[2/2] powerpc/nvdimm: Update vmemmap_populated to check sub-section range
diff mbox series

Message ID 20190910062826.10041-2-aneesh.kumar@linux.ibm.com
State New
Headers show
Series
  • [1/2] libnvdimm/altmap: Track namespace boundaries in altmap
Related show

Commit Message

Aneesh Kumar K.V Sept. 10, 2019, 6:28 a.m. UTC
With commit: 7cc7867fb061 ("mm/devm_memremap_pages: enable sub-section remap")
pmem namespaces are remapped in 2M chunks. On architectures like ppc64 we
can map the memmap area using 16MB hugepage size and that can cover
a memory range of 16G.

While enabling new pmem namespaces, since memory is added in sub-section chunks,
before creating a new memmap mapping, kernel should check whether there is an
existing memmap mapping covering the new pmem namespace. Currently, this is
validated by checking whether the section covering the range is already
initialized or not. Considering there can be multiple namespaces in the same
section this can result in wrong validation. Update this to check for
sub-sections in the range. This is done by checking for all pfns in the range we
are mapping.

We could optimize this by checking only just one pfn in each sub-section. But
since this is not fast-path we keep this simple.

Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
---
 arch/powerpc/mm/init_64.c | 45 ++++++++++++++++++++-------------------
 1 file changed, 23 insertions(+), 22 deletions(-)

Comments

Dan Williams Sept. 16, 2019, 5:46 p.m. UTC | #1
On Mon, Sep 9, 2019 at 11:29 PM Aneesh Kumar K.V
<aneesh.kumar@linux.ibm.com> wrote:
>
> With commit: 7cc7867fb061 ("mm/devm_memremap_pages: enable sub-section remap")
> pmem namespaces are remapped in 2M chunks. On architectures like ppc64 we
> can map the memmap area using 16MB hugepage size and that can cover
> a memory range of 16G.
>
> While enabling new pmem namespaces, since memory is added in sub-section chunks,
> before creating a new memmap mapping, kernel should check whether there is an
> existing memmap mapping covering the new pmem namespace. Currently, this is
> validated by checking whether the section covering the range is already
> initialized or not. Considering there can be multiple namespaces in the same
> section this can result in wrong validation. Update this to check for
> sub-sections in the range. This is done by checking for all pfns in the range we
> are mapping.
>
> We could optimize this by checking only just one pfn in each sub-section. But
> since this is not fast-path we keep this simple.
>
> Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
> ---
>  arch/powerpc/mm/init_64.c | 45 ++++++++++++++++++++-------------------
>  1 file changed, 23 insertions(+), 22 deletions(-)
>
> diff --git a/arch/powerpc/mm/init_64.c b/arch/powerpc/mm/init_64.c
> index 4e08246acd79..7710ccdc19a2 100644
> --- a/arch/powerpc/mm/init_64.c
> +++ b/arch/powerpc/mm/init_64.c
> @@ -70,30 +70,24 @@ EXPORT_SYMBOL_GPL(kernstart_addr);
>
>  #ifdef CONFIG_SPARSEMEM_VMEMMAP
>  /*
> - * Given an address within the vmemmap, determine the pfn of the page that
> - * represents the start of the section it is within.  Note that we have to
> - * do this by hand as the proffered address may not be correctly aligned.
> - * Subtraction of non-aligned pointers produces undefined results.
> - */
> -static unsigned long __meminit vmemmap_section_start(unsigned long page)
> -{
> -       unsigned long offset = page - ((unsigned long)(vmemmap));
> -
> -       /* Return the pfn of the start of the section. */
> -       return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
> -}
> -
> -/*
> - * Check if this vmemmap page is already initialised.  If any section
> + * Check if this vmemmap page is already initialised.  If any sub section
>   * which overlaps this vmemmap page is initialised then this page is
>   * initialised already.
>   */
> -static int __meminit vmemmap_populated(unsigned long start, int page_size)
> +
> +static int __meminit vmemmap_populated(unsigned long start, int size)
>  {
> -       unsigned long end = start + page_size;
> -       start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
> +       unsigned long end = start + size;
>
> -       for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
> +       /* start is size aligned and it is always > sizeof(struct page) */
> +       VM_BUG_ON(start & sizeof(struct page));

If start is size aligned why not include that assumption in the VM_BUG_ON()?

Otherwise it seems this patch could be reduced simply by:

s/PAGE_SECTION_MASK/PAGE_SUBSECTION_MASK/
s/PAGES_PER_SECTION/PAGES_PER_SUBSECTION/

...and leave the vmemmap_section_start() function in place? In other
words this path used to guarantee that 'start' was aligned to the
minimum mem-hotplug granularity, the change looks ok on the surface,
but it seems a subtle change in semantics.

Can you get an ack from a powerpc maintainer, or maybe this patch
should route through the powerpc tree?

I'll take patch1 through the nvdimm tree.
Oliver O'Halloran Sept. 17, 2019, 5:47 a.m. UTC | #2
On Tue, Sep 10, 2019 at 4:29 PM Aneesh Kumar K.V
<aneesh.kumar@linux.ibm.com> wrote:
>
> With commit: 7cc7867fb061 ("mm/devm_memremap_pages: enable sub-section remap")
> pmem namespaces are remapped in 2M chunks. On architectures like ppc64 we
> can map the memmap area using 16MB hugepage size and that can cover
> a memory range of 16G.
>
> While enabling new pmem namespaces, since memory is added in sub-section chunks,
> before creating a new memmap mapping, kernel should check whether there is an
> existing memmap mapping covering the new pmem namespace. Currently, this is
> validated by checking whether the section covering the range is already
> initialized or not. Considering there can be multiple namespaces in the same
> section this can result in wrong validation. Update this to check for
> sub-sections in the range. This is done by checking for all pfns in the range we
> are mapping.
>
> We could optimize this by checking only just one pfn in each sub-section. But
> since this is not fast-path we keep this simple.
>
> Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
> ---
>  arch/powerpc/mm/init_64.c | 45 ++++++++++++++++++++-------------------
>  1 file changed, 23 insertions(+), 22 deletions(-)
>
> diff --git a/arch/powerpc/mm/init_64.c b/arch/powerpc/mm/init_64.c
> index 4e08246acd79..7710ccdc19a2 100644
> --- a/arch/powerpc/mm/init_64.c
> +++ b/arch/powerpc/mm/init_64.c
> @@ -70,30 +70,24 @@ EXPORT_SYMBOL_GPL(kernstart_addr);
>
>  #ifdef CONFIG_SPARSEMEM_VMEMMAP
>  /*
> - * Given an address within the vmemmap, determine the pfn of the page that
> - * represents the start of the section it is within.  Note that we have to
> - * do this by hand as the proffered address may not be correctly aligned.
> - * Subtraction of non-aligned pointers produces undefined results.
> - */
> -static unsigned long __meminit vmemmap_section_start(unsigned long page)
> -{
> -       unsigned long offset = page - ((unsigned long)(vmemmap));
> -
> -       /* Return the pfn of the start of the section. */
> -       return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
> -}

If you want to go with Dan's suggestion of keeping the function and
using PAGE_SUBSECTION_MASK then can you fold the pfn_to_page() below
into vmemmap_section_start()? The current behaviour of returning a pfn
makes no sense to me.

> - * Check if this vmemmap page is already initialised.  If any section
> + * Check if this vmemmap page is already initialised.  If any sub section
>   * which overlaps this vmemmap page is initialised then this page is
>   * initialised already.
>   */
> -static int __meminit vmemmap_populated(unsigned long start, int page_size)
> +
> +static int __meminit vmemmap_populated(unsigned long start, int size)
>  {
> -       unsigned long end = start + page_size;
> -       start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
> +       unsigned long end = start + size;

> -       for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
> +       /* start is size aligned and it is always > sizeof(struct page) */
> +       VM_BUG_ON(start & sizeof(struct page));

Shouldn't the test be: start & (sizeof(struct page) - 1)?
VM_BUG_ON(start != ALIGN(start, page_size)) would be clearer.

> +       for (; start < end; start += sizeof(struct page))
> +               /*
> +                * pfn valid check here is intended to really check
> +                * whether we have any subsection already initialized
> +                * in this range. We keep it simple by checking every
> +                * pfn in the range.
> +                */
>                 if (pfn_valid(page_to_pfn((struct page *)start)))
>                         return 1;

Having a few lines of separation between the for () and the loop body
always looks a bit sketch, even if it's just a comment. Wrapping the
block in braces or moving the comment above the loop is probably a
good idea.

Looks fine otherwise
Aneesh Kumar K.V Sept. 17, 2019, 9:54 a.m. UTC | #3
On 9/16/19 11:16 PM, Dan Williams wrote:
> On Mon, Sep 9, 2019 at 11:29 PM Aneesh Kumar K.V
> <aneesh.kumar@linux.ibm.com> wrote:
>>
>> With commit: 7cc7867fb061 ("mm/devm_memremap_pages: enable sub-section remap")
>> pmem namespaces are remapped in 2M chunks. On architectures like ppc64 we
>> can map the memmap area using 16MB hugepage size and that can cover
>> a memory range of 16G.
>>
>> While enabling new pmem namespaces, since memory is added in sub-section chunks,
>> before creating a new memmap mapping, kernel should check whether there is an
>> existing memmap mapping covering the new pmem namespace. Currently, this is
>> validated by checking whether the section covering the range is already
>> initialized or not. Considering there can be multiple namespaces in the same
>> section this can result in wrong validation. Update this to check for
>> sub-sections in the range. This is done by checking for all pfns in the range we
>> are mapping.
>>
>> We could optimize this by checking only just one pfn in each sub-section. But
>> since this is not fast-path we keep this simple.
>>
>> Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
>> ---
>>   arch/powerpc/mm/init_64.c | 45 ++++++++++++++++++++-------------------
>>   1 file changed, 23 insertions(+), 22 deletions(-)
>>
>> diff --git a/arch/powerpc/mm/init_64.c b/arch/powerpc/mm/init_64.c
>> index 4e08246acd79..7710ccdc19a2 100644
>> --- a/arch/powerpc/mm/init_64.c
>> +++ b/arch/powerpc/mm/init_64.c
>> @@ -70,30 +70,24 @@ EXPORT_SYMBOL_GPL(kernstart_addr);
>>
>>   #ifdef CONFIG_SPARSEMEM_VMEMMAP
>>   /*
>> - * Given an address within the vmemmap, determine the pfn of the page that
>> - * represents the start of the section it is within.  Note that we have to
>> - * do this by hand as the proffered address may not be correctly aligned.
>> - * Subtraction of non-aligned pointers produces undefined results.
>> - */
>> -static unsigned long __meminit vmemmap_section_start(unsigned long page)
>> -{
>> -       unsigned long offset = page - ((unsigned long)(vmemmap));
>> -
>> -       /* Return the pfn of the start of the section. */
>> -       return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
>> -}
>> -
>> -/*
>> - * Check if this vmemmap page is already initialised.  If any section
>> + * Check if this vmemmap page is already initialised.  If any sub section
>>    * which overlaps this vmemmap page is initialised then this page is
>>    * initialised already.
>>    */
>> -static int __meminit vmemmap_populated(unsigned long start, int page_size)
>> +
>> +static int __meminit vmemmap_populated(unsigned long start, int size)
>>   {
>> -       unsigned long end = start + page_size;
>> -       start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
>> +       unsigned long end = start + size;
>>
>> -       for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
>> +       /* start is size aligned and it is always > sizeof(struct page) */
>> +       VM_BUG_ON(start & sizeof(struct page));
> 
> If start is size aligned why not include that assumption in the VM_BUG_ON()?
> 
> Otherwise it seems this patch could be reduced simply by:
> 
> s/PAGE_SECTION_MASK/PAGE_SUBSECTION_MASK/
> s/PAGES_PER_SECTION/PAGES_PER_SUBSECTION/
> 
> ...and leave the vmemmap_section_start() function in place? In other
> words this path used to guarantee that 'start' was aligned to the
> minimum mem-hotplug granularity, the change looks ok on the surface,
> but it seems a subtle change in semantics.
> 


How about this?

/*
  * Given an address within the vmemmap, determine the page that
  * represents the start of the subsection it is within.  Note that we 
have to
  * do this by hand as the proffered address may not be correctly aligned.
  * Subtraction of non-aligned pointers produces undefined results.
  */
static struct page * __meminit vmemmap_subsection_start(unsigned long 
vmemmap_addr)
{
	unsigned long start_pfn;
	unsigned long offset = vmemmap_addr - ((unsigned long)(vmemmap));

	/* Return the pfn of the start of the section. */
	start_pfn = (offset / sizeof(struct page)) & PAGE_SUBSECTION_MASK;
	return pfn_to_page(start_pfn);
}

/*
  * Since memory is added in sub-section chunks, before creating a new 
vmemmap
  * mapping, the kernel should check whether there is an existing memmap 
mapping
  * covering the new subsection added. This is needed because kernel can map
  * vmemmap area using 16MB pages which will cover a memory range of 
16G. Such
  * a range covers multiple subsections (2M)
  *
  * If any subsection in the 16G range mapped by vmemmap is valid we 
consider the
  * vmemmap populated (There is a page table entry already present). We 
can't do
  * a page table lookup here because with the hash translation we don't keep
  * vmemmap details in linux page table.
  */
static int __meminit vmemmap_populated(unsigned long vmemmap_addr, int 
vmemmap_map_size)
{
	struct page *start;
	unsigned long vmemmap_end = vmemmap_addr + vmemmap_map_size;
	start = vmemmap_subsection_start(vmemmap_addr);

	for (; (unsigned long)start < vmemmap_end; start += PAGES_PER_SUBSECTION)
		/*
		 * pfn valid check here is intended to really check
		 * whether we have any subsection already initialized
		 * in this range.
		 */
		if (pfn_valid(page_to_pfn(start)))
			return 1;

	return 0;
}




> Can you get an ack from a powerpc maintainer, or maybe this patch
> should route through the powerpc tree?
> 

I guess this can go via powerpc tree. I will send an update V2 patch for 
this alone and work with Michael to pick that up.

> I'll take patch1 through the nvdimm tree.
> 

Thanks.

-aneesh

Patch
diff mbox series

diff --git a/arch/powerpc/mm/init_64.c b/arch/powerpc/mm/init_64.c
index 4e08246acd79..7710ccdc19a2 100644
--- a/arch/powerpc/mm/init_64.c
+++ b/arch/powerpc/mm/init_64.c
@@ -70,30 +70,24 @@  EXPORT_SYMBOL_GPL(kernstart_addr);
 
 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 /*
- * Given an address within the vmemmap, determine the pfn of the page that
- * represents the start of the section it is within.  Note that we have to
- * do this by hand as the proffered address may not be correctly aligned.
- * Subtraction of non-aligned pointers produces undefined results.
- */
-static unsigned long __meminit vmemmap_section_start(unsigned long page)
-{
-	unsigned long offset = page - ((unsigned long)(vmemmap));
-
-	/* Return the pfn of the start of the section. */
-	return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
-}
-
-/*
- * Check if this vmemmap page is already initialised.  If any section
+ * Check if this vmemmap page is already initialised.  If any sub section
  * which overlaps this vmemmap page is initialised then this page is
  * initialised already.
  */
-static int __meminit vmemmap_populated(unsigned long start, int page_size)
+
+static int __meminit vmemmap_populated(unsigned long start, int size)
 {
-	unsigned long end = start + page_size;
-	start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
+	unsigned long end = start + size;
 
-	for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
+	/* start is size aligned and it is always > sizeof(struct page) */
+	VM_BUG_ON(start & sizeof(struct page));
+	for (; start < end; start += sizeof(struct page))
+		/*
+		 * pfn valid check here is intended to really check
+		 * whether we have any subsection already initialized
+		 * in this range. We keep it simple by checking every
+		 * pfn in the range.
+		 */
 		if (pfn_valid(page_to_pfn((struct page *)start)))
 			return 1;
 
@@ -201,6 +195,12 @@  int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
 		void *p = NULL;
 		int rc;
 
+		/*
+		 * This vmemmap range is backing different subsections. If any
+		 * of that subsection is marked valid, that means we already
+		 * have initialized a page table covering this range and hence
+		 * the vmemmap range is populated.
+		 */
 		if (vmemmap_populated(start, page_size))
 			continue;
 
@@ -290,9 +290,10 @@  void __ref vmemmap_free(unsigned long start, unsigned long end,
 		struct page *page;
 
 		/*
-		 * the section has already be marked as invalid, so
-		 * vmemmap_populated() true means some other sections still
-		 * in this page, so skip it.
+		 * We have already marked the subsection we are trying to remove
+		 * invalid. So if we want to remove the vmemmap range, we
+		 * need to make sure there is no subsection marked valid
+		 * in this range.
 		 */
 		if (vmemmap_populated(start, page_size))
 			continue;