Re: [net-next v5 PATCH] page_pool: handle page recycle for NUMA_NO_NODE condition

[Ilias Apalodimas <ilias.apalodimas@linaro.org>]

Hi Jesper,

Looking at the overall path again, i still need we need to reconsider 
pool->p.nid semantics.

As i said i like the patch and the whole functionality and code seems fine,
but here's the current situation.
If a user sets pool->p.nid == NUMA_NO_NODE and wants to use
page_pool_update_nid() the whole behavior feels a liitle odd.
page_pool_update_nid() first check will always be true since .nid =
NUMA_NO_NODE). Then we'll update this to a real nid. So we end up overwriting
what the user initially coded in. 
This is close to what i proposed in the previous mails on this thread. Always
store a real nid even if the user explicitly requests NUMA_NO_NODE.

So  semantics is still a problem. I'll stick to what we initially suggested.
1. We either *always* store a real nid
or 
2. If NUMA_NO_NODE is present ignore every other check and recycle the memory
blindly. 

Regards
/Ilias

On Fri, Dec 20, 2019 at 06:06:49PM +0200, Ilias Apalodimas wrote:
> On Fri, Dec 20, 2019 at 04:22:54PM +0100, Jesper Dangaard Brouer wrote:
> > On Fri, 20 Dec 2019 12:49:37 +0200
> > Ilias Apalodimas <ilias.apalodimas@linaro.org> wrote:
> > 
> > > On Fri, Dec 20, 2019 at 11:41:16AM +0100, Jesper Dangaard Brouer wrote:
> > > > On Fri, 20 Dec 2019 12:23:14 +0200
> > > > Ilias Apalodimas <ilias.apalodimas@linaro.org> wrote:
> > > >   
> > > > > Hi Jesper, 
> > > > > 
> > > > > I like the overall approach since this moves the check out of  the hotpath. 
> > > > > @Saeed, since i got no hardware to test this on, would it be possible to check
> > > > > that it still works fine for mlx5?
> > > > > 
> > > > > [...]  
> > > > > > +	struct ptr_ring *r = &pool->ring;
> > > > > > +	struct page *page;
> > > > > > +	int pref_nid; /* preferred NUMA node */
> > > > > > +
> > > > > > +	/* Quicker fallback, avoid locks when ring is empty */
> > > > > > +	if (__ptr_ring_empty(r))
> > > > > > +		return NULL;
> > > > > > +
> > > > > > +	/* Softirq guarantee CPU and thus NUMA node is stable. This,
> > > > > > +	 * assumes CPU refilling driver RX-ring will also run RX-NAPI.
> > > > > > +	 */
> > > > > > +	pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;    
> > > > > 
> > > > > One of the use cases for this is that during the allocation we are not
> > > > > guaranteed to pick up the correct NUMA node. 
> > > > > This will get automatically fixed once the driver starts recycling packets. 
> > > > > 
> > > > > I don't feel strongly about this, since i don't usually like hiding value
> > > > > changes from the user but, would it make sense to move this into 
> > > > > __page_pool_alloc_pages_slow() and change the pool->p.nid?
> > > > > 
> > > > > Since alloc_pages_node() will replace NUMA_NO_NODE with numa_mem_id()
> > > > > regardless, why not store the actual node in our page pool information?
> > > > > You can then skip this and check pool->p.nid == numa_mem_id(), regardless of
> > > > > what's configured.   
> > > > 
> > > > This single code line helps support that drivers can control the nid
> > > > themselves.  This is a feature that is only used my mlx5 AFAIK.
> > > > 
> > > > I do think that is useful to allow the driver to "control" the nid, as
> > > > pinning/preferring the pages to come from the NUMA node that matches
> > > > the PCI-e controller hardware is installed in does have benefits.  
> > > 
> > > Sure you can keep the if statement as-is, it won't break anything. 
> > > Would we want to store the actual numa id in pool->p.nid if the user
> > > selects 'NUMA_NO_NODE'?
> >  
> > No. pool->p.nid should stay as NUMA_NO_NODE, because that makes it
> > dynamic.  If someone moves an RX IRQ to another CPU on another NUMA
> > node, then this 'NUMA_NO_NODE' setting makes pages transitioned
> > automatically.
> Ok this assumed that drivers were going to use page_pool_nid_changed(), but with
> the current code we don't have to force them to do that. Let's keep this as-is.
> 
> I'll be running a few more tests  and wait in case Saeed gets a chance to test
> it and send my reviewed-by
> 
> Cheers
> /Ilias
> > 
> > -- 
> > Best regards,
> >   Jesper Dangaard Brouer
> >   MSc.CS, Principal Kernel Engineer at Red Hat
> >   LinkedIn: http://www.linkedin.com/in/brouer
> >