On Thu, Oct 4, 2012 at 10:23 PM, Min-gyu Kim [off-list ref] wrote:

quoted

-----Original Message-----
From: kvm-owner at vger.kernel.org [mailto:kvm-owner at vger.kernel.org] On
Behalf Of Christoffer Dall
Sent: Monday, October 01, 2012 6:11 PM
To: kvm at vger.kernel.org; linux-arm-kernel at lists.infradead.org;
kvmarm at lists.cs.columbia.edu
Cc: Marc Zyngier
Subject: [PATCH v2 06/14] KVM: ARM: Memory virtualization setup

+static void stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache
*cache,
+                        phys_addr_t addr, const pte_t *new_pte) {
+     pgd_t *pgd;
+     pud_t *pud;
+     pmd_t *pmd;
+     pte_t *pte, old_pte;
+
+     /* Create 2nd stage page table mapping - Level 1 */
+     pgd = kvm->arch.pgd + pgd_index(addr);
+     pud = pud_offset(pgd, addr);
+     if (pud_none(*pud)) {
+             if (!cache)
+                     return; /* ignore calls from kvm_set_spte_hva */
+             pmd = mmu_memory_cache_alloc(cache);
+             pud_populate(NULL, pud, pmd);
+             pmd += pmd_index(addr);
+             get_page(virt_to_page(pud));
+     } else
+             pmd = pmd_offset(pud, addr);
+
+     /* Create 2nd stage page table mapping - Level 2 */
+     if (pmd_none(*pmd)) {
+             if (!cache)
+                     return; /* ignore calls from kvm_set_spte_hva */
+             pte = mmu_memory_cache_alloc(cache);
+             clean_pte_table(pte);
+             pmd_populate_kernel(NULL, pmd, pte);
+             pte += pte_index(addr);
+             get_page(virt_to_page(pmd));
+     } else
+             pte = pte_offset_kernel(pmd, addr);
+
+     /* Create 2nd stage page table mapping - Level 3 */
+     old_pte = *pte;
+     set_pte_ext(pte, *new_pte, 0);
+     if (pte_present(old_pte))
+             __kvm_tlb_flush_vmid(kvm);
+     else
+             get_page(virt_to_page(pte));
+}

I'm not sure about the 3-level page table, but isn't it necessary to
clean the page table for 2nd level?
There are two mmu_memory_cache_alloc calls. One has following clean_pte_table
and the other doesn't have.

hmm, it probably is - I couldn't really find the common case where
this is done in the kernel normally (except for some custom loop in
ioremap and idmap), but I added this fix:

diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
index 5394a52..f11ba27f 100644
--- a/arch/arm/kvm/mmu.c
+++ b/arch/arm/kvm/mmu.c

@@ -430,6 +430,7 @@ static void stage2_set_pte(struct kvm *kvm, struct

kvm_mmu_memory_cache *cache,
 		if (!cache)
 			return; /* ignore calls from kvm_set_spte_hva */
 		pmd = mmu_memory_cache_alloc(cache);
+		clean_dcache_area(pmd, PTRS_PER_PMD * sizeof(pmd_t));
 		pud_populate(NULL, pud, pmd);
 		pmd += pmd_index(addr);
 		get_page(virt_to_page(pud));

And why do you ignore calls from kvm_set_spte_hva? It is supposed to happen when
host moves the page, right? Then you ignore the case because it can be handled
later when fault actually happens? Is there any other reason that I miss?

kvm_set_spte_hva tells us that a page at some IPA is going to be
backed by another physical page, which means we must adjust the stage
2 mapping. However, if we don't have that page mapped in the stage 2
page table, we don't need to do anything, and certainly don't want to
start allocating unnecessary level2 and level3 page tables.


Thanks!
-Christoffer