Messages in this thread | | | Date | Wed, 29 Apr 2015 21:44:28 +0200 | From | Manfred Spraul <> | Subject | Re: [PATCH v2] ipc/mqueue: remove STATE_PENDING |
| |
Hi Davidlohr,
On 04/28/2015 06:59 PM, Davidlohr Bueso wrote: > On Tue, 2015-04-28 at 18:43 +0200, Peter Zijlstra wrote: >> Well, if you can 'guarantee' the cmpxchg will not fail, you can then >> rely on the fact that cmpxchg implies a full barrier, which would >> obviate the need for the wmb. > Yes, assuming it implies barriers on both sides. And we could obviously > remove the need for pairing. With wake_q being local to wq_sleep() I > cannot see duplicate tasks trying to add themselves in the list. Failed > cmpxchg should only occur when users start misusing the wake_q. > > Manfred, do you have any objections to this? Perhaps I've missed the > real purpose of the barriers. I don't remember the details either, so let's check what should happen:
CPU1: sender copies message to kernel memory aaaa CPU1: sender does receiver->msg = message; ** barrier 1 CPU1: sender does receiver->state = STATE_READY;
CPU2: receiver notices receiver->state = STATE_READY; ** barrier 2 CPU2: receiver reads receiver->msg bbbb CPU2: receiver reads *receiver->msg
Failures would be: - write to receiver->state is visible before the write to receiver->msg or to *receiver->msg ** barrier 1 needs to be an smp_wmb() - cpu 2 reads receiver->msg before receiver->state ** barrier 2 needs to be an smp_rmb().
As far as I can see, no barrier is needed in pos aaaa or bbbb.
With regards to failed cmpxchg(): I don't see that mqueue could cause it by itself.
Who is allowed to use wake_q? If it is permitted to use wake_q for e.g. timeout/signal delivery wakeup, then that user might have a pending wakeup stored in the task struct.
-- Manfred
| |