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

[Yuanhan Liu]

On Wed, Dec 23, 2015 at 11:26:17PM +0100, Thomas Monjalon wrote:
> 2015-12-23 10:09, Yuanhan Liu:
> > On Wed, Dec 23, 2015 at 09:55:54AM +0800, Peter Xu wrote:
> > > On Tue, Dec 22, 2015 at 04:38:30PM +0000, Xie, Huawei wrote:
> > > > On 12/22/2015 7:39 PM, Peter Xu wrote:
> > > > > I tried to unbind one of the virtio net device, I see the PCI entry
> > > > > still there.
> > > > >
> > > > > Before unbind:
> > > > >
> > > > > [root at vm proc]# lspci -k -s 00:03.0
> > > > > 00:03.0 Ethernet controller: Red Hat, Inc Virtio network device
> > > > >         Subsystem: Red Hat, Inc Device 0001
> > > > >         Kernel driver in use: virtio-pci
> > > > > [root at vm proc]# cat /proc/ioports | grep c060-c07f
> > > > >   c060-c07f : 0000:00:03.0
> > > > >     c060-c07f : virtio-pci
> > > > >
> > > > > After unbind:
> > > > >
> > > > > [root at vm proc]# lspci -k -s 00:03.0
> > > > > 00:03.0 Ethernet controller: Red Hat, Inc Virtio network device
> > > > >         Subsystem: Red Hat, Inc Device 0001
> > > > > [root at vm proc]# cat /proc/ioports | grep c060-c07f
> > > > >   c060-c07f : 0000:00:03.0
> > > > >
> > > > > So... does this means that it is an alternative to black list
> > > > > solution?
> > > > Oh, we could firstly check if this port is manipulated by kernel driver
> > > > in virtio_resource_init/eth_virtio_dev_init, as long as it is not too late.
> > 
> > Why can't we simply quit at pci_scan_one, once finding that it's not
> > bond to uio (or similar stuff)? That would be generic enough, that we
> > don't have to do similar checks for each new pmd driver.
> > 
> > Or, am I missing something?
> 
> UIO is not needed to make virtio works (without interrupt support).
> Sometimes it may be required to avoid using kernel modules.

While dig the git history, I found the commit:

    commit da978dfdc43b59e290a46d7ece5fd19ce79a1162
    Author: Ouyang Changchun <changchun.ouyang at intel.com>
    Date:   Mon Feb 9 09:14:06 2015 +0800
    
        virtio: use port IO to get PCI resource
    
        Make virtio not require UIO for some security reasons, this is to match
        6WIND's virtio-net-pmd.
    
        Signed-off-by: Changchun Ouyang <changchun.ouyang at intel.com>
        Acked-by: Huawei Xie <huawei.xie at intel.com>

The commit log is not well written, giving no explanation about the
"some security reasons".

Anyway, I see it now that it's kind of a design.


Note that my first patch set about enabling virtio 1.0 [0] sets the
RTE_PCI_DRV_NEED_MAPPING flag, which somehow implies that uio is a
must, otherwise, eal init would fail at pci_map_device().

So that my pathset breaks the un-documented rule, and I need fix it.
How about adding a wrapper, say rte_pci_map_device(), and exporting
it, so that virtio pmd driver could map resources when necessary?

[0]: http://dpdk.org/dev/patchwork/bundle/yliu/virtio-1.0-pmd/


> > > I guess there might be two problems? Which are:
> > > 
> > > 1. How user avoid DPDK taking over virtio devices that they do not
> > >    want for IO (chooses which device to use)
> > 
> > Isn't that what's the 'binding/unbinding' for?
> 
> Binding is, sometimes, required.

We may need fix the doc then. As the doc says it's a must:

    3.6. Binding and Unbinding Network Ports to/from the Kernel Modules

    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.


	--yliu

> But does it mean DPDK should use every available ports?
> That's the default and may be configured with blacklist/whitelist.
> 
> > > 2. Driver conflict between virtio PMD in DPDK, and virtio-pci in
> > >    kernel (happens on every virtio device that DPDK uses)
> > 
> > If you unbinded the kernel driver first, which is the suggested (or
> > must?) way to use DPDK, that will not happen.