svcrdma: Poll CQs in "workqueue" mode
authorChuck Lever <chuck.lever@oracle.com>
Tue, 7 Feb 2017 16:59:04 +0000 (11:59 -0500)
committerJ. Bruce Fields <bfields@redhat.com>
Wed, 8 Feb 2017 19:42:01 +0000 (14:42 -0500)
svcrdma calls svc_xprt_put() in its completion handlers, which
currently run in IRQ context.

However, svc_xprt_put() is meant to be invoked in process context,
not in IRQ context. After the last transport reference is gone, it
directly calls a transport release function that expects to run in
process context.

Change the CQ polling modes to IB_POLL_WORKQUEUE so that svcrdma
invokes svc_xprt_put() only in process context. As an added benefit,
bottom half-disabled spin locking can be eliminated from I/O paths.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
net/sunrpc/xprtrdma/svc_rdma_transport.c

index b9ccd73631a9b369328f1bae4e921c1a5b7f4103..f7b2daf72a86582807798379ac3be336b061a958 100644 (file)
@@ -606,12 +606,12 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
 
        dprintk("svcrdma: rqstp=%p\n", rqstp);
 
-       spin_lock_bh(&rdma_xprt->sc_rq_dto_lock);
+       spin_lock(&rdma_xprt->sc_rq_dto_lock);
        if (!list_empty(&rdma_xprt->sc_read_complete_q)) {
                ctxt = list_first_entry(&rdma_xprt->sc_read_complete_q,
                                        struct svc_rdma_op_ctxt, list);
                list_del(&ctxt->list);
-               spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
+               spin_unlock(&rdma_xprt->sc_rq_dto_lock);
                rdma_read_complete(rqstp, ctxt);
                goto complete;
        } else if (!list_empty(&rdma_xprt->sc_rq_dto_q)) {
@@ -623,7 +623,7 @@ int svc_rdma_recvfrom(struct svc_rqst *rqstp)
                clear_bit(XPT_DATA, &xprt->xpt_flags);
                ctxt = NULL;
        }
-       spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock);
+       spin_unlock(&rdma_xprt->sc_rq_dto_lock);
        if (!ctxt) {
                /* This is the EAGAIN path. The svc_recv routine will
                 * return -EAGAIN, the nfsd thread will go to call into
index 87b8b5a103249b8a815169f1e557df0b017489e6..ab2fd5377f944b4a5edeb5ccdce5ada1fb453fff 100644 (file)
@@ -188,7 +188,7 @@ struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
 {
        struct svc_rdma_op_ctxt *ctxt = NULL;
 
-       spin_lock_bh(&xprt->sc_ctxt_lock);
+       spin_lock(&xprt->sc_ctxt_lock);
        xprt->sc_ctxt_used++;
        if (list_empty(&xprt->sc_ctxts))
                goto out_empty;
@@ -196,7 +196,7 @@ struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
        ctxt = list_first_entry(&xprt->sc_ctxts,
                                struct svc_rdma_op_ctxt, list);
        list_del(&ctxt->list);
-       spin_unlock_bh(&xprt->sc_ctxt_lock);
+       spin_unlock(&xprt->sc_ctxt_lock);
 
 out:
        ctxt->count = 0;
@@ -208,15 +208,15 @@ out_empty:
        /* Either pre-allocation missed the mark, or send
         * queue accounting is broken.
         */
-       spin_unlock_bh(&xprt->sc_ctxt_lock);
+       spin_unlock(&xprt->sc_ctxt_lock);
 
        ctxt = alloc_ctxt(xprt, GFP_NOIO);
        if (ctxt)
                goto out;
 
-       spin_lock_bh(&xprt->sc_ctxt_lock);
+       spin_lock(&xprt->sc_ctxt_lock);
        xprt->sc_ctxt_used--;
-       spin_unlock_bh(&xprt->sc_ctxt_lock);
+       spin_unlock(&xprt->sc_ctxt_lock);
        WARN_ONCE(1, "svcrdma: empty RDMA ctxt list?\n");
        return NULL;
 }
@@ -253,10 +253,10 @@ void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
                for (i = 0; i < ctxt->count; i++)
                        put_page(ctxt->pages[i]);
 
-       spin_lock_bh(&xprt->sc_ctxt_lock);
+       spin_lock(&xprt->sc_ctxt_lock);
        xprt->sc_ctxt_used--;
        list_add(&ctxt->list, &xprt->sc_ctxts);
-       spin_unlock_bh(&xprt->sc_ctxt_lock);
+       spin_unlock(&xprt->sc_ctxt_lock);
 }
 
 static void svc_rdma_destroy_ctxts(struct svcxprt_rdma *xprt)
@@ -921,14 +921,14 @@ struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
 {
        struct svc_rdma_fastreg_mr *frmr = NULL;
 
-       spin_lock_bh(&rdma->sc_frmr_q_lock);
+       spin_lock(&rdma->sc_frmr_q_lock);
        if (!list_empty(&rdma->sc_frmr_q)) {
                frmr = list_entry(rdma->sc_frmr_q.next,
                                  struct svc_rdma_fastreg_mr, frmr_list);
                list_del_init(&frmr->frmr_list);
                frmr->sg_nents = 0;
        }
-       spin_unlock_bh(&rdma->sc_frmr_q_lock);
+       spin_unlock(&rdma->sc_frmr_q_lock);
        if (frmr)
                return frmr;
 
@@ -941,10 +941,10 @@ void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
        if (frmr) {
                ib_dma_unmap_sg(rdma->sc_cm_id->device,
                                frmr->sg, frmr->sg_nents, frmr->direction);
-               spin_lock_bh(&rdma->sc_frmr_q_lock);
+               spin_lock(&rdma->sc_frmr_q_lock);
                WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
                list_add(&frmr->frmr_list, &rdma->sc_frmr_q);
-               spin_unlock_bh(&rdma->sc_frmr_q_lock);
+               spin_unlock(&rdma->sc_frmr_q_lock);
        }
 }
 
@@ -1026,13 +1026,13 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
                goto errout;
        }
        newxprt->sc_sq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_sq_depth,
-                                       0, IB_POLL_SOFTIRQ);
+                                       0, IB_POLL_WORKQUEUE);
        if (IS_ERR(newxprt->sc_sq_cq)) {
                dprintk("svcrdma: error creating SQ CQ for connect request\n");
                goto errout;
        }
        newxprt->sc_rq_cq = ib_alloc_cq(dev, newxprt, newxprt->sc_rq_depth,
-                                       0, IB_POLL_SOFTIRQ);
+                                       0, IB_POLL_WORKQUEUE);
        if (IS_ERR(newxprt->sc_rq_cq)) {
                dprintk("svcrdma: error creating RQ CQ for connect request\n");
                goto errout;