Messages in this thread | | | Date | Fri, 20 Feb 2009 10:44:59 -0800 | From | "Paul E. McKenney" <> | Subject | Re: [PATCH URCU formal] Add liveness checks to user-level RCU model. |
| |
By the way, this model does -not- cover Mathieu's scenario where a very large number of reading processes continually arrive using an obsolete value of the urcu_gp_ctr variable. Something about having finite memory on all the machines I have access to. ;-)
So, here is an informal outline of why forward progress is guaranteed in this situation:
o We assume that if one CPU makes a change to a shared variable, all other CPUs see the change in a finite period of time. [CPUs like the semi-SMP Blackfins will need to make this be the case in software. If you cannot make this assumption, you cannot prove that -any- non-trivial SMP algorithm will complete.]
o We assume that any given CPU executes code at a finite rate. Of course, I would be very interested in hearing of any real hardware that violates this assumption. ;-)
o We assume that all RCU read-side critical sections complete in finite time (which can be violated if there are CPU-bound real-time-priority processes unless you have priority boosting, a scenario I am excluding -- but if this is a real problem, we need to add priority boosting).
o Then there must be a finite number of threads that have observed the old value of urcu_gp_ctr.
o Any subsequent RCU read-side critical sections will use the new value of urcu_gp_ctr.
o The total time consumed by these threads' RCU read-side critical sections will then be finite.
o The updater will therefore eventually observe all the threads that used the old value of urcu_gp_ctr to have exited all of their RCU read-side critical sections.
So given the three assumptions above, forward progress is guaranteed. Note that although we don't explicitly assume a finite number of threads, the finite-CPU-speed assumption makes it difficult to see how the updater scans the per-thread variables in finite time. ;-)
Thanx, Paul
On Fri, Feb 20, 2009 at 09:39:27AM -0800, Paul E. McKenney wrote: > Break all potentially infinite loops in both urcu_reader() and > urcu_updater(), ensure that urcu_reader() will process any memory barriers > that urcu_updater() might issue, and formulate a "never" claim that checks > to make sure that if either urcu_reader() or urcu_updater() completes, > then the other will eventually also complete. Since urcu_reader() > now has a finite number of steps, it must eventually complete. > > Also replace the code at the end of urcu_reader() that previously absorbed > late memory-barrier requests from urcu_updater with code in urcu_writer() > that checks to see if urcu_reader() has completed. > > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> > --- > formal-model/urcu.sh | 26 ++++++++++- > formal-model/urcu.spin | 113 ++++++++++++++++++++++------------------------- > 2 files changed, 77 insertions(+), 62 deletions(-) > > diff --git a/formal-model/urcu.sh b/formal-model/urcu.sh > index b76c764..5e525ec 100644 > --- a/formal-model/urcu.sh > +++ b/formal-model/urcu.sh > @@ -1,3 +1,25 @@ > +#!/bin/sh > +# > +# Compiles and runs the urcu.spin Promela model. > +# > +# This program is free software; you can redistribute it and/or modify > +# it under the terms of the GNU General Public License as published by > +# the Free Software Foundation; either version 2 of the License, or > +# (at your option) any later version. > +# > +# This program is distributed in the hope that it will be useful, > +# but WITHOUT ANY WARRANTY; without even the implied warranty of > +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > +# GNU General Public License for more details. > +# > +# You should have received a copy of the GNU General Public License > +# along with this program; if not, write to the Free Software > +# Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. > +# > +# Copyright (C) IBM Corporation, 2009 > +# > +# Authors: Paul E. McKenney <paulmck@linux.vnet.ibm.com> > + > spin -a urcu.spin > -cc -DSAFETY -o pan pan.c > -./pan > +cc -o pan pan.c > +./pan -a > diff --git a/formal-model/urcu.spin b/formal-model/urcu.spin > index 3e18457..cf1f670 100644 > --- a/formal-model/urcu.spin > +++ b/formal-model/urcu.spin > @@ -40,21 +40,28 @@ bit updater_done = 0; > byte urcu_gp_ctr = 1; > byte urcu_active_readers = 0; > > +/* must be at least the number of smp_mb() statements in urcu_updater(). */ > +#define N_SMP_MB 9 > + > /* Model the RCU read-side critical section. */ > > proctype urcu_reader() > { > bit done = 0; > + byte i; > bit mbok; > byte tmp; > byte tmp_removed; > byte tmp_free; > > /* Absorb any early requests for memory barriers. */ > + i = 0; > do > - :: need_mb == 1 -> > + :: i < N_SMP_MB && need_mb == 1 -> > need_mb = 0; > - :: !updater_done -> skip; > + i++; > + :: i < N_SMP_MB -> > + i++; > :: 1 -> break; > od; > > @@ -91,12 +98,17 @@ proctype urcu_reader() > if > :: (tmp & RCU_GP_CTR_NEST_MASK) == 0 -> > tmp = urcu_gp_ctr; > + i = 0; > do > - :: (reader_progress[1] + > + :: i < N_SMP_MB && > + (reader_progress[1] + > reader_progress[2] + > - reader_progress[3] == 0) && need_mb == 1 -> > + reader_progress[3] == 0) && > + need_mb == 1 -> > need_mb = 0; > - :: !updater_done -> skip; > + i++; > + :: i < N_SMP_MB -> > + i++; > :: 1 -> break; > od; > urcu_active_readers = tmp; > @@ -151,10 +163,13 @@ proctype urcu_reader() > if > :: mbok == 1 -> > /* We get here if mb processing is safe. */ > + i = 0; > do > - :: need_mb == 1 -> > + :: i < N_SMP_MB && need_mb == 1 -> > need_mb = 0; > - :: !updater_done -> skip; > + i++; > + :: i < N_SMP_MB -> > + i++; > :: 1 -> break; > od; > :: else -> skip; > @@ -169,36 +184,30 @@ proctype urcu_reader() > :: else -> skip; > fi > od; > + > + /* Make sure any concurrent grace period was sufficiently long. */ > assert((tmp_free == 0) || (tmp_removed == 1)); > > /* Reader has completed. */ > reader_done = 1; > - > - /* Process any late-arriving memory-barrier requests. */ > - do > - :: need_mb == 1 -> > - need_mb = 0; > - :: !updater_done -> skip; > - :: 1 -> break; > - od; > } > > +#define smp_mb() \ > + need_mb = 1; \ > + do \ > + :: !reader_done && need_mb == 1 -> skip; \ > + :: reader_done -> break; \ > + :: need_mb == 0 -> break; \ > + od > + > /* Model the RCU update process. */ > > proctype urcu_updater() > { > /* prior synchronize_rcu(), second counter flip. */ > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > urcu_gp_ctr = urcu_gp_ctr + RCU_GP_CTR_BIT; > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > do > :: 1 -> > if > @@ -209,27 +218,15 @@ proctype urcu_updater() > :: else -> break; > fi; > od; > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > > /* Removal statement, e.g., list_del_rcu(). */ > removed = 1; > > /* current synchronize_rcu(), first counter flip. */ > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > urcu_gp_ctr = urcu_gp_ctr + RCU_GP_CTR_BIT; > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > do > :: 1 -> > if > @@ -240,24 +237,12 @@ proctype urcu_updater() > :: else -> break; > fi; > od; > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > > /* current synchronize_rcu(), second counter flip. */ > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > urcu_gp_ctr = urcu_gp_ctr + RCU_GP_CTR_BIT; > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > do > :: 1 -> > if > @@ -268,11 +253,7 @@ proctype urcu_updater() > :: else -> break; > fi; > od; > - need_mb = 1; > - do > - :: need_mb == 1 -> skip; > - :: need_mb == 0 -> break; > - od; > + smp_mb(); > > /* free-up step, e.g., kfree(). */ > free = 1; > @@ -299,3 +280,15 @@ init { > run urcu_updater(); > } > } > + > +/* Require that both reader and updater eventually get done. */ > + > +never { > + do > + :: skip; > + :: reader_done != 0 || updater_done != 0 -> break; > + od; > +accept: do > + :: reader_done == 0 || updater_done == 0; > + od; > +} > -- > 1.5.2.5 >
| |