Re: [PATCH v3 1/2] eal/pci: introduce a PCI driver flag

[Thomas Monjalon <thomas@monjalon.net>]

13/06/2017 09:24, Jerin Jacob:
> From: Thomas Monjalon <thomas@monjalon.net>
> > 13/06/2017 06:43, Jerin Jacob:
> > > From: Thomas Monjalon <thomas@monjalon.net>
> > > > 09/06/2017 12:27, Jerin Jacob:
> > > > > Some ethdev devices like nicvf thunderx PMD need special treatment for
> > > > > Secondary queue set(SQS) PCIe VF devices, where, it expects to not unmap
> > > > > or free the memory without registering the ethdev subsystem.
> > > > > 
> > > > > Introducing a new RTE_PCI_DRV_KEEP_MAPPED_RES
> > > > > PCI driver flag to request PCI subsystem to not unmap the mapped PCI
> > > > > resources(PCI BAR address) if unsupported device detected.
> > > > > 
> > > > > Suggested-by: Ferruh Yigit <ferruh.yigit@intel.com>
> > > > > Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
> > > > [...]
> > > > > --- a/lib/librte_eal/common/eal_common_pci.c
> > > > > +++ b/lib/librte_eal/common/eal_common_pci.c
> > > > > @@ -221,7 +221,12 @@ rte_pci_probe_one_driver(struct rte_pci_driver *dr,
> > > > >  	ret = dr->probe(dr, dev);
> > > > >  	if (ret) {
> > > > >  		dev->driver = NULL;
> > > > > -		if (dr->drv_flags & RTE_PCI_DRV_NEED_MAPPING)
> > > > > +		if ((dr->drv_flags & RTE_PCI_DRV_NEED_MAPPING) &&
> > > > > +			/* Don't unmap if device is unsupported and
> > > > > +			 * driver needs mapped resources.
> > > > > +			 */
> > > > > +			!(ret > 0 &&
> > > > > +				(dr->drv_flags & RTE_PCI_DRV_KEEP_MAPPED_RES)))
> > > > >  			rte_pci_unmap_device(dev);
> > > > >  	}
> > > > >  
> > > > > --- a/lib/librte_eal/common/include/rte_pci.h
> > > > > +++ b/lib/librte_eal/common/include/rte_pci.h
> > > > > +/** Device driver needs to keep mapped resources if unsupported dev detected */
> > > > > +#define RTE_PCI_DRV_KEEP_MAPPED_RES 0x0020
> > > > 
> > > > If I understand well, you want to map resources but not probe it?
> > > 
> > > Yes.
> > > 
> > > > Shouldn't it be less hacky to probe it as a (new) null class?
> > > 
> > > The Vendor and Class ID is same for those device too so we need to map
> > > the PCI bar and have access to know the class of device. If you are concerned about
> > > if it an common code change, My first version was without common code change.
> > > http://dpdk.org/dev/patchwork/patch/24983/
> > > 
> > > Ferruh would like to have flag scheme, I think it make sense for
> > > PMD maintenance perspective.
> > 
> > Yes
> > 
> > My idea was to have a new class of device interface to reserve those
> > resources, so the probe function would succeed.
> > Do you think it would be a good idea?
> 
> Currently Kernel PF code creates 12 SRIOV VF devices per port(one VF device has
> 8 queues === 96 queues(12VFs) for 96 cores(thunderx max cores)), out of that 1 VF
> device is _primary_ which mapped to dpdk ethdev port. If probe succeeds for
> another 11 VFs then too may ethdev NULL ports show up. We can support
> up to 12 ports(12*12 VF = 144 ports). I think, it is not good from
> end user perceptive. We already have unsupported device concept in eal device
> framework(when probe returns > 0). IMHO, it OK to keep as it for
> simplicity.

Thanks for the explanation.
So we have a flag RTE_PCI_DRV_KEEP_MAPPED_RES for this kind of device.
Maybe someone else will think to another usage of a null interface,
so I would like to keep this idea floating in the air, just in case :)