aboutsummaryrefslogtreecommitdiff
path: root/drivers/infiniband
diff options
context:
space:
mode:
authorMike Marciniszyn2011-01-10 17:42:22 -0800
committerRoland Dreier2011-01-10 17:42:22 -0800
commit994bcd28a36af1413381dfe0aac065e2cbc2af40 (patch)
tree26f4685a39d94dbfbab346862061ef6f577653f2 /drivers/infiniband
parent2a600f14d25fda341b5633c75cc50a7574fc1007 (diff)
IB/qib: Issue pre-emptive NAKs on eager buffer overflow
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>
Diffstat (limited to 'drivers/infiniband')
-rw-r--r--drivers/infiniband/hw/qib/qib_driver.c141
1 files changed, 137 insertions, 4 deletions
diff --git a/drivers/infiniband/hw/qib/qib_driver.c b/drivers/infiniband/hw/qib/qib_driver.c
index 816a6bdc0b1c..23e584f4c36c 100644
--- a/drivers/infiniband/hw/qib/qib_driver.c
+++ b/drivers/infiniband/hw/qib/qib_driver.c
@@ -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);