Re: [PATCH] genirq/affinity: report extra vectors on uneven nodes

["Derrick, Jonathan" <jonathan.derrick@intel.com>]

[-- Attachment #1: Type: text/plain, Size: 3268 bytes --]

On Thu, 2019-08-08 at 10:32 -0600, Keith Busch wrote:
> On Thu, Aug 08, 2019 at 09:04:28AM +0200, Thomas Gleixner wrote:
> > On Wed, 7 Aug 2019, Jon Derrick wrote:
> > > The current irq spreading algorithm spreads vectors amongst cpus evenly
> > > per node. If a node has more cpus than another node, the extra vectors
> > > being spread may not be reported back to the caller.
> > > 
> > > This is most apparent with the NVMe driver and nr_cpus < vectors, where
> > > the underreporting results in the caller's WARN being triggered:
> > > 
> > > irq_build_affinity_masks()
> > > ...
> > > 	if (nr_present < numvecs)
> > > 		WARN_ON(nr_present + nr_others < numvecs);
> > > 
> > > Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
> > > ---
> > >  kernel/irq/affinity.c | 7 +++++--
> > >  1 file changed, 5 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/kernel/irq/affinity.c b/kernel/irq/affinity.c
> > > index 4352b08ae48d..9beafb8c7e92 100644
> > > --- a/kernel/irq/affinity.c
> > > +++ b/kernel/irq/affinity.c
> > > @@ -127,7 +127,8 @@ static int __irq_build_affinity_masks(unsigned int startvec,
> > >  	}
> > >  
> > >  	for_each_node_mask(n, nodemsk) {
> > > -		unsigned int ncpus, v, vecs_to_assign, vecs_per_node;
> > > +		unsigned int ncpus, v, vecs_to_assign, total_vecs_to_assign,
> > > +			vecs_per_node;
> > >  
> > >  		/* Spread the vectors per node */
> > >  		vecs_per_node = (numvecs - (curvec - firstvec)) / nodes;
> > > @@ -141,14 +142,16 @@ static int __irq_build_affinity_masks(unsigned int startvec,
> > >  
> > >  		/* Account for rounding errors */
> > >  		extra_vecs = ncpus - vecs_to_assign * (ncpus / vecs_to_assign);
> > > +		total_vecs_to_assign = vecs_to_assign + extra_vecs;
> > >  
> > > -		for (v = 0; curvec < last_affv && v < vecs_to_assign;
> > > +		for (v = 0; curvec < last_affv && v < total_vecs_to_assign;
> > >  		     curvec++, v++) {
> > >  			cpus_per_vec = ncpus / vecs_to_assign;
> > >  
> > >  			/* Account for extra vectors to compensate rounding errors */
> > >  			if (extra_vecs) {
> > >  				cpus_per_vec++;
> > > +				v++;
> > >  				--extra_vecs;
> > >  			}
> > >  			irq_spread_init_one(&masks[curvec].mask, nmsk,
> > > -- 
> 
> This looks like it will break the spread to non-present CPUs since
> it's not accurately reporting how many vectors were assigned for the
> present spread.
> 
> I think the real problem is the spread's vecs_per_node doesn't account
> which nodes contribute more CPUs than others. For example:
> 
>   Node 0 has 32 CPUs
>   Node 1 has 8 CPUs
>   Assign 32 vectors
> 
> The current algorithm assigns 16 vectors to node 0 because vecs_per_node
> is calculated as 32 vectors / 2 nodes on the first iteration. The
> subsequent iteration for node 1 gets 8 vectors because it has only 8
> CPUs, leaving 8 vectors unassigned.
> 
> A more fair spread would give node 0 the remaining 8 vectors. This
> optimization, however, is a bit more complex than the current algorithm,
> which is probably why it wasn't done, so I think the warning should just
> be removed.

It does get a bit complex for the rare scenario in this case
Maybe just an informational warning rather than a stackdumping warning

[-- Attachment #2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 3278 bytes --]