Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Jerome Glisse]

On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> > On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> [...]
> > > What would be the best mmu notifier interface to use where there are no
> > > start/end calls?
> > > Or, is the best solution to add the start/end calls as is done in later
> > > versions of the code?  If that is the suggestion, has there been any change
> > > in invalidate start/end semantics that we should take into account?
> > 
> > start/end would be the one to add, 4.4 seems broken in respect to THP
> > and mmu notification. Another solution is to fix user of mmu notifier,
> > they were only a handful back then. For instance properly adjust the
> > address to match first address covered by pmd or pud and passing down
> > correct page size to mmu_notifier_invalidate_page() would allow to fix
> > this easily.
> > 
> > This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> > with an invalid one (either poison, migration or swap) inside the
> > function. So anyone racing would synchronize on those special entry
> > hence why it is fine to delay mmu_notifier_invalidate_page() to after
> > dropping the page table lock.
> > 
> > Adding start/end might the solution with less code churn as you would
> > only need to change try_to_unmap_one().
> 
> What about dependencies? 369ea8242c0fb sounds like it needs work for all
> notifiers need to be updated as well.

This commit remove mmu_notifier_invalidate_page() hence why everything
need to be updated. But in 4.4 you can get away with just adding start/
end and keep around mmu_notifier_invalidate_page() to minimize disruption.

So the new semantic in 369ea8242c0fb is that all page table changes are
bracketed with mmu notifier start/end calls and invalidate_range right
after tlb flush. This simplify thing and make it more reliable for mmu
notifier users like IOMMU or ODP or GPUs drivers.


> Anyway, I am wondering why we haven't see any bugs coming from
> incomplete range invalidation. How would those exhibit?

Reading back the 4.4 code try_to_unmap() can only be call against a
huge tlb page and only when migrating one ie through migrate_pages()
So this highly limit the cases where issues would happen. I believe
no one use hugetlb fs has backing for guest memory so xen and kvm
would never face such case.

So what is left is ODP, i915, radeon, amd gpu, SGI and IOMMU drivers.

The IOMMU drivers are never use that way AFAICT in 4.4 their was no
drivers upstream for a PCIE device that would support ATS/PASID and
thus the notifier path would never be use. Back then the only device
was AMD APU AFAIK which were never really use with that features
due to lack of mature userspace to use this.

For i915,radeon and amd GPU we would never see this either as the
mmu notifier is only use for uptr GEM object which are only use
to upload texture with either anonymous vma or file back vma. I
never heard of an xorg server or ddx or mesa drivers which would
use hugetlb fs.

So the only ones that might have issues AFAICT are ODP and SGI.
I am unsure on how likely either can be use in conjunction with a
hugetlb fs. CCing maintainers for those so they could comment.


The symptoms would either be memory corruption ie RDMA or SGI would
write to the old huge page and not the new one. Or even harder to
spot use of stall/invalid data ie RDMA or SGI are reading from
the old huge page and instead of the new one.

Corruption to hugetlbfs page can likely go unoticed.

Cheers,
Jérôme

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Mike Kravetz]

On 08/29/2018 02:11 PM, Jerome Glisse wrote:
> On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
>> On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
>>> On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
>> [...]
>>>> What would be the best mmu notifier interface to use where there are no
>>>> start/end calls?
>>>> Or, is the best solution to add the start/end calls as is done in later
>>>> versions of the code?  If that is the suggestion, has there been any change
>>>> in invalidate start/end semantics that we should take into account?
>>>
>>> start/end would be the one to add, 4.4 seems broken in respect to THP
>>> and mmu notification. Another solution is to fix user of mmu notifier,
>>> they were only a handful back then. For instance properly adjust the
>>> address to match first address covered by pmd or pud and passing down
>>> correct page size to mmu_notifier_invalidate_page() would allow to fix
>>> this easily.
>>>
>>> This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
>>> with an invalid one (either poison, migration or swap) inside the
>>> function. So anyone racing would synchronize on those special entry
>>> hence why it is fine to delay mmu_notifier_invalidate_page() to after
>>> dropping the page table lock.
>>>
>>> Adding start/end might the solution with less code churn as you would
>>> only need to change try_to_unmap_one().
>>
>> What about dependencies? 369ea8242c0fb sounds like it needs work for all
>> notifiers need to be updated as well.
> 
> This commit remove mmu_notifier_invalidate_page() hence why everything
> need to be updated. But in 4.4 you can get away with just adding start/
> end and keep around mmu_notifier_invalidate_page() to minimize disruption.
> 
> So the new semantic in 369ea8242c0fb is that all page table changes are
> bracketed with mmu notifier start/end calls and invalidate_range right
> after tlb flush. This simplify thing and make it more reliable for mmu
> notifier users like IOMMU or ODP or GPUs drivers.

Here is what I came up with by adding the start/end calls to the 4.4 version
of try_to_unmap_one.  Note that this assumes/uses the new routine
adjust_range_if_pmd_sharing_possible to adjust the notifier/flush range if
huge pmd sharing is possible.  I changed the mmu_notifier_invalidate_page
to a mmu_notifier_invalidate_range, but am not sure if that needs to happen
earlier in the routine (like right after tlb flush as you said above).
Does this look reasonable?

diff --git a/mm/rmap.c b/mm/rmap.c
index b577fbb98d4b..7ba8bfeddb4b 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1302,11 +1302,30 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	pte_t pteval;
 	spinlock_t *ptl;
 	int ret = SWAP_AGAIN;
+	unsigned long start = address, end;
 	enum ttu_flags flags = (enum ttu_flags)arg;
 
 	/* munlock has nothing to gain from examining un-locked vmas */
 	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
-		goto out;
+		return ret;
+
+	/*
+	 * For THP, we have to assume the worse case ie pmd for invalidation.
+	 * For hugetlb, it could be much worse if we need to do pud
+	 * invalidation in the case of pmd sharing.
+	 *
+	 * Note that the page can not be free in this function as call of
+	 * try_to_unmap() must hold a reference on the page.
+	 */
+	end = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
+	if (PageHuge(page)) {
+		/*
+		 * If sharing is possible, start and end will be adjusted
+		 * accordingly.
+		 */
+		adjust_range_if_pmd_sharing_possible(vma, &start, &end);
+	}
+	mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
 
 	pte = page_check_address(page, mm, address, &ptl, 0);
 	if (!pte)
@@ -1334,6 +1353,29 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 		}
   	}
 
+	if (PageHuge(page) && huge_pmd_unshare(mm, &address, pte)) {
+		/*
+		 * huge_pmd_unshare unmapped an entire PMD page.  There is
+		 * no way of knowing exactly which PMDs may be cached for
+		 * this mm, so flush them all.  start/end were already
+		 * adjusted to cover this range.
+		 */
+		flush_cache_range(vma, start, end);
+		flush_tlb_range(vma, start, end);
+
+		/*
+		 * The ref count of the PMD page was dropped which is part
+		 * of the way map counting is done for shared PMDs.  When
+		 * there is no other sharing, huge_pmd_unshare returns false
+		 * and we will unmap the actual page and drop map count
+		 * to zero.
+		 *
+		 * Note that huge_pmd_unshare modified address and is likely
+		 * not what you would expect.
+		 */
+		goto out_unmap;
+	}
+
 	/* Nuke the page table entry. */
 	flush_cache_page(vma, address, page_to_pfn(page));
 	if (should_defer_flush(mm, flags)) {
@@ -1424,10 +1466,11 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	page_cache_release(page);
 
 out_unmap:
-	pte_unmap_unlock(pte, ptl);
 	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
-		mmu_notifier_invalidate_page(mm, address);
+		mmu_notifier_invalidate_range(mm, start, end);
+	pte_unmap_unlock(pte, ptl);
 out:
+	mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
 	return ret;
 }
 
-- 
Mike Kravetz

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Michal Hocko]

On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
> On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> > On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> > > On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> > [...]
> > > > What would be the best mmu notifier interface to use where there are no
> > > > start/end calls?
> > > > Or, is the best solution to add the start/end calls as is done in later
> > > > versions of the code?  If that is the suggestion, has there been any change
> > > > in invalidate start/end semantics that we should take into account?
> > > 
> > > start/end would be the one to add, 4.4 seems broken in respect to THP
> > > and mmu notification. Another solution is to fix user of mmu notifier,
> > > they were only a handful back then. For instance properly adjust the
> > > address to match first address covered by pmd or pud and passing down
> > > correct page size to mmu_notifier_invalidate_page() would allow to fix
> > > this easily.
> > > 
> > > This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> > > with an invalid one (either poison, migration or swap) inside the
> > > function. So anyone racing would synchronize on those special entry
> > > hence why it is fine to delay mmu_notifier_invalidate_page() to after
> > > dropping the page table lock.
> > > 
> > > Adding start/end might the solution with less code churn as you would
> > > only need to change try_to_unmap_one().
> > 
> > What about dependencies? 369ea8242c0fb sounds like it needs work for all
> > notifiers need to be updated as well.
> 
> This commit remove mmu_notifier_invalidate_page() hence why everything
> need to be updated. But in 4.4 you can get away with just adding start/
> end and keep around mmu_notifier_invalidate_page() to minimize disruption.

OK, this is really interesting. I was really worried to change the
semantic of the mmu notifiers in stable kernels because this is really
a hard to review change and high risk for anybody running those old
kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
into the range scope API then this sounds like the best way forward.

So just to make sure we are at the same page. Does this sounds goo for
stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
top. What do you think?


>From 70a2285b058073eeb2971b94b7e6c8067d2d161a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?J=C3=A9r=C3=B4me=20Glisse?= <jglisse@redhat.com>
Date: Thu, 31 Aug 2017 17:17:27 -0400
Subject: [PATCH] mm/rmap: update to new mmu_notifier semantic v2
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

commit 369ea8242c0fb5239b4ddf0dc568f694bd244de4 upstrea.

Please note that this patch differs from the mainline because we do not
really replace mmu_notifier_invalidate_page by mmu_notifier_invalidate_range
because that requires changes to most of existing mmu notifiers. We also
do not want to change the semantic of this API in old kernels. Anyway
Jerome has suggested that it should be sufficient to simply wrap
mmu_notifier_invalidate_page by *_invalidate_range_start()/end() to fix
invalidation of larger than pte mappings (e.g. THP/hugetlb pages during
migration). We need this change to handle large (hugetlb/THP) pages
migration properly.

Note that because we can not presume the pmd value or pte value we have
to assume the worst and unconditionaly report an invalidation as
happening.

Changed since v2:
  - try_to_unmap_one() only one call to mmu_notifier_invalidate_range()
  - compute end with PAGE_SIZE << compound_order(page)
  - fix PageHuge() case in try_to_unmap_one()

Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Bernhard Held <berny156@gmx.de>
Cc: Adam Borowski <kilobyte@angband.pl>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Wanpeng Li <kernellwp@gmail.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: axie <axie@amd.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Michal Hocko <mhocko@suse.com> # backport to 4.4
---
 mm/rmap.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/mm/rmap.c b/mm/rmap.c
index 1bceb49aa214..364d245e6411 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1324,6 +1324,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	pte_t pteval;
 	spinlock_t *ptl;
 	int ret = SWAP_AGAIN;
+	unsigned long start = address, end;
 	enum ttu_flags flags = (enum ttu_flags)arg;
 
 	/* munlock has nothing to gain from examining un-locked vmas */
@@ -1356,6 +1357,14 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 		}
   	}
 
+	/*
+	 * We have to assume the worse case ie pmd for invalidation. Note that
+	 * the page can not be free in this function as call of try_to_unmap()
+	 * must hold a reference on the page.
+	 */
+	end = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
+	mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
+
 	/* Nuke the page table entry. */
 	flush_cache_page(vma, address, page_to_pfn(page));
 	if (should_defer_flush(mm, flags)) {
@@ -1449,6 +1458,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	pte_unmap_unlock(pte, ptl);
 	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
 		mmu_notifier_invalidate_page(mm, address);
+	mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
 out:
 	return ret;
 }
-- 
2.18.0

-- 
Michal Hocko
SUSE Labs

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Jerome Glisse]

On Thu, Aug 30, 2018 at 12:56:16PM +0200, Michal Hocko wrote:
> On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
> > On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> > > On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> > > > On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> > > [...]
> > > > > What would be the best mmu notifier interface to use where there are no
> > > > > start/end calls?
> > > > > Or, is the best solution to add the start/end calls as is done in later
> > > > > versions of the code?  If that is the suggestion, has there been any change
> > > > > in invalidate start/end semantics that we should take into account?
> > > > 
> > > > start/end would be the one to add, 4.4 seems broken in respect to THP
> > > > and mmu notification. Another solution is to fix user of mmu notifier,
> > > > they were only a handful back then. For instance properly adjust the
> > > > address to match first address covered by pmd or pud and passing down
> > > > correct page size to mmu_notifier_invalidate_page() would allow to fix
> > > > this easily.
> > > > 
> > > > This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> > > > with an invalid one (either poison, migration or swap) inside the
> > > > function. So anyone racing would synchronize on those special entry
> > > > hence why it is fine to delay mmu_notifier_invalidate_page() to after
> > > > dropping the page table lock.
> > > > 
> > > > Adding start/end might the solution with less code churn as you would
> > > > only need to change try_to_unmap_one().
> > > 
> > > What about dependencies? 369ea8242c0fb sounds like it needs work for all
> > > notifiers need to be updated as well.
> > 
> > This commit remove mmu_notifier_invalidate_page() hence why everything
> > need to be updated. But in 4.4 you can get away with just adding start/
> > end and keep around mmu_notifier_invalidate_page() to minimize disruption.
> 
> OK, this is really interesting. I was really worried to change the
> semantic of the mmu notifiers in stable kernels because this is really
> a hard to review change and high risk for anybody running those old
> kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
> into the range scope API then this sounds like the best way forward.
> 
> So just to make sure we are at the same page. Does this sounds goo for
> stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
> top. What do you think?

You need to invalidate outside page table lock so before the call to
page_check_address(). For instance like below patch, which also only
do the range invalidation for huge page which would avoid too much of
a behavior change for user of mmu notifier.

>From 1be4109cfbf1c475ad67a5a57c87c74fd183ab1d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?J=C3=A9r=C3=B4me=20Glisse?= <jglisse@redhat.com>
Date: Thu, 31 Aug 2017 17:17:27 -0400
Subject: [PATCH] mm/rmap: update to new mmu_notifier semantic v2
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

commit 369ea8242c0fb5239b4ddf0dc568f694bd244de4 upstrea.

Please note that this patch differs from the mainline because we do not
really replace mmu_notifier_invalidate_page by mmu_notifier_invalidate_range
because that requires changes to most of existing mmu notifiers. We also
do not want to change the semantic of this API in old kernels. Anyway
Jérôme has suggested that it should be sufficient to simply wrap
mmu_notifier_invalidate_page by *_invalidate_range_start()/end() to fix
invalidation of larger than pte mappings (e.g. THP/hugetlb pages during
migration). We need this change to handle large (hugetlb/THP) pages
migration properly.

Note that because we can not presume the pmd value or pte value we have
to assume the worst and unconditionaly report an invalidation as
happening.

Changed since v2:
  - try_to_unmap_one() only one call to mmu_notifier_invalidate_range()
  - compute end with PAGE_SIZE << compound_order(page)
  - fix PageHuge() case in try_to_unmap_one()

Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Bernhard Held <berny156@gmx.de>
Cc: Adam Borowski <kilobyte@angband.pl>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Wanpeng Li <kernellwp@gmail.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: axie <axie@amd.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Michal Hocko <mhocko@suse.com> # backport to 4.4
---
 mm/rmap.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)

diff --git a/mm/rmap.c b/mm/rmap.c
index b577fbb98d4b..a77f15dc0cf1 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1302,15 +1302,30 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	pte_t pteval;
 	spinlock_t *ptl;
 	int ret = SWAP_AGAIN;
+	unsigned long start = address, end;
 	enum ttu_flags flags = (enum ttu_flags)arg;
 
 	/* munlock has nothing to gain from examining un-locked vmas */
 	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
 		goto out;
 
+	if (unlikely(PageHuge(page))) {
+		/*
+		 * We have to assume the worse case ie pmd for invalidation.
+		 * Note that the page can not be free in this function as call
+		 * of try_to_unmap() must hold a reference on the page.
+		 *
+		 * This is ok to invalidate even if are not unmapping anything
+		 * ie below page_check_address() returning NULL.
+		 */
+		end = min(vma->vm_end, start + (PAGE_SIZE <<
+						compound_order(page)));
+		mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
+	}
+
 	pte = page_check_address(page, mm, address, &ptl, 0);
 	if (!pte)
-		goto out;
+		goto out_notify;
 
 	/*
 	 * If the page is mlock()d, we cannot swap it out.
@@ -1427,6 +1442,9 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	pte_unmap_unlock(pte, ptl);
 	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
 		mmu_notifier_invalidate_page(mm, address);
+out_notify:
+	if (unlikely(PageHuge(page)))
+		mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
 out:
 	return ret;
 }
-- 
2.17.1

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Michal Hocko]

On Thu 30-08-18 10:08:25, Jerome Glisse wrote:
> On Thu, Aug 30, 2018 at 12:56:16PM +0200, Michal Hocko wrote:
> > On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
> > > On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> > > > On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> > > > > On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> > > > [...]
> > > > > > What would be the best mmu notifier interface to use where there are no
> > > > > > start/end calls?
> > > > > > Or, is the best solution to add the start/end calls as is done in later
> > > > > > versions of the code?  If that is the suggestion, has there been any change
> > > > > > in invalidate start/end semantics that we should take into account?
> > > > > 
> > > > > start/end would be the one to add, 4.4 seems broken in respect to THP
> > > > > and mmu notification. Another solution is to fix user of mmu notifier,
> > > > > they were only a handful back then. For instance properly adjust the
> > > > > address to match first address covered by pmd or pud and passing down
> > > > > correct page size to mmu_notifier_invalidate_page() would allow to fix
> > > > > this easily.
> > > > > 
> > > > > This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> > > > > with an invalid one (either poison, migration or swap) inside the
> > > > > function. So anyone racing would synchronize on those special entry
> > > > > hence why it is fine to delay mmu_notifier_invalidate_page() to after
> > > > > dropping the page table lock.
> > > > > 
> > > > > Adding start/end might the solution with less code churn as you would
> > > > > only need to change try_to_unmap_one().
> > > > 
> > > > What about dependencies? 369ea8242c0fb sounds like it needs work for all
> > > > notifiers need to be updated as well.
> > > 
> > > This commit remove mmu_notifier_invalidate_page() hence why everything
> > > need to be updated. But in 4.4 you can get away with just adding start/
> > > end and keep around mmu_notifier_invalidate_page() to minimize disruption.
> > 
> > OK, this is really interesting. I was really worried to change the
> > semantic of the mmu notifiers in stable kernels because this is really
> > a hard to review change and high risk for anybody running those old
> > kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
> > into the range scope API then this sounds like the best way forward.
> > 
> > So just to make sure we are at the same page. Does this sounds goo for
> > stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
> > top. What do you think?
> 
> You need to invalidate outside page table lock so before the call to
> page_check_address(). For instance like below patch, which also only
> do the range invalidation for huge page which would avoid too much of
> a behavior change for user of mmu notifier.

Right. I would rather not make this PageHuge special though. So the
fixed version should be.

>From c05849f6789ec36e2ff11adcd8fa6cfb05e870a9 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?J=C3=A9r=C3=B4me=20Glisse?= <jglisse@redhat.com>
Date: Thu, 31 Aug 2017 17:17:27 -0400
Subject: [PATCH] mm/rmap: update to new mmu_notifier semantic v2
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

commit 369ea8242c0fb5239b4ddf0dc568f694bd244de4 upstrea.

Please note that this patch differs from the mainline because we do not
really replace mmu_notifier_invalidate_page by mmu_notifier_invalidate_range
because that requires changes to most of existing mmu notifiers. We also
do not want to change the semantic of this API in old kernels. Anyway
Jerome has suggested that it should be sufficient to simply wrap
mmu_notifier_invalidate_page by *_invalidate_range_start()/end() to fix
invalidation of larger than pte mappings (e.g. THP/hugetlb pages during
migration). We need this change to handle large (hugetlb/THP) pages
migration properly.

Note that because we can not presume the pmd value or pte value we have
to assume the worst and unconditionaly report an invalidation as
happening.

Changed since v2:
  - try_to_unmap_one() only one call to mmu_notifier_invalidate_range()
  - compute end with PAGE_SIZE << compound_order(page)
  - fix PageHuge() case in try_to_unmap_one()

Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Bernhard Held <berny156@gmx.de>
Cc: Adam Borowski <kilobyte@angband.pl>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Cc: Wanpeng Li <kernellwp@gmail.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Takashi Iwai <tiwai@suse.de>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: axie <axie@amd.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Michal Hocko <mhocko@suse.com> # backport to 4.4
---
 mm/rmap.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/mm/rmap.c b/mm/rmap.c
index 1bceb49aa214..aba994f55d6c 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1324,12 +1324,21 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	pte_t pteval;
 	spinlock_t *ptl;
 	int ret = SWAP_AGAIN;
+	unsigned long start = address, end;
 	enum ttu_flags flags = (enum ttu_flags)arg;
 
 	/* munlock has nothing to gain from examining un-locked vmas */
 	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
 		goto out;
 
+	/*
+	 * We have to assume the worse case ie pmd for invalidation. Note that
+	 * the page can not be free in this function as call of try_to_unmap()
+	 * must hold a reference on the page.
+	 */
+	end = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
+	mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
+
 	pte = page_check_address(page, mm, address, &ptl, 0);
 	if (!pte)
 		goto out;
@@ -1450,6 +1459,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
 	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
 		mmu_notifier_invalidate_page(mm, address);
 out:
+	mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
 	return ret;
 }
 
-- 
2.18.0

-- 
Michal Hocko
SUSE Labs

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Jerome Glisse]

On Thu, Aug 30, 2018 at 06:19:52PM +0200, Michal Hocko wrote:
> On Thu 30-08-18 10:08:25, Jerome Glisse wrote:
> > On Thu, Aug 30, 2018 at 12:56:16PM +0200, Michal Hocko wrote:
> > > On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
> > > > On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> > > > > On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> > > > > > On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> > > > > [...]
> > > > > > > What would be the best mmu notifier interface to use where there are no
> > > > > > > start/end calls?
> > > > > > > Or, is the best solution to add the start/end calls as is done in later
> > > > > > > versions of the code?  If that is the suggestion, has there been any change
> > > > > > > in invalidate start/end semantics that we should take into account?
> > > > > > 
> > > > > > start/end would be the one to add, 4.4 seems broken in respect to THP
> > > > > > and mmu notification. Another solution is to fix user of mmu notifier,
> > > > > > they were only a handful back then. For instance properly adjust the
> > > > > > address to match first address covered by pmd or pud and passing down
> > > > > > correct page size to mmu_notifier_invalidate_page() would allow to fix
> > > > > > this easily.
> > > > > > 
> > > > > > This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> > > > > > with an invalid one (either poison, migration or swap) inside the
> > > > > > function. So anyone racing would synchronize on those special entry
> > > > > > hence why it is fine to delay mmu_notifier_invalidate_page() to after
> > > > > > dropping the page table lock.
> > > > > > 
> > > > > > Adding start/end might the solution with less code churn as you would
> > > > > > only need to change try_to_unmap_one().
> > > > > 
> > > > > What about dependencies? 369ea8242c0fb sounds like it needs work for all
> > > > > notifiers need to be updated as well.
> > > > 
> > > > This commit remove mmu_notifier_invalidate_page() hence why everything
> > > > need to be updated. But in 4.4 you can get away with just adding start/
> > > > end and keep around mmu_notifier_invalidate_page() to minimize disruption.
> > > 
> > > OK, this is really interesting. I was really worried to change the
> > > semantic of the mmu notifiers in stable kernels because this is really
> > > a hard to review change and high risk for anybody running those old
> > > kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
> > > into the range scope API then this sounds like the best way forward.
> > > 
> > > So just to make sure we are at the same page. Does this sounds goo for
> > > stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
> > > top. What do you think?
> > 
> > You need to invalidate outside page table lock so before the call to
> > page_check_address(). For instance like below patch, which also only
> > do the range invalidation for huge page which would avoid too much of
> > a behavior change for user of mmu notifier.
> 
> Right. I would rather not make this PageHuge special though. So the
> fixed version should be.

Why not testing for huge ? Only huge is broken and thus only that
need the extra range invalidation. Doing the double invalidation
for single page is bit overkill.

Also below is bogus you need to add a out_notify: label to avoid
an inbalance in start/end callback.

> 
> From c05849f6789ec36e2ff11adcd8fa6cfb05e870a9 Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?J=C3=A9r=C3=B4me=20Glisse?= <jglisse@redhat.com>
> Date: Thu, 31 Aug 2017 17:17:27 -0400
> Subject: [PATCH] mm/rmap: update to new mmu_notifier semantic v2
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
> 
> commit 369ea8242c0fb5239b4ddf0dc568f694bd244de4 upstrea.
> 
> Please note that this patch differs from the mainline because we do not
> really replace mmu_notifier_invalidate_page by mmu_notifier_invalidate_range
> because that requires changes to most of existing mmu notifiers. We also
> do not want to change the semantic of this API in old kernels. Anyway
> Jerome has suggested that it should be sufficient to simply wrap
> mmu_notifier_invalidate_page by *_invalidate_range_start()/end() to fix
> invalidation of larger than pte mappings (e.g. THP/hugetlb pages during
> migration). We need this change to handle large (hugetlb/THP) pages
> migration properly.
> 
> Note that because we can not presume the pmd value or pte value we have
> to assume the worst and unconditionaly report an invalidation as
> happening.
> 
> Changed since v2:
>   - try_to_unmap_one() only one call to mmu_notifier_invalidate_range()
>   - compute end with PAGE_SIZE << compound_order(page)
>   - fix PageHuge() case in try_to_unmap_one()
> 
> Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
> Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
> Cc: Dan Williams <dan.j.williams@intel.com>
> Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
> Cc: Bernhard Held <berny156@gmx.de>
> Cc: Adam Borowski <kilobyte@angband.pl>
> Cc: Radim Krčmář <rkrcmar@redhat.com>
> Cc: Wanpeng Li <kernellwp@gmail.com>
> Cc: Paolo Bonzini <pbonzini@redhat.com>
> Cc: Takashi Iwai <tiwai@suse.de>
> Cc: Nadav Amit <nadav.amit@gmail.com>
> Cc: Mike Galbraith <efault@gmx.de>
> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> Cc: axie <axie@amd.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> Signed-off-by: Michal Hocko <mhocko@suse.com> # backport to 4.4
> ---
>  mm/rmap.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 1bceb49aa214..aba994f55d6c 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1324,12 +1324,21 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>  	pte_t pteval;
>  	spinlock_t *ptl;
>  	int ret = SWAP_AGAIN;
> +	unsigned long start = address, end;
>  	enum ttu_flags flags = (enum ttu_flags)arg;
>  
>  	/* munlock has nothing to gain from examining un-locked vmas */
>  	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
>  		goto out;
>  
> +	/*
> +	 * We have to assume the worse case ie pmd for invalidation. Note that
> +	 * the page can not be free in this function as call of try_to_unmap()
> +	 * must hold a reference on the page.
> +	 */
> +	end = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
> +	mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
> +
>  	pte = page_check_address(page, mm, address, &ptl, 0);
>  	if (!pte)
>  		goto out;

Instead

>  		goto out_notify;

> @@ -1450,6 +1459,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>  	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
>  		mmu_notifier_invalidate_page(mm, address);

+out_notify:

> +	mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
>  out:
>  	return ret;
>  }
>  
> -- 
> 2.18.0
> 
> -- 
> Michal Hocko
> SUSE Labs

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Mike Kravetz]

On 08/30/2018 09:57 AM, Jerome Glisse wrote:
> On Thu, Aug 30, 2018 at 06:19:52PM +0200, Michal Hocko wrote:
>> On Thu 30-08-18 10:08:25, Jerome Glisse wrote:
>>> On Thu, Aug 30, 2018 at 12:56:16PM +0200, Michal Hocko wrote:
>>>> On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
>>>>> On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
>>>>>> On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
>>>>>>> On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
>>>>>> [...]
>>>>>>>> What would be the best mmu notifier interface to use where there are no
>>>>>>>> start/end calls?
>>>>>>>> Or, is the best solution to add the start/end calls as is done in later
>>>>>>>> versions of the code?  If that is the suggestion, has there been any change
>>>>>>>> in invalidate start/end semantics that we should take into account?
>>>>>>>
>>>>>>> start/end would be the one to add, 4.4 seems broken in respect to THP
>>>>>>> and mmu notification. Another solution is to fix user of mmu notifier,
>>>>>>> they were only a handful back then. For instance properly adjust the
>>>>>>> address to match first address covered by pmd or pud and passing down
>>>>>>> correct page size to mmu_notifier_invalidate_page() would allow to fix
>>>>>>> this easily.
>>>>>>>
>>>>>>> This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
>>>>>>> with an invalid one (either poison, migration or swap) inside the
>>>>>>> function. So anyone racing would synchronize on those special entry
>>>>>>> hence why it is fine to delay mmu_notifier_invalidate_page() to after
>>>>>>> dropping the page table lock.
>>>>>>>
>>>>>>> Adding start/end might the solution with less code churn as you would
>>>>>>> only need to change try_to_unmap_one().
>>>>>>
>>>>>> What about dependencies? 369ea8242c0fb sounds like it needs work for all
>>>>>> notifiers need to be updated as well.
>>>>>
>>>>> This commit remove mmu_notifier_invalidate_page() hence why everything
>>>>> need to be updated. But in 4.4 you can get away with just adding start/
>>>>> end and keep around mmu_notifier_invalidate_page() to minimize disruption.
>>>>
>>>> OK, this is really interesting. I was really worried to change the
>>>> semantic of the mmu notifiers in stable kernels because this is really
>>>> a hard to review change and high risk for anybody running those old
>>>> kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
>>>> into the range scope API then this sounds like the best way forward.
>>>>
>>>> So just to make sure we are at the same page. Does this sounds goo for
>>>> stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
>>>> top. What do you think?
>>>
>>> You need to invalidate outside page table lock so before the call to
>>> page_check_address(). For instance like below patch, which also only
>>> do the range invalidation for huge page which would avoid too much of
>>> a behavior change for user of mmu notifier.
>>
>> Right. I would rather not make this PageHuge special though. So the
>> fixed version should be.
> 
> Why not testing for huge ? Only huge is broken and thus only that
> need the extra range invalidation. Doing the double invalidation
> for single page is bit overkill.

I am a bit confused, and hope this does not add to any confusion by others.

IIUC, the patch below does not attempt to 'fix' anything.  It is simply
there to add the start/end notifiers to the v4.4 version of this routine
so that a subsequent patch can use them (with modified ranges) to handle
unmapping a shared pmd huge page.  That is the mainline fix which started
this thread.

Since we are only/mostly interested in fixing the shared pmd issue in
4.4, how about just adding the start/end notifiers to the very specific
case where pmd sharing is possible?

I can see the value in trying to back port dependent patches such as this
so that stable releases look more like mainline.  However, I am not sure of
the value in this case as this patch was part of a larger set changing
notifier semantics.

-- 
Mike Kravetz

> Also below is bogus you need to add a out_notify: label to avoid
> an inbalance in start/end callback.
> 
>>
>> From c05849f6789ec36e2ff11adcd8fa6cfb05e870a9 Mon Sep 17 00:00:00 2001
>> From: =?UTF-8?q?J=C3=A9r=C3=B4me=20Glisse?= <jglisse@redhat.com>
>> Date: Thu, 31 Aug 2017 17:17:27 -0400
>> Subject: [PATCH] mm/rmap: update to new mmu_notifier semantic v2
>> MIME-Version: 1.0
>> Content-Type: text/plain; charset=UTF-8
>> Content-Transfer-Encoding: 8bit
>>
>> commit 369ea8242c0fb5239b4ddf0dc568f694bd244de4 upstrea.
>>
>> Please note that this patch differs from the mainline because we do not
>> really replace mmu_notifier_invalidate_page by mmu_notifier_invalidate_range
>> because that requires changes to most of existing mmu notifiers. We also
>> do not want to change the semantic of this API in old kernels. Anyway
>> Jerome has suggested that it should be sufficient to simply wrap
>> mmu_notifier_invalidate_page by *_invalidate_range_start()/end() to fix
>> invalidation of larger than pte mappings (e.g. THP/hugetlb pages during
>> migration). We need this change to handle large (hugetlb/THP) pages
>> migration properly.
>>
>> Note that because we can not presume the pmd value or pte value we have
>> to assume the worst and unconditionaly report an invalidation as
>> happening.
>>
>> Changed since v2:
>>   - try_to_unmap_one() only one call to mmu_notifier_invalidate_range()
>>   - compute end with PAGE_SIZE << compound_order(page)
>>   - fix PageHuge() case in try_to_unmap_one()
>>
>> Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
>> Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
>> Cc: Dan Williams <dan.j.williams@intel.com>
>> Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
>> Cc: Bernhard Held <berny156@gmx.de>
>> Cc: Adam Borowski <kilobyte@angband.pl>
>> Cc: Radim Krčmář <rkrcmar@redhat.com>
>> Cc: Wanpeng Li <kernellwp@gmail.com>
>> Cc: Paolo Bonzini <pbonzini@redhat.com>
>> Cc: Takashi Iwai <tiwai@suse.de>
>> Cc: Nadav Amit <nadav.amit@gmail.com>
>> Cc: Mike Galbraith <efault@gmx.de>
>> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
>> Cc: axie <axie@amd.com>
>> Cc: Andrew Morton <akpm@linux-foundation.org>
>> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
>> Signed-off-by: Michal Hocko <mhocko@suse.com> # backport to 4.4
>> ---
>>  mm/rmap.c | 10 ++++++++++
>>  1 file changed, 10 insertions(+)
>>
>> diff --git a/mm/rmap.c b/mm/rmap.c
>> index 1bceb49aa214..aba994f55d6c 100644
>> --- a/mm/rmap.c
>> +++ b/mm/rmap.c
>> @@ -1324,12 +1324,21 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>>  	pte_t pteval;
>>  	spinlock_t *ptl;
>>  	int ret = SWAP_AGAIN;
>> +	unsigned long start = address, end;
>>  	enum ttu_flags flags = (enum ttu_flags)arg;
>>  
>>  	/* munlock has nothing to gain from examining un-locked vmas */
>>  	if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
>>  		goto out;
>>  
>> +	/*
>> +	 * We have to assume the worse case ie pmd for invalidation. Note that
>> +	 * the page can not be free in this function as call of try_to_unmap()
>> +	 * must hold a reference on the page.
>> +	 */
>> +	end = min(vma->vm_end, start + (PAGE_SIZE << compound_order(page)));
>> +	mmu_notifier_invalidate_range_start(vma->vm_mm, start, end);
>> +
>>  	pte = page_check_address(page, mm, address, &ptl, 0);
>>  	if (!pte)
>>  		goto out;
> 
> Instead
> 
>>  		goto out_notify;
> 
>> @@ -1450,6 +1459,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
>>  	if (ret != SWAP_FAIL && ret != SWAP_MLOCK && !(flags & TTU_MUNLOCK))
>>  		mmu_notifier_invalidate_page(mm, address);
> 
> +out_notify:
> 
>> +	mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
>>  out:
>>  	return ret;
>>  }
>>  
>> -- 
>> 2.18.0
>>
>> -- 
>> Michal Hocko
>> SUSE Labs

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Jerome Glisse]

On Thu, Aug 30, 2018 at 11:05:16AM -0700, Mike Kravetz wrote:
> On 08/30/2018 09:57 AM, Jerome Glisse wrote:
> > On Thu, Aug 30, 2018 at 06:19:52PM +0200, Michal Hocko wrote:
> >> On Thu 30-08-18 10:08:25, Jerome Glisse wrote:
> >>> On Thu, Aug 30, 2018 at 12:56:16PM +0200, Michal Hocko wrote:
> >>>> On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
> >>>>> On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> >>>>>> On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> >>>>>>> On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> >>>>>> [...]
> >>>>>>>> What would be the best mmu notifier interface to use where there are no
> >>>>>>>> start/end calls?
> >>>>>>>> Or, is the best solution to add the start/end calls as is done in later
> >>>>>>>> versions of the code?  If that is the suggestion, has there been any change
> >>>>>>>> in invalidate start/end semantics that we should take into account?
> >>>>>>>
> >>>>>>> start/end would be the one to add, 4.4 seems broken in respect to THP
> >>>>>>> and mmu notification. Another solution is to fix user of mmu notifier,
> >>>>>>> they were only a handful back then. For instance properly adjust the
> >>>>>>> address to match first address covered by pmd or pud and passing down
> >>>>>>> correct page size to mmu_notifier_invalidate_page() would allow to fix
> >>>>>>> this easily.
> >>>>>>>
> >>>>>>> This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> >>>>>>> with an invalid one (either poison, migration or swap) inside the
> >>>>>>> function. So anyone racing would synchronize on those special entry
> >>>>>>> hence why it is fine to delay mmu_notifier_invalidate_page() to after
> >>>>>>> dropping the page table lock.
> >>>>>>>
> >>>>>>> Adding start/end might the solution with less code churn as you would
> >>>>>>> only need to change try_to_unmap_one().
> >>>>>>
> >>>>>> What about dependencies? 369ea8242c0fb sounds like it needs work for all
> >>>>>> notifiers need to be updated as well.
> >>>>>
> >>>>> This commit remove mmu_notifier_invalidate_page() hence why everything
> >>>>> need to be updated. But in 4.4 you can get away with just adding start/
> >>>>> end and keep around mmu_notifier_invalidate_page() to minimize disruption.
> >>>>
> >>>> OK, this is really interesting. I was really worried to change the
> >>>> semantic of the mmu notifiers in stable kernels because this is really
> >>>> a hard to review change and high risk for anybody running those old
> >>>> kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
> >>>> into the range scope API then this sounds like the best way forward.
> >>>>
> >>>> So just to make sure we are at the same page. Does this sounds goo for
> >>>> stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
> >>>> top. What do you think?
> >>>
> >>> You need to invalidate outside page table lock so before the call to
> >>> page_check_address(). For instance like below patch, which also only
> >>> do the range invalidation for huge page which would avoid too much of
> >>> a behavior change for user of mmu notifier.
> >>
> >> Right. I would rather not make this PageHuge special though. So the
> >> fixed version should be.
> > 
> > Why not testing for huge ? Only huge is broken and thus only that
> > need the extra range invalidation. Doing the double invalidation
> > for single page is bit overkill.
> 
> I am a bit confused, and hope this does not add to any confusion by others.
> 
> IIUC, the patch below does not attempt to 'fix' anything.  It is simply
> there to add the start/end notifiers to the v4.4 version of this routine
> so that a subsequent patch can use them (with modified ranges) to handle
> unmapping a shared pmd huge page.  That is the mainline fix which started
> this thread.
> 
> Since we are only/mostly interested in fixing the shared pmd issue in
> 4.4, how about just adding the start/end notifiers to the very specific
> case where pmd sharing is possible?
> 
> I can see the value in trying to back port dependent patches such as this
> so that stable releases look more like mainline.  However, I am not sure of
> the value in this case as this patch was part of a larger set changing
> notifier semantics.

For all intents and purposes this is not a backport of the original
patch so maybe we should just drop the commit reference and just
explains that it is there to fix mmu notifier in respect to huge page
migration.

The original patches fix more than this case because newer featurers
like THP migration, THP swapping, ... added more cases where things
would have been wrong. But in 4.4 frame there is only huge tlb fs
migration.

Cheers,
Jérôme

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Michal Hocko]

On Thu 30-08-18 14:39:44, Jerome Glisse wrote:
> On Thu, Aug 30, 2018 at 11:05:16AM -0700, Mike Kravetz wrote:
> > On 08/30/2018 09:57 AM, Jerome Glisse wrote:
> > > On Thu, Aug 30, 2018 at 06:19:52PM +0200, Michal Hocko wrote:
> > >> On Thu 30-08-18 10:08:25, Jerome Glisse wrote:
> > >>> On Thu, Aug 30, 2018 at 12:56:16PM +0200, Michal Hocko wrote:
> > >>>> On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
> > >>>>> On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> > >>>>>> On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> > >>>>>>> On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> > >>>>>> [...]
> > >>>>>>>> What would be the best mmu notifier interface to use where there are no
> > >>>>>>>> start/end calls?
> > >>>>>>>> Or, is the best solution to add the start/end calls as is done in later
> > >>>>>>>> versions of the code?  If that is the suggestion, has there been any change
> > >>>>>>>> in invalidate start/end semantics that we should take into account?
> > >>>>>>>
> > >>>>>>> start/end would be the one to add, 4.4 seems broken in respect to THP
> > >>>>>>> and mmu notification. Another solution is to fix user of mmu notifier,
> > >>>>>>> they were only a handful back then. For instance properly adjust the
> > >>>>>>> address to match first address covered by pmd or pud and passing down
> > >>>>>>> correct page size to mmu_notifier_invalidate_page() would allow to fix
> > >>>>>>> this easily.
> > >>>>>>>
> > >>>>>>> This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> > >>>>>>> with an invalid one (either poison, migration or swap) inside the
> > >>>>>>> function. So anyone racing would synchronize on those special entry
> > >>>>>>> hence why it is fine to delay mmu_notifier_invalidate_page() to after
> > >>>>>>> dropping the page table lock.
> > >>>>>>>
> > >>>>>>> Adding start/end might the solution with less code churn as you would
> > >>>>>>> only need to change try_to_unmap_one().
> > >>>>>>
> > >>>>>> What about dependencies? 369ea8242c0fb sounds like it needs work for all
> > >>>>>> notifiers need to be updated as well.
> > >>>>>
> > >>>>> This commit remove mmu_notifier_invalidate_page() hence why everything
> > >>>>> need to be updated. But in 4.4 you can get away with just adding start/
> > >>>>> end and keep around mmu_notifier_invalidate_page() to minimize disruption.
> > >>>>
> > >>>> OK, this is really interesting. I was really worried to change the
> > >>>> semantic of the mmu notifiers in stable kernels because this is really
> > >>>> a hard to review change and high risk for anybody running those old
> > >>>> kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
> > >>>> into the range scope API then this sounds like the best way forward.
> > >>>>
> > >>>> So just to make sure we are at the same page. Does this sounds goo for
> > >>>> stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
> > >>>> top. What do you think?
> > >>>
> > >>> You need to invalidate outside page table lock so before the call to
> > >>> page_check_address(). For instance like below patch, which also only
> > >>> do the range invalidation for huge page which would avoid too much of
> > >>> a behavior change for user of mmu notifier.
> > >>
> > >> Right. I would rather not make this PageHuge special though. So the
> > >> fixed version should be.
> > > 
> > > Why not testing for huge ? Only huge is broken and thus only that
> > > need the extra range invalidation. Doing the double invalidation
> > > for single page is bit overkill.
> > 
> > I am a bit confused, and hope this does not add to any confusion by others.
> > 
> > IIUC, the patch below does not attempt to 'fix' anything.  It is simply
> > there to add the start/end notifiers to the v4.4 version of this routine
> > so that a subsequent patch can use them (with modified ranges) to handle
> > unmapping a shared pmd huge page.  That is the mainline fix which started
> > this thread.
> > 
> > Since we are only/mostly interested in fixing the shared pmd issue in
> > 4.4, how about just adding the start/end notifiers to the very specific
> > case where pmd sharing is possible?
> > 
> > I can see the value in trying to back port dependent patches such as this
> > so that stable releases look more like mainline.  However, I am not sure of
> > the value in this case as this patch was part of a larger set changing
> > notifier semantics.
> 
> For all intents and purposes this is not a backport of the original
> patch so maybe we should just drop the commit reference and just
> explains that it is there to fix mmu notifier in respect to huge page
> migration.
> 
> The original patches fix more than this case because newer featurers
> like THP migration, THP swapping, ... added more cases where things
> would have been wrong. But in 4.4 frame there is only huge tlb fs
> migration.

And THP migration is still a problem with 4.4 AFAICS. All other cases
simply split the huge page but THP migration keeps it in one piece and
as such it is theoretically broken as you have explained. So I would
stick with what I posted with some more clarifications in the changelog
if you think it is appropriate (suggestions welcome).
-- 
Michal Hocko
SUSE Labs

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Jerome Glisse]

On Mon, Sep 03, 2018 at 07:56:54AM +0200, Michal Hocko wrote:
> On Thu 30-08-18 14:39:44, Jerome Glisse wrote:
> > On Thu, Aug 30, 2018 at 11:05:16AM -0700, Mike Kravetz wrote:
> > > On 08/30/2018 09:57 AM, Jerome Glisse wrote:
> > > > On Thu, Aug 30, 2018 at 06:19:52PM +0200, Michal Hocko wrote:
> > > >> On Thu 30-08-18 10:08:25, Jerome Glisse wrote:
> > > >>> On Thu, Aug 30, 2018 at 12:56:16PM +0200, Michal Hocko wrote:
> > > >>>> On Wed 29-08-18 17:11:07, Jerome Glisse wrote:
> > > >>>>> On Wed, Aug 29, 2018 at 08:39:06PM +0200, Michal Hocko wrote:
> > > >>>>>> On Wed 29-08-18 14:14:25, Jerome Glisse wrote:
> > > >>>>>>> On Wed, Aug 29, 2018 at 10:24:44AM -0700, Mike Kravetz wrote:
> > > >>>>>> [...]
> > > >>>>>>>> What would be the best mmu notifier interface to use where there are no
> > > >>>>>>>> start/end calls?
> > > >>>>>>>> Or, is the best solution to add the start/end calls as is done in later
> > > >>>>>>>> versions of the code?  If that is the suggestion, has there been any change
> > > >>>>>>>> in invalidate start/end semantics that we should take into account?
> > > >>>>>>>
> > > >>>>>>> start/end would be the one to add, 4.4 seems broken in respect to THP
> > > >>>>>>> and mmu notification. Another solution is to fix user of mmu notifier,
> > > >>>>>>> they were only a handful back then. For instance properly adjust the
> > > >>>>>>> address to match first address covered by pmd or pud and passing down
> > > >>>>>>> correct page size to mmu_notifier_invalidate_page() would allow to fix
> > > >>>>>>> this easily.
> > > >>>>>>>
> > > >>>>>>> This is ok because user of try_to_unmap_one() replace the pte/pmd/pud
> > > >>>>>>> with an invalid one (either poison, migration or swap) inside the
> > > >>>>>>> function. So anyone racing would synchronize on those special entry
> > > >>>>>>> hence why it is fine to delay mmu_notifier_invalidate_page() to after
> > > >>>>>>> dropping the page table lock.
> > > >>>>>>>
> > > >>>>>>> Adding start/end might the solution with less code churn as you would
> > > >>>>>>> only need to change try_to_unmap_one().
> > > >>>>>>
> > > >>>>>> What about dependencies? 369ea8242c0fb sounds like it needs work for all
> > > >>>>>> notifiers need to be updated as well.
> > > >>>>>
> > > >>>>> This commit remove mmu_notifier_invalidate_page() hence why everything
> > > >>>>> need to be updated. But in 4.4 you can get away with just adding start/
> > > >>>>> end and keep around mmu_notifier_invalidate_page() to minimize disruption.
> > > >>>>
> > > >>>> OK, this is really interesting. I was really worried to change the
> > > >>>> semantic of the mmu notifiers in stable kernels because this is really
> > > >>>> a hard to review change and high risk for anybody running those old
> > > >>>> kernels. If we can keep the mmu_notifier_invalidate_page and wrap them
> > > >>>> into the range scope API then this sounds like the best way forward.
> > > >>>>
> > > >>>> So just to make sure we are at the same page. Does this sounds goo for
> > > >>>> stable 4.4. backport? Mike's hugetlb pmd shared fixup can be applied on
> > > >>>> top. What do you think?
> > > >>>
> > > >>> You need to invalidate outside page table lock so before the call to
> > > >>> page_check_address(). For instance like below patch, which also only
> > > >>> do the range invalidation for huge page which would avoid too much of
> > > >>> a behavior change for user of mmu notifier.
> > > >>
> > > >> Right. I would rather not make this PageHuge special though. So the
> > > >> fixed version should be.
> > > > 
> > > > Why not testing for huge ? Only huge is broken and thus only that
> > > > need the extra range invalidation. Doing the double invalidation
> > > > for single page is bit overkill.
> > > 
> > > I am a bit confused, and hope this does not add to any confusion by others.
> > > 
> > > IIUC, the patch below does not attempt to 'fix' anything.  It is simply
> > > there to add the start/end notifiers to the v4.4 version of this routine
> > > so that a subsequent patch can use them (with modified ranges) to handle
> > > unmapping a shared pmd huge page.  That is the mainline fix which started
> > > this thread.
> > > 
> > > Since we are only/mostly interested in fixing the shared pmd issue in
> > > 4.4, how about just adding the start/end notifiers to the very specific
> > > case where pmd sharing is possible?
> > > 
> > > I can see the value in trying to back port dependent patches such as this
> > > so that stable releases look more like mainline.  However, I am not sure of
> > > the value in this case as this patch was part of a larger set changing
> > > notifier semantics.
> > 
> > For all intents and purposes this is not a backport of the original
> > patch so maybe we should just drop the commit reference and just
> > explains that it is there to fix mmu notifier in respect to huge page
> > migration.
> > 
> > The original patches fix more than this case because newer featurers
> > like THP migration, THP swapping, ... added more cases where things
> > would have been wrong. But in 4.4 frame there is only huge tlb fs
> > migration.
> 
> And THP migration is still a problem with 4.4 AFAICS. All other cases
> simply split the huge page but THP migration keeps it in one piece and
> as such it is theoretically broken as you have explained. So I would
> stick with what I posted with some more clarifications in the changelog
> if you think it is appropriate (suggestions welcome).

Reading code there is no THP migration in 4.4 only huge tlb migration.
Look at handle_mm_fault which do not know how to handle swap pmd, only
the huge tlb fs fault handler knows how to handle those. Hence why i
was checking for huge tlb exactly as page_check_address() to only range
invalidate for huge tlb fs migration.

But i am fine with doing the range invalidation with all.

Cheers,
Jérôme

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Mike Kravetz]

On 09/04/2018 07:00 AM, Jerome Glisse wrote:
> On Mon, Sep 03, 2018 at 07:56:54AM +0200, Michal Hocko wrote:
>> On Thu 30-08-18 14:39:44, Jerome Glisse wrote:
>>> For all intents and purposes this is not a backport of the original
>>> patch so maybe we should just drop the commit reference and just
>>> explains that it is there to fix mmu notifier in respect to huge page
>>> migration.
>>>
>>> The original patches fix more than this case because newer featurers
>>> like THP migration, THP swapping, ... added more cases where things
>>> would have been wrong. But in 4.4 frame there is only huge tlb fs
>>> migration.
>>
>> And THP migration is still a problem with 4.4 AFAICS. All other cases
>> simply split the huge page but THP migration keeps it in one piece and
>> as such it is theoretically broken as you have explained. So I would
>> stick with what I posted with some more clarifications in the changelog
>> if you think it is appropriate (suggestions welcome).
> 
> Reading code there is no THP migration in 4.4 only huge tlb migration.
> Look at handle_mm_fault which do not know how to handle swap pmd, only
> the huge tlb fs fault handler knows how to handle those. Hence why i
> was checking for huge tlb exactly as page_check_address() to only range
> invalidate for huge tlb fs migration.

I agree with Jérôme that THP migration was added after 4.4.  But, I could
be missing something.

> But i am fine with doing the range invalidation with all.

Since the shared pmd patch which will ultimately go on top of this needs
the PageHuge checks, my preference would be Jérôme's patch.

However, IMO I am not certain we really need/want a separate patch.  We
could just add the notifiers to the shared pmd patch.  Back porting the
shared pmd patch will also require some fixup.

Either would work.  I'll admit I do not know what stable maintainers would
prefer.
-- 
Mike Kravetz

Re: [PATCH v6 1/2] mm: migration: fix migration of huge PMD shared pages

[Michal Hocko]

On Tue 04-09-18 10:00:36, Jerome Glisse wrote:
> On Mon, Sep 03, 2018 at 07:56:54AM +0200, Michal Hocko wrote:
[...]
> > And THP migration is still a problem with 4.4 AFAICS. All other cases
> > simply split the huge page but THP migration keeps it in one piece and
> > as such it is theoretically broken as you have explained. So I would
> > stick with what I posted with some more clarifications in the changelog
> > if you think it is appropriate (suggestions welcome).
> 
> Reading code there is no THP migration in 4.4 only huge tlb migration.

Meh, you are right. For some reason I misread unmap_and_move_huge_page
to be also for THP. Sorry for the conusion. My fault!

Then it would be indeed safer to use your backport.
-- 
Michal Hocko
SUSE Labs