Re: [PATCH v4 4/4] mm/page_table_check: check entries at pmd levels

[Wei Xu <weixugc@google.com>]

On Wed, Jan 26, 2022 at 10:36 AM Pasha Tatashin
<pasha.tatashin@soleen.com> wrote:
>
> syzbot detected a case where the page table counters were not properly
> updated.
>
> syzkaller login:  ------------[ cut here ]------------
> kernel BUG at mm/page_table_check.c:162!
> invalid opcode: 0000 [#1] PREEMPT SMP KASAN
> CPU: 0 PID: 3099 Comm: pasha Not tainted 5.16.0+ #48
> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIO4
> RIP: 0010:__page_table_check_zero+0x159/0x1a0
> Code: 7d 3a b2 ff 45 39 f5 74 2a e8 43 38 b2 ff 4d 85 e4 01
> RSP: 0018:ffff888010667418 EFLAGS: 00010293
> RAX: 0000000000000000 RBX: 0000000000000001 RCX: 0000000000
> RDX: ffff88800cea8680 RSI: ffffffff81becaf9 RDI: 0000000003
> RBP: ffff888010667450 R08: 0000000000000001 R09: 0000000000
> R10: ffffffff81becaab R11: 0000000000000001 R12: ffff888008
> R13: 0000000000000001 R14: 0000000000000200 R15: dffffc0000
> FS:  0000000000000000(0000) GS:ffff888035e00000(0000) knlG0
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 00007ffd875cad00 CR3: 00000000094ce000 CR4: 0000000000
> DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000
> DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000
> Call Trace:
>  <TASK>
>  free_pcp_prepare+0x3be/0xaa0
>  free_unref_page+0x1c/0x650
>  ? trace_hardirqs_on+0x6a/0x1d0
>  free_compound_page+0xec/0x130
>  free_transhuge_page+0x1be/0x260
>  __put_compound_page+0x90/0xd0
>  release_pages+0x54c/0x1060
>  ? filemap_remove_folio+0x161/0x210
>  ? lock_downgrade+0x720/0x720
>  ? __put_page+0x150/0x150
>  ? filemap_free_folio+0x164/0x350
>  __pagevec_release+0x7c/0x110
>  shmem_undo_range+0x85e/0x1250
> ...
>
> The repro involved having a huge page that is split due to uprobe event
> temporarily replacing one of the pages in the huge page. Later the huge
> page was combined again, but the counters were off, as the PTE level
> was not properly updated.
>
> Make sure that when PMD is cleared and prior to freeing the level the
> PTEs are updated.
>
> Fixes: df4e817b7108 ("mm: page table check")
>
> Signed-off-by: Pasha Tatashin <pasha.tatashin@soleen.com>
> ---
>  include/linux/page_table_check.h | 18 ++++++++++++++++++
>  mm/khugepaged.c                  |  3 +++
>  mm/page_table_check.c            | 21 +++++++++++++++++++++
>  3 files changed, 42 insertions(+)
>
> diff --git a/include/linux/page_table_check.h b/include/linux/page_table_check.h
> index 38cace1da7b6..e88bbe37727b 100644
> --- a/include/linux/page_table_check.h
> +++ b/include/linux/page_table_check.h
> @@ -26,6 +26,8 @@ void __page_table_check_pmd_set(struct mm_struct *mm, unsigned long addr,
>                                 pmd_t *pmdp, pmd_t pmd);
>  void __page_table_check_pud_set(struct mm_struct *mm, unsigned long addr,
>                                 pud_t *pudp, pud_t pud);
> +void __page_table_check_pmd_clear_full(struct mm_struct *mm, unsigned long addr,
> +                                      pmd_t pmd);
>
>  static inline void page_table_check_alloc(struct page *page, unsigned int order)
>  {
> @@ -100,6 +102,16 @@ static inline void page_table_check_pud_set(struct mm_struct *mm,
>         __page_table_check_pud_set(mm, addr, pudp, pud);
>  }
>
> +static inline void page_table_check_pmd_clear_full(struct mm_struct *mm,
> +                                                  unsigned long addr,
> +                                                  pmd_t pmd)
> +{
> +       if (static_branch_likely(&page_table_check_disabled))
> +               return;
> +
> +       __page_table_check_pmd_clear_full(mm, addr, pmd);
> +}
> +
>  #else
>
>  static inline void page_table_check_alloc(struct page *page, unsigned int order)
> @@ -143,5 +155,11 @@ static inline void page_table_check_pud_set(struct mm_struct *mm,
>  {
>  }
>
> +static inline void page_table_check_pmd_clear_full(struct mm_struct *mm,
> +                                                  unsigned long addr,
> +                                                  pmd_t pmd)
> +{
> +}
> +
>  #endif /* CONFIG_PAGE_TABLE_CHECK */
>  #endif /* __LINUX_PAGE_TABLE_CHECK_H */
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 30e59e4af272..d84977c6dc0d 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -16,6 +16,7 @@
>  #include <linux/hashtable.h>
>  #include <linux/userfaultfd_k.h>
>  #include <linux/page_idle.h>
> +#include <linux/page_table_check.h>
>  #include <linux/swapops.h>
>  #include <linux/shmem_fs.h>
>
> @@ -1422,10 +1423,12 @@ static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *v
>         spinlock_t *ptl;
>         pmd_t pmd;
>
> +       mmap_assert_write_locked(mm);
>         ptl = pmd_lock(vma->vm_mm, pmdp);
>         pmd = pmdp_collapse_flush(vma, addr, pmdp);
>         spin_unlock(ptl);
>         mm_dec_nr_ptes(mm);
> +       page_table_check_pmd_clear_full(mm, addr, pmd);

pmdp_collapse_flush() already calls page_table_check_pmd_clear() via
pmdp_huge_get_and_clean().  Both pmdp_table_check_pmd_clear() and
page_table_check_pmd_clear_full() can call
__page_table_check_pmd_clear(). If that happens, then the page table
check counters can be messed up.  Certainly, there is no bug here
because the pmd is not huge and __page_table_check_pmd_clear() should
be skipped in both calls. But it would be better to avoid this
unnecessary subtlety by renaming page_table_check_pmd_clear_full() to
page_table_check_clear_pte_range() and not calling
__page_table_check_pmd_clear() there.  To make the code even more
clear, __page_table_check_pmd_clear() can also be renamed as
__page_table_check_huge_pmd_clear() (similar for its callers).

>         pte_free(mm, pmd_pgtable(pmd));
>  }
>
> diff --git a/mm/page_table_check.c b/mm/page_table_check.c
> index c61d7ebe13b1..251f95a808b4 100644
> --- a/mm/page_table_check.c
> +++ b/mm/page_table_check.c
> @@ -247,3 +247,24 @@ void __page_table_check_pud_set(struct mm_struct *mm, unsigned long addr,
>         }
>  }
>  EXPORT_SYMBOL(__page_table_check_pud_set);
> +
> +void __page_table_check_pmd_clear_full(struct mm_struct *mm, unsigned long addr,
> +                                      pmd_t pmd)
> +{
> +       if (&init_mm == mm)
> +               return;
> +
> +       if (!pmd_bad(pmd) && !pmd_leaf(pmd)) {
> +               pte_t *ptep = pte_offset_map(&pmd, addr);
> +               unsigned long i;
> +
> +               pte_unmap(ptep);
> +               for (i = 0; i < PTRS_PER_PTE; i++) {
> +                       __page_table_check_pte_clear(mm, addr, *ptep);
> +                       addr += PAGE_SIZE;
> +                       ptep++;
> +               }
> +       } else {
> +               __page_table_check_pmd_clear(mm, addr, pmd);
> +       }
> +}
> --
> 2.35.0.rc0.227.g00780c9af4-goog
>