]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
selftests/landlock: Check that coredump sockets stay unrestricted
authorGünther Noack <gnoack3000@gmail.com>
Fri, 27 Mar 2026 16:48:35 +0000 (17:48 +0100)
committerMickaël Salaün <mic@digikod.net>
Tue, 7 Apr 2026 16:51:10 +0000 (18:51 +0200)
Even when a process is restricted with the new
LANDLOCK_ACCESS_FS_RESOLVE_UNIX right, the kernel can continue writing
its coredump to the configured coredump socket.

In the test, we create a local server and rewire the system to write
coredumps into it.  We then create a child process within a Landlock
domain where LANDLOCK_ACCESS_FS_RESOLVE_UNIX is restricted and make
the process crash.  The test uses SO_PEERCRED to check that the
connecting client process is the expected one.

Includes a fix by Mickaël Salaün for setting the EUID to 0 (see [1]).

Link[1]: https://lore.kernel.org/all/20260218.ohth8theu8Yi@digikod.net/
Suggested-by: Mickaël Salaün <mic@digikod.net>
Signed-off-by: Günther Noack <gnoack3000@gmail.com>
Link: https://lore.kernel.org/r/20260327164838.38231-11-gnoack3000@gmail.com
Signed-off-by: Mickaël Salaün <mic@digikod.net>
tools/testing/selftests/landlock/fs_test.c

index 3dad643741f709197cbdd2d87606459575a7994b..af0f0b16129aba0cb8600ec66d129322f4312c3e 100644 (file)
@@ -22,6 +22,7 @@
 #include <sys/ioctl.h>
 #include <sys/mount.h>
 #include <sys/prctl.h>
+#include <sys/resource.h>
 #include <sys/sendfile.h>
 #include <sys/socket.h>
 #include <sys/stat.h>
@@ -4928,6 +4929,148 @@ TEST_F(scoped_domains, unix_seqpacket_connect_to_child_full)
 #undef USE_SENDTO
 #undef ENFORCE_ALL
 
+static void read_core_pattern(struct __test_metadata *const _metadata,
+                             char *buf, size_t buf_size)
+{
+       int fd;
+       ssize_t ret;
+
+       fd = open("/proc/sys/kernel/core_pattern", O_RDONLY | O_CLOEXEC);
+       ASSERT_LE(0, fd);
+
+       ret = read(fd, buf, buf_size - 1);
+       ASSERT_LE(0, ret);
+       EXPECT_EQ(0, close(fd));
+
+       buf[ret] = '\0';
+}
+
+static void set_core_pattern(struct __test_metadata *const _metadata,
+                            const char *pattern)
+{
+       int fd;
+       size_t len = strlen(pattern);
+
+       /*
+        * Writing to /proc/sys/kernel/core_pattern requires EUID 0 because
+        * sysctl_perm() checks that, ignoring capabilities like
+        * CAP_SYS_ADMIN or CAP_DAC_OVERRIDE.
+        *
+        * Switching EUID clears the dumpable flag, which must be restored
+        * afterwards to allow coredumps.
+        */
+       set_cap(_metadata, CAP_SETUID);
+       ASSERT_EQ(0, seteuid(0));
+       clear_cap(_metadata, CAP_SETUID);
+
+       fd = open("/proc/sys/kernel/core_pattern", O_WRONLY | O_CLOEXEC);
+       ASSERT_LE(0, fd)
+       {
+               TH_LOG("Failed to open core_pattern for writing: %s",
+                      strerror(errno));
+       }
+
+       ASSERT_EQ(len, write(fd, pattern, len));
+       EXPECT_EQ(0, close(fd));
+
+       set_cap(_metadata, CAP_SETUID);
+       ASSERT_EQ(0, seteuid(getuid()));
+       clear_cap(_metadata, CAP_SETUID);
+
+       /* Restore dumpable flag cleared by seteuid(). */
+       ASSERT_EQ(0, prctl(PR_SET_DUMPABLE, 1, 0, 0, 0));
+}
+
+FIXTURE(coredump)
+{
+       char original_core_pattern[256];
+};
+
+FIXTURE_SETUP(coredump)
+{
+       disable_caps(_metadata);
+       read_core_pattern(_metadata, self->original_core_pattern,
+                         sizeof(self->original_core_pattern));
+}
+
+FIXTURE_TEARDOWN_PARENT(coredump)
+{
+       set_core_pattern(_metadata, self->original_core_pattern);
+}
+
+/*
+ * Test that even when a process is restricted with
+ * LANDLOCK_ACCESS_FS_RESOLVE_UNIX, the kernel can still initiate a connection
+ * to the coredump socket on the processes' behalf.
+ */
+TEST_F_FORK(coredump, socket_not_restricted)
+{
+       static const char core_pattern[] = "@/tmp/landlock_coredump_test.sock";
+       const char *const sock_path = core_pattern + 1;
+       int srv_fd, conn_fd, status;
+       pid_t child_pid;
+       struct ucred cred;
+       socklen_t cred_len = sizeof(cred);
+       char buf[4096];
+
+       /* Set up the coredump server socket. */
+       unlink(sock_path);
+       srv_fd = set_up_named_unix_server(_metadata, SOCK_STREAM, sock_path);
+
+       /* Point coredumps at our socket. */
+       set_core_pattern(_metadata, core_pattern);
+
+       /* Restrict LANDLOCK_ACCESS_FS_RESOLVE_UNIX. */
+       drop_access_rights(_metadata,
+                          &(struct landlock_ruleset_attr){
+                                  .handled_access_fs =
+                                          LANDLOCK_ACCESS_FS_RESOLVE_UNIX,
+                          });
+
+       /* Fork a child that crashes. */
+       child_pid = fork();
+       ASSERT_LE(0, child_pid);
+       if (child_pid == 0) {
+               struct rlimit rl = {
+                       .rlim_cur = RLIM_INFINITY,
+                       .rlim_max = RLIM_INFINITY,
+               };
+
+               ASSERT_EQ(0, setrlimit(RLIMIT_CORE, &rl));
+
+               /* Crash on purpose. */
+               kill(getpid(), SIGSEGV);
+               _exit(1);
+       }
+
+       /*
+        * Accept the coredump connection.  If Landlock incorrectly denies the
+        * kernel's coredump connect, accept() will block forever, so the test
+        * would time out.
+        */
+       conn_fd = accept(srv_fd, NULL, NULL);
+       ASSERT_LE(0, conn_fd);
+
+       /* Check that the connection came from the crashing child. */
+       ASSERT_EQ(0, getsockopt(conn_fd, SOL_SOCKET, SO_PEERCRED, &cred,
+                               &cred_len));
+       EXPECT_EQ(child_pid, cred.pid);
+
+       /* Drain the coredump data so the kernel can finish. */
+       while (read(conn_fd, buf, sizeof(buf)) > 0)
+               ;
+
+       EXPECT_EQ(0, close(conn_fd));
+
+       /* Wait for the child and verify it coredumped. */
+       ASSERT_EQ(child_pid, waitpid(child_pid, &status, 0));
+       ASSERT_TRUE(WIFSIGNALED(status));
+       ASSERT_TRUE(WCOREDUMP(status));
+
+       EXPECT_EQ(0, close(srv_fd));
+       EXPECT_EQ(0, unlink(sock_path));
+}
+
 /* clang-format off */
 FIXTURE(layout1_bind) {};
 /* clang-format on */