On Wed, Aug 08, 2018 at 03:42:33PM +0200, Oscar Salvador wrote: > On Wed, Aug 08, 2018 at 10:08:41AM +0200, David Hildenbrand wrote: > > Then it is maybe time to cleary distinguish both types of memory, as > > they are fundamentally different when it comes to online/offline behavior. > > > > Ordinary ram: > > add_memory ... > > online_pages ... > > offline_pages > > remove_memory > > > > Device memory > > add_device_memory ... > > remove_device_memory > > > > So adding/removing from the zone and stuff can be handled there. > > Uhm, I have been thinking about this. > Maybe we could do something like (completely untested): > > > == memory_hotplug code == > > int add_device_memory(int nid, unsigned long start, unsigned long size, > struct vmem_altmap *altmap, bool mapping) > { > int ret; > unsigned long start_pfn = PHYS_PFN(start); > unsigned long nr_pages = size >> PAGE_SHIFT; > > mem_hotplug_begin(); > if (mapping) > ret = arch_add_memory(nid, start, size, altmap, false) > else > ret = add_pages(nid, start_pfn, nr_pages, altmap, false): > > if (!ret) { > pgdata_t *pgdata = NODE_DATA(nid); > struct zone *zone = pgdata->node_zones[ZONE_DEVICE]; > > online_mem_sections(start_pfn, start_pfn + nr_pages); > move_pfn_range_to_zone(zone, start_pfn, nr_pages, altmap); > } > mem_hotplug_done(); > > return ret; > } > > int del_device_memory(int nid, unsigned long start, unsigned long size, > struct vmem_altmap *altmap, bool mapping) > { > int ret; > unsigned long start_pfn = PHYS_PFN(start); > unsigned long nr_pages = size >> PAGE_SHIFT; > pgdata_t *pgdata = NODE_DATA(nid); > struct zone *zone = pgdata->node_zones[ZONE_DEVICE]; > > mem_hotplug_begin(); > > offline_mem_sections(start_pfn, start_pfn + nr_pages); > __shrink_pages(zone, start_pfn, start_pfn + nr_pages, nr_pages); > > if (mapping) > ret = arch_remove_memory(nid, start, size, altmap) > else > ret = __remove_pages(nid, start_pfn, nr_pages, altmap) > > mem_hotplug_done(); > > return ret; > } > > === > > And then, HMM/devm code could use it. > > For example: > > hmm_devmem_pages_create(): > > ... > ... > if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC) > linear_mapping = true; > else > linear_mapping = false; > > ret = add_device_memory(nid, align_start, align_size, NULL, linear_mapping); > if (ret) > goto error_add_memory; > ... > ... > > > hmm_devmem_release: > > ... > ... > if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY) > mapping = false; > else > mapping = true; > > del_device_memory(nid, start_pfn << PAGE_SHIFT, npages << PAGE_SHIFT, > NULL, > mapping); > ... > ... > > > In this way, we do not need to play tricks in HMM/devm code, we just need to > call those functions when adding/removing memory. Note that Dan did post patches that already go in that direction (unifying code between devm and HMM). I think they are in Andrew queue, looks for mm: Rework hmm to use devm_memremap_pages and other fixes > > We would still have to figure out a way to go for the release_mem_region_adjustable() stuff though. Yes. Cheers, Jérôme