[PATCH v1 0/5] mm: place pages to the freelist tail when onling and undoing isolation
From: David Hildenbrand
Date: Mon Sep 28 2020 - 14:21:31 EST
When adding separate memory blocks via add_memory*() and onlining them
immediately, the metadata (especially the memmap) of the next block will be
placed onto one of the just added+onlined block. This creates a chain
of unmovable allocations: If the last memory block cannot get
offlined+removed() so will all dependent ones. We directly have unmovable
allocations all over the place.
This can be observed quite easily using virtio-mem, however, it can also
be observed when using DIMMs. The freshly onlined pages will usually be
placed to the head of the freelists, meaning they will be allocated next,
turning the just-added memory usually immediately un-removable. The
fresh pages are cold, prefering to allocate others (that might be hot)
also feels to be the natural thing to do.
It also applies to the hyper-v balloon xen-balloon, and ppc64 dlpar: when
adding separate, successive memory blocks, each memory block will have
unmovable allocations on them - for example gigantic pages will fail to
allocate.
While the ZONE_NORMAL doesn't provide any guarantees that memory can get
offlined+removed again (any kind of fragmentation with unmovable
allocations is possible), there are many scenarios (hotplugging a lot of
memory, running workload, hotunplug some memory/as much as possible) where
we can offline+remove quite a lot with this patchset.
a) To visualize the problem, a very simple example:
Start a VM with 4GB and 8GB of virtio-mem memory:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000033fffffff 9G online yes 32-103
Memory block size: 128M
Total online memory: 12G
Total offline memory: 0B
Then try to unplug as much as possible using virtio-mem. Observe which
memory blocks are still around. Without this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
0x0000000148000000-0x000000014fffffff 128M online yes 41
0x0000000158000000-0x000000015fffffff 128M online yes 43
0x0000000168000000-0x000000016fffffff 128M online yes 45
0x0000000178000000-0x000000017fffffff 128M online yes 47
0x0000000188000000-0x0000000197ffffff 256M online yes 49-50
0x00000001a0000000-0x00000001a7ffffff 128M online yes 52
0x00000001b0000000-0x00000001b7ffffff 128M online yes 54
0x00000001c0000000-0x00000001c7ffffff 128M online yes 56
0x00000001d0000000-0x00000001d7ffffff 128M online yes 58
0x00000001e0000000-0x00000001e7ffffff 128M online yes 60
0x00000001f0000000-0x00000001f7ffffff 128M online yes 62
0x0000000200000000-0x0000000207ffffff 128M online yes 64
0x0000000210000000-0x0000000217ffffff 128M online yes 66
0x0000000220000000-0x0000000227ffffff 128M online yes 68
0x0000000230000000-0x0000000237ffffff 128M online yes 70
0x0000000240000000-0x0000000247ffffff 128M online yes 72
0x0000000250000000-0x0000000257ffffff 128M online yes 74
0x0000000260000000-0x0000000267ffffff 128M online yes 76
0x0000000270000000-0x0000000277ffffff 128M online yes 78
0x0000000280000000-0x0000000287ffffff 128M online yes 80
0x0000000290000000-0x0000000297ffffff 128M online yes 82
0x00000002a0000000-0x00000002a7ffffff 128M online yes 84
0x00000002b0000000-0x00000002b7ffffff 128M online yes 86
0x00000002c0000000-0x00000002c7ffffff 128M online yes 88
0x00000002d0000000-0x00000002d7ffffff 128M online yes 90
0x00000002e0000000-0x00000002e7ffffff 128M online yes 92
0x00000002f0000000-0x00000002f7ffffff 128M online yes 94
0x0000000300000000-0x0000000307ffffff 128M online yes 96
0x0000000310000000-0x0000000317ffffff 128M online yes 98
0x0000000320000000-0x0000000327ffffff 128M online yes 100
0x0000000330000000-0x000000033fffffff 256M online yes 102-103
Memory block size: 128M
Total online memory: 8.1G
Total offline memory: 0B
With this patch set:
[root@localhost ~]# lsmem
RANGE SIZE STATE REMOVABLE BLOCK
0x0000000000000000-0x00000000bfffffff 3G online yes 0-23
0x0000000100000000-0x000000013fffffff 1G online yes 32-39
Memory block size: 128M
Total online memory: 4G
Total offline memory: 0B
All memory can get unplugged, all memory block can get removed. Of course,
no workload ran and the system was basically idle, but it highlights the
issue - the fairly deterministic chain of unmovable allocations. When a
huge page for the 2MB memmap is needed, a just-onlined 4MB page will
be split. The remaining 2MB page will be used for the memmap of the next
memory block. So one memory block will hold the memmap of the two following
memory blocks. Finally the pages of the last-onlined memory block will get
used for the next bigger allocations - if any allocation is unmovable,
all dependent memory blocks cannot get unplugged and removed until that
allocation is gone.
Note that with bigger memory blocks (e.g., 256MB), *all* memory
blocks are dependent and none can get unplugged again!
b) Experiment with memory intensive workload
I performed an experiment with an older version of this patch set
(before we used undo_isolate_page_range() in online_pages():
Hotplug 56GB to a VM with an initial 4GB, onlining all memory to
ZONE_NORMAL right from the kernel when adding it. I then run various
memory intensive workloads that consume most system memory for a total of
45 minutes. Once finished, I try to unplug as much memory as possible.
With this change, I am able to remove via virtio-mem (adding individual
128MB memory blocks) 413 out of 448 added memory blocks. Via individual
(256MB) DIMMs 380 out of 448 added memory blocks. (I don't have any numbers
without this patchset, but looking at the above example, it's at most half
of the 448 memory blocks for virtio-mem, and most probably none for DIMMs).
Again, there are workloads that might behave very differently due to the
nature of ZONE_NORMAL.
This change also affects (besodes memory onlining):
- Other users of undo_isolate_page_range(): Pages are always placed to the
tail.
-- When memory offlining fails
-- When memory isolation fails after having isolated some pageblocks
-- When alloc_contig_range() either succeeds or fails
- Other users of __putback_isolated_page(): Pages are always placed to the
tail.
-- Free page reporting
- Other users of __free_pages_core()
-- AFAIKs, any memory that is getting exposed to the buddy during boot.
IIUC we will now usually allocate memory from lower addresses within
a zone first (especially during boot).
- Other users of generic_online_page()
-- Hyper-V balloon
RFC -> v1:
- Tweak some patch descriptions
- "mm/page_alloc: place pages to tail in __putback_isolated_page()"
-- FOP_TO_TAIL now has higher precedence than page shuffling
-- Add a note that nothing should rely on FOP_TO_TAIL for correctness
- "mm/page_alloc: always move pages to the tail of the freelist in
unset_migratetype_isolate()"
-- Use "bool" parameter for move_freepages_block() as requested
- "mm/page_alloc: place pages to tail in __free_pages_core()"
-- Eliminate set_page_refcounted() + page_ref_dec() and add a comment
- "mm/memory_hotplug: update comment regarding zone shuffling"
-- Added
David Hildenbrand (5):
mm/page_alloc: convert "report" flag of __free_one_page() to a proper
flag
mm/page_alloc: place pages to tail in __putback_isolated_page()
mm/page_alloc: always move pages to the tail of the freelist in
unset_migratetype_isolate()
mm/page_alloc: place pages to tail in __free_pages_core()
mm/memory_hotplug: update comment regarding zone shuffling
include/linux/page-isolation.h | 4 +-
mm/memory_hotplug.c | 11 ++--
mm/page_alloc.c | 114 ++++++++++++++++++++++++---------
mm/page_isolation.c | 12 +++-
4 files changed, 97 insertions(+), 44 deletions(-)
--
2.26.2