2 * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <rdma/ib_smi.h>
35 #include <rdma/ib_verbs.h>
41 * qib_ud_loopback - handle send on loopback QPs
42 * @sqp: the sending QP
43 * @swqe: the send work request
45 * This is called from qib_make_ud_req() to forward a WQE addressed
47 * Note that the receive interrupt handler may be calling qib_ud_rcv()
48 * while this is being called.
50 static void qib_ud_loopback(struct rvt_qp
*sqp
, struct rvt_swqe
*swqe
)
52 struct qib_ibport
*ibp
= to_iport(sqp
->ibqp
.device
, sqp
->port_num
);
53 struct qib_pportdata
*ppd
= ppd_from_ibp(ibp
);
54 struct qib_devdata
*dd
= ppd
->dd
;
55 struct rvt_dev_info
*rdi
= &dd
->verbs_dev
.rdi
;
57 struct rdma_ah_attr
*ah_attr
;
59 struct rvt_sge_state ssge
;
63 enum ib_qp_type sqptype
, dqptype
;
66 qp
= rvt_lookup_qpn(rdi
, &ibp
->rvp
, swqe
->ud_wr
.remote_qpn
);
68 ibp
->rvp
.n_pkt_drops
++;
72 sqptype
= sqp
->ibqp
.qp_type
== IB_QPT_GSI
?
73 IB_QPT_UD
: sqp
->ibqp
.qp_type
;
74 dqptype
= qp
->ibqp
.qp_type
== IB_QPT_GSI
?
75 IB_QPT_UD
: qp
->ibqp
.qp_type
;
77 if (dqptype
!= sqptype
||
78 !(ib_rvt_state_ops
[qp
->state
] & RVT_PROCESS_RECV_OK
)) {
79 ibp
->rvp
.n_pkt_drops
++;
83 ah_attr
= &ibah_to_rvtah(swqe
->ud_wr
.ah
)->attr
;
84 ppd
= ppd_from_ibp(ibp
);
86 if (qp
->ibqp
.qp_num
> 1) {
91 pkey1
= qib_get_pkey(ibp
, sqp
->s_pkey_index
);
92 pkey2
= qib_get_pkey(ibp
, qp
->s_pkey_index
);
93 if (unlikely(!qib_pkey_ok(pkey1
, pkey2
))) {
94 lid
= ppd
->lid
| (rdma_ah_get_path_bits(ah_attr
) &
95 ((1 << ppd
->lmc
) - 1));
96 qib_bad_pkey(ibp
, pkey1
,
97 rdma_ah_get_sl(ah_attr
),
98 sqp
->ibqp
.qp_num
, qp
->ibqp
.qp_num
,
100 cpu_to_be16(rdma_ah_get_dlid(ah_attr
)));
106 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
107 * Qkeys with the high order bit set mean use the
108 * qkey from the QP context instead of the WR (see 10.2.5).
110 if (qp
->ibqp
.qp_num
) {
113 qkey
= (int)swqe
->ud_wr
.remote_qkey
< 0 ?
114 sqp
->qkey
: swqe
->ud_wr
.remote_qkey
;
115 if (unlikely(qkey
!= qp
->qkey
))
120 * A GRH is expected to precede the data even if not
121 * present on the wire.
123 length
= swqe
->length
;
124 memset(&wc
, 0, sizeof(wc
));
125 wc
.byte_len
= length
+ sizeof(struct ib_grh
);
127 if (swqe
->wr
.opcode
== IB_WR_SEND_WITH_IMM
) {
128 wc
.wc_flags
= IB_WC_WITH_IMM
;
129 wc
.ex
.imm_data
= swqe
->wr
.ex
.imm_data
;
132 spin_lock_irqsave(&qp
->r_lock
, flags
);
135 * Get the next work request entry to find where to put the data.
137 if (qp
->r_flags
& RVT_R_REUSE_SGE
)
138 qp
->r_flags
&= ~RVT_R_REUSE_SGE
;
142 ret
= qib_get_rwqe(qp
, 0);
144 rvt_rc_error(qp
, IB_WC_LOC_QP_OP_ERR
);
148 if (qp
->ibqp
.qp_num
== 0)
149 ibp
->rvp
.n_vl15_dropped
++;
153 /* Silently drop packets which are too big. */
154 if (unlikely(wc
.byte_len
> qp
->r_len
)) {
155 qp
->r_flags
|= RVT_R_REUSE_SGE
;
156 ibp
->rvp
.n_pkt_drops
++;
160 if (rdma_ah_get_ah_flags(ah_attr
) & IB_AH_GRH
) {
162 const struct ib_global_route
*grd
= rdma_ah_read_grh(ah_attr
);
164 qib_make_grh(ibp
, &grh
, grd
, 0, 0);
165 qib_copy_sge(&qp
->r_sge
, &grh
,
167 wc
.wc_flags
|= IB_WC_GRH
;
169 rvt_skip_sge(&qp
->r_sge
, sizeof(struct ib_grh
), true);
170 ssge
.sg_list
= swqe
->sg_list
+ 1;
171 ssge
.sge
= *swqe
->sg_list
;
172 ssge
.num_sge
= swqe
->wr
.num_sge
;
175 u32 len
= sge
->length
;
179 if (len
> sge
->sge_length
)
180 len
= sge
->sge_length
;
182 qib_copy_sge(&qp
->r_sge
, sge
->vaddr
, len
, 1);
185 sge
->sge_length
-= len
;
186 if (sge
->sge_length
== 0) {
188 *sge
= *ssge
.sg_list
++;
189 } else if (sge
->length
== 0 && sge
->mr
->lkey
) {
190 if (++sge
->n
>= RVT_SEGSZ
) {
191 if (++sge
->m
>= sge
->mr
->mapsz
)
196 sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].vaddr
;
198 sge
->mr
->map
[sge
->m
]->segs
[sge
->n
].length
;
202 rvt_put_ss(&qp
->r_sge
);
203 if (!test_and_clear_bit(RVT_R_WRID_VALID
, &qp
->r_aflags
))
205 wc
.wr_id
= qp
->r_wr_id
;
206 wc
.status
= IB_WC_SUCCESS
;
207 wc
.opcode
= IB_WC_RECV
;
209 wc
.src_qp
= sqp
->ibqp
.qp_num
;
210 wc
.pkey_index
= qp
->ibqp
.qp_type
== IB_QPT_GSI
?
211 swqe
->ud_wr
.pkey_index
: 0;
212 wc
.slid
= ppd
->lid
| (rdma_ah_get_path_bits(ah_attr
) &
213 ((1 << ppd
->lmc
) - 1));
214 wc
.sl
= rdma_ah_get_sl(ah_attr
);
215 wc
.dlid_path_bits
= rdma_ah_get_dlid(ah_attr
) & ((1 << ppd
->lmc
) - 1);
216 wc
.port_num
= qp
->port_num
;
217 /* Signal completion event if the solicited bit is set. */
218 rvt_cq_enter(ibcq_to_rvtcq(qp
->ibqp
.recv_cq
), &wc
,
219 swqe
->wr
.send_flags
& IB_SEND_SOLICITED
);
220 ibp
->rvp
.n_loop_pkts
++;
222 spin_unlock_irqrestore(&qp
->r_lock
, flags
);
228 * qib_make_ud_req - construct a UD request packet
231 * Assumes the s_lock is held.
233 * Return 1 if constructed; otherwise, return 0.
235 int qib_make_ud_req(struct rvt_qp
*qp
, unsigned long *flags
)
237 struct qib_qp_priv
*priv
= qp
->priv
;
238 struct ib_other_headers
*ohdr
;
239 struct rdma_ah_attr
*ah_attr
;
240 struct qib_pportdata
*ppd
;
241 struct qib_ibport
*ibp
;
242 struct rvt_swqe
*wqe
;
251 if (!(ib_rvt_state_ops
[qp
->state
] & RVT_PROCESS_NEXT_SEND_OK
)) {
252 if (!(ib_rvt_state_ops
[qp
->state
] & RVT_FLUSH_SEND
))
254 /* We are in the error state, flush the work request. */
255 if (qp
->s_last
== READ_ONCE(qp
->s_head
))
257 /* If DMAs are in progress, we can't flush immediately. */
258 if (atomic_read(&priv
->s_dma_busy
)) {
259 qp
->s_flags
|= RVT_S_WAIT_DMA
;
262 wqe
= rvt_get_swqe_ptr(qp
, qp
->s_last
);
263 qib_send_complete(qp
, wqe
, IB_WC_WR_FLUSH_ERR
);
267 /* see post_one_send() */
268 if (qp
->s_cur
== READ_ONCE(qp
->s_head
))
271 wqe
= rvt_get_swqe_ptr(qp
, qp
->s_cur
);
272 next_cur
= qp
->s_cur
+ 1;
273 if (next_cur
>= qp
->s_size
)
276 /* Construct the header. */
277 ibp
= to_iport(qp
->ibqp
.device
, qp
->port_num
);
278 ppd
= ppd_from_ibp(ibp
);
279 ah_attr
= &ibah_to_rvtah(wqe
->ud_wr
.ah
)->attr
;
280 if (rdma_ah_get_dlid(ah_attr
) >= be16_to_cpu(IB_MULTICAST_LID_BASE
)) {
281 if (rdma_ah_get_dlid(ah_attr
) !=
282 be16_to_cpu(IB_LID_PERMISSIVE
))
283 this_cpu_inc(ibp
->pmastats
->n_multicast_xmit
);
285 this_cpu_inc(ibp
->pmastats
->n_unicast_xmit
);
287 this_cpu_inc(ibp
->pmastats
->n_unicast_xmit
);
288 lid
= rdma_ah_get_dlid(ah_attr
) & ~((1 << ppd
->lmc
) - 1);
289 if (unlikely(lid
== ppd
->lid
)) {
290 unsigned long tflags
= *flags
;
292 * If DMAs are in progress, we can't generate
293 * a completion for the loopback packet since
294 * it would be out of order.
295 * XXX Instead of waiting, we could queue a
296 * zero length descriptor so we get a callback.
298 if (atomic_read(&priv
->s_dma_busy
)) {
299 qp
->s_flags
|= RVT_S_WAIT_DMA
;
302 qp
->s_cur
= next_cur
;
303 spin_unlock_irqrestore(&qp
->s_lock
, tflags
);
304 qib_ud_loopback(qp
, wqe
);
305 spin_lock_irqsave(&qp
->s_lock
, tflags
);
307 qib_send_complete(qp
, wqe
, IB_WC_SUCCESS
);
312 qp
->s_cur
= next_cur
;
313 extra_bytes
= -wqe
->length
& 3;
314 nwords
= (wqe
->length
+ extra_bytes
) >> 2;
316 /* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
318 qp
->s_cur_size
= wqe
->length
;
319 qp
->s_cur_sge
= &qp
->s_sge
;
320 qp
->s_srate
= rdma_ah_get_static_rate(ah_attr
);
322 qp
->s_sge
.sge
= wqe
->sg_list
[0];
323 qp
->s_sge
.sg_list
= wqe
->sg_list
+ 1;
324 qp
->s_sge
.num_sge
= wqe
->wr
.num_sge
;
325 qp
->s_sge
.total_len
= wqe
->length
;
327 if (rdma_ah_get_ah_flags(ah_attr
) & IB_AH_GRH
) {
328 /* Header size in 32-bit words. */
329 qp
->s_hdrwords
+= qib_make_grh(ibp
, &priv
->s_hdr
->u
.l
.grh
,
330 rdma_ah_read_grh(ah_attr
),
331 qp
->s_hdrwords
, nwords
);
333 ohdr
= &priv
->s_hdr
->u
.l
.oth
;
335 * Don't worry about sending to locally attached multicast
336 * QPs. It is unspecified by the spec. what happens.
339 /* Header size in 32-bit words. */
341 ohdr
= &priv
->s_hdr
->u
.oth
;
343 if (wqe
->wr
.opcode
== IB_WR_SEND_WITH_IMM
) {
345 ohdr
->u
.ud
.imm_data
= wqe
->wr
.ex
.imm_data
;
346 bth0
= IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE
<< 24;
348 bth0
= IB_OPCODE_UD_SEND_ONLY
<< 24;
349 lrh0
|= rdma_ah_get_sl(ah_attr
) << 4;
350 if (qp
->ibqp
.qp_type
== IB_QPT_SMI
)
351 lrh0
|= 0xF000; /* Set VL (see ch. 13.5.3.1) */
353 lrh0
|= ibp
->sl_to_vl
[rdma_ah_get_sl(ah_attr
)] << 12;
354 priv
->s_hdr
->lrh
[0] = cpu_to_be16(lrh0
);
355 priv
->s_hdr
->lrh
[1] =
356 cpu_to_be16(rdma_ah_get_dlid(ah_attr
)); /* DEST LID */
357 priv
->s_hdr
->lrh
[2] =
358 cpu_to_be16(qp
->s_hdrwords
+ nwords
+ SIZE_OF_CRC
);
361 lid
|= rdma_ah_get_path_bits(ah_attr
) &
362 ((1 << ppd
->lmc
) - 1);
363 priv
->s_hdr
->lrh
[3] = cpu_to_be16(lid
);
365 priv
->s_hdr
->lrh
[3] = IB_LID_PERMISSIVE
;
366 if (wqe
->wr
.send_flags
& IB_SEND_SOLICITED
)
367 bth0
|= IB_BTH_SOLICITED
;
368 bth0
|= extra_bytes
<< 20;
369 bth0
|= qp
->ibqp
.qp_type
== IB_QPT_SMI
? QIB_DEFAULT_P_KEY
:
370 qib_get_pkey(ibp
, qp
->ibqp
.qp_type
== IB_QPT_GSI
?
371 wqe
->ud_wr
.pkey_index
: qp
->s_pkey_index
);
372 ohdr
->bth
[0] = cpu_to_be32(bth0
);
374 * Use the multicast QP if the destination LID is a multicast LID.
376 ohdr
->bth
[1] = rdma_ah_get_dlid(ah_attr
) >=
377 be16_to_cpu(IB_MULTICAST_LID_BASE
) &&
378 rdma_ah_get_dlid(ah_attr
) != be16_to_cpu(IB_LID_PERMISSIVE
) ?
379 cpu_to_be32(QIB_MULTICAST_QPN
) :
380 cpu_to_be32(wqe
->ud_wr
.remote_qpn
);
381 ohdr
->bth
[2] = cpu_to_be32(wqe
->psn
& QIB_PSN_MASK
);
383 * Qkeys with the high order bit set mean use the
384 * qkey from the QP context instead of the WR (see 10.2.5).
386 ohdr
->u
.ud
.deth
[0] = cpu_to_be32((int)wqe
->ud_wr
.remote_qkey
< 0 ?
387 qp
->qkey
: wqe
->ud_wr
.remote_qkey
);
388 ohdr
->u
.ud
.deth
[1] = cpu_to_be32(qp
->ibqp
.qp_num
);
393 qp
->s_flags
&= ~RVT_S_BUSY
;
397 static unsigned qib_lookup_pkey(struct qib_ibport
*ibp
, u16 pkey
)
399 struct qib_pportdata
*ppd
= ppd_from_ibp(ibp
);
400 struct qib_devdata
*dd
= ppd
->dd
;
401 unsigned ctxt
= ppd
->hw_pidx
;
404 pkey
&= 0x7fff; /* remove limited/full membership bit */
406 for (i
= 0; i
< ARRAY_SIZE(dd
->rcd
[ctxt
]->pkeys
); ++i
)
407 if ((dd
->rcd
[ctxt
]->pkeys
[i
] & 0x7fff) == pkey
)
411 * Should not get here, this means hardware failed to validate pkeys.
412 * Punt and return index 0.
418 * qib_ud_rcv - receive an incoming UD packet
419 * @ibp: the port the packet came in on
420 * @hdr: the packet header
421 * @has_grh: true if the packet has a GRH
422 * @data: the packet data
423 * @tlen: the packet length
424 * @qp: the QP the packet came on
426 * This is called from qib_qp_rcv() to process an incoming UD packet
428 * Called at interrupt level.
430 void qib_ud_rcv(struct qib_ibport
*ibp
, struct ib_header
*hdr
,
431 int has_grh
, void *data
, u32 tlen
, struct rvt_qp
*qp
)
433 struct ib_other_headers
*ohdr
;
445 hdrsize
= 8 + 12 + 8; /* LRH + BTH + DETH */
447 ohdr
= &hdr
->u
.l
.oth
;
448 hdrsize
= 8 + 40 + 12 + 8; /* LRH + GRH + BTH + DETH */
450 qkey
= be32_to_cpu(ohdr
->u
.ud
.deth
[0]);
451 src_qp
= be32_to_cpu(ohdr
->u
.ud
.deth
[1]) & RVT_QPN_MASK
;
454 * Get the number of bytes the message was padded by
455 * and drop incomplete packets.
457 pad
= (be32_to_cpu(ohdr
->bth
[0]) >> 20) & 3;
458 if (unlikely(tlen
< (hdrsize
+ pad
+ 4)))
461 tlen
-= hdrsize
+ pad
+ 4;
464 * Check that the permissive LID is only used on QP0
465 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
467 if (qp
->ibqp
.qp_num
) {
468 if (unlikely(hdr
->lrh
[1] == IB_LID_PERMISSIVE
||
469 hdr
->lrh
[3] == IB_LID_PERMISSIVE
))
471 if (qp
->ibqp
.qp_num
> 1) {
474 pkey1
= be32_to_cpu(ohdr
->bth
[0]);
475 pkey2
= qib_get_pkey(ibp
, qp
->s_pkey_index
);
476 if (unlikely(!qib_pkey_ok(pkey1
, pkey2
))) {
479 (be16_to_cpu(hdr
->lrh
[0]) >> 4) &
481 src_qp
, qp
->ibqp
.qp_num
,
482 hdr
->lrh
[3], hdr
->lrh
[1]);
486 if (unlikely(qkey
!= qp
->qkey
))
489 /* Drop invalid MAD packets (see 13.5.3.1). */
490 if (unlikely(qp
->ibqp
.qp_num
== 1 &&
492 (be16_to_cpu(hdr
->lrh
[0]) >> 12) == 15)))
497 /* Drop invalid MAD packets (see 13.5.3.1). */
498 if (tlen
!= 256 || (be16_to_cpu(hdr
->lrh
[0]) >> 12) != 15)
500 smp
= (struct ib_smp
*) data
;
501 if ((hdr
->lrh
[1] == IB_LID_PERMISSIVE
||
502 hdr
->lrh
[3] == IB_LID_PERMISSIVE
) &&
503 smp
->mgmt_class
!= IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
)
508 * The opcode is in the low byte when its in network order
509 * (top byte when in host order).
511 opcode
= be32_to_cpu(ohdr
->bth
[0]) >> 24;
512 if (qp
->ibqp
.qp_num
> 1 &&
513 opcode
== IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE
) {
514 wc
.ex
.imm_data
= ohdr
->u
.ud
.imm_data
;
515 wc
.wc_flags
= IB_WC_WITH_IMM
;
517 } else if (opcode
== IB_OPCODE_UD_SEND_ONLY
) {
524 * A GRH is expected to precede the data even if not
525 * present on the wire.
527 wc
.byte_len
= tlen
+ sizeof(struct ib_grh
);
530 * Get the next work request entry to find where to put the data.
532 if (qp
->r_flags
& RVT_R_REUSE_SGE
)
533 qp
->r_flags
&= ~RVT_R_REUSE_SGE
;
537 ret
= qib_get_rwqe(qp
, 0);
539 rvt_rc_error(qp
, IB_WC_LOC_QP_OP_ERR
);
543 if (qp
->ibqp
.qp_num
== 0)
544 ibp
->rvp
.n_vl15_dropped
++;
548 /* Silently drop packets which are too big. */
549 if (unlikely(wc
.byte_len
> qp
->r_len
)) {
550 qp
->r_flags
|= RVT_R_REUSE_SGE
;
554 qib_copy_sge(&qp
->r_sge
, &hdr
->u
.l
.grh
,
555 sizeof(struct ib_grh
), 1);
556 wc
.wc_flags
|= IB_WC_GRH
;
558 rvt_skip_sge(&qp
->r_sge
, sizeof(struct ib_grh
), true);
559 qib_copy_sge(&qp
->r_sge
, data
, wc
.byte_len
- sizeof(struct ib_grh
), 1);
560 rvt_put_ss(&qp
->r_sge
);
561 if (!test_and_clear_bit(RVT_R_WRID_VALID
, &qp
->r_aflags
))
563 wc
.wr_id
= qp
->r_wr_id
;
564 wc
.status
= IB_WC_SUCCESS
;
565 wc
.opcode
= IB_WC_RECV
;
569 wc
.pkey_index
= qp
->ibqp
.qp_type
== IB_QPT_GSI
?
570 qib_lookup_pkey(ibp
, be32_to_cpu(ohdr
->bth
[0])) : 0;
571 wc
.slid
= be16_to_cpu(hdr
->lrh
[3]);
572 wc
.sl
= (be16_to_cpu(hdr
->lrh
[0]) >> 4) & 0xF;
573 dlid
= be16_to_cpu(hdr
->lrh
[1]);
575 * Save the LMC lower bits if the destination LID is a unicast LID.
577 wc
.dlid_path_bits
= dlid
>= be16_to_cpu(IB_MULTICAST_LID_BASE
) ? 0 :
578 dlid
& ((1 << ppd_from_ibp(ibp
)->lmc
) - 1);
579 wc
.port_num
= qp
->port_num
;
580 /* Signal completion event if the solicited bit is set. */
581 rvt_cq_enter(ibcq_to_rvtcq(qp
->ibqp
.recv_cq
), &wc
,
582 ib_bth_is_solicited(ohdr
));
586 ibp
->rvp
.n_pkt_drops
++;