Re: [PATCH v1] mm: recheck PageCompound() before printing out info of head page

From: John Hubbard
Date: Thu Apr 14 2022 - 21:08:22 EST


On 4/14/22 16:59, Naoya Horiguchi wrote:
From: Naoya Horiguchi <naoya.horiguchi@xxxxxxx>

The following VM_BUG_ON_FOLIO() is triggered when memory error event
happens on the (thp/folio) pages which are about to be freed:

[ 1160.232771] page:00000000b36a8a0f refcount:1 mapcount:0 mapping:0000000000000000 index:0x1 pfn:0x16a000
[ 1160.236916] page:00000000b36a8a0f refcount:0 mapcount:0 mapping:0000000000000000 index:0x1 pfn:0x16a000
[ 1160.240684] flags: 0x57ffffc0800000(hwpoison|node=1|zone=2|lastcpupid=0x1fffff)
[ 1160.243458] raw: 0057ffffc0800000 dead000000000100 dead000000000122 0000000000000000
[ 1160.246268] raw: 0000000000000001 0000000000000000 00000000ffffffff 0000000000000000
[ 1160.249197] page dumped because: VM_BUG_ON_FOLIO(!folio_test_large(folio))
[ 1160.251815] ------------[ cut here ]------------
[ 1160.253438] kernel BUG at include/linux/mm.h:788!
[ 1160.256162] invalid opcode: 0000 [#1] PREEMPT SMP PTI
[ 1160.258172] CPU: 2 PID: 115368 Comm: mceinj.sh Tainted: G E 5.18.0-rc1-v5.18-rc1-220404-2353-005-g83111+ #3
[ 1160.262049] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.15.0-1.fc35 04/01/2014
[ 1160.265103] RIP: 0010:dump_page.cold+0x27e/0x2bd
[ 1160.266757] Code: fe ff ff 48 c7 c6 81 f1 5a 98 e9 4c fe ff ff 48 c7 c6 a1 95 59 98 e9 40 fe ff ff 48 c7 c6 50 bf 5a 98 48 89 ef e8 9d 04 6d ff <0f> 0b 41 f7 c4 ff 0f 00 00 0f 85 9f fd ff ff 49 8b 04 24 a9 00 00
[ 1160.273180] RSP: 0018:ffffaa2c4d59fd18 EFLAGS: 00010292
[ 1160.274969] RAX: 000000000000003e RBX: 0000000000000001 RCX: 0000000000000000
[ 1160.277263] RDX: 0000000000000001 RSI: ffffffff985995a1 RDI: 00000000ffffffff
[ 1160.279571] RBP: ffffdc9c45a80000 R08: 0000000000000000 R09: 00000000ffffdfff
[ 1160.281794] R10: ffffaa2c4d59fb08 R11: ffffffff98940d08 R12: ffffdc9c45a80000
[ 1160.283920] R13: ffffffff985b6f94 R14: 0000000000000000 R15: ffffdc9c45a80000
[ 1160.286641] FS: 00007eff54ce1740(0000) GS:ffff99c67bd00000(0000) knlGS:0000000000000000
[ 1160.289498] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 1160.291106] CR2: 00005628381a5f68 CR3: 0000000104712003 CR4: 0000000000170ee0
[ 1160.293031] Call Trace:
[ 1160.293724] <TASK>
[ 1160.294334] get_hwpoison_page+0x47d/0x570
[ 1160.295474] memory_failure+0x106/0xaa0
[ 1160.296474] ? security_capable+0x36/0x50
[ 1160.297524] hard_offline_page_store+0x43/0x80
[ 1160.298684] kernfs_fop_write_iter+0x11c/0x1b0
[ 1160.299829] new_sync_write+0xf9/0x160
[ 1160.300810] vfs_write+0x209/0x290
[ 1160.301835] ksys_write+0x4f/0xc0
[ 1160.302718] do_syscall_64+0x3b/0x90
[ 1160.303664] entry_SYSCALL_64_after_hwframe+0x44/0xae
[ 1160.304981] RIP: 0033:0x7eff54b018b7

As shown in the RIP address, this VM_BUG_ON in folio_entire_mapcount() is
called from dump_page("hwpoison: unhandlable page") in get_any_page().
The below explains the mechanism of the race:

CPU 0 CPU 1

memory_failure
get_hwpoison_page
get_any_page
dump_page
compound = PageCompound
free_pages_prepare
page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP
folio_entire_mapcount
VM_BUG_ON_FOLIO(!folio_test_large(folio))

So in order to avoid the BUG_ON, recheck PageCompound() before __dump_page()
prints out information about the head page. Precisely saying, this does
not close up the race, but mitigates the impact to just inaccurate info,
which should be acceptable.

Actually, I don't think it really closes the race even to that extent,
because a machine check (MCE, on x86) could come in at any point, so
rechecking (without locks) won't really completely fix such a bug, right?


Fixes: 74e8ee4708a8 ("mm: Turn head_compound_mapcount() into folio_entire_mapcount()")
Signed-off-by: Naoya Horiguchi <naoya.horiguchi@xxxxxxx>
---
I'm actually not sure how widely the race condition could affect.
If we can consider this as a bug specific to memory_failure(), simply
replacing the dump_page() with simple printk() might be a proper fix

I think we should have separate, limited, special case functions that
are safe for use by exotic things like the MCE handler. And so just
doing a printk there seems better to me.

(I confirmed that that fix also works). But if any other dump_page()'s
caller could race with freeing thp/folio, the more generic fix
(I'm suggesting in this patch) would be desirable.

*If* there are other races, they should be addressed separately, IMHO.
Because MCE should be a special case.

---
mm/debug.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/debug.c b/mm/debug.c
index bef329bf28f0..fce63b200bea 100644
--- a/mm/debug.c
+++ b/mm/debug.c
@@ -93,7 +93,7 @@ static void __dump_page(struct page *page)
pr_warn("page:%p refcount:%d mapcount:%d mapping:%p index:%#lx pfn:%#lx\n",
page, page_ref_count(head), mapcount, mapping,
page_to_pgoff(page), page_to_pfn(page));
- if (compound) {
+ if (compound && PageCompound(page)) {

If people decide to keep this approach, then I'd go with something
that makes it clear, like:

/* Re-check, in a semi-doomed attempt to avoid racing with MCE: */
compound = PageCompound(page);

:)

thanks,
--
John Hubbard
NVIDIA

pr_warn("head:%p order:%u compound_mapcount:%d compound_pincount:%d\n",
head, compound_order(head),
folio_entire_mapcount(folio),