tmpfs: Allow xattrs in the trusted namespace if creds has CAP_SYS_ADMIN.

This is needed to support the overlay opaque attribute.

PiperOrigin-RevId: 328552985

pkg/abi/linux/xattr.go

@@ -23,6 +23,9 @@ const (
 	XATTR_CREATE  = 1
 	XATTR_REPLACE = 2
 
+	XATTR_TRUSTED_PREFIX     = "trusted."
+	XATTR_TRUSTED_PREFIX_LEN = len(XATTR_TRUSTED_PREFIX)
+
 	XATTR_USER_PREFIX     = "user."
 	XATTR_USER_PREFIX_LEN = len(XATTR_USER_PREFIX)
 )

pkg/sentry/fsimpl/overlay/filesystem.go

@@ -30,7 +30,7 @@ import (
 // _OVL_XATTR_OPAQUE is an extended attribute key whose value is set to "y" for
 // opaque directories.
 // Linux: fs/overlayfs/overlayfs.h:OVL_XATTR_OPAQUE
-const _OVL_XATTR_OPAQUE = "trusted.overlay.opaque"
+const _OVL_XATTR_OPAQUE = linux.XATTR_TRUSTED_PREFIX + "overlay.opaque"
 
 func isWhiteout(stat *linux.Statx) bool {
 	return stat.Mode&linux.S_IFMT == linux.S_IFCHR && stat.RdevMajor == 0 && stat.RdevMinor == 0

pkg/sentry/fsimpl/tmpfs/tmpfs.go

@@ -631,49 +631,65 @@ func (i *inode) listxattr(size uint64) ([]string, error) {
 }
 
 func (i *inode) getxattr(creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) {
-	if err := i.checkPermissions(creds, vfs.MayRead); err != nil {
+	if err := i.checkXattrPermissions(creds, opts.Name, vfs.MayRead); err != nil {
 		return "", err
 	}
-	if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) {
-		return "", syserror.EOPNOTSUPP
-	}
-	if !i.userXattrSupported() {
-		return "", syserror.ENODATA
-	}
 	return i.xattrs.Getxattr(opts)
 }
 
 func (i *inode) setxattr(creds *auth.Credentials, opts *vfs.SetxattrOptions) error {
-	if err := i.checkPermissions(creds, vfs.MayWrite); err != nil {
+	if err := i.checkXattrPermissions(creds, opts.Name, vfs.MayWrite); err != nil {
 		return err
 	}
-	if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) {
-		return syserror.EOPNOTSUPP
-	}
-	if !i.userXattrSupported() {
-		return syserror.EPERM
-	}
 	return i.xattrs.Setxattr(opts)
 }
 
 func (i *inode) removexattr(creds *auth.Credentials, name string) error {
-	if err := i.checkPermissions(creds, vfs.MayWrite); err != nil {
+	if err := i.checkXattrPermissions(creds, name, vfs.MayWrite); err != nil {
 		return err
 	}
-	if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) {
-		return syserror.EOPNOTSUPP
-	}
-	if !i.userXattrSupported() {
-		return syserror.EPERM
-	}
 	return i.xattrs.Removexattr(name)
 }
 
-// Extended attributes in the user.* namespace are only supported for regular
-// files and directories.
-func (i *inode) userXattrSupported() bool {
-	filetype := linux.S_IFMT & atomic.LoadUint32(&i.mode)
-	return filetype == linux.S_IFREG || filetype == linux.S_IFDIR
+func (i *inode) checkXattrPermissions(creds *auth.Credentials, name string, ats vfs.AccessTypes) error {
+	switch {
+	case ats&vfs.MayRead == vfs.MayRead:
+		if err := i.checkPermissions(creds, vfs.MayRead); err != nil {
+			return err
+		}
+	case ats&vfs.MayWrite == vfs.MayWrite:
+		if err := i.checkPermissions(creds, vfs.MayWrite); err != nil {
+			return err
+		}
+	default:
+		panic(fmt.Sprintf("checkXattrPermissions called with impossible AccessTypes: %v", ats))
+	}
+
+	switch {
+	case strings.HasPrefix(name, linux.XATTR_TRUSTED_PREFIX):
+		// The trusted.* namespace can only be accessed by privileged
+		// users.
+		if creds.HasCapability(linux.CAP_SYS_ADMIN) {
+			return nil
+		}
+		if ats&vfs.MayWrite == vfs.MayWrite {
+			return syserror.EPERM
+		}
+		return syserror.ENODATA
+	case strings.HasPrefix(name, linux.XATTR_USER_PREFIX):
+		// Extended attributes in the user.* namespace are only
+		// supported for regular files and directories.
+		filetype := linux.S_IFMT & atomic.LoadUint32(&i.mode)
+		if filetype == linux.S_IFREG || filetype == linux.S_IFDIR {
+			return nil
+		}
+		if ats&vfs.MayWrite == vfs.MayWrite {
+			return syserror.EPERM
+		}
+		return syserror.ENODATA
+
+	}
+	return syserror.EOPNOTSUPP
 }
 
 // fileDescription is embedded by tmpfs implementations of

test/syscalls/linux/memfd.cc

@@ -14,12 +14,10 @@
 
 #include <errno.h>
 #include <fcntl.h>
-#include <linux/magic.h>
 #include <linux/memfd.h>
 #include <linux/unistd.h>
 #include <string.h>
 #include <sys/mman.h>
-#include <sys/statfs.h>
 #include <sys/syscall.h>
 
 #include <vector>
@@ -53,6 +51,7 @@ namespace {
 #define F_SEAL_GROW 0x0004
 #define F_SEAL_WRITE 0x0008
 
+using ::gvisor::testing::IsTmpfs;
 using ::testing::StartsWith;
 
 const std::string kMemfdName = "some-memfd";
@@ -444,20 +443,6 @@ TEST(MemfdTest, SealsAreInodeLevelProperties) {
   EXPECT_THAT(ftruncate(memfd3.get(), kPageSize), SyscallFailsWithErrno(EPERM));
 }
 
-PosixErrorOr<bool> IsTmpfs(const std::string& path) {
-  struct statfs stat;
-  if (statfs(path.c_str(), &stat)) {
-    if (errno == ENOENT) {
-      // Nothing at path, don't raise this as an error. Instead, just report no
-      // tmpfs at path.
-      return false;
-    }
-    return PosixError(errno,
-                      absl::StrFormat("statfs(\"%s\", %#p)", path, &stat));
-  }
-  return stat.f_type == TMPFS_MAGIC;
-}
-
 // Tmpfs files also support seals, but are created with F_SEAL_SEAL.
 TEST(MemfdTest, TmpfsFilesHaveSealSeal) {
   SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(IsTmpfs("/tmp")));

test/syscalls/linux/xattr.cc

@@ -28,6 +28,7 @@
 #include "test/syscalls/linux/file_base.h"
 #include "test/util/capability_util.h"
 #include "test/util/file_descriptor.h"
+#include "test/util/fs_util.h"
 #include "test/util/posix_error.h"
 #include "test/util/temp_path.h"
 #include "test/util/test_util.h"
@@ -37,6 +38,8 @@ namespace testing {
 
 namespace {
 
+using ::gvisor::testing::IsTmpfs;
+
 class XattrTest : public FileTest {};
 
 TEST_F(XattrTest, XattrNonexistentFile) {
@@ -604,6 +607,77 @@ TEST_F(XattrTest, XattrWithFD) {
   EXPECT_THAT(fremovexattr(fd.get(), name), SyscallSucceeds());
 }
 
+TEST_F(XattrTest, TrustedNamespaceWithCapSysAdmin) {
+  // Trusted namespace not supported in VFS1.
+  SKIP_IF(IsRunningWithVFS1());
+
+  // TODO(b/66162845): Only gVisor tmpfs currently supports trusted namespace.
+  SKIP_IF(IsRunningOnGvisor() &&
+          !ASSERT_NO_ERRNO_AND_VALUE(IsTmpfs(test_file_name_)));
+
+  // Setting/Getting in the trusted namespace requires CAP_SYS_ADMIN.
+  SKIP_IF(!ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_SYS_ADMIN)));
+
+  const char* path = test_file_name_.c_str();
+  const char name[] = "trusted.test";
+
+  // Set.
+  char val = 'a';
+  size_t size = sizeof(val);
+  EXPECT_THAT(setxattr(path, name, &val, size, /*flags=*/0), SyscallSucceeds());
+
+  // Get.
+  char got = '\0';
+  EXPECT_THAT(getxattr(path, name, &got, size), SyscallSucceedsWithValue(size));
+  EXPECT_EQ(val, got);
+
+  // List.
+  char list[sizeof(name)];
+  EXPECT_THAT(listxattr(path, list, sizeof(list)),
+              SyscallSucceedsWithValue(sizeof(name)));
+  EXPECT_STREQ(list, name);
+
+  // Remove.
+  EXPECT_THAT(removexattr(path, name), SyscallSucceeds());
+
+  // Get should now return ENODATA.
+  EXPECT_THAT(getxattr(path, name, &got, size), SyscallFailsWithErrno(ENODATA));
+}
+
+TEST_F(XattrTest, TrustedNamespaceWithoutCapSysAdmin) {
+  // Trusted namespace not supported in VFS1.
+  SKIP_IF(IsRunningWithVFS1());
+
+  // TODO(b/66162845): Only gVisor tmpfs currently supports trusted namespace.
+  SKIP_IF(IsRunningOnGvisor() &&
+          !ASSERT_NO_ERRNO_AND_VALUE(IsTmpfs(test_file_name_)));
+
+  // Drop CAP_SYS_ADMIN if we have it.
+  if (ASSERT_NO_ERRNO_AND_VALUE(HaveCapability(CAP_SYS_ADMIN))) {
+    EXPECT_NO_ERRNO(SetCapability(CAP_SYS_ADMIN, false));
+  }
+
+  const char* path = test_file_name_.c_str();
+  const char name[] = "trusted.test";
+
+  // Set fails.
+  char val = 'a';
+  size_t size = sizeof(val);
+  EXPECT_THAT(setxattr(path, name, &val, size, /*flags=*/0),
+              SyscallFailsWithErrno(EPERM));
+
+  // Get fails.
+  char got = '\0';
+  EXPECT_THAT(getxattr(path, name, &got, size), SyscallFailsWithErrno(ENODATA));
+
+  // List still works, but returns no items.
+  char list[sizeof(name)];
+  EXPECT_THAT(listxattr(path, list, sizeof(list)), SyscallSucceedsWithValue(0));
+
+  // Remove fails.
+  EXPECT_THAT(removexattr(path, name), SyscallFailsWithErrno(EPERM));
+}
+
 }  // namespace
 
 }  // namespace testing

test/util/fs_util.cc

@@ -15,7 +15,11 @@
 #include "test/util/fs_util.h"
 
 #include <dirent.h>
+#ifndef __fuchsia__
+#include <linux/magic.h>
+#endif  // __fuchsia__
 #include <sys/stat.h>
+#include <sys/statfs.h>
 #include <sys/types.h>
 #include <unistd.h>
 
@@ -629,5 +633,21 @@ PosixErrorOr<std::string> ProcessExePath(int pid) {
   return ReadLink(absl::StrCat("/proc/", pid, "/exe"));
 }
 
+#ifndef __fuchsia__
+PosixErrorOr<bool> IsTmpfs(const std::string& path) {
+  struct statfs stat;
+  if (statfs(path.c_str(), &stat)) {
+    if (errno == ENOENT) {
+      // Nothing at path, don't raise this as an error. Instead, just report no
+      // tmpfs at path.
+      return false;
+    }
+    return PosixError(errno,
+                      absl::StrFormat("statfs(\"%s\", %#p)", path, &stat));
+  }
+  return stat.f_type == TMPFS_MAGIC;
+}
+#endif  // __fuchsia__
+
 }  // namespace testing
 }  // namespace gvisor

test/util/fs_util.h

@@ -17,6 +17,7 @@
 
 #include <dirent.h>
 #include <sys/stat.h>
+#include <sys/statfs.h>
 #include <sys/types.h>
 #include <unistd.h>
 
@@ -178,6 +179,11 @@ std::string CleanPath(absl::string_view path);
 // Returns the full path to the executable of the given pid or a PosixError.
 PosixErrorOr<std::string> ProcessExePath(int pid);
 
+#ifndef __fuchsia__
+// IsTmpfs returns true if the file at path is backed by tmpfs.
+PosixErrorOr<bool> IsTmpfs(const std::string& path);
+#endif  // __fucshia__
+
 namespace internal {
 // Not part of the public API.
 std::string JoinPathImpl(std::initializer_list<absl::string_view> paths);