[v7,04/24] mm: Dont assume page-table invariance during faults

From: Peter Zijlstra <peterz@infradead.org>

One of the side effects of speculating on faults (without holding
mmap_sem) is that we can race with free_pgtables() and therefore we
cannot assume the page-tables will stick around.

Remove the reliance on the pte pointer.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>

In most of the case pte_unmap_same() was returning 1, which meaning that
do_swap_page() should do its processing. So in most of the case there will
be no impact.

Now regarding the case where pte_unmap_safe() was returning 0, and thus
do_swap_page return 0 too, this happens when the page has already been
swapped back. This may happen before do_swap_page() get called or while in
the call to do_swap_page(). In that later case, the check done when
swapin_readahead() returns will detect that case.

The worst case would be that a page fault is occuring on 2 threads at the
same time on the same swapped out page. In that case one thread will take
much time looping in __read_swap_cache_async(). But in the regular page
fault path, this is even worse since the thread would wait for semaphore to
be released before starting anything.

[Remove only if !CONFIG_SPECULATIVE_PAGE_FAULT]
Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
---
 mm/memory.c | 4 ++++
 1 file changed, 4 insertions(+)

On Tue, Feb 06, 2018 at 05:49:50PM +0100, Laurent Dufour wrote:
> From: Peter Zijlstra <peterz@infradead.org>
> 
> One of the side effects of speculating on faults (without holding
> mmap_sem) is that we can race with free_pgtables() and therefore we
> cannot assume the page-tables will stick around.
> 
> Remove the reliance on the pte pointer.
> 
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> 
> In most of the case pte_unmap_same() was returning 1, which meaning that
> do_swap_page() should do its processing. So in most of the case there will
> be no impact.
> 
> Now regarding the case where pte_unmap_safe() was returning 0, and thus
> do_swap_page return 0 too, this happens when the page has already been
> swapped back. This may happen before do_swap_page() get called or while in
> the call to do_swap_page(). In that later case, the check done when
> swapin_readahead() returns will detect that case.
> 
> The worst case would be that a page fault is occuring on 2 threads at the
> same time on the same swapped out page. In that case one thread will take
> much time looping in __read_swap_cache_async(). But in the regular page
> fault path, this is even worse since the thread would wait for semaphore to
> be released before starting anything.
> 
> [Remove only if !CONFIG_SPECULATIVE_PAGE_FAULT]
> Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>

I have a great deal of trouble connecting all of the words above to the
contents of the patch.

>  
> +#ifndef CONFIG_SPECULATIVE_PAGE_FAULT
>  /*
>   * handle_pte_fault chooses page fault handler according to an entry which was
>   * read non-atomically.  Before making any commitment, on those architectures
> @@ -2311,6 +2312,7 @@ static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
>  	pte_unmap(page_table);
>  	return same;
>  }
> +#endif /* CONFIG_SPECULATIVE_PAGE_FAULT */
>  
>  static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
>  {
> @@ -2898,11 +2900,13 @@ int do_swap_page(struct vm_fault *vmf)
>  		swapcache = page;
>  	}
>  
> +#ifndef CONFIG_SPECULATIVE_PAGE_FAULT
>  	if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
>  		if (page)
>  			put_page(page);
>  		goto out;
>  	}
> +#endif
>  

This feels to me like we want:

#ifdef CONFIG_SPECULATIVE_PAGE_FAULT
[current code]
#else
/*
 * Some words here which explains why we always want to return this
 * value if we support speculative page faults.
 */
static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
				pte_t *page_table, pte_t orig_pte)
{
	return 1;
}
#endif

instead of cluttering do_swap_page with an ifdef.

On 06/02/2018 21:28, Matthew Wilcox wrote:
> On Tue, Feb 06, 2018 at 05:49:50PM +0100, Laurent Dufour wrote:
>> From: Peter Zijlstra <peterz@infradead.org>
>>
>> One of the side effects of speculating on faults (without holding
>> mmap_sem) is that we can race with free_pgtables() and therefore we
>> cannot assume the page-tables will stick around.
>>
>> Remove the reliance on the pte pointer.
>>
>> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
>>
>> In most of the case pte_unmap_same() was returning 1, which meaning that
>> do_swap_page() should do its processing. So in most of the case there will
>> be no impact.
>>
>> Now regarding the case where pte_unmap_safe() was returning 0, and thus
>> do_swap_page return 0 too, this happens when the page has already been
>> swapped back. This may happen before do_swap_page() get called or while in
>> the call to do_swap_page(). In that later case, the check done when
>> swapin_readahead() returns will detect that case.
>>
>> The worst case would be that a page fault is occuring on 2 threads at the
>> same time on the same swapped out page. In that case one thread will take
>> much time looping in __read_swap_cache_async(). But in the regular page
>> fault path, this is even worse since the thread would wait for semaphore to
>> be released before starting anything.
>>
>> [Remove only if !CONFIG_SPECULATIVE_PAGE_FAULT]
>> Signed-off-by: Laurent Dufour <ldufour@linux.vnet.ibm.com>
> 
> I have a great deal of trouble connecting all of the words above to the
> contents of the patch.

Thanks for pushing forward here, this raised some doubts on my side.

I reviewed that part of code, and I think I could now change the way
pte_unmap_safe() is checking for the pte's value. Since we now have all the
needed details in the vm_fault structure, I will pass it to
pte_unamp_same() and deal with the VMA checks when locking for the pte as
it is done in the other part of the page fault handler by calling
pte_spinlock().

This means that this patch will be dropped, and pte_unmap_same() will become :

static inline int pte_unmap_same(struct vm_fault *vmf, int *same)
{
	int ret = 0;

	*same = 1;
#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
	if (sizeof(pte_t) > sizeof(unsigned long)) {
		if (pte_spinlock(vmf)) {
			*same = pte_same(*vmf->pte, vmf->orig_pte);
			spin_unlock(vmf->ptl);
		}
		else
			ret = VM_FAULT_RETRY;
	}
#endif
	pte_unmap(vmf->pte);
	return ret;
}

Laurent.

> 
>>  
>> +#ifndef CONFIG_SPECULATIVE_PAGE_FAULT
>>  /*
>>   * handle_pte_fault chooses page fault handler according to an entry which was
>>   * read non-atomically.  Before making any commitment, on those architectures
>> @@ -2311,6 +2312,7 @@ static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
>>  	pte_unmap(page_table);
>>  	return same;
>>  }
>> +#endif /* CONFIG_SPECULATIVE_PAGE_FAULT */
>>  
>>  static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
>>  {
>> @@ -2898,11 +2900,13 @@ int do_swap_page(struct vm_fault *vmf)
>>  		swapcache = page;
>>  	}
>>  
>> +#ifndef CONFIG_SPECULATIVE_PAGE_FAULT
>>  	if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
>>  		if (page)
>>  			put_page(page);
>>  		goto out;
>>  	}
>> +#endif
>>  
> 
> This feels to me like we want:
> 
> #ifdef CONFIG_SPECULATIVE_PAGE_FAULT
> [current code]
> #else
> /*
>  * Some words here which explains why we always want to return this
>  * value if we support speculative page faults.
>  */
> static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
> 				pte_t *page_table, pte_t orig_pte)
> {
> 	return 1;
> }
> #endif
> 
> instead of cluttering do_swap_page with an ifdef.
>

On Thu, Feb 08, 2018 at 03:35:58PM +0100, Laurent Dufour wrote:
> I reviewed that part of code, and I think I could now change the way
> pte_unmap_safe() is checking for the pte's value. Since we now have all the
> needed details in the vm_fault structure, I will pass it to
> pte_unamp_same() and deal with the VMA checks when locking for the pte as
> it is done in the other part of the page fault handler by calling
> pte_spinlock().

This does indeed look much better!  Thank you!

> This means that this patch will be dropped, and pte_unmap_same() will become :
> 
> static inline int pte_unmap_same(struct vm_fault *vmf, int *same)
> {
> 	int ret = 0;
> 
> 	*same = 1;
> #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
> 	if (sizeof(pte_t) > sizeof(unsigned long)) {
> 		if (pte_spinlock(vmf)) {
> 			*same = pte_same(*vmf->pte, vmf->orig_pte);
> 			spin_unlock(vmf->ptl);
> 		}
> 		else
> 			ret = VM_FAULT_RETRY;
> 	}
> #endif
> 	pte_unmap(vmf->pte);
> 	return ret;
> }

I'm not a huge fan of auxiliary return values.  Perhaps we could do this
instead:

	ret = pte_unmap_same(vmf);
	if (ret != VM_FAULT_NOTSAME) {
		if (page)
			put_page(page);
		goto out;
	}
	ret = 0;

(we have a lot of unused bits in VM_FAULT_, so adding a new one shouldn't
be a big deal)

On 08/02/2018 16:00, Matthew Wilcox wrote:
> On Thu, Feb 08, 2018 at 03:35:58PM +0100, Laurent Dufour wrote:
>> I reviewed that part of code, and I think I could now change the way
>> pte_unmap_safe() is checking for the pte's value. Since we now have all the
>> needed details in the vm_fault structure, I will pass it to
>> pte_unamp_same() and deal with the VMA checks when locking for the pte as
>> it is done in the other part of the page fault handler by calling
>> pte_spinlock().
> 
> This does indeed look much better!  Thank you!
> 
>> This means that this patch will be dropped, and pte_unmap_same() will become :
>>
>> static inline int pte_unmap_same(struct vm_fault *vmf, int *same)
>> {
>> 	int ret = 0;
>>
>> 	*same = 1;
>> #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
>> 	if (sizeof(pte_t) > sizeof(unsigned long)) {
>> 		if (pte_spinlock(vmf)) {
>> 			*same = pte_same(*vmf->pte, vmf->orig_pte);
>> 			spin_unlock(vmf->ptl);
>> 		}
>> 		else
>> 			ret = VM_FAULT_RETRY;
>> 	}
>> #endif
>> 	pte_unmap(vmf->pte);
>> 	return ret;
>> }
> 
> I'm not a huge fan of auxiliary return values.  Perhaps we could do this
> instead:
> 
> 	ret = pte_unmap_same(vmf);
> 	if (ret != VM_FAULT_NOTSAME) {
> 		if (page)
> 			put_page(page);
> 		goto out;
> 	}
> 	ret = 0;
> 
> (we have a lot of unused bits in VM_FAULT_, so adding a new one shouldn't
> be a big deal)

I do agree, using an auxiliary return value is not a good idea.

What about the following changes based on your suggestion ?

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 7de4323b9e89..0cd31a37bb3d 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1212,6 +1212,7 @@ static inline void clear_page_pfmemalloc(struct page *page)
 #define VM_FAULT_NEEDDSYNC  0x2000     /* ->fault did not modify page tables
                                         * and needs fsync() to complete (for
                                         * synchronous page faults in DAX) */
+#define VM_FAULT_PTNOTSAME 0x4000      /* Page table entries have changed */
 
 #define VM_FAULT_ERROR (VM_FAULT_OOM | VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV | \
                         VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE | \
diff --git a/mm/memory.c b/mm/memory.c
index b7da99c74fef..c9b419f8e4c5 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -2433,21 +2433,30 @@ static inline bool pte_map_lock(struct vm_fault *vmf)
  * parts, do_swap_page must check under lock before unmapping the pte and
  * proceeding (but do_wp_page is only called after already making such a check;
  * and do_anonymous_page can safely check later on).
+ *
+ * pte_unmap_same() returns:
+ *     0                       if the PTE are the same
+ *     VM_FAULT_PTNOTSAME      if the PTE are different
+ *     VM_FAULT_RETRY          if the VMA has changed in our back during
+ *                             a speculative page fault handling.
  */
-static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
-                               pte_t *page_table, pte_t orig_pte)
+static inline int pte_unmap_same(struct vm_fault *vmf)
 {
-       int same = 1;
+       int ret = 0;
+
 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
        if (sizeof(pte_t) > sizeof(unsigned long)) {
-               spinlock_t *ptl = pte_lockptr(mm, pmd);
-               spin_lock(ptl);
-               same = pte_same(*page_table, orig_pte);
-               spin_unlock(ptl);
+               if (pte_spinlock(vmf)) {
+                       if (!pte_same(*vmf->pte, vmf->orig_pte))
+                               ret = VM_FAULT_PTNOTSAME;
+                       spin_unlock(vmf->ptl);
+               }
+               else
+                       ret = VM_FAULT_RETRY;
        }
 #endif
-       pte_unmap(page_table);
-       return same;
+       pte_unmap(vmf->pte);
+       return ret;
 }
 
 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
@@ -3037,7 +3046,7 @@ int do_swap_page(struct vm_fault *vmf)
        pte_t pte;
        int locked;
        int exclusive = 0;
-       int ret = 0;
+       int ret;
        bool vma_readahead = swap_use_vma_readahead();
 
        if (vma_readahead) {
@@ -3045,9 +3054,16 @@ int do_swap_page(struct vm_fault *vmf)
                swapcache = page;
        }
 
-       if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte)) {
+       ret = pte_unmap_same(vmf);
+       if (ret) {
                if (page)
                        put_page(page);
+               /*
+                * In the case the PTE are different, meaning that the
+                * page has already been processed by another CPU, we return 0.
+                */
+               if (ret == VM_FAULT_PTNOTSAME)
+                       ret = 0;
                goto out;
        }

Thanks,
Laurent.