Re: [PATCH -v3 3/4] lib, Make gen_pool memory allocator lockless

From: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
To: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Cc: Huang Ying <ying.huang@intel.com>, Len Brown <lenb@kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	Andi Kleen <andi@firstfloor.org>,
	"Luck, Tony" <tony.luck@intel.com>,
	"linux-acpi@vger.kernel.org" <linux-acpi@vger.kernel.org>,
	Andrew Morton <akpm@linux-foundation.org>
Subject: Re: [PATCH -v3 3/4] lib, Make gen_pool memory allocator lockless
Date: Fri, 15 Apr 2011 13:44:51 -0400	[thread overview]
Message-ID: <20110415174451.GB19644@Krystal> (raw)
In-Reply-To: <20110415174320.GA2283@linux.vnet.ibm.com>

* Paul E. McKenney (paulmck@linux.vnet.ibm.com) wrote:
> On Fri, Apr 15, 2011 at 12:46:29PM -0400, Mathieu Desnoyers wrote:
> > * Huang Ying (ying.huang@intel.com) wrote:
> > > On 04/14/2011 05:07 AM, Mathieu Desnoyers wrote:
> > > > * Huang Ying (ying.huang@intel.com) wrote:
> > > > [...]
> > > >> + * rcu_read_lock and rcu_read_unlock is not used int gen_pool_alloc,
> > > >> + * gen_pool_free, gen_pool_avail and gen_pool_size etc, because chunks
> > > >> + * are only added into pool, not deleted from pool unless the pool
> > > >> + * itself is destroyed.  If chunk will be deleted from pool,
> > > >> + * rcu_read_lock and rcu_read_unlock should be uses in these
> > > >> + * functions.
> > > > 
> > > > So how do you protect between pool destruction and adding chunks into
> > > > the pool ?
> > > 
> > > Because the pool itself will be freed when destruction, we need some
> > > mechanism outside of pool.  For example, if gen_pool_add() is called via
> > > device file IOCTL, we must un-register the device file first, and
> > > destroy the pool after the last reference to device has gone.
> > 
> > I am concerned about the list_for_each_entry_rcu() (and thus
> > rcu_dereference_raw()) used outside of rcu_read_lock/unlock pairs.
> > Validation infrastructure as recently been added to RCU: it triggers
> > warnings when these situations are encountered in some RCU debugging
> > configurations. The case of RCU list iteration is not covered by the
> > checks, but it would make sense to be aware of it.
> > 
> > So although it seems like your code does not require rcu read lock
> > critical sections, I'd prefer to let Paul McKenney have a look.
> 
> As long as you add elements and never remove them, then you can get
> away with using list_for_each_entry_rcu() outside of an RCU read-side
> critical section.  But please comment this -- it is all too easy
> for someone to decide later to start deleting elements without also
> inserting the needed rcu_read_lock() and rcu_read_unlock() pairs.
> 
> But I have lost the thread -- what code am I supposed to look at?

You can have a look at https://lkml.org/lkml/2011/4/13/56

Thanks!

Mathieu

> 
> 							Thanx, Paul

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com