[patch 3/5] KVM: add kvm_follow_page()

[Marcelo Tosatti <mtosatti-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

[-- Attachment #1: mmu-setpte-atomic --]
[-- Type: text/plain, Size: 4557 bytes --]

In preparation for a mmu spinlock, avoid schedules in mmu_set_spte() 
by using follow_page() instead of get_user_pages().

The fault handling work by get_user_pages() now happens outside the lock.

Signed-off-by: Marcelo Tosatti <mtosatti-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>

Index: kvm.quilt/arch/x86/kvm/mmu.c
===================================================================
--- kvm.quilt.orig/arch/x86/kvm/mmu.c
+++ kvm.quilt/arch/x86/kvm/mmu.c
@@ -907,7 +907,9 @@ static void mmu_set_spte(struct kvm_vcpu
 	if (!(pte_access & ACC_EXEC_MASK))
 		spte |= PT64_NX_MASK;
 
-	page = gfn_to_page(vcpu->kvm, gfn);
+	page = kvm_follow_page(vcpu->kvm, gfn);
+	if (!page)
+		return;
 
 	spte |= PT_PRESENT_MASK;
 	if (pte_access & ACC_USER_MASK)
@@ -983,8 +985,11 @@ static int nonpaging_map(struct kvm_vcpu
 		table = __va(table_addr);
 
 		if (level == 1) {
+			struct page *page;
+			page = gfn_to_page(vcpu->kvm, gfn);
 			mmu_set_spte(vcpu, &table[index], ACC_ALL, ACC_ALL,
 				     0, write, 1, &pt_write, gfn);
+			kvm_release_page_clean(page);
 			return pt_write || is_io_pte(table[index]);
 		}
 
Index: kvm.quilt/include/linux/kvm_host.h
===================================================================
--- kvm.quilt.orig/include/linux/kvm_host.h
+++ kvm.quilt/include/linux/kvm_host.h
@@ -163,6 +163,7 @@ int kvm_arch_set_memory_region(struct kv
 				int user_alloc);
 gfn_t unalias_gfn(struct kvm *kvm, gfn_t gfn);
 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn);
+struct page *kvm_follow_page(struct kvm *kvm, gfn_t gfn);
 void kvm_release_page_clean(struct page *page);
 void kvm_release_page_dirty(struct page *page);
 int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
Index: kvm.quilt/virt/kvm/kvm_main.c
===================================================================
--- kvm.quilt.orig/virt/kvm/kvm_main.c
+++ kvm.quilt/virt/kvm/kvm_main.c
@@ -453,6 +453,25 @@ static unsigned long gfn_to_hva(struct k
 	return (slot->userspace_addr + (gfn - slot->base_gfn) * PAGE_SIZE);
 }
 
+struct page *kvm_follow_page(struct kvm *kvm, gfn_t gfn)
+{
+	unsigned long addr;
+	struct vm_area_struct *vma;
+
+	addr = gfn_to_hva(kvm, gfn);
+	/* MMIO access */
+	if (kvm_is_error_hva(addr)) {
+		get_page(bad_page);
+		return bad_page;
+	}
+
+	vma = find_vma(current->mm, addr);
+	if (!vma)
+		return NULL;
+
+	return follow_page(vma, addr, FOLL_GET|FOLL_TOUCH);
+}
+
 /*
  * Requires current->mm->mmap_sem to be held
  */
Index: kvm.quilt/arch/x86/kvm/paging_tmpl.h
===================================================================
--- kvm.quilt.orig/arch/x86/kvm/paging_tmpl.h
+++ kvm.quilt/arch/x86/kvm/paging_tmpl.h
@@ -67,6 +67,7 @@ struct guest_walker {
 	gfn_t table_gfn[PT_MAX_FULL_LEVELS];
 	pt_element_t ptes[PT_MAX_FULL_LEVELS];
 	gpa_t pte_gpa[PT_MAX_FULL_LEVELS];
+	struct page *page;
 	unsigned pt_access;
 	unsigned pte_access;
 	gfn_t gfn;
@@ -203,14 +204,18 @@ walk:
 		--walker->level;
 	}
 
+	walker->page = gfn_to_page(vcpu->kvm, walker->gfn);
+
 	if (write_fault && !is_dirty_pte(pte)) {
 		bool ret;
 
 		mark_page_dirty(vcpu->kvm, table_gfn);
 		ret = FNAME(cmpxchg_gpte)(vcpu->kvm, table_gfn, index, pte,
 			    pte|PT_DIRTY_MASK);
-		if (ret)
+		if (ret) {
+			kvm_release_page_clean(walker->page);
 			goto walk;
+		}
 		pte |= PT_DIRTY_MASK;
 		mutex_lock(&vcpu->kvm->lock);
 		kvm_mmu_pte_write(vcpu, pte_gpa, (u8 *)&pte, sizeof(pte));
@@ -323,8 +328,10 @@ static u64 *FNAME(fetch)(struct kvm_vcpu
 			r = kvm_read_guest_atomic(vcpu->kvm,
 						  walker->pte_gpa[level - 2],
 				       		  &curr_pte, sizeof(curr_pte));
-			if (r || curr_pte != walker->ptes[level - 2])
-				return NULL;
+			if (r || curr_pte != walker->ptes[level - 2]) {
+				shadow_ent = NULL;
+				goto out;
+			}
 		}
 		shadow_addr = __pa(shadow_page->spt);
 		shadow_pte = shadow_addr | PT_PRESENT_MASK | PT_ACCESSED_MASK
@@ -336,7 +343,8 @@ static u64 *FNAME(fetch)(struct kvm_vcpu
 		     user_fault, write_fault,
 		     walker->ptes[walker->level-1] & PT_DIRTY_MASK,
 		     ptwrite, walker->gfn);
-
+out:
+	kvm_release_page_clean(walker->page);
 	return shadow_ent;
 }
 
@@ -425,6 +433,7 @@ static gpa_t FNAME(gva_to_gpa)(struct kv
 	if (r) {
 		gpa = gfn_to_gpa(walker.gfn);
 		gpa |= vaddr & ~PAGE_MASK;
+		kvm_release_page_clean(walker.page);
 	}
 
 	return gpa;

-- 


-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2005.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/