On Wed, Dec 09, 2020 at 11:32:26AM +0100, Oscar Salvador wrote: > On Wed, Dec 02, 2020 at 11:05:53AM +0100, David Hildenbrand wrote: > > If you take a look at generic_online_page() there are some things that > > won't be done for our vmemmap pages > > > > 1. kernel_map_pages(page, 1 << order, 1); > > > > We're accessing these pages already when initializing the memmap. We > > might have to explicitly map these vmemmap pages at some point. Might > > require some thought. Did you test with debug pagealloc? > > I always try to run with all debug stuff enabled, but I definitely > did not enable debug_pagealloc. > I will have a look at it. > > > 2. totalram_pages_add(1UL << order); > > > > We should add/remove the vmemmap pages manually from totalram I guess. > > Yes, we should. That was a clear oversight. Looking closer, I do not think we have to account those into totalram. I might be mistaken but looking at memblock_free_all, it seems we only account to totalram_pages those ranges laying in memblock.memory filtering out memblock.reserved. And it seems that the pages we use for pglist_data structs (the ones we handle in register_page_bootmem_info_node) fall in memblock.reserved. -- Oscar Salvador SUSE L3
