Messages in this thread | | | Date | Thu, 15 Mar 2012 13:28:03 +0100 (CET) | From | Thomas Gleixner <> | Subject | Re: [patch 0/5] seqlock consolidation |
| |
On Thu, 15 Mar 2012, Al Viro wrote:
> On Thu, Mar 15, 2012 at 11:44:22AM -0000, Thomas Gleixner wrote: > > > Aside of that replacing open coded constructs with proper functions is > > a worthwhile cleanup by itself. > > Provided that those are proper primitives to start with... I don't like > it - most of ->d_lock uses are _not_ related to ->d_seq, to start with > and then we get an interesting mix of functions that do and do not assume > the lock already taken, etc. > > Could you describe RT patch problems in more details? I really don't like > this solution - interface is overcomplicated and doesn't fit well...
On RT the spinlock sections (except the raw_spinlock) ones are becoming preemtible as we replace the spinlocks by "sleeping spinlock", a PI aware rtmutex.
So now assume the following:
spin_lock(d->d_lock); .... d->d_seq++;
-> Preemption read_seqcount_begin()
repeat: ret = ACCESS_ONCE(sl->sequence); if (unlikely(ret & 1)) { cpu_relax(); goto repeat; } smp_rmb();
Now when the preempting task has higher prioriy, the thing loops forever. Fail ....
The only way out is to lock d->d_lock which is contended, so the reader side boosts the writer and waits for the lock to be released. Though with the open coded seqlock we have no idea which lock we need to take.
Any better ideas ?
Thanks,
tglx
| |