Some trace shows lots of bdi_dirty=0 lines where it's actually some
small value if w/o the accounting errors in the per-cpu bdi stats.

In this case the max pause time should really be set to the smallest
(non-zero) value to avoid IO queue underrun and improve throughput.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

On a system with 1 local mount and 1 NFS mount, if the NFS server
becomes not responding when dd to the NFS mount, the NFS dirty pages may
exceed the global dirty limit and _every_ task involving writing will be
blocked. The whole system appears unresponsive.

The workaround is to permit through the bdi's that only has a small
number of dirty pages. The number chosen (bdi_stat_error pages) is not
enough to enable the local disk to run in optimal throughput, however is
enough to make the system responsive on a broken NFS mount. The user can
then kill the dirtiers on the NFS mount and increase the global dirty
limit to bring up the local disk's throughput.

It risks allowing dirty pages to grow much larger than the global dirty
limit when there are 1000+ mounts, however that's very unlikely to happen,
especially in low memory profiles.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

We do "floating proportions" to let active devices to grow its target
share of dirty pages and stalled/inactive devices to decrease its target
share over time.

It works well except in the case of "an inactive disk suddenly goes
busy", where the initial target share may be too small. To mitigate
this, bdi_position_ratio() has the below line to raise a small
bdi_thresh when it's safe to do so, so that the disk be feed with enough
dirty pages for efficient IO and in turn fast rampup of bdi_thresh:

        bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);

balance_dirty_pages() normally does negative feedback control which
adjusts ratelimit to balance the bdi dirty pages around the target.
In some extreme cases when that is not enough, it will have to block
the tasks completely until the bdi dirty pages drop below bdi_thresh.
Acked-by: NJan Kara <jack@suse.cz>
Acked-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

They are not used any more.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

The sleep based balance_dirty_pages() can pause at most MAX_PAUSE=200ms
on every 1 4KB-page, which means it cannot throttle a task under
4KB/200ms=20KB/s. So when there are more than 512 dd writing to a
10MB/s USB stick, its bdi dirty pages could grow out of control.

Even if we can increase MAX_PAUSE, the minimal (task_ratelimit = 1)
means a limit of 4KB/s.
                                                       
They can eventually be safeguarded by the global limit check 
(nr_dirty < dirty_thresh). However if someone is also writing to an 
HDD at the same time, it'll get poor HDD write performance.
                                                       
We at least want to maintain good write performance for other devices
when one device is attacked by some "massive parallel" workload, or
suffers from slow write bandwidth, or somehow get stalled due to some 
error condition (eg. NFS server not responding).

For a stalled device, we need to completely block its dirtiers, too,
before its bdi dirty pages grow all the way up to the global limit and
leave no space for the other functional devices.

So change the loop exit condition to

	/*
	 * Always enforce global dirty limit; also enforce bdi dirty limit
	 * if the normal max_pause sleeps cannot keep things under control.
	 */
	if (nr_dirty < dirty_thresh &&
	    (bdi_dirty < bdi_thresh || bdi->dirty_ratelimit > 1))
		break;

which can be further simplified to

	if (task_ratelimit)
		break;
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

There is no reason why task in balance_dirty_pages() shouldn't be killable
and it helps in recovering from some error conditions (like when filesystem
goes in error state and cannot accept writeback anymore but we still want to
kill processes using it to be able to unmount it).

There will be follow up patches to further abort the generic_perform_write()
and other filesystem write loops, to avoid large write + SIGKILL combination
exceeding the dirty limit and possibly strange OOM.
Reported-by: NKazuya Mio <k-mio@sx.jp.nec.com>
Tested-by: NKazuya Mio <k-mio@sx.jp.nec.com>
Reviewed-by: NNeil Brown <neilb@suse.de>
Reviewed-by: NKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

In balance_dirty_pages() task_ratelimit may be not initialized
(initialization skiped by goto pause), and then used when calling
tracing hook.

Fix it by moving the task_ratelimit assignment before goto pause.
Reported-by: NWitold Baryluk <baryluk@smp.if.uj.edu.pl>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Looks like someone got distracted after adding the comment characters.
Signed-off-by: NJohannes Weiner <jweiner@redhat.com>
Acked-by: NPeter Zijlstra <peterz@infradead.org>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The files changed within are only using the EXPORT_SYMBOL
macro variants.  They are not using core modular infrastructure
and hence don't need module.h but only the export.h header.
Signed-off-by: NPaul Gortmaker <paul.gortmaker@windriver.com>

This creates a new 'reason' field in a wb_writeback_work
structure, which unambiguously identifies who initiates
writeback activity.  A 'wb_reason' enumeration has been
added to writeback.h, to enumerate the possible reasons.

The 'writeback_work_class' and tracepoint event class and
'writeback_queue_io' tracepoints are updated to include the
symbolic 'reason' in all trace events.

And the 'writeback_inodes_sbXXX' family of routines has had
a wb_stats parameter added to them, so callers can specify
why writeback is being started.
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NCurt Wohlgemuth <curtw@google.com>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Useful for analyzing the dynamics of the throttling algorithms and
debugging user reported problems.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

It helps understand how various throttle bandwidths are updated.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Fix powerpc compile warnings

mm/page-writeback.c: In function 'bdi_position_ratio':
mm/page-writeback.c:622:3: warning: comparison of distinct pointer types lacks a cast [enabled by default]
page-writeback.c:635:4: warning: comparison of distinct pointer types lacks a cast [enabled by default]

Also fix gcc "uninitialized var" warnings.
Reported-by: NStephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Keep a minimal pool of dirty pages for each bdi, so that the disk IO
queues won't underrun. Also gently increase a small bdi_thresh to avoid
it stuck in 0 for some light dirtied bdi.

It's particularly useful for JBOD and small memory system.

It may result in (pos_ratio > 1) at the setpoint and push the dirty
pages high. This is more or less intended because the bdi is in the
danger of IO queue underflow.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

The dirty pause time shall ultimately be controlled by adjusting
nr_dirtied_pause, since there is relationship

	pause = pages_dirtied / task_ratelimit

Assuming

	pages_dirtied ~= nr_dirtied_pause
	task_ratelimit ~= dirty_ratelimit

We get

	nr_dirtied_pause ~= dirty_ratelimit * desired_pause

Here dirty_ratelimit is preferred over task_ratelimit because it's
more stable.

It's also important to limit possible large transitional errors:

- bw is changing quickly
- pages_dirtied << nr_dirtied_pause on entering dirty exceeded area
- pages_dirtied >> nr_dirtied_pause on btrfs (to be improved by a
  separate fix, but still expect non-trivial errors)

So we end up using the above formula inside clamp_val().

The best test case for this code is to run 100 "dd bs=4M" tasks on
btrfs and check its pause time distribution.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Apply two policies to scale down the max pause time for

1) small number of concurrent dirtiers
2) small memory system (comparing to storage bandwidth)

MAX_PAUSE=200ms may only be suitable for high end servers with lots of
concurrent dirtiers, where the large pause time can reduce much overheads.

Otherwise, smaller pause time is desirable whenever possible, so as to
get good responsiveness and smooth user experiences. It's actually
required for good disk utilization in the case when all the dirty pages
can be synced to disk within MAX_PAUSE=200ms.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

As proposed by Chris, Dave and Jan, don't start foreground writeback IO
inside balance_dirty_pages(). Instead, simply let it idle sleep for some
time to throttle the dirtying task. In the mean while, kick off the
per-bdi flusher thread to do background writeback IO.

RATIONALS
=========

- disk seeks on concurrent writeback of multiple inodes (Dave Chinner)

  If every thread doing writes and being throttled start foreground
  writeback, it leads to N IO submitters from at least N different
  inodes at the same time, end up with N different sets of IO being
  issued with potentially zero locality to each other, resulting in
  much lower elevator sort/merge efficiency and hence we seek the disk
  all over the place to service the different sets of IO.
  OTOH, if there is only one submission thread, it doesn't jump between
  inodes in the same way when congestion clears - it keeps writing to
  the same inode, resulting in large related chunks of sequential IOs
  being issued to the disk. This is more efficient than the above
  foreground writeback because the elevator works better and the disk
  seeks less.

- lock contention and cache bouncing on concurrent IO submitters (Dave Chinner)

  With this patchset, the fs_mark benchmark on a 12-drive software RAID0 goes
  from CPU bound to IO bound, freeing "3-4 CPUs worth of spinlock contention".

  * "CPU usage has dropped by ~55%", "it certainly appears that most of
    the CPU time saving comes from the removal of contention on the
    inode_wb_list_lock" (IMHO at least 10% comes from the reduction of
    cacheline bouncing, because the new code is able to call much less
    frequently into balance_dirty_pages() and hence access the global
    page states)

  * the user space "App overhead" is reduced by 20%, by avoiding the
    cacheline pollution by the complex writeback code path

  * "for a ~5% throughput reduction", "the number of write IOs have
    dropped by ~25%", and the elapsed time reduced from 41:42.17 to
    40:53.23.

  * On a simple test of 100 dd, it reduces the CPU %system time from 30% to 3%,
    and improves IO throughput from 38MB/s to 42MB/s.

- IO size too small for fast arrays and too large for slow USB sticks

  The write_chunk used by current balance_dirty_pages() cannot be
  directly set to some large value (eg. 128MB) for better IO efficiency.
  Because it could lead to more than 1 second user perceivable stalls.
  Even the current 4MB write size may be too large for slow USB sticks.
  The fact that balance_dirty_pages() starts IO on itself couples the
  IO size to wait time, which makes it hard to do suitable IO size while
  keeping the wait time under control.

  Now it's possible to increase writeback chunk size proportional to the
  disk bandwidth. In a simple test of 50 dd's on XFS, 1-HDD, 3GB ram,
  the larger writeback size dramatically reduces the seek count to 1/10
  (far beyond my expectation) and improves the write throughput by 24%.

- long block time in balance_dirty_pages() hurts desktop responsiveness

  Many of us may have the experience: it often takes a couple of seconds
  or even long time to stop a heavy writing dd/cp/tar command with
  Ctrl-C or "kill -9".

- IO pipeline broken by bumpy write() progress

  There are a broad class of "loop {read(buf); write(buf);}" applications
  whose read() pipeline will be under-utilized or even come to a stop if
  the write()s have long latencies _or_ don't progress in a constant rate.
  The current threshold based throttling inherently transfers the large
  low level IO completion fluctuations to bumpy application write()s,
  and further deteriorates with increasing number of dirtiers and/or bdi's.

  For example, when doing 50 dd's + 1 remote rsync to an XFS partition,
  the rsync progresses very bumpy in legacy kernel, and throughput is
  improved by 67% by this patchset. (plus the larger write chunk size,
  it will be 93% speedup).

  The new rate based throttling can support 1000+ dd's with excellent
  smoothness, low latency and low overheads.

For the above reasons, it's much better to do IO-less and low latency
pauses in balance_dirty_pages().

Jan Kara, Dave Chinner and me explored the scheme to let
balance_dirty_pages() wait for enough writeback IO completions to
safeguard the dirty limit. However it's found to have two problems:

- in large NUMA systems, the per-cpu counters may have big accounting
  errors, leading to big throttle wait time and jitters.

- NFS may kill large amount of unstable pages with one single COMMIT.
  Because NFS server serves COMMIT with expensive fsync() IOs, it is
  desirable to delay and reduce the number of COMMITs. So it's not
  likely to optimize away such kind of bursty IO completions, and the
  resulted large (and tiny) stall times in IO completion based throttling.

So here is a pause time oriented approach, which tries to control the
pause time in each balance_dirty_pages() invocations, by controlling
the number of pages dirtied before calling balance_dirty_pages(), for
smooth and efficient dirty throttling:

- avoid useless (eg. zero pause time) balance_dirty_pages() calls
- avoid too small pause time (less than   4ms, which burns CPU power)
- avoid too large pause time (more than 200ms, which hurts responsiveness)
- avoid big fluctuations of pause times

It can control pause times at will. The default policy (in a followup
patch) will be to do ~10ms pauses in 1-dd case, and increase to ~100ms
in 1000-dd case.

BEHAVIOR CHANGE
===============

(1) dirty threshold

Users will notice that the applications will get throttled once crossing
the global (background + dirty)/2=15% threshold, and then balanced around
17.5%. Before patch, the behavior is to just throttle it at 20% dirtyable
memory in 1-dd case.

Since the task will be soft throttled earlier than before, it may be
perceived by end users as performance "slow down" if his application
happens to dirty more than 15% dirtyable memory.

(2) smoothness/responsiveness

Users will notice a more responsive system during heavy writeback.
"killall dd" will take effect instantly.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Add two fields to task_struct.

1) account dirtied pages in the individual tasks, for accuracy
2) per-task balance_dirty_pages() call intervals, for flexibility

The balance_dirty_pages() call interval (ie. nr_dirtied_pause) will
scale near-sqrt to the safety gap between dirty pages and threshold.

The main problem of per-task nr_dirtied is, if 1k+ tasks start dirtying
pages at exactly the same time, each task will be assigned a large
initial nr_dirtied_pause, so that the dirty threshold will be exceeded
long before each task reached its nr_dirtied_pause and hence call
balance_dirty_pages().

The solution is to watch for the number of pages dirtied on each CPU in
between the calls into balance_dirty_pages(). If it exceeds ratelimit_pages
(3% dirty threshold), force call balance_dirty_pages() for a chance to
set bdi->dirty_exceeded. In normal situations, this safeguarding
condition is not expected to trigger at all.

On the sqrt in dirty_poll_interval():

It will serve as an initial guess when dirty pages are still in the
freerun area.

When dirty pages are floating inside the dirty control scope [freerun,
limit], a followup patch will use some refined dirty poll interval to
get the desired pause time.

   thresh-dirty (MB)    sqrt
		   1      16
		   2      22
		   4      32
		   8      45
		  16      64
		  32      90
		  64     128
		 128     181
		 256     256
		 512     362
		1024     512

The above table means, given 1MB (or 1GB) gap and the dd tasks polling
balance_dirty_pages() on every 16 (or 512) pages, the dirty limit won't
be exceeded as long as there are less than 16 (or 512) concurrent dd's.

So sqrt naturally leads to less overheads and more safe concurrent tasks
for large memory servers, which have large (thresh-freerun) gaps.

peter: keep the per-CPU ratelimit for safeguarding the 1k+ tasks case

CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: NAndrea Righi <andrea@betterlinux.com>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

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: NWu Fengguang <fengguang.wu@intel.com>

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: NWu Fengguang <fengguang.wu@intel.com>

No behavior change.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

bdi_position_ratio() provides a scale factor to bdi->dirty_ratelimit, so
that the resulted task rate limit can drive the dirty pages back to the
global/bdi setpoints.

Old scheme is,
                                          |
                           free run area  |  throttle area
  ----------------------------------------+---------------------------->
                                    thresh^                  dirty pages

New scheme is,

  ^ task rate limit
  |
  |            *
  |             *
  |              *
  |[free run]      *      [smooth throttled]
  |                  *
  |                     *
  |                         *
  ..bdi->dirty_ratelimit..........*
  |                               .     *
  |                               .          *
  |                               .              *
  |                               .                 *
  |                               .                    *
  +-------------------------------.-----------------------*------------>
                          setpoint^                  limit^  dirty pages

The slope of the bdi control line should be

1) large enough to pull the dirty pages to setpoint reasonably fast

2) small enough to avoid big fluctuations in the resulted pos_ratio and
   hence task ratelimit

Since the fluctuation range of the bdi dirty pages is typically observed
to be within 1-second worth of data, the bdi control line's slope is
selected to be a linear function of bdi write bandwidth, so that it can
adapt to slow/fast storage devices well.

Assume the bdi control line

	pos_ratio = 1.0 + k * (dirty - bdi_setpoint)

where k is the negative slope.

If targeting for 12.5% fluctuation range in pos_ratio when dirty pages
are fluctuating in range

	[bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2],

we get slope

	k = - 1 / (8 * write_bw)

Let pos_ratio(x_intercept) = 0, we get the parameter used in code:

	x_intercept = bdi_setpoint + 8 * write_bw

The global/bdi slopes are nicely complementing each other when the
system has only one major bdi (indicated by bdi_thresh ~= thresh):

1) slope of global control line    => scaling to the control scope size
2) slope of main bdi control line  => scaling to the writeout bandwidth

so that

- in memory tight systems, (1) becomes strong enough to squeeze dirty
  pages inside the control scope

- in large memory systems where the "gravity" of (1) for pulling the
  dirty pages to setpoint is too weak, (2) can back (1) up and drive
  dirty pages to bdi_setpoint ~= setpoint reasonably fast.

Unfortunately in JBOD setups, the fluctuation range of bdi threshold
is related to memory size due to the interferences between disks.  In
this case, the bdi slope will be weighted sum of write_bw and bdi_thresh.

Given equations

        span = x_intercept - bdi_setpoint
        k = df/dx = - 1 / span

and the extremum values

        span = bdi_thresh
        dx = bdi_thresh

we get

        df = - dx / span = - 1.0

That means, when bdi_dirty deviates bdi_thresh up, pos_ratio and hence
task ratelimit will fluctuate by -100%.

peter: use 3rd order polynomial for the global control line

CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Introduce the BDI_DIRTIED counter. It will be used for estimating the
bdi's dirty bandwidth.

CC: Jan Kara <jack@suse.cz>
CC: Michael Rubin <mrubin@google.com>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Revert the pass-good area introduced in ffd1f609 ("writeback:
introduce max-pause and pass-good dirty limits") and make the max-pause
area smaller and safe.

This fixes ~30% performance regression in the ext3 data=writeback
fio_mmap_randwrite_64k/fio_mmap_randrw_64k test cases, where there are
12 JBOD disks, on each disk runs 8 concurrent tasks doing reads+writes.

Using deadline scheduler also has a regression, but not that big as CFQ,
so this suggests we have some write starvation.

The test logs show that

- the disks are sometimes under utilized

- global dirty pages sometimes rush high to the pass-good area for
  several hundred seconds, while in the mean time some bdi dirty pages
  drop to very low value (bdi_dirty << bdi_thresh).  Then suddenly the
  global dirty pages dropped under global dirty threshold and bdi_dirty
  rush very high (for example, 2 times higher than bdi_thresh). During
  which time balance_dirty_pages() is not called at all.

So the problems are

1) The random writes progress so slow that they break the assumption of
   the max-pause logic that "8 pages per 200ms is typically more than
   enough to curb heavy dirtiers".

2) The max-pause logic ignored task_bdi_thresh and thus opens the possibility
   for some bdi's to over dirty pages, leading to (bdi_dirty >> bdi_thresh)
   and then (bdi_thresh >> bdi_dirty) for others.

3) The higher max-pause/pass-good thresholds somehow leads to the bad
   swing of dirty pages.

The fix is to allow the task to slightly dirty over task_bdi_thresh, but
no way to exceed bdi_dirty and/or global dirty_thresh.

Tests show that it fixed the JBOD regression completely (both behavior
and performance), while still being able to cut down large pause times
in balance_dirty_pages() for single-disk cases.
Reported-by: NLi Shaohua <shaohua.li@intel.com>
Tested-by: NLi Shaohua <shaohua.li@intel.com>
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

NR_WRITTEN is now accounted at block IO enqueue time, which is not very
accurate as to common understanding.  This moves NR_WRITTEN accounting to
the IO completion time and makes it more consistent with BDI_WRITTEN,
which is used for bandwidth estimation.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>
Cc: Michael Rubin <mrubin@google.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

radix_tree_tagged() is lockless - it reads from a member of the raid-tree
root node.  It does not require any protection.
Signed-off-by: NKonstantin Khlebnikov <khlebnikov@openvz.org>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

We set bdi->dirty_exceeded (and thus ratelimiting code starts to
call balance_dirty_pages() every 8 pages) when a per-bdi limit is
exceeded or global limit is exceeded. But per-bdi limit also depends
on the task. Thus different tasks reach the limit on that bdi at
different levels of dirty pages. The result is that with current code
bdi->dirty_exceeded ping-ponged between 1 and 0 depending on which task
just got into balance_dirty_pages().

We fix the issue by clearing bdi->dirty_exceeded only when per-bdi amount
of dirty pages drops below the threshold (7/8 * bdi_dirty_limit) where task
limits already do not have any influence.

Impact:  The end result is, the dirty pages are kept more tightly under
control, with the average number slightly lowered than before.  This
reduces the risk to throttle light dirtiers and hence more responsive.
However it may add overheads by enforcing balance_dirty_pages() calls
on every 8 pages when there are 2+ heavy dirtiers.

CC: Andrew Morton <akpm@linux-foundation.org>
CC: Christoph Hellwig <hch@infradead.org>
CC: Dave Chinner <david@fromorbit.com>
CC: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Add trace event balance_dirty_state for showing the global dirty page
counts and thresholds at each global_dirty_limits() invocation.  This
will cover the callers throttle_vm_writeout(), over_bground_thresh()
and each balance_dirty_pages() loop.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

The max-pause limit helps to keep the sleep time inside
balance_dirty_pages() within MAX_PAUSE=200ms. The 200ms max sleep means
per task rate limit of 8pages/200ms=160KB/s when dirty exceeded, which
normally is enough to stop dirtiers from continue pushing the dirty
pages high, unless there are a sufficient large number of slow dirtiers
(eg. 500 tasks doing 160KB/s will still sum up to 80MB/s, exceeding the
write bandwidth of a slow disk and hence accumulating more and more dirty
pages).

The pass-good limit helps to let go of the good bdi's in the presence of
a blocked bdi (ie. NFS server not responding) or slow USB disk which for
some reason build up a large number of initial dirty pages that refuse
to go away anytime soon.

For example, given two bdi's A and B and the initial state

	bdi_thresh_A = dirty_thresh / 2
	bdi_thresh_B = dirty_thresh / 2
	bdi_dirty_A  = dirty_thresh / 2
	bdi_dirty_B  = dirty_thresh / 2

Then A get blocked, after a dozen seconds

	bdi_thresh_A = 0
	bdi_thresh_B = dirty_thresh
	bdi_dirty_A  = dirty_thresh / 2
	bdi_dirty_B  = dirty_thresh / 2

The (bdi_dirty_B < bdi_thresh_B) test is now useless and the dirty pages
will be effectively throttled by condition (nr_dirty < dirty_thresh).
This has two problems:
(1) we lose the protections for light dirtiers
(2) balance_dirty_pages() effectively becomes IO-less because the
    (bdi_nr_reclaimable > bdi_thresh) test won't be true. This is good
    for IO, but balance_dirty_pages() loses an important way to break
    out of the loop which leads to more spread out throttle delays.

DIRTY_PASSGOOD_AREA can eliminate the above issues. The only problem is,
DIRTY_PASSGOOD_AREA needs to be defined as 2 to fully cover the above
example while this patch uses the more conservative value 8 so as not to
surprise people with too many dirty pages than expected.

The max-pause limit won't noticeably impact the speed dirty pages are
knocked down when there is a sudden drop of global/bdi dirty thresholds.
Because the heavy dirties will be throttled below 160KB/s which is slow
enough. It does help to avoid long dirty throttle delays and especially
will make light dirtiers more responsive.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

The start of a heavy weight application (ie. KVM) may instantly knock
down determine_dirtyable_memory() if the swap is not enabled or full.
global_dirty_limits() and bdi_dirty_limit() will in turn get global/bdi
dirty thresholds that are _much_ lower than the global/bdi dirty pages.

balance_dirty_pages() will then heavily throttle all dirtiers including
the light ones, until the dirty pages drop below the new dirty thresholds.
During this _deep_ dirty-exceeded state, the system may appear rather
unresponsive to the users.

About "deep" dirty-exceeded: task_dirty_limit() assigns 1/8 lower dirty
threshold to heavy dirtiers than light ones, and the dirty pages will
be throttled around the heavy dirtiers' dirty threshold and reasonably
below the light dirtiers' dirty threshold. In this state, only the heavy
dirtiers will be throttled and the dirty pages are carefully controlled
to not exceed the light dirtiers' dirty threshold. However if the
threshold itself suddenly drops below the number of dirty pages, the
light dirtiers will get heavily throttled.

So introduce global_dirty_limit for tracking the global dirty threshold
with policies

- follow downwards slowly
- follow up in one shot

global_dirty_limit can effectively mask out the impact of sudden drop of
dirtyable memory. It will be used in the next patch for two new type of
dirty limits. Note that the new dirty limits are not going to avoid
throttling the light dirtiers, but could limit their sleep time to 200ms.
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Introduce

	nr_dirty = NR_FILE_DIRTY + NR_WRITEBACK + NR_UNSTABLE_NFS

in order to simplify many tests in the following patches.

balance_dirty_pages() will eventually care only about the dirty sums
besides nr_writeback.
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

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: NWu Fengguang <fengguang.wu@intel.com>

Introduce the BDI_WRITTEN counter. It will be used for estimating the
bdi's write bandwidth.

Peter Zijlstra <a.p.zijlstra@chello.nl>:
Move BDI_WRITTEN accounting into __bdi_writeout_inc().
This will cover and fix fuse, which only calls bdi_writeout_inc().

CC: Michael Rubin <mrubin@google.com>
Reviewed-by: NKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Pass struct wb_writeback_work all the way down to writeback_sb_inodes(),
and initialize the struct writeback_control there.

struct writeback_control is basically designed to control writeback of a
single file, but we keep abuse it for writing multiple files in
writeback_sb_inodes() and its callers.

It immediately clean things up, e.g. suddenly wbc.nr_to_write vs
work->nr_pages starts to make sense, and instead of saving and restoring
pages_skipped in writeback_sb_inodes it can always start with a clean
zero value.

It also makes a neat IO pattern change: large dirty files are now
written in the full 4MB writeback chunk size, rather than whatever
remained quota in wbc->nr_to_write.
Acked-by: NJan Kara <jack@suse.cz>
Proposed-by: NChristoph Hellwig <hch@infradead.org>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

This helps prevent tmpfs dirtiers from skewing the per-cpu bdp_ratelimits.
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

This avoids unnecessary checks and dirty throttling on tmpfs/ramfs.

Notes about the tmpfs/ramfs behavior changes:

As for 2.6.36 and older kernels, the tmpfs writes will sleep inside
balance_dirty_pages() as long as we are over the (dirty+background)/2
global throttle threshold.  This is because both the dirty pages and
threshold will be 0 for tmpfs/ramfs. Hence this test will always
evaluate to TRUE:

                dirty_exceeded =
                        (bdi_nr_reclaimable + bdi_nr_writeback >= bdi_thresh)
                        || (nr_reclaimable + nr_writeback >= dirty_thresh);

For 2.6.37, someone complained that the current logic does not allow the
users to set vm.dirty_ratio=0.  So commit 4cbec4c8 changed the test to

                dirty_exceeded =
                        (bdi_nr_reclaimable + bdi_nr_writeback > bdi_thresh)
                        || (nr_reclaimable + nr_writeback > dirty_thresh);

So 2.6.37 will behave differently for tmpfs/ramfs: it will never get
throttled unless the global dirty threshold is exceeded (which is very
unlikely to happen; once happen, will block many tasks).

I'd say that the 2.6.36 behavior is very bad for tmpfs/ramfs. It means
for a busy writing server, tmpfs write()s may get livelocked! The
"inadvertent" throttling can hardly bring help to any workload because
of its "either no throttling, or get throttled to death" property.

So based on 2.6.37, this patch won't bring more noticeable changes.

CC: Hugh Dickins <hughd@google.com>
Acked-by: NRik van Riel <riel@redhat.com>
Acked-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Reviewed-by: NMinchan Kim <minchan.kim@gmail.com>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

Clarify the bdi_dirty_limit() comment.
Acked-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: NJan Kara <jack@suse.cz>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

sync(2) is performed in two stages: the WB_SYNC_NONE sync and the
WB_SYNC_ALL sync. Identify the first stage with .tagged_writepages and
do livelock prevention for it, too.

Jan's commit f446daae ("mm: implement writeback livelock avoidance
using page tagging") is a partial fix in that it only fixed the
WB_SYNC_ALL phase livelock.

Although ext4 is tested to no longer livelock with commit f446daae,
it may due to some "redirty_tail() after pages_skipped" effect which
is by no means a guarantee for _all_ the file systems.

Note that writeback_inodes_sb() is called by not only sync(), they are
treated the same because the other callers also need livelock prevention.

Impact:  It changes the order in which pages/inodes are synced to disk.
Now in the WB_SYNC_NONE stage, it won't proceed to write the next inode
until finished with the current inode.
Acked-by: NJan Kara <jack@suse.cz>
CC: Dave Chinner <david@fromorbit.com>
Signed-off-by: NWu Fengguang <fengguang.wu@intel.com>

For range-cyclic writeback (e.g.  kupdate), the writeback code sets a
continuation point of the next writeback to mapping->writeback_index which
is set the page after the last written page.  This happens so that we
evenly write the whole file even if pages in it get continuously
redirtied.

However, in some cases, sequential writer is writing in the middle of the
page and it just redirties the last written page by continuing from that.
For example with an application which uses a file as a big ring buffer we
see:

[1st writeback session]
       ...
       flush-8:0-2743  4571: block_bio_queue: 8,0 W 94898514 + 8
       flush-8:0-2743  4571: block_bio_queue: 8,0 W 94898522 + 8
       flush-8:0-2743  4571: block_bio_queue: 8,0 W 94898530 + 8
       flush-8:0-2743  4571: block_bio_queue: 8,0 W 94898538 + 8
       flush-8:0-2743  4571: block_bio_queue: 8,0 W 94898546 + 8
     kworker/0:1-11    4571: block_rq_issue: 8,0 W 0 () 94898514 + 40
>>     flush-8:0-2743  4571: block_bio_queue: 8,0 W 94898554 + 8
>>     flush-8:0-2743  4571: block_rq_issue: 8,0 W 0 () 94898554 + 8

[2nd writeback session after 35sec]
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94898562 + 8
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94898570 + 8
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94898578 + 8
       ...
     kworker/0:1-11    4606: block_rq_issue: 8,0 W 0 () 94898562 + 640
     kworker/0:1-11    4606: block_rq_issue: 8,0 W 0 () 94899202 + 72
       ...
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94899962 + 8
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94899970 + 8
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94899978 + 8
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94899986 + 8
       flush-8:0-2743  4606: block_bio_queue: 8,0 W 94899994 + 8
     kworker/0:1-11    4606: block_rq_issue: 8,0 W 0 () 94899962 + 40
>>     flush-8:0-2743  4606: block_bio_queue: 8,0 W 94898554 + 8
>>     flush-8:0-2743  4606: block_rq_issue: 8,0 W 0 () 94898554 + 8

So we seeked back to 94898554 after we wrote all the pages at the end of
the file.

This extra seek seems unnecessary.  If we continue writeback from the last
written page, we can avoid it and do not cause harm to other cases.  The
original intent of even writeout over the whole file is preserved and if
the page does not get redirtied pagevec_lookup_tag() just skips it.

As an exceptional case, when I/O error happens, set done_index to the next
page as the comment in the code suggests.
Tested-by: NWu Fengguang <fengguang.wu@intel.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NJan Kara <jack@suse.cz>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

invalidate_mapping_pages is very big hint to reclaimer.  It means user
doesn't want to use the page any more.  So in order to prevent working set
page eviction, this patch move the page into tail of inactive list by
PG_reclaim.

Please, remember that pages in inactive list are working set as well as
active list.  If we don't move pages into inactive list's tail, pages near
by tail of inactive list can be evicted although we have a big clue about
useless pages.  It's totally bad.

Now PG_readahead/PG_reclaim is shared.  fe3cba17 added ClearPageReclaim
into clear_page_dirty_for_io for preventing fast reclaiming readahead
marker page.

In this series, PG_reclaim is used by invalidated page, too.  If VM find
the page is invalidated and it's dirty, it sets PG_reclaim to reclaim
asap.  Then, when the dirty page will be writeback,
clear_page_dirty_for_io will clear PG_reclaim unconditionally.  It
disturbs this serie's goal.

I think it's okay to clear PG_readahead when the page is dirty, not
writeback time.  So this patch moves ClearPageReadahead.  In v4,
ClearPageReadahead in set_page_dirty has a problem which is reported by
Steven Barrett.  It's due to compound page.  Some driver(ex, audio) calls
set_page_dirty with compound page which isn't on LRU.  but my patch does
ClearPageRelcaim on compound page.  In non-CONFIG_PAGEFLAGS_EXTENDED, it
breaks PageTail flag.

I think it doesn't affect THP and pass my test with THP enabling but Cced
Andrea for double check.
Signed-off-by: NMinchan Kim <minchan.kim@gmail.com>
Reported-by: NSteven Barrett <damentz@liquorix.net>
Reviewed-by: NJohannes Weiner <hannes@cmpxchg.org>
Acked-by: NRik van Riel <riel@redhat.com>
Acked-by: NMel Gorman <mel@csn.ul.ie>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>