答复: [PATCH 01/11] Initialize the mapping of KASan shadow memory

On Oct 19, 2017 at 19:09, Russell King - ARM Linux [mailto:linux at armlinux.org.uk] wrote:

On Thu, Oct 12, 2017 at 02:42:49AM +0300, Dmitry Osipenko wrote:

quoted

On 11.10.2017 11:22, Abbott Liu wrote:

quoted

+void __init kasan_map_early_shadow(pgd_t *pgdp)
+{
+	int i;
+	unsigned long start = KASAN_SHADOW_START;
+	unsigned long end = KASAN_SHADOW_END;
+	unsigned long addr;
+	unsigned long next;
+	pgd_t *pgd;
+
+	for (i = 0; i < PTRS_PER_PTE; i++)
+		set_pte_at(&init_mm, KASAN_SHADOW_START + i*PAGE_SIZE,
+			&kasan_zero_pte[i], pfn_pte(
+				virt_to_pfn(kasan_zero_page),
+				__pgprot(_L_PTE_DEFAULT | L_PTE_DIRTY | L_PTE_XN)));

Shouldn't all __pgprot's contain L_PTE_MT_WRITETHROUGH ?

One of the architecture restrictions is that the cache attributes of
all aliases should match (but there is a specific workaround that
permits this, provided that the dis-similar mappings aren't accessed
without certain intervening instructions.)

Why should it be L_PTE_MT_WRITETHROUGH, and not the same cache
attributes as the lowmem mapping?

Here is mapping the kasan shadow which is used at the early stage of kernel start(from start
of start_kernel to paging_init). At this stage we only read the kasan shadows, never write the
kasan shadows which is initialized to be zero. 

We will map the kasan shadows again with flags PAGE_KERNEL:
pte_t * __meminit kasan_pte_populate(pmd_t *pmd, unsigned long addr, int node)
{
	pte_t *pte = pte_offset_kernel(pmd, addr);
	if (pte_none(*pte)) {
		pte_t entry;
		void *p = kasan_alloc_block(PAGE_SIZE, node);
 		if (!p)
			return NULL;
		entry = pfn_pte(virt_to_pfn(p), __pgprot(pgprot_val(PAGE_KERNEL)));
			set_pte_at(&init_mm, addr, pte, entry);
	}
	return pte;
}