Skip to contentSkip to conversation

RDMA/iw_cxgb4: initiate CLOSE when entering TERM

PATCH Applied Krishnamraju Eraparaju 2 participants 2 messages Feb 4, 2020 – Feb 11, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RDMA/iw_cxgb4: initiate CLOSE when entering TERM

- As per draft-hilland-iwarp-verbs-v1.0, sec 6.2.3, always initiate a CLOSE when entering into TERM state.

- In c4iw_modify_qp(), disconnect operation should only be performed when the modify_qp call is invoked from ib_core. And all other internal modify_qp calls(invoked within iw_cxgb4) that needs 'disconnect' should call c4iw_ep_disconnect() explicitly after modify_qp. Otherwise, deadlocks like below can occur:

 Call Trace:
  schedule+0x2f/0xa0
  schedule_preempt_disabled+0xa/0x10
  __mutex_lock.isra.5+0x2d0/0x4a0
  c4iw_ep_disconnect+0x39/0x430    => tries to reacquire ep lock again
  c4iw_modify_qp+0x468/0x10d0
  rx_data+0x218/0x570              => acquires ep lock
  process_work+0x5f/0x70
  process_one_work+0x1a7/0x3b0
  worker_thread+0x30/0x390
  kthread+0x112/0x130
  ret_from_fork+0x35/0x40
Fixes d2c33370ae73 ("RDMA/iw_cxgb4: Always disconnect when QP is

transitioning to TERMINATE state")

---
 drivers/infiniband/hw/cxgb4/cm.c | 4 ++++
 drivers/infiniband/hw/cxgb4/qp.c | 4 ++--
 2 files changed, 6 insertions(+), 2 deletions(-)
drivers/infiniband/hw/cxgb4/cm.c +4 -0
index ee1182f9b627..d69dece3b1d5 100644
@@ -3036,6 +3036,10 @@ static int terminate(struct c4iw_dev *dev, struct sk_buff *skb)
 				       C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
 		}
 
+		/* As per draft-hilland-iwarp-verbs-v1.0, sec 6.2.3,
+		 * when entering the TERM state the RNIC MUST initiate a CLOSE.
+		 */
+		c4iw_ep_disconnect(ep, 1, GFP_KERNEL);
 		c4iw_put_ep(&ep->com);
 	} else
 		pr_warn("TERM received tid %u no ep/qp\n", tid);
drivers/infiniband/hw/cxgb4/qp.c +2 -2
index bbcac539777a..89ac2f9ae6dd 100644
@@ -1948,10 +1948,10 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
 			qhp->attr.layer_etype = attrs->layer_etype;
 			qhp->attr.ecode = attrs->ecode;
 			ep = qhp->ep;
-			c4iw_get_ep(&ep->com);
-			disconnect = 1;
 			if (!internal) {
+				c4iw_get_ep(&ep->com);
 				terminate = 1;
+				disconnect = 1;
 			} else {
 				terminate = qhp->attr.send_term;
 				ret = rdma_fini(rhp, qhp, ep);
Signature · 1 line
2.23.0.rc0
#1 lore raw
Quoted 22 lines · Krishnamraju Eraparaju

- As per draft-hilland-iwarp-verbs-v1.0, sec 6.2.3, always initiate a CLOSE when entering into TERM state.

- In c4iw_modify_qp(), disconnect operation should only be performed when the modify_qp call is invoked from ib_core. And all other internal modify_qp calls(invoked within iw_cxgb4) that needs 'disconnect' should call c4iw_ep_disconnect() explicitly after modify_qp. Otherwise, deadlocks like below can occur:

 Call Trace:
  schedule+0x2f/0xa0
  schedule_preempt_disabled+0xa/0x10
  __mutex_lock.isra.5+0x2d0/0x4a0
  c4iw_ep_disconnect+0x39/0x430    => tries to reacquire ep lock again
  c4iw_modify_qp+0x468/0x10d0
  rx_data+0x218/0x570              => acquires ep lock
  process_work+0x5f/0x70
  process_one_work+0x1a7/0x3b0
  worker_thread+0x30/0x390
  kthread+0x112/0x130
  ret_from_fork+0x35/0x40
Quoted 7 lines
Fixes d2c33370ae73 ("RDMA/iw_cxgb4: Always disconnect when QP is

transitioning to TERMINATE state")

---
 drivers/infiniband/hw/cxgb4/cm.c | 4 ++++
 drivers/infiniband/hw/cxgb4/qp.c | 4 ++--
 2 files changed, 6 insertions(+), 2 deletions(-)

Applied to for-rc, thanks

Jason