On Wed, Sep 26, 2018 at 11:25 AM Alexander Duyck <alexander.h.duyck@xxxxxxxxxxxxxxx> wrote: > > > > 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 > >> drivers. > > > > 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 > > duplication. > > 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 > hotplug lock. > > 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 > parallel. 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 to > 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 [1]. 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. [1]: https://patchwork.kernel.org/patch/10527229/ just focus on the mm/sparse-vmemmap.c changes at the end.
