WSL2-Linux-Kernel/include/linux/backing-dev.h

360 строки
10 KiB
C
Исходник Обычный вид История

/*
* include/linux/backing-dev.h
*
* low-level device information and state which is propagated up through
* to high-level code.
*/
#ifndef _LINUX_BACKING_DEV_H
#define _LINUX_BACKING_DEV_H
#include <linux/percpu_counter.h>
#include <linux/log2.h>
#include <linux/flex_proportions.h>
#include <linux/kernel.h>
#include <linux/fs.h>
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
#include <linux/sched.h>
#include <linux/timer.h>
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
#include <linux/writeback.h>
#include <linux/atomic.h>
#include <linux/sysctl.h>
writeback: replace custom worker pool implementation with unbound workqueue Writeback implements its own worker pool - each bdi can be associated with a worker thread which is created and destroyed dynamically. The worker thread for the default bdi is always present and serves as the "forker" thread which forks off worker threads for other bdis. there's no reason for writeback to implement its own worker pool when using unbound workqueue instead is much simpler and more efficient. This patch replaces custom worker pool implementation in writeback with an unbound workqueue. The conversion isn't too complicated but the followings are worth mentioning. * bdi_writeback->last_active, task and wakeup_timer are removed. delayed_work ->dwork is added instead. Explicit timer handling is no longer necessary. Everything works by either queueing / modding / flushing / canceling the delayed_work item. * bdi_writeback_thread() becomes bdi_writeback_workfn() which runs off bdi_writeback->dwork. On each execution, it processes bdi->work_list and reschedules itself if there are more things to do. The function also handles low-mem condition, which used to be handled by the forker thread. If the function is running off a rescuer thread, it only writes out limited number of pages so that the rescuer can serve other bdis too. This preserves the flusher creation failure behavior of the forker thread. * INIT_LIST_HEAD(&bdi->bdi_list) is used to tell bdi_writeback_workfn() about on-going bdi unregistration so that it always drains work_list even if it's running off the rescuer. Note that the original code was broken in this regard. Under memory pressure, a bdi could finish unregistration with non-empty work_list. * The default bdi is no longer special. It now is treated the same as any other bdi and bdi_cap_flush_forker() is removed. * BDI_pending is no longer used. Removed. * Some tracepoints become non-applicable. The following TPs are removed - writeback_nothread, writeback_wake_thread, writeback_wake_forker_thread, writeback_thread_start, writeback_thread_stop. Everything, including devices coming and going away and rescuer operation under simulated memory pressure, seems to work fine in my test setup. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Jens Axboe <axboe@kernel.dk> Cc: Fengguang Wu <fengguang.wu@intel.com> Cc: Jeff Moyer <jmoyer@redhat.com>
2013-04-02 06:08:06 +04:00
#include <linux/workqueue.h>
struct page;
struct device;
struct dentry;
/*
* Bits in backing_dev_info.state
*/
enum bdi_state {
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
BDI_wb_alloc, /* Default embedded wb allocated */
BDI_async_congested, /* The async (write) queue is getting full */
BDI_sync_congested, /* The sync queue is getting full */
BDI_registered, /* bdi_register() was done */
BDI_writeback_running, /* Writeback is in progress */
BDI_unused, /* Available bits start here */
};
typedef int (congested_fn)(void *, int);
enum bdi_stat_item {
BDI_RECLAIMABLE,
BDI_WRITEBACK,
BDI_DIRTIED,
BDI_WRITTEN,
NR_BDI_STAT_ITEMS
};
#define BDI_STAT_BATCH (8*(1+ilog2(nr_cpu_ids)))
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
struct bdi_writeback {
struct backing_dev_info *bdi; /* our parent bdi */
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
unsigned int nr;
unsigned long last_old_flush; /* last old data flush */
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
writeback: replace custom worker pool implementation with unbound workqueue Writeback implements its own worker pool - each bdi can be associated with a worker thread which is created and destroyed dynamically. The worker thread for the default bdi is always present and serves as the "forker" thread which forks off worker threads for other bdis. there's no reason for writeback to implement its own worker pool when using unbound workqueue instead is much simpler and more efficient. This patch replaces custom worker pool implementation in writeback with an unbound workqueue. The conversion isn't too complicated but the followings are worth mentioning. * bdi_writeback->last_active, task and wakeup_timer are removed. delayed_work ->dwork is added instead. Explicit timer handling is no longer necessary. Everything works by either queueing / modding / flushing / canceling the delayed_work item. * bdi_writeback_thread() becomes bdi_writeback_workfn() which runs off bdi_writeback->dwork. On each execution, it processes bdi->work_list and reschedules itself if there are more things to do. The function also handles low-mem condition, which used to be handled by the forker thread. If the function is running off a rescuer thread, it only writes out limited number of pages so that the rescuer can serve other bdis too. This preserves the flusher creation failure behavior of the forker thread. * INIT_LIST_HEAD(&bdi->bdi_list) is used to tell bdi_writeback_workfn() about on-going bdi unregistration so that it always drains work_list even if it's running off the rescuer. Note that the original code was broken in this regard. Under memory pressure, a bdi could finish unregistration with non-empty work_list. * The default bdi is no longer special. It now is treated the same as any other bdi and bdi_cap_flush_forker() is removed. * BDI_pending is no longer used. Removed. * Some tracepoints become non-applicable. The following TPs are removed - writeback_nothread, writeback_wake_thread, writeback_wake_forker_thread, writeback_thread_start, writeback_thread_stop. Everything, including devices coming and going away and rescuer operation under simulated memory pressure, seems to work fine in my test setup. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Jens Axboe <axboe@kernel.dk> Cc: Fengguang Wu <fengguang.wu@intel.com> Cc: Jeff Moyer <jmoyer@redhat.com>
2013-04-02 06:08:06 +04:00
struct delayed_work dwork; /* work item used for writeback */
struct list_head b_dirty; /* dirty inodes */
struct list_head b_io; /* parked for writeback */
struct list_head b_more_io; /* parked for more writeback */
writeback: split inode_wb_list_lock into bdi_writeback.list_lock Split the global inode_wb_list_lock into a per-bdi_writeback list_lock, as it's currently the most contended lock in the system for metadata heavy workloads. It won't help for single-filesystem workloads for which we'll need the I/O-less balance_dirty_pages, but at least we can dedicate a cpu to spinning on each bdi now for larger systems. Based on earlier patches from Nick Piggin and Dave Chinner. It reduces lock contentions to 1/4 in this test case: 10 HDD JBOD, 100 dd on each disk, XFS, 6GB ram lock_stat version 0.3 ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- class name con-bounces contentions waittime-min waittime-max waittime-total acq-bounces acquisitions holdtime-min holdtime-max holdtime-total ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- vanilla 2.6.39-rc3: inode_wb_list_lock: 42590 44433 0.12 147.74 144127.35 252274 886792 0.08 121.34 917211.23 ------------------ inode_wb_list_lock 2 [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85 inode_wb_list_lock 34 [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49 inode_wb_list_lock 12893 [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0 inode_wb_list_lock 10702 [<ffffffff8115afef>] writeback_single_inode+0x16d/0x20a ------------------ inode_wb_list_lock 2 [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85 inode_wb_list_lock 19 [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49 inode_wb_list_lock 5550 [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0 inode_wb_list_lock 8511 [<ffffffff8115b4ad>] writeback_sb_inodes+0x10f/0x157 2.6.39-rc3 + patch: &(&wb->list_lock)->rlock: 11383 11657 0.14 151.69 40429.51 90825 527918 0.11 145.90 556843.37 ------------------------ &(&wb->list_lock)->rlock 10 [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86 &(&wb->list_lock)->rlock 1493 [<ffffffff8115b1ed>] writeback_inodes_wb+0x3d/0x150 &(&wb->list_lock)->rlock 3652 [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f &(&wb->list_lock)->rlock 1412 [<ffffffff8115a38e>] writeback_single_inode+0x17f/0x223 ------------------------ &(&wb->list_lock)->rlock 3 [<ffffffff8110b5af>] bdi_lock_two+0x46/0x4b &(&wb->list_lock)->rlock 6 [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86 &(&wb->list_lock)->rlock 2061 [<ffffffff8115af97>] __mark_inode_dirty+0x173/0x1cf &(&wb->list_lock)->rlock 2629 [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f hughd@google.com: fix recursive lock when bdi_lock_two() is called with new the same as old akpm@linux-foundation.org: cleanup bdev_inode_switch_bdi() comment Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-04-22 04:19:44 +04:00
spinlock_t list_lock; /* protects the b_* lists */
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
};
struct backing_dev_info {
struct list_head bdi_list;
unsigned long ra_pages; /* max readahead in PAGE_CACHE_SIZE units */
unsigned long state; /* Always use atomic bitops on this */
unsigned int capabilities; /* Device capabilities */
congested_fn *congested_fn; /* Function pointer if device is md/dm */
void *congested_data; /* Pointer to aux data for congested func */
char *name;
struct percpu_counter bdi_stat[NR_BDI_STAT_ITEMS];
mm: per device dirty threshold Scale writeback cache per backing device, proportional to its writeout speed. By decoupling the BDI dirty thresholds a number of problems we currently have will go away, namely: - mutual interference starvation (for any number of BDIs); - deadlocks with stacked BDIs (loop, FUSE and local NFS mounts). It might be that all dirty pages are for a single BDI while other BDIs are idling. By giving each BDI a 'fair' share of the dirty limit, each one can have dirty pages outstanding and make progress. A global threshold also creates a deadlock for stacked BDIs; when A writes to B, and A generates enough dirty pages to get throttled, B will never start writeback until the dirty pages go away. Again, by giving each BDI its own 'independent' dirty limit, this problem is avoided. So the problem is to determine how to distribute the total dirty limit across the BDIs fairly and efficiently. A DBI that has a large dirty limit but does not have any dirty pages outstanding is a waste. What is done is to keep a floating proportion between the DBIs based on writeback completions. This way faster/more active devices get a larger share than slower/idle devices. [akpm@linux-foundation.org: fix warnings] [hugh@veritas.com: Fix occasional hang when a task couldn't get out of balance_dirty_pages] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-17 10:25:50 +04:00
writeback: bdi write bandwidth estimation The estimation value will start from 100MB/s and adapt to the real bandwidth in seconds. It tries to update the bandwidth only when disk is fully utilized. Any inactive period of more than one second will be skipped. The estimated bandwidth will be reflecting how fast the device can writeout when _fully utilized_, and won't drop to 0 when it goes idle. The value will remain constant at disk idle time. At busy write time, if not considering fluctuations, it will also remain high unless be knocked down by possible concurrent reads that compete for the disk time and bandwidth with async writes. The estimation is not done purely in the flusher because there is no guarantee for write_cache_pages() to return timely to update bandwidth. The bdi->avg_write_bandwidth smoothing is very effective for filtering out sudden spikes, however may be a little biased in long term. The overheads are low because the bdi bandwidth update only occurs at 200ms intervals. The 200ms update interval is suitable, because it's not possible to get the real bandwidth for the instance at all, due to large fluctuations. The NFS commits can be as large as seconds worth of data. One XFS completion may be as large as half second worth of data if we are going to increase the write chunk to half second worth of data. In ext4, fluctuations with time period of around 5 seconds is observed. And there is another pattern of irregular periods of up to 20 seconds on SSD tests. That's why we are not only doing the estimation at 200ms intervals, but also averaging them over a period of 3 seconds and then go further to do another level of smoothing in avg_write_bandwidth. CC: Li Shaohua <shaohua.li@intel.com> CC: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2010-08-29 21:22:30 +04:00
unsigned long bw_time_stamp; /* last time write bw is updated */
writeback: dirty rate control It's all about bdi->dirty_ratelimit, which aims to be (write_bw / N) when there are N dd tasks. On write() syscall, use bdi->dirty_ratelimit ============================================ balance_dirty_pages(pages_dirtied) { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); pause = pages_dirtied / task_ratelimit; sleep(pause); } On every 200ms, update bdi->dirty_ratelimit =========================================== bdi_update_dirty_ratelimit() { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate; bdi->dirty_ratelimit = balanced_dirty_ratelimit } Estimation of balanced bdi->dirty_ratelimit =========================================== balanced task_ratelimit ----------------------- balance_dirty_pages() needs to throttle tasks dirtying pages such that the total amount of dirty pages stays below the specified dirty limit in order to avoid memory deadlocks. Furthermore we desire fairness in that tasks get throttled proportionally to the amount of pages they dirty. IOW we want to throttle tasks such that we match the dirty rate to the writeout bandwidth, this yields a stable amount of dirty pages: dirty_rate == write_bw (1) The fairness requirement gives us: task_ratelimit = balanced_dirty_ratelimit == write_bw / N (2) where N is the number of dd tasks. We don't know N beforehand, but still can estimate balanced_dirty_ratelimit within 200ms. Start by throttling each dd task at rate task_ratelimit = task_ratelimit_0 (3) (any non-zero initial value is OK) After 200ms, we measured dirty_rate = # of pages dirtied by all dd's / 200ms write_bw = # of pages written to the disk / 200ms For the aggressive dd dirtiers, the equality holds dirty_rate == N * task_rate == N * task_ratelimit_0 (4) Or task_ratelimit_0 == dirty_rate / N (5) Now we conclude that the balanced task ratelimit can be estimated by write_bw balanced_dirty_ratelimit = task_ratelimit_0 * ---------- (6) dirty_rate Because with (4) and (5) we can get the desired equality (1): write_bw balanced_dirty_ratelimit == (dirty_rate / N) * ---------- dirty_rate == write_bw / N Then using the balanced task ratelimit we can compute task pause times like: task_pause = task->nr_dirtied / task_ratelimit task_ratelimit with position control ------------------------------------ However, while the above gives us means of matching the dirty rate to the writeout bandwidth, it at best provides us with a stable dirty page count (assuming a static system). In order to control the dirty page count such that it is high enough to provide performance, but does not exceed the specified limit we need another control. The dirty position control works by extending (2) to task_ratelimit = balanced_dirty_ratelimit * pos_ratio (7) where pos_ratio is a negative feedback function that subjects to 1) f(setpoint) = 1.0 2) df/dx < 0 That is, if the dirty pages are ABOVE the setpoint, we throttle each task a bit more HEAVY than balanced_dirty_ratelimit, so that the dirty pages are created less fast than they are cleaned, thus DROP to the setpoints (and the reverse). Based on (7) and the assumption that both dirty_ratelimit and pos_ratio remains CONSTANT for the past 200ms, we get task_ratelimit_0 = balanced_dirty_ratelimit * pos_ratio (8) Putting (8) into (6), we get the formula used in bdi_update_dirty_ratelimit(): write_bw balanced_dirty_ratelimit *= pos_ratio * ---------- (9) dirty_rate Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-06-12 20:51:31 +04:00
unsigned long dirtied_stamp;
writeback: bdi write bandwidth estimation The estimation value will start from 100MB/s and adapt to the real bandwidth in seconds. It tries to update the bandwidth only when disk is fully utilized. Any inactive period of more than one second will be skipped. The estimated bandwidth will be reflecting how fast the device can writeout when _fully utilized_, and won't drop to 0 when it goes idle. The value will remain constant at disk idle time. At busy write time, if not considering fluctuations, it will also remain high unless be knocked down by possible concurrent reads that compete for the disk time and bandwidth with async writes. The estimation is not done purely in the flusher because there is no guarantee for write_cache_pages() to return timely to update bandwidth. The bdi->avg_write_bandwidth smoothing is very effective for filtering out sudden spikes, however may be a little biased in long term. The overheads are low because the bdi bandwidth update only occurs at 200ms intervals. The 200ms update interval is suitable, because it's not possible to get the real bandwidth for the instance at all, due to large fluctuations. The NFS commits can be as large as seconds worth of data. One XFS completion may be as large as half second worth of data if we are going to increase the write chunk to half second worth of data. In ext4, fluctuations with time period of around 5 seconds is observed. And there is another pattern of irregular periods of up to 20 seconds on SSD tests. That's why we are not only doing the estimation at 200ms intervals, but also averaging them over a period of 3 seconds and then go further to do another level of smoothing in avg_write_bandwidth. CC: Li Shaohua <shaohua.li@intel.com> CC: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2010-08-29 21:22:30 +04:00
unsigned long written_stamp; /* pages written at bw_time_stamp */
unsigned long write_bandwidth; /* the estimated write bandwidth */
unsigned long avg_write_bandwidth; /* further smoothed write bw */
writeback: dirty rate control It's all about bdi->dirty_ratelimit, which aims to be (write_bw / N) when there are N dd tasks. On write() syscall, use bdi->dirty_ratelimit ============================================ balance_dirty_pages(pages_dirtied) { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); pause = pages_dirtied / task_ratelimit; sleep(pause); } On every 200ms, update bdi->dirty_ratelimit =========================================== bdi_update_dirty_ratelimit() { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate; bdi->dirty_ratelimit = balanced_dirty_ratelimit } Estimation of balanced bdi->dirty_ratelimit =========================================== balanced task_ratelimit ----------------------- balance_dirty_pages() needs to throttle tasks dirtying pages such that the total amount of dirty pages stays below the specified dirty limit in order to avoid memory deadlocks. Furthermore we desire fairness in that tasks get throttled proportionally to the amount of pages they dirty. IOW we want to throttle tasks such that we match the dirty rate to the writeout bandwidth, this yields a stable amount of dirty pages: dirty_rate == write_bw (1) The fairness requirement gives us: task_ratelimit = balanced_dirty_ratelimit == write_bw / N (2) where N is the number of dd tasks. We don't know N beforehand, but still can estimate balanced_dirty_ratelimit within 200ms. Start by throttling each dd task at rate task_ratelimit = task_ratelimit_0 (3) (any non-zero initial value is OK) After 200ms, we measured dirty_rate = # of pages dirtied by all dd's / 200ms write_bw = # of pages written to the disk / 200ms For the aggressive dd dirtiers, the equality holds dirty_rate == N * task_rate == N * task_ratelimit_0 (4) Or task_ratelimit_0 == dirty_rate / N (5) Now we conclude that the balanced task ratelimit can be estimated by write_bw balanced_dirty_ratelimit = task_ratelimit_0 * ---------- (6) dirty_rate Because with (4) and (5) we can get the desired equality (1): write_bw balanced_dirty_ratelimit == (dirty_rate / N) * ---------- dirty_rate == write_bw / N Then using the balanced task ratelimit we can compute task pause times like: task_pause = task->nr_dirtied / task_ratelimit task_ratelimit with position control ------------------------------------ However, while the above gives us means of matching the dirty rate to the writeout bandwidth, it at best provides us with a stable dirty page count (assuming a static system). In order to control the dirty page count such that it is high enough to provide performance, but does not exceed the specified limit we need another control. The dirty position control works by extending (2) to task_ratelimit = balanced_dirty_ratelimit * pos_ratio (7) where pos_ratio is a negative feedback function that subjects to 1) f(setpoint) = 1.0 2) df/dx < 0 That is, if the dirty pages are ABOVE the setpoint, we throttle each task a bit more HEAVY than balanced_dirty_ratelimit, so that the dirty pages are created less fast than they are cleaned, thus DROP to the setpoints (and the reverse). Based on (7) and the assumption that both dirty_ratelimit and pos_ratio remains CONSTANT for the past 200ms, we get task_ratelimit_0 = balanced_dirty_ratelimit * pos_ratio (8) Putting (8) into (6), we get the formula used in bdi_update_dirty_ratelimit(): write_bw balanced_dirty_ratelimit *= pos_ratio * ---------- (9) dirty_rate Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-06-12 20:51:31 +04:00
/*
* The base dirty throttle rate, re-calculated on every 200ms.
* All the bdi tasks' dirty rate will be curbed under it.
writeback: stabilize bdi->dirty_ratelimit There are some imperfections in balanced_dirty_ratelimit. 1) large fluctuations The dirty_rate used for computing balanced_dirty_ratelimit is merely averaged in the past 200ms (very small comparing to the 3s estimation period for write_bw), which makes rather dispersed distribution of balanced_dirty_ratelimit. It's pretty hard to average out the singular points by increasing the estimation period. Considering that the averaging technique will introduce very undesirable time lags, I give it up totally. (btw, the 3s write_bw averaging time lag is much more acceptable because its impact is one-way and therefore won't lead to oscillations.) The more practical way is filtering -- most singular balanced_dirty_ratelimit points can be filtered out by remembering some prev_balanced_rate and prev_prev_balanced_rate. However the more reliable way is to guard balanced_dirty_ratelimit with task_ratelimit. 2) due to truncates and fs redirties, the (write_bw <=> dirty_rate) match could become unbalanced, which may lead to large systematical errors in balanced_dirty_ratelimit. The truncates, due to its possibly bumpy nature, can hardly be compensated smoothly. So let's face it. When some over-estimated balanced_dirty_ratelimit brings dirty_ratelimit high, dirty pages will go higher than the setpoint. task_ratelimit will in turn become lower than dirty_ratelimit. So if we consider both balanced_dirty_ratelimit and task_ratelimit and update dirty_ratelimit only when they are on the same side of dirty_ratelimit, the systematical errors in balanced_dirty_ratelimit won't be able to bring dirty_ratelimit far away. The balanced_dirty_ratelimit estimation may also be inaccurate near @limit or @freerun, however is less an issue. 3) since we ultimately want to - keep the fluctuations of task ratelimit as small as possible - keep the dirty pages around the setpoint as long time as possible the update policy used for (2) also serves the above goals nicely: if for some reason the dirty pages are high (task_ratelimit < dirty_ratelimit), and dirty_ratelimit is low (dirty_ratelimit < balanced_dirty_ratelimit), there is no point to bring up dirty_ratelimit in a hurry only to hurt both the above two goals. So, we make use of task_ratelimit to limit the update of dirty_ratelimit in two ways: 1) avoid changing dirty rate when it's against the position control target (the adjusted rate will slow down the progress of dirty pages going back to setpoint). 2) limit the step size. task_ratelimit is changing values step by step, leaving a consistent trace comparing to the randomly jumping balanced_dirty_ratelimit. task_ratelimit also has the nice smaller errors in stable state and typically larger errors when there are big errors in rate. So it's a pretty good limiting factor for the step size of dirty_ratelimit. Note that bdi->dirty_ratelimit is always tracking balanced_dirty_ratelimit. task_ratelimit is merely used as a limiting factor. Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-08-27 01:53:24 +04:00
* @dirty_ratelimit tracks the estimated @balanced_dirty_ratelimit
* in small steps and is much more smooth/stable than the latter.
writeback: dirty rate control It's all about bdi->dirty_ratelimit, which aims to be (write_bw / N) when there are N dd tasks. On write() syscall, use bdi->dirty_ratelimit ============================================ balance_dirty_pages(pages_dirtied) { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); pause = pages_dirtied / task_ratelimit; sleep(pause); } On every 200ms, update bdi->dirty_ratelimit =========================================== bdi_update_dirty_ratelimit() { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate; bdi->dirty_ratelimit = balanced_dirty_ratelimit } Estimation of balanced bdi->dirty_ratelimit =========================================== balanced task_ratelimit ----------------------- balance_dirty_pages() needs to throttle tasks dirtying pages such that the total amount of dirty pages stays below the specified dirty limit in order to avoid memory deadlocks. Furthermore we desire fairness in that tasks get throttled proportionally to the amount of pages they dirty. IOW we want to throttle tasks such that we match the dirty rate to the writeout bandwidth, this yields a stable amount of dirty pages: dirty_rate == write_bw (1) The fairness requirement gives us: task_ratelimit = balanced_dirty_ratelimit == write_bw / N (2) where N is the number of dd tasks. We don't know N beforehand, but still can estimate balanced_dirty_ratelimit within 200ms. Start by throttling each dd task at rate task_ratelimit = task_ratelimit_0 (3) (any non-zero initial value is OK) After 200ms, we measured dirty_rate = # of pages dirtied by all dd's / 200ms write_bw = # of pages written to the disk / 200ms For the aggressive dd dirtiers, the equality holds dirty_rate == N * task_rate == N * task_ratelimit_0 (4) Or task_ratelimit_0 == dirty_rate / N (5) Now we conclude that the balanced task ratelimit can be estimated by write_bw balanced_dirty_ratelimit = task_ratelimit_0 * ---------- (6) dirty_rate Because with (4) and (5) we can get the desired equality (1): write_bw balanced_dirty_ratelimit == (dirty_rate / N) * ---------- dirty_rate == write_bw / N Then using the balanced task ratelimit we can compute task pause times like: task_pause = task->nr_dirtied / task_ratelimit task_ratelimit with position control ------------------------------------ However, while the above gives us means of matching the dirty rate to the writeout bandwidth, it at best provides us with a stable dirty page count (assuming a static system). In order to control the dirty page count such that it is high enough to provide performance, but does not exceed the specified limit we need another control. The dirty position control works by extending (2) to task_ratelimit = balanced_dirty_ratelimit * pos_ratio (7) where pos_ratio is a negative feedback function that subjects to 1) f(setpoint) = 1.0 2) df/dx < 0 That is, if the dirty pages are ABOVE the setpoint, we throttle each task a bit more HEAVY than balanced_dirty_ratelimit, so that the dirty pages are created less fast than they are cleaned, thus DROP to the setpoints (and the reverse). Based on (7) and the assumption that both dirty_ratelimit and pos_ratio remains CONSTANT for the past 200ms, we get task_ratelimit_0 = balanced_dirty_ratelimit * pos_ratio (8) Putting (8) into (6), we get the formula used in bdi_update_dirty_ratelimit(): write_bw balanced_dirty_ratelimit *= pos_ratio * ---------- (9) dirty_rate Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-06-12 20:51:31 +04:00
*/
unsigned long dirty_ratelimit;
writeback: stabilize bdi->dirty_ratelimit There are some imperfections in balanced_dirty_ratelimit. 1) large fluctuations The dirty_rate used for computing balanced_dirty_ratelimit is merely averaged in the past 200ms (very small comparing to the 3s estimation period for write_bw), which makes rather dispersed distribution of balanced_dirty_ratelimit. It's pretty hard to average out the singular points by increasing the estimation period. Considering that the averaging technique will introduce very undesirable time lags, I give it up totally. (btw, the 3s write_bw averaging time lag is much more acceptable because its impact is one-way and therefore won't lead to oscillations.) The more practical way is filtering -- most singular balanced_dirty_ratelimit points can be filtered out by remembering some prev_balanced_rate and prev_prev_balanced_rate. However the more reliable way is to guard balanced_dirty_ratelimit with task_ratelimit. 2) due to truncates and fs redirties, the (write_bw <=> dirty_rate) match could become unbalanced, which may lead to large systematical errors in balanced_dirty_ratelimit. The truncates, due to its possibly bumpy nature, can hardly be compensated smoothly. So let's face it. When some over-estimated balanced_dirty_ratelimit brings dirty_ratelimit high, dirty pages will go higher than the setpoint. task_ratelimit will in turn become lower than dirty_ratelimit. So if we consider both balanced_dirty_ratelimit and task_ratelimit and update dirty_ratelimit only when they are on the same side of dirty_ratelimit, the systematical errors in balanced_dirty_ratelimit won't be able to bring dirty_ratelimit far away. The balanced_dirty_ratelimit estimation may also be inaccurate near @limit or @freerun, however is less an issue. 3) since we ultimately want to - keep the fluctuations of task ratelimit as small as possible - keep the dirty pages around the setpoint as long time as possible the update policy used for (2) also serves the above goals nicely: if for some reason the dirty pages are high (task_ratelimit < dirty_ratelimit), and dirty_ratelimit is low (dirty_ratelimit < balanced_dirty_ratelimit), there is no point to bring up dirty_ratelimit in a hurry only to hurt both the above two goals. So, we make use of task_ratelimit to limit the update of dirty_ratelimit in two ways: 1) avoid changing dirty rate when it's against the position control target (the adjusted rate will slow down the progress of dirty pages going back to setpoint). 2) limit the step size. task_ratelimit is changing values step by step, leaving a consistent trace comparing to the randomly jumping balanced_dirty_ratelimit. task_ratelimit also has the nice smaller errors in stable state and typically larger errors when there are big errors in rate. So it's a pretty good limiting factor for the step size of dirty_ratelimit. Note that bdi->dirty_ratelimit is always tracking balanced_dirty_ratelimit. task_ratelimit is merely used as a limiting factor. Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-08-27 01:53:24 +04:00
unsigned long balanced_dirty_ratelimit;
writeback: dirty rate control It's all about bdi->dirty_ratelimit, which aims to be (write_bw / N) when there are N dd tasks. On write() syscall, use bdi->dirty_ratelimit ============================================ balance_dirty_pages(pages_dirtied) { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); pause = pages_dirtied / task_ratelimit; sleep(pause); } On every 200ms, update bdi->dirty_ratelimit =========================================== bdi_update_dirty_ratelimit() { task_ratelimit = bdi->dirty_ratelimit * bdi_position_ratio(); balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate; bdi->dirty_ratelimit = balanced_dirty_ratelimit } Estimation of balanced bdi->dirty_ratelimit =========================================== balanced task_ratelimit ----------------------- balance_dirty_pages() needs to throttle tasks dirtying pages such that the total amount of dirty pages stays below the specified dirty limit in order to avoid memory deadlocks. Furthermore we desire fairness in that tasks get throttled proportionally to the amount of pages they dirty. IOW we want to throttle tasks such that we match the dirty rate to the writeout bandwidth, this yields a stable amount of dirty pages: dirty_rate == write_bw (1) The fairness requirement gives us: task_ratelimit = balanced_dirty_ratelimit == write_bw / N (2) where N is the number of dd tasks. We don't know N beforehand, but still can estimate balanced_dirty_ratelimit within 200ms. Start by throttling each dd task at rate task_ratelimit = task_ratelimit_0 (3) (any non-zero initial value is OK) After 200ms, we measured dirty_rate = # of pages dirtied by all dd's / 200ms write_bw = # of pages written to the disk / 200ms For the aggressive dd dirtiers, the equality holds dirty_rate == N * task_rate == N * task_ratelimit_0 (4) Or task_ratelimit_0 == dirty_rate / N (5) Now we conclude that the balanced task ratelimit can be estimated by write_bw balanced_dirty_ratelimit = task_ratelimit_0 * ---------- (6) dirty_rate Because with (4) and (5) we can get the desired equality (1): write_bw balanced_dirty_ratelimit == (dirty_rate / N) * ---------- dirty_rate == write_bw / N Then using the balanced task ratelimit we can compute task pause times like: task_pause = task->nr_dirtied / task_ratelimit task_ratelimit with position control ------------------------------------ However, while the above gives us means of matching the dirty rate to the writeout bandwidth, it at best provides us with a stable dirty page count (assuming a static system). In order to control the dirty page count such that it is high enough to provide performance, but does not exceed the specified limit we need another control. The dirty position control works by extending (2) to task_ratelimit = balanced_dirty_ratelimit * pos_ratio (7) where pos_ratio is a negative feedback function that subjects to 1) f(setpoint) = 1.0 2) df/dx < 0 That is, if the dirty pages are ABOVE the setpoint, we throttle each task a bit more HEAVY than balanced_dirty_ratelimit, so that the dirty pages are created less fast than they are cleaned, thus DROP to the setpoints (and the reverse). Based on (7) and the assumption that both dirty_ratelimit and pos_ratio remains CONSTANT for the past 200ms, we get task_ratelimit_0 = balanced_dirty_ratelimit * pos_ratio (8) Putting (8) into (6), we get the formula used in bdi_update_dirty_ratelimit(): write_bw balanced_dirty_ratelimit *= pos_ratio * ---------- (9) dirty_rate Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-06-12 20:51:31 +04:00
struct fprop_local_percpu completions;
mm: per device dirty threshold Scale writeback cache per backing device, proportional to its writeout speed. By decoupling the BDI dirty thresholds a number of problems we currently have will go away, namely: - mutual interference starvation (for any number of BDIs); - deadlocks with stacked BDIs (loop, FUSE and local NFS mounts). It might be that all dirty pages are for a single BDI while other BDIs are idling. By giving each BDI a 'fair' share of the dirty limit, each one can have dirty pages outstanding and make progress. A global threshold also creates a deadlock for stacked BDIs; when A writes to B, and A generates enough dirty pages to get throttled, B will never start writeback until the dirty pages go away. Again, by giving each BDI its own 'independent' dirty limit, this problem is avoided. So the problem is to determine how to distribute the total dirty limit across the BDIs fairly and efficiently. A DBI that has a large dirty limit but does not have any dirty pages outstanding is a waste. What is done is to keep a floating proportion between the DBIs based on writeback completions. This way faster/more active devices get a larger share than slower/idle devices. [akpm@linux-foundation.org: fix warnings] [hugh@veritas.com: Fix occasional hang when a task couldn't get out of balance_dirty_pages] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-17 10:25:50 +04:00
int dirty_exceeded;
unsigned int min_ratio;
unsigned int max_ratio, max_prop_frac;
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
struct bdi_writeback wb; /* default writeback info for this bdi */
spinlock_t wb_lock; /* protects work_list */
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
struct list_head work_list;
struct device *dev;
struct timer_list laptop_mode_wb_timer;
#ifdef CONFIG_DEBUG_FS
struct dentry *debug_dir;
struct dentry *debug_stats;
#endif
};
int bdi_init(struct backing_dev_info *bdi);
void bdi_destroy(struct backing_dev_info *bdi);
__printf(3, 4)
int bdi_register(struct backing_dev_info *bdi, struct device *parent,
const char *fmt, ...);
int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev);
void bdi_unregister(struct backing_dev_info *bdi);
int bdi_setup_and_register(struct backing_dev_info *, char *, unsigned int);
void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
enum wb_reason reason);
void bdi_start_background_writeback(struct backing_dev_info *bdi);
writeback: replace custom worker pool implementation with unbound workqueue Writeback implements its own worker pool - each bdi can be associated with a worker thread which is created and destroyed dynamically. The worker thread for the default bdi is always present and serves as the "forker" thread which forks off worker threads for other bdis. there's no reason for writeback to implement its own worker pool when using unbound workqueue instead is much simpler and more efficient. This patch replaces custom worker pool implementation in writeback with an unbound workqueue. The conversion isn't too complicated but the followings are worth mentioning. * bdi_writeback->last_active, task and wakeup_timer are removed. delayed_work ->dwork is added instead. Explicit timer handling is no longer necessary. Everything works by either queueing / modding / flushing / canceling the delayed_work item. * bdi_writeback_thread() becomes bdi_writeback_workfn() which runs off bdi_writeback->dwork. On each execution, it processes bdi->work_list and reschedules itself if there are more things to do. The function also handles low-mem condition, which used to be handled by the forker thread. If the function is running off a rescuer thread, it only writes out limited number of pages so that the rescuer can serve other bdis too. This preserves the flusher creation failure behavior of the forker thread. * INIT_LIST_HEAD(&bdi->bdi_list) is used to tell bdi_writeback_workfn() about on-going bdi unregistration so that it always drains work_list even if it's running off the rescuer. Note that the original code was broken in this regard. Under memory pressure, a bdi could finish unregistration with non-empty work_list. * The default bdi is no longer special. It now is treated the same as any other bdi and bdi_cap_flush_forker() is removed. * BDI_pending is no longer used. Removed. * Some tracepoints become non-applicable. The following TPs are removed - writeback_nothread, writeback_wake_thread, writeback_wake_forker_thread, writeback_thread_start, writeback_thread_stop. Everything, including devices coming and going away and rescuer operation under simulated memory pressure, seems to work fine in my test setup. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Jens Axboe <axboe@kernel.dk> Cc: Fengguang Wu <fengguang.wu@intel.com> Cc: Jeff Moyer <jmoyer@redhat.com>
2013-04-02 06:08:06 +04:00
void bdi_writeback_workfn(struct work_struct *work);
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
int bdi_has_dirty_io(struct backing_dev_info *bdi);
void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi);
writeback: split inode_wb_list_lock into bdi_writeback.list_lock Split the global inode_wb_list_lock into a per-bdi_writeback list_lock, as it's currently the most contended lock in the system for metadata heavy workloads. It won't help for single-filesystem workloads for which we'll need the I/O-less balance_dirty_pages, but at least we can dedicate a cpu to spinning on each bdi now for larger systems. Based on earlier patches from Nick Piggin and Dave Chinner. It reduces lock contentions to 1/4 in this test case: 10 HDD JBOD, 100 dd on each disk, XFS, 6GB ram lock_stat version 0.3 ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- class name con-bounces contentions waittime-min waittime-max waittime-total acq-bounces acquisitions holdtime-min holdtime-max holdtime-total ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- vanilla 2.6.39-rc3: inode_wb_list_lock: 42590 44433 0.12 147.74 144127.35 252274 886792 0.08 121.34 917211.23 ------------------ inode_wb_list_lock 2 [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85 inode_wb_list_lock 34 [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49 inode_wb_list_lock 12893 [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0 inode_wb_list_lock 10702 [<ffffffff8115afef>] writeback_single_inode+0x16d/0x20a ------------------ inode_wb_list_lock 2 [<ffffffff81165da5>] bdev_inode_switch_bdi+0x29/0x85 inode_wb_list_lock 19 [<ffffffff8115bd0b>] inode_wb_list_del+0x22/0x49 inode_wb_list_lock 5550 [<ffffffff8115bb53>] __mark_inode_dirty+0x170/0x1d0 inode_wb_list_lock 8511 [<ffffffff8115b4ad>] writeback_sb_inodes+0x10f/0x157 2.6.39-rc3 + patch: &(&wb->list_lock)->rlock: 11383 11657 0.14 151.69 40429.51 90825 527918 0.11 145.90 556843.37 ------------------------ &(&wb->list_lock)->rlock 10 [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86 &(&wb->list_lock)->rlock 1493 [<ffffffff8115b1ed>] writeback_inodes_wb+0x3d/0x150 &(&wb->list_lock)->rlock 3652 [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f &(&wb->list_lock)->rlock 1412 [<ffffffff8115a38e>] writeback_single_inode+0x17f/0x223 ------------------------ &(&wb->list_lock)->rlock 3 [<ffffffff8110b5af>] bdi_lock_two+0x46/0x4b &(&wb->list_lock)->rlock 6 [<ffffffff8115b189>] inode_wb_list_del+0x5f/0x86 &(&wb->list_lock)->rlock 2061 [<ffffffff8115af97>] __mark_inode_dirty+0x173/0x1cf &(&wb->list_lock)->rlock 2629 [<ffffffff8115a8e9>] writeback_sb_inodes+0x123/0x16f hughd@google.com: fix recursive lock when bdi_lock_two() is called with new the same as old akpm@linux-foundation.org: cleanup bdev_inode_switch_bdi() comment Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-04-22 04:19:44 +04:00
void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2);
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
extern spinlock_t bdi_lock;
extern struct list_head bdi_list;
writeback: replace custom worker pool implementation with unbound workqueue Writeback implements its own worker pool - each bdi can be associated with a worker thread which is created and destroyed dynamically. The worker thread for the default bdi is always present and serves as the "forker" thread which forks off worker threads for other bdis. there's no reason for writeback to implement its own worker pool when using unbound workqueue instead is much simpler and more efficient. This patch replaces custom worker pool implementation in writeback with an unbound workqueue. The conversion isn't too complicated but the followings are worth mentioning. * bdi_writeback->last_active, task and wakeup_timer are removed. delayed_work ->dwork is added instead. Explicit timer handling is no longer necessary. Everything works by either queueing / modding / flushing / canceling the delayed_work item. * bdi_writeback_thread() becomes bdi_writeback_workfn() which runs off bdi_writeback->dwork. On each execution, it processes bdi->work_list and reschedules itself if there are more things to do. The function also handles low-mem condition, which used to be handled by the forker thread. If the function is running off a rescuer thread, it only writes out limited number of pages so that the rescuer can serve other bdis too. This preserves the flusher creation failure behavior of the forker thread. * INIT_LIST_HEAD(&bdi->bdi_list) is used to tell bdi_writeback_workfn() about on-going bdi unregistration so that it always drains work_list even if it's running off the rescuer. Note that the original code was broken in this regard. Under memory pressure, a bdi could finish unregistration with non-empty work_list. * The default bdi is no longer special. It now is treated the same as any other bdi and bdi_cap_flush_forker() is removed. * BDI_pending is no longer used. Removed. * Some tracepoints become non-applicable. The following TPs are removed - writeback_nothread, writeback_wake_thread, writeback_wake_forker_thread, writeback_thread_start, writeback_thread_stop. Everything, including devices coming and going away and rescuer operation under simulated memory pressure, seems to work fine in my test setup. Signed-off-by: Tejun Heo <tj@kernel.org> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Jens Axboe <axboe@kernel.dk> Cc: Fengguang Wu <fengguang.wu@intel.com> Cc: Jeff Moyer <jmoyer@redhat.com>
2013-04-02 06:08:06 +04:00
extern struct workqueue_struct *bdi_wq;
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
static inline int wb_has_dirty_io(struct bdi_writeback *wb)
{
return !list_empty(&wb->b_dirty) ||
!list_empty(&wb->b_io) ||
!list_empty(&wb->b_more_io);
}
static inline void __add_bdi_stat(struct backing_dev_info *bdi,
enum bdi_stat_item item, s64 amount)
{
__percpu_counter_add(&bdi->bdi_stat[item], amount, BDI_STAT_BATCH);
}
static inline void __inc_bdi_stat(struct backing_dev_info *bdi,
enum bdi_stat_item item)
{
__add_bdi_stat(bdi, item, 1);
}
static inline void inc_bdi_stat(struct backing_dev_info *bdi,
enum bdi_stat_item item)
{
unsigned long flags;
local_irq_save(flags);
__inc_bdi_stat(bdi, item);
local_irq_restore(flags);
}
static inline void __dec_bdi_stat(struct backing_dev_info *bdi,
enum bdi_stat_item item)
{
__add_bdi_stat(bdi, item, -1);
}
static inline void dec_bdi_stat(struct backing_dev_info *bdi,
enum bdi_stat_item item)
{
unsigned long flags;
local_irq_save(flags);
__dec_bdi_stat(bdi, item);
local_irq_restore(flags);
}
static inline s64 bdi_stat(struct backing_dev_info *bdi,
enum bdi_stat_item item)
{
return percpu_counter_read_positive(&bdi->bdi_stat[item]);
}
static inline s64 __bdi_stat_sum(struct backing_dev_info *bdi,
enum bdi_stat_item item)
{
return percpu_counter_sum_positive(&bdi->bdi_stat[item]);
}
static inline s64 bdi_stat_sum(struct backing_dev_info *bdi,
enum bdi_stat_item item)
{
s64 sum;
unsigned long flags;
local_irq_save(flags);
sum = __bdi_stat_sum(bdi, item);
local_irq_restore(flags);
return sum;
}
extern void bdi_writeout_inc(struct backing_dev_info *bdi);
/*
* maximal error of a stat counter.
*/
static inline unsigned long bdi_stat_error(struct backing_dev_info *bdi)
{
#ifdef CONFIG_SMP
return nr_cpu_ids * BDI_STAT_BATCH;
#else
return 1;
#endif
}
int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio);
int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned int max_ratio);
/*
* Flags in backing_dev_info::capability
*
* The first three flags control whether dirty pages will contribute to the
* VM's accounting and whether writepages() should be called for dirty pages
* (something that would not, for example, be appropriate for ramfs)
*
* WARNING: these flags are closely related and should not normally be
* used separately. The BDI_CAP_NO_ACCT_AND_WRITEBACK combines these
* three flags into a single convenience macro.
*
* BDI_CAP_NO_ACCT_DIRTY: Dirty pages shouldn't contribute to accounting
* BDI_CAP_NO_WRITEBACK: Don't write pages back
* BDI_CAP_NO_ACCT_WB: Don't automatically account writeback pages
*
* These flags let !MMU mmap() govern direct device mapping vs immediate
* copying more easily for MAP_PRIVATE, especially for ROM filesystems.
*
* BDI_CAP_MAP_COPY: Copy can be mapped (MAP_PRIVATE)
* BDI_CAP_MAP_DIRECT: Can be mapped directly (MAP_SHARED)
* BDI_CAP_READ_MAP: Can be mapped for reading
* BDI_CAP_WRITE_MAP: Can be mapped for writing
* BDI_CAP_EXEC_MAP: Can be mapped for execution
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 07:26:32 +04:00
*
* BDI_CAP_SWAP_BACKED: Count shmem/tmpfs objects as swap-backed.
*/
#define BDI_CAP_NO_ACCT_DIRTY 0x00000001
#define BDI_CAP_NO_WRITEBACK 0x00000002
#define BDI_CAP_MAP_COPY 0x00000004
#define BDI_CAP_MAP_DIRECT 0x00000008
#define BDI_CAP_READ_MAP 0x00000010
#define BDI_CAP_WRITE_MAP 0x00000020
#define BDI_CAP_EXEC_MAP 0x00000040
#define BDI_CAP_NO_ACCT_WB 0x00000080
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 07:26:32 +04:00
#define BDI_CAP_SWAP_BACKED 0x00000100
bdi: allow block devices to say that they require stable page writes This patchset ("stable page writes, part 2") makes some key modifications to the original 'stable page writes' patchset. First, it provides creators (devices and filesystems) of a backing_dev_info a flag that declares whether or not it is necessary to ensure that page contents cannot change during writeout. It is no longer assumed that this is true of all devices (which was never true anyway). Second, the flag is used to relaxed the wait_on_page_writeback calls so that wait only occurs if the device needs it. Third, it fixes up the remaining disk-backed filesystems to use this improved conditional-wait logic to provide stable page writes on those filesystems. It is hoped that (for people not using checksumming devices, anyway) this patchset will give back unnecessary performance decreases since the original stable page write patchset went into 3.0. Sorry about not fixing it sooner. Complaints were registered by several people about the long write latencies introduced by the original stable page write patchset. Generally speaking, the kernel ought to allocate as little extra memory as possible to facilitate writeout, but for people who simply cannot wait, a second page stability strategy is (re)introduced: snapshotting page contents. The waiting behavior is still the default strategy; to enable page snapshotting, a superblock flag (MS_SNAP_STABLE) must be set. This flag is used to bandaid^Henable stable page writeback on ext3[1], and is not used anywhere else. Given that there are already a few storage devices and network FSes that have rolled their own page stability wait/page snapshot code, it would be nice to move towards consolidating all of these. It seems possible that iscsi and raid5 may wish to use the new stable page write support to enable zero-copy writeout. Thank you to Jan Kara for helping fix a couple more filesystems. Per Andrew Morton's request, here are the result of using dbench to measure latencies on ext2: 3.8.0-rc3: Operation Count AvgLat MaxLat ---------------------------------------- WriteX 109347 0.028 59.817 ReadX 347180 0.004 3.391 Flush 15514 29.828 287.283 Throughput 57.429 MB/sec 4 clients 4 procs max_latency=287.290 ms 3.8.0-rc3 + patches: WriteX 105556 0.029 4.273 ReadX 335004 0.005 4.112 Flush 14982 30.540 298.634 Throughput 55.4496 MB/sec 4 clients 4 procs max_latency=298.650 ms As you can see, for ext2 the maximum write latency decreases from ~60ms on a laptop hard disk to ~4ms. I'm not sure why the flush latencies increase, though I suspect that being able to dirty pages faster gives the flusher more work to do. On ext4, the average write latency decreases as well as all the maximum latencies: 3.8.0-rc3: WriteX 85624 0.152 33.078 ReadX 272090 0.010 61.210 Flush 12129 36.219 168.260 Throughput 44.8618 MB/sec 4 clients 4 procs max_latency=168.276 ms 3.8.0-rc3 + patches: WriteX 86082 0.141 30.928 ReadX 273358 0.010 36.124 Flush 12214 34.800 165.689 Throughput 44.9941 MB/sec 4 clients 4 procs max_latency=165.722 ms XFS seems to exhibit similar latency improvements as ext2: 3.8.0-rc3: WriteX 125739 0.028 104.343 ReadX 399070 0.005 4.115 Flush 17851 25.004 131.390 Throughput 66.0024 MB/sec 4 clients 4 procs max_latency=131.406 ms 3.8.0-rc3 + patches: WriteX 123529 0.028 6.299 ReadX 392434 0.005 4.287 Flush 17549 25.120 188.687 Throughput 64.9113 MB/sec 4 clients 4 procs max_latency=188.704 ms ...and btrfs, just to round things out, also shows some latency decreases: 3.8.0-rc3: WriteX 67122 0.083 82.355 ReadX 212719 0.005 2.828 Flush 9547 47.561 147.418 Throughput 35.3391 MB/sec 4 clients 4 procs max_latency=147.433 ms 3.8.0-rc3 + patches: WriteX 64898 0.101 71.631 ReadX 206673 0.005 7.123 Flush 9190 47.963 219.034 Throughput 34.0795 MB/sec 4 clients 4 procs max_latency=219.044 ms Before this patchset, all filesystems would block, regardless of whether or not it was necessary. ext3 would wait, but still generate occasional checksum errors. The network filesystems were left to do their own thing, so they'd wait too. After this patchset, all the disk filesystems except ext3 and btrfs will wait only if the hardware requires it. ext3 (if necessary) snapshots pages instead of blocking, and btrfs provides its own bdi so the mm will never wait. Network filesystems haven't been touched, so either they provide their own wait code, or they don't block at all. The blocking behavior is back to what it was before 3.0 if you don't have a disk requiring stable page writes. This patchset has been tested on 3.8.0-rc3 on x64 with ext3, ext4, and xfs. I've spot-checked 3.8.0-rc4 and seem to be getting the same results as -rc3. [1] The alternative fixes to ext3 include fixing the locking order and page bit handling like we did for ext4 (but then why not just use ext4?), or setting PG_writeback so early that ext3 becomes extremely slow. I tried that, but the number of write()s I could initiate dropped by nearly an order of magnitude. That was a bit much even for the author of the stable page series! :) This patch: Creates a per-backing-device flag that tracks whether or not pages must be held immutable during writeout. Eventually it will be used to waive wait_for_page_writeback() if nothing requires stable pages. Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Artem Bityutskiy <dedekind1@gmail.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Eric Van Hensbergen <ericvh@gmail.com> Cc: Ron Minnich <rminnich@sandia.gov> Cc: Latchesar Ionkov <lucho@ionkov.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-22 04:42:48 +04:00
#define BDI_CAP_STABLE_WRITES 0x00000200
#define BDI_CAP_VMFLAGS \
(BDI_CAP_READ_MAP | BDI_CAP_WRITE_MAP | BDI_CAP_EXEC_MAP)
#define BDI_CAP_NO_ACCT_AND_WRITEBACK \
(BDI_CAP_NO_WRITEBACK | BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_ACCT_WB)
#if defined(VM_MAYREAD) && \
(BDI_CAP_READ_MAP != VM_MAYREAD || \
BDI_CAP_WRITE_MAP != VM_MAYWRITE || \
BDI_CAP_EXEC_MAP != VM_MAYEXEC)
#error please change backing_dev_info::capabilities flags
#endif
extern struct backing_dev_info default_backing_dev_info;
extern struct backing_dev_info noop_backing_dev_info;
int writeback_in_progress(struct backing_dev_info *bdi);
static inline int bdi_congested(struct backing_dev_info *bdi, int bdi_bits)
{
if (bdi->congested_fn)
return bdi->congested_fn(bdi->congested_data, bdi_bits);
return (bdi->state & bdi_bits);
}
static inline int bdi_read_congested(struct backing_dev_info *bdi)
{
return bdi_congested(bdi, 1 << BDI_sync_congested);
}
static inline int bdi_write_congested(struct backing_dev_info *bdi)
{
return bdi_congested(bdi, 1 << BDI_async_congested);
}
static inline int bdi_rw_congested(struct backing_dev_info *bdi)
{
return bdi_congested(bdi, (1 << BDI_sync_congested) |
(1 << BDI_async_congested));
}
enum {
BLK_RW_ASYNC = 0,
BLK_RW_SYNC = 1,
};
void clear_bdi_congested(struct backing_dev_info *bdi, int sync);
void set_bdi_congested(struct backing_dev_info *bdi, int sync);
long congestion_wait(int sync, long timeout);
long wait_iff_congested(struct zone *zone, int sync, long timeout);
int pdflush_proc_obsolete(struct ctl_table *table, int write,
void __user *buffer, size_t *lenp, loff_t *ppos);
bdi: allow block devices to say that they require stable page writes This patchset ("stable page writes, part 2") makes some key modifications to the original 'stable page writes' patchset. First, it provides creators (devices and filesystems) of a backing_dev_info a flag that declares whether or not it is necessary to ensure that page contents cannot change during writeout. It is no longer assumed that this is true of all devices (which was never true anyway). Second, the flag is used to relaxed the wait_on_page_writeback calls so that wait only occurs if the device needs it. Third, it fixes up the remaining disk-backed filesystems to use this improved conditional-wait logic to provide stable page writes on those filesystems. It is hoped that (for people not using checksumming devices, anyway) this patchset will give back unnecessary performance decreases since the original stable page write patchset went into 3.0. Sorry about not fixing it sooner. Complaints were registered by several people about the long write latencies introduced by the original stable page write patchset. Generally speaking, the kernel ought to allocate as little extra memory as possible to facilitate writeout, but for people who simply cannot wait, a second page stability strategy is (re)introduced: snapshotting page contents. The waiting behavior is still the default strategy; to enable page snapshotting, a superblock flag (MS_SNAP_STABLE) must be set. This flag is used to bandaid^Henable stable page writeback on ext3[1], and is not used anywhere else. Given that there are already a few storage devices and network FSes that have rolled their own page stability wait/page snapshot code, it would be nice to move towards consolidating all of these. It seems possible that iscsi and raid5 may wish to use the new stable page write support to enable zero-copy writeout. Thank you to Jan Kara for helping fix a couple more filesystems. Per Andrew Morton's request, here are the result of using dbench to measure latencies on ext2: 3.8.0-rc3: Operation Count AvgLat MaxLat ---------------------------------------- WriteX 109347 0.028 59.817 ReadX 347180 0.004 3.391 Flush 15514 29.828 287.283 Throughput 57.429 MB/sec 4 clients 4 procs max_latency=287.290 ms 3.8.0-rc3 + patches: WriteX 105556 0.029 4.273 ReadX 335004 0.005 4.112 Flush 14982 30.540 298.634 Throughput 55.4496 MB/sec 4 clients 4 procs max_latency=298.650 ms As you can see, for ext2 the maximum write latency decreases from ~60ms on a laptop hard disk to ~4ms. I'm not sure why the flush latencies increase, though I suspect that being able to dirty pages faster gives the flusher more work to do. On ext4, the average write latency decreases as well as all the maximum latencies: 3.8.0-rc3: WriteX 85624 0.152 33.078 ReadX 272090 0.010 61.210 Flush 12129 36.219 168.260 Throughput 44.8618 MB/sec 4 clients 4 procs max_latency=168.276 ms 3.8.0-rc3 + patches: WriteX 86082 0.141 30.928 ReadX 273358 0.010 36.124 Flush 12214 34.800 165.689 Throughput 44.9941 MB/sec 4 clients 4 procs max_latency=165.722 ms XFS seems to exhibit similar latency improvements as ext2: 3.8.0-rc3: WriteX 125739 0.028 104.343 ReadX 399070 0.005 4.115 Flush 17851 25.004 131.390 Throughput 66.0024 MB/sec 4 clients 4 procs max_latency=131.406 ms 3.8.0-rc3 + patches: WriteX 123529 0.028 6.299 ReadX 392434 0.005 4.287 Flush 17549 25.120 188.687 Throughput 64.9113 MB/sec 4 clients 4 procs max_latency=188.704 ms ...and btrfs, just to round things out, also shows some latency decreases: 3.8.0-rc3: WriteX 67122 0.083 82.355 ReadX 212719 0.005 2.828 Flush 9547 47.561 147.418 Throughput 35.3391 MB/sec 4 clients 4 procs max_latency=147.433 ms 3.8.0-rc3 + patches: WriteX 64898 0.101 71.631 ReadX 206673 0.005 7.123 Flush 9190 47.963 219.034 Throughput 34.0795 MB/sec 4 clients 4 procs max_latency=219.044 ms Before this patchset, all filesystems would block, regardless of whether or not it was necessary. ext3 would wait, but still generate occasional checksum errors. The network filesystems were left to do their own thing, so they'd wait too. After this patchset, all the disk filesystems except ext3 and btrfs will wait only if the hardware requires it. ext3 (if necessary) snapshots pages instead of blocking, and btrfs provides its own bdi so the mm will never wait. Network filesystems haven't been touched, so either they provide their own wait code, or they don't block at all. The blocking behavior is back to what it was before 3.0 if you don't have a disk requiring stable page writes. This patchset has been tested on 3.8.0-rc3 on x64 with ext3, ext4, and xfs. I've spot-checked 3.8.0-rc4 and seem to be getting the same results as -rc3. [1] The alternative fixes to ext3 include fixing the locking order and page bit handling like we did for ext4 (but then why not just use ext4?), or setting PG_writeback so early that ext3 becomes extremely slow. I tried that, but the number of write()s I could initiate dropped by nearly an order of magnitude. That was a bit much even for the author of the stable page series! :) This patch: Creates a per-backing-device flag that tracks whether or not pages must be held immutable during writeout. Eventually it will be used to waive wait_for_page_writeback() if nothing requires stable pages. Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Adrian Hunter <adrian.hunter@intel.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Artem Bityutskiy <dedekind1@gmail.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Eric Van Hensbergen <ericvh@gmail.com> Cc: Ron Minnich <rminnich@sandia.gov> Cc: Latchesar Ionkov <lucho@ionkov.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-22 04:42:48 +04:00
static inline bool bdi_cap_stable_pages_required(struct backing_dev_info *bdi)
{
return bdi->capabilities & BDI_CAP_STABLE_WRITES;
}
static inline bool bdi_cap_writeback_dirty(struct backing_dev_info *bdi)
{
return !(bdi->capabilities & BDI_CAP_NO_WRITEBACK);
}
static inline bool bdi_cap_account_dirty(struct backing_dev_info *bdi)
{
return !(bdi->capabilities & BDI_CAP_NO_ACCT_DIRTY);
}
static inline bool bdi_cap_account_writeback(struct backing_dev_info *bdi)
{
/* Paranoia: BDI_CAP_NO_WRITEBACK implies BDI_CAP_NO_ACCT_WB */
return !(bdi->capabilities & (BDI_CAP_NO_ACCT_WB |
BDI_CAP_NO_WRITEBACK));
}
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 07:26:32 +04:00
static inline bool bdi_cap_swap_backed(struct backing_dev_info *bdi)
{
return bdi->capabilities & BDI_CAP_SWAP_BACKED;
}
static inline bool mapping_cap_writeback_dirty(struct address_space *mapping)
{
return bdi_cap_writeback_dirty(mapping->backing_dev_info);
}
static inline bool mapping_cap_account_dirty(struct address_space *mapping)
{
return bdi_cap_account_dirty(mapping->backing_dev_info);
}
vmscan: split LRU lists into anon & file sets Split the LRU lists in two, one set for pages that are backed by real file systems ("file") and one for pages that are backed by memory and swap ("anon"). The latter includes tmpfs. The advantage of doing this is that the VM will not have to scan over lots of anonymous pages (which we generally do not want to swap out), just to find the page cache pages that it should evict. This patch has the infrastructure and a basic policy to balance how much we scan the anon lists and how much we scan the file lists. The big policy changes are in separate patches. [lee.schermerhorn@hp.com: collect lru meminfo statistics from correct offset] [kosaki.motohiro@jp.fujitsu.com: prevent incorrect oom under split_lru] [kosaki.motohiro@jp.fujitsu.com: fix pagevec_move_tail() doesn't treat unevictable page] [hugh@veritas.com: memcg swapbacked pages active] [hugh@veritas.com: splitlru: BDI_CAP_SWAP_BACKED] [akpm@linux-foundation.org: fix /proc/vmstat units] [nishimura@mxp.nes.nec.co.jp: memcg: fix handling of shmem migration] [kosaki.motohiro@jp.fujitsu.com: adjust Quicklists field of /proc/meminfo] [kosaki.motohiro@jp.fujitsu.com: fix style issue of get_scan_ratio()] Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-10-19 07:26:32 +04:00
static inline bool mapping_cap_swap_backed(struct address_space *mapping)
{
return bdi_cap_swap_backed(mapping->backing_dev_info);
}
writeback: switch to per-bdi threads for flushing data This gets rid of pdflush for bdi writeout and kupdated style cleaning. pdflush writeout suffers from lack of locality and also requires more threads to handle the same workload, since it has to work in a non-blocking fashion against each queue. This also introduces lumpy behaviour and potential request starvation, since pdflush can be starved for queue access if others are accessing it. A sample ffsb workload that does random writes to files is about 8% faster here on a simple SATA drive during the benchmark phase. File layout also seems a LOT more smooth in vmstat: r b swpd free buff cache si so bi bo in cs us sy id wa 0 1 0 608848 2652 375372 0 0 0 71024 604 24 1 10 48 42 0 1 0 549644 2712 433736 0 0 0 60692 505 27 1 8 48 44 1 0 0 476928 2784 505192 0 0 4 29540 553 24 0 9 53 37 0 1 0 457972 2808 524008 0 0 0 54876 331 16 0 4 38 58 0 1 0 366128 2928 614284 0 0 4 92168 710 58 0 13 53 34 0 1 0 295092 3000 684140 0 0 0 62924 572 23 0 9 53 37 0 1 0 236592 3064 741704 0 0 4 58256 523 17 0 8 48 44 0 1 0 165608 3132 811464 0 0 0 57460 560 21 0 8 54 38 0 1 0 102952 3200 873164 0 0 4 74748 540 29 1 10 48 41 0 1 0 48604 3252 926472 0 0 0 53248 469 29 0 7 47 45 where vanilla tends to fluctuate a lot in the creation phase: r b swpd free buff cache si so bi bo in cs us sy id wa 1 1 0 678716 5792 303380 0 0 0 74064 565 50 1 11 52 36 1 0 0 662488 5864 319396 0 0 4 352 302 329 0 2 47 51 0 1 0 599312 5924 381468 0 0 0 78164 516 55 0 9 51 40 0 1 0 519952 6008 459516 0 0 4 78156 622 56 1 11 52 37 1 1 0 436640 6092 541632 0 0 0 82244 622 54 0 11 48 41 0 1 0 436640 6092 541660 0 0 0 8 152 39 0 0 51 49 0 1 0 332224 6200 644252 0 0 4 102800 728 46 1 13 49 36 1 0 0 274492 6260 701056 0 0 4 12328 459 49 0 7 50 43 0 1 0 211220 6324 763356 0 0 0 106940 515 37 1 10 51 39 1 0 0 160412 6376 813468 0 0 0 8224 415 43 0 6 49 45 1 1 0 85980 6452 886556 0 0 4 113516 575 39 1 11 54 34 0 2 0 85968 6452 886620 0 0 0 1640 158 211 0 0 46 54 A 10 disk test with btrfs performs 26% faster with per-bdi flushing. A SSD based writeback test on XFS performs over 20% better as well, with the throughput being very stable around 1GB/sec, where pdflush only manages 750MB/sec and fluctuates wildly while doing so. Random buffered writes to many files behave a lot better as well, as does random mmap'ed writes. A separate thread is added to sync the super blocks. In the long term, adding sync_supers_bdi() functionality could get rid of this thread again. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2009-09-09 11:08:54 +04:00
static inline int bdi_sched_wait(void *word)
{
schedule();
return 0;
}
#endif /* _LINUX_BACKING_DEV_H */