Re: [PATCH] VFS: br_write_lock locks on possible CPUs other than online CPUs

[Al Viro <viro@ZenIV.linux.org.uk>]

On Tue, Dec 20, 2011 at 08:05:32PM +0530, Srivatsa S. Bhat wrote:

> Sorry but I didn't quite get your point...
> No two cpu hotplug operations can race because of the cpu_hotplug lock they
> use. Hence, if a cpu online operation begins, it has to succeed or fail
> eventually. No other cpu hotplug operation can intervene. Ditto for cpu offline
> operations.
> 
> Hence a CPU_UP_PREPARE event *will* be followed by a corresponding
> CPU_UP_CANCELED or CPU_ONLINE event for the same cpu. (And we ignore the
> CPU_STARTING event that comes in between, on purpose, so as to avoid the race
> with cpu_online_mask). Similar is the story for offline operation.
> 
> And if the notifier grabs the spinlock and keeps it locked between these 2
> points of a cpu hotplug operation, it ensures that our br locks will spin,
> instead of block till the cpu hotplug operation is complete. Isn't this what
> we desired all along? "A non-blocking way to sync br locks with cpu hotplug"?
> 
> Or am I missing something?

The standard reason why losing the timeslice while holding a spinlock means
deadlocks?
CPU1: grabs spinlock
CPU[2..n]: tries to grab the same spinlock, spins
CPU1: does something blocking, process loses timeslice
CPU1: whatever got scheduled there happens to to try and grab the same
spinlock and you are stuck.  At that point *all* CPUs are spinning on
that spinlock and your code that would eventually unlock it has no chance
to get any CPU to run on.

Having the callback grab and release a spinlock is fine (as long as you
don't do anything blocking between these spin_lock/spin_unlock).  Having
it leave with spinlock held, though, means that the area where you can't
block has expanded a whole lot.  As I said, brittle...

A quick grep through the actual callbacks immediately shows e.g.
del_timer_sync() done on CPU_DOWN_PREPARE.  And sysfs_remove_group(),
which leads to outright mutex_lock().  And sysfs_remove_link() (ditto).
And via_cputemp_device_remove() (again, mutex_lock()).  And free_irq().
And perf_event_exit_cpu() (mutex_lock()).  And...

IOW, there are shitloads of deadlocks right there.  If your callback's
position in the chain is earlier than any of those, you are screwed.

No, what I had in mind was different - use the callbacks to maintain a
bitmap that would contain
	a) all CPUs that were online at the moment
	b) ... and not too much else
Updates protected by spinlock; in all cases it gets dropped before the
callback returns.  br_write_lock() grabs that spinlock and iterates over
the set; it *does* leave the spinlock grabbed - that's OK, since all
code between br_write_lock() and br_write_unlock() must be non-blocking
anyway.  br_write_unlock() iterates over the same bitmap (unchanged since
br_write_lock()) and finally drops the spinlock.

AFAICS, what we want in callback is
	CPU_DEAD, CPU_DEAD_FROZEN, CPU_UP_CANCELLED, CPU_UP_CANCELLED_FROZEN:
		grab spinlock
		remove cpu from bitmap
		drop spinlock
	CPU_UP_PREPARE, CPU_UP_PREPARE_FROZEN
		grab spinlock
		add cpu to bitmap
		drop spinlock
That ought to keep bitmap close to cpu_online_mask, which is enough for
our purposes.