]> git.baikalelectronics.ru Git - kernel.git/commit
IB/hfi1: Reserve and collapse CPU cores for contexts
authorSebastian Sanchez <sebastian.sanchez@intel.com>
Mon, 25 Jul 2016 14:54:48 +0000 (07:54 -0700)
committerDoug Ledford <dledford@redhat.com>
Tue, 2 Aug 2016 19:47:07 +0000 (15:47 -0400)
commit07b02e7ed3be6bb3ab0a0c8c34529cbc266e136c
tree3a886c96fe7e2c3ab51840671bf3b7981944ac07
parent83326550a313e648be0e754292feaf2bee99c173
IB/hfi1: Reserve and collapse CPU cores for contexts

Kernel receive queues oversubscribe CPU cores on multi-HFI systems.
To prevent this, the kernel receive queues are separated onto
different cores, and the SDMA engine interrupts are constrained to
a lesser number of cores.

hfi1s_on_numa_node*krcvqs is the number of CPU cores that are
reserved for kernel receive queues for all HFIs. Each HFI initializes
its kernel receive queues to one of the reserved CPU cores. If there
ends up being 0 CPU cores leftover for SDMA engines, use the same
CPU cores as receive contexts.

In addition, general and control contexts are assigned to their own
CPU core, however, both types of contexts tend to have low traffic.
To save CPU cores, collapse general and control contexts to one CPU
core for all HFI units. This change prevents SDMA engine interrupts
from wrapping around general contexts.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/hfi1/affinity.c
drivers/infiniband/hw/hfi1/affinity.h
drivers/infiniband/hw/hfi1/hfi.h
drivers/infiniband/hw/hfi1/init.c