[dpdk-dev] [PATCH v3 3/8] net/failsafe: support probed sub-devices getting

[Gaëtan Rivet]

On Tue, Jan 16, 2018 at 12:27:57PM +0000, Matan Azrad wrote:
> Hi Gaetan
> 
> From: Gaëtan Rivet, Tuesday, January 16, 2018 1:09 PM
> > Hi Matan,
> > 
> > I'n not fond of the commit title, how about:
> > 
> > [PATCH v3 3/8] net/failsafe: add probed etherdev capture
> > 
> > ?
> > 
> OK, no problem.
> 
> > On Tue, Jan 09, 2018 at 02:47:28PM +0000, Matan Azrad wrote:
> > > Previous fail-safe code didn't support getting probed sub-devices and
> > > failed when it tried to probe them.
> > >
> > > Skip fail-safe sub-device probing when it already was probed.
> > >
> > > Signed-off-by: Matan Azrad <matan at mellanox.com>
> > > Cc: Gaetan Rivet <gaetan.rivet at 6wind.com>
> > > ---
> > >  doc/guides/nics/fail_safe.rst       |  5 ++++
> > >  drivers/net/failsafe/failsafe_eal.c | 60
> > > ++++++++++++++++++++++++-------------
> > >  2 files changed, 45 insertions(+), 20 deletions(-)
> > >
> > > diff --git a/doc/guides/nics/fail_safe.rst
> > > b/doc/guides/nics/fail_safe.rst index 5b1b47e..b89e53b 100644
> > > --- a/doc/guides/nics/fail_safe.rst
> > > +++ b/doc/guides/nics/fail_safe.rst
> > > @@ -115,6 +115,11 @@ Fail-safe command line parameters
> > >    order to take only the last line into account (unlike ``exec()``) at every
> > >    probe attempt.
> > >
> > > +.. note::
> > > +
> > > +   In case of whitelist sub-device probed by EAL, fail-safe PMD will take the
> > device
> > > +   as is, which means that EAL device options are taken in this case.
> > > +
> > >  - **mac** parameter [MAC address]
> > >
> > >    This parameter allows the user to set a default MAC address to the
> > > fail-safe diff --git a/drivers/net/failsafe/failsafe_eal.c
> > > b/drivers/net/failsafe/failsafe_eal.c
> > > index 19d26f5..7bc7453 100644
> > > --- a/drivers/net/failsafe/failsafe_eal.c
> > > +++ b/drivers/net/failsafe/failsafe_eal.c
> > > @@ -36,39 +36,59 @@
> > >  #include "failsafe_private.h"
> > >
> > >  static int
> > > +fs_get_port_by_device_name(const char *name, uint16_t *port_id)
> > 
> > The naming convention for the failsafe driver is
> > 
> >       namespace_object_sub-object_action()
> > 
> OK.
> > With an ordering of objects by their scope (std, rte, failsafe, file).
> > Also, "get" as an action is not descriptive enough.
> > 
> Isn't "get by device name" descriptive? 

The endgame is capturing a device that we know we are interested in.
The device name being used for matching is an implementation detail,
which should be abstracted by using a sub-function.

Putting this in the name defeat the reason for using another function.

> > static int
> > fs_ethdev_capture(const char *name, uint16_t *port_id);
> > 
> You miss here the main reason why we need this function instead of using rte_eth_dev_get_port_by_name.
> The reason we need this function is because we want to find the device by the device name and not ethdev name.
> What's about  fs_port_capture_by_device_name?

You are getting a port_id that is only valid for the rte_eth_devices
array, by using the ethdev iterator. You are only looking for an ethdev.

So it doesn't really matter whether you are using the ethdev name or the
device name, in the end you are capturing an ethdev
--> fs_ethdev_capture seems good for me.

Now, I guess you will say that the user would need to know that they
have to provide a device name that would be written in device->name. The
issue here is that you have a leaky abstraction for your function,
forcing this kind of consideration on your function user.

So I'd go further and will ask you to change the `const char *name` to a
`const rte_devargs *da` in the parameters.

> Maybe comparing it to device->devargs->name is better, What do you think?
> 

You are touching at a pretty contentious subject here :) .

Identifying devices is not currently a well-defined function in DPDK.
Some ports (actually, only one model: ConnectX-3) will have several
ports using the same PCI slot. But even ignoring this glaring problem...

As it is, the device->name for PCI will match the name given as a
devargs, so functionally this should not change anything.

Furthermore, you will have devices probed without any devargs. The
fail-safe would thus be unable to capture non-blacklisted devices when
the PCI bus is in blacklist mode.

These not-blacklisted devices actually will have a full-PCI name (DomBDF
format), so a simple match with the one passed in your fail-safe devargs
will fail, ex:

   # A physical port exists at 0000:00:02.0
   testpmd --vdev="net_failsafe,dev(00:02.0)" -- -i

Would fail to capture the device 0000:00:02.0, as this is the name that
the PCI bus would give to this device, in the absence of a user-given
name.

In 18.05, or 18.08 there should be an EAL function that would be able to
identify a device given a specific ID string (very close to an
rte_devargs). Currently, this API does not exist.

You can hack your way around this for the moment, IF you really, really
want: parse your devargs, get the bus, use the bus->parse() function to
get a binary device representation, and compare bytes per bytes the
binary representation given by your devargs and by the device->name.

But this is a hack, and a pretty ugly one at that: you have
no way of knowing the size taken by this binary representation, so you
can restrict yourself to the vdev and PCI bus for the moment and take
the larger of an rte_vdev_driver pointer and an rte_pci_addr....

        {
            union {
                    rte_vdev_driver *drv;
                    struct rte_pci_addr pci_addr;
            } bindev1, bindev2;
            memset(&bindev1, 0, sizeof(bindev1));
            memset(&bindev2, 0, sizeof(bindev2));
            rte_eal_devargs_parse(device->name, da1);
            rte_eal_devargs_parse(your_devstr, da2);
            RTE_ASSERT(da1->bus == rte_bus_find_by_name("pci") ||
                       da1->bus == rte_bus_find_by_name("vdev"));
            RTE_ASSERT(da2->bus == rte_bus_find_by_name("pci") ||
                       da2->bus == rte_bus_find_by_name("vdev"));
            da1->bus->parse(da1->name, &bindev1);
            da1->bus->parse(da2->name, &bindev2);
            if (memcmp(&bindev1, &bindev2, sizeof(bindev1)) == 0) {
                    /* found the device */
            } else {
                    /* not found */
            }
        }

So, really, really ugly. Anyway.

<snip>

> > > +			/* Take control of device probed by EAL options. */
> > > +			DEBUG("Taking control of a probed sub device"
> > > +			      " %d named %s", i, da->name);
> > 
> > In this case, the devargs of the probed device must be copied within the sub-
> > device definition and removed from the EAL using the proper rte_devargs
> > API.
> > 
> > Note that there is no rte_devargs copy function. You can use
> > rte_devargs_parse instead, "parsing" again the original devargs into the sub-
> > device one. It is necessary for complying with internal rte_devargs
> > requirements (da->args being malloc-ed, at the moment, but may evolve).
> > 
> > The rte_eal_devargs_parse function is not easy enough to use right now,
> > you will have to build a devargs string (using snprintf) and submit it.
> > I proposed a change this release for it but it will not make it for 18.02, that
> > would have simplified your implementation.
> > 
> 
> Got you. You right we need to remove the created devargs in fail-safe parse level.
> What do you think about checking it in the parse level and avoid the new devargs creation?
> Also to do the copy in parse level(same method as we are doing in probe level)?
> 

Not sure I follow here, but the new rte_devargs is part of the
sub-device (it is not a pointer, but allocated alongside the
sub_device).

So keep everything here, it is the right place to deal with these
things.

-- 
Gaëtan Rivet
6WIND