[dpdk-dev] [PATCH v2 04/18] eal: add lightweight kvarg parsing utility

[Neil Horman]

On Fri, Mar 23, 2018 at 10:31:22AM +0100, Gaëtan Rivet wrote:
> On Thu, Mar 22, 2018 at 08:53:49PM -0400, Neil Horman wrote:
> > On Thu, Mar 22, 2018 at 05:27:51PM +0100, Gaëtan Rivet wrote:
> > > On Thu, Mar 22, 2018 at 10:10:37AM -0400, Neil Horman wrote:
> > > > On Wed, Mar 21, 2018 at 05:32:24PM +0000, Wiles, Keith wrote:
> > > > > 
> > > > > 
> > > > > > On Mar 21, 2018, at 12:15 PM, Gaetan Rivet <gaetan.rivet at 6wind.com> wrote:
> > > > > > 
> > > > > > This library offers a quick way to parse parameters passed with a
> > > > > > key=value syntax.
> > > > > > 
> > > > > > A single function is needed and finds the relevant element within the
> > > > > > text. No dynamic allocation is performed. It is possible to chain the
> > > > > > parsing of each pairs for quickly scanning a list.
> > > > > > 
> > > > > > This utility is private to the EAL and should allow avoiding having to
> > > > > > move around the more complete librte_kvargs.
> > > > > 
> > > > > What is the big advantage with this code and the librte_kvargs code. Is it just no allocation, rte_kvargs needs to be build before parts of EAL or what?
> > > > > 
> > > > > My concern is we have now two flavors one in EAL and one in librte_kvargs, would it not be more reasonable to improve rte_kvargs to remove your objections? I am all for fast, better, stronger code :-)
> > > > > 
> > > > +1, this really doesn't make much sense to me.  Two parsing routines seems like
> > > > its just asking for us to have to fix parsing bugs in two places.  If allocation
> > > > is a concern, I don't see why you can't just change the malloc in
> > > > rte_kvargs_parse to an automatic allocation on the stack, or a preallocation set
> > > > of kvargs that can be shared from init time.
> > > 
> > > I think the existing allocation scheme is fine for other usages (in
> > > drivers and so on). Not for what I wanted to do.
> > > 
> > Ok, but thats an adressable issue.  you can bifurcate the parse function to an
> > internal function that accepts any preallocated kvargs struct, and export two
> > wrapper functions, one which allocates the struct from the heap, another which
> > allocated automatically on the stack.
> > 
> 
> Sure, everything is possible.
> 
Ok.

> > > >                                               librte_kvargs isn't necessecarily
> > > > the best parsing library ever, but its not bad, and it just seems wrong to go
> > > > re-inventing the wheel.
> > > > 
> > > 
> > > It serves a different purpose than the one I'm pursuing.
> > > 
> > > This helper is lightweight and private. If I wanted to integrate my
> > > needs with librte_kvargs, I would be adding new functionalities, making
> > > it more complex, and for a use-case that is useless for the vast
> > > majority of users of the lib.
> > > 
> > Ok, to that end:
> > 
> > 1) Privacy is not an issue (at least from my understanding of what your doing).
> > If we start with the assumption that librte_kvargs is capable of satisfying your
> > needs (even if its not done in an optimal way), the fact that your version of
> > the function is internal to the library doesn't seem overly relevant, unless
> > theres something critical to that privacy that I'm missing.
> > 
> 
> Privacy is only a point I brought up to say that the impact of this
> function is minimal. People looking to parse their kvargs should not
> have any ambiguity regarding how they should do so. Only librte_kvargs
> is available.
> 
Ok, would you also council others developing dpdk apps to write their own
parsing routines when what they needed was trivial for the existing library?
You are people too :)

> > 2) Lightweight function  seems like something that can be integrated with
> > librte_kvargs.  Looking at it, what may I ask in librte_kvargs is insufficiently
> > non-performant for your needs, specifically?  We talked about the heap
> > allocation above, is there something else? The string duplication perhaps?
> > 
> > 
> 
> Mostly the way to use it.
> The filter strings are
> bus=value,.../class=value,...
> 
> where either bus= list or class= list can be omitted, but at least one
> must appear.
> 
Ok, so whats the problem with using librte_kvargs for that?  Is it that the list
that acts as the value to the key isn't parsed out into its own set of tokens?
That seems entirely addressable.

> I want to read a single kvarg. I do not want to parse the whole string.
> the '/' signifies the end of the current layer.
> 
This makes it seem like librte_kvargs can handle this as a trivial case of its
functionality.

> librte_kvargs does not care about those points. I cannot ask it to only
> read either bus or class, as it would then throw an error for all the
> other keys (which the EAL has necessarily no knowledge of).
> 
But you can ask it to read both, and within your libraries logic make the
determination as to the validitiy of receiving both.  Alternatively you can
modify the valid_keys check in kvargs to be a regex that matches on either bus
or class, or accept an ignore parameter for keys that may appear but should be
ignored in the light of other keys.  Theres lots of options here.

> So I would need to:
> 
>   * Add a custom storage scheme
>   * Add a custom parsing mode stopping at the first kvarg
>   * Add an edge-case to ignore the '/', so as not to throw off the rest
>     of the parsing (least it be considered part of the previous kvarg
>     value field).
> 
> Seeing this, does adding those really specifics functionality help
> librte_kvargs to be more useful and usable? I do not think so.
> 
I think you're overcomplicating this.
How about enhancing librte_kvargs to make parsing configurable
such that invalid keys get ignored rather than generate errors?

> It would only serve to disrupt the library for a marginal use-case, with
> the introduction of edge-cases that will blur the specs of the lib's
> API, making it harder to avoid subtle bugs.
> 
What do you mean "disrupt the library"?  What is its purpose of a library if not
to do the jobs we want it to?  If everyone created their own routine to do
something that a library could do with some modifications, we'd be left with
1000 versions of the same routine.  If the existing library does 99% of what you
want it to, lets ennumerate what the missing %1 is and make the change, not
throw the baby out with the bathwater.

> Only way to do so sanely would be to add rte_parse_kv as part of
> librte_kvargs, as is. But then the whole thing does not make sense IMO:
> no one would care to use it, the maintainance effort is the same, the
> likelyhood of bugs as well (but in the process we would disrupt the
> distribution of librte_kvargs by moving it within the EAL).
> 
> I see no benefit to either solution.
> 
Again, thats an overcomplication.  As I read it, you have a need to interrogate
a key/value string, whos contents may contain invalid keys (for your parsing
purposes), and whos values may themselves be lists, correct?  If so, I don't see
the problem in enhancing libkvargs to:

1) Allow for the parsing routine to ignore invalid keys (or even ignore specific
invalid keys, and trigger on any unknown keys)

2) Allows for the subparsing of lists into their own set of tokens.

> > > If that's really an issue, I'm better off simply removing rte_parse_kv
> > > and writing the parsing by hand within my function. This would be ugly
> > > and tedious, but less than moving librte_kvargs within EAL and changing
> > > it to my needs.
> > I don't think thats necessecary, I just think if you can ennumerate the items
> > that are non-performant for your needs we can make some changes to librte_kvargs
> > to optimize around them, or offer parsing options to avoid them, and in the
> > process avoid some code duplication
> > 
> 
> I think it makes sense to have specialized functions for specialized
> use-cases, and forcing the code to be generic and work with all of them
> will make it more complicated.
> 
This isn't specialized, its trivial.  Its just a trivial case that libkvargs
isn't built to handle at the moment.  Lets get it there.

> The genericity would only be worth it if people actually needed to parse
> the device strings the same way I do. No one has any business doing so.
> This genericity adds complexity and issues, without even being useful in
> the first place.
> 
I really think you're trying to take a short cut here where none is needed, and
I'm sorry, but I can't support that.

Neil

> -- 
> Gaëtan Rivet
> 6WIND
>