On 9/1/21 09:22, Mike Rapoport wrote: > On Tue, Aug 31, 2021 at 06:25:23PM +0000, Edgecombe, Rick P wrote: >> On Tue, 2021-08-31 at 11:55 +0300, Mike Rapoport wrote: >> > On Mon, Aug 30, 2021 at 04:59:19PM -0700, Rick Edgecombe wrote: >> <trim> >> > > -static void * __meminit vmemmap_alloc_block_zero(unsigned long >> > > size, int node) >> > > +static void * __meminit vmemmap_alloc_table(int node) >> > > { >> > > - void *p = vmemmap_alloc_block(size, node); >> > > + void *p; >> > > + if (slab_is_available()) { >> > > + struct page *page = alloc_table_node(GFP_KERNEL | >> > > __GFP_ZERO, node); >> > >> > This change removes __GFP_RETRY_MAYFAIL|__GFP_NOWARN from the >> > original gfp >> > vmemmap_alloc_block() used. >> Oh, yea good point. Hmm, I guess grouped pages could be aware of that >> flag too. Would be a small addition, but it starts to grow >> unfortunately. >> >> > Not sure __GFP_RETRY_MAYFAIL is really needed in >> > vmemmap_alloc_block_zero() >> > at the first place, though. >> Looks like due to a real issue: >> 055e4fd96e95b0eee0d92fd54a26be7f0d3bcad0 That commit added __GFP_REPEAT, but __GFP_RETRY_MAYFAIL these days became subtly different. > I believe the issue was with memory map blocks rather than with page > tables, but since sparse-vmemmap uses the same vmemmap_alloc_block() for > both, the GFP flag got stick with both. > > I'm not really familiar with reclaim internals to say if > __GFP_RETRY_MAYFAIL would help much for order-0 allocation. For costly allocation, __GFP_RETRY_MAYFAIL will try harder, thus the RETRY part is accented. For order-0 the only difference is that it will skip OOM, thus the MAYFAIL part. It usually means there's a fallback. I guess in this case there's no fallback, so allocating without __GFP_RETRY_MAYFAIL would be better. > Vlastimil, can you comment on this? > >> I think it should not affect PKS tables for now, so maybe I can make >> separate logic instead. I'll look into it. Thanks. >> > >> > More broadly, maybe it makes sense to split boot time and memory >> > hotplug >> > paths and use pxd_alloc() for the latter. >> > >> > > + >> > > + if (!page) >> > > + return NULL; >> > > + return page_address(page); >> > > + } >> > > >> > > + p = __earlyonly_bootmem_alloc(node, PAGE_SIZE, PAGE_SIZE, >> > > __pa(MAX_DMA_ADDRESS)); >> > >> > Opportunistically rename to __earlyonly_memblock_alloc()? ;-) >> > >> Heh, I can. Just grepping, there are several other instances of >> foo_bootmem() only calling foo_memblock() pattern scattered about. Or >> maybe I'm missing the distinction. > > Heh, I didn't do s/bootmem/memblock/g, so foo_bootmem() are reminders we > had bootmem allocator once. > Maybe it's a good time to remove them :) > >> <trim> >