Re: [PATCH tip/core/rcu 01/45] rcutorture: Add forward-progress checking for writer

[Josh Triplett]

On Fri, May 09, 2014 at 03:30:04PM -0700, Paul E. McKenney wrote:
> On Fri, May 09, 2014 at 12:32:46PM -0700, Josh Triplett wrote:
> > On Fri, May 09, 2014 at 10:36:58AM -0700, Paul E. McKenney wrote:
> > > On Fri, May 09, 2014 at 08:52:31AM -0700, Josh Triplett wrote:
> > > > On Wed, May 07, 2014 at 04:43:13PM -0700, Paul E. McKenney wrote:
> > > > > On Wed, May 07, 2014 at 02:16:49PM -0700, josh@xxxxxxxxxxxxxxxx wrote:
> > > > > > On Mon, Apr 28, 2014 at 05:24:49PM -0700, Paul E. McKenney wrote:
> > > > > > > From: "Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx>
> > > > > > > 
> > > > > > > The rcutorture output currently does not distinguish between stalls in
> > > > > > > the RCU implementation and stalls in the rcu_torture_writer() kthreads.
> > > > > > > This commit therefore adds some diagnostics to help distinguish between
> > > > > > > these two conditions, at least for the non-SRCU implementations.  (SRCU
> > > > > > > does not provide evidence of update-side forward progress by design.)
> > > > > > > 
> > > > > > > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
> > > > > > 
> > > > > > The concept makes sense, and the writer state annotations seem like a
> > > > > > useful debugging mechanism, but having RCU know about RCU torture types
> > > > > > seems fundamentally wrong.  This mechanism accesses rcu_state, which is
> > > > > > already implementation-specific, so why not just only define the
> > > > > > function for the RCU implementations that support it, and then have a
> > > > > > function pointer in the torture-test structure to report a stall?
> > > > > 
> > > > > Ouch.  It is worse than that!  When running RCU-bh or RCU-sched,
> > > > > the current code incorrectly returns the statistics for RCU.
> > > > > So I do need some way for rcutorture to tell RCU which flavor
> > > > > it is testing.
> > > > > 
> > > > > One thing I could do would be to pass in a pointer to the call_rcu()
> > > > > function (cur_ops->call from rcutorture's viewpoint), then scan the
> > > > > rcu_state structures looking for the selected flavor (rsp->call from
> > > > > tree.c's viewpoint).  In the SRCU and RCU-busted cases, the flavor would
> > > > > not be found, and I could then just set everything to zero.
> > > > > 
> > > > > Does that seem reasonable, or is there a better way to do this?
> > > > 
> > > > That search seems rather too hackish; why not just declare one
> > > > stats-returning function per RCU flavor, and put the pointer to the
> > > > corresponding function in the structure for each test type?
> > > 
> > > The problem is that rcutorture doesn't know anything about the structures,
> > > as those are internal to the implementation.  All it knows is which
> > > functions it is using.  I -could- EXPORT_SYMBOL_GPL() the rcu_state
> > > structures to modules (they are already non-static), then rename
> > > TINY_RCU's rcu_ctrlblk to rcu_state to allow the needed type punning,
> > > then do some special-case thing for SRCU, and put a pointer to whatever
> > > in rcu_torture_ops, but that was feeling a bit hackish as well.
> > > 
> > > So what did you have in mind to allow rcutorture to communicate the
> > > rcuflavor to the underlying RCU implementation?
> > 
> > Rather than EXPORT_SYMBOL_GPLing the rcu_state structures, just
> > EXPORT_SYMBOL_GPL one version of rcutorture_get_gp_data per RCU flavor.
> > (And hide them all behind #ifdef CONFIG_RCU_TORTURE_TEST.)  Then add a
> > get_gp_data field to rcu_torture_ops; if NULL, skip the stats.  (Or put
> > a no-op version in rcutorture.)
> 
> But that would require me to provide these same exports from TINY_RCU,
> which does not need them.
> 
> How about exporting integers identifying the flavors of RCU to rcutorture,
> which rcutorture can then pass to rcutorture_get_gp_data()?  This allows
> TINY_RCU to provide a trivial static inline function.  TREE_RCU and
> TREE_PREEMPT_RCU can keep an array of pointers to the corresponding
> rcu_state structure, with NULL pointers for flavors of RCU that don't
> have any data to provide.
> 
> Would that help?

Either way seems fine: a single function with an extra parameter or a
unique function per flavor.  But if you're going to provide the
RCU-flavor integers, they should be in RCU itself and refer to RCU
flavors, rather than being in rcutorture and refering to test types.

- Josh Triplett
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/