Hello Christophe,

On Tue, Feb 20, 2018 at 06:45:09PM +0100, christophe leroy wrote:
[...]

quoted

-	if (slab_is_available() && (p < virt_to_phys(high_memory)) &&
+	if (slab_is_available() && pfn_valid(__phys_to_pfn(p)) &&

I'm not sure this is equivalent:

high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
#define ARCH_PFN_OFFSET		((unsigned long)(MEMORY_START >> PAGE_SHIFT))
#define pfn_valid(pfn)		((pfn) >= ARCH_PFN_OFFSET && (pfn) < max_mapnr)
set_max_mapnr(max_pfn);

So in the current implementation it checks against max_low_pfn while your
patch checks against max_pfn

	max_low_pfn = max_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
#ifdef CONFIG_HIGHMEM
	max_low_pfn = lowmem_end_addr >> PAGE_SHIFT;
#endif

Good point, I haven't considered CONFIG_HIGHMEM before.

As far as I understand it, in the non-CONFIG_HIGHMEM case:

  - max_low_pfn is set to the same value as max_pfn, so the ioremap
    check should detect the same PFNs as RAM.

and with CONFIG_HIGHMEM:

  - max_low_pfn is set to lowmem_end_addr >> PAGE_SHIFT
  - but max_pfn isn't

So, I think you're right.


While looking through arch/powerpc/mm, I noticed that there's a
page_is_ram function, which simply uses the memblocks directly, on
PPC32. It seems like a good candidate for the RAM check in
__ioremap_caller, except that there's this code, which apparently
trashes memblock 0 completely on non-CONFIG_NEED_MULTIPLE_NODES:

  https://elixir.bootlin.com/linux/v4.16-rc2/source/arch/powerpc/mm/mem.c#L223


Thanks,
Jonathan Neuschäfer