[RFC,0/3] reduce THP fault thrashing

Message

Vlastimil Babka Dec. 11, 2018, 2:29 p.m. UTC

Hi,

this is my attempt at reducing the madvised THP fault local node thrashing by
reclaim+compaction attempts which Andrea reported, by trying to better utilize
recent compaction results. It doesn't introduce any new __GFP_ONLY_COMPACT flag
or add order-specific decisions like Andrea's and David's previous patches, but
it does add __GFP_NORETRY back to madvised THP faults, like they both did
(Patch 1). Patch 2 is based on another Andrea's suggestion, where any
compaction failure is a reason to not try further (not just defered
compaction). Finally, patch 3 introduces defered compaction tracking for async
mode which is what's used for THP faults. Details in respective patch
changelogs.

I haven't tested it yet besides running transhuge-stress and verifying via
tracepoints that defered async compaction does happen. I hope all interested
parties can test the series on their workloads, thanks in advance. I expect
that THP fault success rates will be worse, but hopefully it will also fix
the local node thrashing issue. The success rates can then likely be improved
by making compaction core smarter, but that's a separate topic.

The series is based on v4.20-rc6.

Vlastimil

Vlastimil Babka (3):
  mm, thp: restore __GFP_NORETRY for madvised thp fault allocations
  mm, page_alloc: reclaim for __GFP_NORETRY costly requests only when
    compaction was skipped
  mm, compaction: introduce deferred async compaction

 include/linux/compaction.h        | 10 ++--
 include/linux/mmzone.h            |  6 +--
 include/trace/events/compaction.h | 29 ++++++-----
 mm/compaction.c                   | 80 ++++++++++++++++++-------------
 mm/huge_memory.c                  | 13 +++--
 mm/page_alloc.c                   | 14 +++---
 6 files changed, 84 insertions(+), 68 deletions(-)