Re: [PATCH v2 3/4] mm/mmap/vma_merge: extend invariants, avoid invalid res, vma

[Marek Szyprowski]

On 21.03.2023 21:45, Lorenzo Stoakes wrote:
> Previously, vma was an uninitialised variable which was only definitely
> assigned as a result of the logic covering all possible input cases - for
> it to have remained uninitialised, prev would have to be NULL, and next
> would _have_ to be mergeable.
>
> We now reuse vma to assign curr and next, so to be absolutely explicit,
> ensure this variable is _always_ assigned, and while we're at it remove the
> redundant assignment of both res and vma (if prev is NULL then we simply
> assign to NULL).
>
> In addition, we absolutely do rely on addr == curr->vm_start should curr
> exist, so assert as much.
>
> Signed-off-by: Lorenzo Stoakes <lstoakes@xxxxxxxxx>

This patch has been merged into today's linux next-20230323 as commit 
6426bbcc76be ("mm/mmap/vma_merge: extend invariants, avoid invalid res, 
vma").

Unfortunately it breaks booting of some ARM 32bit machines, like Samsung 
Exynos5422-based Odroid XU3 board. This shortened log shows the issue:

Run /sbin/init as init process
   with arguments:
     /sbin/init
   with environment:
     HOME=/
     TERM=linux
8<--- cut here ---
Unhandled fault: page domain fault (0x01b) at 0xb6f03010
[b6f03010] *pgd=b5e84835
Internal error: : 1b [#1] PREEMPT SMP ARM
Modules linked in:
CPU: 2 PID: 1 Comm: init Not tainted 6.3.0-rc1-00296-g6426bbcc76be #6445
Hardware name: Samsung Exynos (Flattened Device Tree)
PC is at vma_merge+0xa0/0x728
LR is at vma_merge+0x294/0x728
pc : [<c02b08a8>]    lr : [<c02b0a9c>]    psr: a0000013
...
Flags: NzCv  IRQs on  FIQs on  Mode SVC_32  ISA ARM  Segment none
Control: 10c5387d  Table: 4f11406a  DAC: 00000051
...
Process init (pid: 1, stack limit = 0x5219a5c7)
Stack: (0xf0835e30 to 0xf0836000)
...
  vma_merge from mprotect_fixup+0xc8/0x290
  mprotect_fixup from do_mprotect_pkey.constprop.0+0x210/0x338
  do_mprotect_pkey.constprop.0 from ret_fast_syscall+0x0/0x1c
Exception stack(0xf0835fa8 to 0xf0835ff0)
...
---[ end trace 0000000000000000 ]---
note: init[1] exited with irqs disabled
Kernel panic - not syncing: Attempted to kill init! exitcode=0x0000000b

Reverting it on top of linux-next, together with 183b2bced4c9 
("mm/mmap/vma_merge: init cleanup, be explicit about the non-mergeable 
case"), which depends on this patch, fixes the boot issue.


> ---
>   mm/mmap.c | 19 ++++++++++++++-----
>   1 file changed, 14 insertions(+), 5 deletions(-)
>
> diff --git a/mm/mmap.c b/mm/mmap.c
> index 6361baf75601..7aec49c3bc74 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -911,7 +911,7 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>   {
>   	pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
>   	pgoff_t vma_pgoff;
> -	struct vm_area_struct *curr, *next, *res = NULL;
> +	struct vm_area_struct *curr, *next, *res;
>   	struct vm_area_struct *vma, *adjust, *remove, *remove2;
>   	int err = -1;
>   	bool merge_prev = false;
> @@ -940,14 +940,18 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>   		/* Is there a VMA next to a hole (case 1 - 3) or prev (4)? */
>   		next = vma_lookup(mm, end);
>   
> -	/* verify some invariant that must be enforced by the caller */
> +	/*
> +	 * By default, we return prev. Cases 3, 4, 8 will instead return next
> +	 * and cases 3, 8 will also update vma to point at next.
> +	 */
> +	res = vma = prev;
> +
> +	/* Verify some invariant that must be enforced by the caller. */
>   	VM_WARN_ON(prev && addr <= prev->vm_start);
> -	VM_WARN_ON(curr && end > curr->vm_end);
> +	VM_WARN_ON(curr && (addr != curr->vm_start || end > curr->vm_end));
>   	VM_WARN_ON(addr >= end);
>   
>   	if (prev) {
> -		res = prev;
> -		vma = prev;
>   		vma_start = prev->vm_start;
>   		vma_pgoff = prev->vm_pgoff;
>   		/* Can we merge the predecessor? */
> @@ -958,6 +962,7 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>   			vma_prev(vmi);
>   		}
>   	}
> +
>   	/* Can we merge the successor? */
>   	if (next && mpol_equal(policy, vma_policy(next)) &&
>   	    can_vma_merge_before(next, vm_flags,
> @@ -998,6 +1003,10 @@ struct vm_area_struct *vma_merge(struct vma_iterator *vmi, struct mm_struct *mm,
>   			adj_start = -(prev->vm_end - addr);
>   			err = dup_anon_vma(next, prev);
>   		} else {
> +			/*
> +			 * Note that cases 3 and 8 are the ONLY ones where prev
> +			 * is permitted to be (but is not necessarily) NULL.
> +			 */
>   			vma = next;			/* case 3 */
>   			vma_start = addr;
>   			vma_end = next->vm_end;

Best regards
-- 
Marek Szyprowski, PhD
Samsung R&D Institute Poland