kvm: share upper halves among all pagtables

Fixes: #509

Signed-off-by: Lai Jiangshan <jiangshan.ljs@antfin.com>
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>

3 files changed, 26 insertions, 23 deletions

diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go
index dd45ad10b..5979aef97 100644
--- a/pkg/sentry/platform/kvm/kvm.go
+++ b/pkg/sentry/platform/kvm/kvm.go
@@ -158,8 +158,7 @@ func (*KVM) MaxUserAddress() usermem.Addr {
 // NewAddressSpace returns a new pagetable root.
 func (k *KVM) NewAddressSpace(_ interface{}) (platform.AddressSpace, <-chan struct{}, error) {
 	// Allocate page tables and install system mappings.
-	pageTables := pagetables.New(newAllocator())
-	k.machine.mapUpperHalf(pageTables)
+	pageTables := pagetables.NewWithUpper(newAllocator(), k.machine.upperSharedPageTables, ring0.KernelStartAddress)
 
 	// Return the new address space.
 	return &addressSpace{
diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go
index f70d761fd..e2fffc99b 100644
--- a/pkg/sentry/platform/kvm/machine.go
+++ b/pkg/sentry/platform/kvm/machine.go
@@ -41,6 +41,9 @@ type machine struct {
 	// slots are currently being updated, and the caller should retry.
 	nextSlot uint32
 
+	// upperSharedPageTables tracks the read-only shared upper of all the pagetables.
+	upperSharedPageTables *pagetables.PageTables
+
 	// kernel is the set of global structures.
 	kernel ring0.Kernel
 
@@ -199,9 +202,7 @@ func newMachine(vm int) (*machine, error) {
 	log.Debugf("The maximum number of vCPUs is %d.", m.maxVCPUs)
 	m.vCPUsByTID = make(map[uint64]*vCPU)
 	m.vCPUsByID = make([]*vCPU, m.maxVCPUs)
-	m.kernel.Init(ring0.KernelOpts{
-		PageTables: pagetables.New(newAllocator()),
-	}, m.maxVCPUs)
+	m.kernel.Init(m.maxVCPUs)
 
 	// Pull the maximum slots.
 	maxSlots, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(m.fd), _KVM_CHECK_EXTENSION, _KVM_CAP_MAX_MEMSLOTS)
@@ -213,6 +214,13 @@ func newMachine(vm int) (*machine, error) {
 	log.Debugf("The maximum number of slots is %d.", m.maxSlots)
 	m.usedSlots = make([]uintptr, m.maxSlots)
 
+	// Create the upper shared pagetables and kernel(sentry) pagetables.
+	m.upperSharedPageTables = pagetables.New(newAllocator())
+	m.mapUpperHalf(m.upperSharedPageTables)
+	m.upperSharedPageTables.Allocator.(*allocator).base.Drain()
+	m.upperSharedPageTables.MarkReadOnlyShared()
+	m.kernel.PageTables = pagetables.NewWithUpper(newAllocator(), m.upperSharedPageTables, ring0.KernelStartAddress)
+
 	// Apply the physical mappings. Note that these mappings may point to
 	// guest physical addresses that are not actually available. These
 	// physical pages are mapped on demand, see kernel_unsafe.go.
@@ -226,7 +234,6 @@ func newMachine(vm int) (*machine, error) {
 
 		return true // Keep iterating.
 	})
-	m.mapUpperHalf(m.kernel.PageTables)
 
 	var physicalRegionsReadOnly []physicalRegion
 	var physicalRegionsAvailable []physicalRegion
diff --git a/pkg/sentry/platform/kvm/machine_amd64.go b/pkg/sentry/platform/kvm/machine_amd64.go
index a8b729e62..8e03c310d 100644
--- a/pkg/sentry/platform/kvm/machine_amd64.go
+++ b/pkg/sentry/platform/kvm/machine_amd64.go
@@ -432,30 +432,27 @@ func availableRegionsForSetMem() (phyRegions []physicalRegion) {
 	return physicalRegions
 }
 
-var execRegions = func() (regions []region) {
+func (m *machine) mapUpperHalf(pageTable *pagetables.PageTables) {
+	// Map all the executible regions so that all the entry functions
+	// are mapped in the upper half.
 	applyVirtualRegions(func(vr virtualRegion) {
 		if excludeVirtualRegion(vr) || vr.filename == "[vsyscall]" {
 			return
 		}
+
 		if vr.accessType.Execute {
-			regions = append(regions, vr.region)
+			r := vr.region
+			physical, length, ok := translateToPhysical(r.virtual)
+			if !ok || length < r.length {
+				panic("impossible translation")
+			}
+			pageTable.Map(
+				usermem.Addr(ring0.KernelStartAddress|r.virtual),
+				r.length,
+				pagetables.MapOpts{AccessType: usermem.Execute},
+				physical)
 		}
 	})
-	return
-}()
-
-func (m *machine) mapUpperHalf(pageTable *pagetables.PageTables) {
-	for _, r := range execRegions {
-		physical, length, ok := translateToPhysical(r.virtual)
-		if !ok || length < r.length {
-			panic("impossilbe translation")
-		}
-		pageTable.Map(
-			usermem.Addr(ring0.KernelStartAddress|r.virtual),
-			r.length,
-			pagetables.MapOpts{AccessType: usermem.Execute},
-			physical)
-	}
 	for start, end := range m.kernel.EntryRegions() {
 		regionLen := end - start
 		physical, length, ok := translateToPhysical(start)