On Wed, Sep 26, 2018 at 11:25 AM Alexander Duyck
On 9/26/2018 12:55 AM, Michal Hocko wrote:
> On Tue 25-09-18 13:21:24, Alexander Duyck wrote:
>> The ZONE_DEVICE pages were being initialized in two locations. One was with
>> the memory_hotplug lock held and another was outside of that lock. The
>> problem with this is that it was nearly doubling the memory initialization
>> time. Instead of doing this twice, once while holding a global lock and
>> once without, I am opting to defer the initialization to the one outside of
>> the lock. This allows us to avoid serializing the overhead for memory init
>> and we can instead focus on per-node init times.
>> One issue I encountered is that devm_memremap_pages and
>> hmm_devmmem_pages_create were initializing only the pgmap field the same
>> way. One wasn't initializing hmm_data, and the other was initializing it to
>> a poison value. Since this is something that is exposed to the driver in
>> the case of hmm I am opting for a third option and just initializing
>> hmm_data to 0 since this is going to be exposed to unknown third party
> Why cannot you pull move_pfn_range_to_zone out of the hotplug lock? In
> other words why are you making zone device even more special in the
> generic hotplug code when it already has its own means to initialize the
> pfn range by calling move_pfn_range_to_zone. Not to mention the code
So there were a few things I wasn't sure we could pull outside of the
hotplug lock. One specific example is the bits related to resizing the
pgdat and zone. I wanted to avoid pulling those bits outside of the
The other bit that I left inside the hot-plug lock with this approach
was the initialization of the pages that contain the vmemmap.
> That being said I really dislike this patch.
In my mind this was a patch that "killed two birds with one stone". I
had two issues to address, the first one being the fact that we were
performing the memmap_init_zone while holding the hotplug lock, and the
other being the loop that was going through and initializing pgmap in
the hmm and memremap calls essentially added another 20 seconds
(measured for 3TB of memory per node) to the init time. With this patch
I was able to cut my init time per node by that 20 seconds, and then
made it so that we could scale as we added nodes as they could run in
Yeah, at the very least there is no reason for devm_memremap_pages()
to do another loop through all pages, the core should handle this, but
cleaning up the scope of the hotplug lock is needed.
With that said I am open to suggestions if you still feel like I need
follow this up with some additional work. I just want to avoid
introducing any regressions in regards to functionality or performance.
Could we push the hotplug lock deeper to the places that actually need
it? What I found with my initial investigation is that we don't even
need the hotplug lock for the vmemmap initialization with this patch
Alternatively it seems the hotplug lock wants to synchronize changes
to the zone and the page init work. If the hotplug lock was an rwsem
the zone changes would be a write lock, but the init work could be
done as a read lock to allow parallelism. I.e. still provide a sync
point to be able to assert that no hotplug work is in-flight will
holding the write lock, but otherwise allow threads that are touching
independent parts of the memmap to run at the same time.
just focus on the
mm/sparse-vmemmap.c changes at the end.