[PATCH v3 2/2] arm64: Mark kernel page ranges contiguous

On Fri, Feb 26, 2016 at 05:28:26PM +0000, Will Deacon wrote:

On Thu, Feb 25, 2016 at 02:46:54PM -0600, Jeremy Linton wrote:

quoted

On 02/25/2016 10:16 AM, Will Deacon wrote:

quoted

On Fri, Feb 19, 2016 at 11:46:23AM -0600, Jeremy Linton wrote:

quoted

+static void clear_cont_pte_range(pte_t *pte, unsigned long addr)
+{
+	int i;
+
+	pte -= CONT_RANGE_OFFSET(addr);
+	for (i = 0; i < CONT_PTES; i++) {
+		if (pte_cont(*pte))
+			set_pte(pte, pte_mknoncont(*pte));
+		pte++;
+	}
+	flush_tlb_all();

Do you still need this invalidation? I thought the table weren't even
live at this point?

Well it continues to match the calls in alloc_init_p*.

Ok, but if it's not needed (and I don't think it is), then we should
remove the invalidation from there too rather than add more of it.

I agree. Talking to Mark R, I think all the create_pgd_mapping code
should be trimmed of p*d splitting entirely (with some warnings left in
place).

The efi_create_mapping() case is done on efi_mm anyway with a low
address and non-active TTBR0_EL1. As Mark suggested, dropping p*d
splitting should work for this case as well by ignoring the next
consecutive mapping if it overlaps since it's a linear mapping (possibly
some permissions may need to be upgraded). In this case, the TLBI would
be done by efi_create_mapping() or even skipped altogether assuming that
no page tables are changed (only going from invalid to valid).

My preference, however, would be for efi_create_mapping() should use a
different mechanism from create_pgd_mapping (aimed at very early memmap)
like remap_pfn_range() but I haven't looked in detail on whether this is
feasible.

Once we have this clean-up in place, we can rebase the contiguous PTE
patches on top and remove unnecessary TLBI.

-- 
Catalin