jiangyiwen wrote on Sat, Jul 14, 2018: > On 2018/7/14 17:05, Dominique Martinet wrote: > > jiangyiwen wrote on Sat, Jul 14, 2018: > >> When client has multiple threads that issue io requests all the > >> time, and the server has a very good performance, it may cause > >> cpu is running in the irq context for a long time because it can > >> check virtqueue has buf in the *while* loop. > >> > >> So we should keep chan->lock in the whole loop. > > > > Hmm, this is generally bad practice to hold a spin lock for long. > > In general, spin locks are meant to protect data, not code. > > > > I'd want some numbers to decide on this one, even if I think this > > particular case is safe (e.g. this cannot dead-lock) > > > > Actually, the loop will not hold a spin lock for long, because other > threads will not issue new requests in this case. In addition, > virtio-blk or virtio-scsi also use this solution, I guess it may also > encounter this problem before.
Fair enough. If you do have some numbers to give though (throughput and/or iops before/after) I'd still be really curious. > >> chan->ring_bufs_avail = 1; > >> - spin_unlock_irqrestore(&chan->lock, flags); > >> /* Wakeup if anyone waiting for VirtIO ring space. */ > >> wake_up(chan->vc_wq); > > > > In particular, the wake up here echoes to wait events that will > > immediately try to grab the lock, and will needlessly spin on it until > > this thread is done. > > If we do go this way I'd want setting chan->ring_bufs_avail to be done > > just before unlocking and the wakeup to be done just after unlocking out > > of the loop iff we processed at least one iteration here. > > I can move the wakeup operation after the unlocking. Like what I said > above, I think this loop will not execute for long. Please do, you listed virtio_blk as doing this and they have the same kind of pattern with a req_done bool and only restarting stopped queues if they processed something -- Dominique