arm64: kernel panic in paging_init()

[Catalin Marinas]

On 4 Feb 2014, at 18:57, Mark Salter <msalter at redhat.com> wrote:
> On Tue, 2014-02-04 at 12:14 +0000, Catalin Marinas wrote:
>> On Mon, Feb 03, 2014 at 08:50:49PM +0000, Mark Salter wrote:
>>> I'm seeing the following panic in paging init. This is on the foundation
>>> model with a modified dtb memory node which has a non section-aligned
>>> bank:
>>> 	memory at 80000000 {
>>> 		device_type = "memory";
>>> 		reg = <0x00000000 0x80000000 0 0x20000000>,
>>> 		      <0x00000000 0xa0300000 0 0x1fd00000>;
>>> 	};
>>> 
>>> I only see this with 64k pagesize configured. What happens is the
>>> non section-aligned bank causes alloc_init_pte() to allocate a page
>>> for the new pte from the end of the first bank (the failing address
>>> 0xfffffe001fff0000 [0x9fff0000 phys]). This should be a valid page
>>> since it was mapped during the create_mapping() call for the first
>>> memory bank. A flush_tlb_all() added to the end of create_mapping()
>>> makes the panic go away so I think the problem is something stale
>>> cached before the page with the failing address was mapped.
>> 
>> I think it goes like this:
>> 
>> head.S maps enough memory to get started but using 64K pages rather than
>> 512M sections with a single pgd entry and several ptes. This never gets
>> to the end of the first block (just up to KERNEL_END).
>> 
>> create_mapping() realises it can do a 512M section mapping, overriding
>> the original table pgd entry with a block one. The memblock limit is set
>> correctly PGDIR_SIZE but create_mapping, when it replaces the table pgd
>> with a block one doesn't do any TLB invalidation.
>> 
>> So I wouldn't do a TLB invalidation all the time but only when the old
>> pmd was set. Please give the patch below a try, I only compiled it (I'll
>> add some text afterwards):
> 
> Yes, that works. Thanks.

Can I add your tested-by?

Thanks.

Catalin