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

[Ilias Apalodimas]

Hi Jesper,

On Mon, Dec 23, 2019 at 05:52:57PM +0100, Jesper Dangaard Brouer wrote:
> On Mon, 23 Dec 2019 09:57:00 +0200
> Ilias Apalodimas <ilias.apalodimas@xxxxxxxxxx> wrote:
> 
> > 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.
> 
> As soon as driver uses page_pool_update_nid() than means they want to
> control the NUMA placement explicitly.  As soon as that happens, it is
> the drivers responsibility and choice, and page_pool API must respect
> that (and not automatically change that behind drivers back).
> 
> 
> > 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. 
> > 
> 
> Hmm... I actually disagree with both 1 and 2.
> 
> My semantics proposal:
> If driver configures page_pool with NUMA_NO_NODE, then page_pool tried
> to help get the best default performance. (Which according to
> performance measurements is to have RX-pages belong to the NUMA node
> RX-processing runs on).
> 
> The reason I want this behavior is that during driver init/boot, it can
> easily happen that a driver allocates RX-pages from wrong NUMA node.
> This will cause a performance slowdown, that normally doesn't happen,
> because without a cache (like page_pool) RX-pages would fairly quickly
> transition over to the RX NUMA node (instead we keep recycling these,
> in your case #2, where you suggest recycle blindly in case of
> NUMA_NO_NODE). IMHO page_pool should hide this border-line case from
> driver developers.

Yea #2 has different semantics than the one you propose. So if he chooses
NUMA_NO_NODE, i'd expect the machines(s) the driver sits on, are not NUMA-aware.
Think specific SoC's, i'd never expect PCI cards to use that.
As i said i don't feel strongly about this anyway, it's just another case i had
under consideration but i like what you propose more. I'll try to add 
documentation on page_pool API and describe the semantics you have in mind.


Thanks
/Ilias

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