Re: [PATCH RFC 09/12] xen/pvticketlock: Xen implementation for PV ticket locks

[Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>]

On Sun, Sep 26, 2010 at 03:34:55PM -0700, Jeremy Fitzhardinge wrote:
>  On 09/26/2010 04:39 AM, Srivatsa Vaddagiri wrote:
> > On Fri, Jul 16, 2010 at 06:03:04PM -0700, Jeremy Fitzhardinge wrote:
> >> Replace the old Xen implementation of PV spinlocks with and implementation
> >> of xen_lock_spinning and xen_unlock_kick.
> > I see that the old implementation took care of a spinlock() call being
> > interrupted by another spinlock (in interrupt handler), by saving/restoring 
> > old lock of interest. We don't seem to be doing that in this new version?
> > Won't that lead to loss of wakeup -> hang?

Sorry about coming back late on this, but as I was looking at the most recent
version of pv-ticketlocks, this came up in my mind again ..

> No, interrupts are disabled while waiting to take the lock, so it isn't
> possible for an interrupt to come in.

Where are we disabling interrupts? Is it in xen_poll_irq()?

>  With the old-style locks it was
> reasonable to leave interrupts enabled while spinning, but with ticket
> locks it isn't.
> 
> (I haven some prototype patches to implement nested spinning of ticket
> locks,

Hmm ..where is nested spinning allowed/possible? Process context will
disable interrupts/bh from wanting the same (spin-)lock it is trying to
acquire?

> by allowing the nested taker to steal the queue position of the
> outer lock-taker, and switch its ticket with a later one.  But there's a
> fundamental problem with the idea: each lock taker needs to take a
> ticket.  If you don't allow nesting, then the max amount of tickets
> needed = number of cpus-1; however, with nesting, the max number of
> tickets = ncpus * max-nesting-depth, so the size of the ticket type must
> be larger for a given number of cpus, or the max number of cpus must be
> reduced.)

- vatsa