lkml.org 
[lkml]   [2011]   [Jun]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [PATCH 5/5] CFQ: use proper locking for cache of last hit cic
    From
    Date
    On Wed, 2011-06-08 at 14:42 -0400, Vivek Goyal wrote:
    > On Wed, Jun 08, 2011 at 08:18:44PM +0200, Paul Bolle wrote:
    > > On Mon, 2011-06-06 at 05:06 +0200, Jens Axboe wrote:
    > > > On 2011-06-05 18:26, Paul Bolle wrote:
    > > > > @@ -2704,8 +2706,13 @@ static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
    > > > > smp_wmb();
    > > > > cic->key = cfqd_dead_key(cfqd);
    > > > >
    > > > > - if (ioc->last_cic == cic)
    > > > > + spin_lock_irqsave(&ioc->lock, flags);
    > > > > + rcu_read_lock();
    > > > > + last_cic = rcu_dereference(ioc->last_cic);
    > > > > + rcu_read_unlock();
    > > > > + if (last_cic == cic)
    > > > > rcu_assign_pointer(ioc->last_cic, NULL);
    > > > > + spin_unlock_irqrestore(&ioc->lock, flags);
    > > >
    > > > We don't need the ioc->lock for checking the cache, it would in fact
    > > > defeat the purpose of using RCU.
    > >
    > > Just to show that I'm RCU-challenged, is that because:
    > > 1) my use of locking on ioc->lock defends for a race that is not
    > > actually possible; or
    > > 2) the worst thing that could happen is that some new and correct value
    > > of ioc->last_cic will be replaced with NULL, which is simply not a big
    > > deal?
    >
    > I don't understand this point.

    That could be because I don't really get all the RCU voodoo, nor how
    this all interacts with the io_contect->lock here, so I probably asked
    an impossible question.

    > All ioc->ioc_data updates are under
    > ioc->lock except the one __cfq_exit_single_io_context() and that's what
    > jens patch fixed. So clearly there was atleast one race where we were
    > doing a value update without taking appropriate lock.
    >
    > Why do you think that some new and correct value will be replaced
    > by NULL?

    Jens updated the code to
    if (rcu_dereference(ioc->ioc_data) == cic) {
    spin_lock(&ioc->lock);
    rcu_assign_pointer(ioc->ioc_data, NULL);
    spin_unlock(&ioc->lock);
    }

    I basically suggested (except for an apparently useless
    rcu_read_lock() / rcu_read_unlock() pair and using spin_lock_irqsave()
    etc.):
    spin_lock(&ioc->lock);
    if (rcu_dereference(ioc->ioc_data) == cic)
    rcu_assign_pointer(ioc->ioc_data, NULL);
    spin_unlock(&ioc->lock);

    Ie, I thought that reads of and updates to ioc->ioc_data should be done
    with ioc->lock held. My reasoning was that in Jens code ioc->ioc_data
    might already be updated (by another thread, or whatever) and thus not
    be equal to cic by the time it's updated to NULL. See, in my
    understanding ioc->ioc_data could be equal to cic when it's
    rcu_dereference()'d, but unequal to cic by the time it's
    rcu_assign_pointer()'d to NULL.


    Paul Bolle



    \
     
     \ /
      Last update: 2011-06-08 21:35    [W:0.025 / U:29.700 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site