Re: [nft PATCH v2 2/2] src: get rid of printf

[Florian Westphal <fw@strlen.de>]

Phil Sutter <phil@nwl.cc> wrote:
> On Thu, Sep 21, 2017 at 05:51:12PM +0200, Pablo Neira Ayuso wrote:
> > On Thu, Sep 21, 2017 at 05:43:09PM +0200, Florian Westphal wrote:
> > > Phil Sutter <phil@nwl.cc> wrote:
> > > > With the proposed implementation of nft_print(), this will be
> > > > problematic: nft_run_cmd_from_buffer() waits for the command to finish
> > > > before printing all the output at once. This obviously breaks monitor
> > > > which runs endlessly.
> > > > 
> > > > Maybe we should reference nft_print via a function pointer in
> > > > output_ctx? This would allow to configure a different nft_print
> > > > implementation which flushes the buffer to stdout immediately.
> > > > 
> > > > What do you think?
> > > 
> > > I wondered the same thing.
> > > 
> > > I have following issue: When using new typeof keyword then listing
> > > a set definition should output something like
> > > 
> > > set foo { type typeof(meta iifname) }
> > > 
> > > Because the kernel has no notion of 'meta iifname', we have to
> > > store this information in the kernel so we can read it back during
> > > delinearization.
> > > 
> > > One way to do this would be to store the 'meta iifname' string in the
> > > sets userdata.
> > > 
> > > For that, expr_print() would have to be able to print to a buffer
> > > (or we would need an expr_snprintf or something similar) to extract
> > > the convert struct *expr back to its original name.
> > > 
> > > The other solution would be to stash this in the expression
> > > during parsing but that seems weird as we need to be able to do such
> > > conversion anyway when printing the ruleset, so we merely need
> > > to make this accessible outside of plain printf() to stdout.
> > 
> > Probably the underlying problem is that the monitor code is that not
> > delinearing, ie. transforming from netlink to abstract syntax tree
> > (ast) before printing?
> 
> No, we are talking about different issues (but with a potential common
> solution):
> 
> Florian searches for a way to "convert" an expression from it's binary
> form (a struct expr) into a human readable string (here: "meta
> iifname"). We have this "converter" already in very limited form, namely
> expr_print(). Eric's patch changes that function to make use of
> nft_print(), which effectively allows to define a custom buffer to print
> into.
> 
> Florian, did I get this right?

Yes, thats right.

(Pablo, I do not NEED this function, but, as far is I understood
 you want nft to list

 set name { type typeof(ip saddr) }

back exactly like this, keeping the 'typeof'.  And AFAICS the best
way to do this is to store the human readable "ip saddr" as-is in the
userdata area and then just print that back.

If I misunderstood what you were saying, please clarify.

Ultimately something is needed to annotate in any case because in case
of 'meta iifname . ipv4_addr' we don't have anything to go on when deliarizing as
the string type lacks any length information and we don't have any rule
context from set definition to infer the needed key length information
of the interface name.