All of lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH] resource: re-factor page_is_ram()
@ 2022-06-01 16:32 Vaibhav Jain
  2022-06-10 10:23 ` David Hildenbrand
  2022-07-18 10:48 ` David Hildenbrand
  0 siblings, 2 replies; 6+ messages in thread
From: Vaibhav Jain @ 2022-06-01 16:32 UTC (permalink / raw)
  To: linux-kernel
  Cc: Vaibhav Jain, Aneesh Kumar K . V, Andrew Morton,
	David Hildenbrand, Dan Williams, Miaohe Lin, Muchun Song

Presently page_is_ram() relies on walk_system_ram_range() that performs a walk
on kernel iomem resources hierarchy with a dummy callback __is_ram(). Before
calling find_next_iomem_res(), walk_system_ram_range() does some book-keeping
which can be avoided for page_is_ram() use-case.

Hence this patch proposes to update page_is_ram() to directly call
find_next_iomem_res() with minimal book-keeping needed.

To avoid allocating a 'struct resource' the patch also updates
find_next_iomem_res() to not return -EINVAL in case 'res == NULL'. Instead
out 'struct resource *res' is only populated when its not NULL.

Signed-off-by: Vaibhav Jain <vaibhav@linux.ibm.com>
---
 kernel/resource.c | 19 ++++++++-----------
 1 file changed, 8 insertions(+), 11 deletions(-)

diff --git a/kernel/resource.c b/kernel/resource.c
index 34eaee179689..ecf6b9a50adc 100644
--- a/kernel/resource.c
+++ b/kernel/resource.c
@@ -311,7 +311,7 @@ EXPORT_SYMBOL(release_resource);
  *
  * If a resource is found, returns 0 and @*res is overwritten with the part
  * of the resource that's within [@start..@end]; if none is found, returns
- * -ENODEV.  Returns -EINVAL for invalid parameters.
+ * -ENODEV.
  *
  * @start:	start address of the resource searched for
  * @end:	end address of same resource
@@ -328,9 +328,6 @@ static int find_next_iomem_res(resource_size_t start, resource_size_t end,
 {
 	struct resource *p;
 
-	if (!res)
-		return -EINVAL;
-
 	if (start >= end)
 		return -EINVAL;
 
@@ -356,7 +353,7 @@ static int find_next_iomem_res(resource_size_t start, resource_size_t end,
 		break;
 	}
 
-	if (p) {
+	if (p && res) {
 		/* copy data */
 		*res = (struct resource) {
 			.start = max(start, p->start),
@@ -474,18 +471,18 @@ int walk_system_ram_range(unsigned long start_pfn, unsigned long nr_pages,
 	return ret;
 }
 
-static int __is_ram(unsigned long pfn, unsigned long nr_pages, void *arg)
-{
-	return 1;
-}
-
 /*
  * This generic page_is_ram() returns true if specified address is
  * registered as System RAM in iomem_resource list.
  */
 int __weak page_is_ram(unsigned long pfn)
 {
-	return walk_system_ram_range(pfn, 1, NULL, __is_ram) == 1;
+	const resource_size_t pfn_res = PFN_PHYS(pfn);
+
+	return find_next_iomem_res(pfn_res,
+				   pfn_res + 1,
+				   IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY,
+				   IORES_DESC_NONE, NULL) == 0;
 }
 EXPORT_SYMBOL_GPL(page_is_ram);
 
-- 
2.35.3


^ permalink raw reply related	[flat|nested] 6+ messages in thread

* Re: [PATCH] resource: re-factor page_is_ram()
  2022-06-01 16:32 [PATCH] resource: re-factor page_is_ram() Vaibhav Jain
@ 2022-06-10 10:23 ` David Hildenbrand
  2022-06-16  6:45   ` Vaibhav Jain
  2022-07-18 10:48 ` David Hildenbrand
  1 sibling, 1 reply; 6+ messages in thread
From: David Hildenbrand @ 2022-06-10 10:23 UTC (permalink / raw)
  To: Vaibhav Jain, linux-kernel
  Cc: Aneesh Kumar K . V, Andrew Morton, Dan Williams, Miaohe Lin, Muchun Song

On 01.06.22 18:32, Vaibhav Jain wrote:
> Presently page_is_ram() relies on walk_system_ram_range() that performs a walk
> on kernel iomem resources hierarchy with a dummy callback __is_ram(). Before
> calling find_next_iomem_res(), walk_system_ram_range() does some book-keeping
> which can be avoided for page_is_ram() use-case.
> 
> Hence this patch proposes to update page_is_ram() to directly call
> find_next_iomem_res() with minimal book-keeping needed.

I consider the code harder to get compared to just reusing the
more-generic and expressive walk_system_ram_range().

It somehow feels like we're duplicating the code here just to optimize
out a handful of instructions.

If it doesn't make the code easier to read (at least for me), why do we
care?

-- 
Thanks,

David / dhildenb


^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [PATCH] resource: re-factor page_is_ram()
  2022-06-10 10:23 ` David Hildenbrand
@ 2022-06-16  6:45   ` Vaibhav Jain
  2022-07-18 10:42     ` David Hildenbrand
  0 siblings, 1 reply; 6+ messages in thread
From: Vaibhav Jain @ 2022-06-16  6:45 UTC (permalink / raw)
  To: David Hildenbrand, linux-kernel
  Cc: Aneesh Kumar K . V, Andrew Morton, Dan Williams, Miaohe Lin, Muchun Song

Hi David,

Thanks for looking into this patch,

David Hildenbrand <david@redhat.com> writes:

> On 01.06.22 18:32, Vaibhav Jain wrote:
>> Presently page_is_ram() relies on walk_system_ram_range() that performs a walk
>> on kernel iomem resources hierarchy with a dummy callback __is_ram(). Before
>> calling find_next_iomem_res(), walk_system_ram_range() does some book-keeping
>> which can be avoided for page_is_ram() use-case.
>> 
>> Hence this patch proposes to update page_is_ram() to directly call
>> find_next_iomem_res() with minimal book-keeping needed.
>
> I consider the code harder to get compared to just reusing the
> more-generic and expressive walk_system_ram_range()
>
> It somehow feels like we're duplicating the code here just to optimize
> out a handful of instructions.

The only reason for existence of dummy callback __is_ram() is for
page_is_ram() to be able to use walk_system_ram_range(). For
page_is_ram() usecase what walk_system_ram_range() essentially does is
to iterate over find_next_iomem_res() and call __is_ram() which is not
really needed to page_is_ram().

The improvement to the gcc (v12.1.1) generated code (x86_64) for
page_is_ram is quite evident.

With the patch:
   0x0000000000000920 <+0>:	call   0x925 <page_is_ram+5>
   0x0000000000000925 <+5>:	shl    $0xc,%rdi
   0x0000000000000929 <+9>:	xor    %r8d,%r8d
   0x000000000000092c <+12>:	xor    %ecx,%ecx
   0x000000000000092e <+14>:	mov    $0x81000200,%edx
   0x0000000000000933 <+19>:	lea    0x1(%rdi),%rsi
   0x0000000000000937 <+23>:	call   0x7e0 <find_next_iomem_res>
   0x000000000000093c <+28>:	test   %eax,%eax
   0x000000000000093e <+30>:	sete   %al
   0x0000000000000941 <+33>:	movzbl %al,%eax
   0x0000000000000944 <+36>:	ret
   0x0000000000000945 <+37>:	int3

Without the patch:
   0x0000000000001000 <+0>:	call   0x1005 <page_is_ram+5>
   0x0000000000001005 <+5>:	shl    $0xc,%rdi
   0x0000000000001009 <+9>:	lea    0xfff(%rdi),%rsi
   0x0000000000001010 <+16>:	cmp    %rsi,%rdi
   0x0000000000001013 <+19>:	jae    0x1064 <page_is_ram+100>
   0x0000000000001015 <+21>:	sub    $0x40,%rsp
   0x0000000000001019 <+25>:	xor    %ecx,%ecx
   0x000000000000101b <+27>:	mov    $0x81000200,%edx
   0x0000000000001020 <+32>:	mov    %rsp,%r8
   0x0000000000001023 <+35>:	call   0x7e0 <find_next_iomem_res>
   0x0000000000001028 <+40>:	test   %eax,%eax
   0x000000000000102a <+42>:	jne    0x105a <page_is_ram+90>
   0x000000000000102c <+44>:	mov    (%rsp),%rax
   0x0000000000001030 <+48>:	mov    $0x1,%ecx
   0x0000000000001035 <+53>:	lea    0xfff(%rax),%rdx
   0x000000000000103c <+60>:	mov    0x8(%rsp),%rax
   0x0000000000001041 <+65>:	shr    $0xc,%rdx
   0x0000000000001045 <+69>:	add    $0x1,%rax
   0x0000000000001049 <+73>:	shr    $0xc,%rax
   0x000000000000104d <+77>:	cmp    %rax,%rdx
   0x0000000000001050 <+80>:	jae    0x105a <page_is_ram+90>
   0x0000000000001052 <+82>:	mov    %ecx,%eax
   0x0000000000001054 <+84>:	add    $0x40,%rsp
   0x0000000000001058 <+88>:	ret
   0x0000000000001059 <+89>:	int3
   0x000000000000105a <+90>:	xor    %ecx,%ecx
   0x000000000000105c <+92>:	add    $0x40,%rsp
   0x0000000000001060 <+96>:	mov    %ecx,%eax
   0x0000000000001062 <+98>:	ret
   0x0000000000001063 <+99>:	int3
   0x0000000000001064 <+100>:	xor    %eax,%eax
   0x0000000000001066 <+102>:	ret
   0x0000000000001067 <+103>:	int3

Looking at the disassembly above, gcc has inlined both walk_system_ram_range()
and __is_ram() in page_is_ram(). This ends up in page_is_ram() calling
find_next_iomem_res() directly anyways with bunch of book-keeping
afterwards which can be avoided.

>
> If it doesn't make the code easier to read (at least for me), why do we
> care?
IMHO, calling find_next_iomem_res() from page_is_ram() instead of
calling walk_system_ram_range() makes it easy to trace the path of
page_is_ram(). Also the dummy callback makes the code flow seems strange
initially.

-- 
Cheers
~ Vaibhav

^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [PATCH] resource: re-factor page_is_ram()
  2022-06-16  6:45   ` Vaibhav Jain
@ 2022-07-18 10:42     ` David Hildenbrand
  0 siblings, 0 replies; 6+ messages in thread
From: David Hildenbrand @ 2022-07-18 10:42 UTC (permalink / raw)
  To: Vaibhav Jain, linux-kernel
  Cc: Aneesh Kumar K . V, Andrew Morton, Dan Williams, Miaohe Lin, Muchun Song

[sorry for the late reply]

> 
> The improvement to the gcc (v12.1.1) generated code (x86_64) for
> page_is_ram is quite evident.
> 
> With the patch:
>    0x0000000000000920 <+0>:	call   0x925 <page_is_ram+5>
>    0x0000000000000925 <+5>:	shl    $0xc,%rdi
>    0x0000000000000929 <+9>:	xor    %r8d,%r8d
>    0x000000000000092c <+12>:	xor    %ecx,%ecx
>    0x000000000000092e <+14>:	mov    $0x81000200,%edx
>    0x0000000000000933 <+19>:	lea    0x1(%rdi),%rsi
>    0x0000000000000937 <+23>:	call   0x7e0 <find_next_iomem_res>
>    0x000000000000093c <+28>:	test   %eax,%eax
>    0x000000000000093e <+30>:	sete   %al
>    0x0000000000000941 <+33>:	movzbl %al,%eax
>    0x0000000000000944 <+36>:	ret
>    0x0000000000000945 <+37>:	int3
> 
> Without the patch:
>    0x0000000000001000 <+0>:	call   0x1005 <page_is_ram+5>
>    0x0000000000001005 <+5>:	shl    $0xc,%rdi
>    0x0000000000001009 <+9>:	lea    0xfff(%rdi),%rsi
>    0x0000000000001010 <+16>:	cmp    %rsi,%rdi
>    0x0000000000001013 <+19>:	jae    0x1064 <page_is_ram+100>
>    0x0000000000001015 <+21>:	sub    $0x40,%rsp
>    0x0000000000001019 <+25>:	xor    %ecx,%ecx
>    0x000000000000101b <+27>:	mov    $0x81000200,%edx
>    0x0000000000001020 <+32>:	mov    %rsp,%r8
>    0x0000000000001023 <+35>:	call   0x7e0 <find_next_iomem_res>
>    0x0000000000001028 <+40>:	test   %eax,%eax
>    0x000000000000102a <+42>:	jne    0x105a <page_is_ram+90>
>    0x000000000000102c <+44>:	mov    (%rsp),%rax
>    0x0000000000001030 <+48>:	mov    $0x1,%ecx
>    0x0000000000001035 <+53>:	lea    0xfff(%rax),%rdx
>    0x000000000000103c <+60>:	mov    0x8(%rsp),%rax
>    0x0000000000001041 <+65>:	shr    $0xc,%rdx
>    0x0000000000001045 <+69>:	add    $0x1,%rax
>    0x0000000000001049 <+73>:	shr    $0xc,%rax
>    0x000000000000104d <+77>:	cmp    %rax,%rdx
>    0x0000000000001050 <+80>:	jae    0x105a <page_is_ram+90>
>    0x0000000000001052 <+82>:	mov    %ecx,%eax
>    0x0000000000001054 <+84>:	add    $0x40,%rsp
>    0x0000000000001058 <+88>:	ret
>    0x0000000000001059 <+89>:	int3
>    0x000000000000105a <+90>:	xor    %ecx,%ecx
>    0x000000000000105c <+92>:	add    $0x40,%rsp
>    0x0000000000001060 <+96>:	mov    %ecx,%eax
>    0x0000000000001062 <+98>:	ret
>    0x0000000000001063 <+99>:	int3
>    0x0000000000001064 <+100>:	xor    %eax,%eax
>    0x0000000000001066 <+102>:	ret
>    0x0000000000001067 <+103>:	int3
> 
> Looking at the disassembly above, gcc has inlined both walk_system_ram_range()
> and __is_ram() in page_is_ram(). This ends up in page_is_ram() calling
> find_next_iomem_res() directly anyways with bunch of book-keeping
> afterwards which can be avoided.

We usually don't care about such micro-optimizations unless you can
showcase actual performance numbers. Otherwise we'd have constant,
unnecessary code-churn all over the place.

Most probably, all that list walking dominates the runtime either way.

Feel free to proof me wrong ;)

> >>
>> If it doesn't make the code easier to read (at least for me), why do we
>> care?
> IMHO, calling find_next_iomem_res() from page_is_ram() instead of
> calling walk_system_ram_range() makes it easy to trace the path of
> page_is_ram(). Also the dummy callback makes the code flow seems strange
> initially.
> 

I'm not convinced, but I don't care enough to object. I'll add more
review feedback to the patch.

-- 
Thanks,

David / dhildenb


^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [PATCH] resource: re-factor page_is_ram()
  2022-06-01 16:32 [PATCH] resource: re-factor page_is_ram() Vaibhav Jain
  2022-06-10 10:23 ` David Hildenbrand
@ 2022-07-18 10:48 ` David Hildenbrand
  2022-07-18 18:00   ` Dan Williams
  1 sibling, 1 reply; 6+ messages in thread
From: David Hildenbrand @ 2022-07-18 10:48 UTC (permalink / raw)
  To: Vaibhav Jain, linux-kernel
  Cc: Aneesh Kumar K . V, Andrew Morton, Dan Williams, Miaohe Lin, Muchun Song

On 01.06.22 18:32, Vaibhav Jain wrote:
> Presently page_is_ram() relies on walk_system_ram_range() that performs a walk
> on kernel iomem resources hierarchy with a dummy callback __is_ram(). Before
> calling find_next_iomem_res(), walk_system_ram_range() does some book-keeping
> which can be avoided for page_is_ram() use-case.
> 
> Hence this patch proposes to update page_is_ram() to directly call
> find_next_iomem_res() with minimal book-keeping needed.
> 
> To avoid allocating a 'struct resource' the patch also updates
> find_next_iomem_res() to not return -EINVAL in case 'res == NULL'. Instead
> out 'struct resource *res' is only populated when its not NULL.
> 
> Signed-off-by: Vaibhav Jain <vaibhav@linux.ibm.com>
> ---
>  kernel/resource.c | 19 ++++++++-----------
>  1 file changed, 8 insertions(+), 11 deletions(-)
> 
> diff --git a/kernel/resource.c b/kernel/resource.c
> index 34eaee179689..ecf6b9a50adc 100644
> --- a/kernel/resource.c
> +++ b/kernel/resource.c
> @@ -311,7 +311,7 @@ EXPORT_SYMBOL(release_resource);
>   *
>   * If a resource is found, returns 0 and @*res is overwritten with the part
>   * of the resource that's within [@start..@end]; if none is found, returns
> - * -ENODEV.  Returns -EINVAL for invalid parameters.
> + * -ENODEV.
>   *

There is still another -EINVAL in that function ...

>   * @start:	start address of the resource searched for
>   * @end:	end address of same resource
> @@ -328,9 +328,6 @@ static int find_next_iomem_res(resource_size_t start, resource_size_t end,
>  {
>  	struct resource *p;
>  
> -	if (!res)
> -		return -EINVAL;
> -
>  	if (start >= end)
>  		return -EINVAL;

As all callers guarantee that, we might just remove it.

>  
> @@ -356,7 +353,7 @@ static int find_next_iomem_res(resource_size_t start, resource_size_t end,
>  		break;
>  	}
>  
> -	if (p) {
> +	if (p && res) {
>  		/* copy data */
>  		*res = (struct resource) {
>  			.start = max(start, p->start),
> @@ -474,18 +471,18 @@ int walk_system_ram_range(unsigned long start_pfn, unsigned long nr_pages,
>  	return ret;
>  }
>  
> -static int __is_ram(unsigned long pfn, unsigned long nr_pages, void *arg)
> -{
> -	return 1;
> -}
> -
>  /*
>   * This generic page_is_ram() returns true if specified address is
>   * registered as System RAM in iomem_resource list.
>   */
>  int __weak page_is_ram(unsigned long pfn)
>  {
> -	return walk_system_ram_range(pfn, 1, NULL, __is_ram) == 1;
> +	const resource_size_t pfn_res = PFN_PHYS(pfn);
> +
> +	return find_next_iomem_res(pfn_res,
> +				   pfn_res + 1,
> +				   IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY,
> +				   IORES_DESC_NONE, NULL) == 0;
>  }
>  EXPORT_SYMBOL_GPL(page_is_ram);
>  

What about

a) A cleanup patch upfront that removes both -EINVAL cases from
find_next_iomem_res() followed by

b) The actual change to page_is_ram()

?

-- 
Thanks,

David / dhildenb


^ permalink raw reply	[flat|nested] 6+ messages in thread

* Re: [PATCH] resource: re-factor page_is_ram()
  2022-07-18 10:48 ` David Hildenbrand
@ 2022-07-18 18:00   ` Dan Williams
  0 siblings, 0 replies; 6+ messages in thread
From: Dan Williams @ 2022-07-18 18:00 UTC (permalink / raw)
  To: David Hildenbrand, Vaibhav Jain, linux-kernel
  Cc: Aneesh Kumar K . V, Andrew Morton, Dan Williams, Miaohe Lin, Muchun Song

David Hildenbrand wrote:
> On 01.06.22 18:32, Vaibhav Jain wrote:
> > Presently page_is_ram() relies on walk_system_ram_range() that performs a walk
> > on kernel iomem resources hierarchy with a dummy callback __is_ram(). Before
> > calling find_next_iomem_res(), walk_system_ram_range() does some book-keeping
> > which can be avoided for page_is_ram() use-case.
> > 
> > Hence this patch proposes to update page_is_ram() to directly call
> > find_next_iomem_res() with minimal book-keeping needed.
> > 
> > To avoid allocating a 'struct resource' the patch also updates
> > find_next_iomem_res() to not return -EINVAL in case 'res == NULL'. Instead
> > out 'struct resource *res' is only populated when its not NULL.
> > 
> > Signed-off-by: Vaibhav Jain <vaibhav@linux.ibm.com>
> > ---
> >  kernel/resource.c | 19 ++++++++-----------
> >  1 file changed, 8 insertions(+), 11 deletions(-)
> > 
> > diff --git a/kernel/resource.c b/kernel/resource.c
> > index 34eaee179689..ecf6b9a50adc 100644
> > --- a/kernel/resource.c
> > +++ b/kernel/resource.c
> > @@ -311,7 +311,7 @@ EXPORT_SYMBOL(release_resource);
> >   *
> >   * If a resource is found, returns 0 and @*res is overwritten with the part
> >   * of the resource that's within [@start..@end]; if none is found, returns
> > - * -ENODEV.  Returns -EINVAL for invalid parameters.
> > + * -ENODEV.
> >   *
> 
> There is still another -EINVAL in that function ...
> 
> >   * @start:	start address of the resource searched for
> >   * @end:	end address of same resource
> > @@ -328,9 +328,6 @@ static int find_next_iomem_res(resource_size_t start, resource_size_t end,
> >  {
> >  	struct resource *p;
> >  
> > -	if (!res)
> > -		return -EINVAL;
> > -
> >  	if (start >= end)
> >  		return -EINVAL;
> 
> As all callers guarantee that, we might just remove it.
> 
> >  
> > @@ -356,7 +353,7 @@ static int find_next_iomem_res(resource_size_t start, resource_size_t end,
> >  		break;
> >  	}
> >  
> > -	if (p) {
> > +	if (p && res) {
> >  		/* copy data */
> >  		*res = (struct resource) {
> >  			.start = max(start, p->start),
> > @@ -474,18 +471,18 @@ int walk_system_ram_range(unsigned long start_pfn, unsigned long nr_pages,
> >  	return ret;
> >  }
> >  
> > -static int __is_ram(unsigned long pfn, unsigned long nr_pages, void *arg)
> > -{
> > -	return 1;
> > -}
> > -
> >  /*
> >   * This generic page_is_ram() returns true if specified address is
> >   * registered as System RAM in iomem_resource list.
> >   */
> >  int __weak page_is_ram(unsigned long pfn)
> >  {
> > -	return walk_system_ram_range(pfn, 1, NULL, __is_ram) == 1;
> > +	const resource_size_t pfn_res = PFN_PHYS(pfn);
> > +
> > +	return find_next_iomem_res(pfn_res,
> > +				   pfn_res + 1,
> > +				   IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY,
> > +				   IORES_DESC_NONE, NULL) == 0;

I tend to agree with David that this change makes the page_is_ram()
harder to read. I think the problem is that the "next" nature of
find_next_iomem_res() is meant to be handled by the caller. So it really
should be called find_iomem_res().

> >  }
> >  EXPORT_SYMBOL_GPL(page_is_ram);
> >  
> 
> What about
> 
> a) A cleanup patch upfront that removes both -EINVAL cases from
> find_next_iomem_res() followed by

...a patch to rename find_next_iomem_res()

> 
> b) The actual change to page_is_ram()
> 
> ?

^ permalink raw reply	[flat|nested] 6+ messages in thread

end of thread, other threads:[~2022-07-18 18:00 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-06-01 16:32 [PATCH] resource: re-factor page_is_ram() Vaibhav Jain
2022-06-10 10:23 ` David Hildenbrand
2022-06-16  6:45   ` Vaibhav Jain
2022-07-18 10:42     ` David Hildenbrand
2022-07-18 10:48 ` David Hildenbrand
2022-07-18 18:00   ` Dan Williams

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.