Hello Dan,
> > > + /* The request submission function */
> > > +static int virtio_pmem_flush(struct nd_region *nd_region)
> > > +{
> > > + int err;
[...]
> > > + init_waitqueue_head(&req->host_acked);
> > > + init_waitqueue_head(&req->wq_buf);
> > > +
> > > + spin_lock_irqsave(&vpmem->pmem_lock, flags);
> > > + sg_init_one(&sg, req->name, strlen(req->name));
> > > + sgs[0] = &sg;
> > > + sg_init_one(&ret, &req->ret, sizeof(req->ret));
> > > + sgs[1] = &ret;
[...]
> > > + spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> > > + /* When host has read buffer, this completes via host_ack */
> > > + wait_event(req->host_acked, req->done);
> >
> > Hmm, this seems awkward if this is called from pmem_make_request. If
> > we need to wait for completion that should be managed by the guest
> > block layer. I.e. make_request should just queue request and then
> > trigger bio_endio() when the response comes back.
>
> We are plugging VIRTIO based flush callback for virtio_pmem driver. If pmem
> driver (pmem_make_request) has to queue request we have to plug "blk_mq_ops"
> callbacks for corresponding VIRTIO vqs. AFAICU there is no existing
> multiqueue
> code merged for pmem driver yet, though i could see patches by Dave upstream.
>
I thought about this and with current infrastructure "make_request" releases spinlock
and makes current thread/task. All Other threads are free to call 'make_request'/flush
and similarly wait by releasing the lock. This actually works like a queue of threads
waiting for notifications from host.
Current pmem code do not have multiqueue support and I am not sure if core pmem code
needs it. Adding multiqueue support just for virtio-pmem and not for pmem in same driver
will be confusing or require alot of tweaking.
Could you please give your suggestions on this.
Thanks,
Pankaj
