Skip to content
Commit a49bd4d7 authored by Michal Hocko's avatar Michal Hocko Committed by Linus Torvalds
Browse files

mm, numa: rework do_pages_move

Patch series "unclutter thp migration"

Motivation:

THP migration is hacked into the generic migration with rather
surprising semantic.  The migration allocation callback is supposed to
check whether the THP can be migrated at once and if that is not the
case then it allocates a simple page to migrate.  unmap_and_move then
fixes that up by splitting the THP into small pages while moving the
head page to the newly allocated order-0 page.  Remaining pages are
moved to the LRU list by split_huge_page.  The same happens if the THP
allocation fails.  This is really ugly and error prone [2].

I also believe that split_huge_page to the LRU lists is inherently wrong
because all tail pages are not migrated.  Some callers will just work
around that by retrying (e.g.  memory hotplug).  There are other pfn
walkers which are simply broken though.  e.g. madvise_inject_error will
migrate head and then advances next pfn by the huge page size.
do_move_page_to_node_array, queue_pages_range (migrate_pages, mbind),
will simply split the THP before migration if the THP migration is not
supported then falls back to single page migration but it doesn't handle
tail pages if the THP migration path is not able to allocate a fresh THP
so we end up with ENOMEM and fail the whole migration which is a
questionable behavior.  Page compaction doesn't try to migrate large
pages so it should be immune.

The first patch reworks do_pages_move which relies on a very ugly
calling semantic when the return status is pushed to the migration path
via private pointer.  It uses pre allocated fixed size batching to
achieve that.  We simply cannot do the same if a THP is to be split
during the migration path which is done in the patch 3.  Patch 2 is
follow up cleanup which removes the mentioned return status calling
convention ugliness.

On a side note:

There are some semantic issues I have encountered on the way when
working on patch 1 but I am not addressing them here.  E.g. trying to
move THP tail pages will result in either success or EBUSY (the later
one more likely once we isolate head from the LRU list).  Hugetlb
reports EACCESS on tail pages.  Some errors are reported via status
parameter but migration failures are not even though the original
`reason' argument suggests there was an intention to do so.  From a
quick look into git history this never worked.  I have tried to keep the
semantic unchanged.

Then there is a relatively minor thing that the page isolation might
fail because of pages not being on the LRU - e.g. because they are
sitting on the per-cpu LRU caches.  Easily fixable.

This patch (of 3):

do_pages_move is supposed to move user defined memory (an array of
addresses) to the user defined numa nodes (an array of nodes one for
each address).  The user provided status array then contains resulting
numa node for each address or an error.  The semantic of this function
is little bit confusing because only some errors are reported back.
Notably migrate_pages error is only reported via the return value.  This
patch doesn't try to address these semantic nuances but rather change
the underlying implementation.

Currently we are processing user input (which can be really large) in
batches which are stored to a temporarily allocated page.  Each address
is resolved to its struct page and stored to page_to_node structure
along with the requested target numa node.  The array of these
structures is then conveyed down the page migration path via private
argument.  new_page_node then finds the corresponding structure and
allocates the proper target page.

What is the problem with the current implementation and why to change
it? Apart from being quite ugly it also doesn't cope with unexpected
pages showing up on the migration list inside migrate_pages path.  That
doesn't happen currently but the follow up patch would like to make the
thp migration code more clear and that would need to split a THP into
the list for some cases.

How does the new implementation work? Well, instead of batching into a
fixed size array we simply batch all pages that should be migrated to
the same node and isolate all of them into a linked list which doesn't
require any additional storage.  This should work reasonably well
because page migration usually migrates larger ranges of memory to a
specific node.  So the common case should work equally well as the
current implementation.  Even if somebody constructs an input where the
target numa nodes would be interleaved we shouldn't see a large
performance impact because page migration alone doesn't really benefit
from batching.  mmap_sem batching for the lookup is quite questionable
and isolate_lru_page which would benefit from batching is not using it
even in the current implementation.

Link: http://lkml.kernel.org/r/20180103082555.14592-2-mhocko@kernel.org


Signed-off-by: default avatarMichal Hocko <mhocko@suse.com>
Acked-by: default avatarKirill A. Shutemov <kirill@shutemov.name>
Reviewed-by: default avatarAndrew Morton <akpm@linux-foundation.org>
Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com>
Cc: Zi Yan <zi.yan@cs.rutgers.edu>
Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Reale <ar@linux.vnet.ibm.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: default avatarAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
parent bfc6b1ca
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment