aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorRalph Campbell <ralph.campbell@qlogic.com>2007-07-25 14:08:28 -0400
committerRoland Dreier <rolandd@cisco.com>2007-10-09 23:05:49 -0400
commit4ee97180ac76deb5a715ac45b7d7516e6ee82ae7 (patch)
tree6683d1c34d3f36271a9d8275a645ce67222ffc56
parent210d6ca3db058cd1d6e6fd235ee3e25d6ac221cd (diff)
IB/ipath: Change UD to queue work requests like RC & UC
The code to post UD sends tried to process work requests at the time ib_post_send() is called without using a WQE queue. This was fine as long as HW resources were available for sending a packet. This patch changes UD to be handled more like RC and UC and shares more code. Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com> Signed-off-by: Roland Dreier <rolandd@cisco.com>
Diffstat
-rw-r--r--drivers/infiniband/hw/ipath/ipath_qp.c11
-rw-r--r--drivers/infiniband/hw/ipath/ipath_rc.c61
-rw-r--r--drivers/infiniband/hw/ipath/ipath_ruc.c308
-rw-r--r--drivers/infiniband/hw/ipath/ipath_uc.c77
-rw-r--r--drivers/infiniband/hw/ipath/ipath_ud.c372
-rw-r--r--drivers/infiniband/hw/ipath/ipath_verbs.c241
-rw-r--r--drivers/infiniband/hw/ipath/ipath_verbs.h35
7 files changed, 494 insertions, 611 deletions
diff --git a/drivers/infiniband/hw/ipath/ipath_qp.c b/drivers/infiniband/hw/ipath/ipath_qp.c
index 1324b35ff1f8..a8c4a6b03d7a 100644
--- a/drivers/infiniband/hw/ipath/ipath_qp.c
+++ b/drivers/infiniband/hw/ipath/ipath_qp.c
@@ -338,6 +338,7 @@ static void ipath_reset_qp(struct ipath_qp *qp)
338 qp->s_busy = 0; 338 qp->s_busy = 0;
339 qp->s_flags &= IPATH_S_SIGNAL_REQ_WR; 339 qp->s_flags &= IPATH_S_SIGNAL_REQ_WR;
340 qp->s_hdrwords = 0; 340 qp->s_hdrwords = 0;
341 qp->s_wqe = NULL;
341 qp->s_psn = 0; 342 qp->s_psn = 0;
342 qp->r_psn = 0; 343 qp->r_psn = 0;
343 qp->r_msn = 0; 344 qp->r_msn = 0;
@@ -751,6 +752,9 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
751 switch (init_attr->qp_type) { 752 switch (init_attr->qp_type) {
752 case IB_QPT_UC: 753 case IB_QPT_UC:
753 case IB_QPT_RC: 754 case IB_QPT_RC:
755 case IB_QPT_UD:
756 case IB_QPT_SMI:
757 case IB_QPT_GSI:
754 sz = sizeof(struct ipath_sge) * 758 sz = sizeof(struct ipath_sge) *
755 init_attr->cap.max_send_sge + 759 init_attr->cap.max_send_sge +
756 sizeof(struct ipath_swqe); 760 sizeof(struct ipath_swqe);
@@ -759,10 +763,6 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
759 ret = ERR_PTR(-ENOMEM); 763 ret = ERR_PTR(-ENOMEM);
760 goto bail; 764 goto bail;
761 } 765 }
762 /* FALLTHROUGH */
763 case IB_QPT_UD:
764 case IB_QPT_SMI:
765 case IB_QPT_GSI:
766 sz = sizeof(*qp); 766 sz = sizeof(*qp);
767 if (init_attr->srq) { 767 if (init_attr->srq) {
768 struct ipath_srq *srq = to_isrq(init_attr->srq); 768 struct ipath_srq *srq = to_isrq(init_attr->srq);
@@ -805,8 +805,7 @@ struct ib_qp *ipath_create_qp(struct ib_pd *ibpd,
805 spin_lock_init(&qp->r_rq.lock); 805 spin_lock_init(&qp->r_rq.lock);
806 atomic_set(&qp->refcount, 0); 806 atomic_set(&qp->refcount, 0);
807 init_waitqueue_head(&qp->wait); 807 init_waitqueue_head(&qp->wait);
808 tasklet_init(&qp->s_task, ipath_do_ruc_send, 808 tasklet_init(&qp->s_task, ipath_do_send, (unsigned long)qp);
809 (unsigned long)qp);
810 INIT_LIST_HEAD(&qp->piowait); 809 INIT_LIST_HEAD(&qp->piowait);
811 INIT_LIST_HEAD(&qp->timerwait); 810 INIT_LIST_HEAD(&qp->timerwait);
812 qp->state = IB_QPS_RESET; 811 qp->state = IB_QPS_RESET;
diff --git a/drivers/infiniband/hw/ipath/ipath_rc.c b/drivers/infiniband/hw/ipath/ipath_rc.c
index 46744ea2babd..53259daeb4f8 100644
--- a/drivers/infiniband/hw/ipath/ipath_rc.c
+++ b/drivers/infiniband/hw/ipath/ipath_rc.c
@@ -81,9 +81,8 @@ static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
81 * Note that we are in the responder's side of the QP context. 81 * Note that we are in the responder's side of the QP context.
82 * Note the QP s_lock must be held. 82 * Note the QP s_lock must be held.
83 */ 83 */
84static int ipath_make_rc_ack(struct ipath_qp *qp, 84static int ipath_make_rc_ack(struct ipath_ibdev *dev, struct ipath_qp *qp,
85 struct ipath_other_headers *ohdr, 85 struct ipath_other_headers *ohdr, u32 pmtu)
86 u32 pmtu, u32 *bth0p, u32 *bth2p)
87{ 86{
88 struct ipath_ack_entry *e; 87 struct ipath_ack_entry *e;
89 u32 hwords; 88 u32 hwords;
@@ -192,8 +191,7 @@ static int ipath_make_rc_ack(struct ipath_qp *qp,
192 } 191 }
193 qp->s_hdrwords = hwords; 192 qp->s_hdrwords = hwords;
194 qp->s_cur_size = len; 193 qp->s_cur_size = len;
195 *bth0p = bth0 | (1 << 22); /* Set M bit */ 194 ipath_make_ruc_header(dev, qp, ohdr, bth0, bth2);
196 *bth2p = bth2;
197 return 1; 195 return 1;
198 196
199bail: 197bail:
@@ -203,32 +201,39 @@ bail:
203/** 201/**
204 * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC) 202 * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
205 * @qp: a pointer to the QP 203 * @qp: a pointer to the QP
206 * @ohdr: a pointer to the IB header being constructed
207 * @pmtu: the path MTU
208 * @bth0p: pointer to the BTH opcode word
209 * @bth2p: pointer to the BTH PSN word
210 * 204 *
211 * Return 1 if constructed; otherwise, return 0. 205 * Return 1 if constructed; otherwise, return 0.
212 * Note the QP s_lock must be held and interrupts disabled.
213 */ 206 */
214int ipath_make_rc_req(struct ipath_qp *qp, 207int ipath_make_rc_req(struct ipath_qp *qp)
215 struct ipath_other_headers *ohdr,
216 u32 pmtu, u32 *bth0p, u32 *bth2p)
217{ 208{
218 struct ipath_ibdev *dev = to_idev(qp->ibqp.device); 209 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
210 struct ipath_other_headers *ohdr;
219 struct ipath_sge_state *ss; 211 struct ipath_sge_state *ss;
220 struct ipath_swqe *wqe; 212 struct ipath_swqe *wqe;
221 u32 hwords; 213 u32 hwords;
222 u32 len; 214 u32 len;
223 u32 bth0; 215 u32 bth0;
224 u32 bth2; 216 u32 bth2;
217 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
225 char newreq; 218 char newreq;
219 unsigned long flags;
220 int ret = 0;
221
222 ohdr = &qp->s_hdr.u.oth;
223 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
224 ohdr = &qp->s_hdr.u.l.oth;
225
226 /*
227 * The lock is needed to synchronize between the sending tasklet,
228 * the receive interrupt handler, and timeout resends.
229 */
230 spin_lock_irqsave(&qp->s_lock, flags);
226 231
227 /* Sending responses has higher priority over sending requests. */ 232 /* Sending responses has higher priority over sending requests. */
228 if ((qp->r_head_ack_queue != qp->s_tail_ack_queue || 233 if ((qp->r_head_ack_queue != qp->s_tail_ack_queue ||
229 (qp->s_flags & IPATH_S_ACK_PENDING) || 234 (qp->s_flags & IPATH_S_ACK_PENDING) ||
230 qp->s_ack_state != OP(ACKNOWLEDGE)) && 235 qp->s_ack_state != OP(ACKNOWLEDGE)) &&
231 ipath_make_rc_ack(qp, ohdr, pmtu, bth0p, bth2p)) 236 ipath_make_rc_ack(dev, qp, ohdr, pmtu))
232 goto done; 237 goto done;
233 238
234 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) || 239 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
@@ -560,13 +565,12 @@ int ipath_make_rc_req(struct ipath_qp *qp,
560 qp->s_hdrwords = hwords; 565 qp->s_hdrwords = hwords;
561 qp->s_cur_sge = ss; 566 qp->s_cur_sge = ss;
562 qp->s_cur_size = len; 567 qp->s_cur_size = len;
563 *bth0p = bth0 | (qp->s_state << 24); 568 ipath_make_ruc_header(dev, qp, ohdr, bth0 | (qp->s_state << 24), bth2);
564 *bth2p = bth2;
565done: 569done:
566 return 1; 570 ret = 1;
567
568bail: 571bail:
569 return 0; 572 spin_unlock_irqrestore(&qp->s_lock, flags);
573 return ret;
570} 574}
571 575
572/** 576/**
@@ -627,7 +631,7 @@ static void send_rc_ack(struct ipath_qp *qp)
627 /* 631 /*
628 * If we can send the ACK, clear the ACK state. 632 * If we can send the ACK, clear the ACK state.
629 */ 633 */
630 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) { 634 if (ipath_verbs_send(qp, &hdr, hwords, NULL, 0) == 0) {
631 dev->n_unicast_xmit++; 635 dev->n_unicast_xmit++;
632 goto done; 636 goto done;
633 } 637 }
@@ -757,7 +761,9 @@ void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
757 wc->vendor_err = 0; 761 wc->vendor_err = 0;
758 wc->byte_len = 0; 762 wc->byte_len = 0;
759 wc->qp = &qp->ibqp; 763 wc->qp = &qp->ibqp;
764 wc->imm_data = 0;
760 wc->src_qp = qp->remote_qpn; 765 wc->src_qp = qp->remote_qpn;
766 wc->wc_flags = 0;
761 wc->pkey_index = 0; 767 wc->pkey_index = 0;
762 wc->slid = qp->remote_ah_attr.dlid; 768 wc->slid = qp->remote_ah_attr.dlid;
763 wc->sl = qp->remote_ah_attr.sl; 769 wc->sl = qp->remote_ah_attr.sl;
@@ -1041,7 +1047,9 @@ static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode,
1041 wc.vendor_err = 0; 1047 wc.vendor_err = 0;
1042 wc.byte_len = 0; 1048 wc.byte_len = 0;
1043 wc.qp = &qp->ibqp; 1049 wc.qp = &qp->ibqp;
1050 wc.imm_data = 0;
1044 wc.src_qp = qp->remote_qpn; 1051 wc.src_qp = qp->remote_qpn;
1052 wc.wc_flags = 0;
1045 wc.pkey_index = 0; 1053 wc.pkey_index = 0;
1046 wc.slid = qp->remote_ah_attr.dlid; 1054 wc.slid = qp->remote_ah_attr.dlid;
1047 wc.sl = qp->remote_ah_attr.sl; 1055 wc.sl = qp->remote_ah_attr.sl;
@@ -1454,6 +1462,19 @@ static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1454 goto send_ack; 1462 goto send_ack;
1455 } 1463 }
1456 /* 1464 /*
1465 * Try to send a simple ACK to work around a Mellanox bug
1466 * which doesn't accept a RDMA read response or atomic
1467 * response as an ACK for earlier SENDs or RDMA writes.
1468 */
1469 if (qp->r_head_ack_queue == qp->s_tail_ack_queue &&
1470 !(qp->s_flags & IPATH_S_ACK_PENDING) &&
1471 qp->s_ack_state == OP(ACKNOWLEDGE)) {
1472 spin_unlock_irqrestore(&qp->s_lock, flags);
1473 qp->r_nak_state = 0;
1474 qp->r_ack_psn = qp->s_ack_queue[i].psn - 1;
1475 goto send_ack;
1476 }
1477 /*
1457 * Resend the RDMA read or atomic op which 1478 * Resend the RDMA read or atomic op which
1458 * ACKs this duplicate request. 1479 * ACKs this duplicate request.
1459 */ 1480 */
diff --git a/drivers/infiniband/hw/ipath/ipath_ruc.c b/drivers/infiniband/hw/ipath/ipath_ruc.c
index c69c25239443..4b6b7ee8e5c1 100644
--- a/drivers/infiniband/hw/ipath/ipath_ruc.c
+++ b/drivers/infiniband/hw/ipath/ipath_ruc.c
@@ -31,6 +31,8 @@
31 * SOFTWARE. 31 * SOFTWARE.
32 */ 32 */
33 33
34#include <linux/spinlock.h>
35
34#include "ipath_verbs.h" 36#include "ipath_verbs.h"
35#include "ipath_kernel.h" 37#include "ipath_kernel.h"
36 38
@@ -106,27 +108,30 @@ void ipath_insert_rnr_queue(struct ipath_qp *qp)
106 spin_unlock_irqrestore(&dev->pending_lock, flags); 108 spin_unlock_irqrestore(&dev->pending_lock, flags);
107} 109}
108 110
109static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe) 111/**
112 * ipath_init_sge - Validate a RWQE and fill in the SGE state
113 * @qp: the QP
114 *
115 * Return 1 if OK.
116 */
117int ipath_init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
118 u32 *lengthp, struct ipath_sge_state *ss)
110{ 119{
111 int user = to_ipd(qp->ibqp.pd)->user;
112 int i, j, ret; 120 int i, j, ret;
113 struct ib_wc wc; 121 struct ib_wc wc;
114 122
115 qp->r_len = 0; 123 *lengthp = 0;
116 for (i = j = 0; i < wqe->num_sge; i++) { 124 for (i = j = 0; i < wqe->num_sge; i++) {
117 if (wqe->sg_list[i].length == 0) 125 if (wqe->sg_list[i].length == 0)
118 continue; 126 continue;
119 /* Check LKEY */ 127 /* Check LKEY */
120 if ((user && wqe->sg_list[i].lkey == 0) || 128 if (!ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
121 !ipath_lkey_ok(qp, &qp->r_sg_list[j], &wqe->sg_list[i], 129 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
122 IB_ACCESS_LOCAL_WRITE))
123 goto bad_lkey; 130 goto bad_lkey;
124 qp->r_len += wqe->sg_list[i].length; 131 *lengthp += wqe->sg_list[i].length;
125 j++; 132 j++;
126 } 133 }
127 qp->r_sge.sge = qp->r_sg_list[0]; 134 ss->num_sge = j;
128 qp->r_sge.sg_list = qp->r_sg_list + 1;
129 qp->r_sge.num_sge = j;
130 ret = 1; 135 ret = 1;
131 goto bail; 136 goto bail;
132 137
@@ -172,6 +177,8 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
172 u32 tail; 177 u32 tail;
173 int ret; 178 int ret;
174 179
180 qp->r_sge.sg_list = qp->r_sg_list;
181
175 if (qp->ibqp.srq) { 182 if (qp->ibqp.srq) {
176 srq = to_isrq(qp->ibqp.srq); 183 srq = to_isrq(qp->ibqp.srq);
177 handler = srq->ibsrq.event_handler; 184 handler = srq->ibsrq.event_handler;
@@ -199,7 +206,8 @@ int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only)
199 wqe = get_rwqe_ptr(rq, tail); 206 wqe = get_rwqe_ptr(rq, tail);
200 if (++tail >= rq->size) 207 if (++tail >= rq->size)
201 tail = 0; 208 tail = 0;
202 } while (!wr_id_only && !init_sge(qp, wqe)); 209 } while (!wr_id_only && !ipath_init_sge(qp, wqe, &qp->r_len,
210 &qp->r_sge));
203 qp->r_wr_id = wqe->wr_id; 211 qp->r_wr_id = wqe->wr_id;
204 wq->tail = tail; 212 wq->tail = tail;
205 213
@@ -239,9 +247,9 @@ bail:
239 247
240/** 248/**
241 * ipath_ruc_loopback - handle UC and RC lookback requests 249 * ipath_ruc_loopback - handle UC and RC lookback requests
242 * @sqp: the loopback QP 250 * @sqp: the sending QP
243 * 251 *
244 * This is called from ipath_do_uc_send() or ipath_do_rc_send() to 252 * This is called from ipath_do_send() to
245 * forward a WQE addressed to the same HCA. 253 * forward a WQE addressed to the same HCA.
246 * Note that although we are single threaded due to the tasklet, we still 254 * Note that although we are single threaded due to the tasklet, we still
247 * have to protect against post_send(). We don't have to worry about 255 * have to protect against post_send(). We don't have to worry about
@@ -450,40 +458,18 @@ again:
450 wc.byte_len = wqe->length; 458 wc.byte_len = wqe->length;
451 wc.qp = &qp->ibqp; 459 wc.qp = &qp->ibqp;
452 wc.src_qp = qp->remote_qpn; 460 wc.src_qp = qp->remote_qpn;
453 /* XXX do we know which pkey matched? Only needed for GSI. */
454 wc.pkey_index = 0; 461 wc.pkey_index = 0;
455 wc.slid = qp->remote_ah_attr.dlid; 462 wc.slid = qp->remote_ah_attr.dlid;
456 wc.sl = qp->remote_ah_attr.sl; 463 wc.sl = qp->remote_ah_attr.sl;
457 wc.dlid_path_bits = 0; 464 wc.dlid_path_bits = 0;
465 wc.port_num = 1;
458 /* Signal completion event if the solicited bit is set. */ 466 /* Signal completion event if the solicited bit is set. */
459 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 467 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
460 wqe->wr.send_flags & IB_SEND_SOLICITED); 468 wqe->wr.send_flags & IB_SEND_SOLICITED);
461 469
462send_comp: 470send_comp:
463 sqp->s_rnr_retry = sqp->s_rnr_retry_cnt; 471 sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
464 472 ipath_send_complete(sqp, wqe, IB_WC_SUCCESS);
465 if (!(sqp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
466 (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
467 wc.wr_id = wqe->wr.wr_id;
468 wc.status = IB_WC_SUCCESS;
469 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
470 wc.vendor_err = 0;
471 wc.byte_len = wqe->length;
472 wc.qp = &sqp->ibqp;
473 wc.src_qp = 0;
474 wc.pkey_index = 0;
475 wc.slid = 0;
476 wc.sl = 0;
477 wc.dlid_path_bits = 0;
478 wc.port_num = 0;
479 ipath_cq_enter(to_icq(sqp->ibqp.send_cq), &wc, 0);
480 }
481
482 /* Update s_last now that we are finished with the SWQE */
483 spin_lock_irqsave(&sqp->s_lock, flags);
484 if (++sqp->s_last >= sqp->s_size)
485 sqp->s_last = 0;
486 spin_unlock_irqrestore(&sqp->s_lock, flags);
487 goto again; 473 goto again;
488 474
489done: 475done:
@@ -491,13 +477,11 @@ done:
491 wake_up(&qp->wait); 477 wake_up(&qp->wait);
492} 478}
493 479
494static int want_buffer(struct ipath_devdata *dd) 480static void want_buffer(struct ipath_devdata *dd)
495{ 481{
496 set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl); 482 set_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
497 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 483 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
498 dd->ipath_sendctrl); 484 dd->ipath_sendctrl);
499
500 return 0;
501} 485}
502 486
503/** 487/**
@@ -507,14 +491,11 @@ static int want_buffer(struct ipath_devdata *dd)
507 * 491 *
508 * Called when we run out of PIO buffers. 492 * Called when we run out of PIO buffers.
509 */ 493 */
510static void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev) 494static void ipath_no_bufs_available(struct ipath_qp *qp,
495 struct ipath_ibdev *dev)
511{ 496{
512 unsigned long flags; 497 unsigned long flags;
513 498
514 spin_lock_irqsave(&dev->pending_lock, flags);
515 if (list_empty(&qp->piowait))
516 list_add_tail(&qp->piowait, &dev->piowait);
517 spin_unlock_irqrestore(&dev->pending_lock, flags);
518 /* 499 /*
519 * Note that as soon as want_buffer() is called and 500 * Note that as soon as want_buffer() is called and
520 * possibly before it returns, ipath_ib_piobufavail() 501 * possibly before it returns, ipath_ib_piobufavail()
@@ -524,101 +505,14 @@ static void ipath_no_bufs_available(struct ipath_qp *qp, struct ipath_ibdev *dev
524 * We leave the busy flag set so that another post send doesn't 505 * We leave the busy flag set so that another post send doesn't
525 * try to put the same QP on the piowait list again. 506 * try to put the same QP on the piowait list again.
526 */ 507 */
508 spin_lock_irqsave(&dev->pending_lock, flags);
509 list_add_tail(&qp->piowait, &dev->piowait);
510 spin_unlock_irqrestore(&dev->pending_lock, flags);
527 want_buffer(dev->dd); 511 want_buffer(dev->dd);
528 dev->n_piowait++; 512 dev->n_piowait++;
529} 513}
530 514
531/** 515/**
532 * ipath_post_ruc_send - post RC and UC sends
533 * @qp: the QP to post on
534 * @wr: the work request to send
535 */
536int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr)
537{
538 struct ipath_swqe *wqe;
539 unsigned long flags;
540 u32 next;
541 int i, j;
542 int acc;
543 int ret;
544
545 /*
546 * Don't allow RDMA reads or atomic operations on UC or
547 * undefined operations.
548 * Make sure buffer is large enough to hold the result for atomics.
549 */
550 if (qp->ibqp.qp_type == IB_QPT_UC) {
551 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ) {
552 ret = -EINVAL;
553 goto bail;
554 }
555 } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) {
556 ret = -EINVAL;
557 goto bail;
558 } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
559 (wr->num_sge == 0 ||
560 wr->sg_list[0].length < sizeof(u64) ||
561 wr->sg_list[0].addr & (sizeof(u64) - 1))) {
562 ret = -EINVAL;
563 goto bail;
564 } else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) {
565 ret = -EINVAL;
566 goto bail;
567 }
568 /* IB spec says that num_sge == 0 is OK. */
569 if (wr->num_sge > qp->s_max_sge) {
570 ret = -ENOMEM;
571 goto bail;
572 }
573 spin_lock_irqsave(&qp->s_lock, flags);
574 next = qp->s_head + 1;
575 if (next >= qp->s_size)
576 next = 0;
577 if (next == qp->s_last) {
578 spin_unlock_irqrestore(&qp->s_lock, flags);
579 ret = -EINVAL;
580 goto bail;
581 }
582
583 wqe = get_swqe_ptr(qp, qp->s_head);
584 wqe->wr = *wr;
585 wqe->ssn = qp->s_ssn++;
586 wqe->sg_list[0].mr = NULL;
587 wqe->sg_list[0].vaddr = NULL;
588 wqe->sg_list[0].length = 0;
589 wqe->sg_list[0].sge_length = 0;
590 wqe->length = 0;
591 acc = wr->opcode >= IB_WR_RDMA_READ ? IB_ACCESS_LOCAL_WRITE : 0;
592 for (i = 0, j = 0; i < wr->num_sge; i++) {
593 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
594 spin_unlock_irqrestore(&qp->s_lock, flags);
595 ret = -EINVAL;
596 goto bail;
597 }
598 if (wr->sg_list[i].length == 0)
599 continue;
600 if (!ipath_lkey_ok(qp, &wqe->sg_list[j], &wr->sg_list[i],
601 acc)) {
602 spin_unlock_irqrestore(&qp->s_lock, flags);
603 ret = -EINVAL;
604 goto bail;
605 }
606 wqe->length += wr->sg_list[i].length;
607 j++;
608 }
609 wqe->wr.num_sge = j;
610 qp->s_head = next;
611 spin_unlock_irqrestore(&qp->s_lock, flags);
612
613 ipath_do_ruc_send((unsigned long) qp);
614
615 ret = 0;
616
617bail:
618 return ret;
619}
620
621/**
622 * ipath_make_grh - construct a GRH header 516 * ipath_make_grh - construct a GRH header
623 * @dev: a pointer to the ipath device 517 * @dev: a pointer to the ipath device
624 * @hdr: a pointer to the GRH header being constructed 518 * @hdr: a pointer to the GRH header being constructed
@@ -648,39 +542,66 @@ u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
648 return sizeof(struct ib_grh) / sizeof(u32); 542 return sizeof(struct ib_grh) / sizeof(u32);
649} 543}
650 544
545void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
546 struct ipath_other_headers *ohdr,
547 u32 bth0, u32 bth2)
548{
549 u16 lrh0;
550 u32 nwords;
551 u32 extra_bytes;
552
553 /* Construct the header. */
554 extra_bytes = -qp->s_cur_size & 3;
555 nwords = (qp->s_cur_size + extra_bytes) >> 2;
556 lrh0 = IPATH_LRH_BTH;
557 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
558 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
559 &qp->remote_ah_attr.grh,
560 qp->s_hdrwords, nwords);
561 lrh0 = IPATH_LRH_GRH;
562 }
563 lrh0 |= qp->remote_ah_attr.sl << 4;
564 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
565 qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
566 qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
567 qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
568 bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
569 bth0 |= extra_bytes << 20;
570 ohdr->bth[0] = cpu_to_be32(bth0 | (1 << 22));
571 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
572 ohdr->bth[2] = cpu_to_be32(bth2);
573}
574
651/** 575/**
652 * ipath_do_ruc_send - perform a send on an RC or UC QP 576 * ipath_do_send - perform a send on a QP
653 * @data: contains a pointer to the QP 577 * @data: contains a pointer to the QP
654 * 578 *
655 * Process entries in the send work queue until credit or queue is 579 * Process entries in the send work queue until credit or queue is
656 * exhausted. Only allow one CPU to send a packet per QP (tasklet). 580 * exhausted. Only allow one CPU to send a packet per QP (tasklet).
657 * Otherwise, after we drop the QP s_lock, two threads could send 581 * Otherwise, two threads could send packets out of order.
658 * packets out of order.
659 */ 582 */
660void ipath_do_ruc_send(unsigned long data) 583void ipath_do_send(unsigned long data)
661{ 584{
662 struct ipath_qp *qp = (struct ipath_qp *)data; 585 struct ipath_qp *qp = (struct ipath_qp *)data;
663 struct ipath_ibdev *dev = to_idev(qp->ibqp.device); 586 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
664 unsigned long flags; 587 int (*make_req)(struct ipath_qp *qp);
665 u16 lrh0;
666 u32 nwords;
667 u32 extra_bytes;
668 u32 bth0;
669 u32 bth2;
670 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
671 struct ipath_other_headers *ohdr;
672 588
673 if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy)) 589 if (test_and_set_bit(IPATH_S_BUSY, &qp->s_busy))
674 goto bail; 590 goto bail;
675 591
676 if (unlikely(qp->remote_ah_attr.dlid == dev->dd->ipath_lid)) { 592 if ((qp->ibqp.qp_type == IB_QPT_RC ||
593 qp->ibqp.qp_type == IB_QPT_UC) &&
594 qp->remote_ah_attr.dlid == dev->dd->ipath_lid) {
677 ipath_ruc_loopback(qp); 595 ipath_ruc_loopback(qp);
678 goto clear; 596 goto clear;
679 } 597 }
680 598
681 ohdr = &qp->s_hdr.u.oth; 599 if (qp->ibqp.qp_type == IB_QPT_RC)
682 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH) 600 make_req = ipath_make_rc_req;
683 ohdr = &qp->s_hdr.u.l.oth; 601 else if (qp->ibqp.qp_type == IB_QPT_UC)
602 make_req = ipath_make_uc_req;
603 else
604 make_req = ipath_make_ud_req;
684 605
685again: 606again:
686 /* Check for a constructed packet to be sent. */ 607 /* Check for a constructed packet to be sent. */
@@ -689,9 +610,8 @@ again:
689 * If no PIO bufs are available, return. An interrupt will 610 * If no PIO bufs are available, return. An interrupt will
690 * call ipath_ib_piobufavail() when one is available. 611 * call ipath_ib_piobufavail() when one is available.
691 */ 612 */
692 if (ipath_verbs_send(dev->dd, qp->s_hdrwords, 613 if (ipath_verbs_send(qp, &qp->s_hdr, qp->s_hdrwords,
693 (u32 *) &qp->s_hdr, qp->s_cur_size, 614 qp->s_cur_sge, qp->s_cur_size)) {
694 qp->s_cur_sge)) {
695 ipath_no_bufs_available(qp, dev); 615 ipath_no_bufs_available(qp, dev);
696 goto bail; 616 goto bail;
697 } 617 }
@@ -700,54 +620,42 @@ again:
700 qp->s_hdrwords = 0; 620 qp->s_hdrwords = 0;
701 } 621 }
702 622
703 /* 623 if (make_req(qp))
704 * The lock is needed to synchronize between setting 624 goto again;
705 * qp->s_ack_state, resend timer, and post_send(). 625clear:
706 */ 626 clear_bit(IPATH_S_BUSY, &qp->s_busy);
707 spin_lock_irqsave(&qp->s_lock, flags); 627bail:;
708 628}
709 if (!((qp->ibqp.qp_type == IB_QPT_RC) ?
710 ipath_make_rc_req(qp, ohdr, pmtu, &bth0, &bth2) :
711 ipath_make_uc_req(qp, ohdr, pmtu, &bth0, &bth2))) {
712 /*
713 * Clear the busy bit before unlocking to avoid races with
714 * adding new work queue items and then failing to process
715 * them.
716 */
717 clear_bit(IPATH_S_BUSY, &qp->s_busy);
718 spin_unlock_irqrestore(&qp->s_lock, flags);
719 goto bail;
720 }
721 629
722 spin_unlock_irqrestore(&qp->s_lock, flags); 630void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
631 enum ib_wc_status status)
632{
633 u32 last = qp->s_last;
723 634
724 /* Construct the header. */ 635 if (++last == qp->s_size)
725 extra_bytes = (4 - qp->s_cur_size) & 3; 636 last = 0;
726 nwords = (qp->s_cur_size + extra_bytes) >> 2; 637 qp->s_last = last;
727 lrh0 = IPATH_LRH_BTH;
728 if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
729 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
730 &qp->remote_ah_attr.grh,
731 qp->s_hdrwords, nwords);
732 lrh0 = IPATH_LRH_GRH;
733 }
734 lrh0 |= qp->remote_ah_attr.sl << 4;
735 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
736 qp->s_hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
737 qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
738 SIZE_OF_CRC);
739 qp->s_hdr.lrh[3] = cpu_to_be16(dev->dd->ipath_lid);
740 bth0 |= ipath_get_pkey(dev->dd, qp->s_pkey_index);
741 bth0 |= extra_bytes << 20;
742 ohdr->bth[0] = cpu_to_be32(bth0);
743 ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
744 ohdr->bth[2] = cpu_to_be32(bth2);
745 638
746 /* Check for more work to do. */ 639 /* See ch. 11.2.4.1 and 10.7.3.1 */
747 goto again; 640 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
641 (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
642 status != IB_WC_SUCCESS) {
643 struct ib_wc wc;
748 644
749clear: 645 wc.wr_id = wqe->wr.wr_id;
750 clear_bit(IPATH_S_BUSY, &qp->s_busy); 646 wc.status = status;
751bail: 647 wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
752 return; 648 wc.vendor_err = 0;
649 wc.byte_len = wqe->length;
650 wc.imm_data = 0;
651 wc.qp = &qp->ibqp;
652 wc.src_qp = 0;
653 wc.wc_flags = 0;
654 wc.pkey_index = 0;
655 wc.slid = 0;
656 wc.sl = 0;
657 wc.dlid_path_bits = 0;
658 wc.port_num = 0;
659 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
660 }
753} 661}
diff --git a/drivers/infiniband/hw/ipath/ipath_uc.c b/drivers/infiniband/hw/ipath/ipath_uc.c
index 8380fbc50d2c..767beb903c25 100644
--- a/drivers/infiniband/hw/ipath/ipath_uc.c
+++ b/drivers/infiniband/hw/ipath/ipath_uc.c
@@ -37,72 +37,40 @@
37/* cut down ridiculously long IB macro names */ 37/* cut down ridiculously long IB macro names */
38#define OP(x) IB_OPCODE_UC_##x 38#define OP(x) IB_OPCODE_UC_##x
39 39
40static void complete_last_send(struct ipath_qp *qp, struct ipath_swqe *wqe,
41 struct ib_wc *wc)
42{
43 if (++qp->s_last == qp->s_size)
44 qp->s_last = 0;
45 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) ||
46 (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
47 wc->wr_id = wqe->wr.wr_id;
48 wc->status = IB_WC_SUCCESS;
49 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
50 wc->vendor_err = 0;
51 wc->byte_len = wqe->length;
52 wc->qp = &qp->ibqp;
53 wc->src_qp = qp->remote_qpn;
54 wc->pkey_index = 0;
55 wc->slid = qp->remote_ah_attr.dlid;
56 wc->sl = qp->remote_ah_attr.sl;
57 wc->dlid_path_bits = 0;
58 wc->port_num = 0;
59 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 0);
60 }
61}
62
63/** 40/**
64 * ipath_make_uc_req - construct a request packet (SEND, RDMA write) 41 * ipath_make_uc_req - construct a request packet (SEND, RDMA write)
65 * @qp: a pointer to the QP 42 * @qp: a pointer to the QP
66 * @ohdr: a pointer to the IB header being constructed
67 * @pmtu: the path MTU
68 * @bth0p: pointer to the BTH opcode word
69 * @bth2p: pointer to the BTH PSN word
70 * 43 *
71 * Return 1 if constructed; otherwise, return 0. 44 * Return 1 if constructed; otherwise, return 0.
72 * Note the QP s_lock must be held and interrupts disabled.
73 */ 45 */
74int ipath_make_uc_req(struct ipath_qp *qp, 46int ipath_make_uc_req(struct ipath_qp *qp)
75 struct ipath_other_headers *ohdr,
76 u32 pmtu, u32 *bth0p, u32 *bth2p)
77{ 47{
48 struct ipath_other_headers *ohdr;
78 struct ipath_swqe *wqe; 49 struct ipath_swqe *wqe;
79 u32 hwords; 50 u32 hwords;
80 u32 bth0; 51 u32 bth0;
81 u32 len; 52 u32 len;
82 struct ib_wc wc; 53 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
54 int ret = 0;
83 55
84 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) 56 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK))
85 goto done; 57 goto done;
86 58
59 ohdr = &qp->s_hdr.u.oth;
60 if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
61 ohdr = &qp->s_hdr.u.l.oth;
62
87 /* header size in 32-bit words LRH+BTH = (8+12)/4. */ 63 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
88 hwords = 5; 64 hwords = 5;
89 bth0 = 1 << 22; /* Set M bit */ 65 bth0 = 1 << 22; /* Set M bit */
90 66
91 /* Get the next send request. */ 67 /* Get the next send request. */
92 wqe = get_swqe_ptr(qp, qp->s_last); 68 wqe = get_swqe_ptr(qp, qp->s_cur);
69 qp->s_wqe = NULL;
93 switch (qp->s_state) { 70 switch (qp->s_state) {
94 default: 71 default:
95 /*
96 * Signal the completion of the last send
97 * (if there is one).
98 */
99 if (qp->s_last != qp->s_tail) {
100 complete_last_send(qp, wqe, &wc);
101 wqe = get_swqe_ptr(qp, qp->s_last);
102 }
103
104 /* Check if send work queue is empty. */ 72 /* Check if send work queue is empty. */
105 if (qp->s_tail == qp->s_head) 73 if (qp->s_cur == qp->s_head)
106 goto done; 74 goto done;
107 /* 75 /*
108 * Start a new request. 76 * Start a new request.
@@ -131,6 +99,9 @@ int ipath_make_uc_req(struct ipath_qp *qp,
131 } 99 }
132 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 100 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
133 bth0 |= 1 << 23; 101 bth0 |= 1 << 23;
102 qp->s_wqe = wqe;
103 if (++qp->s_cur >= qp->s_size)
104 qp->s_cur = 0;
134 break; 105 break;
135 106
136 case IB_WR_RDMA_WRITE: 107 case IB_WR_RDMA_WRITE:
@@ -157,13 +128,14 @@ int ipath_make_uc_req(struct ipath_qp *qp,
157 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 128 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
158 bth0 |= 1 << 23; 129 bth0 |= 1 << 23;
159 } 130 }
131 qp->s_wqe = wqe;
132 if (++qp->s_cur >= qp->s_size)
133 qp->s_cur = 0;
160 break; 134 break;
161 135
162 default: 136 default:
163 goto done; 137 goto done;
164 } 138 }
165 if (++qp->s_tail >= qp->s_size)
166 qp->s_tail = 0;
167 break; 139 break;
168 140
169 case OP(SEND_FIRST): 141 case OP(SEND_FIRST):
@@ -185,6 +157,9 @@ int ipath_make_uc_req(struct ipath_qp *qp,
185 } 157 }
186 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 158 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
187 bth0 |= 1 << 23; 159 bth0 |= 1 << 23;
160 qp->s_wqe = wqe;
161 if (++qp->s_cur >= qp->s_size)
162 qp->s_cur = 0;
188 break; 163 break;
189 164
190 case OP(RDMA_WRITE_FIRST): 165 case OP(RDMA_WRITE_FIRST):
@@ -207,18 +182,22 @@ int ipath_make_uc_req(struct ipath_qp *qp,
207 if (wqe->wr.send_flags & IB_SEND_SOLICITED) 182 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
208 bth0 |= 1 << 23; 183 bth0 |= 1 << 23;
209 } 184 }
185 qp->s_wqe = wqe;
186 if (++qp->s_cur >= qp->s_size)
187 qp->s_cur = 0;
210 break; 188 break;
211 } 189 }
212 qp->s_len -= len; 190 qp->s_len -= len;
213 qp->s_hdrwords = hwords; 191 qp->s_hdrwords = hwords;
214 qp->s_cur_sge = &qp->s_sge; 192 qp->s_cur_sge = &qp->s_sge;
215 qp->s_cur_size = len; 193 qp->s_cur_size = len;
216 *bth0p = bth0 | (qp->s_state << 24); 194 ipath_make_ruc_header(to_idev(qp->ibqp.device),
217 *bth2p = qp->s_next_psn++ & IPATH_PSN_MASK; 195 qp, ohdr, bth0 | (qp->s_state << 24),
218 return 1; 196 qp->s_next_psn++ & IPATH_PSN_MASK);
197 ret = 1;
219 198
220done: 199done:
221 return 0; 200 return ret;
222} 201}
223 202
224/** 203/**
diff --git a/drivers/infiniband/hw/ipath/ipath_ud.c b/drivers/infiniband/hw/ipath/ipath_ud.c
index f9a3338a5fb7..34c4a0a5be31 100644
--- a/drivers/infiniband/hw/ipath/ipath_ud.c
+++ b/drivers/infiniband/hw/ipath/ipath_ud.c
@@ -36,68 +36,17 @@
36#include "ipath_verbs.h" 36#include "ipath_verbs.h"
37#include "ipath_kernel.h" 37#include "ipath_kernel.h"
38 38
39static int init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
40 u32 *lengthp, struct ipath_sge_state *ss)
41{
42 int user = to_ipd(qp->ibqp.pd)->user;
43 int i, j, ret;
44 struct ib_wc wc;
45
46 *lengthp = 0;
47 for (i = j = 0; i < wqe->num_sge; i++) {
48 if (wqe->sg_list[i].length == 0)
49 continue;
50 /* Check LKEY */
51 if ((user && wqe->sg_list[i].lkey == 0) ||
52 !ipath_lkey_ok(qp, j ? &ss->sg_list[j - 1] : &ss->sge,
53 &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
54 goto bad_lkey;
55 *lengthp += wqe->sg_list[i].length;
56 j++;
57 }
58 ss->num_sge = j;
59 ret = 1;
60 goto bail;
61
62bad_lkey:
63 wc.wr_id = wqe->wr_id;
64 wc.status = IB_WC_LOC_PROT_ERR;
65 wc.opcode = IB_WC_RECV;
66 wc.vendor_err = 0;
67 wc.byte_len = 0;
68 wc.imm_data = 0;
69 wc.qp = &qp->ibqp;
70 wc.src_qp = 0;
71 wc.wc_flags = 0;
72 wc.pkey_index = 0;
73 wc.slid = 0;
74 wc.sl = 0;
75 wc.dlid_path_bits = 0;
76 wc.port_num = 0;
77 /* Signal solicited completion event. */
78 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
79 ret = 0;
80bail:
81 return ret;
82}
83
84/** 39/**
85 * ipath_ud_loopback - handle send on loopback QPs 40 * ipath_ud_loopback - handle send on loopback QPs
86 * @sqp: the QP 41 * @sqp: the sending QP
87 * @ss: the SGE state 42 * @swqe: the send work request
88 * @length: the length of the data to send
89 * @wr: the work request
90 * @wc: the work completion entry
91 * 43 *
92 * This is called from ipath_post_ud_send() to forward a WQE addressed 44 * This is called from ipath_make_ud_req() to forward a WQE addressed
93 * to the same HCA. 45 * to the same HCA.
94 * Note that the receive interrupt handler may be calling ipath_ud_rcv() 46 * Note that the receive interrupt handler may be calling ipath_ud_rcv()
95 * while this is being called. 47 * while this is being called.
96 */ 48 */
97static void ipath_ud_loopback(struct ipath_qp *sqp, 49static void ipath_ud_loopback(struct ipath_qp *sqp, struct ipath_swqe *swqe)
98 struct ipath_sge_state *ss,
99 u32 length, struct ib_send_wr *wr,
100 struct ib_wc *wc)
101{ 50{
102 struct ipath_ibdev *dev = to_idev(sqp->ibqp.device); 51 struct ipath_ibdev *dev = to_idev(sqp->ibqp.device);
103 struct ipath_qp *qp; 52 struct ipath_qp *qp;
@@ -110,12 +59,18 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
110 struct ipath_rwq *wq; 59 struct ipath_rwq *wq;
111 struct ipath_rwqe *wqe; 60 struct ipath_rwqe *wqe;
112 void (*handler)(struct ib_event *, void *); 61 void (*handler)(struct ib_event *, void *);
62 struct ib_wc wc;
113 u32 tail; 63 u32 tail;
114 u32 rlen; 64 u32 rlen;
65 u32 length;
115 66
116 qp = ipath_lookup_qpn(&dev->qp_table, wr->wr.ud.remote_qpn); 67 qp = ipath_lookup_qpn(&dev->qp_table, swqe->wr.wr.ud.remote_qpn);
117 if (!qp) 68 if (!qp) {
118 return; 69 dev->n_pkt_drops++;
70 goto send_comp;
71 }
72
73 rsge.sg_list = NULL;
119 74
120 /* 75 /*
121 * Check that the qkey matches (except for QP0, see 9.6.1.4.1). 76 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
@@ -123,39 +78,34 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
123 * qkey from the QP context instead of the WR (see 10.2.5). 78 * qkey from the QP context instead of the WR (see 10.2.5).
124 */ 79 */
125 if (unlikely(qp->ibqp.qp_num && 80 if (unlikely(qp->ibqp.qp_num &&
126 ((int) wr->wr.ud.remote_qkey < 0 81 ((int) swqe->wr.wr.ud.remote_qkey < 0 ?
127 ? qp->qkey : wr->wr.ud.remote_qkey) != qp->qkey)) { 82 sqp->qkey : swqe->wr.wr.ud.remote_qkey) != qp->qkey)) {
128 /* XXX OK to lose a count once in a while. */ 83 /* XXX OK to lose a count once in a while. */
129 dev->qkey_violations++; 84 dev->qkey_violations++;
130 dev->n_pkt_drops++; 85 dev->n_pkt_drops++;
131 goto done; 86 goto drop;
132 } 87 }
133 88
134 /* 89 /*
135 * A GRH is expected to preceed the data even if not 90 * A GRH is expected to preceed the data even if not
136 * present on the wire. 91 * present on the wire.
137 */ 92 */
138 wc->byte_len = length + sizeof(struct ib_grh); 93 length = swqe->length;
94 wc.byte_len = length + sizeof(struct ib_grh);
139 95
140 if (wr->opcode == IB_WR_SEND_WITH_IMM) { 96 if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
141 wc->wc_flags = IB_WC_WITH_IMM; 97 wc.wc_flags = IB_WC_WITH_IMM;
142 wc->imm_data = wr->imm_data; 98 wc.imm_data = swqe->wr.imm_data;
143 } else { 99 } else {
144 wc->wc_flags = 0; 100 wc.wc_flags = 0;
145 wc->imm_data = 0; 101 wc.imm_data = 0;
146 } 102 }
147 103
148 if (wr->num_sge > 1) {
149 rsge.sg_list = kmalloc((wr->num_sge - 1) *
150 sizeof(struct ipath_sge),
151 GFP_ATOMIC);
152 } else
153 rsge.sg_list = NULL;
154
155 /* 104 /*
156 * Get the next work request entry to find where to put the data. 105 * This would be a lot simpler if we could call ipath_get_rwqe()
157 * Note that it is safe to drop the lock after changing rq->tail 106 * but that uses state that the receive interrupt handler uses
158 * since ipath_post_receive() won't fill the empty slot. 107 * so we would need to lock out receive interrupts while doing
108 * local loopback.
159 */ 109 */
160 if (qp->ibqp.srq) { 110 if (qp->ibqp.srq) {
161 srq = to_isrq(qp->ibqp.srq); 111 srq = to_isrq(qp->ibqp.srq);
@@ -167,32 +117,53 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
167 rq = &qp->r_rq; 117 rq = &qp->r_rq;
168 } 118 }
169 119
120 if (rq->max_sge > 1) {
121 /*
122 * XXX We could use GFP_KERNEL if ipath_do_send()
123 * was always called from the tasklet instead of
124 * from ipath_post_send().
125 */
126 rsge.sg_list = kmalloc((rq->max_sge - 1) *
127 sizeof(struct ipath_sge),
128 GFP_ATOMIC);
129 if (!rsge.sg_list) {
130 dev->n_pkt_drops++;
131 goto drop;
132 }
133 }
134
135 /*
136 * Get the next work request entry to find where to put the data.
137 * Note that it is safe to drop the lock after changing rq->tail
138 * since ipath_post_receive() won't fill the empty slot.
139 */
170 spin_lock_irqsave(&rq->lock, flags); 140 spin_lock_irqsave(&rq->lock, flags);
171 wq = rq->wq; 141 wq = rq->wq;
172 tail = wq->tail; 142 tail = wq->tail;
173 while (1) { 143 /* Validate tail before using it since it is user writable. */
174 if (unlikely(tail == wq->head)) { 144 if (tail >= rq->size)
175 spin_unlock_irqrestore(&rq->lock, flags); 145 tail = 0;
176 dev->n_pkt_drops++; 146 if (unlikely(tail == wq->head)) {
177 goto bail_sge; 147 spin_unlock_irqrestore(&rq->lock, flags);
178 } 148 dev->n_pkt_drops++;
179 /* Make sure entry is read after head index is read. */ 149 goto drop;
180 smp_rmb(); 150 }
181 wqe = get_rwqe_ptr(rq, tail); 151 wqe = get_rwqe_ptr(rq, tail);
182 if (++tail >= rq->size) 152 if (!ipath_init_sge(qp, wqe, &rlen, &rsge)) {
183 tail = 0; 153 spin_unlock_irqrestore(&rq->lock, flags);
184 if (init_sge(qp, wqe, &rlen, &rsge)) 154 dev->n_pkt_drops++;
185 break; 155 goto drop;
186 wq->tail = tail;
187 } 156 }
188 /* Silently drop packets which are too big. */ 157 /* Silently drop packets which are too big. */
189 if (wc->byte_len > rlen) { 158 if (wc.byte_len > rlen) {
190 spin_unlock_irqrestore(&rq->lock, flags); 159 spin_unlock_irqrestore(&rq->lock, flags);
191 dev->n_pkt_drops++; 160 dev->n_pkt_drops++;
192 goto bail_sge; 161 goto drop;
193 } 162 }
163 if (++tail >= rq->size)
164 tail = 0;
194 wq->tail = tail; 165 wq->tail = tail;
195 wc->wr_id = wqe->wr_id; 166 wc.wr_id = wqe->wr_id;
196 if (handler) { 167 if (handler) {
197 u32 n; 168 u32 n;
198 169
@@ -221,13 +192,13 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
221 } else 192 } else
222 spin_unlock_irqrestore(&rq->lock, flags); 193 spin_unlock_irqrestore(&rq->lock, flags);
223 194
224 ah_attr = &to_iah(wr->wr.ud.ah)->attr; 195 ah_attr = &to_iah(swqe->wr.wr.ud.ah)->attr;
225 if (ah_attr->ah_flags & IB_AH_GRH) { 196 if (ah_attr->ah_flags & IB_AH_GRH) {
226 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh)); 197 ipath_copy_sge(&rsge, &ah_attr->grh, sizeof(struct ib_grh));
227 wc->wc_flags |= IB_WC_GRH; 198 wc.wc_flags |= IB_WC_GRH;
228 } else 199 } else
229 ipath_skip_sge(&rsge, sizeof(struct ib_grh)); 200 ipath_skip_sge(&rsge, sizeof(struct ib_grh));
230 sge = &ss->sge; 201 sge = swqe->sg_list;
231 while (length) { 202 while (length) {
232 u32 len = sge->length; 203 u32 len = sge->length;
233 204
@@ -241,8 +212,8 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
241 sge->length -= len; 212 sge->length -= len;
242 sge->sge_length -= len; 213 sge->sge_length -= len;
243 if (sge->sge_length == 0) { 214 if (sge->sge_length == 0) {
244 if (--ss->num_sge) 215 if (--swqe->wr.num_sge)
245 *sge = *ss->sg_list++; 216 sge++;
246 } else if (sge->length == 0 && sge->mr != NULL) { 217 } else if (sge->length == 0 && sge->mr != NULL) {
247 if (++sge->n >= IPATH_SEGSZ) { 218 if (++sge->n >= IPATH_SEGSZ) {
248 if (++sge->m >= sge->mr->mapsz) 219 if (++sge->m >= sge->mr->mapsz)
@@ -256,123 +227,60 @@ static void ipath_ud_loopback(struct ipath_qp *sqp,
256 } 227 }
257 length -= len; 228 length -= len;
258 } 229 }
259 wc->status = IB_WC_SUCCESS; 230 wc.status = IB_WC_SUCCESS;
260 wc->opcode = IB_WC_RECV; 231 wc.opcode = IB_WC_RECV;
261 wc->vendor_err = 0; 232 wc.vendor_err = 0;
262 wc->qp = &qp->ibqp; 233 wc.qp = &qp->ibqp;
263 wc->src_qp = sqp->ibqp.qp_num; 234 wc.src_qp = sqp->ibqp.qp_num;
264 /* XXX do we know which pkey matched? Only needed for GSI. */ 235 /* XXX do we know which pkey matched? Only needed for GSI. */
265 wc->pkey_index = 0; 236 wc.pkey_index = 0;
266 wc->slid = dev->dd->ipath_lid | 237 wc.slid = dev->dd->ipath_lid |
267 (ah_attr->src_path_bits & 238 (ah_attr->src_path_bits &
268 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1)); 239 ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1));
269 wc->sl = ah_attr->sl; 240 wc.sl = ah_attr->sl;
270 wc->dlid_path_bits = 241 wc.dlid_path_bits =
271 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1); 242 ah_attr->dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
243 wc.port_num = 1;
272 /* Signal completion event if the solicited bit is set. */ 244 /* Signal completion event if the solicited bit is set. */
273 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), wc, 245 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
274 wr->send_flags & IB_SEND_SOLICITED); 246 swqe->wr.send_flags & IB_SEND_SOLICITED);
275 247drop:
276bail_sge:
277 kfree(rsge.sg_list); 248 kfree(rsge.sg_list);
278done:
279 if (atomic_dec_and_test(&qp->refcount)) 249 if (atomic_dec_and_test(&qp->refcount))
280 wake_up(&qp->wait); 250 wake_up(&qp->wait);
251send_comp:
252 ipath_send_complete(sqp, swqe, IB_WC_SUCCESS);
281} 253}
282 254
283/** 255/**
284 * ipath_post_ud_send - post a UD send on QP 256 * ipath_make_ud_req - construct a UD request packet
285 * @qp: the QP 257 * @qp: the QP
286 * @wr: the work request
287 * 258 *
288 * Note that we actually send the data as it is posted instead of putting 259 * Return 1 if constructed; otherwise, return 0.
289 * the request into a ring buffer. If we wanted to use a ring buffer,
290 * we would need to save a reference to the destination address in the SWQE.
291 */ 260 */
292int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr) 261int ipath_make_ud_req(struct ipath_qp *qp)
293{ 262{
294 struct ipath_ibdev *dev = to_idev(qp->ibqp.device); 263 struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
295 struct ipath_other_headers *ohdr; 264 struct ipath_other_headers *ohdr;
296 struct ib_ah_attr *ah_attr; 265 struct ib_ah_attr *ah_attr;
297 struct ipath_sge_state ss; 266 struct ipath_swqe *wqe;
298 struct ipath_sge *sg_list;
299 struct ib_wc wc;
300 u32 hwords;
301 u32 nwords; 267 u32 nwords;
302 u32 len;
303 u32 extra_bytes; 268 u32 extra_bytes;
304 u32 bth0; 269 u32 bth0;
305 u16 lrh0; 270 u16 lrh0;
306 u16 lid; 271 u16 lid;
307 int i; 272 int ret = 0;
308 int ret;
309 273
310 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)) { 274 if (unlikely(!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK)))
311 ret = 0;
312 goto bail; 275 goto bail;
313 }
314 276
315 if (wr->wr.ud.ah->pd != qp->ibqp.pd) { 277 if (qp->s_cur == qp->s_head)
316 ret = -EPERM;
317 goto bail; 278 goto bail;
318 }
319 279
320 /* IB spec says that num_sge == 0 is OK. */ 280 wqe = get_swqe_ptr(qp, qp->s_cur);
321 if (wr->num_sge > qp->s_max_sge) {
322 ret = -EINVAL;
323 goto bail;
324 }
325
326 if (wr->num_sge > 1) {
327 sg_list = kmalloc((qp->s_max_sge - 1) * sizeof(*sg_list),
328 GFP_ATOMIC);
329 if (!sg_list) {
330 ret = -ENOMEM;
331 goto bail;
332 }
333 } else
334 sg_list = NULL;
335
336 /* Check the buffer to send. */
337 ss.sg_list = sg_list;
338 ss.sge.mr = NULL;
339 ss.sge.vaddr = NULL;
340 ss.sge.length = 0;
341 ss.sge.sge_length = 0;
342 ss.num_sge = 0;
343 len = 0;
344 for (i = 0; i < wr->num_sge; i++) {
345 /* Check LKEY */
346 if (to_ipd(qp->ibqp.pd)->user && wr->sg_list[i].lkey == 0) {
347 ret = -EINVAL;
348 goto bail;
349 }
350
351 if (wr->sg_list[i].length == 0)
352 continue;
353 if (!ipath_lkey_ok(qp, ss.num_sge ?
354 sg_list + ss.num_sge - 1 : &ss.sge,
355 &wr->sg_list[i], 0)) {
356 ret = -EINVAL;
357 goto bail;
358 }
359 len += wr->sg_list[i].length;
360 ss.num_sge++;
361 }
362 /* Check for invalid packet size. */
363 if (len > dev->dd->ipath_ibmtu) {
364 ret = -EINVAL;
365 goto bail;
366 }
367 extra_bytes = (4 - len) & 3;
368 nwords = (len + extra_bytes) >> 2;
369 281
370 /* Construct the header. */ 282 /* Construct the header. */
371 ah_attr = &to_iah(wr->wr.ud.ah)->attr; 283 ah_attr = &to_iah(wqe->wr.wr.ud.ah)->attr;
372 if (ah_attr->dlid == 0) {
373 ret = -EINVAL;
374 goto bail;
375 }
376 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) { 284 if (ah_attr->dlid >= IPATH_MULTICAST_LID_BASE) {
377 if (ah_attr->dlid != IPATH_PERMISSIVE_LID) 285 if (ah_attr->dlid != IPATH_PERMISSIVE_LID)
378 dev->n_multicast_xmit++; 286 dev->n_multicast_xmit++;
@@ -383,64 +291,53 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
383 lid = ah_attr->dlid & 291 lid = ah_attr->dlid &
384 ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1); 292 ~((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
385 if (unlikely(lid == dev->dd->ipath_lid)) { 293 if (unlikely(lid == dev->dd->ipath_lid)) {
386 /* 294 ipath_ud_loopback(qp, wqe);
387 * Pass in an uninitialized ib_wc to save stack
388 * space.
389 */
390 ipath_ud_loopback(qp, &ss, len, wr, &wc);
391 goto done; 295 goto done;
392 } 296 }
393 } 297 }
298
299 extra_bytes = -wqe->length & 3;
300 nwords = (wqe->length + extra_bytes) >> 2;
301
302 /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
303 qp->s_hdrwords = 7;
304 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM)
305 qp->s_hdrwords++;
306 qp->s_cur_size = wqe->length;
307 qp->s_cur_sge = &qp->s_sge;
308 qp->s_wqe = wqe;
309 qp->s_sge.sge = wqe->sg_list[0];
310 qp->s_sge.sg_list = wqe->sg_list + 1;
311 qp->s_sge.num_sge = wqe->wr.num_sge;
312
394 if (ah_attr->ah_flags & IB_AH_GRH) { 313 if (ah_attr->ah_flags & IB_AH_GRH) {
395 /* Header size in 32-bit words. */ 314 /* Header size in 32-bit words. */
396 hwords = 17; 315 qp->s_hdrwords += ipath_make_grh(dev, &qp->s_hdr.u.l.grh,
316 &ah_attr->grh,
317 qp->s_hdrwords, nwords);
397 lrh0 = IPATH_LRH_GRH; 318 lrh0 = IPATH_LRH_GRH;
398 ohdr = &qp->s_hdr.u.l.oth; 319 ohdr = &qp->s_hdr.u.l.oth;
399 qp->s_hdr.u.l.grh.version_tclass_flow =
400 cpu_to_be32((6 << 28) |
401 (ah_attr->grh.traffic_class << 20) |
402 ah_attr->grh.flow_label);
403 qp->s_hdr.u.l.grh.paylen =
404 cpu_to_be16(((wr->opcode ==
405 IB_WR_SEND_WITH_IMM ? 6 : 5) +
406 nwords + SIZE_OF_CRC) << 2);
407 /* next_hdr is defined by C8-7 in ch. 8.4.1 */
408 qp->s_hdr.u.l.grh.next_hdr = 0x1B;
409 qp->s_hdr.u.l.grh.hop_limit = ah_attr->grh.hop_limit;
410 /* The SGID is 32-bit aligned. */
411 qp->s_hdr.u.l.grh.sgid.global.subnet_prefix =
412 dev->gid_prefix;
413 qp->s_hdr.u.l.grh.sgid.global.interface_id =
414 dev->dd->ipath_guid;
415 qp->s_hdr.u.l.grh.dgid = ah_attr->grh.dgid;
416 /* 320 /*
417 * Don't worry about sending to locally attached multicast 321 * Don't worry about sending to locally attached multicast
418 * QPs. It is unspecified by the spec. what happens. 322 * QPs. It is unspecified by the spec. what happens.
419 */ 323 */
420 } else { 324 } else {
421 /* Header size in 32-bit words. */ 325 /* Header size in 32-bit words. */
422 hwords = 7;
423 lrh0 = IPATH_LRH_BTH; 326 lrh0 = IPATH_LRH_BTH;
424 ohdr = &qp->s_hdr.u.oth; 327 ohdr = &qp->s_hdr.u.oth;
425 } 328 }
426 if (wr->opcode == IB_WR_SEND_WITH_IMM) { 329 if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
427 ohdr->u.ud.imm_data = wr->imm_data; 330 ohdr->u.ud.imm_data = wqe->wr.imm_data;
428 wc.imm_data = wr->imm_data;
429 hwords += 1;
430 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24; 331 bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
431 } else if (wr->opcode == IB_WR_SEND) { 332 } else
432 wc.imm_data = 0;
433 bth0 = IB_OPCODE_UD_SEND_ONLY << 24; 333 bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
434 } else {
435 ret = -EINVAL;
436 goto bail;
437 }
438 lrh0 |= ah_attr->sl << 4; 334 lrh0 |= ah_attr->sl << 4;
439 if (qp->ibqp.qp_type == IB_QPT_SMI) 335 if (qp->ibqp.qp_type == IB_QPT_SMI)
440 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */ 336 lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
441 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0); 337 qp->s_hdr.lrh[0] = cpu_to_be16(lrh0);
442 qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */ 338 qp->s_hdr.lrh[1] = cpu_to_be16(ah_attr->dlid); /* DEST LID */
443 qp->s_hdr.lrh[2] = cpu_to_be16(hwords + nwords + SIZE_OF_CRC); 339 qp->s_hdr.lrh[2] = cpu_to_be16(qp->s_hdrwords + nwords +
340 SIZE_OF_CRC);
444 lid = dev->dd->ipath_lid; 341 lid = dev->dd->ipath_lid;
445 if (lid) { 342 if (lid) {
446 lid |= ah_attr->src_path_bits & 343 lid |= ah_attr->src_path_bits &
@@ -448,7 +345,7 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
448 qp->s_hdr.lrh[3] = cpu_to_be16(lid); 345 qp->s_hdr.lrh[3] = cpu_to_be16(lid);
449 } else 346 } else
450 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE; 347 qp->s_hdr.lrh[3] = IB_LID_PERMISSIVE;
451 if (wr->send_flags & IB_SEND_SOLICITED) 348 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
452 bth0 |= 1 << 23; 349 bth0 |= 1 << 23;
453 bth0 |= extra_bytes << 20; 350 bth0 |= extra_bytes << 20;
454 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY : 351 bth0 |= qp->ibqp.qp_type == IB_QPT_SMI ? IPATH_DEFAULT_P_KEY :
@@ -460,38 +357,20 @@ int ipath_post_ud_send(struct ipath_qp *qp, struct ib_send_wr *wr)
460 ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE && 357 ohdr->bth[1] = ah_attr->dlid >= IPATH_MULTICAST_LID_BASE &&
461 ah_attr->dlid != IPATH_PERMISSIVE_LID ? 358 ah_attr->dlid != IPATH_PERMISSIVE_LID ?
462 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) : 359 __constant_cpu_to_be32(IPATH_MULTICAST_QPN) :
463 cpu_to_be32(wr->wr.ud.remote_qpn); 360 cpu_to_be32(wqe->wr.wr.ud.remote_qpn);
464 /* XXX Could lose a PSN count but not worth locking */
465 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK); 361 ohdr->bth[2] = cpu_to_be32(qp->s_next_psn++ & IPATH_PSN_MASK);
466 /* 362 /*
467 * Qkeys with the high order bit set mean use the 363 * Qkeys with the high order bit set mean use the
468 * qkey from the QP context instead of the WR (see 10.2.5). 364 * qkey from the QP context instead of the WR (see 10.2.5).
469 */ 365 */
470 ohdr->u.ud.deth[0] = cpu_to_be32((int)wr->wr.ud.remote_qkey < 0 ? 366 ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->wr.wr.ud.remote_qkey < 0 ?
471 qp->qkey : wr->wr.ud.remote_qkey); 367 qp->qkey : wqe->wr.wr.ud.remote_qkey);
472 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num); 368 ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
473 if (ipath_verbs_send(dev->dd, hwords, (u32 *) &qp->s_hdr,
474 len, &ss))
475 dev->n_no_piobuf++;
476 369
477done: 370done:
478 /* Queue the completion status entry. */ 371 if (++qp->s_cur >= qp->s_size)
479 if (!(qp->s_flags & IPATH_S_SIGNAL_REQ_WR) || 372 qp->s_cur = 0;
480 (wr->send_flags & IB_SEND_SIGNALED)) { 373 ret = 1;
481 wc.wr_id = wr->wr_id;
482 wc.status = IB_WC_SUCCESS;
483 wc.vendor_err = 0;
484 wc.opcode = IB_WC_SEND;
485 wc.byte_len = len;
486 wc.qp = &qp->ibqp;
487 wc.src_qp = 0;
488 wc.wc_flags = 0;
489 /* XXX initialize other fields? */
490 ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
491 }
492 kfree(sg_list);
493
494 ret = 0;
495 374
496bail: 375bail:
497 return ret; 376 return ret;
@@ -673,6 +552,7 @@ void ipath_ud_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
673 */ 552 */
674 wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 : 553 wc.dlid_path_bits = dlid >= IPATH_MULTICAST_LID_BASE ? 0 :
675 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1); 554 dlid & ((1 << (dev->mkeyprot_resv_lmc & 7)) - 1);
555 wc.port_num = 1;
676 /* Signal completion event if the solicited bit is set. */ 556 /* Signal completion event if the solicited bit is set. */
677 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 557 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
678 (ohdr->bth[0] & 558 (ohdr->bth[0] &
diff --git a/drivers/infiniband/hw/ipath/ipath_verbs.c b/drivers/infiniband/hw/ipath/ipath_verbs.c
index 559d4a662937..3cc82b62b3c5 100644
--- a/drivers/infiniband/hw/ipath/ipath_verbs.c
+++ b/drivers/infiniband/hw/ipath/ipath_verbs.c
@@ -231,6 +231,103 @@ void ipath_skip_sge(struct ipath_sge_state *ss, u32 length)
231} 231}
232 232
233/** 233/**
234 * ipath_post_one_send - post one RC, UC, or UD send work request
235 * @qp: the QP to post on
236 * @wr: the work request to send
237 */
238static int ipath_post_one_send(struct ipath_qp *qp, struct ib_send_wr *wr)
239{
240 struct ipath_swqe *wqe;
241 u32 next;
242 int i;
243 int j;
244 int acc;
245 int ret;
246 unsigned long flags;
247
248 spin_lock_irqsave(&qp->s_lock, flags);
249
250 /* Check that state is OK to post send. */
251 if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK))
252 goto bail_inval;
253
254 /* IB spec says that num_sge == 0 is OK. */
255 if (wr->num_sge > qp->s_max_sge)
256 goto bail_inval;
257
258 /*
259 * Don't allow RDMA reads or atomic operations on UC or
260 * undefined operations.
261 * Make sure buffer is large enough to hold the result for atomics.
262 */
263 if (qp->ibqp.qp_type == IB_QPT_UC) {
264 if ((unsigned) wr->opcode >= IB_WR_RDMA_READ)
265 goto bail_inval;
266 } else if (qp->ibqp.qp_type == IB_QPT_UD) {
267 /* Check UD opcode */
268 if (wr->opcode != IB_WR_SEND &&
269 wr->opcode != IB_WR_SEND_WITH_IMM)
270 goto bail_inval;
271 /* Check UD destination address PD */
272 if (qp->ibqp.pd != wr->wr.ud.ah->pd)
273 goto bail_inval;
274 } else if ((unsigned) wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD)
275 goto bail_inval;
276 else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP &&
277 (wr->num_sge == 0 ||
278 wr->sg_list[0].length < sizeof(u64) ||
279 wr->sg_list[0].addr & (sizeof(u64) - 1)))
280 goto bail_inval;
281 else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic)
282 goto bail_inval;
283
284 next = qp->s_head + 1;
285 if (next >= qp->s_size)
286 next = 0;
287 if (next == qp->s_last)
288 goto bail_inval;
289
290 wqe = get_swqe_ptr(qp, qp->s_head);
291 wqe->wr = *wr;
292 wqe->ssn = qp->s_ssn++;
293 wqe->length = 0;
294 if (wr->num_sge) {
295 acc = wr->opcode >= IB_WR_RDMA_READ ?
296 IB_ACCESS_LOCAL_WRITE : 0;
297 for (i = 0, j = 0; i < wr->num_sge; i++) {
298 u32 length = wr->sg_list[i].length;
299 int ok;
300
301 if (length == 0)
302 continue;
303 ok = ipath_lkey_ok(qp, &wqe->sg_list[j],
304 &wr->sg_list[i], acc);
305 if (!ok)
306 goto bail_inval;
307 wqe->length += length;
308 j++;
309 }
310 wqe->wr.num_sge = j;
311 }
312 if (qp->ibqp.qp_type == IB_QPT_UC ||
313 qp->ibqp.qp_type == IB_QPT_RC) {
314 if (wqe->length > 0x80000000U)
315 goto bail_inval;
316 } else if (wqe->length > to_idev(qp->ibqp.device)->dd->ipath_ibmtu)
317 goto bail_inval;
318 qp->s_head = next;
319
320 ret = 0;
321 goto bail;
322
323bail_inval:
324 ret = -EINVAL;
325bail:
326 spin_unlock_irqrestore(&qp->s_lock, flags);
327 return ret;
328}
329
330/**
234 * ipath_post_send - post a send on a QP 331 * ipath_post_send - post a send on a QP
235 * @ibqp: the QP to post the send on 332 * @ibqp: the QP to post the send on
236 * @wr: the list of work requests to post 333 * @wr: the list of work requests to post
@@ -244,35 +341,17 @@ static int ipath_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
244 struct ipath_qp *qp = to_iqp(ibqp); 341 struct ipath_qp *qp = to_iqp(ibqp);
245 int err = 0; 342 int err = 0;
246 343
247 /* Check that state is OK to post send. */
248 if (!(ib_ipath_state_ops[qp->state] & IPATH_POST_SEND_OK)) {
249 *bad_wr = wr;
250 err = -EINVAL;
251 goto bail;
252 }
253
254 for (; wr; wr = wr->next) { 344 for (; wr; wr = wr->next) {
255 switch (qp->ibqp.qp_type) { 345 err = ipath_post_one_send(qp, wr);
256 case IB_QPT_UC:
257 case IB_QPT_RC:
258 err = ipath_post_ruc_send(qp, wr);
259 break;
260
261 case IB_QPT_SMI:
262 case IB_QPT_GSI:
263 case IB_QPT_UD:
264 err = ipath_post_ud_send(qp, wr);
265 break;
266
267 default:
268 err = -EINVAL;
269 }
270 if (err) { 346 if (err) {
271 *bad_wr = wr; 347 *bad_wr = wr;
272 break; 348 goto bail;
273 } 349 }
274 } 350 }
275 351
352 /* Try to do the send work in the caller's context. */
353 ipath_do_send((unsigned long) qp);
354
276bail: 355bail:
277 return err; 356 return err;
278} 357}
@@ -641,11 +720,11 @@ static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
641 u32 len = ss->sge.length; 720 u32 len = ss->sge.length;
642 u32 off; 721 u32 off;
643 722
644 BUG_ON(len == 0);
645 if (len > length) 723 if (len > length)
646 len = length; 724 len = length;
647 if (len > ss->sge.sge_length) 725 if (len > ss->sge.sge_length)
648 len = ss->sge.sge_length; 726 len = ss->sge.sge_length;
727 BUG_ON(len == 0);
649 /* If the source address is not aligned, try to align it. */ 728 /* If the source address is not aligned, try to align it. */
650 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1); 729 off = (unsigned long)ss->sge.vaddr & (sizeof(u32) - 1);
651 if (off) { 730 if (off) {
@@ -767,30 +846,15 @@ static void copy_io(u32 __iomem *piobuf, struct ipath_sge_state *ss,
767 __raw_writel(last, piobuf); 846 __raw_writel(last, piobuf);
768} 847}
769 848
770/** 849static int ipath_verbs_send_pio(struct ipath_qp *qp, u32 *hdr, u32 hdrwords,
771 * ipath_verbs_send - send a packet 850 struct ipath_sge_state *ss, u32 len,
772 * @dd: the infinipath device 851 u32 plen, u32 dwords)
773 * @hdrwords: the number of words in the header
774 * @hdr: the packet header
775 * @len: the length of the packet in bytes
776 * @ss: the SGE to send
777 */
778int ipath_verbs_send(struct ipath_devdata *dd, u32 hdrwords,
779 u32 *hdr, u32 len, struct ipath_sge_state *ss)
780{ 852{
853 struct ipath_devdata *dd = to_idev(qp->ibqp.device)->dd;
781 u32 __iomem *piobuf; 854 u32 __iomem *piobuf;
782 unsigned flush_wc; 855 unsigned flush_wc;
783 u32 plen;
784 int ret; 856 int ret;
785 857
786 /* +1 is for the qword padding of pbc */
787 plen = hdrwords + ((len + 3) >> 2) + 1;
788 if (unlikely((plen << 2) > dd->ipath_ibmaxlen)) {
789 ret = -EINVAL;
790 goto bail;
791 }
792
793 /* Get a PIO buffer to use. */
794 piobuf = ipath_getpiobuf(dd, NULL); 858 piobuf = ipath_getpiobuf(dd, NULL);
795 if (unlikely(piobuf == NULL)) { 859 if (unlikely(piobuf == NULL)) {
796 ret = -EBUSY; 860 ret = -EBUSY;
@@ -831,13 +895,10 @@ int ipath_verbs_send(struct ipath_devdata *dd, u32 hdrwords,
831 /* The common case is aligned and contained in one segment. */ 895 /* The common case is aligned and contained in one segment. */
832 if (likely(ss->num_sge == 1 && len <= ss->sge.length && 896 if (likely(ss->num_sge == 1 && len <= ss->sge.length &&
833 !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) { 897 !((unsigned long)ss->sge.vaddr & (sizeof(u32) - 1)))) {
834 u32 dwords;
835 u32 *addr = (u32 *) ss->sge.vaddr; 898 u32 *addr = (u32 *) ss->sge.vaddr;
836 899
837 /* Update address before sending packet. */ 900 /* Update address before sending packet. */
838 update_sge(ss, len); 901 update_sge(ss, len);
839 /* Need to round up for the last dword in the packet. */
840 dwords = (len + 3) >> 2;
841 if (flush_wc) { 902 if (flush_wc) {
842 __iowrite32_copy(piobuf, addr, dwords - 1); 903 __iowrite32_copy(piobuf, addr, dwords - 1);
843 /* must flush early everything before trigger word */ 904 /* must flush early everything before trigger word */
@@ -851,11 +912,37 @@ int ipath_verbs_send(struct ipath_devdata *dd, u32 hdrwords,
851 } 912 }
852 copy_io(piobuf, ss, len, flush_wc); 913 copy_io(piobuf, ss, len, flush_wc);
853done: 914done:
915 if (qp->s_wqe)
916 ipath_send_complete(qp, qp->s_wqe, IB_WC_SUCCESS);
854 ret = 0; 917 ret = 0;
855bail: 918bail:
856 return ret; 919 return ret;
857} 920}
858 921
922/**
923 * ipath_verbs_send - send a packet
924 * @qp: the QP to send on
925 * @hdr: the packet header
926 * @hdrwords: the number of words in the header
927 * @ss: the SGE to send
928 * @len: the length of the packet in bytes
929 */
930int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
931 u32 hdrwords, struct ipath_sge_state *ss, u32 len)
932{
933 u32 plen;
934 int ret;
935 u32 dwords = (len + 3) >> 2;
936
937 /* +1 is for the qword padding of pbc */
938 plen = hdrwords + dwords + 1;
939
940 ret = ipath_verbs_send_pio(qp, (u32 *) hdr, hdrwords,
941 ss, len, plen, dwords);
942
943 return ret;
944}
945
859int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords, 946int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
860 u64 *rwords, u64 *spkts, u64 *rpkts, 947 u64 *rwords, u64 *spkts, u64 *rpkts,
861 u64 *xmit_wait) 948 u64 *xmit_wait)
@@ -864,7 +951,6 @@ int ipath_snapshot_counters(struct ipath_devdata *dd, u64 *swords,
864 951
865 if (!(dd->ipath_flags & IPATH_INITTED)) { 952 if (!(dd->ipath_flags & IPATH_INITTED)) {
866 /* no hardware, freeze, etc. */ 953 /* no hardware, freeze, etc. */
867 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
868 ret = -EINVAL; 954 ret = -EINVAL;
869 goto bail; 955 goto bail;
870 } 956 }
@@ -890,48 +976,44 @@ bail:
890int ipath_get_counters(struct ipath_devdata *dd, 976int ipath_get_counters(struct ipath_devdata *dd,
891 struct ipath_verbs_counters *cntrs) 977 struct ipath_verbs_counters *cntrs)
892{ 978{
979 struct ipath_cregs const *crp = dd->ipath_cregs;
893 int ret; 980 int ret;
894 981
895 if (!(dd->ipath_flags & IPATH_INITTED)) { 982 if (!(dd->ipath_flags & IPATH_INITTED)) {
896 /* no hardware, freeze, etc. */ 983 /* no hardware, freeze, etc. */
897 ipath_dbg("unit %u not usable\n", dd->ipath_unit);
898 ret = -EINVAL; 984 ret = -EINVAL;
899 goto bail; 985 goto bail;
900 } 986 }
901 cntrs->symbol_error_counter = 987 cntrs->symbol_error_counter =
902 ipath_snap_cntr(dd, dd->ipath_cregs->cr_ibsymbolerrcnt); 988 ipath_snap_cntr(dd, crp->cr_ibsymbolerrcnt);
903 cntrs->link_error_recovery_counter = 989 cntrs->link_error_recovery_counter =
904 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkerrrecovcnt); 990 ipath_snap_cntr(dd, crp->cr_iblinkerrrecovcnt);
905 /* 991 /*
906 * The link downed counter counts when the other side downs the 992 * The link downed counter counts when the other side downs the
907 * connection. We add in the number of times we downed the link 993 * connection. We add in the number of times we downed the link
908 * due to local link integrity errors to compensate. 994 * due to local link integrity errors to compensate.
909 */ 995 */
910 cntrs->link_downed_counter = 996 cntrs->link_downed_counter =
911 ipath_snap_cntr(dd, dd->ipath_cregs->cr_iblinkdowncnt); 997 ipath_snap_cntr(dd, crp->cr_iblinkdowncnt);
912 cntrs->port_rcv_errors = 998 cntrs->port_rcv_errors =
913 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rxdroppktcnt) + 999 ipath_snap_cntr(dd, crp->cr_rxdroppktcnt) +
914 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvovflcnt) + 1000 ipath_snap_cntr(dd, crp->cr_rcvovflcnt) +
915 ipath_snap_cntr(dd, dd->ipath_cregs->cr_portovflcnt) + 1001 ipath_snap_cntr(dd, crp->cr_portovflcnt) +
916 ipath_snap_cntr(dd, dd->ipath_cregs->cr_err_rlencnt) + 1002 ipath_snap_cntr(dd, crp->cr_err_rlencnt) +
917 ipath_snap_cntr(dd, dd->ipath_cregs->cr_invalidrlencnt) + 1003 ipath_snap_cntr(dd, crp->cr_invalidrlencnt) +
918 ipath_snap_cntr(dd, dd->ipath_cregs->cr_erricrccnt) + 1004 ipath_snap_cntr(dd, crp->cr_errlinkcnt) +
919 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errvcrccnt) + 1005 ipath_snap_cntr(dd, crp->cr_erricrccnt) +
920 ipath_snap_cntr(dd, dd->ipath_cregs->cr_errlpcrccnt) + 1006 ipath_snap_cntr(dd, crp->cr_errvcrccnt) +
921 ipath_snap_cntr(dd, dd->ipath_cregs->cr_badformatcnt) + 1007 ipath_snap_cntr(dd, crp->cr_errlpcrccnt) +
1008 ipath_snap_cntr(dd, crp->cr_badformatcnt) +
922 dd->ipath_rxfc_unsupvl_errs; 1009 dd->ipath_rxfc_unsupvl_errs;
923 cntrs->port_rcv_remphys_errors = 1010 cntrs->port_rcv_remphys_errors =
924 ipath_snap_cntr(dd, dd->ipath_cregs->cr_rcvebpcnt); 1011 ipath_snap_cntr(dd, crp->cr_rcvebpcnt);
925 cntrs->port_xmit_discards = 1012 cntrs->port_xmit_discards = ipath_snap_cntr(dd, crp->cr_unsupvlcnt);
926 ipath_snap_cntr(dd, dd->ipath_cregs->cr_unsupvlcnt); 1013 cntrs->port_xmit_data = ipath_snap_cntr(dd, crp->cr_wordsendcnt);
927 cntrs->port_xmit_data = 1014 cntrs->port_rcv_data = ipath_snap_cntr(dd, crp->cr_wordrcvcnt);
928 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt); 1015 cntrs->port_xmit_packets = ipath_snap_cntr(dd, crp->cr_pktsendcnt);
929 cntrs->port_rcv_data = 1016 cntrs->port_rcv_packets = ipath_snap_cntr(dd, crp->cr_pktrcvcnt);
930 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
931 cntrs->port_xmit_packets =
932 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
933 cntrs->port_rcv_packets =
934 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
935 cntrs->local_link_integrity_errors = 1017 cntrs->local_link_integrity_errors =
936 (dd->ipath_flags & IPATH_GPIO_ERRINTRS) ? 1018 (dd->ipath_flags & IPATH_GPIO_ERRINTRS) ?
937 dd->ipath_lli_errs : dd->ipath_lli_errors; 1019 dd->ipath_lli_errs : dd->ipath_lli_errors;
@@ -1045,8 +1127,9 @@ static int ipath_query_port(struct ib_device *ibdev,
1045 u8 port, struct ib_port_attr *props) 1127 u8 port, struct ib_port_attr *props)
1046{ 1128{
1047 struct ipath_ibdev *dev = to_idev(ibdev); 1129 struct ipath_ibdev *dev = to_idev(ibdev);
1130 struct ipath_devdata *dd = dev->dd;
1048 enum ib_mtu mtu; 1131 enum ib_mtu mtu;
1049 u16 lid = dev->dd->ipath_lid; 1132 u16 lid = dd->ipath_lid;
1050 u64 ibcstat; 1133 u64 ibcstat;
1051 1134
1052 memset(props, 0, sizeof(*props)); 1135 memset(props, 0, sizeof(*props));
@@ -1054,16 +1137,16 @@ static int ipath_query_port(struct ib_device *ibdev,
1054 props->lmc = dev->mkeyprot_resv_lmc & 7; 1137 props->lmc = dev->mkeyprot_resv_lmc & 7;
1055 props->sm_lid = dev->sm_lid; 1138 props->sm_lid = dev->sm_lid;
1056 props->sm_sl = dev->sm_sl; 1139 props->sm_sl = dev->sm_sl;
1057 ibcstat = dev->dd->ipath_lastibcstat; 1140 ibcstat = dd->ipath_lastibcstat;
1058 props->state = ((ibcstat >> 4) & 0x3) + 1; 1141 props->state = ((ibcstat >> 4) & 0x3) + 1;
1059 /* See phys_state_show() */ 1142 /* See phys_state_show() */
1060 props->phys_state = ipath_cvt_physportstate[ 1143 props->phys_state = ipath_cvt_physportstate[
1061 dev->dd->ipath_lastibcstat & 0xf]; 1144 dd->ipath_lastibcstat & 0xf];
1062 props->port_cap_flags = dev->port_cap_flags; 1145 props->port_cap_flags = dev->port_cap_flags;
1063 props->gid_tbl_len = 1; 1146 props->gid_tbl_len = 1;
1064 props->max_msg_sz = 0x80000000; 1147 props->max_msg_sz = 0x80000000;
1065 props->pkey_tbl_len = ipath_get_npkeys(dev->dd); 1148 props->pkey_tbl_len = ipath_get_npkeys(dd);
1066 props->bad_pkey_cntr = ipath_get_cr_errpkey(dev->dd) - 1149 props->bad_pkey_cntr = ipath_get_cr_errpkey(dd) -
1067 dev->z_pkey_violations; 1150 dev->z_pkey_violations;
1068 props->qkey_viol_cntr = dev->qkey_violations; 1151 props->qkey_viol_cntr = dev->qkey_violations;
1069 props->active_width = IB_WIDTH_4X; 1152 props->active_width = IB_WIDTH_4X;
@@ -1073,12 +1156,12 @@ static int ipath_query_port(struct ib_device *ibdev,
1073 props->init_type_reply = 0; 1156 props->init_type_reply = 0;
1074 1157
1075 /* 1158 /*
1076 * Note: the chips support a maximum MTU of 4096, but the driver 1159 * Note: the chip supports a maximum MTU of 4096, but the driver
1077 * hasn't implemented this feature yet, so set the maximum value 1160 * hasn't implemented this feature yet, so set the maximum value
1078 * to 2048. 1161 * to 2048.
1079 */ 1162 */
1080 props->max_mtu = IB_MTU_2048; 1163 props->max_mtu = IB_MTU_2048;
1081 switch (dev->dd->ipath_ibmtu) { 1164 switch (dd->ipath_ibmtu) {
1082 case 4096: 1165 case 4096:
1083 mtu = IB_MTU_4096; 1166 mtu = IB_MTU_4096;
1084 break; 1167 break;
@@ -1427,9 +1510,7 @@ static int disable_timer(struct ipath_devdata *dd)
1427{ 1510{
1428 /* Disable GPIO bit 2 interrupt */ 1511 /* Disable GPIO bit 2 interrupt */
1429 if (dd->ipath_flags & IPATH_GPIO_INTR) { 1512 if (dd->ipath_flags & IPATH_GPIO_INTR) {
1430 u64 val;
1431 /* Disable GPIO bit 2 interrupt */ 1513 /* Disable GPIO bit 2 interrupt */
1432 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_gpio_mask);
1433 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT)); 1514 dd->ipath_gpio_mask &= ~((u64) (1 << IPATH_GPIO_PORT0_BIT));
1434 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask, 1515 ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_mask,
1435 dd->ipath_gpio_mask); 1516 dd->ipath_gpio_mask);
diff --git a/drivers/infiniband/hw/ipath/ipath_verbs.h b/drivers/infiniband/hw/ipath/ipath_verbs.h
index 1a24c6a4a814..619ad728b07b 100644
--- a/drivers/infiniband/hw/ipath/ipath_verbs.h
+++ b/drivers/infiniband/hw/ipath/ipath_verbs.h
@@ -42,6 +42,8 @@
42#include <rdma/ib_pack.h> 42#include <rdma/ib_pack.h>
43#include <rdma/ib_user_verbs.h> 43#include <rdma/ib_user_verbs.h>
44 44
45#include "ipath_kernel.h"
46
45#define IPATH_MAX_RDMA_ATOMIC 4 47#define IPATH_MAX_RDMA_ATOMIC 4
46 48
47#define QPN_MAX (1 << 24) 49#define QPN_MAX (1 << 24)
@@ -59,6 +61,7 @@
59 */ 61 */
60#define IB_CQ_NONE (IB_CQ_NEXT_COMP + 1) 62#define IB_CQ_NONE (IB_CQ_NEXT_COMP + 1)
61 63
64/* AETH NAK opcode values */
62#define IB_RNR_NAK 0x20 65#define IB_RNR_NAK 0x20
63#define IB_NAK_PSN_ERROR 0x60 66#define IB_NAK_PSN_ERROR 0x60
64#define IB_NAK_INVALID_REQUEST 0x61 67#define IB_NAK_INVALID_REQUEST 0x61
@@ -66,6 +69,7 @@
66#define IB_NAK_REMOTE_OPERATIONAL_ERROR 0x63 69#define IB_NAK_REMOTE_OPERATIONAL_ERROR 0x63
67#define IB_NAK_INVALID_RD_REQUEST 0x64 70#define IB_NAK_INVALID_RD_REQUEST 0x64
68 71
72/* Flags for checking QP state (see ib_ipath_state_ops[]) */
69#define IPATH_POST_SEND_OK 0x01 73#define IPATH_POST_SEND_OK 0x01
70#define IPATH_POST_RECV_OK 0x02 74#define IPATH_POST_RECV_OK 0x02
71#define IPATH_PROCESS_RECV_OK 0x04 75#define IPATH_PROCESS_RECV_OK 0x04
@@ -239,7 +243,7 @@ struct ipath_mregion {
239 */ 243 */
240struct ipath_sge { 244struct ipath_sge {
241 struct ipath_mregion *mr; 245 struct ipath_mregion *mr;
242 void *vaddr; /* current pointer into the segment */ 246 void *vaddr; /* kernel virtual address of segment */
243 u32 sge_length; /* length of the SGE */ 247 u32 sge_length; /* length of the SGE */
244 u32 length; /* remaining length of the segment */ 248 u32 length; /* remaining length of the segment */
245 u16 m; /* current index: mr->map[m] */ 249 u16 m; /* current index: mr->map[m] */
@@ -407,6 +411,7 @@ struct ipath_qp {
407 u32 s_ssn; /* SSN of tail entry */ 411 u32 s_ssn; /* SSN of tail entry */
408 u32 s_lsn; /* limit sequence number (credit) */ 412 u32 s_lsn; /* limit sequence number (credit) */
409 struct ipath_swqe *s_wq; /* send work queue */ 413 struct ipath_swqe *s_wq; /* send work queue */
414 struct ipath_swqe *s_wqe;
410 struct ipath_rq r_rq; /* receive work queue */ 415 struct ipath_rq r_rq; /* receive work queue */
411 struct ipath_sge r_sg_list[0]; /* verified SGEs */ 416 struct ipath_sge r_sg_list[0]; /* verified SGEs */
412}; 417};
@@ -683,8 +688,8 @@ void ipath_sqerror_qp(struct ipath_qp *qp, struct ib_wc *wc);
683 688
684void ipath_get_credit(struct ipath_qp *qp, u32 aeth); 689void ipath_get_credit(struct ipath_qp *qp, u32 aeth);
685 690
686int ipath_verbs_send(struct ipath_devdata *dd, u32 hdrwords, 691int ipath_verbs_send(struct ipath_qp *qp, struct ipath_ib_header *hdr,
687 u32 *hdr, u32 len, struct ipath_sge_state *ss); 692 u32 hdrwords, struct ipath_sge_state *ss, u32 len);
688 693
689void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int sig); 694void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int sig);
690 695
@@ -692,8 +697,6 @@ void ipath_copy_sge(struct ipath_sge_state *ss, void *data, u32 length);
692 697
693void ipath_skip_sge(struct ipath_sge_state *ss, u32 length); 698void ipath_skip_sge(struct ipath_sge_state *ss, u32 length);
694 699
695int ipath_post_ruc_send(struct ipath_qp *qp, struct ib_send_wr *wr);
696
697void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr, 700void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
698 int has_grh, void *data, u32 tlen, struct ipath_qp *qp); 701 int has_grh, void *data, u32 tlen, struct ipath_qp *qp);
699 702
@@ -733,6 +736,8 @@ int ipath_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr);
733 736
734int ipath_destroy_srq(struct ib_srq *ibsrq); 737int ipath_destroy_srq(struct ib_srq *ibsrq);
735 738
739void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int sig);
740
736int ipath_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry); 741int ipath_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry);
737 742
738struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries, int comp_vector, 743struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries, int comp_vector,
@@ -782,18 +787,28 @@ int ipath_mmap(struct ib_ucontext *context, struct vm_area_struct *vma);
782 787
783void ipath_insert_rnr_queue(struct ipath_qp *qp); 788void ipath_insert_rnr_queue(struct ipath_qp *qp);
784 789
790int ipath_init_sge(struct ipath_qp *qp, struct ipath_rwqe *wqe,
791 u32 *lengthp, struct ipath_sge_state *ss);
792
785int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only); 793int ipath_get_rwqe(struct ipath_qp *qp, int wr_id_only);
786 794
787u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr, 795u32 ipath_make_grh(struct ipath_ibdev *dev, struct ib_grh *hdr,
788 struct ib_global_route *grh, u32 hwords, u32 nwords); 796 struct ib_global_route *grh, u32 hwords, u32 nwords);
789 797
790void ipath_do_ruc_send(unsigned long data); 798void ipath_make_ruc_header(struct ipath_ibdev *dev, struct ipath_qp *qp,
799 struct ipath_other_headers *ohdr,
800 u32 bth0, u32 bth2);
801
802void ipath_do_send(unsigned long data);
803
804void ipath_send_complete(struct ipath_qp *qp, struct ipath_swqe *wqe,
805 enum ib_wc_status status);
806
807int ipath_make_rc_req(struct ipath_qp *qp);
791 808
792int ipath_make_rc_req(struct ipath_qp *qp, struct ipath_other_headers *ohdr, 809int ipath_make_uc_req(struct ipath_qp *qp);
793 u32 pmtu, u32 *bth0p, u32 *bth2p);
794 810
795int ipath_make_uc_req(struct ipath_qp *qp, struct ipath_other_headers *ohdr, 811int ipath_make_ud_req(struct ipath_qp *qp);
796 u32 pmtu, u32 *bth0p, u32 *bth2p);
797 812
798int ipath_register_ib_device(struct ipath_devdata *); 813int ipath_register_ib_device(struct ipath_devdata *);
799 814
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-28 18:41:06 -0500