Re: [dpdk-dev] [PATCH] doc: document the new devargs syntax

[Thomas Monjalon <thomas@monjalon.net>]

23/01/2018 17:08, Gaëtan Rivet:
> Hi Yuanhan, Thomas,
> 
> On Tue, Jan 23, 2018 at 03:29:34PM +0100, Thomas Monjalon wrote:
> > 23/01/2018 13:46, Yuanhan Liu:
> > > On Thu, Jan 18, 2018 at 10:46:23AM +0100, Gaëtan Rivet wrote:
> > > > On Thu, Jan 18, 2018 at 09:46:29AM +0100, Thomas Monjalon wrote:
> > > > > 18/01/2018 08:35, Yuanhan Liu:
> > > > > > On Wed, Jan 17, 2018 at 12:34:08PM +0000, Ferruh Yigit wrote:
> > > > > > > So does it make sense to separate them logically? Perhaps as "device identifier"
> > > > > > > and "device args".
> > > > > > 
> > > > > > Then I think it returns back to the old issue: how could we identify a
> > > > > > port when the bus id (say BDF for PCI bus) is not enough for identifying
> > > > > > a port? Such case could happen when a single NIC has 2 ports sharing
> > > > > > the same BDF. It could also happen with the VF representors that will
> > > > > > be introduced shortly.
> > > > > 
> > > > > Yes, the device matching syntax must include bus category, class category
> > > > > and driver category. So any device can be identified in future.
> > > > > 
> > > > > But I think Ferruh is talking about separating device matching
> > > > > (which is described in this proposal) and device settings
> > > > > (which are usually mixed in -w and --vdev options).
> > > > > I agree there are different things and may be separate.
> > > > > They could share the same syntax (bus/class/driver) but be separate
> > > > > with a semicolon:
> > > > > 	matching;settings
> > > >
> > > > Can you give an example?
> > > 
> > > Let's take port addition in OVS-DPDK as an example. It happens in 2
> > > steps:
> > > - port lookup (if port is already probed)
> > > - dev attachment (if lookup fails)
> > > 
> > > And also let's assume we need probe a ConnectX-3 port. Note that for
> > > ConnectX-3, there are 2 ports sharing the same PCI addr. Thus, PCI
> > > BDF is not enough. And let's assume we use another extra property
> > > "port".
> > > 
> > > If the proposal described in this patch is being used, the devarg
> > > would look like following:
> > > 
> > >     bus=pci,id=04:00.0/class=eth,port=0/driver=mlx4,mlx4_arg_A=val,...
> > > 
> > > Then "bus=pci,id=04:00.0/class=eth,port=0" will be used for lookup,
> > > It means we are looking for a port with PCI BDF == 04:00.0 AND
> > > port == 0 (the first port of the 2 ports).
> > > 
> > > Note that in my proposal the driver category is not intended for lookup.
> > > If any properties needed be looked in the driver category, they would
> > > probably need be elevated to the class category.
> > 
> > It is not my thought.
> > I think we should be able to use bus, class and driver properties for lookup.
> > We can imagine doing a lookup on a driver specific id, which is not
> > candidate to elevation to the class category.
> 
> This means having a new set of ops for drivers to implement (get / set
> on specific properties -- configuration items).

Just new ops to parse the string.
Then the driver is free to do whatever he wants internally.

> > > If port not found, then the whole string will be used for dev attachment.
> > > It means we are attaching a port with PCI BDF == 04.00.0 AND
> > > port == 0 (the 2nd port will not be attached).
> > > 
> > > 
> > > And here is how the devargs would look like if "matching;settings" is
> > > being used:
> > > 
> > >     bus=pci,id=04:00.0/class=eth,port=0;bus=pci,id=04:00.0/class=eth,port=0/driver=mlx4,mlx4_arg_A=val,...
> > > 
> > > The part before ";" will be used for lookup and the later part will be
> > > used for attachment. It should work. It just looks redundant.
> > 
> > It does not have to be redundant.
> > It can be:
> > 	bus=pci,id=04:00.0/class=eth,port=0;driver=mlx4,mlx4_arg1=settings1,...
> > 
> 
> Did you mean
> 
> > 	bus=pci,id=04:00.0/class=eth,port=0;class=driver,name=mlx4,mlx4_arg1=settings1,...

No :)
There are 3 categories: bus, class and driver.
class is for eth, crypto, event, etc.

> Here? Or is it that you "elevated" driver to be a property of the eth
> class, and then immediately chained with driver parameters without
> declaring the new driver class?

No I think you misunderstand.
I re-use the same syntax for matching and settings.
Overview is: bus/class/driver;bus/class/driver
where first part is for matching, and second part is for settings if any.
Another overview is: matching;settings

> > Another example, setting the MAC address:
> > 	bus=pci,id=04:00.0/class=eth,port=0;class=eth,mac=00:11:22:33:44:55
> 
> So, I guess this ";" syntax is meant for a user to provide once and for
> all a device string: perhaps on the command line, or programmatically.
> It would be used first for EAL init, then reused as-is (the entire
> string) for lookup / port matching afterward.
> 
> I think this is forcing the user to keep in mind a logic that should be
> abstracted away ("Here I am writing for init time, here I am writing for
> matching -- but I need to put it at the same place for 'reasons'").
> 
> I think mashing those two concepts together introduce complexity, and I
> think keeping them separate is user hostile as the devargs that was used
> for initializing a device cannot be re-used afterward for matching the
> device that resulted from this initialization string.
> 
> Drivers answers to a specific API (ethdev, cryptodev, ...), to create
> standardized objects in response to parameters that are given to them
> for init. I think matching properties should be restricted to higher
> classes (bus, eth/crypto), while the driver class should be left
> free-form and to the responsibility of the PMD itself (while having the
> proper libraries for helping parsing safely, thus driving developpers
> toward similar syntaxes, while not forcing them in those).
> 
> Match could be performed on bus / eth classes only, while init could
> use elements of the three classes. For simplicity, the same syntax rules
> could be enforced at all level, or for flexibility some leeway could be
> left on the most specific (driver).

I think it is more generic to allow bus/class/driver for matching
and for settings.
You seem to disagree about having driver-specific matching properties.
I argue it will address all future corner cases.
You say driver settings should be free form while I propose just
key=value syntax for them; it is a kind of free form.