On Wed, Nov 13, 2024 at 12:08:01PM +0100, Pablo Neira Ayuso wrote:
> Hi Phil,
>
> On Tue, Nov 05, 2024 at 06:07:50PM +0100, Phil Sutter wrote:
> > On Tue, Nov 05, 2024 at 05:51:35PM +0100, Pablo Neira Ayuso wrote:
> > > Hi Phil,
> > >
> > > On Tue, Nov 05, 2024 at 02:35:16PM +0100, Phil Sutter wrote:
> > > > On Thu, Oct 31, 2024 at 11:04:11PM +0100, Pablo Neira Ayuso wrote:
> > > > > Update json parser to collapse {add,create} element commands to reduce
> > > > > memory consumption in the case of large sets defined by one element per
> > > > > command:
> > > > >
> > > > > {"nftables": [{"add": {"element": {"family": "ip", "table": "x", "name":
> > > > > "y", "elem": [{"set": ["1.1.0.0"]}]}}},...]}
> > > >
> > > > Thanks for the fix!
> > > >
> > > > > Add CTX_F_COLLAPSED flag to report that command has been collapsed.
> > > >
> > > > I had come up with a similar solution (but did not find time to submit
> > > > it last week). My solution to the "what to return" problem was to
> > > > introduce a 'static struct cmd cmd_nop' and return its address. Your
> > > > flag way is fine, too from my PoV.
> > >
> > > OK, I'm going to push it out then.
> > >
> > > > > This patch reduces memory consumption by ~32% this case.
> > > > >
> > > > > Fixes: 20f1c60ac8c8 ("src: collapse set element commands from parser")
> > > > > Reported-by: Eric Garver <eric@xxxxxxxxxxx>
> > > > > Signed-off-by: Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx>
> > > > > ---
> > > > > Side note: While profiling, I can still see lots json objects, this
> > > > > results in memory consumption that is 5 times than native
> > > > > representation. Error reporting is also lagging behind, it should be
> > > > > possible to add a json_t pointer to struct location to relate
> > > > > expressions and json objects.
> > > >
> > > > I can have a look at mem use if I find spare time (TM).
> > >
> > > I understand, that is always the issue.
> >
> > It's on my TODO at least, let's hope for the best.
>
> Thanks, we are already discussing this in a different thread.
>
> > > > We already record links between struct cmd and json_t objects for echo
> > > > mode (and only then). The problem with error reporting in my opinion is
> > > > the lack of location data in json_t. You might remember, I tried to
> > > > extend libjansson to our needs but my MR[1] is being ignored for more
> > > > than a year now. Should we just ship an extended copy in nftables?
> > >
> > > Do you still have the link with your proposal around? I don't find it
> > > in my notes anymore.
> >
> > Ah, prolly forgot to resolve that [1] above:
> >
> > https://github.com/akheron/jansson/pull/662
> >
> > > IIRC the rejection came from concerns about increasing memory usage
> > > for our specific usecase, that was an extra pointer to store location,
> > > correct?
> >
> > That and lack of interest in the feature in general. See the linked !461
> > for some feedback. The uncommented implementation in !662 hides
> > everything behind a decoder flag and avoids any memory overhead if not
> > enabled. The only remaining concern I can't address is: "we don't see
> > this as an important feature that should be included in Jansson."
>
> Your changeset for jansson look relative small, coding style look also
> correct, and you have addressed the memory consumption issue.
>
> It is a pity, I wonder what people do when json in this case? They
> have to scavenge in the json soup to find for non-syntactic errors as
> we have to do with it?
>
> Having said this, I like jansson implementation, it is neat.
>
> Caching a copy of libjansson in the nftables tree does not look like a
> way to go IMO.
>
> Is there a chance to poke them again / refresh your pull request?
I rebased the merge request and added a comment. Let's see if someone
reacts.
> Otherwise, we will have to search for alternative path to improve the
> existing situation.
If they don't accept, I see no way other than forking their lib or
reimplementing functionality in another way.
Cheers, Phil
