file-syncfs: Avoid deadlock

This is a four-player problem (here for diskfs):

- One runs sync, which calls diskfs_S_file_syncfs, which triggers
  fsys_syncfs on all active translators, while keeping the
  translator_ihash_lock lock.

- One of the active translators is hung, for some reason

- Another ext2fs thread is trying to call fshelp_set_active_translator from
  dir_lookup. It is stuck on trying to acquire translator_ihash_lock, and it
  holds the np.

- The ext2fs thread running diskfs_sync_everything tries to lock that
  np, while holding the nodecache_lock.

In the end everything is locked. While diskfs_S_file_syncfs can as well just
atomically get the list of active translators, and then call fsys_syncfs
without keeping translator_ihash_lock held.

1 files changed, 19 insertions, 17 deletions

diff --git a/libdiskfs/file-syncfs.c b/libdiskfs/file-syncfs.c
index 8cd4003b..d271ab1e 100644
--- a/libdiskfs/file-syncfs.c
+++ b/libdiskfs/file-syncfs.c
@@ -19,33 +19,35 @@
 #include "fs_S.h"
 #include <hurd/fsys.h>
 
-struct args
-{
-  int wait;
-};
-
-static error_t
-helper (void *cookie, const char *name, mach_port_t control)
-{
-  struct args *args = cookie;
-  (void) name;
-  fsys_syncfs (control, args->wait, 1);
-  return 0;
-}
-
 /* Implement file_syncfs as described in <hurd/fs.defs>. */
 kern_return_t
 diskfs_S_file_syncfs (struct protid *cred,
 		      int wait,
 		      int dochildren)
 {
-  struct args args = { wait };
-
   if (!cred)
     return EOPNOTSUPP;
 
   if (dochildren)
-    fshelp_map_active_translators (helper, &args);
+    {
+      error_t err;
+      char *n = NULL;
+      size_t n_len = 0;
+      mach_port_t *c;
+      size_t c_count, i;
+
+      err = fshelp_get_active_translators (&n, &n_len, &c, &c_count);
+      if (err)
+	return err;
+      free(n);
+
+      for (i = 0; i < c_count; i++)
+	fsys_syncfs (c[i], wait, 1);
+
+      free(c);
+      if (err)
+	return err;
+    }
 
   if (diskfs_synchronous)
     wait = 1;