Re: [PATCH 4/4] x86/mm: replace GFP_ATOMIC with GFP_KERNEL for direct map allocations

[Mike Rapoport]

On Thu, Nov 11, 2021 at 07:19:40AM -0800, Dave Hansen wrote:
> On 11/11/21 3:02 AM, Mike Rapoport wrote:
> > From: Mike Rapoport <rppt@xxxxxxxxxxxxx>
> > 
> > The allocations of the direct map pages are mostly happen very early during
> > the system boot and they use either the page table cache in brk area of bss
> > or memblock.
> > 
> > The few callers that effectively use page allocator for the direct map
> > updates are gart_iommu_init() and memory hotplug. Neither of them happen in
> > an atomic context so there is no reason to use GFP_ATOMIC for these
> > allocations.
> > 
> > Replace GFP_ATOMIC with GFP_KERNEL to avoid using atomic reserves for
> > allocations that do not require that.
> 
> I usually think of the biggest downside of GFP_ATOMIC as being that it
> fails more often.  But, since we tend not to be low on memory in early
> boot, GFP_ATOMIC and GFP_KERNEL end up being pretty close in actual
> behavior.
> 
> These allocations also get exposed via init_extra_mapping_*().  But,
> those are used via early_initcall()s where GFP_KERNEL is fine too.

Right, I forgot to mention them in the changelog...

> Those are a bit worrying because they're in somewhat nice code, like the
> Numascale APIC code.  I'm not sure how much use it sees these days.
>
> I guess if this goes wrong somehow, we'll get some nice splats to tell
> us what happened.
> 
> Was this motivated by anything in particular?  Or is it a pure cleanup?

The trigger was the discussion about PKS protection for the kernel page
tables, but for now I'd say it's pure cleanup.

-- 
Sincerely yours,
Mike.