Re: Recurring warning in page_copy_sane (inside copy_page_to_iter) when running stress tests involving drop_caches

From: Lech Perczak
Date: Thu May 16 2019 - 11:20:27 EST

W dniu 15.05.2019 oÂ17:18, Matthew Wilcox pisze:
> On Wed, May 15, 2019 at 08:02:17AM -0700, Eric Dumazet wrote:
>> On Wed, May 15, 2019 at 7:43 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>>> You're seeing a race between page_address(page) being called twice.
>>> Between those two calls, something has caused the page to be removed from
>>> the page_address_map() list. Eric's patch avoids calling page_address(),
>>> so apply it and be happy.
>> Hmm... wont the kmap_atomic() done later, after page_copy_sane() would
>> suffer from the race ?
>> It seems there is a real bug somewhere to fix.
> No. page_address() called before the kmap_atomic() will look through
> the list of mappings and see if that page is mapped somewhere. We unmap
> lazily, so all it takes to trigger this race is that the page _has_
> been mapped before, and its mapping gets torn down during this call.
> While the page is kmapped, its mapping cannot be torn down.

And that's the answer I'm really glad to hear.
In the meantime I've set up a test run with CONFIG_HIGHMEM disabled
to be extra sure, however quite expectedly it robs the system of 256MiB
of accessible memory.

Unfortunatly, looking through defconfigs for 32-bit ARM, CONFIG_HIGHMEM
is enabled in quite a few of them, i.MX included.

I'll pick up the patch then and drop it when it gets included in 4.19.y.

With kind regards,