container: improve capability handling

This cleans up preserving caps for expansion and correctly sets privileged caps.

Signed-off-by: Ophestra <cat@gensokyo.uk>

3 files changed, 36 insertions, 26 deletions

diff --git a/container/container_test.go b/container/container_test.go
index f34035b2..48b6e8a4 100644
--- a/container/container_test.go
+++ b/container/container_test.go
@@ -537,7 +537,6 @@ func testContainerCancel(
 	containerExtra func(c *container.Container),
 	waitCheck func(ps *os.ProcessState, waitErr error),
 ) {
-	t.Parallel()
 	ctx, cancel := context.WithCancel(t.Context())
 
 	c := helperNewContainer(ctx, "block")
diff --git a/container/init.go b/container/init.go
index 76a77828..78c95439 100644
--- a/container/init.go
+++ b/container/init.go
@@ -186,19 +186,24 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	if err := k.setDumpable(ext.SUID_DUMP_USER); err != nil {
 		k.fatalf(msg, "cannot set SUID_DUMP_USER: %v", err)
 	}
-	if err := k.writeFile(fhs.Proc+"self/uid_map",
-		append([]byte{}, strconv.Itoa(param.Uid)+" "+strconv.Itoa(param.HostUid)+" 1\n"...),
-		0); err != nil {
+	if err := k.writeFile(
+		fhs.Proc+"self/uid_map",
+		[]byte(strconv.Itoa(param.Uid)+" "+strconv.Itoa(param.HostUid)+" 1\n"),
+		0,
+	); err != nil {
 		k.fatalf(msg, "%v", err)
 	}
-	if err := k.writeFile(fhs.Proc+"self/setgroups",
+	if err := k.writeFile(
+		fhs.Proc+"self/setgroups",
 		[]byte("deny\n"),
-		0); err != nil && !os.IsNotExist(err) {
+		0,
+	); err != nil && !os.IsNotExist(err) {
 		k.fatalf(msg, "%v", err)
 	}
 	if err := k.writeFile(fhs.Proc+"self/gid_map",
-		append([]byte{}, strconv.Itoa(param.Gid)+" "+strconv.Itoa(param.HostGid)+" 1\n"...),
-		0); err != nil {
+		[]byte(strconv.Itoa(param.Gid)+" "+strconv.Itoa(param.HostGid)+" 1\n"),
+		0,
+	); err != nil {
 		k.fatalf(msg, "%v", err)
 	}
 	if err := k.setDumpable(ext.SUID_DUMP_DISABLE); err != nil {
@@ -323,11 +328,16 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 		}
 	}
 
+	var keepCaps []uintptr
+	if param.Privileged {
+		keepCaps = append(keepCaps, CAP_SYS_ADMIN, CAP_SETPCAP)
+	}
+
 	if err := k.capAmbientClearAll(); err != nil {
 		k.fatalf(msg, "cannot clear the ambient capability set: %v", err)
 	}
-	for i := uintptr(0); i <= lastcap; i++ {
-		if param.Privileged && i == CAP_SYS_ADMIN {
+	for i := range lastcap + 1 {
+		if slices.Contains(keepCaps, i) {
 			continue
 		}
 		if err := k.capBoundingSetDrop(i); err != nil {
@@ -336,20 +346,23 @@ func initEntrypoint(k syscallDispatcher, msg message.Msg) {
 	}
 
 	var keep [2]uint32
-	if param.Privileged {
-		keep[capToIndex(CAP_SYS_ADMIN)] |= capToMask(CAP_SYS_ADMIN)
-
-		if err := k.capAmbientRaise(CAP_SYS_ADMIN); err != nil {
-			k.fatalf(msg, "cannot raise CAP_SYS_ADMIN: %v", err)
-		}
+	for _, c := range keepCaps {
+		keep[capToIndex(c)] |= capToMask(c)
 	}
+
 	if err := k.capset(
 		&capHeader{_LINUX_CAPABILITY_VERSION_3, 0},
-		&[2]capData{{0, keep[0], keep[0]}, {0, keep[1], keep[1]}},
+		&[2]capData{{keep[0], keep[0], keep[0]}, {keep[1], keep[1], keep[1]}},
 	); err != nil {
 		k.fatalf(msg, "cannot capset: %v", err)
 	}
 
+	for _, c := range keepCaps {
+		if err := k.capAmbientRaise(c); err != nil {
+			k.fatalf(msg, "cannot raise %#x: %v", c, err)
+		}
+	}
+
 	if !param.SeccompDisable {
 		rules := param.SeccompRules
 		if len(rules) == 0 { // non-empty rules slice always overrides presets
diff --git a/container/init_test.go b/container/init_test.go
index a27391a2..23d0a7b3 100644
--- a/container/init_test.go
+++ b/container/init_test.go
@@ -1624,7 +1624,6 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x5)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x6)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x7)}, nil, nil),
-				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x8)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x9)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xa)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xb)}, nil, nil),
@@ -1656,8 +1655,9 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x26)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x27)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x28)}, nil, nil),
+				call("capset", stub.ExpectArgs{&capHeader{_LINUX_CAPABILITY_VERSION_3, 0}, &[2]capData{{0x200100, 0x200100, 0x200100}, {0, 0, 0}}}, nil, nil),
 				call("capAmbientRaise", stub.ExpectArgs{uintptr(0x15)}, nil, stub.UniqueError(19)),
-				call("fatalf", stub.ExpectArgs{"cannot raise CAP_SYS_ADMIN: %v", []any{stub.UniqueError(19)}}, nil, nil),
+				call("fatalf", stub.ExpectArgs{"cannot raise %#x: %v", []any{uintptr(0x15), stub.UniqueError(19)}}, nil, nil),
 			},
 		}, nil},
 
@@ -1731,7 +1731,6 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x5)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x6)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x7)}, nil, nil),
-				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x8)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x9)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xa)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xb)}, nil, nil),
@@ -1763,8 +1762,7 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x26)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x27)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x28)}, nil, nil),
-				call("capAmbientRaise", stub.ExpectArgs{uintptr(0x15)}, nil, nil),
-				call("capset", stub.ExpectArgs{&capHeader{_LINUX_CAPABILITY_VERSION_3, 0}, &[2]capData{{0, 0x200000, 0x200000}, {0, 0, 0}}}, nil, stub.UniqueError(17)),
+				call("capset", stub.ExpectArgs{&capHeader{_LINUX_CAPABILITY_VERSION_3, 0}, &[2]capData{{0x200100, 0x200100, 0x200100}, {0, 0, 0}}}, nil, stub.UniqueError(17)),
 				call("fatalf", stub.ExpectArgs{"cannot capset: %v", []any{stub.UniqueError(17)}}, nil, nil),
 			},
 		}, nil},
@@ -1839,7 +1837,6 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x5)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x6)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x7)}, nil, nil),
-				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x8)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x9)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xa)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xb)}, nil, nil),
@@ -1871,8 +1868,9 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x26)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x27)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x28)}, nil, nil),
+				call("capset", stub.ExpectArgs{&capHeader{_LINUX_CAPABILITY_VERSION_3, 0}, &[2]capData{{0x200100, 0x200100, 0x200100}, {0, 0, 0}}}, nil, nil),
 				call("capAmbientRaise", stub.ExpectArgs{uintptr(0x15)}, nil, nil),
-				call("capset", stub.ExpectArgs{&capHeader{_LINUX_CAPABILITY_VERSION_3, 0}, &[2]capData{{0, 0x200000, 0x200000}, {0, 0, 0}}}, nil, nil),
+				call("capAmbientRaise", stub.ExpectArgs{uintptr(0x8)}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"resolving presets %#x", []any{std.FilterPreset(0xf)}}, nil, nil),
 				call("seccompLoad", stub.ExpectArgs{seccomp.Preset(0xf, 0), seccomp.ExportFlag(0)}, nil, stub.UniqueError(15)),
 				call("fatalf", stub.ExpectArgs{"cannot load syscall filter: %v", []any{stub.UniqueError(15)}}, nil, nil),
@@ -2699,7 +2697,6 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x5)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x6)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x7)}, nil, nil),
-				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x8)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x9)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xa)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0xb)}, nil, nil),
@@ -2731,8 +2728,9 @@ func TestInitEntrypoint(t *testing.T) {
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x26)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x27)}, nil, nil),
 				call("capBoundingSetDrop", stub.ExpectArgs{uintptr(0x28)}, nil, nil),
+				call("capset", stub.ExpectArgs{&capHeader{_LINUX_CAPABILITY_VERSION_3, 0}, &[2]capData{{0x200100, 0x200100, 0x200100}, {0, 0, 0}}}, nil, nil),
 				call("capAmbientRaise", stub.ExpectArgs{uintptr(0x15)}, nil, nil),
-				call("capset", stub.ExpectArgs{&capHeader{_LINUX_CAPABILITY_VERSION_3, 0}, &[2]capData{{0, 0x200000, 0x200000}, {0, 0, 0}}}, nil, nil),
+				call("capAmbientRaise", stub.ExpectArgs{uintptr(0x8)}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"resolving presets %#x", []any{std.FilterPreset(0xf)}}, nil, nil),
 				call("seccompLoad", stub.ExpectArgs{seccomp.Preset(0xf, 0), seccomp.ExportFlag(0)}, nil, nil),
 				call("verbosef", stub.ExpectArgs{"%d filter rules loaded", []any{73}}, nil, nil),