Re: [PATCH] mm/memblock: fix potential issue in memblock_search_pfn_nid()

[Wei Yang <richard.weiyang@gmail.com>]

On Wed, Apr 04, 2018 at 07:45:53AM +0200, Michal Hocko wrote:
>On Wed 04-04-18 09:33:57, Wei Yang wrote:
>> On Tue, Apr 03, 2018 at 01:30:41PM +0200, Michal Hocko wrote:
>> >On Mon 02-04-18 09:50:26, Wei Yang wrote:
>> >> On Fri, Mar 30, 2018 at 01:57:27PM -0700, Andrew Morton wrote:
>> >> >On Fri, 30 Mar 2018 11:30:55 +0800 Wei Yang <richard.weiyang@gmail.com> wrote:
>> >> >
>> >> >> memblock_search_pfn_nid() returns the nid and the [start|end]_pfn of the
>> >> >> memory region where pfn sits in. While the calculation of start_pfn has
>> >> >> potential issue when the regions base is not page aligned.
>> >> >> 
>> >> >> For example, we assume PAGE_SHIFT is 12 and base is 0x1234. Current
>> >> >> implementation would return 1 while this is not correct.
>> >> >
>> >> >Why is this not correct?  The caller might want the pfn of the page
>> >> >which covers the base?
>> >> >
>> >> 
>> >> Hmm... the only caller of memblock_search_pfn_nid() is __early_pfn_to_nid(),
>> >> which returns the nid of a pfn and save the [start_pfn, end_pfn] with in the
>> >> same memory region to a cache. So this looks not a good practice to store
>> >> un-exact pfn in the cache.
>> >> 
>> >> >> This patch fixes this by using PFN_UP().
>> >> >> 
>> >> >> The original commit is commit e76b63f80d93 ("memblock, numa: binary search
>> >> >> node id") and merged in v3.12.
>> >> >> 
>> >> >> Signed-off-by: Wei Yang <richard.weiyang@gmail.com>
>> >> >> Cc: 3.12+ <stable@vger.kernel.org>
>> >> >
>> >> >Please fully describe the runtime effects of a bug when fixing that
>> >> >bug.  This description doesn't give enough justification for merging
>> >> >the patch into mainline, let alone -stable.
>> >> 
>> >> Since PFN_UP() and PFN_DOWN() differs when the address is not page aligned, in
>> >> theory we may have two situations like below.
>> >
>> >Have you ever seen a HW that would report page unaligned memory ranges?
>> >Is this even possible?
>> 
>> No, so we don't need to handle this case?
>
>Memblock code is subtle enough to not touch it for something that
>doesn't really exist. We do have some alignment assumptions all over
>so if we have weird configurations with page non-aligned regions of
>memory then we should do the rounding when this is detected rathert than
>spread them all over random places.

Well, I got you concern and I agree to have some assumption to make the code
easy to read or write.

But, my point is current code doesn't keep this assumption always. For
example, __next_mem_pfn_range() and memblock_search_pfn_nid() treat the pfn
range differently. The inconsistency may confuse audience and harm the kernel
in the long turn. 

To not spread this calculation over random places, how about introducing a
helper function memblock_pfn_range() to consolidate the place and help
audience in the future.

For example, https://patchwork.kernel.org/patch/10323845/ tries to optimize
the page initialization which rely on the memblock pfn range. If we could
consolidate the calculation in one place, would it be more convenient to
maintain the kernel?

>-- 
>Michal Hocko
>SUSE Labs

-- 
Wei Yang
Help you, Help me