@@ -75,6 +75,7 @@ struct svc_rdma_op_ctxt {
struct svc_rdma_fastreg_mr *frmr;
int hdr_count;
struct xdr_buf arg;
+ struct ib_cqe cqe;
struct list_head dto_q;
enum ib_wr_opcode wr_op;
enum ib_wc_status wc_status;
@@ -174,7 +175,6 @@ struct svcxprt_rdma {
struct work_struct sc_work;
};
/* sc_flags */
-#define RDMAXPRT_RQ_PENDING 1
#define RDMAXPRT_SQ_PENDING 2
#define RDMAXPRT_CONN_PENDING 3
@@ -68,7 +68,6 @@ static void svc_rdma_detach(struct svc_xprt *xprt);
static void svc_rdma_free(struct svc_xprt *xprt);
static int svc_rdma_has_wspace(struct svc_xprt *xprt);
static int svc_rdma_secure_port(struct svc_rqst *);
-static void rq_cq_reap(struct svcxprt_rdma *xprt);
static void sq_cq_reap(struct svcxprt_rdma *xprt);
static DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
@@ -413,7 +412,6 @@ static void dto_tasklet_func(unsigned long data)
list_del_init(&xprt->sc_dto_q);
spin_unlock_irqrestore(&dto_lock, flags);
- rq_cq_reap(xprt);
sq_cq_reap(xprt);
svc_xprt_put(&xprt->sc_xprt);
@@ -422,93 +420,49 @@ static void dto_tasklet_func(unsigned long data)
spin_unlock_irqrestore(&dto_lock, flags);
}
-/*
- * Receive Queue Completion Handler
- *
- * Since an RQ completion handler is called on interrupt context, we
- * need to defer the handling of the I/O to a tasklet
- */
-static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
-{
- struct svcxprt_rdma *xprt = cq_context;
- unsigned long flags;
-
- /* Guard against unconditional flush call for destroyed QP */
- if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
- return;
-
- /*
- * Set the bit regardless of whether or not it's on the list
- * because it may be on the list already due to an SQ
- * completion.
- */
- set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
-
- /*
- * If this transport is not already on the DTO transport queue,
- * add it
- */
- spin_lock_irqsave(&dto_lock, flags);
- if (list_empty(&xprt->sc_dto_q)) {
- svc_xprt_get(&xprt->sc_xprt);
- list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
- }
- spin_unlock_irqrestore(&dto_lock, flags);
-
- /* Tasklet does all the work to avoid irqsave locks. */
- tasklet_schedule(&dto_tasklet);
-}
-
-/*
- * rq_cq_reap - Process the RQ CQ.
+/**
+ * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
+ * @cq: completion queue
+ * @wc: completed WR
*
- * Take all completing WC off the CQE and enqueue the associated DTO
- * context on the dto_q for the transport.
- *
- * Note that caller must hold a transport reference.
*/
-static void rq_cq_reap(struct svcxprt_rdma *xprt)
+static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
{
- int ret;
- struct ib_wc wc;
- struct svc_rdma_op_ctxt *ctxt = NULL;
-
- if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
- return;
+ struct svcxprt_rdma *xprt = cq->cq_context;
+ struct ib_cqe *cqe = wc->wr_cqe;
+ struct svc_rdma_op_ctxt *ctxt;
- ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
- atomic_inc(&rdma_stat_rq_poll);
+ /* WARNING: Only wc->wr_cqe and wc->status *
+ * are reliable at this point. */
+ ctxt = container_of(cqe, struct svc_rdma_op_ctxt, cqe);
+ ctxt->wc_status = wc->status;
+ svc_rdma_unmap_dma(ctxt);
- while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
- ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
- ctxt->wc_status = wc.status;
- ctxt->byte_len = wc.byte_len;
- svc_rdma_unmap_dma(ctxt);
- if (wc.status != IB_WC_SUCCESS) {
- /* Close the transport */
- dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
- set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
- svc_rdma_put_context(ctxt, 1);
- svc_xprt_put(&xprt->sc_xprt);
- continue;
- }
- spin_lock_bh(&xprt->sc_rq_dto_lock);
- list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
- spin_unlock_bh(&xprt->sc_rq_dto_lock);
- svc_xprt_put(&xprt->sc_xprt);
- }
+ if (wc->status != IB_WC_SUCCESS)
+ goto flushed;
- if (ctxt)
- atomic_inc(&rdma_stat_rq_prod);
+ /* All wc fields are now known to be valid */
+ ctxt->byte_len = wc->byte_len;
+ spin_lock(&xprt->sc_rq_dto_lock);
+ list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
+ spin_unlock(&xprt->sc_rq_dto_lock);
set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
- /*
- * If data arrived before established event,
- * don't enqueue. This defers RPC I/O until the
- * RDMA connection is complete.
- */
- if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
- svc_xprt_enqueue(&xprt->sc_xprt);
+ if (test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
+ goto out;
+ svc_xprt_enqueue(&xprt->sc_xprt);
+ goto out;
+
+flushed:
+ if (wc->status != IB_WC_WR_FLUSH_ERR)
+ pr_warn("svcrdma: receive: %s (%u/0x%x)\n",
+ ib_wc_status_msg(wc->status),
+ wc->status, wc->vendor_err);
+ set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
+ svc_rdma_put_context(ctxt, 1);
+
+out:
+ svc_xprt_put(&xprt->sc_xprt);
}
/*
@@ -681,6 +635,7 @@ int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
ctxt = svc_rdma_get_context(xprt);
buflen = 0;
ctxt->direction = DMA_FROM_DEVICE;
+ ctxt->cqe.done = svc_rdma_wc_receive;
for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
if (sge_no >= xprt->sc_max_sge) {
pr_err("svcrdma: Too many sges (%d)\n", sge_no);
@@ -705,7 +660,7 @@ int svc_rdma_post_recv(struct svcxprt_rdma *xprt, gfp_t flags)
recv_wr.next = NULL;
recv_wr.sg_list = &ctxt->sge[0];
recv_wr.num_sge = ctxt->count;
- recv_wr.wr_id = (u64)(unsigned long)ctxt;
+ recv_wr.wr_cqe = &ctxt->cqe;
svc_xprt_get(&xprt->sc_xprt);
ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
@@ -1094,12 +1049,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
dprintk("svcrdma: error creating SQ CQ for connect request\n");
goto errout;
}
- cq_attr.cqe = newxprt->sc_rq_depth;
- newxprt->sc_rq_cq = ib_create_cq(dev,
- rq_comp_handler,
- cq_event_handler,
- newxprt,
- &cq_attr);
+ newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
+ 0, IB_POLL_SOFTIRQ);
if (IS_ERR(newxprt->sc_rq_cq)) {
dprintk("svcrdma: error creating RQ CQ for connect request\n");
goto errout;
@@ -1193,7 +1144,6 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
* miss the first message
*/
ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
- ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
/* Accept Connection */
set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
@@ -1337,7 +1287,7 @@ static void __svc_rdma_free(struct work_struct *work)
ib_destroy_cq(rdma->sc_sq_cq);
if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
- ib_destroy_cq(rdma->sc_rq_cq);
+ ib_free_cq(rdma->sc_rq_cq);
if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
ib_dealloc_pd(rdma->sc_pd);
Calling ib_poll_cq() to sort through WCs during a completion is a common pattern amongst RDMA consumers. Since commit 14d3a3b2498e ("IB: add a proper completion queue abstraction"), WC sorting can be handled by the IB core. By converting to this new API, svcrdma is made a better neighbor to other RDMA consumers, as it allows the core to schedule the delivery of completions more fairly amongst all active consumers. Receive completions no longer use the dto_tasklet. Each polled Receive WC is now handled individually in soft IRQ context. The server transport's rdma_stat_rq_poll and rdma_stat_rq_prod metrics are no longer updated. As far as I can tell, ib_alloc_cq() does not enable a consumer to specify a callback for handling async CQ events. This may cause a problem with handling spurious connection loss, though I think Receive completion flushing should be enough to force the server to clean up properly anyway. Signed-off-by: Chuck Lever <chuck.lever@oracle.com> --- include/linux/sunrpc/svc_rdma.h | 2 net/sunrpc/xprtrdma/svc_rdma_transport.c | 130 +++++++++--------------------- 2 files changed, 41 insertions(+), 91 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html