From: Moni Shoua <monis@xxxxxxxxxxxx>
Page fault events are processed in a workqueue context. Since each QP
can have up to two concurrent unrelated page-faults, one for requester
and one for responder, page-fault handling can be done in parallel.
Achieve this by changing the workqueue to be multi-threaded.
The number of threads is the same as the number of command interface
channels to avoid command interface bottlenecks.
In addition to multi-threads, change the workqueue flags to give it high
priority.
Stress benchmark shows that before this change 85% of page faults were
waiting in queue 8 seconds or more while after the change 98% of page
faults were waiting in queue 64 milliseconds or less. The number of threads
was chosen as the number of channels to the command interface.
Fixes: d9aaed838765 ("{net,IB}/mlx5: Refactor page fault handling")
Signed-off-by: Moni Shoua <monis@xxxxxxxxxxxx>
Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxxxx>
---
drivers/net/ethernet/mellanox/mlx5/core/eq.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eq.c b/drivers/net/ethernet/mellanox/mlx5/core/eq.c
index c1e1a16a9b07..aeab0c4f60f4 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eq.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eq.c
@@ -356,8 +356,9 @@ static int init_pf_ctx(struct mlx5_eq_pagefault *pf_ctx, const char *name)
spin_lock_init(&pf_ctx->lock);
INIT_WORK(&pf_ctx->work, eq_pf_action);
- pf_ctx->wq = alloc_ordered_workqueue(name,
- WQ_MEM_RECLAIM);
+ pf_ctx->wq = alloc_workqueue(name,
+ WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
+ MLX5_NUM_CMD_EQE);
if (!pf_ctx->wq)
return -ENOMEM;
--
2.19.1
