Re: [RESEND PATCH v1] mm/vmalloc: fix page mapping if vm_area_alloc_pages() with high order fallback to order 0

[Michal Hocko <mhocko@suse.com>]

On Fri 16-08-24 19:46:26, Hailong Liu wrote:
> On Fri, 16. Aug 12:13, Uladzislau Rezki wrote:
> > On Fri, Aug 16, 2024 at 05:12:32PM +0800, Hailong Liu wrote:
> > > On Thu, 15. Aug 22:07, Andrew Morton wrote:
> > > > On Fri, 9 Aug 2024 11:41:42 +0200 Uladzislau Rezki <urezki@gmail.com> wrote:
> > > >
> > > > > > > Acked-by: Barry Song <baohua@kernel.org>
> > > > > > >
> > > > > > > because we already have a fallback here:
> > > > > > >
> > > > > > > void *__vmalloc_node_range_noprof :
> > > > > > >
> > > > > > > fail:
> > > > > > >         if (shift > PAGE_SHIFT) {
> > > > > > >                 shift = PAGE_SHIFT;
> > > > > > >                 align = real_align;
> > > > > > >                 size = real_size;
> > > > > > >                 goto again;
> > > > > > >         }
> > > > > >
> > > > > > This really deserves a comment because this is not really clear at all.
> > > > > > The code is also fragile and it would benefit from some re-org.
> > > > > >
> > > > > > Thanks for the fix.
> > > > > >
> > > > > > Acked-by: Michal Hocko <mhocko@suse.com>
> > > > > >
> > > > > I agree. This is only clear for people who know the code. A "fallback"
> > > > > to order-0 should be commented.
> > > >
> > > > It's been a week.  Could someone please propose a fixup patch to add
> > > > this comment?
> > >
> > > Hi Andrew:
> > >
> > > Do you mean that I need to send a v2 patch with the the comments included?
> > >
> > It is better to post v2.
> Got it.
> 
> >
> > But before, could you please comment on:
> >
> > in case of order-0, bulk path may easily fail and fallback to the single
> > page allocator. If an request is marked as NO_FAIL, i am talking about
> > order-0 request, your change breaks GFP_NOFAIL for !order.
> >
> > Am i missing something obvious?
> For order-0, alloc_pages(GFP_X | __GFP_NOFAIL, 0), buddy allocator will handle
> the flag correctly. IMO we don't need to handle the flag here.

Let me clarify what I would like to have clarified:

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 6b783baf12a1..fea90a39f5c5 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -3510,13 +3510,13 @@ void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
 EXPORT_SYMBOL_GPL(vmap_pfn);
 #endif /* CONFIG_VMAP_PFN */
 
+/* GFP_NOFAIL semantic is implemented by __vmalloc_node_range_noprof */
 static inline unsigned int
 vm_area_alloc_pages(gfp_t gfp, int nid,
 		unsigned int order, unsigned int nr_pages, struct page **pages)
 {
 	unsigned int nr_allocated = 0;
-	gfp_t alloc_gfp = gfp;
-	bool nofail = gfp & __GFP_NOFAIL;
+	gfp_t alloc_gfp = gfp & ~ __GFP_NOFAIL;
 	struct page *page;
 	int i;
 
@@ -3527,9 +3527,6 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
 	 * more permissive.
 	 */
 	if (!order) {
-		/* bulk allocator doesn't support nofail req. officially */
-		gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
-
 		while (nr_allocated < nr_pages) {
 			unsigned int nr, nr_pages_request;
 
@@ -3547,12 +3544,12 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
 			 * but mempolicy wants to alloc memory by interleaving.
 			 */
 			if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
-				nr = alloc_pages_bulk_array_mempolicy_noprof(bulk_gfp,
+				nr = alloc_pages_bulk_array_mempolicy_noprof(alloc_gfp,
 							nr_pages_request,
 							pages + nr_allocated);
 
 			else
-				nr = alloc_pages_bulk_array_node_noprof(bulk_gfp, nid,
+				nr = alloc_pages_bulk_array_node_noprof(alloc_gfp, nid,
 							nr_pages_request,
 							pages + nr_allocated);
 
@@ -3566,13 +3563,6 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
 			if (nr != nr_pages_request)
 				break;
 		}
-	} else if (gfp & __GFP_NOFAIL) {
-		/*
-		 * Higher order nofail allocations are really expensive and
-		 * potentially dangerous (pre-mature OOM, disruptive reclaim
-		 * and compaction etc.
-		 */
-		alloc_gfp &= ~__GFP_NOFAIL;
 	}
 
 	/* High-order pages or fallback path if "bulk" fails. */
-- 
Michal Hocko
SUSE Labs