Re: [PATCH 2/2] hwpoison, memory_hotplug: lock folio before unmap hwpoisoned folio

[David Hildenbrand <david@redhat.com>]

On 14.01.25 10:08, mawupeng wrote:
> 
> 
> On 2025/1/13 20:35, David Hildenbrand wrote:
>> On 13.01.25 09:27, Wupeng Ma wrote:
>>> From: Ma Wupeng <mawupeng1@huawei.com>
>>>
>>> Commit b15c87263a69 ("hwpoison, memory_hotplug: allow hwpoisoned pages to
>>> be offlined) add page poison checks in do_migrate_range in order to make
>>> offline hwpoisoned page possible by introducing isolate_lru_page and
>>> try_to_unmap for hwpoisoned page. However folio lock must be held before
>>> calling try_to_unmap. Add it to fix this problem.
>>>
>>> Waring will be produced if filio is not locked during unmap:
>>>
>>>     ------------[ cut here ]------------
>>>     kernel BUG at ./include/linux/swapops.h:400!
>>>     Internal error: Oops - BUG: 00000000f2000800 [#1] PREEMPT SMP
>>>     Modules linked in:
>>>     CPU: 4 UID: 0 PID: 411 Comm: bash Tainted: G        W          6.13.0-rc1-00016-g3c434c7ee82a-dirty #41
>>>     Tainted: [W]=WARN
>>>     Hardware name: QEMU QEMU Virtual Machine, BIOS 0.0.0 02/06/2015
>>>     pstate: 40400005 (nZcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>>>     pc : try_to_unmap_one+0xb08/0xd3c
>>>     lr : try_to_unmap_one+0x3dc/0xd3c
>>>     Call trace:
>>>      try_to_unmap_one+0xb08/0xd3c (P)
>>>      try_to_unmap_one+0x3dc/0xd3c (L)
>>>      rmap_walk_anon+0xdc/0x1f8
>>>      rmap_walk+0x3c/0x58
>>>      try_to_unmap+0x88/0x90
>>>      unmap_poisoned_folio+0x30/0xa8
>>>      do_migrate_range+0x4a0/0x568
>>>      offline_pages+0x5a4/0x670
>>>      memory_block_action+0x17c/0x374
>>>      memory_subsys_offline+0x3c/0x78
>>>      device_offline+0xa4/0xd0
>>>      state_store+0x8c/0xf0
>>>      dev_attr_store+0x18/0x2c
>>>      sysfs_kf_write+0x44/0x54
>>>      kernfs_fop_write_iter+0x118/0x1a8
>>>      vfs_write+0x3a8/0x4bc
>>>      ksys_write+0x6c/0xf8
>>>      __arm64_sys_write+0x1c/0x28
>>>      invoke_syscall+0x44/0x100
>>>      el0_svc_common.constprop.0+0x40/0xe0
>>>      do_el0_svc+0x1c/0x28
>>>      el0_svc+0x30/0xd0
>>>      el0t_64_sync_handler+0xc8/0xcc
>>>      el0t_64_sync+0x198/0x19c
>>>     Code: f9407be0 b5fff320 d4210000 17ffff97 (d4210000)
>>>     ---[ end trace 0000000000000000 ]---
>>>
>>> Fixes: b15c87263a69 ("hwpoison, memory_hotplug: allow hwpoisoned pages to be offlined")
>>> Signed-off-by: Ma Wupeng <mawupeng1@huawei.com>
>>> ---
>>>    mm/memory_hotplug.c | 6 +++++-
>>>    1 file changed, 5 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
>>> index 330668d37e44..9bedecfc3577 100644
>>> --- a/mm/memory_hotplug.c
>>> +++ b/mm/memory_hotplug.c
>>> @@ -1805,8 +1805,12 @@ static void do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
>>>                (folio_test_large(folio) && folio_test_has_hwpoisoned(folio))) {
>>>                if (WARN_ON(folio_test_lru(folio)))
>>>                    folio_isolate_lru(folio);
>>> -            if (folio_mapped(folio))
>>> +            if (folio_mapped(folio)) {
>>> +                folio_lock(folio);
>>>                    unmap_poisoned_folio(folio, TTU_IGNORE_MLOCK | TTU_HWPOISON);
>>> +                folio_unlock(folio);
>>> +            }
>>
>> The comment above says "have elevated reference counts", but I I wonder if this code could race with un-poisoning (although probably a rare event).
>>
>>
>> If there is an elevated reference already, why not move that chunk after the folio_try_get() and just drop the comment that describes the implied magic?
>>
>> I mean, migration of hwpoison is dangerous either way, so that's not the biggest problem I guess :)
> 
> AFAICT during memory_failure, the refcount will not be cleared for poisoned page in order to keep it away from free buddy pages.
> So move that chunk after the folio_try_get() do seems nice and poisoned free pages do seems weird to me.

If the "free" poisoned pages have a raised refcount, folio_try_get() 
would be able to grab them to essentially skip them (not mapped).

If the "free" poisoned pages don't have a raised refcount (impossible 
right now IIRC), folio_try_get() would skip, them, which would be the 
right thing to do.

So I think that should work.

-- 
Cheers,

David / dhildenb