Re: [PATCH v1] mm/memory_hotplug: Easier calculation to get pages to next section boundary

From: Baoquan He <bhe@redhat.com>
To: Wei Yang <richardw.yang@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>,
	linux-kernel@vger.kernel.org, linux-mm@kvack.org,
	Segher Boessenkool <segher@kernel.crashing.org>,
	Andrew Morton <akpm@linux-foundation.org>,
	Michal Hocko <mhocko@kernel.org>,
	Oscar Salvador <osalvador@suse.de>
Subject: Re: [PATCH v1] mm/memory_hotplug: Easier calculation to get pages to next section boundary
Date: Thu, 6 Feb 2020 08:37:36 +0800	[thread overview]
Message-ID: <20200206003736.GI8965@MiWiFi-R3L-srv> (raw)
In-Reply-To: <20200206001317.GH8965@MiWiFi-R3L-srv>

On 02/06/20 at 08:13am, Baoquan He wrote:
> On 02/06/20 at 07:50am, Wei Yang wrote:
> > On Thu, Feb 06, 2020 at 07:19:45AM +0800, Wei Yang wrote:
> > >On Wed, Feb 05, 2020 at 02:52:51PM +0100, David Hildenbrand wrote:
> > >>Let's use a calculation that's easier to understand and calculates the
> > >>same result. Reusing existing macros makes this look nicer.
> > >>
> > >>We always want to have the number of pages (> 0) to the next section
> > >>boundary, starting from the current pfn.
> > >>
> > >>Suggested-by: Segher Boessenkool <segher@kernel.crashing.org>
> > >>Cc: Andrew Morton <akpm@linux-foundation.org>
> > >>Cc: Michal Hocko <mhocko@kernel.org>
> > >>Cc: Oscar Salvador <osalvador@suse.de>
> > >>Cc: Baoquan He <bhe@redhat.com>
> > >>Cc: Wei Yang <richardw.yang@linux.intel.com>
> > >>Signed-off-by: David Hildenbrand <david@redhat.com>
> > >
> > >Reviewed-by: Wei Yang <richardw.yang@linux.intel.com>
> > >
> > >BTW, I got one question about hotplug size requirement.
> > >
> > >I thought the hotplug range should be section size aligned, while taking a
> > >look into current code function check_hotplug_memory_range() guard the range.
> 
> A good question. The current code should be block size aligned. I
> remember in some places we assume each block comprise all the sections.
> Can't imagine one or some of them are half section filled.

I could be wrong, half filled block may not cause problem. 

> 
> It truly has a risk that system ram is very huge to make the block
> size is 2G, someone try to insert a 1G memory board. However, it should
> only exist in experiment environment, e.g build a guest with enough ram,
> then hot add 1G DIMM. In reality, we don't need to worry about it, at
> least what I saw is 512G order of magnitude.
> 
> > >
> > >This function says the range should be block_size aligned. And if I am
> > >correct, block size on x86 should be in the range
> > >
> > >    [MIN_MEMORY_BLOCK_SIZE, MEM_SIZE_FOR_LARGE_BLOCK]
> > >    
> > >And MIN_MEMORY_BLOCK_SIZE is section size.
> 
> No, if I got it right, the range on x86 is
> [MIN_MEMORY_BLOCK_SIZE, MAX_BLOCK_SIZE].
> 
> MEM_SIZE_FOR_LARGE_BLOCK is the starting point from which block size can
> be adjusted. Otherwise it's MIN_MEMORY_BLOCK_SIZE.
> 
> /* Amount of ram needed to start using large blocks */                                                                                            
> #define MEM_SIZE_FOR_LARGE_BLOCK (64UL << 30)
> 
> > >
> > >Seems currently we support subsection hotplug? Then how a subsection range got
> > >hotplug? Or this patch is a pre-requisite?
> 
> The sub-section hotplug feature was added by your colleague Dan
> Williams. It intends to fix a nvdimms issue that nvdimms device could be
> mapped into a non section size aligned starting address. And nvdimms
> makes use of the existing memory hotplug mechanism to manage pages.
> Not sure if we are saying the same thing.
> 
> > >
> > 
> > One more question is we support hot-add subsection memory but not support
> > hot-online subsection memory.
> > 
> > Is my understanding correct?
> > 
> > -- 
> > Wei Yang
> > Help you, Help me
> > 