RDMA/core: Provide RDMA DIM support for ULPs

Added the interface in the infiniband driver that applies the rdma_dim
adaptive moderation. There is now a special function for allocating an
ib_cq that uses rdma_dim.

Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over
NVMf between two equal end-hosts with 56 cores across a Mellanox switch
using null_blk device:

READS without DIM:
blk size | BW       | IOPS | 99th percentile latency  | 99.99th latency
512B     | 3.8GiB/s | 7.7M | 1401  usec               | 2442  usec
4k       | 7.0GiB/s | 1.8M | 4817  usec               | 6587  usec
64k      | 10.7GiB/s| 175k | 9896  usec               | 10028 usec

IO WRITES without DIM:
blk size | BW       | IOPS | 99th percentile latency  | 99.99th latency
512B     | 3.6GiB/s | 7.5M | 1434  usec               | 2474  usec
4k       | 6.3GiB/s | 1.6M | 938   usec               | 1221  usec
64k      | 10.7GiB/s| 175k | 8979  usec               | 12780 usec

IO READS with DIM:
blk size | BW       | IOPS | 99th percentile latency  | 99.99th latency
512B     | 4GiB/s   | 8.2M | 816    usec              | 889   usec
4k       | 10.1GiB/s| 2.65M| 3359   usec              | 5080  usec
64k      | 10.7GiB/s| 175k | 9896   usec              | 10028 usec

IO WRITES with DIM:
blk size | BW       | IOPS  | 99th percentile latency | 99.99th latency
512B     | 3.9GiB/s | 8.1M  | 799   usec              | 922   usec
4k       | 9.6GiB/s | 2.5M  | 717   usec              | 1004  usec
64k      | 10.7GiB/s| 176k  | 8586  usec              | 12256 usec

The rdma_dim algorithm was designed to measure the effectiveness of
moderation on the flow in a general way and thus should be appropriate
for all RDMA storage protocols.

rdma_dim is configured to be the default option based on performance
improvement seen after extensive tests.

Signed-off-by: Yamin Friedman <yaminf@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

diff --git a/drivers/infiniband/core/cq.c b/drivers/infiniband/core/cq.c
index 00d70f166209a7bda17e228fbd343c8526c2b9de..ffd6e24109d52953bdc446f33d7f329675061d7c 100644 (file)
--- a/drivers/infiniband/core/cq.c
+++ b/drivers/infiniband/core/cq.c
@@ -18,6 +18,40 @@
 #define IB_POLL_FLAGS \
        (IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS)
 
+static void ib_cq_rdma_dim_work(struct work_struct *w)
+{
+       struct dim *dim = container_of(w, struct dim, work);
+       struct ib_cq *cq = dim->priv;
+
+       u16 usec = rdma_dim_prof[dim->profile_ix].usec;
+       u16 comps = rdma_dim_prof[dim->profile_ix].comps;
+
+       dim->state = DIM_START_MEASURE;
+
+       cq->device->ops.modify_cq(cq, comps, usec);
+}
+
+static void rdma_dim_init(struct ib_cq *cq)
+{
+       struct dim *dim;
+
+       if (!cq->device->ops.modify_cq || !cq->device->use_cq_dim ||
+           cq->poll_ctx == IB_POLL_DIRECT)
+               return;
+
+       dim = kzalloc(sizeof(struct dim), GFP_KERNEL);
+       if (!dim)
+               return;
+
+       dim->state = DIM_START_MEASURE;
+       dim->tune_state = DIM_GOING_RIGHT;
+       dim->profile_ix = RDMA_DIM_START_PROFILE;
+       dim->priv = cq;
+       cq->dim = dim;
+
+       INIT_WORK(&dim->work, ib_cq_rdma_dim_work);
+}
+
 static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *wcs,
                           int batch)
 {
@@ -78,6 +112,7 @@ static void ib_cq_completion_direct(struct ib_cq *cq, void *private)
 static int ib_poll_handler(struct irq_poll *iop, int budget)
 {
        struct ib_cq *cq = container_of(iop, struct ib_cq, iop);
+       struct dim *dim = cq->dim;
        int completed;
 
        completed = __ib_process_cq(cq, budget, cq->wc, IB_POLL_BATCH);
@@ -87,6 +122,9 @@ static int ib_poll_handler(struct irq_poll *iop, int budget)
                        irq_poll_sched(&cq->iop);
        }
 
+       if (dim)
+               rdma_dim(dim, completed);
+
        return completed;
 }
 
@@ -105,6 +143,8 @@ static void ib_cq_poll_work(struct work_struct *work)
        if (completed >= IB_POLL_BUDGET_WORKQUEUE ||
            ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
                queue_work(cq->comp_wq, &cq->work);
+       else if (cq->dim)
+               rdma_dim(cq->dim, completed);
 }
 
 static void ib_cq_completion_workqueue(struct ib_cq *cq, void *private)
@@ -161,6 +201,8 @@ struct ib_cq *__ib_alloc_cq_user(struct ib_device *dev, void *private,
 
        rdma_restrack_kadd(&cq->res);
 
+       rdma_dim_init(cq);
+
        switch (cq->poll_ctx) {
        case IB_POLL_DIRECT:
                cq->comp_handler = ib_cq_completion_direct;
@@ -223,6 +265,9 @@ void ib_free_cq_user(struct ib_cq *cq, struct ib_udata *udata)
 
        rdma_restrack_del(&cq->res);
        cq->device->ops.destroy_cq(cq, udata);
+       if (cq->dim)
+               cancel_work_sync(&cq->dim->work);
+       kfree(cq->dim);
        kfree(cq->wc);
        kfree(cq);
 }

diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h
index 4053be51b7fa33a79158646d309283eabb614a5b..c5f8a9f1706374fece3b342a3b995d9e31ca18d5 100644 (file)
--- a/include/rdma/ib_verbs.h
+++ b/include/rdma/ib_verbs.h
@@ -61,6 +61,7 @@
 #include <linux/cgroup_rdma.h>
 #include <linux/irqflags.h>
 #include <linux/preempt.h>
+#include <linux/dim.h>
 #include <uapi/rdma/ib_user_verbs.h>
 #include <rdma/rdma_counter.h>
 #include <rdma/restrack.h>
@@ -1509,6 +1510,7 @@ struct ib_cq {
                struct work_struct      work;
        };
        struct workqueue_struct *comp_wq;
+       struct dim *dim;
        /*
         * Implementation details of the RDMA core, don't use in drivers:
         */
@@ -2576,6 +2578,8 @@ struct ib_device {
        u16                          is_switch:1;
        /* Indicates kernel verbs support, should not be used in drivers */
        u16                          kverbs_provider:1;
+       /* CQ adaptive moderation (RDMA DIM) */
+       u16                          use_cq_dim:1;
        u8                           node_type;
        u8                           phys_port_cnt;
        struct ib_device_attr        attrs;