Re: [PATCH 1/1] resource: Fixed iomem resource release failed on release_mem_region_adjustable() when memory node or cpu node hot-remove.

[Kani, Toshi]

On Fri, 2018-06-22 at 18:19 -0500, Bjorn Helgaas wrote:
> [+cc Toshi]
> 
> On Fri, Jun 22, 2018 at 08:01:38PM +0800, guomin chen wrote:
> > We've got a bug report indicating the hot-remove node resource
> > release failed,when the memory on this node is divided into
> > several sections.because the release_mem_region_adjustable()
> > can only release one resource that must be [start,end].
> 
> Can you please include a URL for the bug report?  That's useful for
> additional details and gives hints about how future changes in this
> area might be tested.
> 
> release_mem_region_adjustable() and the only call to it were added
> by Toshi (cc'd):
> 
>   825f787bb496 ("resource: add release_mem_region_adjustable()")
>   fe74ebb106a5 ("mm: change __remove_pages() to call release_mem_region_adjustable()")
> 
> > In my case, the BIOS supports faulty memory isolation. if BIOS
> > detected bad memory block, the BIOS will isolates this badblock.
> > And set this badblock memory to EfiUnusableMemory in EFI memory map
> > base on UEFI 2.7 spec.For example in my system, the memory range on
> > node2 is [mem 0x0000080000000000-0x00000807ffffffff].but the BIOS
> > detected the [8004e000000-8004e0fffff] is a badblock memory.
> > So the memory on node2 seem like this:
> >    		80000000000-8004dffffff : System RAM
> > 		8004e000000-8004e0fffff : Unusable memory
> > 		8004e100000-807ffffffff : System RAM
> > 
> > Now, when offline the cpu node2,the kernel will try to release
> > ioresource [mem 0x0000080000000000-0x00000807ffffffff]. at this
> > time, the kernel will release failed,and output error message:
> > "Unable to release resource <0x0000080000000000-0x00000807ffffffff>
> > (-22)".

Hmm... this BIO implementation does not sound right to me.  When EFI
memory map has a hole, ACPI memory object should have the same hole in
its resource info.  If so, this node2 memory hot-delete would have
requested a deletion of 2 separate System RAM ranges, not a single range
including this hole.  This is also important for memory hot-add.  ACPI
memory object needs to describe this hole in order for the OS to avoid
adding this badblock memory.

Thanks,
-Toshi