Implement /proc/sys/net/ipv4/ip_local_port_range

Speeds up the socket stress tests by a couple orders of magnitude.

PiperOrigin-RevId: 361721050

6 files changed, 233 insertions, 7 deletions

diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go
index 52061175f..bbe282c03 100644
--- a/pkg/sentry/fs/proc/sys_net.go
+++ b/pkg/sentry/fs/proc/sys_net.go
@@ -17,6 +17,7 @@ package proc
 import (
 	"fmt"
 	"io"
+	"math"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
@@ -26,6 +27,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/fs/ramfs"
 	"gvisor.dev/gvisor/pkg/sentry/inet"
 	"gvisor.dev/gvisor/pkg/sync"
+	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/tcpip/network/ipv4"
 	"gvisor.dev/gvisor/pkg/usermem"
 	"gvisor.dev/gvisor/pkg/waiter"
@@ -498,6 +500,120 @@ func (f *ipForwardingFile) Write(ctx context.Context, _ *fs.File, src usermem.IO
 	return n, f.stack.SetForwarding(ipv4.ProtocolNumber, *f.ipf.enabled)
 }
 
+// portRangeInode implements fs.InodeOperations. It provides and allows
+// modification of the range of ephemeral ports that IPv4 and IPv6 sockets
+// choose from.
+//
+// +stateify savable
+type portRangeInode struct {
+	fsutil.SimpleFileInode
+
+	stack inet.Stack `state:"wait"`
+
+	// start and end store the port range. We must save/restore this here,
+	// since a netstack instance is created on restore.
+	start *uint16
+	end   *uint16
+}
+
+func newPortRangeInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode {
+	ipf := &portRangeInode{
+		SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0644), linux.PROC_SUPER_MAGIC),
+		stack:           s,
+	}
+	sattr := fs.StableAttr{
+		DeviceID:  device.ProcDevice.DeviceID(),
+		InodeID:   device.ProcDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.SpecialFile,
+	}
+	return fs.NewInode(ctx, ipf, msrc, sattr)
+}
+
+// Truncate implements fs.InodeOperations.Truncate. Truncate is called when
+// O_TRUNC is specified for any kind of existing Dirent but is not called via
+// (f)truncate for proc files.
+func (*portRangeInode) Truncate(context.Context, *fs.Inode, int64) error {
+	return nil
+}
+
+// +stateify savable
+type portRangeFile struct {
+	fsutil.FileGenericSeek          `state:"nosave"`
+	fsutil.FileNoIoctl              `state:"nosave"`
+	fsutil.FileNoMMap               `state:"nosave"`
+	fsutil.FileNoSplice             `state:"nosave"`
+	fsutil.FileNoopFlush            `state:"nosave"`
+	fsutil.FileNoopFsync            `state:"nosave"`
+	fsutil.FileNoopRelease          `state:"nosave"`
+	fsutil.FileNotDirReaddir        `state:"nosave"`
+	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
+	waiter.AlwaysReady              `state:"nosave"`
+
+	inode *portRangeInode
+}
+
+// GetFile implements fs.InodeOperations.GetFile.
+func (in *portRangeInode) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	flags.Pread = true
+	flags.Pwrite = true
+	return fs.NewFile(ctx, dirent, flags, &portRangeFile{
+		inode: in,
+	}), nil
+}
+
+// Read implements fs.FileOperations.Read.
+func (pf *portRangeFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset != 0 {
+		return 0, io.EOF
+	}
+
+	if pf.inode.start == nil {
+		start, end := pf.inode.stack.PortRange()
+		pf.inode.start = &start
+		pf.inode.end = &end
+	}
+
+	contents := fmt.Sprintf("%d %d\n", *pf.inode.start, *pf.inode.end)
+	n, err := dst.CopyOut(ctx, []byte(contents))
+	return int64(n), err
+}
+
+// Write implements fs.FileOperations.Write.
+//
+// Offset is ignored, multiple writes are not supported.
+func (pf *portRangeFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	if src.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	// Only consider size of one memory page for input for performance
+	// reasons.
+	src = src.TakeFirst(usermem.PageSize - 1)
+
+	ports := make([]int32, 2)
+	n, err := usermem.CopyInt32StringsInVec(ctx, src.IO, src.Addrs, ports, src.Opts)
+	if err != nil {
+		return 0, err
+	}
+
+	// Port numbers must be uint16s.
+	if ports[0] < 0 || ports[1] < 0 || ports[0] > math.MaxUint16 || ports[1] > math.MaxUint16 {
+		return 0, syserror.EINVAL
+	}
+
+	if err := pf.inode.stack.SetPortRange(uint16(ports[0]), uint16(ports[1])); err != nil {
+		return 0, err
+	}
+	if pf.inode.start == nil {
+		pf.inode.start = new(uint16)
+		pf.inode.end = new(uint16)
+	}
+	*pf.inode.start = uint16(ports[0])
+	*pf.inode.end = uint16(ports[1])
+	return n, nil
+}
+
 func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode {
 	contents := map[string]*fs.Inode{
 		// Add tcp_sack.
@@ -506,12 +622,15 @@ func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s ine
 		// Add ip_forward.
 		"ip_forward": newIPForwardingInode(ctx, msrc, s),
 
+		// Allow for configurable ephemeral port ranges. Note that this
+		// controls ports for both IPv4 and IPv6 sockets.
+		"ip_local_port_range": newPortRangeInode(ctx, msrc, s),
+
 		// The following files are simple stubs until they are
 		// implemented in netstack, most of these files are
 		// configuration related. We use the value closest to the
 		// actual netstack behavior or any empty file, all of these
 		// files will have mode 0444 (read-only for all users).
-		"ip_local_port_range":     newStaticProcInode(ctx, msrc, []byte("16000   65535")),
 		"ip_local_reserved_ports": newStaticProcInode(ctx, msrc, []byte("")),
 		"ipfrag_time":             newStaticProcInode(ctx, msrc, []byte("30")),
 		"ip_nonlocal_bind":        newStaticProcInode(ctx, msrc, []byte("0")),
diff --git a/pkg/sentry/fsimpl/proc/tasks_sys.go b/pkg/sentry/fsimpl/proc/tasks_sys.go
index fd7823daa..fb274b78e 100644
--- a/pkg/sentry/fsimpl/proc/tasks_sys.go
+++ b/pkg/sentry/fsimpl/proc/tasks_sys.go
@@ -17,6 +17,7 @@ package proc
 import (
 	"bytes"
 	"fmt"
+	"math"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
@@ -69,17 +70,17 @@ func (fs *filesystem) newSysNetDir(ctx context.Context, root *auth.Credentials,
 	if stack := k.RootNetworkNamespace().Stack(); stack != nil {
 		contents = map[string]kernfs.Inode{
 			"ipv4": fs.newStaticDir(ctx, root, map[string]kernfs.Inode{
-				"tcp_recovery": fs.newInode(ctx, root, 0644, &tcpRecoveryData{stack: stack}),
-				"tcp_rmem":     fs.newInode(ctx, root, 0644, &tcpMemData{stack: stack, dir: tcpRMem}),
-				"tcp_sack":     fs.newInode(ctx, root, 0644, &tcpSackData{stack: stack}),
-				"tcp_wmem":     fs.newInode(ctx, root, 0644, &tcpMemData{stack: stack, dir: tcpWMem}),
-				"ip_forward":   fs.newInode(ctx, root, 0444, &ipForwarding{stack: stack}),
+				"ip_forward":          fs.newInode(ctx, root, 0444, &ipForwarding{stack: stack}),
+				"ip_local_port_range": fs.newInode(ctx, root, 0644, &portRange{stack: stack}),
+				"tcp_recovery":        fs.newInode(ctx, root, 0644, &tcpRecoveryData{stack: stack}),
+				"tcp_rmem":            fs.newInode(ctx, root, 0644, &tcpMemData{stack: stack, dir: tcpRMem}),
+				"tcp_sack":            fs.newInode(ctx, root, 0644, &tcpSackData{stack: stack}),
+				"tcp_wmem":            fs.newInode(ctx, root, 0644, &tcpMemData{stack: stack, dir: tcpWMem}),
 
 				// The following files are simple stubs until they are implemented in
 				// netstack, most of these files are configuration related. We use the
 				// value closest to the actual netstack behavior or any empty file, all
 				// of these files will have mode 0444 (read-only for all users).
-				"ip_local_port_range":     fs.newInode(ctx, root, 0444, newStaticFile("16000   65535")),
 				"ip_local_reserved_ports": fs.newInode(ctx, root, 0444, newStaticFile("")),
 				"ipfrag_time":             fs.newInode(ctx, root, 0444, newStaticFile("30")),
 				"ip_nonlocal_bind":        fs.newInode(ctx, root, 0444, newStaticFile("0")),
@@ -421,3 +422,68 @@ func (ipf *ipForwarding) Write(ctx context.Context, src usermem.IOSequence, offs
 	}
 	return n, nil
 }
+
+// portRange implements vfs.WritableDynamicBytesSource for
+// /proc/sys/net/ipv4/ip_local_port_range.
+//
+// +stateify savable
+type portRange struct {
+	kernfs.DynamicBytesFile
+
+	stack inet.Stack `state:"wait"`
+
+	// start and end store the port range. We must save/restore this here,
+	// since a netstack instance is created on restore.
+	start *uint16
+	end   *uint16
+}
+
+var _ vfs.WritableDynamicBytesSource = (*portRange)(nil)
+
+// Generate implements vfs.DynamicBytesSource.Generate.
+func (pr *portRange) Generate(ctx context.Context, buf *bytes.Buffer) error {
+	if pr.start == nil {
+		start, end := pr.stack.PortRange()
+		pr.start = &start
+		pr.end = &end
+	}
+	_, err := fmt.Fprintf(buf, "%d %d\n", *pr.start, *pr.end)
+	return err
+}
+
+// Write implements vfs.WritableDynamicBytesSource.Write.
+func (pr *portRange) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	if offset != 0 {
+		// No need to handle partial writes thus far.
+		return 0, syserror.EINVAL
+	}
+	if src.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	// Limit input size so as not to impact performance if input size is
+	// large.
+	src = src.TakeFirst(usermem.PageSize - 1)
+
+	ports := make([]int32, 2)
+	n, err := usermem.CopyInt32StringsInVec(ctx, src.IO, src.Addrs, ports, src.Opts)
+	if err != nil {
+		return 0, err
+	}
+
+	// Port numbers must be uint16s.
+	if ports[0] < 0 || ports[1] < 0 || ports[0] > math.MaxUint16 || ports[1] > math.MaxUint16 {
+		return 0, syserror.EINVAL
+	}
+
+	if err := pr.stack.SetPortRange(uint16(ports[0]), uint16(ports[1])); err != nil {
+		return 0, err
+	}
+	if pr.start == nil {
+		pr.start = new(uint16)
+		pr.end = new(uint16)
+	}
+	*pr.start = uint16(ports[0])
+	*pr.end = uint16(ports[1])
+	return n, nil
+}
diff --git a/pkg/sentry/inet/inet.go b/pkg/sentry/inet/inet.go
index f31277d30..6b71bd3a9 100644
--- a/pkg/sentry/inet/inet.go
+++ b/pkg/sentry/inet/inet.go
@@ -93,6 +93,14 @@ type Stack interface {
 
 	// SetForwarding enables or disables packet forwarding between NICs.
 	SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error
+
+	// PortRange returns the UDP and TCP inclusive range of ephemeral ports
+	// used in both IPv4 and IPv6.
+	PortRange() (uint16, uint16)
+
+	// SetPortRange sets the UDP and TCP IPv4 and IPv6 ephemeral port range
+	// (inclusive).
+	SetPortRange(start uint16, end uint16) error
 }
 
 // Interface contains information about a network interface.
diff --git a/pkg/sentry/inet/test_stack.go b/pkg/sentry/inet/test_stack.go
index 9ebeba8a3..03e2608c2 100644
--- a/pkg/sentry/inet/test_stack.go
+++ b/pkg/sentry/inet/test_stack.go
@@ -164,3 +164,15 @@ func (s *TestStack) SetForwarding(protocol tcpip.NetworkProtocolNumber, enable b
 	s.IPForwarding = enable
 	return nil
 }
+
+// PortRange implements inet.Stack.PortRange.
+func (*TestStack) PortRange() (uint16, uint16) {
+	// Use the default Linux values per net/ipv4/af_inet.c:inet_init_net().
+	return 32768, 28232
+}
+
+// SetPortRange implements inet.Stack.SetPortRange.
+func (*TestStack) SetPortRange(start uint16, end uint16) error {
+	// No-op.
+	return nil
+}
diff --git a/pkg/sentry/socket/hostinet/stack.go b/pkg/sentry/socket/hostinet/stack.go
index e6323244c..5bcf92e14 100644
--- a/pkg/sentry/socket/hostinet/stack.go
+++ b/pkg/sentry/socket/hostinet/stack.go
@@ -504,3 +504,14 @@ func (s *Stack) Forwarding(protocol tcpip.NetworkProtocolNumber) bool {
 func (s *Stack) SetForwarding(tcpip.NetworkProtocolNumber, bool) error {
 	return syserror.EACCES
 }
+
+// PortRange implements inet.Stack.PortRange.
+func (*Stack) PortRange() (uint16, uint16) {
+	// Use the default Linux values per net/ipv4/af_inet.c:inet_init_net().
+	return 32768, 28232
+}
+
+// SetPortRange implements inet.Stack.SetPortRange.
+func (*Stack) SetPortRange(start uint16, end uint16) error {
+	return syserror.EACCES
+}
diff --git a/pkg/sentry/socket/netstack/stack.go b/pkg/sentry/socket/netstack/stack.go
index 71c3bc034..b215067cf 100644
--- a/pkg/sentry/socket/netstack/stack.go
+++ b/pkg/sentry/socket/netstack/stack.go
@@ -478,3 +478,13 @@ func (s *Stack) SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool)
 	}
 	return nil
 }
+
+// PortRange implements inet.Stack.PortRange.
+func (s *Stack) PortRange() (uint16, uint16) {
+	return s.Stack.PortRange()
+}
+
+// SetPortRange implements inet.Stack.SetPortRange.
+func (s *Stack) SetPortRange(start uint16, end uint16) error {
+	return syserr.TranslateNetstackError(s.Stack.SetPortRange(start, end)).ToError()
+}