Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

[Matthew Wilcox]

On Fri, May 11, 2018 at 12:10:25AM +0800, Huaisheng Ye wrote:
> -#define __GFP_DMA	((__force gfp_t)___GFP_DMA)
> -#define __GFP_HIGHMEM	((__force gfp_t)___GFP_HIGHMEM)
> -#define __GFP_DMA32	((__force gfp_t)___GFP_DMA32)
> +#define __GFP_DMA	((__force gfp_t)OPT_ZONE_DMA ^ ZONE_NORMAL)
> +#define __GFP_HIGHMEM	((__force gfp_t)ZONE_MOVABLE ^ ZONE_NORMAL)
> +#define __GFP_DMA32	((__force gfp_t)OPT_ZONE_DMA32 ^ ZONE_NORMAL)

No, you've made gfp_zone even more complex than it already is.
If you can't use OPT_ZONE_HIGHMEM here, then this is a waste of time.

>  static inline enum zone_type gfp_zone(gfp_t flags)
>  {
>  	enum zone_type z;
> -	int bit = (__force int) (flags & GFP_ZONEMASK);
> +	z = ((__force unsigned int)flags & ___GFP_ZONE_MASK) ^ ZONE_NORMAL;
> +
> +	if (z > OPT_ZONE_HIGHMEM)
> +		z = OPT_ZONE_HIGHMEM +
> +			!!((__force unsigned int)flags & ___GFP_MOVABLE);
>  
> -	z = (GFP_ZONE_TABLE >> (bit * GFP_ZONES_SHIFT)) &
> -					 ((1 << GFP_ZONES_SHIFT) - 1);
> -	VM_BUG_ON((GFP_ZONE_BAD >> bit) & 1);
> +	VM_BUG_ON(z > ZONE_MOVABLE);
>  	return z;
>  }

RE: [External] Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

[Huaisheng HS1 Ye]

> From: owner-linux-mm@kvack.org [mailto:owner-linux-mm@kvack.org] On Behalf Of Matthew
> Wilcox
> On Fri, May 11, 2018 at 12:10:25AM +0800, Huaisheng Ye wrote:
> > -#define __GFP_DMA	((__force gfp_t)___GFP_DMA)
> > -#define __GFP_HIGHMEM	((__force gfp_t)___GFP_HIGHMEM)
> > -#define __GFP_DMA32	((__force gfp_t)___GFP_DMA32)
> > +#define __GFP_DMA	((__force gfp_t)OPT_ZONE_DMA ^ ZONE_NORMAL)
> > +#define __GFP_HIGHMEM	((__force gfp_t)ZONE_MOVABLE ^ ZONE_NORMAL)
> > +#define __GFP_DMA32	((__force gfp_t)OPT_ZONE_DMA32 ^ ZONE_NORMAL)
> 
> No, you've made gfp_zone even more complex than it already is.
> If you can't use OPT_ZONE_HIGHMEM here, then this is a waste of time.
> 
Dear Matthew,

The reason why I don't use OPT_ZONE_HIGHMEM for __GFP_HIGHMEM	 directly is that, for x86_64 platform there is no CONFIG_HIGHMEM, so OPT_ZONE_HIGHMEM shall always be equal to ZONE_NORMAL.

For gfp_zone it is impossible to distinguish the meaning of lowest 3 bits in flags. How can gfp_zone to understand it comes from OPT_ZONE_HIGHMEM or ZONE_NORMAL?
And the most pained thing is that, if __GFP_HIGHMEM with movable flag enabled, it means that ZONE_MOVABLE shall be returned.
That is different from ZONE_DMA, ZONE_DMA32 and ZONE_NORMAL.

I was thinking...
Whether it is possible to use other judgement condition to decide OPT_ZONE_HIGHMEM or ZONE_MOVABLE shall be returned from gfp_zone.

Sincerely,
Huaisheng Ye


> >  static inline enum zone_type gfp_zone(gfp_t flags)
> >  {
> >  	enum zone_type z;
> > -	int bit = (__force int) (flags & GFP_ZONEMASK);
> > +	z = ((__force unsigned int)flags & ___GFP_ZONE_MASK) ^ ZONE_NORMAL;
> > +
> > +	if (z > OPT_ZONE_HIGHMEM)
> > +		z = OPT_ZONE_HIGHMEM +
> > +			!!((__force unsigned int)flags & ___GFP_MOVABLE);
> >
> > -	z = (GFP_ZONE_TABLE >> (bit * GFP_ZONES_SHIFT)) &
> > -					 ((1 << GFP_ZONES_SHIFT) - 1);
> > -	VM_BUG_ON((GFP_ZONE_BAD >> bit) & 1);
> > +	VM_BUG_ON(z > ZONE_MOVABLE);
> >  	return z;
> >  }

Re: [External] Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

[Matthew Wilcox]

On Fri, May 11, 2018 at 03:24:34AM +0000, Huaisheng HS1 Ye wrote:
> > From: owner-linux-mm@kvack.org [mailto:owner-linux-mm@kvack.org] On Behalf Of Matthew
> > Wilcox
> > On Fri, May 11, 2018 at 12:10:25AM +0800, Huaisheng Ye wrote:
> > > -#define __GFP_DMA	((__force gfp_t)___GFP_DMA)
> > > -#define __GFP_HIGHMEM	((__force gfp_t)___GFP_HIGHMEM)
> > > -#define __GFP_DMA32	((__force gfp_t)___GFP_DMA32)
> > > +#define __GFP_DMA	((__force gfp_t)OPT_ZONE_DMA ^ ZONE_NORMAL)
> > > +#define __GFP_HIGHMEM	((__force gfp_t)ZONE_MOVABLE ^ ZONE_NORMAL)
> > > +#define __GFP_DMA32	((__force gfp_t)OPT_ZONE_DMA32 ^ ZONE_NORMAL)
> > 
> > No, you've made gfp_zone even more complex than it already is.
> > If you can't use OPT_ZONE_HIGHMEM here, then this is a waste of time.
> > 
> Dear Matthew,
> 
> The reason why I don't use OPT_ZONE_HIGHMEM for __GFP_HIGHMEM	 directly is that, for x86_64 platform there is no CONFIG_HIGHMEM, so OPT_ZONE_HIGHMEM shall always be equal to ZONE_NORMAL.

Right.  On 64-bit platforms, if somebody asks for HIGHMEM, they should
get NORMAL pages.

> For gfp_zone it is impossible to distinguish the meaning of lowest 3 bits in flags. How can gfp_zone to understand it comes from OPT_ZONE_HIGHMEM or ZONE_NORMAL?
> And the most pained thing is that, if __GFP_HIGHMEM with movable flag enabled, it means that ZONE_MOVABLE shall be returned.
> That is different from ZONE_DMA, ZONE_DMA32 and ZONE_NORMAL.

The point of this exercise is to actually encode the zone number in
the bottom bits of the GFP flags instead of something which has to be
interpreted into a zone number.  When somebody sets __GFP_MOVABLE, they
should also be setting ZONE_MOVABLE:

-#define __GFP_MOVABLE   ((__force gfp_t)___GFP_MOVABLE)  /* ZONE_MOVABLE allowed */
+#define __GFP_MOVABLE   ((__force gfp_t)(___GFP_MOVABLE | (ZONE_MOVABLE ^ ZONE_NORMAL)))

One thing that does need to change is:

-#define GFP_HIGHUSER_MOVABLE    (GFP_HIGHUSER | __GFP_MOVABLE)
+#define GFP_HIGHUSER_MOVABLE    (GFP_USER | __GFP_MOVABLE)

otherwise we'll be OR'ing ZONE_MOVABLE and ZONE_HIGHMEM together.

> I was thinking...
> Whether it is possible to use other judgement condition to decide OPT_ZONE_HIGHMEM or ZONE_MOVABLE shall be returned from gfp_zone.
> 
> Sincerely,
> Huaisheng Ye
> 
> 
> > >  static inline enum zone_type gfp_zone(gfp_t flags)
> > >  {
> > >  	enum zone_type z;
> > > -	int bit = (__force int) (flags & GFP_ZONEMASK);
> > > +	z = ((__force unsigned int)flags & ___GFP_ZONE_MASK) ^ ZONE_NORMAL;
> > > +
> > > +	if (z > OPT_ZONE_HIGHMEM)
> > > +		z = OPT_ZONE_HIGHMEM +
> > > +			!!((__force unsigned int)flags & ___GFP_MOVABLE);
> > >
> > > -	z = (GFP_ZONE_TABLE >> (bit * GFP_ZONES_SHIFT)) &
> > > -					 ((1 << GFP_ZONES_SHIFT) - 1);
> > > -	VM_BUG_ON((GFP_ZONE_BAD >> bit) & 1);
> > > +	VM_BUG_ON(z > ZONE_MOVABLE);
> > >  	return z;
> > >  }
> 

RE: [External] Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

[Huaisheng HS1 Ye]

> From: Matthew Wilcox [mailto:willy@infradead.org]
> Sent: Friday, May 11, 2018 9:26 PM> 
> On Fri, May 11, 2018 at 03:24:34AM +0000, Huaisheng HS1 Ye wrote:
> > > From: owner-linux-mm@kvack.org [mailto:owner-linux-mm@kvack.org] On Behalf Of
> Matthew
> > > Wilcox
> > > On Fri, May 11, 2018 at 12:10:25AM +0800, Huaisheng Ye wrote:
> > > > -#define __GFP_DMA	((__force gfp_t)___GFP_DMA)
> > > > -#define __GFP_HIGHMEM	((__force gfp_t)___GFP_HIGHMEM)
> > > > -#define __GFP_DMA32	((__force gfp_t)___GFP_DMA32)
> > > > +#define __GFP_DMA	((__force gfp_t)OPT_ZONE_DMA ^ ZONE_NORMAL)
> > > > +#define __GFP_HIGHMEM	((__force gfp_t)ZONE_MOVABLE ^ ZONE_NORMAL)
> > > > +#define __GFP_DMA32	((__force gfp_t)OPT_ZONE_DMA32 ^ ZONE_NORMAL)
> > >
> > > No, you've made gfp_zone even more complex than it already is.
> > > If you can't use OPT_ZONE_HIGHMEM here, then this is a waste of time.
> > >
> > Dear Matthew,
> >
> > The reason why I don't use OPT_ZONE_HIGHMEM for __GFP_HIGHMEM	 directly is that,
> for x86_64 platform there is no CONFIG_HIGHMEM, so OPT_ZONE_HIGHMEM shall always be
> equal to ZONE_NORMAL.
> 
> Right.  On 64-bit platforms, if somebody asks for HIGHMEM, they should
> get NORMAL pages.
> 
> > For gfp_zone it is impossible to distinguish the meaning of lowest 3 bits in flags.
> How can gfp_zone to understand it comes from OPT_ZONE_HIGHMEM or ZONE_NORMAL?
> > And the most pained thing is that, if __GFP_HIGHMEM with movable flag enabled, it
> means that ZONE_MOVABLE shall be returned.
> > That is different from ZONE_DMA, ZONE_DMA32 and ZONE_NORMAL.
> 
> The point of this exercise is to actually encode the zone number in
> the bottom bits of the GFP flags instead of something which has to be
> interpreted into a zone number.  When somebody sets __GFP_MOVABLE, they
> should also be setting ZONE_MOVABLE:
> 
> -#define __GFP_MOVABLE   ((__force gfp_t)___GFP_MOVABLE)  /* ZONE_MOVABLE allowed */
> +#define __GFP_MOVABLE   ((__force gfp_t)(___GFP_MOVABLE | (ZONE_MOVABLE ^ ZONE_NORMAL)))
> 
I am afraid we couldn't do that, because __GFP_MOVABLE would be used potentially with other __GFPs like __GFP_DMA and __GFP_DMA32.
Let's go back to the previous example.
We assume ZONE_DMA equals to 0, and ZONE_DMA32 equals to 1. After encoding with ZONE_NORMAL (which equals to 2), we could get that.

#define __GFP_DMA		((__force gfp_t)OPT_ZONE_DMA ^ ZONE_NORMAL)
#define __GFP_DMA32	((__force gfp_t)OPT_ZONE_DMA32 ^ ZONE_NORMAL)
__GPF_DMA	= 0b 0010
__GPF_DMA32	= 0b 0011

We assume ZONE_MOVABLE equals to 3,
#define __GFP_MOVABLE   ((__force gfp_t)(___GFP_MOVABLE | (ZONE_MOVABLE ^ ZONE_NORMAL)))
__GFP_MOVABLE = 0b 1001

If we OR'ing __GFP_MOVABLE and either __GFP_DMA or __GFP_DMA32, we could get same result as '0b 1011'.
This is unacceptable, because inline function gfp_zone couldn't distinguish that is a request of ZONE_DMA or ZONE_DMA32 from parameter flags.

Once more, I think if we want to encode ZONE_MOVABLE to __GFP_MOVABLE, then the operation of __GFP_MOVABLE OR'ing with any other __GFP* would have risk.

Sincerely,
Huaisheng Ye

> One thing that does need to change is:
> 
> -#define GFP_HIGHUSER_MOVABLE    (GFP_HIGHUSER | __GFP_MOVABLE)
> +#define GFP_HIGHUSER_MOVABLE    (GFP_USER | __GFP_MOVABLE)
> 
> otherwise we'll be OR'ing ZONE_MOVABLE and ZONE_HIGHMEM together.
> 
> > I was thinking...
> > Whether it is possible to use other judgement condition to decide OPT_ZONE_HIGHMEM
> or ZONE_MOVABLE shall be returned from gfp_zone.
> >
> > Sincerely,
> > Huaisheng Ye
> >
> >
> > > >  static inline enum zone_type gfp_zone(gfp_t flags)
> > > >  {
> > > >  	enum zone_type z;
> > > > -	int bit = (__force int) (flags & GFP_ZONEMASK);
> > > > +	z = ((__force unsigned int)flags & ___GFP_ZONE_MASK) ^ ZONE_NORMAL;
> > > > +
> > > > +	if (z > OPT_ZONE_HIGHMEM)
> > > > +		z = OPT_ZONE_HIGHMEM +
> > > > +			!!((__force unsigned int)flags & ___GFP_MOVABLE);
> > > >
> > > > -	z = (GFP_ZONE_TABLE >> (bit * GFP_ZONES_SHIFT)) &
> > > > -					 ((1 << GFP_ZONES_SHIFT) - 1);
> > > > -	VM_BUG_ON((GFP_ZONE_BAD >> bit) & 1);
> > > > +	VM_BUG_ON(z > ZONE_MOVABLE);
> > > >  	return z;
> > > >  }
> >

Re: [External] Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

[Matthew Wilcox]

On Sat, May 12, 2018 at 11:35:00AM +0000, Huaisheng HS1 Ye wrote:
> > The point of this exercise is to actually encode the zone number in
> > the bottom bits of the GFP flags instead of something which has to be
> > interpreted into a zone number.  When somebody sets __GFP_MOVABLE, they
> > should also be setting ZONE_MOVABLE:
> > 
> > -#define __GFP_MOVABLE   ((__force gfp_t)___GFP_MOVABLE)  /* ZONE_MOVABLE allowed */
> > +#define __GFP_MOVABLE   ((__force gfp_t)(___GFP_MOVABLE | (ZONE_MOVABLE ^ ZONE_NORMAL)))
> > 
> I am afraid we couldn't do that, because __GFP_MOVABLE would be used potentially with other __GFPs like __GFP_DMA and __GFP_DMA32.

That's not a combination that makes much sense.  I know it's permitted today
(and it has the effect of being a no-op), but when you think about it, it
doesn't actually make any sense.

RE: [External] Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

[Huaisheng HS1 Ye]

> From: Matthew Wilcox [mailto:willy@infradead.org]
> Sent: Saturday, May 12, 2018 10:23 PM> 
> On Sat, May 12, 2018 at 11:35:00AM +0000, Huaisheng HS1 Ye wrote:
> > > The point of this exercise is to actually encode the zone number in
> > > the bottom bits of the GFP flags instead of something which has to be
> > > interpreted into a zone number.  When somebody sets __GFP_MOVABLE, they
> > > should also be setting ZONE_MOVABLE:
> > >
> > > -#define __GFP_MOVABLE   ((__force gfp_t)___GFP_MOVABLE)  /* ZONE_MOVABLE allowed
> */
> > > +#define __GFP_MOVABLE   ((__force gfp_t)(___GFP_MOVABLE | (ZONE_MOVABLE ^
> ZONE_NORMAL)))
> > >
> > I am afraid we couldn't do that, because __GFP_MOVABLE would be used potentially
> with other __GFPs like __GFP_DMA and __GFP_DMA32.
> 
> That's not a combination that makes much sense.  I know it's permitted today
> (and it has the effect of being a no-op), but when you think about it, it
> doesn't actually make any sense.

Yes, you are right.
After checking almost all references of __GFP_MOVABLE and other __GFP_* flags, perhaps I was far to get excessive pursuit of logical correctness.
For those nonsense combinations, I should ignore them.
Current GFP_ZONE_TABLE can ensure all logical correctness. That makes me want to pursue same effect.

Next, I will revise the patch according to your advice, then try to get overall testing result as far as possible.
There are many combinations because of a lot of conditions in file system and drivers. Hope I could test all things related to the lower 4 bits of gfp.

Sincerely,
Huaisheng Ye

RE: [External] Re: [PATCH v1] include/linux/gfp.h: getting rid of GFP_ZONE_TABLE/BAD

[Huaisheng HS1 Ye]

> From: Matthew Wilcox [mailto:willy@infradead.org]
> Sent: Friday, May 11, 2018 9:26 PM
> On Fri, May 11, 2018 at 03:24:34AM +0000, Huaisheng HS1 Ye wrote:
> > > From: owner-linux-mm@kvack.org [mailto:owner-linux-mm@kvack.org] On Behalf Of
> Matthew
> > > Wilcox
> > > On Fri, May 11, 2018 at 12:10:25AM +0800, Huaisheng Ye wrote:
> > > > -#define __GFP_DMA	((__force gfp_t)___GFP_DMA)
> > > > -#define __GFP_HIGHMEM	((__force gfp_t)___GFP_HIGHMEM)
> > > > -#define __GFP_DMA32	((__force gfp_t)___GFP_DMA32)
> > > > +#define __GFP_DMA	((__force gfp_t)OPT_ZONE_DMA ^ ZONE_NORMAL)
> > > > +#define __GFP_HIGHMEM	((__force gfp_t)ZONE_MOVABLE ^ ZONE_NORMAL)
> > > > +#define __GFP_DMA32	((__force gfp_t)OPT_ZONE_DMA32 ^ ZONE_NORMAL)
> > >
> > > No, you've made gfp_zone even more complex than it already is.
> > > If you can't use OPT_ZONE_HIGHMEM here, then this is a waste of time.
> > >
> > Dear Matthew,
> >
> > The reason why I don't use OPT_ZONE_HIGHMEM for __GFP_HIGHMEM	 directly is that,
> for x86_64 platform there is no CONFIG_HIGHMEM, so OPT_ZONE_HIGHMEM shall always be
> equal to ZONE_NORMAL.
> 
> Right.  On 64-bit platforms, if somebody asks for HIGHMEM, they should
> get NORMAL pages.
> 
> > For gfp_zone it is impossible to distinguish the meaning of lowest 3 bits in flags.
> How can gfp_zone to understand it comes from OPT_ZONE_HIGHMEM or ZONE_NORMAL?
> > And the most pained thing is that, if __GFP_HIGHMEM with movable flag enabled, it
> means that ZONE_MOVABLE shall be returned.
> > That is different from ZONE_DMA, ZONE_DMA32 and ZONE_NORMAL.
> 
> The point of this exercise is to actually encode the zone number in
> the bottom bits of the GFP flags instead of something which has to be
> interpreted into a zone number.  When somebody sets __GFP_MOVABLE, they
> should also be setting ZONE_MOVABLE:
> 
> -#define __GFP_MOVABLE   ((__force gfp_t)___GFP_MOVABLE)  /* ZONE_MOVABLE allowed */
> +#define __GFP_MOVABLE   ((__force gfp_t)(___GFP_MOVABLE | (ZONE_MOVABLE ^ ZONE_NORMAL)))
> 
> One thing that does need to change is:
> 
> -#define GFP_HIGHUSER_MOVABLE    (GFP_HIGHUSER | __GFP_MOVABLE)
> +#define GFP_HIGHUSER_MOVABLE    (GFP_USER | __GFP_MOVABLE)
> 
> otherwise we'll be OR'ing ZONE_MOVABLE and ZONE_HIGHMEM together.

Dear Matthew,

After thinking it over and over, I am afraid there is something needs to be discussed here.
You know current X86_64 config file of kernel doesn't enable CONFIG_HIGHMEM, that is to say from this below,

#define __GFP_HIGHMEM	((__force gfp_t)OPT_ZONE_HIGHMEM ^ ZONE_NORMAL)

__GFP_HIGHMEM should equal to 0b0000, same as the value of ZONE_NORMAL gets encoded.
If we define __GFP_MOVABLE like this,

#define __GFP_MOVABLE   ((__force gfp_t)(___GFP_MOVABLE | (ZONE_MOVABLE ^ ZONE_NORMAL)))

Just like your introduced before, with this modification when somebody sets __GFP_MOVABLE, they should also be setting ZONE_MOVABLE.
That brings us a problem, current mm (GFP_ZONE_TABLE) treats __GFP_MOVABLE as ZONE_NORMAL with movable policy, if without __GFP_HIGHMEM.
The mm shall allocate a page or pages from migrate movable list of ZONE_NORMAL's freelist.
So that conflicts with this modification. And I have checked current kernel, some of function directly set parameter gfp like this.

For example, in fs/ext4/extents.c __read_extent_tree_block,
	bh = sb_getblk_gfp(inode->i_sb, pblk, __GFP_MOVABLE | GFP_NOFS);

for these situations, I think only modify GFP_HIGHUSER_MOVABLE is not enough. I am preparing a workaround to solve this in the V2 patch.
Later I will upload it to email loop.

Sincerely,
Huaisheng Ye


> > I was thinking...
> > Whether it is possible to use other judgement condition to decide OPT_ZONE_HIGHMEM
> or ZONE_MOVABLE shall be returned from gfp_zone.
> >
> > Sincerely,
> > Huaisheng Ye
> >
> >
> > > >  static inline enum zone_type gfp_zone(gfp_t flags)
> > > >  {
> > > >  	enum zone_type z;
> > > > -	int bit = (__force int) (flags & GFP_ZONEMASK);
> > > > +	z = ((__force unsigned int)flags & ___GFP_ZONE_MASK) ^ ZONE_NORMAL;
> > > > +
> > > > +	if (z > OPT_ZONE_HIGHMEM)
> > > > +		z = OPT_ZONE_HIGHMEM +
> > > > +			!!((__force unsigned int)flags & ___GFP_MOVABLE);
> > > >
> > > > -	z = (GFP_ZONE_TABLE >> (bit * GFP_ZONES_SHIFT)) &
> > > > -					 ((1 << GFP_ZONES_SHIFT) - 1);
> > > > -	VM_BUG_ON((GFP_ZONE_BAD >> bit) & 1);
> > > > +	VM_BUG_ON(z > ZONE_MOVABLE);
> > > >  	return z;
> > > >  }
> >