sgi-gru: get_user_page changes

PATCH Superseded Bharath Vedartham 4 participants 14 messages Jul 21, 2019 – Jul 22, 2019

Subsystems (2)

Memory management Char and misc drivers

lore

Share & tools

lore

Bharath Vedartham

author

2019-07-21 15:58 UTC #0 lore raw

cover sgi-gru: get_user_page changes

This patch series incorporates a few changes in the get_user_page usage of sgi-gru.

The main change is the first patch, which is a trivial one line change to convert put_page to put_user_page to enable tracking of get_user_pages.

The second patch removes an uneccessary ifdef of CONFIG_HUGETLB.

The third patch adds __get_user_pages_fast in atomic_pte_lookup to retrive a physical user page in an atomic context instead of manually walking up the page tables like the current code does. This patch should be subject to more review from the gup people.

drivers/misc/sgi-gru/* builds after this patch series. But I do not have the hardware to verify these changes.

The first patch implements gup tracking in the current code. This is to be tested as to check whether gup tracking works properly. Currently, in the upstream kernels put_user_page simply calls put_page. But that is to change in the future. Any suggestions as to how to test this code?

The implementation of gup tracking is in:
https://github.com/johnhubbard/linux/tree/gup_dma_core

We could test it by applying the first patch to the above tree and test it.

More details are in the individual changelogs.

Bharath Vedartham (3):
sgi-gru: Convert put_page() to get_user_page*()
sgi-gru: Remove CONFIG_HUGETLB_PAGE ifdef
sgi-gru: Use __get_user_pages_fast in atomic_pte_lookup

 drivers/misc/sgi-gru/grufault.c | 50 +++++++----------------------------------
 1 file changed, 8 insertions(+), 42 deletions(-)

Signature · 1 line

2.7.4

Bharath Vedartham

author

2019-07-21 15:58 UTC #1 lore raw

PATCH 1/3 sgi-gru: Convert put_page() to get_user_page*()

For pages that were retained via get_user_pages*(), release those pages via the new put_user_page*() routines, instead of via put_page().

This is part a tree-wide conversion, as described in commit fc1d8e7cca2d ("mm: introduce put_user_page*(), placeholder versions").

Cc Ira Weiny <[email protected]>

Cc John Hubbard <[email protected]>

Cc Jérôme Glisse <[email protected]>

Cc Greg Kroah-Hartman <[email protected]>

Cc Dimitri Sivanich <[email protected]>

Cc Arnd Bergmann <[email protected]>

Cc [email protected]

Signed-off-by Bharath Vedartham <[email protected]>

---
 drivers/misc/sgi-gru/grufault.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

drivers/misc/sgi-gru/grufault.c +1 -1

index 4b713a8..61b3447 100644
@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
 		return -EFAULT;
 	*paddr = page_to_phys(page);
-	put_page(page);
+	put_user_page(page);
 	return 0;
 }

Signature · 1 line

2.7.4

John Hubbard

Reviewed-by on v1 1/3

2019-07-22 02:25 UTC #2 lore raw

Re: [PATCH 1/3] sgi-gru: Convert put_page() to get_user_page*()

Quoted 15 lines · Bharath Vedartham

For pages that were retained via get_user_pages*(), release those pages via the new put_user_page*() routines, instead of via put_page().
This is part a tree-wide conversion, as described in commit fc1d8e7cca2d ("mm: introduce put_user_page*(), placeholder versions").
Cc Ira Weiny <[email protected]>
Cc John Hubbard <[email protected]>
Cc Jérôme Glisse <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Arnd Bergmann <[email protected]>
Cc [email protected]
Cc [email protected]
Signed-off-by Bharath Vedartham <[email protected]>

Quoted 4 lines

---
 drivers/misc/sgi-gru/grufault.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Reviewed-by John Hubbard <[email protected]>

thanks,

Signature · 2 lines

John Hubbard
NVIDIA

Quoted 14 lines

drivers/misc/sgi-gru/grufault.c +1 -1

index 4b713a8..61b3447 100644
@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
 		return -EFAULT;
 	*paddr = page_to_phys(page);
-	put_page(page);
+	put_user_page(page);
 	return 0;
 }

Bharath Vedartham

author

2019-07-22 17:47 UTC #3 lore raw

Re: [PATCH 1/3] sgi-gru: Convert put_page() to get_user_page*()

Quoted 21 lines · John Hubbard

Quoted 15 lines · Bharath Vedartham
For pages that were retained via get_user_pages*(), release those pages via the new put_user_page*() routines, instead of via put_page().
This is part a tree-wide conversion, as described in commit fc1d8e7cca2d ("mm: introduce put_user_page*(), placeholder versions").
Cc Ira Weiny <[email protected]>
Cc John Hubbard <[email protected]>
Cc Jérôme Glisse <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Arnd Bergmann <[email protected]>
Cc [email protected]
Cc [email protected]
Signed-off-by Bharath Vedartham <[email protected]>
Quoted 4 lines
---
 drivers/misc/sgi-gru/grufault.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Quoted 1 line

Reviewed-by John Hubbard <[email protected]>

Thanks!

Quoted 19 lines

thanks,

Signature · 2 lines

John Hubbard
NVIDIA

Quoted 14 lines

drivers/misc/sgi-gru/grufault.c +1 -1

index 4b713a8..61b3447 100644
@@ -188,7 +188,7 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
 		return -EFAULT;
 	*paddr = page_to_phys(page);
-	put_page(page);
+	put_user_page(page);
 	return 0;
 }

Bharath Vedartham

author

2019-07-21 15:58 UTC #4 lore raw

PATCH 2/3 sgi-gru: Remove CONFIG_HUGETLB_PAGE ifdef

In reply to Bharath Vedartham

is_vm_hugetlb_page has checks for whether CONFIG_HUGETLB_PAGE is defined or not. If CONFIG_HUGETLB_PAGE is not defined is_vm_hugetlb_page will always return false. There is no need to have an uneccessary CONFIG_HUGETLB_PAGE check in the code.

Cc Ira Weiny <[email protected]>

Cc John Hubbard <[email protected]>

Cc Jérôme Glisse <[email protected]>

Cc Greg Kroah-Hartman <[email protected]>

Cc Dimitri Sivanich <[email protected]>

Cc Arnd Bergmann <[email protected]>

Cc [email protected]

Signed-off-by Bharath Vedartham <[email protected]>

---
 drivers/misc/sgi-gru/grufault.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)

drivers/misc/sgi-gru/grufault.c +3 -8

index 61b3447..75108d2 100644
@@ -180,11 +180,8 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 {
 	struct page *page;
 
-#ifdef CONFIG_HUGETLB_PAGE
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
 		return -EFAULT;
 	*paddr = page_to_phys(page);

@@ -238,11 +235,9 @@ static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
 		return 1;
 
 	*paddr = pte_pfn(pte) << PAGE_SHIFT;
-#ifdef CONFIG_HUGETLB_PAGE
+
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	return 0;
 
 err:

Signature · 1 line

2.7.4

John Hubbard

Reviewed-by on v1 2/3

2019-07-22 02:34 UTC #5 lore raw

Re: [PATCH 2/3] sgi-gru: Remove CONFIG_HUGETLB_PAGE ifdef

Quoted 40 lines · Bharath Vedartham

is_vm_hugetlb_page has checks for whether CONFIG_HUGETLB_PAGE is defined or not. If CONFIG_HUGETLB_PAGE is not defined is_vm_hugetlb_page will always return false. There is no need to have an uneccessary CONFIG_HUGETLB_PAGE check in the code.
Cc Ira Weiny <[email protected]>
Cc John Hubbard <[email protected]>
Cc Jérôme Glisse <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Arnd Bergmann <[email protected]>
Cc [email protected]
Cc [email protected]
Signed-off-by Bharath Vedartham <[email protected]>
---
 drivers/misc/sgi-gru/grufault.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)
drivers/misc/sgi-gru/grufault.c +1 -4
index 61b3447..75108d2 100644
@@ -180,11 +180,8 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 {
 	struct page *page;
 
-#ifdef CONFIG_HUGETLB_PAGE
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
 		return -EFAULT;
 	*paddr = page_to_phys(page);
@@ -238,11 +235,9 @@ static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
 		return 1;
 
 	*paddr = pte_pfn(pte) << PAGE_SHIFT;

drivers/misc/sgi-gru/grufault.c static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr, · line 243

Quoted 10 lines

drivers/misc/sgi-gru/grufault.c +2 -4

-#ifdef CONFIG_HUGETLB_PAGE
+
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	return 0;
 
 err:

Looks like an accurate cleanup to me.

Reviewed-by John Hubbard <[email protected]>

thanks,

Signature · 2 lines

John Hubbard
NVIDIA

William Kucharski

2019-07-22 03:22 UTC #6 lore raw

Re: [PATCH 2/3] sgi-gru: Remove CONFIG_HUGETLB_PAGE ifdef

In reply to Bharath Vedartham

I suspect I'm being massively pedantic here, but the comments for atomic_pte_lookup() note:

* Only supports Intel large pages (2MB only) on x86_64.
* ZZZ - hugepage support is incomplete

That makes me wonder how many systems using this hardware are actually configured with CONFIG_HUGETLB_PAGE.

I ask as in the most common case, this is likely introducing a few extra instructions and possibly an additional branch to a routine that is called per-fault.

So the nit-picky questions are:

1) Does the code really need to be cleaned up in this way?

2) If it does, does it make more sense (given the way pmd_large() is handled now in atomic_pte_lookup()) for this to be coded as:

if (unlikely(is_vm_hugetlb_page(vma)))

	*pageshift = HPAGE_SHIFT;
else
	*pageshift = PAGE_SHIFT;

In all likelihood, these questions are no-ops, and the optimizer may even make my questions completely moot, but I thought I might as well ask anyway.

Quoted 54 lines · Bharath Vedartham

is_vm_hugetlb_page has checks for whether CONFIG_HUGETLB_PAGE is defined or not. If CONFIG_HUGETLB_PAGE is not defined is_vm_hugetlb_page will always return false. There is no need to have an uneccessary CONFIG_HUGETLB_PAGE check in the code.
Cc Ira Weiny <[email protected]>
Cc John Hubbard <[email protected]>
Cc Jérôme Glisse <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Arnd Bergmann <[email protected]>
Cc [email protected]
Cc [email protected]
Signed-off-by Bharath Vedartham <[email protected]>
---
drivers/misc/sgi-gru/grufault.c | 11 +++--------
1 file changed, 3 insertions(+), 8 deletions(-)
drivers/misc/sgi-gru/grufault.c +3 -8
index 61b3447..75108d2 100644
@@ -180,11 +180,8 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 {
 	struct page *page;
 
-#ifdef CONFIG_HUGETLB_PAGE
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
 		return -EFAULT;
 	*paddr = page_to_phys(page);
@@ -238,11 +235,9 @@ static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
 		return 1;
 
 	*paddr = pte_pfn(pte) << PAGE_SHIFT;
-#ifdef CONFIG_HUGETLB_PAGE
+
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	return 0;
 
err:
Signature · 1 line
2.7.4

Bharath Vedartham

author

2019-07-22 17:50 UTC #7 lore raw

Re: [PATCH 2/3] sgi-gru: Remove CONFIG_HUGETLB_PAGE ifdef

Quoted 13 lines · William Kucharski

I suspect I'm being massively pedantic here, but the comments for atomic_pte_lookup() note:
* Only supports Intel large pages (2MB only) on x86_64.
* ZZZ - hugepage support is incomplete
That makes me wonder how many systems using this hardware are actually configured with CONFIG_HUGETLB_PAGE.
I ask as in the most common case, this is likely introducing a few extra instructions and possibly an additional branch to a routine that is called per-fault.
So the nit-picky questions are:
1) Does the code really need to be cleaned up in this way?

Quoted 9 lines

2) If it does, does it make more sense (given the way pmd_large() is handled now in atomic_pte_lookup()) for this to be coded as:
if (unlikely(is_vm_hugetlb_page(vma)))
	*pageshift = HPAGE_SHIFT;
else
	*pageshift = PAGE_SHIFT;
In all likelihood, these questions are no-ops, and the optimizer may even make my questions completely moot, but I thought I might as well ask anyway.

That sounds reasonable. I am not really sure as to how much of an improvement it would be, the condition will be evaluated eitherways AFAIK? Eitherways, the ternary operator does not look good. I ll make a version 2 of this.

Quoted 55 lines

Quoted 54 lines · Bharath Vedartham
is_vm_hugetlb_page has checks for whether CONFIG_HUGETLB_PAGE is defined or not. If CONFIG_HUGETLB_PAGE is not defined is_vm_hugetlb_page will always return false. There is no need to have an uneccessary CONFIG_HUGETLB_PAGE check in the code.
Cc Ira Weiny <[email protected]>
Cc John Hubbard <[email protected]>
Cc Jérôme Glisse <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Arnd Bergmann <[email protected]>
Cc [email protected]
Cc [email protected]
Signed-off-by Bharath Vedartham <[email protected]>
---
drivers/misc/sgi-gru/grufault.c | 11 +++--------
1 file changed, 3 insertions(+), 8 deletions(-)
drivers/misc/sgi-gru/grufault.c +3 -8
index 61b3447..75108d2 100644
@@ -180,11 +180,8 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 {
 	struct page *page;
 
-#ifdef CONFIG_HUGETLB_PAGE
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
 		return -EFAULT;
 	*paddr = page_to_phys(page);
@@ -238,11 +235,9 @@ static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
 		return 1;
 
 	*paddr = pte_pfn(pte) << PAGE_SHIFT;
-#ifdef CONFIG_HUGETLB_PAGE
+
 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
-#else
-	*pageshift = PAGE_SHIFT;
-#endif
+
 	return 0;
 
err:
Signature · 1 line
2.7.4

William Kucharski

Reviewed-by

2019-07-22 22:54 UTC #8 lore raw

Re: [PATCH 2/3] sgi-gru: Remove CONFIG_HUGETLB_PAGE ifdef

Quoted 12 lines · Bharath Vedartham

Quoted 5 lines
In all likelihood, these questions are no-ops, and the optimizer may even make my questions completely moot, but I thought I might as well ask anyway.
That sounds reasonable. I am not really sure as to how much of an improvement it would be, the condition will be evaluated eitherways AFAIK? Eitherways, the ternary operator does not look good. I ll make a version 2 of this.

In THEORY the "unlikely" hints to the compiler that that leg of the "if" can be made the branch and jump leg, though in reality optimization is much more complex than that.

Still, the unlikely() call is also nicely self-documenting as to what the expected outcome is.

Reviewed-by William Kucharski <[email protected]>

Bharath Vedartham

author

2019-07-21 15:58 UTC #9 lore raw

PATCH 3/3 sgi-gru: Use __get_user_pages_fast in atomic_pte_lookup

In reply to Bharath Vedartham

*pte_lookup functions get the physical address for a given virtual address by getting a physical page using gup and use page_to_phys to get the physical address.

Currently, atomic_pte_lookup manually walks the page tables. If this function fails to get a physical page, it will fall back too non_atomic_pte_lookup to get a physical page which uses the slow gup path to get the physical page.

Instead of manually walking the page tables use __get_user_pages_fast which does the same thing and it does not fall back to the slow gup path.

This is largely inspired from kvm code. kvm uses __get_user_pages_fast in hva_to_pfn_fast function which can run in an atomic context.

Cc Ira Weiny <[email protected]>

Cc John Hubbard <[email protected]>

Cc Jérôme Glisse <[email protected]>

Cc Greg Kroah-Hartman <[email protected]>

Cc Dimitri Sivanich <[email protected]>

Cc Arnd Bergmann <[email protected]>

Cc [email protected]

Signed-off-by Bharath Vedartham <[email protected]>

---
 drivers/misc/sgi-gru/grufault.c | 39 +++++----------------------------------
 1 file changed, 5 insertions(+), 34 deletions(-)

drivers/misc/sgi-gru/grufault.c +5 ~~-34~~

index 75108d2..121c9a4 100644
@@ -202,46 +202,17 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
 	int write, unsigned long *paddr, int *pageshift)
 {
-	pgd_t *pgdp;
-	p4d_t *p4dp;
-	pud_t *pudp;
-	pmd_t *pmdp;
-	pte_t pte;
-
-	pgdp = pgd_offset(vma->vm_mm, vaddr);
-	if (unlikely(pgd_none(*pgdp)))
-		goto err;
-
-	p4dp = p4d_offset(pgdp, vaddr);
-	if (unlikely(p4d_none(*p4dp)))
-		goto err;
-
-	pudp = pud_offset(p4dp, vaddr);
-	if (unlikely(pud_none(*pudp)))
-		goto err;
+	struct page *page;
 
-	pmdp = pmd_offset(pudp, vaddr);
-	if (unlikely(pmd_none(*pmdp)))
-		goto err;
-#ifdef CONFIG_X86_64
-	if (unlikely(pmd_large(*pmdp)))
-		pte = *(pte_t *) pmdp;
-	else
-#endif
-		pte = *pte_offset_kernel(pmdp, vaddr);
+	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
 
-	if (unlikely(!pte_present(pte) ||
-		     (write && (!pte_write(pte) || !pte_dirty(pte)))))
+	if (!__get_user_pages_fast(vaddr, 1, write, &page))
 		return 1;
 
-	*paddr = pte_pfn(pte) << PAGE_SHIFT;
-
-	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
+	*paddr = page_to_phys(page);
+	put_user_page(page);
 
 	return 0;
-
-err:
-	return 1;
 }
 
 static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,

Signature · 1 line

2.7.4

John Hubbard

2019-07-22 02:32 UTC #10 lore raw

Re: [PATCH 3/3] sgi-gru: Use __get_user_pages_fast in atomic_pte_lookup

Quoted 61 lines · Bharath Vedartham

*pte_lookup functions get the physical address for a given virtual address by getting a physical page using gup and use page_to_phys to get the physical address.
Currently, atomic_pte_lookup manually walks the page tables. If this function fails to get a physical page, it will fall back too non_atomic_pte_lookup to get a physical page which uses the slow gup path to get the physical page.
Instead of manually walking the page tables use __get_user_pages_fast which does the same thing and it does not fall back to the slow gup path.
This is largely inspired from kvm code. kvm uses __get_user_pages_fast in hva_to_pfn_fast function which can run in an atomic context.
Cc Ira Weiny <[email protected]>
Cc John Hubbard <[email protected]>
Cc Jérôme Glisse <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Arnd Bergmann <[email protected]>
Cc [email protected]
Cc [email protected]
Signed-off-by Bharath Vedartham <[email protected]>
---
 drivers/misc/sgi-gru/grufault.c | 39 +++++----------------------------------
 1 file changed, 5 insertions(+), 34 deletions(-)
drivers/misc/sgi-gru/grufault.c +1 ~~-22~~
index 75108d2..121c9a4 100644
@@ -202,46 +202,17 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
 	int write, unsigned long *paddr, int *pageshift)
 {
-	pgd_t *pgdp;
-	p4d_t *p4dp;
-	pud_t *pudp;
-	pmd_t *pmdp;
-	pte_t pte;
-
-	pgdp = pgd_offset(vma->vm_mm, vaddr);
-	if (unlikely(pgd_none(*pgdp)))
-		goto err;
-
-	p4dp = p4d_offset(pgdp, vaddr);
-	if (unlikely(p4d_none(*p4dp)))
-		goto err;
-
-	pudp = pud_offset(p4dp, vaddr);
-	if (unlikely(pud_none(*pudp)))
-		goto err;
+	struct page *page;
 
-	pmdp = pmd_offset(pudp, vaddr);
-	if (unlikely(pmd_none(*pmdp)))
-		goto err;
-#ifdef CONFIG_X86_64
-	if (unlikely(pmd_large(*pmdp)))

drivers/misc/sgi-gru/grufault.c static int non_atomic_pte_lookup(struct vm_area_struct *vma, · line 210

Quoted 10 lines

drivers/misc/sgi-gru/grufault.c +2 -6

-		pte = *(pte_t *) pmdp;
-	else
-#endif
-		pte = *pte_offset_kernel(pmdp, vaddr);
+	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
 
-	if (unlikely(!pte_present(pte) ||
-		     (write && (!pte_write(pte) || !pte_dirty(pte)))))
+	if (!__get_user_pages_fast(vaddr, 1, write, &page))
 		return 1;

Let's please use numeric, not boolean comparison, for the return value of gup.

Also, optional: as long as you're there, atomic_pte_lookup() ought to either return a bool (true == success) or an errno, rather than a numeric zero or one.

Other than that, this looks like a good cleanup, I wonder how many open-coded gup implementations are floating around like this.

thanks,

Signature · 2 lines

John Hubbard
NVIDIA

Quoted 15 lines

drivers/misc/sgi-gru/grufault.c +2 -6

-	*paddr = pte_pfn(pte) << PAGE_SHIFT;
-
-	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
+	*paddr = page_to_phys(page);
+	put_user_page(page);
 
 	return 0;
-
-err:
-	return 1;
 }
 
 static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,

Bharath Vedartham

author

2019-07-22 17:53 UTC #11 lore raw

Re: [PATCH 3/3] sgi-gru: Use __get_user_pages_fast in atomic_pte_lookup

Quoted 68 lines · John Hubbard

Quoted 67 lines · Bharath Vedartham
*pte_lookup functions get the physical address for a given virtual address by getting a physical page using gup and use page_to_phys to get the physical address.
Currently, atomic_pte_lookup manually walks the page tables. If this function fails to get a physical page, it will fall back too non_atomic_pte_lookup to get a physical page which uses the slow gup path to get the physical page.
Instead of manually walking the page tables use __get_user_pages_fast which does the same thing and it does not fall back to the slow gup path.
This is largely inspired from kvm code. kvm uses __get_user_pages_fast in hva_to_pfn_fast function which can run in an atomic context.
Cc Ira Weiny <[email protected]>
Cc John Hubbard <[email protected]>
Cc Jérôme Glisse <[email protected]>
Cc Greg Kroah-Hartman <[email protected]>
Cc Dimitri Sivanich <[email protected]>
Cc Arnd Bergmann <[email protected]>
Cc [email protected]
Cc [email protected]
Signed-off-by Bharath Vedartham <[email protected]>
---
 drivers/misc/sgi-gru/grufault.c | 39 +++++----------------------------------
 1 file changed, 5 insertions(+), 34 deletions(-)
drivers/misc/sgi-gru/grufault.c +2 ~~-26~~
index 75108d2..121c9a4 100644
@@ -202,46 +202,17 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
 static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
 	int write, unsigned long *paddr, int *pageshift)
 {
-	pgd_t *pgdp;
-	p4d_t *p4dp;
-	pud_t *pudp;
-	pmd_t *pmdp;
-	pte_t pte;
-
-	pgdp = pgd_offset(vma->vm_mm, vaddr);
-	if (unlikely(pgd_none(*pgdp)))
-		goto err;
-
-	p4dp = p4d_offset(pgdp, vaddr);
-	if (unlikely(p4d_none(*p4dp)))
-		goto err;
-
-	pudp = pud_offset(p4dp, vaddr);
-	if (unlikely(pud_none(*pudp)))
-		goto err;
+	struct page *page;
 
-	pmdp = pmd_offset(pudp, vaddr);
-	if (unlikely(pmd_none(*pmdp)))
-		goto err;
-#ifdef CONFIG_X86_64
-	if (unlikely(pmd_large(*pmdp)))
-		pte = *(pte_t *) pmdp;
-	else
-#endif
-		pte = *pte_offset_kernel(pmdp, vaddr);
+	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
 

Quoted 7 lines

drivers/misc/sgi-gru/grufault.c static int non_atomic_pte_lookup(struct vm_area_struct *vma, · line 210
Quoted 4 lines
drivers/misc/sgi-gru/grufault.c +1 -2
-	if (unlikely(!pte_present(pte) ||
-		     (write && (!pte_write(pte) || !pte_dirty(pte)))))
+	if (!__get_user_pages_fast(vaddr, 1, write, &page))
 		return 1;
Let's please use numeric, not boolean comparison, for the return value of gup.

Alright then! I ll resubmit it!

Quoted 3 lines

Also, optional: as long as you're there, atomic_pte_lookup() ought to either return a bool (true == success) or an errno, rather than a numeric zero or one.

That makes sense. But the code which uses atomic_pte_lookup uses the return value of 1 for success and failure value of 0 in gru_vtop. That's why I did not mess with the return values in this code. It would require some change in the driver functionality which I am not ready to do :(

Quoted 2 lines

Other than that, this looks like a good cleanup, I wonder how many open-coded gup implementations are floating around like this.

I ll be on the lookout!

Quoted 20 lines

thanks,

Signature · 2 lines

John Hubbard
NVIDIA

Quoted 15 lines

drivers/misc/sgi-gru/grufault.c +2 -6

-	*paddr = pte_pfn(pte) << PAGE_SHIFT;
-
-	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
+	*paddr = page_to_phys(page);
+	put_user_page(page);
 
 	return 0;
-
-err:
-	return 1;
 }
 
 static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,

John Hubbard

2019-07-22 23:06 UTC #12 lore raw

Re: [PATCH 3/3] sgi-gru: Use __get_user_pages_fast in atomic_pte_lookup

Quoted 2 lines · Bharath Vedartham

Quoted 1 line · John Hubbard
On 7/21/19 8:58 AM, Bharath Vedartham wrote:

...

Quoted 7 lines

Quoted 3 lines
Also, optional: as long as you're there, atomic_pte_lookup() ought to either return a bool (true == success) or an errno, rather than a numeric zero or one.
That makes sense. But the code which uses atomic_pte_lookup uses the return value of 1 for success and failure value of 0 in gru_vtop. That's why I did not mess with the return values in this code. It would require some change in the driver functionality which I am not ready to do :(

It's a static function with only one caller. You could just merge in something like this, on top of what you have:

drivers/misc/sgi-gru/grufault.c +10 -8

index 121c9a4ccb94..2f768fc06432 100644
@@ -189,10 +189,11 @@ static int non_atomic_pte_lookup(struct vm_area_struct *vma,
        return 0;
 }
 
-/*
- * atomic_pte_lookup
+/**
+ * atomic_pte_lookup() - Convert a user virtual address to a physical address
+ * @Return: true for success, false for failure. Failure means that the page
+ *         could not be pinned via gup fast.
  *
- * Convert a user virtual address to a physical address
  * Only supports Intel large pages (2MB only) on x86_64.
  *     ZZZ - hugepage support is incomplete
  *

@@ -207,12 +208,12 @@ static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
        *pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
 
        if (!__get_user_pages_fast(vaddr, 1, write, &page))
-               return 1;
+               return false;
 
        *paddr = page_to_phys(page);
        put_user_page(page);
 
-       return 0;
+       return true;
 }
 
 static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,

@@ -221,7 +222,8 @@ static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,
        struct mm_struct *mm = gts->ts_mm;
        struct vm_area_struct *vma;
        unsigned long paddr;
-       int ret, ps;
+       int ps;
+       bool success;
 
        vma = find_vma(mm, vaddr);
        if (!vma)

@@ -232,8 +234,8 @@ static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,
         * context.
         */
        rmb();  /* Must/check ms_range_active before loading PTEs */
-       ret = atomic_pte_lookup(vma, vaddr, write, &paddr, &ps);
-       if (ret) {
+       success = atomic_pte_lookup(vma, vaddr, write, &paddr, &ps);
+       if (!success) {
                if (atomic)
                        goto upm;
                if (non_atomic_pte_lookup(vma, vaddr, write, &paddr, &ps))

thanks,

Signature · 2 lines

John Hubbard
NVIDIA

Ira Weiny

Reviewed-by on the v1 series

2019-07-22 04:49 UTC #13 lore raw

In reply to Bharath Vedartham

Quoted 21 lines · Bharath Vedartham

This patch series incorporates a few changes in the get_user_page usage of sgi-gru.
The main change is the first patch, which is a trivial one line change to convert put_page to put_user_page to enable tracking of get_user_pages.
The second patch removes an uneccessary ifdef of CONFIG_HUGETLB.
The third patch adds __get_user_pages_fast in atomic_pte_lookup to retrive a physical user page in an atomic context instead of manually walking up the page tables like the current code does. This patch should be subject to more review from the gup people.
drivers/misc/sgi-gru/* builds after this patch series. But I do not have the hardware to verify these changes.
The first patch implements gup tracking in the current code. This is to be tested as to check whether gup tracking works properly. Currently, in the upstream kernels put_user_page simply calls put_page. But that is to change in the future. Any suggestions as to how to test this code?

Quoted 8 lines

The implementation of gup tracking is in:
https://github.com/johnhubbard/linux/tree/gup_dma_core
We could test it by applying the first patch to the above tree and test it.
More details are in the individual changelogs.
Bharath Vedartham (3):

I don't have an opinion on the second patch regarding any performance concerns since I'm not familiar with this hardware.

But from a GUP POV For the series.

Reviewed-by Ira Weiny <[email protected]>

Quoted 10 lines

sgi-gru: Convert put_page() to get_user_page*()
sgi-gru: Remove CONFIG_HUGETLB_PAGE ifdef
sgi-gru: Use __get_user_pages_fast in atomic_pte_lookup
 drivers/misc/sgi-gru/grufault.c | 50 +++++++----------------------------------
 1 file changed, 8 insertions(+), 42 deletions(-)
Signature · 1 line
2.7.4