On 2024/5/13 23:34, Yang Shi wrote: > On Fri, May 10, 2024 at 9:31 PM Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote: >> >> When I did memory failure tests recently, below panic occurs: >> >> kernel BUG at include/linux/mm.h:1135! >> invalid opcode: 0000 [#1] PREEMPT SMP NOPTI >> CPU: 9 PID: 137 Comm: kswapd1 Not tainted 6.9.0-rc4-00491-gd5ce28f156fe-dirty #14 >> RIP: 0010:shrink_huge_zero_page_scan+0x168/0x1a0 >> RSP: 0018:ffff9933c6c57bd0 EFLAGS: 00000246 >> RAX: 000000000000003e RBX: 0000000000000000 RCX: ffff88f61fc5c9c8 >> RDX: 0000000000000000 RSI: 0000000000000027 RDI: ffff88f61fc5c9c0 >> RBP: ffffcd7c446b0000 R08: ffffffff9a9405f0 R09: 0000000000005492 >> R10: 00000000000030ea R11: ffffffff9a9405f0 R12: 0000000000000000 >> R13: 0000000000000000 R14: 0000000000000000 R15: ffff88e703c4ac00 >> FS: 0000000000000000(0000) GS:ffff88f61fc40000(0000) knlGS:0000000000000000 >> CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 >> CR2: 000055f4da6e9878 CR3: 0000000c71048000 CR4: 00000000000006f0 >> Call Trace: >> <TASK> >> do_shrink_slab+0x14f/0x6a0 >> shrink_slab+0xca/0x8c0 >> shrink_node+0x2d0/0x7d0 >> balance_pgdat+0x33a/0x720 >> kswapd+0x1f3/0x410 >> kthread+0xd5/0x100 >> ret_from_fork+0x2f/0x50 >> ret_from_fork_asm+0x1a/0x30 >> </TASK> >> Modules linked in: mce_inject hwpoison_inject >> ---[ end trace 0000000000000000 ]--- >> RIP: 0010:shrink_huge_zero_page_scan+0x168/0x1a0 >> RSP: 0018:ffff9933c6c57bd0 EFLAGS: 00000246 >> RAX: 000000000000003e RBX: 0000000000000000 RCX: ffff88f61fc5c9c8 >> RDX: 0000000000000000 RSI: 0000000000000027 RDI: ffff88f61fc5c9c0 >> RBP: ffffcd7c446b0000 R08: ffffffff9a9405f0 R09: 0000000000005492 >> R10: 00000000000030ea R11: ffffffff9a9405f0 R12: 0000000000000000 >> R13: 0000000000000000 R14: 0000000000000000 R15: ffff88e703c4ac00 >> FS: 0000000000000000(0000) GS:ffff88f61fc40000(0000) knlGS:0000000000000000 >> CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 >> CR2: 000055f4da6e9878 CR3: 0000000c71048000 CR4: 00000000000006f0 >> >> The root cause is that HWPoison flag will be set for huge_zero_folio >> without increasing the folio refcnt. But then unpoison_memory() will >> decrease the folio refcnt unexpectly as it appears like a successfully >> hwpoisoned folio leading to VM_BUG_ON_PAGE(page_ref_count(page) == 0) >> when releasing huge_zero_folio. >> >> Fix this issue by marking huge_zero_folio reserved. So unpoison_memory() >> will skip this page. This will make it consistent with ZERO_PAGE case too. > > If I read the code correctly, unpoison_memory() should not dec > refcount for huge zero page by calling put_page_testzero(). The huge > zero page's real refcount is actually maintained separately by > huge_zero_refcount. It is different from the regular refount in struct > folio, see get_huge_zero_page(). Sure. Huge zero folio should be skipped in unpoison_memory(). It's not supported anyway. I marked huge_zero_folio reserved in order to let unpoison_memory() skip it by folio_test_reserved(folio) check. But as David points out, the use of PG_reserve is limited, so I will find another way to fix the issue. Thanks. .
