RDS: IB: Add vector spreading for cqs

Based on available device vectors, allocate cqs accordingly to
get better spread of completion vectors which helps performace
great deal..

Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
This commit is contained in:
Santosh Shilimkar 2016-07-04 16:16:36 -07:00
parent 09b2b8f528
commit be2f76eacc
3 changed files with 53 additions and 3 deletions

View File

@ -111,6 +111,9 @@ static void rds_ib_dev_free(struct work_struct *work)
kfree(i_ipaddr);
}
if (rds_ibdev->vector_load)
kfree(rds_ibdev->vector_load);
kfree(rds_ibdev);
}
@ -159,6 +162,14 @@ static void rds_ib_add_one(struct ib_device *device)
rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
rds_ibdev->vector_load = kzalloc(sizeof(int) * device->num_comp_vectors,
GFP_KERNEL);
if (!rds_ibdev->vector_load) {
pr_err("RDS/IB: %s failed to allocate vector memory\n",
__func__);
goto put_dev;
}
rds_ibdev->dev = device;
rds_ibdev->pd = ib_alloc_pd(device, 0);
if (IS_ERR(rds_ibdev->pd)) {

View File

@ -185,6 +185,10 @@ struct rds_ib_connection {
/* Endpoint role in connection */
bool i_active_side;
/* Send/Recv vectors */
int i_scq_vector;
int i_rcq_vector;
};
/* This assumes that atomic_t is at least 32 bits */
@ -227,6 +231,7 @@ struct rds_ib_device {
spinlock_t spinlock; /* protect the above */
atomic_t refcount;
struct work_struct free_work;
int *vector_load;
};
#define ibdev_to_node(ibdev) dev_to_node(ibdev->dma_device)

View File

@ -358,6 +358,28 @@ static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
tasklet_schedule(&ic->i_send_tasklet);
}
static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
{
int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
int index = rds_ibdev->dev->num_comp_vectors - 1;
int i;
for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
if (rds_ibdev->vector_load[i] < min) {
index = i;
min = rds_ibdev->vector_load[i];
}
}
rds_ibdev->vector_load[index]++;
return index;
}
static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
{
rds_ibdev->vector_load[index]--;
}
/*
* This needs to be very careful to not leave IS_ERR pointers around for
* cleanup to trip over.
@ -399,25 +421,30 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
/* Protection domain and memory range */
ic->i_pd = rds_ibdev->pd;
ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
cq_attr.comp_vector = ic->i_scq_vector;
ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
rds_ib_cq_event_handler, conn,
&cq_attr);
if (IS_ERR(ic->i_send_cq)) {
ret = PTR_ERR(ic->i_send_cq);
ic->i_send_cq = NULL;
ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
rdsdebug("ib_create_cq send failed: %d\n", ret);
goto out;
}
ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
cq_attr.cqe = ic->i_recv_ring.w_nr;
cq_attr.comp_vector = ic->i_rcq_vector;
ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
rds_ib_cq_event_handler, conn,
&cq_attr);
if (IS_ERR(ic->i_recv_cq)) {
ret = PTR_ERR(ic->i_recv_cq);
ic->i_recv_cq = NULL;
ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
rdsdebug("ib_create_cq recv failed: %d\n", ret);
goto out;
}
@ -780,10 +807,17 @@ void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
/* first destroy the ib state that generates callbacks */
if (ic->i_cm_id->qp)
rdma_destroy_qp(ic->i_cm_id);
if (ic->i_send_cq)
if (ic->i_send_cq) {
if (ic->rds_ibdev)
ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
ib_destroy_cq(ic->i_send_cq);
if (ic->i_recv_cq)
}
if (ic->i_recv_cq) {
if (ic->rds_ibdev)
ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
ib_destroy_cq(ic->i_recv_cq);
}
/* then free the resources that ib callbacks use */
if (ic->i_send_hdrs)