Re: [PATCH v8 13/13] nvmet: Optionally use PCI P2P memory

[Logan Gunthorpe]

On 2018-10-01 3:34 p.m., Sagi Grimberg wrote:
>> +
>> +	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
>> +		pci_p2pdma_remove_client(&ctrl->p2p_clients, nvmet_ns_dev(ns));
>> +		nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE, 0, 0);
> 
> Hi Logan, what is this event here?

Oops, that must have been from a bad rebase.... Will Fix.

>> +		if (sq->qid && p2p_dev) {
>> +			req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
>> +						       req->transfer_len);
>> +			if (req->sg) {
>> +				req->p2p_dev = p2p_dev;
>> +				return 0;
>> +			}
> 
> Would be useful to comment that we fall to normal sgl allocation.

Ok.

>> +/*
>> + * If allow_p2pmem is set, we will try to use P2P memory for the SGL lists for
>> + * Î/O commands. This requires the PCI p2p device to be compatible with the
>> + * backing device for every namespace on this controller.
>> + */
>> +static void nvmet_setup_p2pmem(struct nvmet_ctrl *ctrl, struct nvmet_req *req)
>> +{
>> +	struct nvmet_ns *ns;
>> +	int ret;
>> +
>> +	if (!req->port->use_p2pmem || !req->p2p_client)
>> +		return;
> 
> Nit, IMO would be better to check at the call-site, but not a hard
> must...

I'd rather keep the logic for whether to enable p2pmem in it's own
function. nvme_alloc_ctrl() is already very long and complicated.

> I still do not fully understand why p2p_dev has to be ctrl-wide and not
> per namespace. Sorry to keep bringing this up (again). But if people are
> OK with it then I guess I can stop asking about this...

Because you never answered my question back in March[1] (which I think
you've answered below)....

> I think that at some point we said that this looks like it should fall
> back to host memory for those namespaces.. when we allocate the sgl we
> already assigned a namespace to the request (nvmet_req_init).

I did not realize the namespace would be available at this time. I guess
I can give this a try, but it's going to be a fairly big change from
what's presented here... Though, I agree it'll probably be an improvement.

Logan

[1]
https://lore.kernel.org/lkml/7163af93-2f37-a8b6-986a-3cb2e62bee29@xxxxxxxxxxxx/T/#u