2 * Copyright (c) 2006 QLogic, Inc. 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 "ipath_verbs.h"
35 #include "ipath_common.h"
37 /* cut down ridiculously long IB macro names */
38 #define OP(x) IB_OPCODE_UC_##x
40 static void complete_last_send(struct ipath_qp *qp, struct ipath_swqe *wqe,
43 if (++qp->s_last == qp->s_size)
45 if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
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];
51 wc->byte_len = wqe->length;
52 wc->qp_num = qp->ibqp.qp_num;
53 wc->src_qp = qp->remote_qpn;
55 wc->slid = qp->remote_ah_attr.dlid;
56 wc->sl = qp->remote_ah_attr.sl;
57 wc->dlid_path_bits = 0;
59 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 0);
61 wqe = get_swqe_ptr(qp, qp->s_last);
65 * ipath_make_uc_req - construct a request packet (SEND, RDMA write)
66 * @qp: a pointer to the QP
67 * @ohdr: a pointer to the IB header being constructed
69 * @bth0p: pointer to the BTH opcode word
70 * @bth2p: pointer to the BTH PSN word
72 * Return 1 if constructed; otherwise, return 0.
73 * Note the QP s_lock must be held and interrupts disabled.
75 int ipath_make_uc_req(struct ipath_qp *qp,
76 struct ipath_other_headers *ohdr,
77 u32 pmtu, u32 *bth0p, u32 *bth2p)
79 struct ipath_swqe *wqe;
85 if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK))
88 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
92 /* Get the next send request. */
93 wqe = get_swqe_ptr(qp, qp->s_last);
94 switch (qp->s_state) {
97 * Signal the completion of the last send
100 if (qp->s_last != qp->s_tail)
101 complete_last_send(qp, wqe, &wc);
103 /* Check if send work queue is empty. */
104 if (qp->s_tail == qp->s_head)
107 * Start a new request.
109 qp->s_psn = wqe->psn = qp->s_next_psn;
110 qp->s_sge.sge = wqe->sg_list[0];
111 qp->s_sge.sg_list = wqe->sg_list + 1;
112 qp->s_sge.num_sge = wqe->wr.num_sge;
113 qp->s_len = len = wqe->length;
114 switch (wqe->wr.opcode) {
116 case IB_WR_SEND_WITH_IMM:
118 qp->s_state = OP(SEND_FIRST);
122 if (wqe->wr.opcode == IB_WR_SEND)
123 qp->s_state = OP(SEND_ONLY);
126 OP(SEND_ONLY_WITH_IMMEDIATE);
127 /* Immediate data comes after the BTH */
128 ohdr->u.imm_data = wqe->wr.imm_data;
131 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
135 case IB_WR_RDMA_WRITE:
136 case IB_WR_RDMA_WRITE_WITH_IMM:
137 ohdr->u.rc.reth.vaddr =
138 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
139 ohdr->u.rc.reth.rkey =
140 cpu_to_be32(wqe->wr.wr.rdma.rkey);
141 ohdr->u.rc.reth.length = cpu_to_be32(len);
142 hwords += sizeof(struct ib_reth) / 4;
144 qp->s_state = OP(RDMA_WRITE_FIRST);
148 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
149 qp->s_state = OP(RDMA_WRITE_ONLY);
152 OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
153 /* Immediate data comes after the RETH */
154 ohdr->u.rc.imm_data = wqe->wr.imm_data;
156 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
164 if (++qp->s_tail >= qp->s_size)
169 qp->s_state = OP(SEND_MIDDLE);
171 case OP(SEND_MIDDLE):
177 if (wqe->wr.opcode == IB_WR_SEND)
178 qp->s_state = OP(SEND_LAST);
180 qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
181 /* Immediate data comes after the BTH */
182 ohdr->u.imm_data = wqe->wr.imm_data;
185 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
189 case OP(RDMA_WRITE_FIRST):
190 qp->s_state = OP(RDMA_WRITE_MIDDLE);
192 case OP(RDMA_WRITE_MIDDLE):
198 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
199 qp->s_state = OP(RDMA_WRITE_LAST);
202 OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
203 /* Immediate data comes after the BTH */
204 ohdr->u.imm_data = wqe->wr.imm_data;
206 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
212 qp->s_hdrwords = hwords;
213 qp->s_cur_sge = &qp->s_sge;
214 qp->s_cur_size = len;
215 *bth0p = bth0 | (qp->s_state << 24);
216 *bth2p = qp->s_next_psn++ & IPATH_PSN_MASK;
224 * ipath_uc_rcv - handle an incoming UC packet
225 * @dev: the device the packet came in on
226 * @hdr: the header of the packet
227 * @has_grh: true if the packet has a GRH
228 * @data: the packet data
229 * @tlen: the length of the packet
230 * @qp: the QP for this packet.
232 * This is called from ipath_qp_rcv() to process an incoming UC packet
234 * Called at interrupt level.
236 void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
237 int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
239 struct ipath_other_headers *ohdr;
245 u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
246 struct ib_reth *reth;
252 hdrsize = 8 + 12; /* LRH + BTH */
253 psn = be32_to_cpu(ohdr->bth[2]);
256 ohdr = &hdr->u.l.oth;
257 hdrsize = 8 + 40 + 12; /* LRH + GRH + BTH */
259 * The header with GRH is 60 bytes and the
260 * core driver sets the eager header buffer
261 * size to 56 bytes so the last 4 bytes of
262 * the BTH header (PSN) is in the data buffer.
265 ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
266 if (header_in_data) {
267 psn = be32_to_cpu(((__be32 *) data)[0]);
268 data += sizeof(__be32);
270 psn = be32_to_cpu(ohdr->bth[2]);
273 * The opcode is in the low byte when its in network order
274 * (top byte when in host order).
276 opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
281 /* Compare the PSN verses the expected PSN. */
282 if (unlikely(ipath_cmp24(psn, qp->r_psn) != 0)) {
284 * Handle a sequence error.
285 * Silently drop any current message.
289 qp->r_state = OP(SEND_LAST);
293 case OP(SEND_ONLY_WITH_IMMEDIATE):
296 case OP(RDMA_WRITE_FIRST):
297 case OP(RDMA_WRITE_ONLY):
298 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
307 /* Check for opcode sequence errors. */
308 switch (qp->r_state) {
310 case OP(SEND_MIDDLE):
311 if (opcode == OP(SEND_MIDDLE) ||
312 opcode == OP(SEND_LAST) ||
313 opcode == OP(SEND_LAST_WITH_IMMEDIATE))
317 case OP(RDMA_WRITE_FIRST):
318 case OP(RDMA_WRITE_MIDDLE):
319 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
320 opcode == OP(RDMA_WRITE_LAST) ||
321 opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
326 if (opcode == OP(SEND_FIRST) ||
327 opcode == OP(SEND_ONLY) ||
328 opcode == OP(SEND_ONLY_WITH_IMMEDIATE) ||
329 opcode == OP(RDMA_WRITE_FIRST) ||
330 opcode == OP(RDMA_WRITE_ONLY) ||
331 opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
336 /* OK, process the packet. */
340 case OP(SEND_ONLY_WITH_IMMEDIATE):
342 if (qp->r_reuse_sge) {
344 qp->r_sge = qp->s_rdma_sge;
345 } else if (!ipath_get_rwqe(qp, 0)) {
349 /* Save the WQE so we can reuse it in case of an error. */
350 qp->s_rdma_sge = qp->r_sge;
352 if (opcode == OP(SEND_ONLY))
354 else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE))
357 case OP(SEND_MIDDLE):
358 /* Check for invalid length PMTU or posted rwqe len. */
359 if (unlikely(tlen != (hdrsize + pmtu + 4))) {
364 qp->r_rcv_len += pmtu;
365 if (unlikely(qp->r_rcv_len > qp->r_len)) {
370 ipath_copy_sge(&qp->r_sge, data, pmtu);
373 case OP(SEND_LAST_WITH_IMMEDIATE):
375 if (header_in_data) {
376 wc.imm_data = *(__be32 *) data;
377 data += sizeof(__be32);
379 /* Immediate data comes after BTH */
380 wc.imm_data = ohdr->u.imm_data;
383 wc.wc_flags = IB_WC_WITH_IMM;
387 /* Get the number of bytes the message was padded by. */
388 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
389 /* Check for invalid length. */
390 /* XXX LAST len should be >= 1 */
391 if (unlikely(tlen < (hdrsize + pad + 4))) {
396 /* Don't count the CRC. */
397 tlen -= (hdrsize + pad + 4);
398 wc.byte_len = tlen + qp->r_rcv_len;
399 if (unlikely(wc.byte_len > qp->r_len)) {
404 /* XXX Need to free SGEs */
406 ipath_copy_sge(&qp->r_sge, data, tlen);
407 wc.wr_id = qp->r_wr_id;
408 wc.status = IB_WC_SUCCESS;
409 wc.opcode = IB_WC_RECV;
411 wc.qp_num = qp->ibqp.qp_num;
412 wc.src_qp = qp->remote_qpn;
414 wc.slid = qp->remote_ah_attr.dlid;
415 wc.sl = qp->remote_ah_attr.sl;
416 wc.dlid_path_bits = 0;
418 /* Signal completion event if the solicited bit is set. */
419 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
421 __constant_cpu_to_be32(1 << 23)) != 0);
424 case OP(RDMA_WRITE_FIRST):
425 case OP(RDMA_WRITE_ONLY):
426 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): /* consume RWQE */
428 /* RETH comes after BTH */
430 reth = &ohdr->u.rc.reth;
432 reth = (struct ib_reth *)data;
433 data += sizeof(*reth);
435 hdrsize += sizeof(*reth);
436 qp->r_len = be32_to_cpu(reth->length);
438 if (qp->r_len != 0) {
439 u32 rkey = be32_to_cpu(reth->rkey);
440 u64 vaddr = be64_to_cpu(reth->vaddr);
444 ok = ipath_rkey_ok(dev, &qp->r_sge, qp->r_len,
446 IB_ACCESS_REMOTE_WRITE);
452 qp->r_sge.sg_list = NULL;
453 qp->r_sge.sge.mr = NULL;
454 qp->r_sge.sge.vaddr = NULL;
455 qp->r_sge.sge.length = 0;
456 qp->r_sge.sge.sge_length = 0;
458 if (unlikely(!(qp->qp_access_flags &
459 IB_ACCESS_REMOTE_WRITE))) {
463 if (opcode == OP(RDMA_WRITE_ONLY))
465 else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
468 case OP(RDMA_WRITE_MIDDLE):
469 /* Check for invalid length PMTU or posted rwqe len. */
470 if (unlikely(tlen != (hdrsize + pmtu + 4))) {
474 qp->r_rcv_len += pmtu;
475 if (unlikely(qp->r_rcv_len > qp->r_len)) {
479 ipath_copy_sge(&qp->r_sge, data, pmtu);
482 case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
484 /* Get the number of bytes the message was padded by. */
485 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
486 /* Check for invalid length. */
487 /* XXX LAST len should be >= 1 */
488 if (unlikely(tlen < (hdrsize + pad + 4))) {
492 /* Don't count the CRC. */
493 tlen -= (hdrsize + pad + 4);
494 if (unlikely(tlen + qp->r_rcv_len != qp->r_len)) {
500 else if (!ipath_get_rwqe(qp, 1)) {
504 if (header_in_data) {
505 wc.imm_data = *(__be32 *) data;
506 data += sizeof(__be32);
508 /* Immediate data comes after BTH */
509 wc.imm_data = ohdr->u.imm_data;
512 wc.wc_flags = IB_WC_WITH_IMM;
516 case OP(RDMA_WRITE_LAST):
518 /* Get the number of bytes the message was padded by. */
519 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
520 /* Check for invalid length. */
521 /* XXX LAST len should be >= 1 */
522 if (unlikely(tlen < (hdrsize + pad + 4))) {
526 /* Don't count the CRC. */
527 tlen -= (hdrsize + pad + 4);
528 if (unlikely(tlen + qp->r_rcv_len != qp->r_len)) {
532 ipath_copy_sge(&qp->r_sge, data, tlen);
536 /* Drop packet for unknown opcodes. */
541 qp->r_state = opcode;