From mboxrd@z Thu Jan 1 00:00:00 1970 From: Neil Horman Subject: Re: [PATCH v3 6/6] DPDK changes for accommodating ENIC PMD Date: Mon, 24 Nov 2014 12:19:42 -0500 Message-ID: <20141124171942.GC7532@hmsreliant.think-freely.org> References: <1416758899-1351-1-git-send-email-ssujith@cisco.com> <1416758899-1351-7-git-send-email-ssujith@cisco.com> <20141124001744.GA7751@localhost.localdomain> <20141124113315.GA6038@hmsreliant.think-freely.org> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable Cc: "dev-VfR2kkLFssw@public.gmane.org" , "Prasad Rao \(prrao\)" To: "Sujith Sankar (ssujith)" Return-path: Content-Disposition: inline In-Reply-To: List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces-VfR2kkLFssw@public.gmane.org Sender: "dev" On Mon, Nov 24, 2014 at 04:12:48PM +0000, Sujith Sankar (ssujith) wrote: >=20 >=20 > On 24/11/14 5:03 pm, "Neil Horman" wrote: >=20 > >On Mon, Nov 24, 2014 at 05:45:54AM +0000, Sujith Sankar (ssujith) wrot= e: > >>=20 > >>=20 > >> On 24/11/14 5:47 am, "Neil Horman" wrote: > >>=20 > >> >On Sun, Nov 23, 2014 at 09:38:19PM +0530, Sujith Sankar wrote: > >> >> Signed-off-by: Sujith Sankar > >> >> --- > >> >> 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(+) > >> >>=20 > >> >> 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=3D4 > >> >> CONFIG_RTE_LIBRTE_I40E_ITR_INTERVAL=3D-1 > >> >> =20 > >> >> # > >> >> +# Compile burst-oriented Cisco ENIC PMD driver > >> >> +# > >> >> +CONFIG_RTE_LIBRTE_ENIC_PMD=3Dy > >> >> + > >> >> +# > >> >> # Compile burst-oriented VIRTIO PMD driver > >> >> # > >> >> CONFIG_RTE_LIBRTE_VIRTIO_PMD=3Dy > >> >> diff --git a/lib/Makefile b/lib/Makefile > >> >> index e3237ff..1911790 100644 > >> >> --- a/lib/Makefile > >> >> +++ b/lib/Makefile > >> >> @@ -43,6 +43,7 @@ DIRS-$(CONFIG_RTE_LIBRTE_CMDLINE) +=3D librte_c= mdline > >> >> DIRS-$(CONFIG_RTE_LIBRTE_ETHER) +=3D librte_ether > >> >> DIRS-$(CONFIG_RTE_LIBRTE_E1000_PMD) +=3D librte_pmd_e1000 > >> >> DIRS-$(CONFIG_RTE_LIBRTE_IXGBE_PMD) +=3D librte_pmd_ixgbe > >> >> +DIRS-$(CONFIG_RTE_LIBRTE_IXGBE_PMD) +=3D librte_pmd_enic > >> >> DIRS-$(CONFIG_RTE_LIBRTE_I40E_PMD) +=3D librte_pmd_i40e > >> >> DIRS-$(CONFIG_RTE_LIBRTE_PMD_BOND) +=3D librte_pmd_bond > >> >> DIRS-$(CONFIG_RTE_LIBRTE_PMD_RING) +=3D 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 =3D reg.offset; > >> >> maps[i].size =3D reg.size; > >> >> dev->mem_resource[i].addr =3D bar_addr; > >> >> + dev->mem_resource[i].len =3D reg.size; > >> >> } > >> >> =20 > >> >> /* 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 pu= t it > >> >earlier > >> >in the series, as you call this function two patches back. > >>=20 > >> Thanks for the comment, Neil. I shall move this to a separate patch= and > >> put it earlier in the series. > >>=20 > >> > > >> >Also, this gives me pause, as it seems like you're working around t= he > >> >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 he= re, > >> >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 expo= se > >> >something > >> >like a file descriptor, you run the risk of multiple accessors raci= ng > >> >trying to > >> >set/unset features and preform operations. It would be better if y= ou > >> >could add > >> >apropriate api calls to vfio interface to set what you want. That = way > >>the > >> >library can add appropriate locking if/when needed > >>=20 > >>=20 > >> 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. That=C2=B9s just adding an ent= ry, and > >> container fd is just passed as an argument. So the following additi= on > >>in > >> eal_pci_vfio.c should be sufficient. Since vfio_cfg is per process,= I > >>do > >> not think that any other checking is required. > >>=20 > >> 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_ma= p); > >> } > >>=20 > >>=20 > >>=20 > >> Does this look alright? Do you think that I=C2=B9ve missed out anyt= hing > >>here? > >>=20 > >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 configure= d > >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 t= he > >iommu > >in a 1:1 fashion. So why do you need to do this again? >=20 > Yes, ideally all the memory allocated using rte_memzone_reserve_aligned= () > should have its mapping already there in IOMMU. The code that adds the > mapping was not there initially, but while testing, I saw that it was > giving DMAR errors (like what=E2=80=99s shown below). Putting it back = solved the > problem. >=20 >=20 > [295042.852620] DMAR:[fault reason 05] PTE Write access is not set > [295044.406282] dmar: DRHD: handling fault status reg 602 > [295044.467981] dmar: DMAR:[DMA Write] Request device [86:00.0] fault a= ddr > 7fbd73b32000 >=20 Ok, so what I hear you saying is that, now that that mapping code is avai= lable in the pci probe path, you can remove your workaround correct? Neil >=20 > Regards, > -Sujith >=20 > > > >Neil > > >=20