Re: [RFC 2/6] mm/migrate_pages: split unmap_and_move() to _unmap() and _move()

["Huang, Ying" <ying.huang@intel.com>]

Alistair Popple <apopple@nvidia.com> writes:

> Yang Shi <shy828301@gmail.com> writes:
>
>> On Tue, Sep 27, 2022 at 1:35 PM John Hubbard <jhubbard@nvidia.com> wrote:
>>>
>>> On 9/26/22 18:51, Huang, Ying wrote:
>>> >>> But there might be other cases which may incur deadlock, for example,
>>> >>> filesystem writeback IIUC. Some filesystems may lock a bunch of pages
>>> >>> then write them back in a batch. The same pages may be on the
>>> >>> migration list and they are also dirty and seen by writeback. I'm not
>>> >>> sure whether I miss something that could prevent such a deadlock from
>>> >>> happening.
>>> >>
>>> >> I'm not overly familiar with that area but I would assume any filesystem
>>> >> code doing this would already have to deal with deadlock potential.
>>> >
>>> > Thank you very much for pointing this out.  I think the deadlock is a
>>> > real issue.  Anyway, we shouldn't forbid other places in kernel to lock
>>> > 2 pages at the same time.
>>> >
>>>
>>> I also agree that we cannot make any rules such as "do not lock > 1 page
>>> at the same time, elsewhere in the kernel", because it is already
>>> happening, for example in page-writeback.c, which locks PAGEVEC_SIZE
>>> (15) pages per batch [1].
>
> That's not really the case though. The inner loop of write_cache_page()
> only ever locks one page at a time, either directly via the
> unlock_page() on L2338 (those goto's are amazing) or indirectly via
> (*writepage)() on L2359.
>
> So there's no deadlock potential there because unlocking any previously
> locked page(s) doesn't depend on obtaining the lock for another page.
> Unless I've missed something?

Yes.  This is my understanding too after checking ext4_writepage().

Best Regards,
Huang, Ying

>>> The only deadlock prevention convention that I see is the convention of
>>> locking the pages in order of ascending address. That only helps if
>>> everything does it that way, and migrate code definitely does not.
>>> However...I thought that up until now, at least, the migrate code relied
>>> on trylock (which can fail, and so migration can fail, too), to avoid
>>> deadlock. Is that changing somehow, I didn't see it?
>>
>> The trylock is used by async mode which does try to avoid blocking.
>> But sync mode does use lock. The current implementation of migration
>> does migrate one page at a time, so it is not a problem.
>>
>>>
>>>
>>> [1] https://elixir.bootlin.com/linux/latest/source/mm/page-writeback.c#L2296
>>>
>>> thanks,
>>>
>>> --
>>> John Hubbard
>>> NVIDIA
>>>
>>> > The simplest solution is to batch page migration only if mode ==
>>> > MIGRATE_ASYNC.  Then we may consider to fall back to non-batch mode if
>>> > mode != MIGRATE_ASYNC and trylock page fails.
>>> >
>>>
>>>