(2013/11/06 11:21), Atsushi Kumagai wrote: > (2013/11/06 5:27), Vivek Goyal wrote: >> On Tue, Nov 05, 2013 at 09:45:32PM +0800, Jingbai Ma wrote: >>> This patch set intend to exclude unnecessary hugepages from vmcore dump file. >>> >>> This patch requires the kernel patch to export necessary data structures into >>> vmcore: "kexec: export hugepage data structure into vmcoreinfo" >>> http://lists.infradead.org/pipermail/kexec/2013-November/009997.html >>> >>> This patch introduce two new dump levels 32 and 64 to exclude all unused and >>> active hugepages. The level to exclude all unnecessary pages will be 127 now. >> >> Interesting. Why hugepages should be treated any differentely than normal >> pages? >> >> If user asked to filter out free page, then it should be filtered and >> it should not matter whether it is a huge page or not? > > I'm making a RFC patch of hugepages filtering based on such policy. > > I attach the prototype version. > It's able to filter out also THPs, and suitable for cyclic processing > because it depends on mem_map and looking up it can be divided into > cycles. This is the same idea as page_is_buddy(). > > So I think it's better. > > @@ -4506,14 +4583,49 @@ __exclude_unnecessary_pages(unsigned long mem_map, > && !isAnon(mapping)) { > if (clear_bit_on_2nd_bitmap_for_kernel(pfn)) > pfn_cache_private++; > + /* > + * NOTE: If THP for cache is introduced, the check for > + * compound pages is needed here. > + */ > } > /* > * Exclude the data page of the user process. > */ > - else if ((info->dump_level & DL_EXCLUDE_USER_DATA) > - && isAnon(mapping)) { > - if (clear_bit_on_2nd_bitmap_for_kernel(pfn)) > - pfn_user++; > + else if (info->dump_level & DL_EXCLUDE_USER_DATA) { > + /* > + * Exclude the anonnymous pages as user pages. > + */ > + if (isAnon(mapping)) { > + if (clear_bit_on_2nd_bitmap_for_kernel(pfn)) > + pfn_user++; > + > + /* > + * Check the compound page > + */ > + if (page_is_hugepage(flags) && compound_order > 0) { > + int i, nr_pages = 1 << compound_order; > + > + for (i = 1; i < nr_pages; ++i) { > + if (clear_bit_on_2nd_bitmap_for_kernel(pfn + i)) > + pfn_user++; > + } > + pfn += nr_pages - 2; > + mem_map += (nr_pages - 1) * SIZE(page); > + } > + } > + /* > + * Exclude the hugetlbfs pages as user pages. > + */ > + else if (hugetlb_dtor == SYMBOL(free_huge_page)) { > + int i, nr_pages = 1 << compound_order; > + > + for (i = 0; i < nr_pages; ++i) { > + if (clear_bit_on_2nd_bitmap_for_kernel(pfn + i)) > + pfn_user++; > + } > + pfn += nr_pages - 1; > + mem_map += (nr_pages - 1) * SIZE(page); > + } > } > /* > * Exclude the hwpoison page. I'm concerned about the case that filtering is not performed to part of mem_map entries not belonging to the current cyclic range. If maximum value of compound_order is larger than maximum value of CONFIG_FORCE_MAX_ZONEORDER, which makedumpfile obtains by ARRAY_LENGTH(zone.free_area), it's necessary to align info->bufsize_cyclic with larger one in check_cyclic_buffer_overrun(). -- Thanks. HATAYAMA, Daisuke
