> From: Long Li
> Sent: Wednesday, January 31, 2018 12:23 PM
> To: Michael Kelley (EOSG) <Michael.H.Kelley@xxxxxxxxxxxxx>; KY Srinivasan
> <kys@xxxxxxxxxxxxx>; Stephen Hemminger <sthemmin@xxxxxxxxxxxxx>;
> martin.petersen@xxxxxxxxxx; devel@xxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> linux-scsi@xxxxxxxxxxxxxxx; James E . J . Bottomley <jejb@xxxxxxxxxxxxxxxxxx>
> Subject: RE: [PATCH 1/1] scsi: storvsc: Spread interrupts when picking a channel for I/O
> requests
>
> > Subject: RE: [PATCH 1/1] scsi: storvsc: Spread interrupts when picking a
> > channel for I/O requests
> >
> > Updated/corrected two email addresses ...
> >
> > > -----Original Message-----
> > > From: Michael Kelley (EOSG)
> > > Sent: Wednesday, January 24, 2018 2:14 PM
> > > To: KY Srinivasan <kys@xxxxxxxxxxxxx>; Stephen Hemminger
> > > <sthemmin@xxxxxxxxxxxxx>; martin.petersen@xxxxxxxxxx;
> > > longi@xxxxxxxxxxxxx; JBottomley@xxxxxxxx;
> > > devel@xxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > > linux-scsi@xxxxxxxxxxxxxxx
> > > Cc: Michael Kelley (EOSG) <Michael.H.Kelley@xxxxxxxxxxxxx>
> > > Subject: [PATCH 1/1] scsi: storvsc: Spread interrupts when picking a
> > > channel for I/O requests
> > >
> > > Update the algorithm in storvsc_do_io to look for a channel starting
> > > with the current CPU + 1 and wrap around (within the current NUMA
> > > node). This spreads VMbus interrupts more evenly across CPUs. Previous
> > > code always started with first CPU in the current NUMA node, skewing
> > > the interrupt load to that CPU.
> > >
> > > Signed-off-by: Michael Kelley <mikelley@xxxxxxxxxxxxx>
> > > ---
> > > drivers/scsi/storvsc_drv.c | 3 ++-
> > > 1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
> > > index e07907d..f3264c4 100644
> > > --- a/drivers/scsi/storvsc_drv.c
> > > +++ b/drivers/scsi/storvsc_drv.c
> > > @@ -1310,7 +1310,8 @@ static int storvsc_do_io(struct hv_device *device,
> > > */
> > > cpumask_and(&alloced_mask, &stor_device-
> > >alloced_cpus,
> > >
> > cpumask_of_node(cpu_to_node(q_num)));
> > > - for_each_cpu(tgt_cpu, &alloced_mask) {
> > > + for_each_cpu_wrap(tgt_cpu, &alloced_mask,
> > > + outgoing_channel->target_cpu + 1) {
>
> Does it work when target_cpu is the last CPU on the system?
>
> Otherwise, looking good.
Yes, it works. for_each_cpu_wrap() correctly wraps in the case where
the 3rd parameter ('start') is one past the end of the mask. Arguably,
we shouldn't rely on that, and should do the wrap to 0 before calling
for_each_cpu_wrap().
>
> > > if (tgt_cpu != outgoing_channel->target_cpu)
> > {
> > > outgoing_channel =
> > > stor_device->stor_chns[tgt_cpu];
> > > --
> > > 1.8.3.1
