On Mon, 4 May 2015 10:48:44 +1000
Dave Chinner [off-list ref] wrote:

quoted

I'm not sure why it does the ndelay() and not just a cpu_relax(), but

Because the code was writtenlong before cpu_relax() existed, just
like it was written long before the generic percpu counter code was
added...

Ah, legacy code.

....

quoted

Now, when PREEMPT_RT is not enabled, that spin_lock() disables
preemption. But for PREEMPT_RT, it does not. Although with my test box I
was not able to produce a task state of all tasks, but I'm assuming that
some task called the xfs_icsb_lock_all_counters() and was preempted by
an RT task and could not finish, causing all callers of that lock to
block indefinitely.

Looking at all users of xfs_icsb_lock_all_counters(), they are leaf
functions and do not call anything that may block on PREEMPT_RT. I
believe the proper fix here is to simply disable preemption in
xfs_icsb_lock_all_counters() when PREEMPT_RT is enabled.

RT is going to have other performance problems that are probably
going to negate the scalability this code provides. If you want a
hack that you can easily backport (as this code now uses the generic
percpu counters) then have a look at fs/xfs/xfs_linux.h:

/*
 * Feature macros (disable/enable)
 */
#ifdef CONFIG_SMP
#define HAVE_PERCPU_SB  /* per cpu superblock counters are a 2.6 feature */
#else
#undef  HAVE_PERCPU_SB  /* per cpu superblock counters are a 2.6 feature */
#endif

You can turn off all that per-cpu code simply by:

-#ifdef CONFIG_SMP
+#if defined(CONFIG_SMP) && !defined(CONFIG_PREEMPT_RT)

Oh, I think I like this the best.

Thanks for the advice.

-- Steve