On Tue, Jun 30, 2020 at 07:09:31PM -0700, Andrew Morton wrote: > On Tue, 30 Jun 2020 12:08:25 -0700 Roman Gushchin <guro@xxxxxx> wrote: > > > On Sun, Jun 28, 2020 at 07:43:45PM +1200, Barry Song wrote: > > > Calling cma_declare_contiguous_nid() with false exact_nid for per-numa > > > reservation can easily cause cma leak and various confusion. > > > For example, mm/hugetlb.c is trying to reserve per-numa cma for gigantic > > > pages. But it can easily leak cma and make users confused when system has > > > memoryless nodes. > > > > > > In case the system has 4 numa nodes, and only numa node0 has memory. > > > if we set hugetlb_cma=4G in bootargs, mm/hugetlb.c will get 4 cma areas > > > for 4 different numa nodes. since exact_nid=false in current code, all > > > 4 numa nodes will get cma successfully from node0, but hugetlb_cma[1 to 3] > > > will never be available to hugepage will only allocate memory from > > > hugetlb_cma[0]. > > > > > > In case the system has 4 numa nodes, both numa node0&2 has memory, other > > > nodes have no memory. > > > if we set hugetlb_cma=4G in bootargs, mm/hugetlb.c will get 4 cma areas > > > for 4 different numa nodes. since exact_nid=false in current code, all > > > 4 numa nodes will get cma successfully from node0 or 2, but hugetlb_cma[1] > > > and [3] will never be available to hugepage as mm/hugetlb.c will only > > > allocate memory from hugetlb_cma[0] and hugetlb_cma[2]. > > > This causes permanent leak of the cma areas which are supposed to be > > > used by memoryless node. > > > > > > Of cource we can workaround the issue by letting mm/hugetlb.c scan all > > > cma areas in alloc_gigantic_page() even node_mask includes node0 only. > > > that means when node_mask includes node0 only, we can get page from > > > hugetlb_cma[1] to hugetlb_cma[3]. But this will cause kernel crash in > > > free_gigantic_page() while it wants to free page by: > > > cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order) > > > > > > On the other hand, exact_nid=false won't consider numa distance, it > > > might be not that useful to leverage cma areas on remote nodes. > > > I feel it is much simpler to make exact_nid true to make everything > > > clear. After that, memoryless nodes won't be able to reserve per-numa > > > CMA from other nodes which have memory. > > > > Totally agree. > > > > Acked-by: Roman Gushchin <guro@xxxxxx> > > > > Do we feel this merits a cc:stable? It would be nice. Thanks!
