On Tue, Sep 15, 2020 at 05:22:22PM -0400, Aristeu Rozanski wrote: > Hi Oscar, Naoya, > > On Mon, Sep 14, 2020 at 12:15:54PM +0200, Oscar Salvador wrote: > > The important bit of this patchset is patch#1, which is a fix to take off > > HWPoison pages off a buddy freelist since it can lead us to having HWPoison > > pages back in the game without no one noticing it. > > So fix it (we did that already for soft_offline_page [1]). > > > > The other patches are clean-ups and not that important, so if anything, > > consider patch#1 for inclusion. > > > > [1] https://patchwork.kernel.org/cover/11704083/ > > I found something strange with your and Naoya's hwpoison rework. We have a > customer with a testcase that basically does: > > p1 = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); > p2 = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); > > madvise(p1, size, MADV_MERGEABLE); > madvise(p2, size, MADV_MERGEABLE); > > memset(p1, 'a', size); > memset(p2, 'a', size); > > madvise(p1, size, MADV_SOFT_OFFLINE); > > madvise(p1, size, MADV_UNMERGEABLE); > madvise(p2, size, MADV_UNMERGEABLE); > > > where size is about 200,000 pages. It works on a x86_64 box (with and without the > hwpoison rework). On ppc64 boxes (tested 3 different ones with at least 250GB memory) > it fails to take a page off the buddy list (page_handle_poison()/take_page_off_buddy()) > (madvise MADV_SOFT_OFFLINE returns -EBUSY). Without the hwpoison rework the test passes. I reproduced the similar -EBUSY with small average x86 VM, where it seems to me a race between page_take_off_buddy() and page allocation. Oscar's debug patch shows the following kernel messages: [ 627.357009] Soft offlining pfn 0x235018 at process virtual address 0x7fd112140000 [ 627.358747] __get_any_page: 0x235018 free buddy page [ 627.359875] page:00000000038b52c9 refcount:0 mapcount:-128 mapping:0000000000000000 index:0x1 pfn:0x235018 [ 627.362002] flags: 0x57ffe000000000() [ 627.362841] raw: 0057ffe000000000 fffff84648d12688 ffff955abffd1dd0 0000000000000000 [ 627.364555] raw: 0000000000000001 0000000000000000 00000000ffffff7f 0000000000000000 [ 627.366258] page dumped because: page_handle_poison [ 627.367357] page->mem_cgroup:ffff9559b6912000 [ 627.368342] page_handle_poison: hugepage_or_freepage failed\xb8n [ 627.368344] soft_offline_free_page: page_handle_poison -EBUSY [ 627.370901] page:00000000038b52c9 refcount:6 mapcount:3 mapping:000000001226bf89 index:0x2710 pfn:0x235018 [ 627.373048] aops:ext4_da_aops ino:c63f3 dentry name:"system.journal" [ 627.374526] flags: 0x57ffe00000201c(uptodate|dirty|lru|private) [ 627.375865] raw: 0057ffe00000201c fffff84648d300c8 ffff955ab8c3f020 ffff955aba5f4ee0 [ 627.377586] raw: 0000000000002710 ffff9559b811fc98 0000000500000002 ffff9559b6912000 [ 627.379308] page dumped because: soft_offline_free_page [ 627.380480] page->mem_cgroup:ffff9559b6912000 CPU 0 CPU 1 get_any_page // returns 0 (free buddy path) soft_offline_free_page the page is allocated page_handle_poison -> fail return -EBUSY I'm still not sure why this issue is invisible before rework patch, but setting migrate type to MIGRATE_ISOLATE during offlining could affect the behavior sensitively. Thanks, Naoya Horiguchi
