Re: [PATCH 11/14] PCI/P2PDMA: dma_map P2PDMA map requests that traverse the host bridge

From: Logan Gunthorpe <logang@deltatee.com>
To: Jason Gunthorpe <jgg@mellanox.com>
Cc: Christoph Hellwig <hch@lst.de>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"linux-pci@vger.kernel.org" <linux-pci@vger.kernel.org>,
	"linux-nvme@lists.infradead.org" <linux-nvme@lists.infradead.org>,
	"linux-rdma@vger.kernel.org" <linux-rdma@vger.kernel.org>,
	Bjorn Helgaas <bhelgaas@google.com>,
	Christian Koenig <Christian.Koenig@amd.com>,
	Sagi Grimberg <sagi@grimberg.me>, Keith Busch <kbusch@kernel.org>,
	Jens Axboe <axboe@fb.com>,
	Dan Williams <dan.j.williams@intel.com>,
	Eric Pilmore <epilmore@gigaio.com>,
	Stephen Bates <sbates@raithlin.com>
Subject: Re: [PATCH 11/14] PCI/P2PDMA: dma_map P2PDMA map requests that traverse the host bridge
Date: Thu, 25 Jul 2019 11:22:39 -0600	[thread overview]
Message-ID: <486e2a25-f6e6-75d4-a8b7-6a38fff8546a@deltatee.com> (raw)
In-Reply-To: <20190725163438.GF7450@mellanox.com>

On 2019-07-25 10:34 a.m., Jason Gunthorpe wrote:
> On Thu, Jul 25, 2019 at 10:00:25AM -0600, Logan Gunthorpe wrote:
>>
>>
>> On 2019-07-25 12:10 a.m., Christoph Hellwig wrote:
>>> On Wed, Jul 24, 2019 at 09:58:59AM -0600, Logan Gunthorpe wrote:
>>>>
>>>>
>>>> On 2019-07-24 12:32 a.m., Christoph Hellwig wrote:
>>>>>>  	struct dev_pagemap *pgmap = sg_page(sg)->pgmap;
>>>>>> +	struct pci_dev *client;
>>>>>> +	int dist;
>>>>>> +
>>>>>> +	client = find_parent_pci_dev(dev);
>>>>>> +	if (WARN_ON_ONCE(!client))
>>>>>> +		return 0;
>>>>>>  
>>>>>> +	dist = upstream_bridge_distance(pgmap->pci_p2pdma_provider,
>>>>>> +					client, NULL);
>>>>>
>>>>> Doing this on every mapping call sounds expensive..
>>>>
>>>> The result of this function is cached in an xarray (per patch 4) so, on
>>>> the hot path, it should just be a single xa_load() which should be a
>>>> relatively fast lookup which is similarly used for other hot path
>>>> operations.
>>>
>>> We don't cache find_parent_pci_dev, though.  So we should probably
>>> export find_parent_pci_dev with a proper namespaces name and cache
>>> that in the caler.
>>
>> Oh, yes, I'll take a look at this. Of the two callers: NVMe should be
>> easy we could just pass the PCI device instead of the struct device.
>> RDMA is significantly more unclear: would you add a pci_dev to struct
>> ib_device? Or maybe we should be able to simply rely on the fact that
>> the DMA device *must* be a PCI device and just use to_pci_dev() directly?
> 
> AFAIK you need to use the ib_device->dma_device and add some kind of
> is_pci_dev to make it safe

Yes, that's my thinking. The dma_device *should* be a PCI device. We can
just be sure by doing is_pci_dev() and failing the mapping if it is not.

So I *think* we should be able to simply replace the
find_parent_pci_dev() with:

if (!dev_is_pci(dev))
     return 0;

client = to_pci_dev(dev);

Which should be fast and reliable.

The alternative is to push this out into the caller which may have a bit
more information (like the nvme driver does).

Logan