Messages in this thread | | | Date | Mon, 23 Sep 2013 14:58:51 +0200 | From | Sebastian Andrzej Siewior <> | Subject | Re: IRQ affinity notifiers vs RT |
| |
On 08/30/2013 11:29 PM, Ben Hutchings wrote: > Sebastian, I saw you came up with a fix for this but apparently without > seeing my earlier message:
Yes Ben, I haven't seen it. If I was on Cc I very sorry for overlooking it.
> On Thu, 2013-07-25 at 00:31 +0100, Ben Hutchings wrote:
>> Workqueue code uses spin_lock_irq() on the workqueue lock, which with >> PREEMPT_RT enabled doesn't actually block IRQs. >> >> In 3.6, the irq_cpu_rmap functions relies on a workqueue flush to finish >> any outstanding notifications before freeing the cpu_rmap that they use. >> This won't be reliable if the notification is scheduled after releasing >> the irq_desc lock. >> >> However, following commit 896f97ea95c1 ('lib: cpu_rmap: avoid flushing >> all workqueues') in 3.8, I think that it is sufficient to do only >> kref_get(&desc->affinity_notify->kref) in __irq_set_affinity_locked() >> and then call schedule_work() in irq_set_affinity() after releasing the >> lock. Something like this (untested): > > Does the following make sense to you?
This was suggested by the original submitter on rt-users@v.k.o (Joe Korty) where I've been made aware of this for the first time. This okay except for the part where the workqueue is not scheduled if calling by the __ function (i.e. the mips case). If I read the code correctly, the CPU goes offline and affinity change should be updated / users notified and this is not the case with this patch.
It is a valid question why only one mips SoC needs the function. If you look at commit 0c3263870f ("MIPS: Octeon: Rewrite interrupt handling code.") you can see that tglx himself made this adjustment so it might be valid :) Therefore I assume that we may get more callers of this function and the workqueue should be executed and I made something simple that works on RT.
> > Ben. >
Sebastian
| |