Re: [PATCH] net: core: page_pool: add user refcnt and reintroduce page_pool_destroy

[Ivan Khoronzhuk]

On Tue, Jul 02, 2019 at 09:58:40PM +0300, Ivan Khoronzhuk wrote:
> On Tue, Jul 02, 2019 at 08:29:07PM +0200, Jesper Dangaard Brouer wrote:
> > On Tue, 2 Jul 2019 18:21:13 +0300
> > Ivan Khoronzhuk <ivan.khoronzhuk@xxxxxxxxxx> wrote:
> >
> > > On Tue, Jul 02, 2019 at 05:10:29PM +0200, Jesper Dangaard Brouer wrote:
> > > > On Tue, 2 Jul 2019 17:56:13 +0300
> > > > Ivan Khoronzhuk <ivan.khoronzhuk@xxxxxxxxxx> wrote:
> > > >
> > > > > On Tue, Jul 02, 2019 at 04:52:30PM +0200, Jesper Dangaard Brouer wrote:
> > > > > >On Tue, 2 Jul 2019 17:44:27 +0300
> > > > > >Ivan Khoronzhuk <ivan.khoronzhuk@xxxxxxxxxx> wrote:
> > > > > >
> > > > > >> On Tue, Jul 02, 2019 at 04:31:39PM +0200, Jesper Dangaard Brouer wrote:
> > > > > >> >From: Ivan Khoronzhuk <ivan.khoronzhuk@xxxxxxxxxx>
> > > > > >> >
> > > > > >> >Jesper recently removed page_pool_destroy() (from driver invocation) and
> > > > > >> >moved shutdown and free of page_pool into xdp_rxq_info_unreg(), in-order to
> > > > > >> >handle in-flight packets/pages. This created an asymmetry in drivers
> > > > > >> >create/destroy pairs.
> > > > > >> >
> > > > > >> >This patch add page_pool user refcnt and reintroduce page_pool_destroy.
> > > > > >> >This serves two purposes, (1) simplify drivers error handling as driver now
> > > > > >> >drivers always calls page_pool_destroy() and don't need to track if
> > > > > >> >xdp_rxq_info_reg_mem_model() was unsuccessful. (2) allow special cases
> > > > > >> >where a single RX-queue (with a single page_pool) provides packets for two
> > > > > >> >net_device'es, and thus needs to register the same page_pool twice with two
> > > > > >> >xdp_rxq_info structures.
> > > > > >>
> > > > > >> As I tend to use xdp level patch there is no more reason to mention (2) case
> > > > > >> here. XDP patch serves it better and can prevent not only obj deletion but also
> > > > > >> pool flush, so, this one patch I could better leave only for (1) case.
> > > > > >
> > > > > >I don't understand what you are saying.
> > > > > >
> > > > > >Do you approve this patch, or do you reject this patch?
> > > > > >
> > > > > It's not reject, it's proposition to use both, XDP and page pool patches,
> > > > > each having its goal.
> > > >
> > > > Just to be clear, if you want this patch to get accepted you have to
> > > > reply with your Signed-off-by (as I wrote).
> > > >
> > > > Maybe we should discuss it in another thread, about why you want two
> > > > solutions to the same problem.
> > >
> > > If it solves same problem I propose to reject this one and use this:
> > > https://lkml.org/lkml/2019/7/2/651
> >
> > No, I propose using this one, and rejecting the other one.
>
> There is at least several arguments against this one (related (2) purpose)
>
> It allows:
> - avoid changes to page_pool/mlx5/netsec
> - save not only allocator obj but allocator "page/buffer flush"
> - buffer flush can be present not only in page_pool but for other allocators
> that can behave differently and not so simple solution.
> - to not limit cpsw/(potentially others) to use "page_pool" allocator only
> ....
>
> This patch better leave also, as it simplifies error path for page_pool and
> have more error prone usage comparing with existent one.
>
> Please, don't limit cpsw and potentially other drivers to use only
> page_pool it can be zca or etc... I don't won't to modify each allocator.
> I propose to add both as by fact they solve different problems with common
> solution.

I can pick up this one but remove description related to (2) and add
appropriate modifications to cpsw.

--
Regards,
Ivan Khoronzhuk