Re: [dpdk-dev] [PATCH 09/11] eal: move PCI table macro

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2016-07-08 09:56, Neil Horman:
> On Fri, Jul 08, 2016 at 10:49:25AM +0200, David Marchand wrote:
> > Hello Thomas, Neil,
> > 
> > (will be back in a couple of days, thanks Thomas for pointing this thread)
> > 
> > 
> > On Thu, Jul 7, 2016 at 6:25 PM, Thomas Monjalon
> > <thomas.monjalon@6wind.com> wrote:
> > > 2016-07-07 12:11, Neil Horman:
> > >> On Thu, Jul 07, 2016 at 05:36:28PM +0200, Thomas Monjalon wrote:
> > >> > Remove include of rte_pci.h in the generic header rte_dev.h
> > >> > and move the macro DRIVER_REGISTER_PCI_TABLE in rte_pci.h.
> > > [...]
> > >>
> > >> This seems strange to me, in that its odd for the driver information export
> > >> macros to be spread out in multiple locations.  Specifically it enjoins the use
> > >> of the DRV_EXP_TAG macro, which helps centralize tag naming.  Perhaps the happy
> > >> medium is to place all the export macros (includnig PMD_REGISTER_DRIVER) into
> > >> its own pmd_register.h header?
> > >
> > > I don't know.
> > > David, your opinion?
> > 
> > - The suggestion I did offline to Thomas was to move pci stuff in pci headers.
> > We are trying to move from the "all pci" code in eal to accomodate for
> > other "buses" / architectures.
> I get that, but I'm not sure that applies here.  The macro in question is
> specific to pci busses, and if there is additional bus information to export, it
> will have its own macro (e.g. DRIVER_REGISTER_USB_TABLE or some such).  While
> I could see that being an argument for putting each macro in with its own bus
> type, I think thats the wrong organization here, in that people writing drivers
> will want to know what export macros are available and will expect to look in a
> single place for it.
> 
> > Having a pci macro in a generic header like rte_dev.h is wrong to me.
> > Moving this to a new header like pmd_register.h sounds like a new
> > generic header with pci specific stuff in it.
> Well, yes, but I see that as no different than rte_ethdev.c or rte_pdump.c.
> both of those files will need to know about all the different types of busses
> you support and have to include those corresponding header files (i.e. they will
> have to include rte_pci.h, rte_usb.h, rte_i2c.h, etc).  This is really no
> different in my mind. 
> 
> > So, I am not sure I follow you Neil.
> > 
> > Can you elaborate ?
> > 
> I suppose the best way to describe it is that while I understand and support the
> desire to separate and abstract bus information away from device function, I
> think theres a pragmatic descision here to prioritize functional domain over
> header inclusion.  That is to say, I think when people are writing a driver, it
> will be helpful to have all the export macros in a single location so they know
> what information they can export, and that includes registration of various bus
> type identifiers.  So a file like pmd_registration.h that includes rte_pci.h,
> rte_usb.h, rte_i2c.h, etc is more useful to a developer, than spreading these
> macros out to those various header files, for the sake of avoiding a potentially
> unneeded include.
> 
> > 
> > - Why do you want to centralise the tag naming ?
> > To avoid collisions ?
> Yes, and to centralize that information.  Since the pmdinfogen tool needs to
> know what those tag names are as well, its useful to keep them in the same area
> to maintain co-ordination.  Its also useful because it means we can use one
> macro to define tag naming convention, instead of having to re-implement or
> dead-reckon it in multiple files.
> 
> > Well, adding those tags should not happen that often and I think we
> > can maintain this with careful reviews.
> I don't agree with that.  This discussion is based on the fact that you expect
> that we will be adding additional bus types in the future correct?  Well, given
> that we have a pci bus specific export macro, I would expect that to proliferate
> to every other bus type as well, and so we can expect to at least have a new
> tag added for every bus that is added, in addition to any other bus agnostic
> information people wish to export (just off hand, looking at the linux modinfo
> section, we might expect module author, module version, alias names, licensing
> infomration, and others to be potential export candidates).  So, depending on
> how much this is adopted, I think we can potentially expect a great deal of
> additional tagging to be needed.

Anyway, this macro do not need rte_pci.h.
So the minimal patch can be to just remove this include.