[dpdk-dev] [PATCH v2] igb_uio: add config option to control reset

[Roberts, Lee A.]

> -----Original Message-----
> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Tan, Jianfeng
> Sent: Thursday, November 02, 2017 8:57 PM
> To: Ferruh Yigit <ferruh.yigit at intel.com>; Thomas Monjalon <thomas at monjalon.net>
> Cc: dev at dpdk.org; stable at dpdk.org; Jingjing Wu <jingjing.wu at intel.com>; Shijith Thotton
> <shijith.thotton at caviumnetworks.com>; Gregory Etelson <gregory at weka.io>; Harish Patil
> <harish.patil at cavium.com>; George Prekas <george.prekas at epfl.ch>; Sergio Gonzalez Monroy
> <sergio.gonzalez.monroy at intel.com>; Rasesh Mody <rasesh.mody at cavium.com>
> Subject: Re: [dpdk-dev] [PATCH v2] igb_uio: add config option to control reset
> 
> 
> 
> On 11/3/2017 8:51 AM, Ferruh Yigit wrote:
> > Adding a compile time configuration option to control device reset done
> > during DPDK application exit.
> >
> > Config option is CONFIG_RTE_EAL_IGB_UIO_RESET and enabled by default,
> > so by default reset will happen. Having this reset is safer to be sure
> > device left in a proper case.
> >
> > But for special cases [1] it is possible to disable the config option
> > to prevent the device reset.
> >
> > [1]
> > http://dpdk.org/ml/archives/dev/2017-November/080927.html
> >
> > Fixes: b58eedfc7dd5 ("igb_uio: issue FLR during open and release of device file")
> > Cc: stable at dpdk.org
> >
> > Signed-off-by: Ferruh Yigit <ferruh.yigit at intel.com>
> 
> Realize that we do have a pci_clear_master() in the release() to disable
> the DMA from device until the next open() will enable the DMA again .
> Here is my:
> 
> Reviewed-by: Jianfeng Tan <jianfeng.tan at intel.com>
> 
> Thanks,
> Jianfeng
> 
> > ---
> > Cc: Jianfeng Tan <jianfeng.tan at intel.com>
> > Cc: Jingjing Wu <jingjing.wu at intel.com>
> > Cc: Shijith Thotton <shijith.thotton at caviumnetworks.com>
> > Cc: Gregory Etelson <gregory at weka.io>
> > Cc: Harish Patil <harish.patil at cavium.com>
> > Cc: George Prekas <george.prekas at epfl.ch>
> > Cc: Sergio Gonzalez Monroy <sergio.gonzalez.monroy at intel.com>
> > Cc: Rasesh Mody <rasesh.mody at cavium.com>
> >
> > v2:
> > * fix typo in commit log
> > ---
> >   config/common_base                        | 1 +
> >   config/common_linuxapp                    | 1 +
> >   lib/librte_eal/linuxapp/igb_uio/igb_uio.c | 2 ++
> >   3 files changed, 4 insertions(+)
> >
> > diff --git a/config/common_base b/config/common_base
> > index 82ee75456..2a9947420 100644
> > --- a/config/common_base
> > +++ b/config/common_base
> > @@ -102,6 +102,7 @@ CONFIG_RTE_LIBEAL_USE_HPET=n
> >   CONFIG_RTE_EAL_ALLOW_INV_SOCKET_ID=n
> >   CONFIG_RTE_EAL_ALWAYS_PANIC_ON_ERROR=n
> >   CONFIG_RTE_EAL_IGB_UIO=n
> > +CONFIG_RTE_EAL_IGB_UIO_RESET=n
> >   CONFIG_RTE_EAL_VFIO=n
> >   CONFIG_RTE_MALLOC_DEBUG=n
> >   CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=n
> > diff --git a/config/common_linuxapp b/config/common_linuxapp
> > index 74c7d64ec..b3a602909 100644
> > --- a/config/common_linuxapp
> > +++ b/config/common_linuxapp
> > @@ -37,6 +37,7 @@ CONFIG_RTE_EXEC_ENV_LINUXAPP=y
> >
> >   CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=y
> >   CONFIG_RTE_EAL_IGB_UIO=y
> > +CONFIG_RTE_EAL_IGB_UIO_RESET=y
> >   CONFIG_RTE_EAL_VFIO=y
> >   CONFIG_RTE_KNI_KMOD=y
> >   CONFIG_RTE_LIBRTE_KNI=y
> > diff --git a/lib/librte_eal/linuxapp/igb_uio/igb_uio.c b/lib/librte_eal/linuxapp/igb_uio/igb_uio.c
> > index fd320d87d..0325722c0 100644
> > --- a/lib/librte_eal/linuxapp/igb_uio/igb_uio.c
> > +++ b/lib/librte_eal/linuxapp/igb_uio/igb_uio.c
> > @@ -360,7 +360,9 @@ igbuio_pci_release(struct uio_info *info, struct inode *inode)
> >   	/* stop the device from further DMA */
> >   	pci_clear_master(dev);
> >
> > +#ifdef RTE_EAL_IGB_UIO_RESET
> >   	pci_reset_function(dev);
> > +#endif
> >
> >   	return 0;
> >   }

A compile time configuration option makes life very difficult for application providers.

Consider the case where an application such as Open vSwitch with DPDK support is being provided
with a Linux distribution.  One would want the Open vSwitch binary to support as many vendor NICs
as possible---without the need to recompile.  With a change such as this, one would need to have
different versions of the kernel igb_uio module to support different NICs.

The Linux kernel is already aware of, and provides work-arounds for, various PCI quirks.
For example, see linux/drivers/pci/quirks.c (http://elixir.free-electrons.com/linux/latest/source/drivers/pci/quirks.c).

At this point in igb_uio.c, one is aware of the struct pci_dev "dev" for the device in question.
Access to the vendor and device information should be simple:

struct pci_dev {
	struct list_head bus_list;	/* node in per-bus list */
	struct pci_bus	*bus;		/* bus this device is on */
	struct pci_bus	*subordinate;	/* bus this device bridges to */

	void		*sysdata;	/* hook for sys-specific extension */
	struct proc_dir_entry *procent;	/* device entry in /proc/bus/pci */
	struct pci_slot	*slot;		/* Physical slot this device is in */

	unsigned int	devfn;		/* encoded device & function index */
	unsigned short	vendor;
	unsigned short	device;
	unsigned short	subsystem_vendor;
	unsigned short	subsystem_device;
...

One could imagine using logic to implement corresponding PCI quirks that can be evaluated
at runtime.  For example (in pseudocode),

     if not (vendor = "Cavium" and device = "bnx2x")
        then pci_reset_function(dev);

There are other possible implementations.  If there are enough quirks, one might have action
functions defined---and a table of function pointers associated with each PMD to select the
proper action.

                                                - Lee Roberts