Re: [dpdk-dev] [Question] How pmd virtio works without UIO?

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2015-12-23 05:13, Xie, Huawei:
> On 12/23/2015 10:57 AM, Yuanhan Liu wrote:
> > On Wed, Dec 23, 2015 at 10:41:57AM +0800, Peter Xu wrote:
> >> On Wed, Dec 23, 2015 at 10:01:35AM +0800, Yuanhan Liu wrote:
> >>> On Tue, Dec 22, 2015 at 05:56:41PM +0800, Peter Xu wrote:
> >>>> On Tue, Dec 22, 2015 at 04:32:46PM +0800, Yuanhan Liu wrote:
> >>>>> Actually, you are right. I mentioned in the last email that this is
> >>>>> for configuration part. To answer your question in this email, you
> >>>>> will not be able to go that further (say initiating virtio pmd) if
> >>>>> you don't unbind the origin virtio-net driver, and bind it to igb_uio
> >>>>> (or something similar).
> >>>>>
> >>>>> The start point is from rte_eal_pci_scan, where the sub-function
> >>>>> pci_san_one just initates a DPDK bond driver.
> >>>> I am not sure whether I do understand your meaning correctly
> >>>> (regarding "you willl not be able to go that furture"): The problem
> >>>> is that, we _can_ run testpmd without unbinding the ports and bind
> >>>> to UIO or something. What we need to do is boot the guest, reserve
> >>>> huge pages, and run testpmd (keeping its kernel driver as
> >>>> "virtio-pci"). In pci_scan_one():
> >>>>
> >>>> 	if (!ret) {
> >>>> 		if (!strcmp(driver, "vfio-pci"))
> >>>> 			dev->kdrv = RTE_KDRV_VFIO;
> >>>> 		else if (!strcmp(driver, "igb_uio"))
> >>>> 			dev->kdrv = RTE_KDRV_IGB_UIO;
> >>>> 		else if (!strcmp(driver, "uio_pci_generic"))
> >>>> 			dev->kdrv = RTE_KDRV_UIO_GENERIC;
> >>>> 		else
> >>>> 			dev->kdrv = RTE_KDRV_UNKNOWN;
> >>>> 	} else
> >>>> 		dev->kdrv = RTE_KDRV_UNKNOWN;
> >>>>
> >>>> I think it should be going to RTE_KDRV_UNKNOWN
> >>>> (driver=="virtio-pci") here.
> >>> Sorry, I simply overlook that. I was thinking it will quit here for
> >>> the RTE_KDRV_UNKNOWN case.
> >>>
> >>>> I tried to run IO and it could work,
> >>>> but I am not sure whether it is safe, and how.
> >>> I also did a quick test then, however, with the virtio 1.0 patchset
> >>> I sent before, which sets the RTE_PCI_DRV_NEED_MAPPING, resulting to
> >>> pci_map_device() failure and virtio pmd is not initiated at all.
> >> Then, will the patch work with ioport way to access virtio devices?
> > Yes.
> >
> >>>> Also, I am not sure whether I need to (at least) unbind the
> >>>> virtio-pci driver, so that there should have no kernel driver
> >>>> running for the virtio device before DPDK using it.
> >>> Why not? That's what the DPDK document asked to do
> >>> (http://dpdk.org/doc/guides/linux_gsg/build_dpdk.html):
> >>>
> >>>     3.6. Binding and Unbinding Network Ports to/from the Kernel Modules
> >>>     
> >>>     As of release 1.4, DPDK applications no longer automatically unbind
> >>>     all supported network ports from the kernel driver in use. Instead,
> >>>     all ports that are to be used by an DPDK application must be bound
> >>>     to the uio_pci_generic, igb_uio or vfio-pci module before the
> >>>     application is run. Any network ports under Linux* control will be
> >>>     ignored by the DPDK poll-mode drivers and cannot be used by the
> >>>     application.
> >> This seems obsolete? since it's not covering ioport.
> > I don't think so. Above is for how to run DPDK applications. ioport
> > is just a (optional) way to access PCI resource in a specific PMD.
> >
> > And, above speicification avoids your concerns, that two drivers try
> > to manipulate same device concurrently, doesn't it?
> >
> > And, it is saying "any network ports under Linux* control will be
> > ignored by the DPDK poll-mode drivers and cannot be used by the
> > application", so that the case you were saying that virtio pmd
> > continues to work without the bind looks like a bug to me.
> >
> > Can anyone confirm that?
> 
> That document isn't accurate. virtio doesn't require binding to UIO
> driver if it uses PORT IO. The PORT IO commit said it is because UIO
> isn't secure, but avoid using uio doesn't bring more security as virtio
> PMD still could ask device to DMA into any memory.
> The thing we at least we might do is fail in virtio_resource_init if
> kernel driver is still manipulating this device. This saves the effort
> users use blacklist option and avoids the driver conflict.

+1 for checking kernel driver in use