diff mbox

dax: improve fix for colliding PMD & PTE entries

Message ID 20170526195932.32178-1-ross.zwisler@linux.intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

Ross Zwisler May 26, 2017, 7:59 p.m. UTC
This commit, which has not yet made it upstream but is in the -mm tree:

    dax: Fix race between colliding PMD & PTE entries

fixed a pair of race conditions where racing DAX PTE and PMD faults could
corrupt page tables.  This fix had two shortcomings which are addressed by
this patch:

1) In the PTE fault handler we only checked for a collision using
pmd_devmap().  The pmd_devmap() check will trigger when we have raced with
a PMD that has real DAX storage, but to account for the case where we
collide with a huge zero page entry we also need to check for
pmd_trans_huge().

2) In the PMD fault handler we only continued with the fault if no PMD at
all was present (pmd_none()).  This is the case when we are faulting in a
PMD for the first time, but there are two other cases to consider.  The
first is that we are servicing a write fault over a PMD huge zero page,
which we detect with pmd_trans_huge().  The second is that we are servicing
a write fault over a DAX PMD with real storage, which we address with
pmd_devmap().

Fix both of these, and instead of manually triggering a fallback in the PMD
collision case instead be consistent with the other collision detection
code in the fault handlers and just retry.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: stable@vger.kernel.org
---

For both the -mm tree and for stable, feel free to squash this with the
original commit if you think that is appropriate.

This has passed targeted testing and an xfstests run.
---
 fs/dax.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

Comments

Jan Kara May 29, 2017, 12:17 p.m. UTC | #1
On Fri 26-05-17 13:59:32, Ross Zwisler wrote:
> This commit, which has not yet made it upstream but is in the -mm tree:
> 
>     dax: Fix race between colliding PMD & PTE entries
> 
> fixed a pair of race conditions where racing DAX PTE and PMD faults could
> corrupt page tables.  This fix had two shortcomings which are addressed by
> this patch:
> 
> 1) In the PTE fault handler we only checked for a collision using
> pmd_devmap().  The pmd_devmap() check will trigger when we have raced with
> a PMD that has real DAX storage, but to account for the case where we
> collide with a huge zero page entry we also need to check for
> pmd_trans_huge().
> 
> 2) In the PMD fault handler we only continued with the fault if no PMD at
> all was present (pmd_none()).  This is the case when we are faulting in a
> PMD for the first time, but there are two other cases to consider.  The
> first is that we are servicing a write fault over a PMD huge zero page,
> which we detect with pmd_trans_huge().  The second is that we are servicing
> a write fault over a DAX PMD with real storage, which we address with
> pmd_devmap().
> 
> Fix both of these, and instead of manually triggering a fallback in the PMD
> collision case instead be consistent with the other collision detection
> code in the fault handlers and just retry.
> 
> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
> Cc: stable@vger.kernel.org

Ugh, right. I forgot zero page in page tables will not pass the devmap
check... You can add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza

> ---
> 
> For both the -mm tree and for stable, feel free to squash this with the
> original commit if you think that is appropriate.
> 
> This has passed targeted testing and an xfstests run.
> ---
>  fs/dax.c | 11 +++++++----
>  1 file changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/dax.c b/fs/dax.c
> index fc62f36..2a6889b 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -1160,7 +1160,7 @@ static int dax_iomap_pte_fault(struct vm_fault *vmf,
>  	 * the PTE we need to set up.  If so just return and the fault will be
>  	 * retried.
>  	 */
> -	if (pmd_devmap(*vmf->pmd)) {
> +	if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
>  		vmf_ret = VM_FAULT_NOPAGE;
>  		goto unlock_entry;
>  	}
> @@ -1411,11 +1411,14 @@ static int dax_iomap_pmd_fault(struct vm_fault *vmf,
>  	/*
>  	 * It is possible, particularly with mixed reads & writes to private
>  	 * mappings, that we have raced with a PTE fault that overlaps with
> -	 * the PMD we need to set up.  If so we just fall back to a PTE fault
> -	 * ourselves.
> +	 * the PMD we need to set up.  If so just return and the fault will be
> +	 * retried.
>  	 */
> -	if (!pmd_none(*vmf->pmd))
> +	if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
> +			!pmd_devmap(*vmf->pmd)) {
> +		result = 0;
>  		goto unlock_entry;
> +	}
>  
>  	/*
>  	 * Note that we don't use iomap_apply here.  We aren't doing I/O, only
> -- 
> 2.9.4
>
diff mbox

Patch

diff --git a/fs/dax.c b/fs/dax.c
index fc62f36..2a6889b 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -1160,7 +1160,7 @@  static int dax_iomap_pte_fault(struct vm_fault *vmf,
 	 * the PTE we need to set up.  If so just return and the fault will be
 	 * retried.
 	 */
-	if (pmd_devmap(*vmf->pmd)) {
+	if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
 		vmf_ret = VM_FAULT_NOPAGE;
 		goto unlock_entry;
 	}
@@ -1411,11 +1411,14 @@  static int dax_iomap_pmd_fault(struct vm_fault *vmf,
 	/*
 	 * It is possible, particularly with mixed reads & writes to private
 	 * mappings, that we have raced with a PTE fault that overlaps with
-	 * the PMD we need to set up.  If so we just fall back to a PTE fault
-	 * ourselves.
+	 * the PMD we need to set up.  If so just return and the fault will be
+	 * retried.
 	 */
-	if (!pmd_none(*vmf->pmd))
+	if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
+			!pmd_devmap(*vmf->pmd)) {
+		result = 0;
 		goto unlock_entry;
+	}
 
 	/*
 	 * Note that we don't use iomap_apply here.  We aren't doing I/O, only