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

[Jesper Dangaard Brouer]

On Tue, 2 Jul 2019 21:58:40 +0300
Ivan Khoronzhuk <ivan.khoronzhuk@xxxxxxxxxx> 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'm trying to limit the scope of your changes, for your special case,
because I'm afraid this more common solution is going to limit our
options, painting ourselves into a corner.

E.g. for correct lifetime handling, I think we actually need to do a
dev_hold() on the net_device. (Changes in f71fec47c2 might not be
enough, but I first need to dig into the details and ask Hellwig about
some details).  Adding that after your patch is more complicated (if
even doable).

E.g. doing dev_hold() on the net_device, can also turn into a
performance advantage, when/if page_pool is extended to also "travel"
into SKBs. (Allowing to elide such dev_hold() calls in netstack).

I also worry about the possible performance impact these changes will
have down the road.  (For the RX/alloc side it should be clear by now
that we gain a lot of performance with the single RX-queue binding and
napi protection).  On the return/free side performance *need* to be
improved (it doesn't scale).  I'm basically looking at different ways
to bulk return pages into the ptr_ring, which requires changes in
page_pool and likely in xdp_allocator structure.  Which your changes
are complicating.

This special use-case, seems confined to your driver. And Ilias told me
that XDP is not really a performance benefit for this driver as the HW
PPS-limit is hit before the XDP and netstack limit.  I ask, does it
make sense to add XDP to this driver, if it complicates the code for
everybody else?

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer