Re: 2.6.2-rc2 nfsd+xfs spins in i_size_read()

[Andrew Morton <akpm@osdl.org>]

Miquel van Smoorenburg <miquels@cistron.nl> wrote:
>
> On Fri, 30 Jan 2004 21:21:55, Miquel van Smoorenburg wrote:
> > I added some extra code to i_size_read() and i_size_write(). First,
> > some debugging code:
> > I then started the test that locks up the kernel, and it printed this:
> > 
> > i_size_write: pid 542: sequence is odd!
> > i_size_write: pid 543: sequence is odd!
> > i_size_write: pid 542: sequence is odd!
> > 
> > i_size_read() seems to be looping - pid 0
> > i_size_read() seems to be looping - pid 0
> > [this keeps on being printed and the kernel is locked up]
> > 
> > It took some time for the i_size_write messages to show up, and they were
> > spaced 10-30 seconds apart, and during that time the server was still
> > up - right until the first i_size_read message.
> > 
> 
> Okay, I added a patch to make the sequence increments atomic. Now
> i_size_write() still sometimes ends up with an odd sequence, but
> i_size_read() doesn't lock up anymore.

Go, Miquel!

> What lock exactly is supposed to protect i_size_write, since it
> appears that i_size_write is being called without proper locking ?
> (Am I right?)

If two CPUs hit i_size_write() at the same time we have a bug.  That
function requires that the caller provide external serialisation, via i_sem.

Try adding this to the start of i_size_write():

	if (down_trylock(&inode->i_sem) == 0) {
		printk("I am buggy\n");
		dump_stack();
		up(&inode->i_sem);
	}