Re: mm/percpu.c: use smarter memory allocation for struct pcpu_alloc_info (crisv32 hang)

[Nicolas Pitre <nicolas.pitre@linaro.org>]

On Mon, 20 Nov 2017, Guenter Roeck wrote:

> On Mon, Nov 20, 2017 at 07:28:21PM -0500, Nicolas Pitre wrote:
> > On Mon, 20 Nov 2017, Guenter Roeck wrote:
> > 
> > > bdata->node_min_pfn=60000 PFN_PHYS(bdata->node_min_pfn)=c0000000 start_off=536000 region=c0536000
> > 
> > If PFN_PHYS(bdata->node_min_pfn)=c0000000 and
> > region=c0536000 that means phys_to_virt() is a no-op.
> > 
> No, it is |= 0x80000000

Then the bootmem registration looks very fishy. If you have:

> I think the problem is the 0x60000 in bdata->node_min_pfn. It is shifted
> left by PFN_PHYS, making it 0xc0000000, which in my understanding is
> a virtual address.

Exact.

#define __pa(x)                 ((unsigned long)(x) & 0x7fffffff)
#define __va(x)                 ((void *)((unsigned long)(x) | 0x80000000))

With that, the only possible physical address range you may have is 
0x40000000 - 0x7fffffff, and it better start at 0x40000000. If that's 
not where your RAM is then something is wrong.

This is in fact a very bad idea to define __va() and __pa() using 
bitwise operations as this hides mistakes like defining physical RAM 
address at 0xc0000000. Instead, it should look like:

#define __pa(x)                 ((unsigned long)(x) - 0x80000000)
#define __va(x)                 ((void *)((unsigned long)(x) + 0x80000000))

This way, bad physical RAM address definitions will be caught 
immediately.

> That doesn't seem to be easy to fix. It seems there is a mixup of physical
> and  virtual addresses in the architecture.

Well... I don't think there is much else to say other than this needs 
fixing.


Nicolas