(bdp_ratelimits, nr); |
Tejun Heo | 97b2782 | 2019-08-26 09:06:56 -0700 | [diff] [blame] | 2707 | |
Matthew Wilcox (Oracle) | 203a315 | 2021-05-04 11:01:10 -0400 | [diff] [blame] | 2708 | mem_cgroup_track_foreign_dirty(folio, wb); |
Edward Shishkin | e3a7cca | 2009-03-31 15:19:39 -0700 | [diff] [blame] | 2709 | } |
| 2710 | } |
| 2711 | |
| 2712 | /* |
Konstantin Khlebnikov | b9ea251 | 2015-04-14 15:45:27 -0700 | [diff] [blame] | 2713 | * Helper function for deaccounting dirty page without writeback. |
| 2714 | * |
Kefeng Wang | 6c77b60 | 2023-06-14 22:36:12 +0800 | [diff] [blame] | 2715 | * Caller must hold folio_memcg_lock(). |
Konstantin Khlebnikov | b9ea251 | 2015-04-14 15:45:27 -0700 | [diff] [blame] | 2716 | */ |
Hugh Dickins | 566d336 | 2022-03-24 18:13:59 -0700 | [diff] [blame] | 2717 | void folio_account_cleaned(struct folio *folio, struct bdi_writeback *wb) |
Konstantin Khlebnikov | b9ea251 | 2015-04-14 15:45:27 -0700 | [diff] [blame] | 2718 | { |
Hugh Dickins | 566d336 | 2022-03-24 18:13:59 -0700 | [diff] [blame] | 2719 | long nr = folio_nr_pages(folio); |
| 2720 | |
| 2721 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); |
| 2722 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); |
| 2723 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); |
| 2724 | task_io_account_cancelled_write(nr * PAGE_SIZE); |
Konstantin Khlebnikov | b9ea251 | 2015-04-14 15:45:27 -0700 | [diff] [blame] | 2725 | } |
Konstantin Khlebnikov | b9ea251 | 2015-04-14 15:45:27 -0700 | [diff] [blame] | 2726 | |
| 2727 | /* |
Kemeng Shi | dc6e0ae | 2024-04-25 21:17:24 +0800 | [diff] [blame] | 2728 | * Mark the folio dirty, and set it dirty in the page cache. |
Matthew Wilcox (Oracle) | 6e1cae88 | 2021-06-28 19:36:15 -0700 | [diff] [blame] | 2729 | * |
Matthew Wilcox (Oracle) | 203a315 | 2021-05-04 11:01:10 -0400 | [diff] [blame] | 2730 | * If warn is true, then emit a warning if the folio is not uptodate and has |
Matthew Wilcox (Oracle) | 6e1cae88 | 2021-06-28 19:36:15 -0700 | [diff] [blame] | 2731 | * not been truncated. |
| 2732 | * |
Matthew Wilcox (Oracle) | 3d84d89 | 2024-04-16 04:17:45 +0100 | [diff] [blame] | 2733 | * The caller must hold folio_memcg_lock(). It is the caller's |
| 2734 | * responsibility to prevent the folio from being truncated while |
| 2735 | * this function is in progress, although it may have been truncated |
| 2736 | * before this function is called. Most callers have the folio locked. |
| 2737 | * A few have the folio blocked from truncation through other means (e.g. |
| 2738 | * zap_vma_pages() has it mapped and is holding the page table lock). |
| 2739 | * When called from mark_buffer_dirty(), the filesystem should hold a |
| 2740 | * reference to the buffer_head that is being marked dirty, which causes |
| 2741 | * try_to_free_buffers() to fail. |
Matthew Wilcox (Oracle) | 6e1cae88 | 2021-06-28 19:36:15 -0700 | [diff] [blame] | 2742 | */ |
Matthew Wilcox (Oracle) | 203a315 | 2021-05-04 11:01:10 -0400 | [diff] [blame] | 2743 | void __folio_mark_dirty(struct folio *folio, struct address_space *mapping, |
Matthew Wilcox (Oracle) | 6e1cae88 | 2021-06-28 19:36:15 -0700 | [diff] [blame] | 2744 | int warn) |
| 2745 | { |
| 2746 | unsigned long flags; |
| 2747 | |
| 2748 | xa_lock_irqsave(&mapping->i_pages, flags); |
Matthew Wilcox (Oracle) | 203a315 | 2021-05-04 11:01:10 -0400 | [diff] [blame] | 2749 | if (folio->mapping) { /* Race with truncate? */ |
| 2750 | WARN_ON_ONCE(warn && !folio_test_uptodate(folio)); |
| 2751 | folio_account_dirtied(folio, mapping); |
| 2752 | __xa_set_mark(&mapping->i_pages, folio_index(folio), |
Matthew Wilcox (Oracle) | 6e1cae88 | 2021-06-28 19:36:15 -0700 | [diff] [blame] | 2753 | PAGECACHE_TAG_DIRTY); |
| 2754 | } |
| 2755 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
| 2756 | } |
| 2757 | |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2758 | /** |
| 2759 | * filemap_dirty_folio - Mark a folio dirty for filesystems which do not use buffer_heads. |
| 2760 | * @mapping: Address space this folio belongs to. |
| 2761 | * @folio: Folio to be marked as dirty. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2762 | * |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2763 | * Filesystems which do not use buffer heads should call this function |
Jianguo Bao | ab428b4 | 2023-09-17 23:04:01 +0800 | [diff] [blame] | 2764 | * from their dirty_folio address space operation. It ignores the |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2765 | * contents of folio_get_private(), so if the filesystem marks individual |
| 2766 | * blocks as dirty, the filesystem should handle that itself. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2767 | * |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2768 | * This is also sometimes used by filesystems which use buffer_heads when |
| 2769 | * a single buffer is being dirtied: we want to set the folio dirty in |
| 2770 | * that case, but not all the buffers. This is a "bottom-up" dirtying, |
Matthew Wilcox (Oracle) | e621900 | 2022-02-09 20:22:12 +0000 | [diff] [blame] | 2771 | * whereas block_dirty_folio() is a "top-down" dirtying. |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2772 | * |
| 2773 | * The caller must ensure this doesn't race with truncation. Most will |
| 2774 | * simply hold the folio lock, but e.g. zap_pte_range() calls with the |
| 2775 | * folio mapped and the pte lock held, which also locks out truncation. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2776 | */ |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2777 | bool filemap_dirty_folio(struct address_space *mapping, struct folio *folio) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2778 | { |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2779 | folio_memcg_lock(folio); |
| 2780 | if (folio_test_set_dirty(folio)) { |
| 2781 | folio_memcg_unlock(folio); |
| 2782 | return false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2783 | } |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2784 | |
| 2785 | __folio_mark_dirty(folio, mapping, !folio_test_private(folio)); |
| 2786 | folio_memcg_unlock(folio); |
| 2787 | |
| 2788 | if (mapping->host) { |
| 2789 | /* !PageAnon && !swapper_space */ |
| 2790 | __mark_inode_dirty(mapping->host, I_DIRTY_PAGES); |
| 2791 | } |
| 2792 | return true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2793 | } |
Matthew Wilcox (Oracle) | 85d4d2e | 2021-05-03 23:30:44 -0400 | [diff] [blame] | 2794 | EXPORT_SYMBOL(filemap_dirty_folio); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2795 | |
Matthew Wilcox (Oracle) | 25ff8b1 | 2021-05-03 10:06:55 -0400 | [diff] [blame] | 2796 | /** |
Matthew Wilcox (Oracle) | cd78ab1 | 2021-05-02 23:22:52 -0400 | [diff] [blame] | 2797 | * folio_redirty_for_writepage - Decline to write a dirty folio. |
| 2798 | * @wbc: The writeback control. |
| 2799 | * @folio: The folio. |
| 2800 | * |
| 2801 | * When a writepage implementation decides that it doesn't want to write |
| 2802 | * @folio for some reason, it should call this function, unlock @folio and |
| 2803 | * return 0. |
| 2804 | * |
| 2805 | * Return: True if we redirtied the folio. False if someone else dirtied |
| 2806 | * it first. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2807 | */ |
Matthew Wilcox (Oracle) | cd78ab1 | 2021-05-02 23:22:52 -0400 | [diff] [blame] | 2808 | bool folio_redirty_for_writepage(struct writeback_control *wbc, |
| 2809 | struct folio *folio) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2810 | { |
Christoph Hellwig | ed2da92 | 2023-06-28 17:31:44 +0200 | [diff] [blame] | 2811 | struct address_space *mapping = folio->mapping; |
Matthew Wilcox (Oracle) | cd78ab1 | 2021-05-02 23:22:52 -0400 | [diff] [blame] | 2812 | long nr = folio_nr_pages(folio); |
Christoph Hellwig | ed2da92 | 2023-06-28 17:31:44 +0200 | [diff] [blame] | 2813 | bool ret; |
Konstantin Khebnikov | 8d38633 | 2015-02-11 15:26:55 -0800 | [diff] [blame] | 2814 | |
Matthew Wilcox (Oracle) | cd78ab1 | 2021-05-02 23:22:52 -0400 | [diff] [blame] | 2815 | wbc->pages_skipped += nr; |
Christoph Hellwig | ed2da92 | 2023-06-28 17:31:44 +0200 | [diff] [blame] | 2816 | ret = filemap_dirty_folio(mapping, folio); |
| 2817 | if (mapping && mapping_can_writeback(mapping)) { |
| 2818 | struct inode *inode = mapping->host; |
| 2819 | struct bdi_writeback *wb; |
| 2820 | struct wb_lock_cookie cookie = {}; |
Matthew Wilcox (Oracle) | cd78ab1 | 2021-05-02 23:22:52 -0400 | [diff] [blame] | 2821 | |
Christoph Hellwig | ed2da92 | 2023-06-28 17:31:44 +0200 | [diff] [blame] | 2822 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
| 2823 | current->nr_dirtied -= nr; |
| 2824 | node_stat_mod_folio(folio, NR_DIRTIED, -nr); |
| 2825 | wb_stat_mod(wb, WB_DIRTIED, -nr); |
| 2826 | unlocked_inode_to_wb_end(inode, &cookie); |
| 2827 | } |
Konstantin Khebnikov | 8d38633 | 2015-02-11 15:26:55 -0800 | [diff] [blame] | 2828 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2829 | } |
Matthew Wilcox (Oracle) | cd78ab1 | 2021-05-02 23:22:52 -0400 | [diff] [blame] | 2830 | EXPORT_SYMBOL(folio_redirty_for_writepage); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2831 | |
Matthew Wilcox (Oracle) | b5e8459 | 2021-04-26 23:53:10 -0400 | [diff] [blame] | 2832 | /** |
| 2833 | * folio_mark_dirty - Mark a folio as being modified. |
| 2834 | * @folio: The folio. |
Wu Fengguang | 6746aff | 2009-09-16 11:50:14 +0200 | [diff] [blame] | 2835 | * |
Matthew Wilcox (Oracle) | 2ca456c | 2022-04-28 14:21:02 -0400 | [diff] [blame] | 2836 | * The folio may not be truncated while this function is running. |
| 2837 | * Holding the folio lock is sufficient to prevent truncation, but some |
| 2838 | * callers cannot acquire a sleeping lock. These callers instead hold |
| 2839 | * the page table lock for a page table which contains at least one page |
| 2840 | * in this folio. Truncation will block on the page table lock as it |
| 2841 | * unmaps pages before removing the folio from its mapping. |
Matthew Wilcox (Oracle) | b5e8459 | 2021-04-26 23:53:10 -0400 | [diff] [blame] | 2842 | * |
| 2843 | * Return: True if the folio was newly dirtied, false if it was already dirty. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2844 | */ |
Matthew Wilcox (Oracle) | b5e8459 | 2021-04-26 23:53:10 -0400 | [diff] [blame] | 2845 | bool folio_mark_dirty(struct folio *folio) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2846 | { |
Matthew Wilcox (Oracle) | b5e8459 | 2021-04-26 23:53:10 -0400 | [diff] [blame] | 2847 | struct address_space *mapping = folio_mapping(folio); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2848 | |
| 2849 | if (likely(mapping)) { |
Minchan Kim | 278df9f | 2011-03-22 16:32:54 -0700 | [diff] [blame] | 2850 | /* |
Vishal Moola (Oracle) | 5a9e347 | 2022-12-21 10:08:48 -0800 | [diff] [blame] | 2851 | * readahead/folio_deactivate could remain |
Matthew Wilcox (Oracle) | 6f31a5a | 2022-02-09 20:22:00 +0000 | [diff] [blame] | 2852 | * PG_readahead/PG_reclaim due to race with folio_end_writeback |
| 2853 | * About readahead, if the folio is written, the flags would be |
Minchan Kim | 278df9f | 2011-03-22 16:32:54 -0700 | [diff] [blame] | 2854 | * reset. So no problem. |
Vishal Moola (Oracle) | 5a9e347 | 2022-12-21 10:08:48 -0800 | [diff] [blame] | 2855 | * About folio_deactivate, if the folio is redirtied, |
Matthew Wilcox (Oracle) | 6f31a5a | 2022-02-09 20:22:00 +0000 | [diff] [blame] | 2856 | * the flag will be reset. So no problem. but if the |
| 2857 | * folio is used by readahead it will confuse readahead |
| 2858 | * and make it restart the size rampup process. But it's |
| 2859 | * a trivial problem. |
Minchan Kim | 278df9f | 2011-03-22 16:32:54 -0700 | [diff] [blame] | 2860 | */ |
Matthew Wilcox (Oracle) | b5e8459 | 2021-04-26 23:53:10 -0400 | [diff] [blame] | 2861 | if (folio_test_reclaim(folio)) |
| 2862 | folio_clear_reclaim(folio); |
Matthew Wilcox (Oracle) | 3a3bae5 | 2022-02-09 20:22:15 +0000 | [diff] [blame] | 2863 | return mapping->a_ops->dirty_folio(mapping, folio); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2864 | } |
Matthew Wilcox (Oracle) | 3a3bae5 | 2022-02-09 20:22:15 +0000 | [diff] [blame] | 2865 | |
| 2866 | return noop_dirty_folio(mapping, folio); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2867 | } |
Matthew Wilcox (Oracle) | b5e8459 | 2021-04-26 23:53:10 -0400 | [diff] [blame] | 2868 | EXPORT_SYMBOL(folio_mark_dirty); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2869 | |
| 2870 | /* |
| 2871 | * set_page_dirty() is racy if the caller has no reference against |
| 2872 | * page->mapping->host, and if the page is unlocked. This is because another |
| 2873 | * CPU could truncate the page off the mapping and then free the mapping. |
| 2874 | * |
| 2875 | * Usually, the page _is_ locked, or the caller is a user-space process which |
| 2876 | * holds a reference on the inode by having an open file. |
| 2877 | * |
| 2878 | * In other cases, the page should be locked before running set_page_dirty(). |
| 2879 | */ |
| 2880 | int set_page_dirty_lock(struct page *page) |
| 2881 | { |
| 2882 | int ret; |
| 2883 | |
Jens Axboe | 7eaceac | 2011-03-10 08:52:07 +0100 | [diff] [blame] | 2884 | lock_page(page); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2885 | ret = set_page_dirty(page); |
| 2886 | unlock_page(page); |
| 2887 | return ret; |
| 2888 | } |
| 2889 | EXPORT_SYMBOL(set_page_dirty_lock); |
| 2890 | |
| 2891 | /* |
Tejun Heo | 11f81be | 2015-05-22 17:13:15 -0400 | [diff] [blame] | 2892 | * This cancels just the dirty bit on the kernel page itself, it does NOT |
| 2893 | * actually remove dirty bits on any mmap's that may be around. It also |
| 2894 | * leaves the page tagged dirty, so any sync activity will still find it on |
| 2895 | * the dirty lists, and in particular, clear_page_dirty_for_io() will still |
| 2896 | * look at the dirty bits in the VM. |
| 2897 | * |
| 2898 | * Doing this should *normally* only ever be done when a page is truncated, |
| 2899 | * and is not actually mapped anywhere at all. However, fs/buffer.c does |
| 2900 | * this when it notices that somebody has cleaned out all the buffers on a |
| 2901 | * page without actually doing it through the VM. Can you say "ext3 is |
| 2902 | * horribly ugly"? Thought you could. |
| 2903 | */ |
Matthew Wilcox (Oracle) | fdaf532 | 2021-03-08 16:43:04 -0500 | [diff] [blame] | 2904 | void __folio_cancel_dirty(struct folio *folio) |
Tejun Heo | 11f81be | 2015-05-22 17:13:15 -0400 | [diff] [blame] | 2905 | { |
Matthew Wilcox (Oracle) | fdaf532 | 2021-03-08 16:43:04 -0500 | [diff] [blame] | 2906 | struct address_space *mapping = folio_mapping(folio); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2907 | |
Christoph Hellwig | f56753a | 2020-09-24 08:51:40 +0200 | [diff] [blame] | 2908 | if (mapping_can_writeback(mapping)) { |
Tejun Heo | 682aa8e | 2015-05-28 14:50:53 -0400 | [diff] [blame] | 2909 | struct inode *inode = mapping->host; |
| 2910 | struct bdi_writeback *wb; |
Greg Thelen | 2e898e4 | 2018-04-20 14:55:42 -0700 | [diff] [blame] | 2911 | struct wb_lock_cookie cookie = {}; |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2912 | |
Matthew Wilcox (Oracle) | fdaf532 | 2021-03-08 16:43:04 -0500 | [diff] [blame] | 2913 | folio_memcg_lock(folio); |
Greg Thelen | 2e898e4 | 2018-04-20 14:55:42 -0700 | [diff] [blame] | 2914 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2915 | |
Matthew Wilcox (Oracle) | fdaf532 | 2021-03-08 16:43:04 -0500 | [diff] [blame] | 2916 | if (folio_test_clear_dirty(folio)) |
Hugh Dickins | 566d336 | 2022-03-24 18:13:59 -0700 | [diff] [blame] | 2917 | folio_account_cleaned(folio, wb); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2918 | |
Greg Thelen | 2e898e4 | 2018-04-20 14:55:42 -0700 | [diff] [blame] | 2919 | unlocked_inode_to_wb_end(inode, &cookie); |
Matthew Wilcox (Oracle) | fdaf532 | 2021-03-08 16:43:04 -0500 | [diff] [blame] | 2920 | folio_memcg_unlock(folio); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2921 | } else { |
Matthew Wilcox (Oracle) | fdaf532 | 2021-03-08 16:43:04 -0500 | [diff] [blame] | 2922 | folio_clear_dirty(folio); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2923 | } |
Tejun Heo | 11f81be | 2015-05-22 17:13:15 -0400 | [diff] [blame] | 2924 | } |
Matthew Wilcox (Oracle) | fdaf532 | 2021-03-08 16:43:04 -0500 | [diff] [blame] | 2925 | EXPORT_SYMBOL(__folio_cancel_dirty); |
Tejun Heo | 11f81be | 2015-05-22 17:13:15 -0400 | [diff] [blame] | 2926 | |
| 2927 | /* |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2928 | * Clear a folio's dirty flag, while caring for dirty memory accounting. |
| 2929 | * Returns true if the folio was previously dirty. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2930 | * |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2931 | * This is for preparing to put the folio under writeout. We leave |
| 2932 | * the folio tagged as dirty in the xarray so that a concurrent |
| 2933 | * write-for-sync can discover it via a PAGECACHE_TAG_DIRTY walk. |
| 2934 | * The ->writepage implementation will run either folio_start_writeback() |
| 2935 | * or folio_mark_dirty(), at which stage we bring the folio's dirty flag |
| 2936 | * and xarray dirty tag back into sync. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2937 | * |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2938 | * This incoherency between the folio's dirty flag and xarray tag is |
| 2939 | * unfortunate, but it only exists while the folio is locked. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2940 | */ |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2941 | bool folio_clear_dirty_for_io(struct folio *folio) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2942 | { |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2943 | struct address_space *mapping = folio_mapping(folio); |
| 2944 | bool ret = false; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2945 | |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2946 | VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio); |
Nick Piggin | 7935289 | 2007-07-19 01:47:22 -0700 | [diff] [blame] | 2947 | |
Christoph Hellwig | f56753a | 2020-09-24 08:51:40 +0200 | [diff] [blame] | 2948 | if (mapping && mapping_can_writeback(mapping)) { |
Tejun Heo | 682aa8e | 2015-05-28 14:50:53 -0400 | [diff] [blame] | 2949 | struct inode *inode = mapping->host; |
| 2950 | struct bdi_writeback *wb; |
Greg Thelen | 2e898e4 | 2018-04-20 14:55:42 -0700 | [diff] [blame] | 2951 | struct wb_lock_cookie cookie = {}; |
Tejun Heo | 682aa8e | 2015-05-28 14:50:53 -0400 | [diff] [blame] | 2952 | |
Linus Torvalds | 7658cc2 | 2006-12-29 10:00:58 -0800 | [diff] [blame] | 2953 | /* |
| 2954 | * Yes, Virginia, this is indeed insane. |
| 2955 | * |
| 2956 | * We use this sequence to make sure that |
| 2957 | * (a) we account for dirty stats properly |
| 2958 | * (b) we tell the low-level filesystem to |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2959 | * mark the whole folio dirty if it was |
Linus Torvalds | 7658cc2 | 2006-12-29 10:00:58 -0800 | [diff] [blame] | 2960 | * dirty in a pagetable. Only to then |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2961 | * (c) clean the folio again and return 1 to |
Linus Torvalds | 7658cc2 | 2006-12-29 10:00:58 -0800 | [diff] [blame] | 2962 | * cause the writeback. |
| 2963 | * |
| 2964 | * This way we avoid all nasty races with the |
| 2965 | * dirty bit in multiple places and clearing |
| 2966 | * them concurrently from different threads. |
| 2967 | * |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2968 | * Note! Normally the "folio_mark_dirty(folio)" |
Linus Torvalds | 7658cc2 | 2006-12-29 10:00:58 -0800 | [diff] [blame] | 2969 | * has no effect on the actual dirty bit - since |
| 2970 | * that will already usually be set. But we |
| 2971 | * need the side effects, and it can help us |
| 2972 | * avoid races. |
| 2973 | * |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2974 | * We basically use the folio "master dirty bit" |
Linus Torvalds | 7658cc2 | 2006-12-29 10:00:58 -0800 | [diff] [blame] | 2975 | * as a serialization point for all the different |
| 2976 | * threads doing their things. |
Linus Torvalds | 7658cc2 | 2006-12-29 10:00:58 -0800 | [diff] [blame] | 2977 | */ |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2978 | if (folio_mkclean(folio)) |
| 2979 | folio_mark_dirty(folio); |
Nick Piggin | 7935289 | 2007-07-19 01:47:22 -0700 | [diff] [blame] | 2980 | /* |
| 2981 | * We carefully synchronise fault handlers against |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2982 | * installing a dirty pte and marking the folio dirty |
Johannes Weiner | 2d6d7f9 | 2015-01-08 14:32:18 -0800 | [diff] [blame] | 2983 | * at this point. We do this by having them hold the |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2984 | * page lock while dirtying the folio, and folios are |
Johannes Weiner | 2d6d7f9 | 2015-01-08 14:32:18 -0800 | [diff] [blame] | 2985 | * always locked coming in here, so we get the desired |
| 2986 | * exclusion. |
Nick Piggin | 7935289 | 2007-07-19 01:47:22 -0700 | [diff] [blame] | 2987 | */ |
Greg Thelen | 2e898e4 | 2018-04-20 14:55:42 -0700 | [diff] [blame] | 2988 | wb = unlocked_inode_to_wb_begin(inode, &cookie); |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2989 | if (folio_test_clear_dirty(folio)) { |
| 2990 | long nr = folio_nr_pages(folio); |
| 2991 | lruvec_stat_mod_folio(folio, NR_FILE_DIRTY, -nr); |
| 2992 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); |
| 2993 | wb_stat_mod(wb, WB_RECLAIMABLE, -nr); |
| 2994 | ret = true; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2995 | } |
Greg Thelen | 2e898e4 | 2018-04-20 14:55:42 -0700 | [diff] [blame] | 2996 | unlocked_inode_to_wb_end(inode, &cookie); |
Greg Thelen | c4843a7 | 2015-05-22 17:13:16 -0400 | [diff] [blame] | 2997 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2998 | } |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 2999 | return folio_test_clear_dirty(folio); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3000 | } |
Matthew Wilcox (Oracle) | 9350f20 | 2021-02-28 16:21:20 -0500 | [diff] [blame] | 3001 | EXPORT_SYMBOL(folio_clear_dirty_for_io); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3002 | |
Jan Kara | 633a2ab | 2021-09-02 14:53:04 -0700 | [diff] [blame] | 3003 | static void wb_inode_writeback_start(struct bdi_writeback *wb) |
| 3004 | { |
| 3005 | atomic_inc(&wb->writeback_inodes); |
| 3006 | } |
| 3007 | |
| 3008 | static void wb_inode_writeback_end(struct bdi_writeback *wb) |
| 3009 | { |
Khazhismel Kumykov | f87904c | 2022-08-01 08:50:34 -0700 | [diff] [blame] | 3010 | unsigned long flags; |
Jan Kara | 633a2ab | 2021-09-02 14:53:04 -0700 | [diff] [blame] | 3011 | atomic_dec(&wb->writeback_inodes); |
Jan Kara | 45a2966 | 2021-09-02 14:53:09 -0700 | [diff] [blame] | 3012 | /* |
| 3013 | * Make sure estimate of writeback throughput gets updated after |
| 3014 | * writeback completed. We delay the update by BANDWIDTH_INTERVAL |
| 3015 | * (which is the interval other bandwidth updates use for batching) so |
| 3016 | * that if multiple inodes end writeback at a similar time, they get |
| 3017 | * batched into one bandwidth update. |
| 3018 | */ |
Khazhismel Kumykov | f87904c | 2022-08-01 08:50:34 -0700 | [diff] [blame] | 3019 | spin_lock_irqsave(&wb->work_lock, flags); |
| 3020 | if (test_bit(WB_registered, &wb->state)) |
| 3021 | queue_delayed_work(bdi_wq, &wb->bw_dwork, BANDWIDTH_INTERVAL); |
| 3022 | spin_unlock_irqrestore(&wb->work_lock, flags); |
Jan Kara | 633a2ab | 2021-09-02 14:53:04 -0700 | [diff] [blame] | 3023 | } |
| 3024 | |
Matthew Wilcox (Oracle) | 2580d55 | 2023-10-04 17:53:17 +0100 | [diff] [blame] | 3025 | bool __folio_end_writeback(struct folio *folio) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3026 | { |
Matthew Wilcox (Oracle) | 269ccca3 | 2021-01-15 23:34:16 -0500 | [diff] [blame] | 3027 | long nr = folio_nr_pages(folio); |
| 3028 | struct address_space *mapping = folio_mapping(folio); |
Matthew Wilcox (Oracle) | 2580d55 | 2023-10-04 17:53:17 +0100 | [diff] [blame] | 3029 | bool ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3030 | |
Matthew Wilcox (Oracle) | 269ccca3 | 2021-01-15 23:34:16 -0500 | [diff] [blame] | 3031 | folio_memcg_lock(folio); |
Huang Ying | 371a096 | 2016-10-07 16:59:30 -0700 | [diff] [blame] | 3032 | if (mapping && mapping_use_writeback_tags(mapping)) { |
Tejun Heo | 9101813 | 2015-05-22 17:13:39 -0400 | [diff] [blame] | 3033 | struct inode *inode = mapping->host; |
| 3034 | struct backing_dev_info *bdi = inode_to_bdi(inode); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3035 | unsigned long flags; |
| 3036 | |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 3037 | xa_lock_irqsave(&mapping->i_pages, flags); |
Matthew Wilcox (Oracle) | 2580d55 | 2023-10-04 17:53:17 +0100 | [diff] [blame] | 3038 | ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback); |
Matthew Wilcox (Oracle) | 7d0795d | 2023-10-04 17:53:16 +0100 | [diff] [blame] | 3039 | __xa_clear_mark(&mapping->i_pages, folio_index(folio), |
| 3040 | PAGECACHE_TAG_WRITEBACK); |
| 3041 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { |
| 3042 | struct bdi_writeback *wb = inode_to_wb(inode); |
Tejun Heo | 9101813 | 2015-05-22 17:13:39 -0400 | [diff] [blame] | 3043 | |
Matthew Wilcox (Oracle) | 7d0795d | 2023-10-04 17:53:16 +0100 | [diff] [blame] | 3044 | wb_stat_mod(wb, WB_WRITEBACK, -nr); |
| 3045 | __wb_writeout_add(wb, nr); |
| 3046 | if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) |
| 3047 | wb_inode_writeback_end(wb); |
Peter Zijlstra | 69cb51d | 2007-10-16 23:25:48 -0700 | [diff] [blame] | 3048 | } |
Dave Chinner | 6c60d2b | 2016-07-26 15:21:50 -0700 | [diff] [blame] | 3049 | |
| 3050 | if (mapping->host && !mapping_tagged(mapping, |
| 3051 | PAGECACHE_TAG_WRITEBACK)) |
| 3052 | sb_clear_inode_writeback(mapping->host); |
| 3053 | |
Matthew Wilcox | b93b016 | 2018-04-10 16:36:56 -0700 | [diff] [blame] | 3054 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3055 | } else { |
Matthew Wilcox (Oracle) | 2580d55 | 2023-10-04 17:53:17 +0100 | [diff] [blame] | 3056 | ret = folio_xor_flags_has_waiters(folio, 1 << PG_writeback); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3057 | } |
Matthew Wilcox (Oracle) | 7d0795d | 2023-10-04 17:53:16 +0100 | [diff] [blame] | 3058 | |
| 3059 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, -nr); |
| 3060 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, -nr); |
| 3061 | node_stat_mod_folio(folio, NR_WRITTEN, nr); |
Matthew Wilcox (Oracle) | 269ccca3 | 2021-01-15 23:34:16 -0500 | [diff] [blame] | 3062 | folio_memcg_unlock(folio); |
Matthew Wilcox (Oracle) | 2580d55 | 2023-10-04 17:53:17 +0100 | [diff] [blame] | 3063 | |
| 3064 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3065 | } |
| 3066 | |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3067 | void __folio_start_writeback(struct folio *folio, bool keep_write) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3068 | { |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3069 | long nr = folio_nr_pages(folio); |
| 3070 | struct address_space *mapping = folio_mapping(folio); |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3071 | int access_ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3072 | |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3073 | VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio); |
| 3074 | |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3075 | folio_memcg_lock(folio); |
Huang Ying | 371a096 | 2016-10-07 16:59:30 -0700 | [diff] [blame] | 3076 | if (mapping && mapping_use_writeback_tags(mapping)) { |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3077 | XA_STATE(xas, &mapping->i_pages, folio_index(folio)); |
Tejun Heo | 9101813 | 2015-05-22 17:13:39 -0400 | [diff] [blame] | 3078 | struct inode *inode = mapping->host; |
| 3079 | struct backing_dev_info *bdi = inode_to_bdi(inode); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3080 | unsigned long flags; |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3081 | bool on_wblist; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3082 | |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 3083 | xas_lock_irqsave(&xas, flags); |
| 3084 | xas_load(&xas); |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3085 | folio_test_set_writeback(folio); |
Dave Chinner | 6c60d2b | 2016-07-26 15:21:50 -0700 | [diff] [blame] | 3086 | |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3087 | on_wblist = mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK); |
Dave Chinner | 6c60d2b | 2016-07-26 15:21:50 -0700 | [diff] [blame] | 3088 | |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3089 | xas_set_mark(&xas, PAGECACHE_TAG_WRITEBACK); |
| 3090 | if (bdi->capabilities & BDI_CAP_WRITEBACK_ACCT) { |
| 3091 | struct bdi_writeback *wb = inode_to_wb(inode); |
Jan Kara | 633a2ab | 2021-09-02 14:53:04 -0700 | [diff] [blame] | 3092 | |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3093 | wb_stat_mod(wb, WB_WRITEBACK, nr); |
| 3094 | if (!on_wblist) |
| 3095 | wb_inode_writeback_start(wb); |
Peter Zijlstra | 69cb51d | 2007-10-16 23:25:48 -0700 | [diff] [blame] | 3096 | } |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3097 | |
| 3098 | /* |
| 3099 | * We can come through here when swapping anonymous |
| 3100 | * folios, so we don't necessarily have an inode to |
| 3101 | * track for sync. |
| 3102 | */ |
| 3103 | if (mapping->host && !on_wblist) |
| 3104 | sb_mark_inode_writeback(mapping->host); |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3105 | if (!folio_test_dirty(folio)) |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 3106 | xas_clear_mark(&xas, PAGECACHE_TAG_DIRTY); |
Namjae Jeon | 1c8349a | 2014-05-12 08:12:25 -0400 | [diff] [blame] | 3107 | if (!keep_write) |
Matthew Wilcox | ff9c745 | 2017-11-22 11:41:23 -0500 | [diff] [blame] | 3108 | xas_clear_mark(&xas, PAGECACHE_TAG_TOWRITE); |
| 3109 | xas_unlock_irqrestore(&xas, flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3110 | } else { |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3111 | folio_test_set_writeback(folio); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3112 | } |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3113 | |
| 3114 | lruvec_stat_mod_folio(folio, NR_WRITEBACK, nr); |
| 3115 | zone_stat_mod_folio(folio, NR_ZONE_WRITE_PENDING, nr); |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3116 | folio_memcg_unlock(folio); |
Matthew Wilcox (Oracle) | b5612c3 | 2023-11-08 20:46:05 +0000 | [diff] [blame] | 3117 | |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3118 | access_ret = arch_make_folio_accessible(folio); |
Claudio Imbrenda | f28d436 | 2020-04-01 21:05:56 -0700 | [diff] [blame] | 3119 | /* |
| 3120 | * If writeback has been triggered on a page that cannot be made |
| 3121 | * accessible, it is too late to recover here. |
| 3122 | */ |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3123 | VM_BUG_ON_FOLIO(access_ret != 0, folio); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3124 | } |
Matthew Wilcox (Oracle) | f143f1e | 2021-04-24 12:00:48 -0400 | [diff] [blame] | 3125 | EXPORT_SYMBOL(__folio_start_writeback); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3126 | |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3127 | /** |
| 3128 | * folio_wait_writeback - Wait for a folio to finish writeback. |
| 3129 | * @folio: The folio to wait for. |
| 3130 | * |
| 3131 | * If the folio is currently being written back to storage, wait for the |
| 3132 | * I/O to complete. |
| 3133 | * |
| 3134 | * Context: Sleeps. Must be called in process context and with |
| 3135 | * no spinlocks held. Caller should hold a reference on the folio. |
| 3136 | * If the folio is not locked, writeback may start again after writeback |
| 3137 | * has finished. |
Yafang Shao | 19343b5 | 2019-05-13 17:23:11 -0700 | [diff] [blame] | 3138 | */ |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3139 | void folio_wait_writeback(struct folio *folio) |
Yafang Shao | 19343b5 | 2019-05-13 17:23:11 -0700 | [diff] [blame] | 3140 | { |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3141 | while (folio_test_writeback(folio)) { |
Matthew Wilcox (Oracle) | b9b0ff6 | 2021-07-13 22:45:50 -0400 | [diff] [blame] | 3142 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
Matthew Wilcox (Oracle) | 101c0bf | 2021-03-04 12:02:54 -0500 | [diff] [blame] | 3143 | folio_wait_bit(folio, PG_writeback); |
Yafang Shao | 19343b5 | 2019-05-13 17:23:11 -0700 | [diff] [blame] | 3144 | } |
| 3145 | } |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3146 | EXPORT_SYMBOL_GPL(folio_wait_writeback); |
Yafang Shao | 19343b5 | 2019-05-13 17:23:11 -0700 | [diff] [blame] | 3147 | |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3148 | /** |
| 3149 | * folio_wait_writeback_killable - Wait for a folio to finish writeback. |
| 3150 | * @folio: The folio to wait for. |
| 3151 | * |
| 3152 | * If the folio is currently being written back to storage, wait for the |
| 3153 | * I/O to complete or a fatal signal to arrive. |
| 3154 | * |
| 3155 | * Context: Sleeps. Must be called in process context and with |
| 3156 | * no spinlocks held. Caller should hold a reference on the folio. |
| 3157 | * If the folio is not locked, writeback may start again after writeback |
| 3158 | * has finished. |
| 3159 | * Return: 0 on success, -EINTR if we get a fatal signal while waiting. |
Matthew Wilcox (Oracle) | e5dbd33 | 2021-03-20 05:40:39 +0000 | [diff] [blame] | 3160 | */ |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3161 | int folio_wait_writeback_killable(struct folio *folio) |
Matthew Wilcox (Oracle) | e5dbd33 | 2021-03-20 05:40:39 +0000 | [diff] [blame] | 3162 | { |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3163 | while (folio_test_writeback(folio)) { |
Matthew Wilcox (Oracle) | b9b0ff6 | 2021-07-13 22:45:50 -0400 | [diff] [blame] | 3164 | trace_folio_wait_writeback(folio, folio_mapping(folio)); |
Matthew Wilcox (Oracle) | 101c0bf | 2021-03-04 12:02:54 -0500 | [diff] [blame] | 3165 | if (folio_wait_bit_killable(folio, PG_writeback)) |
Matthew Wilcox (Oracle) | e5dbd33 | 2021-03-20 05:40:39 +0000 | [diff] [blame] | 3166 | return -EINTR; |
| 3167 | } |
| 3168 | |
| 3169 | return 0; |
| 3170 | } |
Matthew Wilcox (Oracle) | 490e016 | 2021-03-04 11:09:17 -0500 | [diff] [blame] | 3171 | EXPORT_SYMBOL_GPL(folio_wait_writeback_killable); |
Matthew Wilcox (Oracle) | e5dbd33 | 2021-03-20 05:40:39 +0000 | [diff] [blame] | 3172 | |
Darrick J. Wong | 1d1d1a7 | 2013-02-21 16:42:51 -0800 | [diff] [blame] | 3173 | /** |
Matthew Wilcox (Oracle) | a49d0c5 | 2021-03-04 11:25:25 -0500 | [diff] [blame] | 3174 | * folio_wait_stable() - wait for writeback to finish, if necessary. |
| 3175 | * @folio: The folio to wait on. |
Darrick J. Wong | 1d1d1a7 | 2013-02-21 16:42:51 -0800 | [diff] [blame] | 3176 | * |
Matthew Wilcox (Oracle) | a49d0c5 | 2021-03-04 11:25:25 -0500 | [diff] [blame] | 3177 | * This function determines if the given folio is related to a backing |
| 3178 | * device that requires folio contents to be held stable during writeback. |
| 3179 | * If so, then it will wait for any pending writeback to complete. |
| 3180 | * |
| 3181 | * Context: Sleeps. Must be called in process context and with |
| 3182 | * no spinlocks held. Caller should hold a reference on the folio. |
| 3183 | * If the folio is not locked, writeback may start again after writeback |
| 3184 | * has finished. |
Darrick J. Wong | 1d1d1a7 | 2013-02-21 16:42:51 -0800 | [diff] [blame] | 3185 | */ |
Matthew Wilcox (Oracle) | a49d0c5 | 2021-03-04 11:25:25 -0500 | [diff] [blame] | 3186 | void folio_wait_stable(struct folio *folio) |
Darrick J. Wong | 1d1d1a7 | 2013-02-21 16:42:51 -0800 | [diff] [blame] | 3187 | { |
Christoph Hellwig | 762321d | 2023-10-25 16:10:17 +0200 | [diff] [blame] | 3188 | if (mapping_stable_writes(folio_mapping(folio))) |
Matthew Wilcox (Oracle) | a49d0c5 | 2021-03-04 11:25:25 -0500 | [diff] [blame] | 3189 | folio_wait_writeback(folio); |
Darrick J. Wong | 1d1d1a7 | 2013-02-21 16:42:51 -0800 | [diff] [blame] | 3190 | } |
Matthew Wilcox (Oracle) | a49d0c5 | 2021-03-04 11:25:25 -0500 | [diff] [blame] | 3191 | EXPORT_SYMBOL_GPL(folio_wait_stable); |