Merge pull request #3516 from cgwalters/remount-shutdown

Add ostree-shutdown.service: hide /sysroot and make /etc read-only

Author: champtar

Makefile-boot.am

@@ -38,6 +38,7 @@ endif
 if BUILDOPT_SYSTEMD
 systemdsystemunit_DATA = src/boot/ostree-prepare-root.service \
 	src/boot/ostree-remount.service \
+	src/boot/ostree-shutdown.service \
 	src/boot/ostree-boot-complete.service \
 	src/boot/ostree-finalize-staged.service \
 	src/boot/ostree-finalize-staged-hold.service \
@@ -69,6 +70,7 @@ EXTRA_DIST += src/boot/dracut/module-setup.sh \
 	src/boot/ostree-boot-complete.service \
 	src/boot/ostree-prepare-root.service \
 	src/boot/ostree-remount.service \
+	src/boot/ostree-shutdown.service \
 	src/boot/ostree-finalize-staged.service \
 	src/boot/ostree-finalize-staged-hold.service \
 	src/boot/ostree-state-overlay@.service \

src/boot/ostree-shutdown.service

@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: LGPL-2.0+
+
+[Unit]
+Description=OSTree Shutdown
+Documentation=man:ostree(1)
+DefaultDependencies=no
+# Only enabled via generator, but for good measure
+ConditionKernelCommandLine=ostree
+# Run after core mounts
+RequiresMountsFor=/etc /sysroot
+# However, we want to only shut down after `/var` has been umounted.
+# Since this runs via ExecStop, this Before= is actually After= at shutdown
+Before=var.mount
+Conflicts=umount.target
+Before=umount.target
+
+[Service]
+Type=oneshot
+RemainAfterExit=yes
+ExecStop=/usr/lib/ostree/ostree-remount --shutdown
+
+# No [Install] section - we're only enabled via generator

src/libostree/ostree-impl-system-generator.c

@@ -108,6 +108,10 @@ require_internal_units (const char *normal_dir, const char *early_dir, const cha
                  "local-fs.target.requires/ostree-remount.service")
       < 0)
     return glnx_throw_errno_prefix (error, "symlinkat");
+  if (symlinkat (SYSTEM_DATA_UNIT_PATH "/ostree-shutdown.service", normal_dir_dfd,
+                 "local-fs.target.wants/ostree-shutdown.service")
+      < 0)
+    return glnx_throw_errno_prefix (error, "symlinkat");
 
   if (!glnx_shutil_mkdir_p_at (normal_dir_dfd, "multi-user.target.wants", 0755, cancellable, error))
     return FALSE;

src/switchroot/ostree-remount.c

@@ -42,6 +42,12 @@
 #include "ostree-mount-util.h"
 #include "otcore.h"
 
+static gboolean opt_shutdown;
+
+static GOptionEntry options[] = { { "shutdown", 'S', 0, G_OPTION_ARG_NONE, &opt_shutdown,
+                                    "Perform shutdown unmounting", NULL },
+                                  { NULL } };
+
 static void
 do_remount (const char *target, bool writable)
 {
@@ -133,10 +139,61 @@ relabel_dir_for_upper (const char *upper_path, const char *real_path, gboolean i
 #endif
 }
 
+// ostree-prepare-root sets things up so that /sysroot points to the "physical" (real) root in the
+// initramfs, and then with composefs `/` is an overlay+EROFS that holds references to content in
+// that physical filesystem.
+//
+// In a typical mutable system where the OS is in a mutable `/` (or `/usr), systemd explicitly
+// skips unmounting both `/` and `/usr`. It will remount them read-only though - and that's
+// the semantic we want to match here.
+static void
+do_shutdown (void)
+{
+  const char *sysroot = "/sysroot";
+  if (mount (sysroot, sysroot, NULL, MS_REMOUNT | MS_SILENT | MS_RDONLY, NULL) < 0)
+    {
+      // Hopefully at this point nothing has any write references, but if they
+      // do we still want to continue.
+      perror ("Remounting /sysroot read-only");
+    }
+  // And fully detach it from the mountns because otherwise systemd thinks
+  // it can be unmounted, but it can't - it's required by `/` (and in a
+  // composefs setup `/etc`) and possibly `/var`. Again, we only really
+  // care that it got mounted read-only and hence outstanding data flushed.
+  // A better fix in the future would be to teach systemd to honor `-.mount`
+  // having a `Requires=sysroot.mount` meaning we can't unmount the latter.
+  if (umount2 (sysroot, MNT_DETACH) < 0)
+    err (EXIT_FAILURE, "umount(/sysroot)");
+
+  // And finally: /etc
+  // NOTE! This one is intentionally last in that we want to try to make
+  // this read-only, but if it fails, systemd-shutdown will have another
+  // attempt after a process killing spree. If anything happens to be
+  // holding a writable fd at this point, conceptually it would have
+  // created race conditions vs ostree-finalize-staged.service, and so
+  // having this service fail will be a signal that those things need
+  // to be fixed.
+  do_remount ("/etc", false);
+  // Don't add anything else after this.
+}
+
 int
 main (int argc, char *argv[])
 {
   g_autoptr (GError) error = NULL;
+  g_autoptr (GOptionContext) context = g_option_context_new ("");
+  g_option_context_add_main_entries (context, options, NULL);
+  if (!g_option_context_parse (context, &argc, &argv, &error))
+    errx (EXIT_FAILURE, "Error parsing options: %s", error->message);
+
+  // Handle the shutdown option
+  if (opt_shutdown)
+    {
+      do_shutdown ();
+      return 0;
+    }
+  // Otherwise fall through to the default startup
+
   g_autoptr (GVariant) ostree_run_metadata_v = NULL;
   {
     glnx_autofd int fd = open (OTCORE_RUN_BOOTED, O_RDONLY | O_CLOEXEC);