Bug#77857: hurd: write_node assertion failed building emacs

[Marcus Brinkmann]

On Thu, Nov 23, 2000 at 08:26:42PM -0500, Roland McGrath wrote:
> It looks to me like the ext2fs block allocation code (getblk.c) can set
> dn_set_mtime when called from the pager whiled not holding the node lock.
> It will hold node->dn->alloc_lock, but that is all.

I think what you say is correct, especially as my other info (which I sent
before reading this mail) seems to be 100% in agreement with this analysis.
 
> The ext2_getblk code is called sometimes with the node lock held
> (storeinfo.c, diskfs_grow) and sometimes without (pager).  So it cannot
> just take the lock around setting dn_set_?time.  I'm not sure whether the
> node lock really ought to be held around setting them or not.

Looking at write_one_disknode in inode.c, I'm not even sure if holding the
lock would be enough. The stack trace shows that the lock was held during
the assertion failure. I don't see how you can sync the node without locking
it, and expect that nobody changes the dn_set_?times in the meantime. Maybe
the assertion is bogus? Even between calling diskfs_set_node_times and the
assertion is a small time window (which seems to be too small to show up in
any normal situation) where someone can stomp on the time flags while
holding the lock or not.

> From reading
> the ufs code, I cna't tell whether it has the same bug, but I think it does.

Yes, but it has a smaller window between diskfs_set_node_times and
write_node.

write_all_disknodes ()
{
  error_t
    helper (struct node *np)
      {
        diskfs_set_node_times (np);
        write_node (np);
        return 0;
      }

  diskfs_node_iterate (helper);
}

My preliminary and by no means authoritative impression is that the
assertion cannot be guaranteed without holding locks on either side (sync
thread and everyone else).

Thanks,
Marcus

-- 
`Rhubarb is no Egyptian god.' Debian http://www.debian.org brinkmd@debian.org
Marcus Brinkmann              GNU    http://www.gnu.org    marcus@gnu.org
Marcus.Brinkmann@ruhr-uni-bochum.de
http://www.marcus-brinkmann.de