Re: [syzbot] [tmpfs?] kernel BUG in folio_flags

[Matthew Wilcox]

On Thu, Jan 12, 2023 at 02:25:39AM -0800, syzbot wrote:
> ------------[ cut here ]------------
> kernel BUG at include/linux/page-flags.h:317!
> invalid opcode: 0000 [#1] PREEMPT SMP KASAN
> CPU: 0 PID: 10782 Comm: syz-executor.1 Not tainted 6.2.0-rc3-next-20230112-syzkaller #0
> Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 10/26/2022
> RIP: 0010:folio_flags.constprop.0+0x122/0x150 include/linux/page-flags.h:317
> Code: 52 ff ff ff e8 ff f2 cd ff 48 8d 43 ff 49 39 c4 0f 84 40 ff ff ff e8 ed f2 cd ff 48 c7 c6 40 0b 57 8a 4c 89 e7 e8 2e e4 05 00 <0f> 0b e8 57 c2 1b 00 e9 fe fe ff ff e8 4d c2 1b 00 eb a9 4c 89 e7
> RSP: 0018:ffffc90005437b00 EFLAGS: 00010246
> RAX: 0000000000040000 RBX: 0000000000000001 RCX: ffffc90014e22000
> RDX: 0000000000040000 RSI: ffffffff81b3ca02 RDI: 0000000000000000
> RBP: 0000000000000004 R08: 0000000000000000 R09: 0000000000000000
> R10: 0000000000000001 R11: 0000000000000001 R12: ffffea0000b28100
> R13: 0000000000000000 R14: dffffc0000000000 R15: 1ffffd400016502c
> FS:  00007f1e422fd700(0000) GS:ffff8880b9800000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 00005642ce910130 CR3: 000000001cef4000 CR4: 00000000003506f0
> Call Trace:
>  <TASK>
>  folio_test_head include/linux/page-flags.h:780 [inline]
>  folio_test_large include/linux/page-flags.h:801 [inline]
>  folio_nr_pages include/linux/mm.h:1735 [inline]
>  compound_nr include/linux/mm.h:1747 [inline]
>  zero_user_segments.constprop.0+0x262/0x350 include/linux/highmem.h:288
>  shmem_write_end+0x684/0x780 mm/shmem.c:2600

Oh, good one.  shmem isn't actually doing anything wrong here, but
it'll be nice to fix it.  (net -10 lines)

What's happening is that shmem is deliberately calling
zero_user_segments() on a tail page.  zero_user_segments() calls
compound_nr() which used to return 1 when called on a tail page but now
trips this assert.  I didn't intend this change of behaviour, so
let's start by fixing that.

Andrew, this is a fix for "mm: Reimplement compound_nr()".  I'll send
the shmem change separately since it's not a bug fix.

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 57d702fc8677..78f751f3ba5b 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1741,10 +1741,22 @@ static inline long folio_nr_pages(struct folio *folio)
 #endif
 }
 
-/* Returns the number of pages in this potentially compound page. */
+/*
+ * compound_nr() returns the number of pages in this potentially compound
+ * page.  compound_nr() can be called on a tail page, and is defined to
+ * return 1 in that case.
+ */
 static inline unsigned long compound_nr(struct page *page)
 {
-	return folio_nr_pages((struct folio *)page);
+	struct folio *folio = (struct folio *)page;
+
+	if (!test_bit(PG_head, &folio->flags))
+		return 1;
+#ifdef CONFIG_64BIT
+	return folio->_folio_nr_pages;
+#else
+	return 1L << folio->_folio_order;
+#endif
 }
 
 /**
@@ -1753,8 +1765,7 @@ static inline unsigned long compound_nr(struct page *page)
  */
 static inline int thp_nr_pages(struct page *page)
 {
-	VM_BUG_ON_PGFLAGS(PageTail(page), page);
-	return compound_nr(page);
+	return folio_nr_pages((struct folio *)page);
 }
 
 /**