On 7/7/20 12:56 PM, Andrew Morton wrote: > On Tue, 7 Jul 2020 16:02:04 +1200 Barry Song <song.bao.hua@xxxxxxxxxxxxx> wrote: > >> hugetlb_cma[0] can be NULL due to various reasons, for example, node0 has >> no memory. so NULL hugetlb_cma[0] doesn't necessarily mean cma is not >> enabled. gigantic pages might have been reserved on other nodes. > > I'm trying to figure out whether this should be backported into 5.7.1, > but the changelog doesn't describe any known user-visible effects of > the bug. Are there any? Barry must have missed this email. He reported the issue so I was hoping he would reply. Based on the code changes, I believe the following could happen: - Someone uses 'hugetlb_cma=' kernel command line parameter to reserve CMA for gigantic pages. - The system topology is such that no memory is on node 0. Therefore, no CMA can be reserved for gigantic pages on node 0. CMA is reserved on other nodes. - The user also specifies a number of gigantic pages to pre-allocate on the command line with hugepagesz=<gigantic_page_size> hugepages=<N> - The routine which allocates gigantic pages from the bootmem allocator will not detect CMA has been reserved as there is no memory on node 0. Therefore, pages will be pre-allocated from bootmem allocator as well as reserved in CMA. This double allocation (bootmem and CMA) is the worst case scenario. Not sure if this is what Barry saw, and I suspect this would rarely happen. After writing this, I started to think that perhaps command line parsing should be changed. If hugetlb_cma= is specified, it makes no sense to pre-allocate gigantic pages. Therefore, the hugepages=<N> paramemter should be ignored and flagged with a warning if hugetlb_cma= is specified. This could be checked at parsing time and there would be no need for such a check in the allocation code (except for sanity cheching). Thoughts? I just cleaned up the parsing code and could make such a change quite easily. -- Mike Kravetz
