On 07/01/18 at 11:28pm, Pavel Tatashin wrote: > > > So, on the first failure, we even stop trying to populate other > > > sections. No more memory to do so. > > > > This is the thing I worry about. In old sparse_mem_maps_populate_node() > > you can see, when not present or failed to populate, just continue. This > > is the main difference between yours and the old code. The key logic is > > changed here. > > > > I do not see how we can succeed after the first failure. We still > allocate from the same node: > > sparse_mem_map_populate() may fail only if we could not allocate large > enough buffer vmemmap_buf_start earlier. > > This means that in: > sparse_mem_map_populate() > vmemmap_populate() > vmemmap_populate_hugepages() > vmemmap_alloc_block_buf() (no buffer, so call allocator) > vmemmap_alloc_block(size, node); > __earlyonly_bootmem_alloc(node, size, size, __pa(MAX_DMA_ADDRESS)); > memblock_virt_alloc_try_nid_raw() -> Nothing changes for > this call to succeed. So, all consequent calls to > sparse_mem_map_populate() in this node will fail as well. Yes, you are right, it's improvement. Thanks. > > > > > > Forgot mentioning it's the vervion in mm/sparse-vmemmap.c > > Sorry, I do not understand what is vervion. Typo, 'version', should be. Sorry for that.
