Handle cases where the kernel incorrectly reports the filename for an…

… execed binary is "/"

Resolves #3648

CMakeLists.txt

@@ -1105,6 +1105,7 @@ set(BASIC_TESTS
   x86/modify_ldt
   mount_ns_exec
   mount_ns_exec2
+  mount_ns_execveat
   mprotect
   mprotect_heterogenous
   mprotect_none

src/record_syscall.cc

@@ -5526,6 +5526,17 @@ static pair<remote_ptr<void>, remote_ptr<void>> get_exe_entry_interp_base(Task*
  */
 static string try_make_process_file_name(RecordTask* t,
                                          const std::string& file_name) {
+  if (file_name == "/") {
+    // The kernel sometimes returns "/" as the file name of a mapped
+    // segment. (We've seen this when the actual file was a bind-mount
+    // that has been opened with O_PATH and then lazily unmounted, then
+    // the O_PATH fd was execed with `execveat`.)
+    // In that case, we have no idea what the actual file is so
+    // return a name that will never work instead "/proc/.../root/"
+    // which can be stat(), causing confusion.
+    return "";
+  }
+
   char proc_root[32];
   // /proc/<pid>/root has magical properties; not only is it a link, but
   // it links to a view of the filesystem as the process sees it, taking into
@@ -5703,6 +5714,7 @@ static void process_execve(RecordTask* t, TaskSyscallState& syscall_state) {
       continue;
     }
     struct stat st;
+    // Maybe we should use /proc/.../map_files instead?
     string file_name = try_make_process_file_name(t, km.fsname());
     if (stat(file_name.c_str(), &st) != 0) {
       st = km.fake_stat();

src/test/mount_ns_execveat.c

@@ -0,0 +1,30 @@
+/* -*- Mode: C; tab-width: 8; c-basic-offset: 2; indent-tabs-mode: nil; -*- */
+
+#include "nsutils.h"
+#include "util.h"
+
+int main(int argc, char* argv[]) {
+  if ((argc > 1 && strcmp(argv[1], "in_copy") == 0) ||
+      -1 == try_setup_ns(CLONE_NEWNS)) {
+    atomic_puts("EXIT-SUCCESS");
+    return 0;
+  }
+
+  test_assert(0 == mkdir("mountpoint", 0700));
+  test_assert(0 == mount("none", "mountpoint", "tmpfs", 0, NULL));
+
+  int fd = open("mountpoint/the_copy", O_RDWR | O_CREAT | O_EXCL, 0600);
+  test_assert(fd >= 0);
+  test_assert(0 == close(fd));
+  test_assert(0 == mount(argv[0], "mountpoint/the_copy",
+                         "", MS_BIND, NULL));
+  int dst_fd = open("mountpoint/the_copy", O_RDONLY | O_PATH);
+  test_assert(dst_fd != -1);
+
+  test_assert(0 == umount2("mountpoint/the_copy", MNT_DETACH));
+
+  char* const new_argv[] = { "mountpoint/the_copy", "in_copy", NULL };
+  execveat(dst_fd, "", new_argv, environ, AT_EMPTY_PATH);
+  test_assert(0);
+  return 1;
+}