Re: [PATCH v2] virtio-balloon: fix managed page counts when migrating pages between zones

From: David Hildenbrand
Date: Tue Dec 10 2019 - 10:36:59 EST

On 10.12.19 16:24, Michael S. Tsirkin wrote:
> On Tue, Dec 10, 2019 at 02:44:38PM +0100, David Hildenbrand wrote:
>> On 05.12.19 10:24, David Hildenbrand wrote:
>>> In case we have to migrate a ballon page to a newpage of another zone, the
>>> managed page count of both zones is wrong. Paired with memory offlining
>>> (which will adjust the managed page count), we can trigger kernel crashes
>>> and all kinds of different symptoms.
>>>
>>> One way to reproduce:
>>> 1. Start a QEMU guest with 4GB, no NUMA
>>> 2. Hotplug a 1GB DIMM and only the memory to ZONE_NORMAL
>>
>> s/only/online/
>>
>> as requested by Igor.
>>
>>> 3. Inflate the balloon to 1GB
>>> 4. Unplug the DIMM (be quick, otherwise unmovable data ends up on it)
>>> 5. Observe /proc/zoneinfo
>>> Node 0, zone Normal
>>> pages free 16810
>>> min 24848885473806
>>> low 18471592959183339
>>> high 36918337032892872
>>> spanned 262144
>>> present 262144
>>> managed 18446744073709533486
>>> 6. Do anything that requires some memory (e.g., inflate the balloon some
>>> more). The OOM goes crazy and the system crashes
>>> [ 238.324946] Out of memory: Killed process 537 (login) total-vm:27584kB, anon-rss:860kB, file-rss:0kB, shmem-rss:00
>>> [ 238.338585] systemd invoked oom-killer: gfp_mask=0x100cca(GFP_HIGHUSER_MOVABLE), order=0, oom_score_adj=0
>>> [ 238.339420] CPU: 0 PID: 1 Comm: systemd Tainted: G D W 5.4.0-next-20191204+ #75
>>> [ 238.340139] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu4
>>> [ 238.341121] Call Trace:
>>> [ 238.341337] dump_stack+0x8f/0xd0
>>> [ 238.341630] dump_header+0x61/0x5ea
>>> [ 238.341942] oom_kill_process.cold+0xb/0x10
>>> [ 238.342299] out_of_memory+0x24d/0x5a0
>>> [ 238.342625] __alloc_pages_slowpath+0xd12/0x1020
>>> [ 238.343024] __alloc_pages_nodemask+0x391/0x410
>>> [ 238.343407] pagecache_get_page+0xc3/0x3a0
>>> [ 238.343757] filemap_fault+0x804/0xc30
>>> [ 238.344083] ? ext4_filemap_fault+0x28/0x42
>>> [ 238.344444] ext4_filemap_fault+0x30/0x42
>>> [ 238.344789] __do_fault+0x37/0x1a0
>>> [ 238.345087] __handle_mm_fault+0x104d/0x1ab0
>>> [ 238.345450] handle_mm_fault+0x169/0x360
>>> [ 238.345790] do_user_addr_fault+0x20d/0x490
>>> [ 238.346154] do_page_fault+0x31/0x210
>>> [ 238.346468] async_page_fault+0x43/0x50
>>> [ 238.346797] RIP: 0033:0x7f47eba4197e
>>> [ 238.347110] Code: Bad RIP value.
>>> [ 238.347387] RSP: 002b:00007ffd7c0c1890 EFLAGS: 00010293
>>> [ 238.347834] RAX: 0000000000000002 RBX: 000055d196a20a20 RCX: 00007f47eba4197e
>>> [ 238.348437] RDX: 0000000000000033 RSI: 00007ffd7c0c18c0 RDI: 0000000000000004
>>> [ 238.349047] RBP: 00007ffd7c0c1c20 R08: 0000000000000000 R09: 0000000000000033
>>> [ 238.349660] R10: 00000000ffffffff R11: 0000000000000293 R12: 0000000000000001
>>> [ 238.350261] R13: ffffffffffffffff R14: 0000000000000000 R15: 00007ffd7c0c18c0
>>> [ 238.350878] Mem-Info:
>>> [ 238.351085] active_anon:3121 inactive_anon:51 isolated_anon:0
>>> [ 238.351085] active_file:12 inactive_file:7 isolated_file:0
>>> [ 238.351085] unevictable:0 dirty:0 writeback:0 unstable:0
>>> [ 238.351085] slab_reclaimable:5565 slab_unreclaimable:10170
>>> [ 238.351085] mapped:3 shmem:111 pagetables:155 bounce:0
>>> [ 238.351085] free:720717 free_pcp:2 free_cma:0
>>> [ 238.353757] Node 0 active_anon:12484kB inactive_anon:204kB active_file:48kB inactive_file:28kB unevictable:0kB iss
>>> [ 238.355979] Node 0 DMA free:11556kB min:36kB low:48kB high:60kB reserved_highatomic:0KB active_anon:152kB inactivB
>>> [ 238.358345] lowmem_reserve[]: 0 2955 2884 2884 2884
>>> [ 238.358761] Node 0 DMA32 free:2677864kB min:7004kB low:10028kB high:13052kB reserved_highatomic:0KB active_anon:0B
>>> [ 238.361202] lowmem_reserve[]: 0 0 72057594037927865 72057594037927865 72057594037927865
>>> [ 238.361888] Node 0 Normal free:193448kB min:99395541895224kB low:73886371836733356kB high:147673348131571488kB reB
>>> [ 238.364765] lowmem_reserve[]: 0 0 0 0 0
>>> [ 238.365101] Node 0 DMA: 7*4kB (U) 5*8kB (UE) 6*16kB (UME) 2*32kB (UM) 1*64kB (U) 2*128kB (UE) 3*256kB (UME) 2*512B
>>> [ 238.366379] Node 0 DMA32: 0*4kB 1*8kB (U) 2*16kB (UM) 2*32kB (UM) 2*64kB (UM) 1*128kB (U) 1*256kB (U) 1*512kB (U)B
>>> [ 238.367654] Node 0 Normal: 1985*4kB (UME) 1321*8kB (UME) 844*16kB (UME) 524*32kB (UME) 300*64kB (UME) 138*128kB (B
>>> [ 238.369184] Node 0 hugepages_total=0 hugepages_free=0 hugepages_surp=0 hugepages_size=2048kB
>>> [ 238.369915] 130 total pagecache pages
>>> [ 238.370241] 0 pages in swap cache
>>> [ 238.370533] Swap cache stats: add 0, delete 0, find 0/0
>>> [ 238.370981] Free swap = 0kB
>>> [ 238.371239] Total swap = 0kB
>>> [ 238.371488] 1048445 pages RAM
>>> [ 238.371756] 0 pages HighMem/MovableOnly
>>> [ 238.372090] 306992 pages reserved
>>> [ 238.372376] 0 pages cma reserved
>>> [ 238.372661] 0 pages hwpoisoned
>>>
>>> In another instance (older kernel), I was able to observe this
>>> (negative page count :/):
>>> [ 180.896971] Offlined Pages 32768
>>> [ 182.667462] Offlined Pages 32768
>>> [ 184.408117] Offlined Pages 32768
>>> [ 186.026321] Offlined Pages 32768
>>> [ 187.684861] Offlined Pages 32768
>>> [ 189.227013] Offlined Pages 32768
>>> [ 190.830303] Offlined Pages 32768
>>> [ 190.833071] Built 1 zonelists, mobility grouping on. Total pages: -36920272750453009
>>>
>>> In another instance (older kernel), I was no longer able to start any
>>> process:
>>> [root@vm ~]# [ 214.348068] Offlined Pages 32768
>>> [ 215.973009] Offlined Pages 32768
>>> cat /proc/meminfo
>>> -bash: fork: Cannot allocate memory
>>> [root@vm ~]# cat /proc/meminfo
>>> -bash: fork: Cannot allocate memory
>>>
>>> Fix it by properly adjusting the managed page count when migrating if
>>> the zone changed. The managed page count of the zones now looks after
>>> unplug of the DIMM (and after deflating the balloon) just like before
>>> inflating the balloon (and plugging+onlining the DIMM).
>>>
>>> We'll temporarily modify the totalram page count. If this ever becomes a
>>> problem, we can fine tune by providing helpers that don't touch
>>> the totalram pages (e.g., adjust_zone_managed_page_count()).
>>>
>>> Reported-by: Yumei Huang <yuhuang@xxxxxxxxxx>
>>> Fixes: 3dcc0571cd64 ("mm: correctly update zone->managed_pages")
>>> Cc: <stable@xxxxxxxxxxxxxxx> # v3.11+
>>> Cc: "Michael S. Tsirkin" <mst@xxxxxxxxxx>
>>> Cc: Jason Wang <jasowang@xxxxxxxxxx>
>>> Cc: Jiang Liu <liuj97@xxxxxxxxx>
>>> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
>>> Cc: virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx
>>> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
>>> ---
>>>
>>> v1 -> v2:
>>> - Adjust count before enquing newpage (and it possibly gets free form the
>>> balloon)
>>> - Check if the zone changed
>>>
>>> ---
>>> drivers/virtio/virtio_balloon.c | 7 +++++++
>>> 1 file changed, 7 insertions(+)
>>>
>>> diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c
>>> index 15b7f1d8c334..3078e1ac2a8f 100644
>>> --- a/drivers/virtio/virtio_balloon.c
>>> +++ b/drivers/virtio/virtio_balloon.c
>>> @@ -722,6 +722,13 @@ static int virtballoon_migratepage(struct balloon_dev_info *vb_dev_info,
>>>
>>> get_page(newpage); /* balloon reference */
>>>
>>> + /* fixup the managed page count (esp. of the zone) */
>>
>> /*
>> * When we migrate to a different zone, we have to adjust the managed
>> * page count of both involved zones.
>> */
>>
>> as requested by Michael.
>>
>>
>> @Michael, if there are no further comments, shall I resend?
>
> Also, what does it have to do with deflate on oom?
> Why is it true we don't need to do it then?

Let's have a look at the other users in this file:

t460s: ~/git/linux memory_holes $ git grep -C3 adjust_man drivers/virtio/virtio_balloon.c
drivers/virtio/virtio_balloon.c- vb->num_pages += VIRTIO_BALLOON_PAGES_PER_PAGE;
drivers/virtio/virtio_balloon.c- if (!virtio_has_feature(vb->vdev,
drivers/virtio/virtio_balloon.c- VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
drivers/virtio/virtio_balloon.c: adjust_managed_page_count(page, -1);
drivers/virtio/virtio_balloon.c- vb->num_pfns += VIRTIO_BALLOON_PAGES_PER_PAGE;
drivers/virtio/virtio_balloon.c- }
drivers/virtio/virtio_balloon.c-
--
drivers/virtio/virtio_balloon.c- list_for_each_entry_safe(page, next, pages, lru) {
drivers/virtio/virtio_balloon.c- if (!virtio_has_feature(vb->vdev,
drivers/virtio/virtio_balloon.c- VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
drivers/virtio/virtio_balloon.c: adjust_managed_page_count(page, 1);
drivers/virtio/virtio_balloon.c- list_del(&page->lru);
drivers/virtio/virtio_balloon.c- put_page(page); /* balloon reference */
drivers/virtio/virtio_balloon.c- }

In case we don't have VIRTIO_BALLOON_F_DEFLATE_ON_OOM, we don't touch the managed
page count when inflating/deflating, therefore, there is nothing to do when migrating
a page between zones.

Why? Because the designers of VIRTIO_BALLOON_F_DEFLATE_ON_OOM decided to make
it look like the VM has all RAM available (and make it look like it's only
allocated, not "unplugged"). With VIRTIO_BALLOON_F_DEFLATE_ON_OOM, the system
will reclaim memory from the balloon when OOM and, therefore, "shrink" the
balloon automatically.

E.g., with VIRTIO_BALLOON_F_DEFLATE_ON_OOM neither "MemToal" under
/proc/meminfo, nor the managed page count under /proc/zoneinfo will change.
Without VIRTIO_BALLOON_F_DEFLATE_ON_OOM, it will, therefore we have to fixup
only without VIRTIO_BALLOON_F_DEFLATE_ON_OOM.


--
Thanks,

David / dhildenb