On Tue, Jun 06, 2023 at 10:17:02AM +0200, Uladzislau Rezki wrote: > On Tue, Jun 06, 2023 at 09:13:24AM +0200, Vlastimil Babka wrote: > > > > On 6/5/23 22:11, Lorenzo Stoakes wrote: > > > In __vmalloc_area_node() we always warn_alloc() when an allocation > > > performed by vm_area_alloc_pages() fails unless it was due to a pending > > > fatal signal. > > > > > > However, huge page allocations instigated either by vmalloc_huge() or > > > __vmalloc_node_range() (or a caller that invokes this like kvmalloc() or > > > kvmalloc_node()) always falls back to order-0 allocations if the huge page > > > allocation fails. > > > > > > This renders the warning useless and noisy, especially as all callers > > > appear to be aware that this may fallback. This has already resulted in at > > > least one bug report from a user who was confused by this (see link). > > > > > > Therefore, simply update the code to only output this warning for order-0 > > > pages when no fatal signal is pending. > > > > > > Link: https://bugzilla.suse.com/show_bug.cgi?id=1211410 > > > Signed-off-by: Lorenzo Stoakes <lstoakes@xxxxxxxxx> > > > > I think there are more reports of same thing from the btrfs context, that > > appear to be a 6.3 regression > > > > https://bugzilla.kernel.org/show_bug.cgi?id=217466 > > Link: https://lore.kernel.org/all/efa04d56-cd7f-6620-bca7-1df89f49bf4b@xxxxxxxxx/ > > > I had a look at that report. The btrfs complains due to the > fact that a high-order page(1 << 9) can not be obtained. In the > vmalloc code we do not fall to 0-order allocator if there is > a request of getting a high-order. This isn't true, we _do_ fallback to order-0 (this is the basis of my patch), in __vmalloc_node_range():- /* Allocate physical pages and map them into vmalloc space. */ ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node); if (!ret) goto fail; ... fail: if (shift > PAGE_SHIFT) { shift = PAGE_SHIFT; align = real_align; size = real_size; goto again; } With the order being derived from shift, and __vmalloc_area_node() only being called from __vmalloc_node_range(). > > I provided a patch to fallback if a high-order. A reproducer, after > applying the patch, started to get oppses in another places. > > IMO, we should fallback even for high-order requests. Because it is > highly likely it can not be accomplished. > > Any thoughts? > > <snip> > diff --git a/mm/vmalloc.c b/mm/vmalloc.c > index 31ff782d368b..7a06452f7807 100644 > --- a/mm/vmalloc.c > +++ b/mm/vmalloc.c > @@ -2957,14 +2957,18 @@ vm_area_alloc_pages(gfp_t gfp, int nid, > page = alloc_pages(alloc_gfp, order); > else > page = alloc_pages_node(nid, alloc_gfp, order); > + > if (unlikely(!page)) { > - if (!nofail) > - break; > + if (nofail) > + alloc_gfp |= __GFP_NOFAIL; > > - /* fall back to the zero order allocations */ > - alloc_gfp |= __GFP_NOFAIL; > - order = 0; > - continue; > + /* Fall back to the zero order allocations. */ > + if (order || nofail) { > + order = 0; > + continue; > + } > + > + break; > } > > /* > <snip> > > > > -- > Uladzislau Rezki I saw that, it seems to be duplicating the same thing as the original fallback code is (which was originally designed to permit higher order non-__GFP_NOFAIL allocations before trying order-0 __GFP_NOFAIL). I don't think it is really useful to change this as it confuses that logic and duplicates something we already do. Honestly though moreover I think this whole area needs some refactoring.
