IB/rdmavt: Fracture single lock used for posting and processing RWQEs
authorKamenee Arumugam <kamenee.arumugam@intel.com>
Fri, 28 Jun 2019 18:04:30 +0000 (14:04 -0400)
committerJason Gunthorpe <jgg@mellanox.com>
Sat, 29 Jun 2019 01:32:16 +0000 (22:32 -0300)
Usage of single lock prevents fetching posted and processing receive work
queue entries from progressing simultaneously and impacts overall
performance.

Fracture the single lock used for posting and processing Receive Work
Queue Entries (RWQEs) to allow the circular buffer to be filled and
emptied at the same time. Two new spinlocks - one for the producers and
one for the consumers used for posting and processing RWQEs simultaneously
and the two indices are define on two different cache lines. The threshold
count is used to avoid reading other index in different cache line every
time.

Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Kamenee Arumugam <kamenee.arumugam@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/sw/rdmavt/qp.c
drivers/infiniband/sw/rdmavt/rc.c
drivers/infiniband/sw/rdmavt/srq.c
include/rdma/rdmavt_qp.h

index 1384060..200b292 100644 (file)
@@ -58,6 +58,8 @@
 #include "vt.h"
 #include "trace.h"
 
+#define RVT_RWQ_COUNT_THRESHOLD 16
+
 static void rvt_rc_timeout(struct timer_list *t);
 
 /*
@@ -835,7 +837,8 @@ int rvt_alloc_rq(struct rvt_rq *rq, u32 size, int node,
                rq->kwq->curr_wq = rq->kwq->wq;
        }
 
-       spin_lock_init(&rq->lock);
+       spin_lock_init(&rq->kwq->p_lock);
+       spin_lock_init(&rq->kwq->c_lock);
        return 0;
 bail:
        rvt_free_rq(rq);
@@ -892,6 +895,8 @@ static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
        qp->s_tail_ack_queue = 0;
        qp->s_acked_ack_queue = 0;
        qp->s_num_rd_atomic = 0;
+       if (qp->r_rq.kwq)
+               qp->r_rq.kwq->count = qp->r_rq.size;
        qp->r_sge.num_sge = 0;
        atomic_set(&qp->s_reserved_used, 0);
 }
@@ -1097,7 +1102,6 @@ struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
                spin_lock_init(&qp->r_lock);
                spin_lock_init(&qp->s_hlock);
                spin_lock_init(&qp->s_lock);
-               spin_lock_init(&qp->r_rq.lock);
                atomic_set(&qp->refcount, 0);
                atomic_set(&qp->local_ops_pending, 0);
                init_waitqueue_head(&qp->wait);
@@ -1305,7 +1309,7 @@ int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
                struct rvt_rwq *wq = NULL;
                struct rvt_krwq *kwq = NULL;
 
-               spin_lock(&qp->r_rq.lock);
+               spin_lock(&qp->r_rq.kwq->c_lock);
                /* qp->ip used to validate if there is a  user buffer mmaped */
                if (qp->ip) {
                        wq = qp->r_rq.wq;
@@ -1331,7 +1335,7 @@ int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
                        RDMA_WRITE_UAPI_ATOMIC(wq->tail, tail);
                else
                        kwq->tail = tail;
-               spin_unlock(&qp->r_rq.lock);
+               spin_unlock(&qp->r_rq.kwq->c_lock);
        } else if (qp->ibqp.event_handler) {
                ret = 1;
        }
@@ -1780,12 +1784,12 @@ int rvt_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
                        return -EINVAL;
                }
 
-               spin_lock_irqsave(&qp->r_rq.lock, flags);
+               spin_lock_irqsave(&qp->r_rq.kwq->p_lock, flags);
                next = wq->head + 1;
                if (next >= qp->r_rq.size)
                        next = 0;
                if (next == READ_ONCE(wq->tail)) {
-                       spin_unlock_irqrestore(&qp->r_rq.lock, flags);
+                       spin_unlock_irqrestore(&qp->r_rq.kwq->p_lock, flags);
                        *bad_wr = wr;
                        return -ENOMEM;
                }
@@ -1810,7 +1814,7 @@ int rvt_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
                         */
                        smp_store_release(&wq->head, next);
                }
-               spin_unlock_irqrestore(&qp->r_rq.lock, flags);
+               spin_unlock_irqrestore(&qp->r_rq.kwq->p_lock, flags);
        }
        return 0;
 }
@@ -2191,13 +2195,13 @@ int rvt_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
                        return -EINVAL;
                }
 
-               spin_lock_irqsave(&srq->rq.lock, flags);
+               spin_lock_irqsave(&srq->rq.kwq->p_lock, flags);
                wq = srq->rq.kwq;
                next = wq->head + 1;
                if (next >= srq->rq.size)
                        next = 0;
                if (next == READ_ONCE(wq->tail)) {
-                       spin_unlock_irqrestore(&srq->rq.lock, flags);
+                       spin_unlock_irqrestore(&srq->rq.kwq->p_lock, flags);
                        *bad_wr = wr;
                        return -ENOMEM;
                }
@@ -2209,7 +2213,7 @@ int rvt_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
                        wqe->sg_list[i] = wr->sg_list[i];
                /* Make sure queue entry is written before the head index. */
                smp_store_release(&wq->head, next);
-               spin_unlock_irqrestore(&srq->rq.lock, flags);
+               spin_unlock_irqrestore(&srq->rq.kwq->p_lock, flags);
        }
        return 0;
 }
@@ -2266,6 +2270,31 @@ bad_lkey:
 }
 
 /**
+ * get_count - count numbers of request work queue entries
+ * in circular buffer
+ * @rq: data structure for request queue entry
+ * @tail: tail indices of the circular buffer
+ * @head: head indices of the circular buffer
+ *
+ * Return - total number of entries in the circular buffer
+ */
+static u32 get_count(struct rvt_rq *rq, u32 tail, u32 head)
+{
+       u32 count;
+
+       count = head;
+
+       if (count >= rq->size)
+               count = 0;
+       if (count < tail)
+               count += rq->size - tail;
+       else
+               count -= tail;
+
+       return count;
+}
+
+/**
  * get_rvt_head - get head indices of the circular buffer
  * @rq: data structure for request queue entry
  * @ip: the QP
@@ -2298,7 +2327,7 @@ int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only)
 {
        unsigned long flags;
        struct rvt_rq *rq;
-       struct rvt_krwq *kwq;
+       struct rvt_krwq *kwq = NULL;
        struct rvt_rwq *wq;
        struct rvt_srq *srq;
        struct rvt_rwqe *wqe;
@@ -2320,16 +2349,16 @@ int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only)
                ip = qp->ip;
        }
 
-       spin_lock_irqsave(&rq->lock, flags);
+       spin_lock_irqsave(&rq->kwq->c_lock, flags);
        if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
                ret = 0;
                goto unlock;
        }
+       kwq = rq->kwq;
        if (ip) {
                wq = rq->wq;
                tail = RDMA_READ_UAPI_ATOMIC(wq->tail);
        } else {
-               kwq = rq->kwq;
                tail = kwq->tail;
        }
 
@@ -2337,8 +2366,11 @@ int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only)
        if (tail >= rq->size)
                tail = 0;
 
-       head = get_rvt_head(rq, ip);
-       if (unlikely(tail == head)) {
+       if (kwq->count < RVT_RWQ_COUNT_THRESHOLD) {
+               head = get_rvt_head(rq, ip);
+               kwq->count = get_count(rq, tail, head);
+       }
+       if (unlikely(kwq->count == 0)) {
                ret = 0;
                goto unlock;
        }
@@ -2362,36 +2394,31 @@ int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only)
        }
        qp->r_wr_id = wqe->wr_id;
 
+       kwq->count--;
        ret = 1;
        set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
        if (handler) {
-               u32 n;
-
                /*
                 * Validate head pointer value and compute
                 * the number of remaining WQEs.
                 */
-               n = get_rvt_head(rq, ip);
-               if (n >= rq->size)
-                       n = 0;
-               if (n < tail)
-                       n += rq->size - tail;
-               else
-                       n -= tail;
-               if (n < srq->limit) {
-                       struct ib_event ev;
-
-                       srq->limit = 0;
-                       spin_unlock_irqrestore(&rq->lock, flags);
-                       ev.device = qp->ibqp.device;
-                       ev.element.srq = qp->ibqp.srq;
-                       ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
-                       handler(&ev, srq->ibsrq.srq_context);
-                       goto bail;
+               if (kwq->count < srq->limit) {
+                       kwq->count = get_count(rq, tail, get_rvt_head(rq, ip));
+                       if (kwq->count < srq->limit) {
+                               struct ib_event ev;
+
+                               srq->limit = 0;
+                               spin_unlock_irqrestore(&rq->kwq->c_lock, flags);
+                               ev.device = qp->ibqp.device;
+                               ev.element.srq = qp->ibqp.srq;
+                               ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
+                               handler(&ev, srq->ibsrq.srq_context);
+                               goto bail;
+                       }
                }
        }
 unlock:
-       spin_unlock_irqrestore(&rq->lock, flags);
+       spin_unlock_irqrestore(&rq->kwq->c_lock, flags);
 bail:
        return ret;
 }
index 44cc7ee..890d7b7 100644 (file)
@@ -107,27 +107,30 @@ __be32 rvt_compute_aeth(struct rvt_qp *qp)
                u32 head;
                u32 tail;
 
-               /* sanity check pointers before trusting them */
-               if (qp->ip) {
-                       head = RDMA_READ_UAPI_ATOMIC(qp->r_rq.wq->head);
-                       tail = RDMA_READ_UAPI_ATOMIC(qp->r_rq.wq->tail);
-               } else {
-                       head = READ_ONCE(qp->r_rq.kwq->head);
-                       tail = READ_ONCE(qp->r_rq.kwq->tail);
+               credits = READ_ONCE(qp->r_rq.kwq->count);
+               if (credits == 0) {
+                       /* sanity check pointers before trusting them */
+                       if (qp->ip) {
+                               head = RDMA_READ_UAPI_ATOMIC(qp->r_rq.wq->head);
+                               tail = RDMA_READ_UAPI_ATOMIC(qp->r_rq.wq->tail);
+                       } else {
+                               head = READ_ONCE(qp->r_rq.kwq->head);
+                               tail = READ_ONCE(qp->r_rq.kwq->tail);
+                       }
+                       if (head >= qp->r_rq.size)
+                               head = 0;
+                       if (tail >= qp->r_rq.size)
+                               tail = 0;
+                       /*
+                        * Compute the number of credits available (RWQEs).
+                        * There is a small chance that the pair of reads are
+                        * not atomic, which is OK, since the fuzziness is
+                        * resolved as further ACKs go out.
+                        */
+                       credits = head - tail;
+                       if ((int)credits < 0)
+                               credits += qp->r_rq.size;
                }
-               if (head >= qp->r_rq.size)
-                       head = 0;
-               if (tail >= qp->r_rq.size)
-                       tail = 0;
-               /*
-                * Compute the number of credits available (RWQEs).
-                * There is a small chance that the pair of reads are
-                * not atomic, which is OK, since the fuzziness is
-                * resolved as further ACKs go out.
-                */
-               credits = head - tail;
-               if ((int)credits < 0)
-                       credits += qp->r_rq.size;
                /*
                 * Binary search the credit table to find the code to
                 * use.
index d306f65..24fef02 100644 (file)
@@ -206,7 +206,7 @@ int rvt_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
                                goto bail_free;
                }
 
-               spin_lock_irq(&srq->rq.lock);
+               spin_lock_irq(&srq->rq.kwq->c_lock);
                /*
                 * validate head and tail pointer values and compute
                 * the number of remaining WQEs.
@@ -261,7 +261,7 @@ int rvt_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
                srq->rq.size = size;
                if (attr_mask & IB_SRQ_LIMIT)
                        srq->limit = attr->srq_limit;
-               spin_unlock_irq(&srq->rq.lock);
+               spin_unlock_irq(&srq->rq.kwq->c_lock);
 
                vfree(owq);
                kvfree(okwq);
@@ -295,17 +295,17 @@ int rvt_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
                        spin_unlock_irq(&dev->pending_lock);
                }
        } else if (attr_mask & IB_SRQ_LIMIT) {
-               spin_lock_irq(&srq->rq.lock);
+               spin_lock_irq(&srq->rq.kwq->c_lock);
                if (attr->srq_limit >= srq->rq.size)
                        ret = -EINVAL;
                else
                        srq->limit = attr->srq_limit;
-               spin_unlock_irq(&srq->rq.lock);
+               spin_unlock_irq(&srq->rq.kwq->c_lock);
        }
        return ret;
 
 bail_unlock:
-       spin_unlock_irq(&srq->rq.lock);
+       spin_unlock_irq(&srq->rq.kwq->c_lock);
 bail_free:
        rvt_free_rq(&tmp_rq);
        return ret;
index ee55fd0..de5915b 100644 (file)
@@ -180,7 +180,9 @@ struct rvt_swqe {
 
 /**
  * struct rvt_krwq - kernel struct receive work request
+ * @p_lock: lock to protect producer of the kernel buffer
  * @head: index of next entry to fill
+ * @c_lock:lock to protect consumer of the kernel buffer
  * @tail: index of next entry to pull
  * @count: count is aproximate of total receive enteries posted
  * @rvt_rwqe: struct of receive work request queue entry
@@ -190,8 +192,13 @@ struct rvt_swqe {
  * mode user.
  */
 struct rvt_krwq {
+       spinlock_t p_lock;      /* protect producer */
        u32 head;               /* new work requests posted to the head */
+
+       /* protect consumer */
+       spinlock_t c_lock ____cacheline_aligned_in_smp;
        u32 tail;               /* receives pull requests from here. */
+       u32 count;              /* approx count of receive entries posted */
        struct rvt_rwqe *curr_wq;
        struct rvt_rwqe wq[];
 };