Re: [PATCH] mm: check pfn_valid first in zero_resv_unavail

[Dave Young]

On 12/01/17 at 10:42am, Michal Hocko wrote:
> On Fri 01-12-17 17:29:51, Dave Young wrote:
> > On 12/01/17 at 10:19am, Michal Hocko wrote:
> > > On Fri 01-12-17 16:56:57, Dave Young wrote:
> > > > On 11/30/17 at 10:35am, Michal Hocko wrote:
> > > [...]
> > > > > Can we exclude that range from the memblock allocator instead? E.g. what
> > > > > happens if somebody allocates from that range?
> > > > 
> > > > It is a EFI BGRT image buffer provided by firmware, they are reserved
> > > > always and can not be used to allocate memory.
> > > 
> > > Hmm, I see but I was actually suggesting to remove this range from the
> > > memblock allocator altogether (memblock_remove) as it shouldn't be there
> > > in the first place.
> > 
> > Oh, I'm not sure because it is introduced as a way for efi to reserve
> > boot services areas to be persistent across kexec reboot. See
> > drivers/firmware/efi/efi.c: efi_mem_reserve(), BGRT is only one user
> > of it, there is esrt and maybe other users, I do not know if it is safe
> > :(
> 
> Hmm, so it this range ever backed by a valid pfn?

I think it is in normal boot, just it does not appear in e820 across kdump
reboot. For kdump kexec_tools provided e820 the last pfn only covers the kdump
crashkernel ranges thus it is not mapped.

> -- 
> Michal Hocko
> SUSE Labs