]> git.kernelconcepts.de Git - karo-tx-linux.git/commitdiff
IB/qib: Issue pre-emptive NAKs on eager buffer overflow
authorMike Marciniszyn <mike.marciniszyn@qlogic.com>
Tue, 11 Jan 2011 01:42:22 +0000 (17:42 -0800)
committerRoland Dreier <rolandd@cisco.com>
Tue, 11 Jan 2011 01:42:22 +0000 (17:42 -0800)
Under congestion resulting in eager buffer overflow attempt to send
pre-emptive NAKs if header queue entries with TID errors are generated
and a valid header is present.  This prevents long timeouts and flow
restarts if a trailing set of packets are dropped due to eager
overflows.  Pre-emptive NAKs are currently only supported for RDMA
writes.

Signed-off-by: Mike Marciniszyn <mike.marciniszyn@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/qib/qib_driver.c

index 816a6bdc0b1c76473801cd808cf92341ddf065f0..23e584f4c36c772c2647e72f3e60c962a74c23aa 100644 (file)
@@ -289,14 +289,147 @@ static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
  * Returns 1 if error was a CRC, else 0.
  * Needed for some chip's synthesized error counters.
  */
-static u32 qib_rcv_hdrerr(struct qib_pportdata *ppd, u32 ctxt,
-                         u32 eflags, u32 l, u32 etail, __le32 *rhf_addr,
-                         struct qib_message_header *hdr)
+static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
+                         u32 ctxt, u32 eflags, u32 l, u32 etail,
+                         __le32 *rhf_addr, struct qib_message_header *rhdr)
 {
        u32 ret = 0;
 
        if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
                ret = 1;
+       else if (eflags == QLOGIC_IB_RHF_H_TIDERR) {
+               /* For TIDERR and RC QPs premptively schedule a NAK */
+               struct qib_ib_header *hdr = (struct qib_ib_header *) rhdr;
+               struct qib_other_headers *ohdr = NULL;
+               struct qib_ibport *ibp = &ppd->ibport_data;
+               struct qib_qp *qp = NULL;
+               u32 tlen = qib_hdrget_length_in_bytes(rhf_addr);
+               u16 lid  = be16_to_cpu(hdr->lrh[1]);
+               int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
+               u32 qp_num;
+               u32 opcode;
+               u32 psn;
+               int diff;
+               unsigned long flags;
+
+               /* Sanity check packet */
+               if (tlen < 24)
+                       goto drop;
+
+               if (lid < QIB_MULTICAST_LID_BASE) {
+                       lid &= ~((1 << ppd->lmc) - 1);
+                       if (unlikely(lid != ppd->lid))
+                               goto drop;
+               }
+
+               /* Check for GRH */
+               if (lnh == QIB_LRH_BTH)
+                       ohdr = &hdr->u.oth;
+               else if (lnh == QIB_LRH_GRH) {
+                       u32 vtf;
+
+                       ohdr = &hdr->u.l.oth;
+                       if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
+                               goto drop;
+                       vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
+                       if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
+                               goto drop;
+               } else
+                       goto drop;
+
+               /* Get opcode and PSN from packet */
+               opcode = be32_to_cpu(ohdr->bth[0]);
+               opcode >>= 24;
+               psn = be32_to_cpu(ohdr->bth[2]);
+
+               /* Get the destination QP number. */
+               qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK;
+               if (qp_num != QIB_MULTICAST_QPN) {
+                       int ruc_res;
+                       qp = qib_lookup_qpn(ibp, qp_num);
+                       if (!qp)
+                               goto drop;
+
+                       /*
+                        * Handle only RC QPs - for other QP types drop error
+                        * packet.
+                        */
+                       spin_lock(&qp->r_lock);
+
+                       /* Check for valid receive state. */
+                       if (!(ib_qib_state_ops[qp->state] &
+                             QIB_PROCESS_RECV_OK)) {
+                               ibp->n_pkt_drops++;
+                               goto unlock;
+                       }
+
+                       switch (qp->ibqp.qp_type) {
+                       case IB_QPT_RC:
+                               spin_lock_irqsave(&qp->s_lock, flags);
+                               ruc_res =
+                                       qib_ruc_check_hdr(
+                                               ibp, hdr,
+                                               lnh == QIB_LRH_GRH,
+                                               qp,
+                                               be32_to_cpu(ohdr->bth[0]));
+                               if (ruc_res) {
+                                       spin_unlock_irqrestore(&qp->s_lock,
+                                                              flags);
+                                       goto unlock;
+                               }
+                               spin_unlock_irqrestore(&qp->s_lock, flags);
+
+                               /* Only deal with RDMA Writes for now */
+                               if (opcode <
+                                   IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
+                                       diff = qib_cmp24(psn, qp->r_psn);
+                                       if (!qp->r_nak_state && diff >= 0) {
+                                               ibp->n_rc_seqnak++;
+                                               qp->r_nak_state =
+                                                       IB_NAK_PSN_ERROR;
+                                               /* Use the expected PSN. */
+                                               qp->r_ack_psn = qp->r_psn;
+                                               /*
+                                                * Wait to send the sequence
+                                                * NAK until all packets
+                                                * in the receive queue have
+                                                * been processed.
+                                                * Otherwise, we end up
+                                                * propagating congestion.
+                                                */
+                                               if (list_empty(&qp->rspwait)) {
+                                                       qp->r_flags |=
+                                                               QIB_R_RSP_NAK;
+                                                       atomic_inc(
+                                                               &qp->refcount);
+                                                       list_add_tail(
+                                                        &qp->rspwait,
+                                                        &rcd->qp_wait_list);
+                                               }
+                                       } /* Out of sequence NAK */
+                               } /* QP Request NAKs */
+                               break;
+                       case IB_QPT_SMI:
+                       case IB_QPT_GSI:
+                       case IB_QPT_UD:
+                       case IB_QPT_UC:
+                       default:
+                               /* For now don't handle any other QP types */
+                               break;
+                       }
+
+unlock:
+                       spin_unlock(&qp->r_lock);
+                       /*
+                        * Notify qib_destroy_qp() if it is waiting
+                        * for us to finish.
+                        */
+                       if (atomic_dec_and_test(&qp->refcount))
+                               wake_up(&qp->wait);
+               } /* Unicast QP */
+       } /* Valid packet with TIDErr */
+
+drop:
        return ret;
 }
 
@@ -376,7 +509,7 @@ u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
                 * packets; only qibhdrerr should be set.
                 */
                if (unlikely(eflags))
-                       crcs += qib_rcv_hdrerr(ppd, rcd->ctxt, eflags, l,
+                       crcs += qib_rcv_hdrerr(rcd, ppd, rcd->ctxt, eflags, l,
                                               etail, rhf_addr, hdr);
                else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
                        qib_ib_rcv(rcd, hdr, ebuf, tlen);