Re: [dpdk-dev] [PATCH v16 1/3] build: disable/enable drivers in Arm builds

[Bruce Richardson <bruce.richardson@intel.com>]

On Wed, Mar 31, 2021 at 08:39:57AM +0000, Juraj Linkeš wrote:
> 
> 
> > -----Original Message-----
> > From: Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>
> > Sent: Tuesday, March 30, 2021 2:39 AM
> > To: Jerin Jacob <jerinjacobk@gmail.com>; Juraj Linkeš
> > <juraj.linkes@pantheon.tech>
> > Cc: Bruce Richardson <bruce.richardson@intel.com>; Ruifeng Wang
> > <Ruifeng.Wang@arm.com>; vcchunga@amazon.com; Dharmik Thakkar
> > <Dharmik.Thakkar@arm.com>; hemant.agrawal@nxp.com; Ajit Khaparde
> > (ajit.khaparde@broadcom.com) <ajit.khaparde@broadcom.com>;
> > ferruh.yigit@intel.com; aboyer@pensando.io; lironh@marvell.com;
> > dev@dpdk.org; nd <nd@arm.com>; Honnappa Nagarahalli
> > <Honnappa.Nagarahalli@arm.com>; nd <nd@arm.com>
> > Subject: RE: [PATCH v16 1/3] build: disable/enable drivers in Arm builds
> > 
> > <snip>
> > 
> > > > <removed parts which I think are not that relevant>
> > > >
> > > > > > > > > > > > The blocklist is, I think, agreed upon by everyone.
> > > > > > > > > > > > The question is whether we want to support the
> > > > > > > > > > > > allowlist alongside it and there seem to be enough
> > > > > > > > > > > > reasons to do
> > > that.
> > > > > > > > > > > Sorry, may be this is answered already, but, what
> > > > > > > > > > > additional benefit does allowlist provide over the blocklist?
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > VPP could use it:
> > > > > > > > > > https://gerrit.fd.io/r/gitweb?p=vpp.git;a=blob;f=build/e
> > > > > > > > > > xt
> > > > > > > > > > erna
> > > > > > > > > > l/pa
> > > > > > > > > > ckag es/dpdk
> > > > > > > > > > .mk;h=c35ac84c27b19411a0cfdf9a3524fdf36024762c;hb=HEAD
> > > > > > > > > >
> > > > > > > > > > They're disabling almost everything so an allowlist would fit there.
> > > > > > > > > > And they won't need to update the list when a new driver
> > > > > > > > > > is added (which they won't need).
> > > > > > > > > This is different from how we started this discussion. The
> > > > > > > > > current discussion was for DPDK internal use. But the one
> > > > > > > > > you are referencing above is for users of DPDK. I am fine
> > > > > > > > > for providing the allow list for the users of DPDK. But
> > > > > > > > > for DPDK internal, I think block list is
> > > > > > > enough.
> > > > > > > > >
> > > > > > > >
> > > > > > > > That's an interesting suggestion. Jerin, what do you think?
> > > > > > > > Why did you
> > > > > > > want to have an allowlist? Would this work?
> > > > > > >
> > > > > > > # The very reason why VPP chooses to have allow list so that
> > > > > > > they can control what needs to include.
> > > > > > > # Another use case is like, in SoCs have fixed internal
> > > > > > > devices, we can have optimized build for that can have only
> > > > > > > allow list of the drivers that care about # For server market,
> > > > > > > block list makes sense # For embedded SoC, allow list makes sense.
> > > > > > For the embedded SoC, IMO, the upstream project could allow the
> > > > > > compilation
> > > > > for wider set of PMDs/libs. May be the end customer can use the
> > > > > allow list to compile/use what is required?
> > > > >
> > > > > Just to understand, how end customer can enable allow list, if
> > > > > DPDK build system does not support it?
> > > > > Also to understand, If we are supporting blocklist, why not have
> > > > > allowlist (I mean, both of them) as both are required as it caters
> > > > > different use case as mention above. We can not emulate allowlist
> > > > > with blocklist as each version of DPDK will have new libraries and
> > > > > PMD's
> > > which end user has no clue. Right?
> > > >
> > > > > I think, that is the reason why VPP is doing the allow list.
> > > >
> > > > I'm not sure what you mean by this, but to clarify, VPP likely would
> > > > be using
> > > the allowlist in this fashion, but that is not an arm specific
> > > usecase. I think what Honnappa wanted to see was how the allowlist
> > > could be used in an arm usecase (such as using it in an SoC configuration).
> > >
> > > There is nothing arm-specific here. Right? allowlist will be common
> > > and will be used by all architecture. Right?
> > Nothing Arm specific. I think for generic Arm servers platforms we can make
> > sure that we allow for compilation of all the DPDK code. We can go ahead with
> > implementing the allow list.
> > 
> 
> I tried to actually use an allowlist and there are some problems with building apps and tests. When I tried to enable a random driver, such as common/sfc_efx, I've ran into dependency issues with apps:
> app/meson.build:53:3: ERROR:  Tried to get unknown variable "static_rte_bus_vdev".
> 
> This is because bus/vdev is not enabled (the allowlist only enabled net/sfc_efx). When I implemented dependency discovery in apps similar to which exists for drivers/libs, I was then unable to build tests (which is a special case app):
> app/test/meson.build:427:1: ERROR:  Tried to get unknown variable "static_rte_bus_pci".
> 
> This is becasue bus/pci is not enabled. If I understand the code correctly, the test dependencies are not matched to each test, meaning we can't disable the tests for which we don't have dependencies - we can only disable all tests.
> 
> In general, this problem also affects the blocklist, i.e. meson build -Ddisable_drivers=bus/pci produces
> drivers/net/hinic/base/meson.build:34:0: ERROR:  Unknown variable "static_rte_bus_pci".
> For blocklists, this is seldom going to be a problem, since most users won't disable "core" drivers.
> 
> I'm not sure what is the best course of action. We can implement the allowlist and then we'll leave it up to implementers to implement allowlists that produce a working build. We could then optionally address the dependency issues brought by disabled drivers in a separate patch.
> 
> Maybe we can just have a list of "core" drivers that can never be disabled (either using blocklists or allowlists)?
> 
> Bruce, what do you think?
> 
Hi,

from my experience this mainly seems to affect the bus drivers,
specifically pci and vdev buses, which seem to be assumed to be always
enabled. I agree it's probably not a real problem right now, though it
would be nice for some testing purposes to have a build possible with
"disable_drivers=*/*". For a solution I am happy for us to have whichever
is easiest to implement - either refusing disabling of bus/pci and
bus/vdev, or fixing the tests and apps to allow them to be built with
reduced functionality.

Apart from those two drivers, I would hope that disabling everything else
works. If not, we should definitely fix it.

/Bruce