On Sun, Jul 26, 2020 at 05:49:04PM +0100, Matthew Wilcox wrote: > On Fri, Jul 24, 2020 at 02:13:11PM +0300, Kirill A. Shutemov wrote: > > On Thu, Jul 23, 2020 at 03:37:44PM +0800, Hillf Danton wrote: > > > > > > On Tue, 21 Jul 2020 14:11:31 +0300 Kirill A. Shutemov wrote: > > > > On Mon, Jul 20, 2020 at 04:51:44PM -0700, Andrew Morton wrote: > > > > > On Sun, 19 Jul 2020 14:10:19 -0700 syzbot wrote: > > > > > > > > > > > syzbot has found a reproducer for the following issue on: > > > > > > > > > > > > HEAD commit: 4c43049f Add linux-next specific files for 20200716 > > > > > > git tree: linux-next > > > > > > console output: https://syzkaller.appspot.com/x/log.txt?x=12c56087100000 > > > > > > kernel config: https://syzkaller.appspot.com/x/.config?x=2c76d72659687242 > > > > > > dashboard link: https://syzkaller.appspot.com/bug?extid=c48f34012b06c4ac67dd > > > > > > compiler: gcc (GCC) 10.1.0-syz 20200507 > > > > > > syz repro: https://syzkaller.appspot.com/x/repro.syz?x=1344abeb100000 > > > > > > > > > > > > IMPORTANT: if you fix the issue, please add the following tag to the commit: > > > > > > Reported-by: syzbot+c48f34012b06c4ac67dd@xxxxxxxxxxxxxxxxxxxxxxxxx > > > > > > > > > > Thanks. > > > > > > > > > > __handle_mm_fault > > > > > ->pmd_migration_entry_wait > > > > > ->migration_entry_to_page > > > > > > > > > > stumbled onto an unlocked page. > > > > > > > > > > I don't immediately see a cause. Perhaps Matthew's "THP prep patches", > > > > > perhaps something else. > > > > > > > > > > Is it possible to perform a bisection? > > > > > > > > Maybe it's related to the new lock_page_async()? > > > > > > Or is there likely the window that after copy_huge_pmd() the src pmd migrate > > > entry is removed and the page unlocked but the dst is not? > > > > No. > > > > copy_huge_pmd() runs with exclusive mmap_lock on the source side and > > destination side is not running yet. > > The one I'm hitting is huge related though. > > I added this debug: > > +++ b/include/linux/swapops.h > @@ -165,8 +165,9 @@ static inline struct page *device_private_entry_to_page(swp_entry_t entry) > #ifdef CONFIG_MIGRATION > static inline swp_entry_t make_migration_entry(struct page *page, int write) > { > - BUG_ON(!PageLocked(compound_head(page))); > + VM_BUG_ON_PAGE(!PageLocked(page), page); > > +if (PageCompound(page)) printk("pfn %lx order %d\n", page_to_pfn(page), thp_order(thp_head(page))); > return swp_entry(write ? SWP_MIGRATION_WRITE : SWP_MIGRATION_READ, > page_to_pfn(page)); > } > @@ -194,7 +195,11 @@ static inline struct page *migration_entry_to_page(swp_entry_t entry) > * Any use of migration entries may only occur while the > * corresponding page is locked > */ > - BUG_ON(!PageLocked(compound_head(p))); > + if (!PageLocked(p)) { > + dump_page(p, "not locked"); > + printk("swap entry %d.%lx\n", swp_type(entry), swp_offset(entry)); > + BUG(); > + } > return p; > } > > > and got useful output (while running generic/086): > > 1457 086 (20181): drop_caches: 3 > 1457 page:00000000a216ae9a refcount:2 mapcount:0 mapping:000000009ba7bfed index:0x2227 pfn:0x229e7 > 1457 aops:def_blk_aops ino:0 > 1457 flags: 0x4000000000002030(lru|active|private) > 1457 raw: 4000000000002030 fffff5b4416b5a48 fffff5b4408a7988 ffff9e9c34848578 > 1457 raw: 0000000000002227 ffff9e9bd18f0d00 00000002ffffffff 0000000000000000 > 1457 page dumped because: not locked > 1457 swap entry 30.229e7 > 1457 ------------[ cut here ]------------ > 1457 kernel BUG at include/linux/swapops.h:201! > 1457 invalid opcode: 0000 [#1] SMP PTI > 1457 CPU: 3 PID: 646 Comm: check Kdump: loaded Tainted: G W 5.8.0-rc6-00067-gd8b18bdf9870-dirty #355 > 1457 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1 04/01/2014 > 1457 RIP: 0010:__migration_entry_wait+0x109/0x110 > [...] > > Looking back in the trace, I see: > > ... > 1457 pfn 229e5 order 9 > 1457 pfn 229e6 order 9 > 1457 pfn 229e7 order 9 > 1457 pfn 229e8 order 9 > 1457 pfn 229e9 order 9 > ... > > so I would say we have a refcount problem. I've probably made it worse by > creating more THPs, but I don't think I'm the originator of the problem. > > I know very little about the migration code today. I suspect I'm going > to have to learn about it next week. It would be interesting to know if the migration entires ever got removed for pfn. I mean if remove_migration_pte() got called for it. It can be rmap issue too. Maybe it misses PMD on remove_migration_ptes() or something. -- Kirill A. Shutemov
