1 files changed, 29 insertions, 16 deletions

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