On Fri, Dec 11, 2020 at 5:35 PM Oscar Salvador <osalvador@xxxxxxx> wrote: > > On Thu, Dec 10, 2020 at 11:55:20AM +0800, Muchun Song wrote: > > When we free a HugeTLB page to the buddy allocator, we should allocate the > > vmemmap pages associated with it. We can do that in the __free_hugepage() > "vmemmap pages that describe the range" would look better to me, but it is ok. Thanks. > > > +#define GFP_VMEMMAP_PAGE \ > > + (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_HIGH | __GFP_NOWARN) > > > > #ifndef VMEMMAP_HPAGE_SHIFT > > #define VMEMMAP_HPAGE_SHIFT HPAGE_SHIFT > > @@ -197,6 +200,11 @@ > > (__boundary - 1 < (end) - 1) ? __boundary : (end); \ > > }) > > > > +typedef void (*vmemmap_remap_pte_func_t)(struct page *reuse, pte_t *pte, > > + unsigned long start, unsigned long end, > > + void *priv); > > Any reason to not have defined GFP_VMEMMAP_PAGE and the new typedef into > hugetlb_vmemmap.h? Because they can only be used in this hugetlb_vmemmap.c. > > > > +static void vmemmap_restore_pte_range(struct page *reuse, pte_t *pte, > > + unsigned long start, unsigned long end, > > + void *priv) > > +{ > > + pgprot_t pgprot = PAGE_KERNEL; > > + void *from = page_to_virt(reuse); > > + unsigned long addr; > > + struct list_head *pages = priv; > [...] > > + > > + /* > > + * Make sure that any data that writes to the @to is made > > + * visible to the physical page. > > + */ > > + flush_kernel_vmap_range(to, PAGE_SIZE); > > Correct me if I am wrong, but flush_kernel_vmap_range is a NOOP under arches which > do not have ARCH_HAS_FLUSH_KERNEL_DCACHE_PAGE. > Since we only enable support for x86_64, and x86_64 is one of those arches, > could we remove this, and introduced later on in case we enable this feature > on an arch that needs it? OK. Will remove. > > I am not sure if you need to flush the range somehow, as you did in > vmemmap_remap_range. > > > +retry: > > + page = alloc_page(GFP_VMEMMAP_PAGE); > > + if (unlikely(!page)) { > > + msleep(100); > > + /* > > + * We should retry infinitely, because we cannot > > + * handle allocation failures. Once we allocate > > + * vmemmap pages successfully, then we can free > > + * a HugeTLB page. > > + */ > > + goto retry; > > I think this is the trickiest part. > With 2MB HugeTLB pages we only need 6 pages, but with 1GB, the number of pages > we need to allocate increases significantly (4088 pages IIRC). > And you are using __GFP_HIGH, which will allow us to use more memory (by > cutting down the watermark), but it might lead to putting the system > on its knees wrt. memory. > And yes, I know that once we allocate the 4088 pages, 1GB gets freed, but > still. Yeah, it is a problem. How about removing __GFP_HIGH only for 1GB HugeTLB page? > > I would like to hear Michal's thoughts on this one, but I wonder if it makes > sense to not let 1GB-HugeTLB pages be freed. > > -- > Oscar Salvador > SUSE L3 -- Yours, Muchun