RDMA/rxe: Fix FIXME in rxe_udp_encap_recv()
Subsystems (2)
PATCH 1/1 RDMA/rxe: Fix FIXME in rxe_udp_encap_recv()
rxe_udp_encap_recv() drops the reference to rxe->ib_dev taken by rxe_get_dev_from_net() which should be held until each received skb is freed. This patch moves the calls to ib_device_put() to each place a received skb is freed. It also takes references to the ib_device for each cloned skb created to process received multicast packets.
instead of open coding")
--- drivers/infiniband/sw/rxe/rxe_comp.c | 49 ++++++++++++---------------- drivers/infiniband/sw/rxe/rxe_net.c | 12 +++---- drivers/infiniband/sw/rxe/rxe_recv.c | 15 +++++++-- drivers/infiniband/sw/rxe/rxe_resp.c | 3 ++ 4 files changed, 42 insertions(+), 37 deletions(-)
drivers/infiniband/sw/rxe/rxe_comp.c +21 -28
@@ -515,6 +515,7 @@ static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)while ((skb =skb_dequeue (&qp->resp_pkts))) {rxe_drop_ref (qp);kfree_skb (skb); +ib_device_put (qp->ibqp.device); }while ((wqe =queue_head (qp->sq.queue))) {
@@ -527,6 +528,17 @@ static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify) } } +void free_pkt (struct rxe_pkt_info *pkt) +{ +struct sk_buff *skb =PKT_TO_SKB (pkt); +struct rxe_qp *qp = pkt->qp; +struct ib_device *dev = qp->ibqp.device; + +kfree_skb (skb); +rxe_drop_ref (qp); +ib_device_put (dev); +} +int rxe_completer (void *arg) {struct rxe_qp *qp = (struct rxe_qp *)arg;
@@ -624,11 +636,8 @@ int rxe_completer(void *arg)break ;case COMPST_DONE : -if (pkt) { -rxe_drop_ref (pkt->qp); -kfree_skb (skb); - skb =NULL ; - } +if (pkt) +free_pkt (pkt); goto done;case COMPST_EXIT :
@@ -671,12 +680,8 @@ int rxe_completer(void *arg)*/ if (qp->comp.started_retry && !qp->comp.timeout_retry) { -if (pkt) { -rxe_drop_ref (pkt->qp); -kfree_skb (skb); - skb =NULL ; - } - +if (pkt) +free_pkt (pkt); goto done; }
@@ -699,13 +704,8 @@ int rxe_completer(void *arg) qp->comp.started_retry =1 ;rxe_run_task (&qp->req.task,0 ); } - -if (pkt) { -rxe_drop_ref (pkt->qp); -kfree_skb (skb); - skb =NULL ; - } - +if (pkt) +free_pkt (pkt); goto done; }else {
@@ -726,9 +726,7 @@ int rxe_completer(void *arg)mod_timer (&qp->rnr_nak_timer, jiffies +rnrnak_jiffies (aeth_syn (pkt) & ~AETH_TYPE_MASK )); -rxe_drop_ref (pkt->qp); -kfree_skb (skb); - skb =NULL ; +free_pkt (pkt); goto exit; }else { rxe_counter_inc(rxe,
@@ -742,13 +740,8 @@ int rxe_completer(void *arg)WARN_ON_ONCE (wqe->status ==IB_WC_SUCCESS );do_complete (qp, wqe);rxe_qp_error (qp); - -if (pkt) { -rxe_drop_ref (pkt->qp); -kfree_skb (skb); - skb =NULL ; - } - +if (pkt) +free_pkt (pkt); goto exit; } }
drivers/infiniband/sw/rxe/rxe_net.c +5 -7
@@ -152,10 +152,14 @@ static struct dst_entry *rxe_find_route(struct net_device *ndev,static int rxe_udp_encap_recv (struct sock *sk,struct sk_buff *skb) {struct udphdr *udph; +struct rxe_dev *rxe;struct net_device *ndev = skb->dev; -struct rxe_dev *rxe =rxe_get_dev_from_net (ndev);struct rxe_pkt_info *pkt =SKB_TO_PKT (skb); +/* takes a reference on rxe->ib_dev +* drop when skb is freed +*/ + rxe =rxe_get_dev_from_net (ndev);if (!rxe) goto drop;
@@ -174,12 +178,6 @@ static int rxe_udp_encap_recv(struct sock *sk, struct sk_buff *skb)rxe_rcv (skb); -/* -* FIXME: this is in the wrong place, it needs to be done when pkt is -* destroyed -*/ -ib_device_put (&rxe->ib_dev); -return 0 ; drop:kfree_skb (skb);
drivers/infiniband/sw/rxe/rxe_recv.c +13 -2
@@ -266,10 +266,19 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)/* for all but the last qp create a new clone of the * skb and pass to the qp. */ -if (mce->qp_list.next != &mcg->qp_list) +if (mce->qp_list.next != &mcg->qp_list) { per_qp_skb =skb_clone (skb,GFP_ATOMIC ); -else +if (!ib_device_try_get (&rxe->ib_dev)) { +/* shouldn't happen we already have +* one ref for skb. +*/ +pr_warn ("ib_device_try_get failed\n" ); +kfree_skb (per_qp_skb); +continue ; + } + }else { per_qp_skb = skb; + }if (unlikely (!per_qp_skb))continue ;
@@ -288,6 +297,7 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb) err1:kfree_skb (skb); +ib_device_put (&rxe->ib_dev); }/**
@@ -397,4 +407,5 @@ void rxe_rcv(struct sk_buff *skb)rxe_drop_ref (pkt->qp);kfree_skb (skb); +ib_device_put (&rxe->ib_dev); }
drivers/infiniband/sw/rxe/rxe_resp.c +3 -0
@@ -99,6 +99,7 @@ static inline enum resp_states get_req(struct rxe_qp *qp,while ((skb =skb_dequeue (&qp->req_pkts))) {rxe_drop_ref (qp);kfree_skb (skb); +ib_device_put (qp->ibqp.device); }/* go drain recv wr queue */
@@ -1012,6 +1013,7 @@ static enum resp_states cleanup(struct rxe_qp *qp, skb =skb_dequeue (&qp->req_pkts);rxe_drop_ref (qp);kfree_skb (skb); +ib_device_put (qp->ibqp.device); }if (qp->resp.mr) {
@@ -1176,6 +1178,7 @@ static void rxe_drain_req_pkts(struct rxe_qp *qp, bool notify)while ((skb =skb_dequeue (&qp->req_pkts))) {rxe_drop_ref (qp);kfree_skb (skb); +ib_device_put (qp->ibqp.device); }if (notify)
Signature · 1 line
2.27.0
Quoted 9 lines · Bob Pearson
rxe_udp_encap_recv()drops the reference torxe->ib_devtaken byrxe_get_dev_from_net()which should be held until each received skb is freed. This patch moves the calls toib_device_put()to each place a received skb is freed. It also takes references to theib_devicefor each cloned skb created to process received multicast packets.Fixes 4c173f596b3ff ("RDMA/rxe: Use ib_device_get_by_netdev()instead of open coding")
Do not line wrap fixes lines
Quoted 8 lines
drivers/infiniband/sw/rxe/rxe_recv.c +0
-0@@ -266,10 +266,19 @@ static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)/* for all but the last qp create a new clone of the * skb and pass to the qp. */
drivers/infiniband/sw/rxe/rxe_recv.c static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb) · line 276
Quoted 10 lines
drivers/infiniband/sw/rxe/rxe_recv.c +7
-2-if (mce->qp_list.next != &mcg->qp_list) +if (mce->qp_list.next != &mcg->qp_list) { per_qp_skb =skb_clone (skb,GFP_ATOMIC ); -else +if (!ib_device_try_get (&rxe->ib_dev)) { +/* shouldn't happen we already have +* one ref for skb. +*/ +pr_warn ("ib_device_try_get failed\n" ); +kfree_skb (per_qp_skb);
I fixed this to just
+ if (WARN_ON(!ib_device_try_get(&rxe->ib_dev))) {
+ kfree_skb(per_qp_skb);
+ continue;
+ }
drivers/infiniband/sw/rxe/rxe_recv.c static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb) · line 281
Quoted 5 lines
drivers/infiniband/sw/rxe/rxe_recv.c +4
-0+continue ; + } + }else { per_qp_skb = skb; + }
And this had a merge collision with the mcast patch, I fixed it up
Applied to for-next, thanks
Jason