Re: [dpdk-dev] [PATCH v3 6/6] DPDK changes for accommodating ENIC PMD

[Neil Horman <nhorman@tuxdriver.com>]

On Mon, Nov 24, 2014 at 05:45:54AM +0000, Sujith Sankar (ssujith) wrote:
> 
> 
> On 24/11/14 5:47 am, "Neil Horman" <nhorman@tuxdriver.com> wrote:
> 
> >On Sun, Nov 23, 2014 at 09:38:19PM +0530, Sujith Sankar wrote:
> >> Signed-off-by: Sujith Sankar <ssujith@cisco.com>
> >> ---
> >>  config/common_linuxapp                             | 5 +++++
> >>  lib/Makefile                                       | 1 +
> >>  lib/librte_eal/linuxapp/eal/eal_pci_vfio.c         | 7 +++++++
> >>  lib/librte_eal/linuxapp/eal/include/eal_pci_init.h | 1 +
> >>  mk/rte.app.mk                                      | 4 ++++
> >>  5 files changed, 18 insertions(+)
> >> 
> >> diff --git a/config/common_linuxapp b/config/common_linuxapp
> >> index 57b61c9..3c091e7 100644
> >> --- a/config/common_linuxapp
> >> +++ b/config/common_linuxapp
> >> @@ -210,6 +210,11 @@ CONFIG_RTE_LIBRTE_I40E_QUEUE_NUM_PER_VM=4
> >>  CONFIG_RTE_LIBRTE_I40E_ITR_INTERVAL=-1
> >>  
> >>  #
> >> +# Compile burst-oriented Cisco ENIC PMD driver
> >> +#
> >> +CONFIG_RTE_LIBRTE_ENIC_PMD=y
> >> +
> >> +#
> >>  # Compile burst-oriented VIRTIO PMD driver
> >>  #
> >>  CONFIG_RTE_LIBRTE_VIRTIO_PMD=y
> >> diff --git a/lib/Makefile b/lib/Makefile
> >> index e3237ff..1911790 100644
> >> --- a/lib/Makefile
> >> +++ b/lib/Makefile
> >> @@ -43,6 +43,7 @@ DIRS-$(CONFIG_RTE_LIBRTE_CMDLINE) += librte_cmdline
> >>  DIRS-$(CONFIG_RTE_LIBRTE_ETHER) += librte_ether
> >>  DIRS-$(CONFIG_RTE_LIBRTE_E1000_PMD) += librte_pmd_e1000
> >>  DIRS-$(CONFIG_RTE_LIBRTE_IXGBE_PMD) += librte_pmd_ixgbe
> >> +DIRS-$(CONFIG_RTE_LIBRTE_IXGBE_PMD) += librte_pmd_enic
> >>  DIRS-$(CONFIG_RTE_LIBRTE_I40E_PMD) += librte_pmd_i40e
> >>  DIRS-$(CONFIG_RTE_LIBRTE_PMD_BOND) += librte_pmd_bond
> >>  DIRS-$(CONFIG_RTE_LIBRTE_PMD_RING) += librte_pmd_ring
> >> diff --git a/lib/librte_eal/linuxapp/eal/eal_pci_vfio.c
> >>b/lib/librte_eal/linuxapp/eal/eal_pci_vfio.c
> >> index c776ddc..6bf8f2e 100644
> >> --- a/lib/librte_eal/linuxapp/eal/eal_pci_vfio.c
> >> +++ b/lib/librte_eal/linuxapp/eal/eal_pci_vfio.c
> >> @@ -736,6 +736,7 @@ pci_vfio_map_resource(struct rte_pci_device *dev)
> >>  		maps[i].offset = reg.offset;
> >>  		maps[i].size = reg.size;
> >>  		dev->mem_resource[i].addr = bar_addr;
> >> +		dev->mem_resource[i].len = reg.size;
> >>  	}
> >>  
> >>  	/* if secondary process, do not set up interrupts */
> >> @@ -791,4 +792,10 @@ pci_vfio_is_enabled(void)
> >>  {
> >>  	return vfio_cfg.vfio_enabled;
> >>  }
> >> +
> >> +int
> >> +pci_vfio_container_fd(void)
> >> +{
> >> +	return vfio_cfg.vfio_container_fd;
> >> +}
> >You should move this function definition to a separate patch and put it
> >earlier
> >in the series, as you call this function two patches back.
> 
> Thanks for the comment, Neil.  I shall move this to a separate patch and
> put it earlier in the series.
> 
> >
> >Also, this gives me pause, as it seems like you're working around the
> >VFIO api.
> >From what I see, you just use this to get an fd that you can use in an
> >ioctl to
> >set some DMA settings.  First off, theres already a function called
> >pci_vfio_get_container_fd, which does exactly what you are doing here,
> >with
> >additional safety checking.
> >
> >However, even though there is an existing function to do what you want, I
> >would
> >recommend that you not use it for your purposes.  Whenever you expose
> >something
> >like a file descriptor, you run the risk of multiple accessors racing
> >trying to
> >set/unset features and preform operations.  It would be better if you
> >could add
> >apropriate api calls to vfio interface to set what you want.  That way the
> >library can add appropriate locking if/when needed
> 
> 
> I see that vfio_cfg.vfio_container_fd is obtained and stored in
> pci_vfio_enable(), and this is not modified later.
> ENIC PMD needs it to add the IOMMU mapping for buffers used for
> communicating with adapter firmware.  That¹s just adding an entry, and
> container fd is just passed as an argument.  So the following addition in
> eal_pci_vfio.c should be sufficient.  Since vfio_cfg is per process, I do
> not think that any other checking is required.
> 
> int
> pci_vfio_map_dma(struct vfio_iommu_type1_dma_map *dma_map)
> {
> 	return ioctl(vfio_cfg.vfio_container_fd, VFIO_IOMMU_MAP_DMA, dma_map);
> }
> 
> 
> 
> Does this look alright?  Do you think that I¹ve missed out anything here?
> 
It looks better yes, but I'm still confused as to why its necessecary.  Looking
back at your use of the fd, you're adding an iov entry for a buffer allocated
via rte_memzone_reserve, which should come out of the dpdk's configured memory.
In pci_vfio_setup_dma_maps, which is part of the pci probe path in eal library,
all of the DPDK's physically available memory is already mapped into the iommu
in a 1:1 fashion.  So why do you need to do this again?

Neil