Revert "mm, thp: restore node-local hugepage allocations"

This reverts commit 2f0799a0 ("mm, thp: restore node-local
hugepage allocations").

commit 2f0799a0 was rightfully applied to avoid the risk of a
severe regression that was reported by the kernel test robot at the end
of the merge window.  Now we understood the regression was a false
positive and was caused by a significant increase in fairness during a
swap trashing benchmark.  So it's safe to re-apply the fix and continue
improving the code from there.  The benchmark that reported the
regression is very useful, but it provides a meaningful result only when
there is no significant alteration in fairness during the workload.  The
removal of __GFP_THISNODE increased fairness.

__GFP_THISNODE cannot be used in the generic page faults path for new
memory allocations under the MPOL_DEFAULT mempolicy, or the allocation
behavior significantly deviates from what the MPOL_DEFAULT semantics are
supposed to be for THP and 4k allocations alike.

Setting THP defrag to "always" or using MADV_HUGEPAGE (with THP defrag
set to "madvise") has never meant to provide an implicit MPOL_BIND on
the "current" node the task is running on, causing swap storms and
providing a much more aggressive behavior than even zone_reclaim_node =
3.

Any workload who could have benefited from __GFP_THISNODE has now to
enable zone_reclaim_mode=1||2||3.  __GFP_THISNODE implicitly provided
the zone_reclaim_mode behavior, but it only did so if THP was enabled:
if THP was disabled, there would have been no chance to get any 4k page
from the current node if the current node was full of pagecache, which
further shows how this __GFP_THISNODE was misplaced in MADV_HUGEPAGE.
MADV_HUGEPAGE has never been intended to provide any zone_reclaim_mode
semantics, in fact the two are orthogonal, zone_reclaim_mode = 1|2|3
must work exactly the same with MADV_HUGEPAGE set or not.

The performance characteristic of memory depends on the hardware
details.  The numbers below are obtained on Naples/EPYC architecture and
the N/A projection extends them to show what we should aim for in the
future as a good THP NUMA locality default.  The benchmark used
exercises random memory seeks (note: the cost of the page faults is not
part of the measurement).

  D0 THP | D0 4k | D1 THP | D1 4k | D2 THP | D2 4k | D3 THP | D3 4k | ...
  0%     | +43%  | +45%   | +106% | +131%  | +224% | N/A    | N/A

D0 means distance zero (i.e.  local memory), D1 means distance one (i.e.
intra socket memory), D2 means distance two (i.e.  inter socket memory),
etc...

For the guest physical memory allocated by qemu and for guest mode
kernel the performance characteristic of RAM is more complex and an
ideal default could be:

  D0 THP | D1 THP | D0 4k | D2 THP | D1 4k | D3 THP | D2 4k | D3 4k | ...
  0%     | +58%   | +101% | N/A    | +222% | N/A    | N/A   | N/A

NOTE: the N/A are projections and haven't been measured yet, the
measurement in this case is done on a 1950x with only two NUMA nodes.
The THP case here means THP was used both in the host and in the guest.

After applying this commit the THP NUMA locality order that we'll get
out of MADV_HUGEPAGE is this:

  D0 THP | D1 THP | D2 THP | D3 THP | ... | D0 4k | D1 4k | D2 4k | D3 4k | ...

Before this commit it was:

  D0 THP | D0 4k | D1 4k | D2 4k | D3 4k | ...

Even if we ignore the breakage of large workloads that can't fit in a
single node that the __GFP_THISNODE implicit "current node" mbind
caused, the THP NUMA locality order provided by __GFP_THISNODE was still
not the one we shall aim for in the long term (i.e.  the first one at
the top).

After this commit is applied, we can introduce a new allocator multi
order API and to replace those two alloc_pages_vmas calls in the page
fault path, with a single multi order call:

        unsigned int order = (1 << HPAGE_PMD_ORDER) | (1 << 0);
        page = alloc_pages_multi_order(..., &order);
        if (!page)
        	goto out;
        if (!(order & (1 << 0))) {
        	VM_WARN_ON(order != 1 << HPAGE_PMD_ORDER);
        	/* THP fault */
        } else {
        	VM_WARN_ON(order != 1 << 0);
        	/* 4k fallback */
        }

The page allocator logic has to be altered so that when it fails on any
zone with order 9, it has to try again with a order 0 before falling
back to the next zone in the zonelist.

After that we need to do more measurements and evaluate if adding an
opt-in feature for guest mode is worth it, to swap "DN 4k | DN+1 THP"
with "DN+1 THP | DN 4k" at every NUMA distance crossing.

Link: http://lkml.kernel.org/r/20190503223146.2312-3-aarcange@redhat.com


Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Rientjes <rientjes@google.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Stefan Priebe - Profihost AG <s.priebe@profihost.ag>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>