Re: [RFC v3 2/4] mm: move PG_slab flag to page_type

[Hyeonggon Yoo]

On Mon, Jan 30, 2023 at 05:11:48AM +0000, Matthew Wilcox wrote:
> On Mon, Jan 30, 2023 at 01:34:59PM +0900, Hyeonggon Yoo wrote:
> > > Seems like quite some changes to page_type to accomodate SLAB, which is
> > > hopefully going away soon(TM). Could we perhaps avoid that?
> > 
> > If it could be done with less changes, I'll try to avoid that.
> 
> Let me outline the idea I had for removing PG_slab:
> 
> Observe that PG_reserved and PG_slab are mutually exclusive.  Also,
> if PG_reserved is set, no other flags are set.  If PG_slab is set, only
> PG_locked is used.  Many of the flags are only for use by anon/page
> cache pages (eg referenced, uptodate, dirty, lru, active, workingset,
> waiters, error, owner_priv_1, writeback, mappedtodisk, reclaim,
> swapbacked, unevictable, mlocked).
> 
> Redefine PG_reserved as PG_kernel.  Now we can use the other _15_
> flags to indicate pagetype, as long as PG_kernel is set. 

So PG_kernel is a new special flag, I thought it indicates
"not usermappable pages", but considering PG_vmalloc it's not.

> So, eg
> PageSlab() can now be (page->flags & PG_type) == PG_slab where

But if PG_xxx and PG_slab shares same bit, PG_xxx would be confused?

> #define PG_kernel	0x00001
> #define PG_type		(PG_kernel | 0x7fff0)
> #define PG_slab		(PG_kernel | 0x00010)
> #define PG_reserved	(PG_kernel | 0x00020)
> #define PG_buddy	(PG_kernel | 0x00030)
> #define PG_offline	(PG_kernel | 0x00040)
> #define PG_table	(PG_kernel | 0x00050)
> #define PG_guard	(PG_kernel | 0x00060)
> 
> That frees up the existing PG_slab, lets us drop the page_type field
> altogether and gives us space to define all the page types we might
> want (eg PG_vmalloc)
> 
> We'll want to reorganise all the flags which are for anon/file pages
> into a contiguous block.  And now that I think about it, vmalloc pages
> can be mapped to userspace, so they can get marked dirty, so only
> 14 bits are available.  Maybe rearrange to ...
> 
> PG_locked	0x000001
> PG_writeback	0x000002
> PG_head		0x000004

I think slab still needs PG_head,
but it seems to be okay with this layout.
(but these assumpstions are better documented, I think)

> PG_dirty	0x000008
> PG_owner_priv_1	0x000010
> PG_arch_1	0x000020
> PG_private	0x000040
> PG_waiters	0x000080
> PG_kernel	0x000100
> PG_referenced	0x000200
> PG_uptodate	0x000400
> PG_lru		0x000800
> PG_active	0x001000
> PG_workingset	0x002000
> PG_error	0x004000
> PG_private_2	0x008000
> PG_mappedtodisk	0x010000
> PG_reclaim	0x020000
> PG_swapbacked	0x040000
> PG_unevictable	0x080000
> PG_mlocked	0x100000
> 
> ... or something.  There are a number of constraints and it may take
> a few iterations to get this right.  Oh, and if this is the layout
> we use, then:
> 
> PG_type		0x1fff00
> PG_reserved	(PG_kernel | 0x200)
> PG_slab		(PG_kernel | 0x400)
> PG_buddy	(PG_kernel | 0x600)
> PG_offline	(PG_kernel | 0x800)
> PG_table	(PG_kernel | 0xa00)
> PG_guard	(PG_kernel | 0xc00)
> PG_vmalloc	(PG_kernel | 0xe00)

what is PG_vmalloc for, is it just an example for
explaining possible layout?

> This is going to make show_page_flags() more complex :-P
>
> Oh, and while we're doing this, we should just make PG_mlocked
> unconditional.  NOMMU doesn't need the extra space in page flags
> (for what?  their large number of NUMA nodes?)