[PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Jan H. Schönherr]

The function follow_pte_pmd() can theoretically return after having
acquired a PMD lock, even when DAX was not compiled with
CONFIG_FS_DAX_PMD.

Release the PMD lock unconditionally.

Fixes: f729c8c9b24f ("dax: wrprotect pmd_t in dax_mapping_entry_mkclean")
Signed-off-by: Jan H. Schönherr <jschoenh@amazon.de>
---
 fs/dax.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/dax.c b/fs/dax.c
index 9598159..c2ebf10 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -636,8 +636,8 @@ static void dax_mapping_entry_mkclean(struct address_space *mapping,
 			pmd = pmd_mkclean(pmd);
 			set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 unlock_pmd:
-			spin_unlock(ptl);
 #endif
+			spin_unlock(ptl);
 		} else {
 			if (pfn != pte_pfn(*ptep))
 				goto unlock_pte;
-- 
2.9.3.1.gcba166c.dirty

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Matthew Wilcox]

On Thu, Jan 18, 2018 at 02:38:39PM +0100, Jan H. Sch�nherr wrote:
> The function follow_pte_pmd() can theoretically return after having
> acquired a PMD lock, even when DAX was not compiled with
> CONFIG_FS_DAX_PMD.

I don't think it can.  How would a PMD entry get into a DAX VMA if we
compiled the kernel without CONFIG_FS_DAX_PMD?

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Matthew Wilcox]

On Thu, Jan 18, 2018 at 06:07:39AM -0800, Matthew Wilcox wrote:
> On Thu, Jan 18, 2018 at 02:38:39PM +0100, Jan H. Sch�nherr wrote:
> > The function follow_pte_pmd() can theoretically return after having
> > acquired a PMD lock, even when DAX was not compiled with
> > CONFIG_FS_DAX_PMD.
> 
> I don't think it can.  How would a PMD entry get into a DAX VMA if we
> compiled the kernel without CONFIG_FS_DAX_PMD?

How about this patch instead?  Should shut up sparse nicely.

diff --git a/fs/dax.c b/fs/dax.c
index 78b72c48374e..fea1b64d111b 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -586,6 +586,12 @@ pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
 	return address;
 }
 
+#ifdef CONFIG_FS_DAX_PMD
+#define dax_pmd(pmdp)	pmdp
+#else
+#define dax_pmd(pmdp)	0
+#endif
+
 /* Walk all mappings of a given index of a file and writeprotect them */
 static void dax_mapping_entry_mkclean(struct address_space *mapping,
 				      pgoff_t index, unsigned long pfn)
@@ -621,8 +627,7 @@ static void dax_mapping_entry_mkclean(struct address_space *mapping,
 		 *
 		 * See Documentation/vm/mmu_notifier.txt
 		 */
-		if (pmdp) {
-#ifdef CONFIG_FS_DAX_PMD
+		if (dax_pmd(pmdp)) {
 			pmd_t pmd;
 
 			if (pfn != pmd_pfn(*pmdp))
@@ -638,7 +643,6 @@ static void dax_mapping_entry_mkclean(struct address_space *mapping,
 			set_pmd_at(vma->vm_mm, address, pmdp, pmd);
 unlock_pmd:
 			spin_unlock(ptl);
-#endif
 		} else {
 			if (pfn != pte_pfn(*ptep))
 				goto unlock_pte;

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Jan H. Schönherr]

On 01/18/2018 03:07 PM, Matthew Wilcox wrote:
> On Thu, Jan 18, 2018 at 02:38:39PM +0100, Jan H. Schönherr wrote:
>> The function follow_pte_pmd() can theoretically return after having
>> acquired a PMD lock, even when DAX was not compiled with
>> CONFIG_FS_DAX_PMD.
> 
> I don't think it can.  How would a PMD entry get into a DAX VMA if we
> compiled the kernel without CONFIG_FS_DAX_PMD?
> 

Maybe it can not in happy cases. But the PMD parts in follow_pte_pmd() are compiled in
unconditionally. So, if there's an issue elsewhere, and for some weird reason we get a PMD entry
in the page table, it would screw the lock balance.

I haven't run into an actual issue with this, it's just supposed to be defensive.

Regards
Jan

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Jan H. Schönherr]

On 01/18/2018 03:22 PM, Matthew Wilcox wrote:
> On Thu, Jan 18, 2018 at 06:07:39AM -0800, Matthew Wilcox wrote:
>> On Thu, Jan 18, 2018 at 02:38:39PM +0100, Jan H. Schönherr wrote:
>>> The function follow_pte_pmd() can theoretically return after having
>>> acquired a PMD lock, even when DAX was not compiled with
>>> CONFIG_FS_DAX_PMD.
>>
>> I don't think it can.  How would a PMD entry get into a DAX VMA if we
>> compiled the kernel without CONFIG_FS_DAX_PMD?
> 
> How about this patch instead?  Should shut up sparse nicely.

It would still skip the unlock, in case pmdp is !=NULL (and locked) after follow_pte_pmd().
So it wouldn't address, what I intended to address with the patch.

Regards
Jan

> 
> diff --git a/fs/dax.c b/fs/dax.c
> index 78b72c48374e..fea1b64d111b 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -586,6 +586,12 @@ pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
>  	return address;
>  }
>  
> +#ifdef CONFIG_FS_DAX_PMD
> +#define dax_pmd(pmdp)	pmdp
> +#else
> +#define dax_pmd(pmdp)	0
> +#endif
> +
>  /* Walk all mappings of a given index of a file and writeprotect them */
>  static void dax_mapping_entry_mkclean(struct address_space *mapping,
>  				      pgoff_t index, unsigned long pfn)
> @@ -621,8 +627,7 @@ static void dax_mapping_entry_mkclean(struct address_space *mapping,
>  		 *
>  		 * See Documentation/vm/mmu_notifier.txt
>  		 */
> -		if (pmdp) {
> -#ifdef CONFIG_FS_DAX_PMD
> +		if (dax_pmd(pmdp)) {
>  			pmd_t pmd;
>  
>  			if (pfn != pmd_pfn(*pmdp))
> @@ -638,7 +643,6 @@ static void dax_mapping_entry_mkclean(struct address_space *mapping,
>  			set_pmd_at(vma->vm_mm, address, pmdp, pmd);
>  unlock_pmd:
>  			spin_unlock(ptl);
> -#endif
>  		} else {
>  			if (pfn != pte_pfn(*ptep))
>  				goto unlock_pte;
> 

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Jan H. Schönherr]

On 01/18/2018 03:27 PM, Jan H. Schönherr wrote:
> It would still skip the unlock, in case pmdp is !=NULL (and locked) after follow_pte_pmd().
> So it wouldn't address, what I intended to address with the patch.

Small correction: we'd take the "else" branch, which would do some stuff it's not supposed to do
when follow_pte_pmd() returns with pmdp!=NULL. And we'd actually do a unlock, but thinking it's
a PTE, not a PMD.

So, defensive-wise, may point still stands, that we're not correctly handling an unexpectedly
returned PMD entry.

Regards
Jan

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Ross Zwisler]

On Thu, Jan 18, 2018 at 02:38:39PM +0100, Jan H. Sch�nherr wrote:
> The function follow_pte_pmd() can theoretically return after having
> acquired a PMD lock, even when DAX was not compiled with
> CONFIG_FS_DAX_PMD.
> 
> Release the PMD lock unconditionally.
> 
> Fixes: f729c8c9b24f ("dax: wrprotect pmd_t in dax_mapping_entry_mkclean")
> Signed-off-by: Jan H. Sch�nherr <jschoenh@amazon.de>
> ---
>  fs/dax.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/dax.c b/fs/dax.c
> index 9598159..c2ebf10 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -636,8 +636,8 @@ static void dax_mapping_entry_mkclean(struct address_space *mapping,
>  			pmd = pmd_mkclean(pmd);
>  			set_pmd_at(vma->vm_mm, address, pmdp, pmd);
>  unlock_pmd:
> -			spin_unlock(ptl);
>  #endif
> +			spin_unlock(ptl);
>  		} else {
>  			if (pfn != pte_pfn(*ptep))
>  				goto unlock_pte;

Sure, this seems fine to me.  This seems simple and correct - you're right
that we aren't taking the PTL on the PMD conditionally based on whether
CONFIG_DAX_PMD is defined, so it doesn't make sense to release it
conditionally.  I think if we ever hit this lock imbalance we're totally
insane anyway, but it the fix is correct and doesn't mess with our code flow.

You can add:
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Ross Zwisler]

On Thu, Jan 18, 2018 at 09:20:13AM -0700, Ross Zwisler wrote:
> On Thu, Jan 18, 2018 at 02:38:39PM +0100, Jan H. Sch�nherr wrote:
> > The function follow_pte_pmd() can theoretically return after having
> > acquired a PMD lock, even when DAX was not compiled with
> > CONFIG_FS_DAX_PMD.
> > 
> > Release the PMD lock unconditionally.
> > 
> > Fixes: f729c8c9b24f ("dax: wrprotect pmd_t in dax_mapping_entry_mkclean")
> > Signed-off-by: Jan H. Sch�nherr <jschoenh@amazon.de>
> > ---
> >  fs/dax.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/fs/dax.c b/fs/dax.c
> > index 9598159..c2ebf10 100644
> > --- a/fs/dax.c
> > +++ b/fs/dax.c
> > @@ -636,8 +636,8 @@ static void dax_mapping_entry_mkclean(struct address_space *mapping,
> >  			pmd = pmd_mkclean(pmd);
> >  			set_pmd_at(vma->vm_mm, address, pmdp, pmd);
> >  unlock_pmd:
> > -			spin_unlock(ptl);
> >  #endif
> > +			spin_unlock(ptl);
> >  		} else {
> >  			if (pfn != pte_pfn(*ptep))
> >  				goto unlock_pte;
> 
> Sure, this seems fine to me.  This seems simple and correct - you're right
> that we aren't taking the PTL on the PMD conditionally based on whether
> CONFIG_DAX_PMD is defined, so it doesn't make sense to release it
> conditionally.  I think if we ever hit this lock imbalance we're totally
> insane anyway, but it the fix is correct and doesn't mess with our code flow.
> 
> You can add:
> Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>

Ah, I just realized that this patch didn't CC Andrew, and he's the one that
usually takes our DAX patches.

Andrew, can you pick this up?  Here's the fsdevel patchwork:

https://patchwork.kernel.org/patch/10173255/

Thanks,
- Ross

Re: [PATCH] dax: Release PMD lock even when there is no PMD support in DAX

[Jan H. Schönherr]

On 01/25/2018 05:34 PM, Ross Zwisler wrote:
> Ah, I just realized that this patch didn't CC Andrew, and he's the one that
> usually takes our DAX patches.
> 
> Andrew, can you pick this up?  Here's the fsdevel patchwork:
> 
> https://patchwork.kernel.org/patch/10173255/

Thanks for that, I didn't know.

Let me know, if I should resend instead (in case it makes things easier).

Regards
Jan