Re: riscv32 EXT4 splat, 6.8 regression?

[Mike Rapoport]

On Tue, Apr 16, 2024 at 05:31:51PM +0100, Matthew Wilcox wrote:
> On Tue, Apr 16, 2024 at 07:19:27PM +0300, Mike Rapoport wrote:
> > > "last page of the first gigabyte" - why first gigabyte? Do you mean
> > > last page of *last* gigabyte?
> >  
> > With 3G-1G split linear map can map only 1G from 0xc0000000 to 0xffffffff
> > (or 0x00000000 with 32-bit overflow):
> > 
> > [    0.000000]       lowmem : 0xc0000000 - 0x00000000   (1024 MB)
> 
> ... but you can't map that much.  You need to reserve space for (may not
> be exhaustive):
> 
>  - PCI BARs (or other MMIO)
>  - vmap
>  - kmap
>  - percpu
>  - ioremap
>  - modules
>  - fixmap
>  - Maybe EFI runtime services?
> 
> You'll be lucky to get 800MB of ZONE_NORMAL.

But that does not mean that the last page won't get to the buddy

-- 
Sincerely yours,
Mike.