On 2/12/21 9:21 AM, Jason Gunthorpe wrote: > On Fri, Feb 05, 2021 at 09:14:28AM +0100, Nicolas Morey-Chaisemartin wrote: >> The current code computes a number of channels per SRP target and spreads >> them equally across all online NUMA nodes. >> Each channel is then assigned a CPU within this node. >> >> In the case of unbalanced, or even unpopulated nodes, some channels >> do not get a CPU associated and thus do not get connected. >> This causes the SRP connection to fail. >> >> This patch solves the issue by rewriting channel computation and allocation: >> - Drop channel to node/CPU association as it had >> no real effect on locality but added unnecessary complexity. >> - Tweak the number of channels allocated to reduce CPU contention when possible: >> - Up to one channel per CPU (instead of up to 4 by node) >> - At least 4 channels per node, unless ch_count module parameter is used. >> >> Signed-off-by: Nicolas Morey-Chaisemartin <nmoreychaisemartin@xxxxxxxx> >> --- >> drivers/infiniband/ulp/srp/ib_srp.c | 110 ++++++++++++---------------- >> 1 file changed, 45 insertions(+), 65 deletions(-) > > Bart? Oops, this patch had escaped from my attention. Reviewed-by: Bart Van Assche <bvanassche@xxxxxxx>
