IB/hfi1: Refine user process affinity algorithm

When performing process affinity recommendations for MPI ranks, the current
algorithm doesn't take into account multiple HFI units. Also, real
cores and HT cores are not distinguished from one another. Therefore,
all HT cores are recommended to be assigned first within the local NUMA
node before recommending the assignments of cores in other NUMA nodes.
It's ideal to assign all real cores across all NUMA nodes first, then all
HT 1 cores, then all HT 2 cores, and so on to balance CPU workload. CPU
cores in other NUMA nodes could be running interrupt handlers, and this is
not taken into account.

To balance the CPU workload for user processes, the following
recommendation algorithm is used:

 For each user process that is opening a context on HFI Y:
  a) If all cores are assigned to user processes, start assignments all
	 over from the first core
  b) Assign real cores first, then HT cores (First set of HT cores on
	 all physical cores, then second set of HT cores, and, so on) in the
	 following order:

	 1. Same NUMA node as HFI Y and not running an IRQ handler
	 2. Same NUMA node as HFI Y and running an IRQ handler
	 3. Different NUMA node to HFI Y and not running an IRQ handler
	 4. Different NUMA node to HFI Y and running an IRQ handler
  c) Mark core as assigned in the global affinity structure. As user
	 processes are done, remove core assignments from global affinity
	 structure.

This implementation allows an arbitrary number of HT cores and provides
support for multiple HFIs.

This is being included in the kernel rather than user space due to the
fact that user space has no way of knowing the CPU recommendations for
contexts running as part of other jobs.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

1 files changed, 10 insertions, 5 deletions

diff --git a/drivers/infiniband/hw/hfi1/file_ops.c b/drivers/infiniband/hw/hfi1/file_ops.c
index 2f097d942f9c..d7c07bc7bd14 100644
--- a/drivers/infiniband/hw/hfi1/file_ops.c
+++ b/drivers/infiniband/hw/hfi1/file_ops.c
@@ -715,7 +715,7 @@ static int hfi1_file_close(struct inode *inode, struct file *fp)
 	hfi1_user_sdma_free_queues(fdata);
 
 	/* release the cpu */
-	hfi1_put_proc_affinity(dd, fdata->rec_cpu_num);
+	hfi1_put_proc_affinity(fdata->rec_cpu_num);
 
 	/*
 	 * Clear any left over, unhandled events so the next process that
@@ -815,9 +815,10 @@ static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
 		ret = find_shared_ctxt(fp, uinfo);
 		if (ret < 0)
 			goto done_unlock;
-		if (ret)
-			fd->rec_cpu_num = hfi1_get_proc_affinity(
-				fd->uctxt->dd, fd->uctxt->numa_id);
+		if (ret) {
+			fd->rec_cpu_num =
+				hfi1_get_proc_affinity(fd->uctxt->numa_id);
+		}
 	}
 
 	/*
@@ -929,7 +930,11 @@ static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
 	if (ctxt == dd->num_rcv_contexts)
 		return -EBUSY;
 
-	fd->rec_cpu_num = hfi1_get_proc_affinity(dd, -1);
+	/*
+	 * If we don't have a NUMA node requested, preference is towards
+	 * device NUMA node.
+	 */
+	fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
 	if (fd->rec_cpu_num != -1)
 		numa = cpu_to_node(fd->rec_cpu_num);
 	else