srat: harsh hot-pluggable memory check?

srat: harsh hot-pluggable memory check?

[Davidlohr Bueso]

When parsing the memory affinity mappings in arch/x86/mm/srat.c:
acpi_numa_memory_affinity_init() I'm wondering if the hot-pluggable check is too harsh, 
as we consider an error if the hot-pluggable bit is set and CONFIG_MEMORY_HOTPLUG is not.

Based on the ACPI specs (v5):

"If the Enabled bit is set and the Hot Pluggable bit is also set. The
system hardware supports hot-add and hot-remove of this memory
region."

This only mentions that the system supports hot-plugging, and IMHO if the
user decides not to use CONFIG_MEMORY_HOTPLUG, it shouldn't be considered an error.
Therefore would it be ok to drop the check? Or am I missing something?

Thanks,
Davidlohr

Re: srat: harsh hot-pluggable memory check?

[Andi Kleen]

> This only mentions that the system supports hot-plugging, and IMHO if the
> user decides not to use CONFIG_MEMORY_HOTPLUG, it shouldn't be considered an error.
> Therefore would it be ok to drop the check? Or am I missing something?

The very strict checks were originally implemented because various early
BIOS had largely fictional SRATs, and trusting them blindly caused
boot failures or a lot of wasted memory for unnecessary hotplug zones. 
The wasted memory was mainly a problem with the old memory hotplug
implementation that pre-allocated memmaps, that's not a problem anymore.
However there may be still some other failure cases.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only.

Re: srat: harsh hot-pluggable memory check?

[Davidlohr Bueso]

On Thu, 2013-01-10 at 21:02 +0100, Andi Kleen wrote:
> > This only mentions that the system supports hot-plugging, and IMHO if the
> > user decides not to use CONFIG_MEMORY_HOTPLUG, it shouldn't be considered an error.
> > Therefore would it be ok to drop the check? Or am I missing something?
> 
> The very strict checks were originally implemented because various early
> BIOS had largely fictional SRATs, and trusting them blindly caused
> boot failures or a lot of wasted memory for unnecessary hotplug zones. 
> The wasted memory was mainly a problem with the old memory hotplug
> implementation that pre-allocated memmaps, that's not a problem anymore.
> However there may be still some other failure cases.
> 

Would you be willing to take a patch that drops this check then? Or do
you see any other scenario where it would still be valid?

Thanks,
Davidlohr

Re: srat: harsh hot-pluggable memory check?

[Andi Kleen]

On Fri, Jan 11, 2013 at 12:13:50PM -0800, Davidlohr Bueso wrote:
> On Thu, 2013-01-10 at 21:02 +0100, Andi Kleen wrote:
> > > This only mentions that the system supports hot-plugging, and IMHO if the
> > > user decides not to use CONFIG_MEMORY_HOTPLUG, it shouldn't be considered an error.
> > > Therefore would it be ok to drop the check? Or am I missing something?
> > 
> > The very strict checks were originally implemented because various early
> > BIOS had largely fictional SRATs, and trusting them blindly caused
> > boot failures or a lot of wasted memory for unnecessary hotplug zones. 
> > The wasted memory was mainly a problem with the old memory hotplug
> > implementation that pre-allocated memmaps, that's not a problem anymore.
> > However there may be still some other failure cases.
> > 
> 
> Would you be willing to take a patch that drops this check then? Or do
> you see any other scenario where it would still be valid?

I don't maintain this code. 

Personally I would be vary to any changes in this area, unless you
can very clearly demonstrate that the change cannot break any old
system.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only.