Similarly to direct reclaim/compaction, kswapd attempts to combine
reclaim and compaction to attempt making memory allocation of given
order available.

The details differ from direct reclaim e.g. in having high watermark as
a goal.  The code involved in kswapd's reclaim/compaction decisions has
evolved to be quite complex.

Testing reveals that it doesn't actually work in at least one scenario,
and closer inspection suggests that it could be greatly simplified
without compromising on the goal (make high-order page available) or
efficiency (don't reclaim too much).  The simplification relieas of
doing all compaction in kcompactd, which is simply woken up when high
watermarks are reached by kswapd's reclaim.

The scenario where kswapd compaction doesn't work was found with mmtests
test stress-highalloc configured to attempt order-9 allocations without
direct reclaim, just waking up kswapd.  There was no compaction attempt
from kswapd during the whole test.  Some added instrumentation shows
what happens:

 - balance_pgdat() sets end_zone to Normal, as it's not balanced
 - reclaim is attempted on DMA zone, which sets nr_attempted to 99, but
   it cannot reclaim anything, so sc.nr_reclaimed is 0
 - for zones DMA32 and Normal, kswapd_shrink_zone uses testorder=0, so
   it merely checks if high watermarks were reached for base pages.
   This is true, so no reclaim is attempted.  For DMA, testorder=0
   wasn't used, as compaction_suitable() returned COMPACT_SKIPPED
 - even though the pgdat_needs_compaction flag wasn't set to false, no
   compaction happens due to the condition sc.nr_reclaimed >
   nr_attempted being false (as 0 < 99)
 - priority-- due to nr_reclaimed being 0, repeat until priority reaches
   0 pgdat_balanced() is false as only the small zone DMA appears
   balanced (curiously in that check, watermark appears OK and
   compaction_suitable() returns COMPACT_PARTIAL, because a lower
   classzone_idx is used there)

Now, even if it was decided that reclaim shouldn't be attempted on the
DMA zone, the scenario would be the same, as (sc.nr_reclaimed=0 >
nr_attempted=0) is also false.  The condition really should use >= as
the comment suggests.  Then there is a mismatch in the check for setting
pgdat_needs_compaction to false using low watermark, while the rest uses
high watermark, and who knows what other subtlety.  Hopefully this
demonstrates that this is unsustainable.

Luckily we can simplify this a lot.  The reclaim/compaction decisions
make sense for direct reclaim scenario, but in kswapd, our primary goal
is to reach high watermark in order-0 pages.  Afterwards we can attempt
compaction just once.  Unlike direct reclaim, we don't reclaim extra
pages (over the high watermark), the current code already disallows it
for good reasons.

After this patch, we simply wake up kcompactd to process the pgdat,
after we have either succeeded or failed to reach the high watermarks in
kswapd, which goes to sleep.  We pass kswapd's order and classzone_idx,
so kcompactd can apply the same criteria to determine which zones are
worth compacting.  Note that we use the classzone_idx from
wakeup_kswapd(), not balanced_classzone_idx which can include higher
zones that kswapd tried to balance too, but didn't consider them in
pgdat_balanced().

Since kswapd now cannot create high-order pages itself, we need to
adjust how it determines the zones to be balanced.  The key element here
is adding a "highorder" parameter to zone_balanced, which, when set to
false, makes it consider only order-0 watermark instead of the desired
higher order (this was done previously by kswapd_shrink_zone(), but not
elsewhere).  This false is passed for example in pgdat_balanced().
Importantly, wakeup_kswapd() uses true to make sure kswapd and thus
kcompactd are woken up for a high-order allocation failure.

The last thing is to decide what to do with pageblock_skip bitmap
handling.  Compaction maintains a pageblock_skip bitmap to record
pageblocks where isolation recently failed.  This bitmap can be reset by
three ways:

1) direct compaction is restarting after going through the full deferred cycle

2) kswapd goes to sleep, and some other direct compaction has previously
   finished scanning the whole zone and set zone->compact_blockskip_flush.
   Note that a successful direct compaction clears this flag.

3) compaction was invoked manually via trigger in /proc

The case 2) is somewhat fuzzy to begin with, but after introducing
kcompactd we should update it.  The check for direct compaction in 1),
and to set the flush flag in 2) use current_is_kswapd(), which doesn't
work for kcompactd.  Thus, this patch adds bool direct_compaction to
compact_control to use in 2).  For the case 1) we remove the check
completely - unlike the former kswapd compaction, kcompactd does use the
deferred compaction functionality, so flushing tied to restarting from
deferred compaction makes sense here.

Note that when kswapd goes to sleep, kcompactd is woken up, so it will
see the flushed pageblock_skip bits.  This is different from when the
former kswapd compaction observed the bits and I believe it makes more
sense.  Kcompactd can afford to be more thorough than a direct
compaction trying to limit allocation latency, or kswapd whose primary
goal is to reclaim.

For testing, I used stress-highalloc configured to do order-9
allocations with GFP_NOWAIT|__GFP_HIGH|__GFP_COMP, so they relied just
on kswapd/kcompactd reclaim/compaction (the interfering kernel builds in
phases 1 and 2 work as usual):

stress-highalloc
                        4.5-rc1+before          4.5-rc1+after
                             -nodirect              -nodirect
Success 1 Min          1.00 (  0.00%)         5.00 (-66.67%)
Success 1 Mean         1.40 (  0.00%)         6.20 (-55.00%)
Success 1 Max          2.00 (  0.00%)         7.00 (-16.67%)
Success 2 Min          1.00 (  0.00%)         5.00 (-66.67%)
Success 2 Mean         1.80 (  0.00%)         6.40 (-52.38%)
Success 2 Max          3.00 (  0.00%)         7.00 (-16.67%)
Success 3 Min         34.00 (  0.00%)        62.00 (  1.59%)
Success 3 Mean        41.80 (  0.00%)        63.80 (  1.24%)
Success 3 Max         53.00 (  0.00%)        65.00 (  2.99%)

User                          3166.67        3181.09
System                        1153.37        1158.25
Elapsed                       1768.53        1799.37

                            4.5-rc1+before   4.5-rc1+after
                                 -nodirect    -nodirect
Direct pages scanned                32938        32797
Kswapd pages scanned              2183166      2202613
Kswapd pages reclaimed            2152359      2143524
Direct pages reclaimed              32735        32545
Percentage direct scans                1%           1%
THP fault alloc                       579          612
THP collapse alloc                    304          316
THP splits                              0            0
THP fault fallback                    793          778
THP collapse fail                      11           16
Compaction stalls                    1013         1007
Compaction success                     92           67
Compaction failures                   920          939
Page migrate success               238457       721374
Page migrate failure                23021        23469
Compaction pages isolated          504695      1479924
Compaction migrate scanned         661390      8812554
Compaction free scanned          13476658     84327916
Compaction cost                       262          838

After this patch we see improvements in allocation success rate
(especially for phase 3) along with increased compaction activity.  The
compaction stalls (direct compaction) in the interfering kernel builds
(probably THP's) also decreased somewhat thanks to kcompactd activity,
yet THP alloc successes improved a bit.

Note that elapsed and user time isn't so useful for this benchmark,
because of the background interference being unpredictable.  It's just
to quickly spot some major unexpected differences.  System time is
somewhat more useful and that didn't increase.

Also (after adjusting mmtests' ftrace monitor):

Time kswapd awake               2547781     2269241
Time kcompactd awake                  0      119253
Time direct compacting           939937      557649
Time kswapd compacting                0           0
Time kcompactd compacting             0      119099

The decrease of overal time spent compacting appears to not match the
increased compaction stats.  I suspect the tasks get rescheduled and
since the ftrace monitor doesn't see that, the reported time is wall
time, not CPU time.  But arguably direct compactors care about overall
latency anyway, whether busy compacting or waiting for CPU doesn't
matter.  And that latency seems to almost halved.

It's also interesting how much time kswapd spent awake just going
through all the priorities and failing to even try compacting, over and
over.

We can also configure stress-highalloc to perform both direct
reclaim/compaction and wakeup kswapd/kcompactd, by using
GFP_KERNEL|__GFP_HIGH|__GFP_COMP:

stress-highalloc
                        4.5-rc1+before         4.5-rc1+after
                               -direct               -direct
Success 1 Min          4.00 (  0.00%)        9.00 (-50.00%)
Success 1 Mean         8.00 (  0.00%)       10.00 (-19.05%)
Success 1 Max         12.00 (  0.00%)       11.00 ( 15.38%)
Success 2 Min          4.00 (  0.00%)        9.00 (-50.00%)
Success 2 Mean         8.20 (  0.00%)       10.00 (-16.28%)
Success 2 Max         13.00 (  0.00%)       11.00 (  8.33%)
Success 3 Min         75.00 (  0.00%)       74.00 (  1.33%)
Success 3 Mean        75.60 (  0.00%)       75.20 (  0.53%)
Success 3 Max         77.00 (  0.00%)       76.00 (  0.00%)

User                          3344.73       3246.04
System                        1194.24       1172.29
Elapsed                       1838.04       1836.76

                            4.5-rc1+before  4.5-rc1+after
                                   -direct     -direct
Direct pages scanned               125146      120966
Kswapd pages scanned              2119757     2135012
Kswapd pages reclaimed            2073183     2108388
Direct pages reclaimed             124909      120577
Percentage direct scans                5%          5%
THP fault alloc                       599         652
THP collapse alloc                    323         354
THP splits                              0           0
THP fault fallback                    806         793
THP collapse fail                      17          16
Compaction stalls                    2457        2025
Compaction success                    906         518
Compaction failures                  1551        1507
Page migrate success              2031423     2360608
Page migrate failure                32845       40852
Compaction pages isolated         4129761     4802025
Compaction migrate scanned       11996712    21750613
Compaction free scanned         214970969   344372001
Compaction cost                      2271        2694

In this scenario, this patch doesn't change the overall success rate as
direct compaction already tries all it can.  There's however significant
reduction in direct compaction stalls (that is, the number of
allocations that went into direct compaction).  The number of successes
(i.e.  direct compaction stalls that ended up with successful
allocation) is reduced by the same number.  This means the offload to
kcompactd is working as expected, and direct compaction is reduced
either due to detecting contention, or compaction deferred by kcompactd.
In the previous version of this patchset there was some apparent
reduction of success rate, but the changes in this version (such as
using sync compaction only), new baseline kernel, and/or averaging
results from 5 executions (my bet), made this go away.

Ftrace-based stats seem to roughly agree:

Time kswapd awake               2532984     2326824
Time kcompactd awake                  0      257916
Time direct compacting           864839      735130
Time kswapd compacting                0           0
Time kcompactd compacting             0      257585
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: David Rientjes <rientjes@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>