1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 /*
3  * Copyright(c) 2015 - 2018 Intel Corporation.
4  */
5 
6 #include "hfi.h"
7 #include "verbs_txreq.h"
8 #include "qp.h"
9 
10 /* cut down ridiculously long IB macro names */
11 #define OP(x) UC_OP(x)
12 
13 /**
14  * hfi1_make_uc_req - construct a request packet (SEND, RDMA write)
15  * @qp: a pointer to the QP
16  * @ps: the current packet state
17  *
18  * Assume s_lock is held.
19  *
20  * Return 1 if constructed; otherwise, return 0.
21  */
hfi1_make_uc_req(struct rvt_qp * qp,struct hfi1_pkt_state * ps)22 int hfi1_make_uc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
23 {
24 	struct hfi1_qp_priv *priv = qp->priv;
25 	struct ib_other_headers *ohdr;
26 	struct rvt_swqe *wqe;
27 	u32 hwords;
28 	u32 bth0 = 0;
29 	u32 len;
30 	u32 pmtu = qp->pmtu;
31 	int middle = 0;
32 
33 	ps->s_txreq = get_txreq(ps->dev, qp);
34 	if (!ps->s_txreq)
35 		goto bail_no_tx;
36 
37 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
38 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
39 			goto bail;
40 		/* We are in the error state, flush the work request. */
41 		if (qp->s_last == READ_ONCE(qp->s_head))
42 			goto bail;
43 		/* If DMAs are in progress, we can't flush immediately. */
44 		if (iowait_sdma_pending(&priv->s_iowait)) {
45 			qp->s_flags |= RVT_S_WAIT_DMA;
46 			goto bail;
47 		}
48 		clear_ahg(qp);
49 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
50 		rvt_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
51 		goto done_free_tx;
52 	}
53 
54 	if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
55 		/* header size in 32-bit words LRH+BTH = (8+12)/4. */
56 		hwords = 5;
57 		if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
58 			ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
59 		else
60 			ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
61 	} else {
62 		/* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
63 		hwords = 7;
64 		if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
65 		    (hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))))
66 			ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
67 		else
68 			ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
69 	}
70 
71 	/* Get the next send request. */
72 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
73 	qp->s_wqe = NULL;
74 	switch (qp->s_state) {
75 	default:
76 		if (!(ib_rvt_state_ops[qp->state] &
77 		    RVT_PROCESS_NEXT_SEND_OK))
78 			goto bail;
79 		/* Check if send work queue is empty. */
80 		if (qp->s_cur == READ_ONCE(qp->s_head)) {
81 			clear_ahg(qp);
82 			goto bail;
83 		}
84 		/*
85 		 * Local operations are processed immediately
86 		 * after all prior requests have completed.
87 		 */
88 		if (wqe->wr.opcode == IB_WR_REG_MR ||
89 		    wqe->wr.opcode == IB_WR_LOCAL_INV) {
90 			int local_ops = 0;
91 			int err = 0;
92 
93 			if (qp->s_last != qp->s_cur)
94 				goto bail;
95 			if (++qp->s_cur == qp->s_size)
96 				qp->s_cur = 0;
97 			if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
98 				err = rvt_invalidate_rkey(
99 					qp, wqe->wr.ex.invalidate_rkey);
100 				local_ops = 1;
101 			}
102 			rvt_send_complete(qp, wqe, err ? IB_WC_LOC_PROT_ERR
103 							: IB_WC_SUCCESS);
104 			if (local_ops)
105 				atomic_dec(&qp->local_ops_pending);
106 			goto done_free_tx;
107 		}
108 		/*
109 		 * Start a new request.
110 		 */
111 		qp->s_psn = wqe->psn;
112 		qp->s_sge.sge = wqe->sg_list[0];
113 		qp->s_sge.sg_list = wqe->sg_list + 1;
114 		qp->s_sge.num_sge = wqe->wr.num_sge;
115 		qp->s_sge.total_len = wqe->length;
116 		len = wqe->length;
117 		qp->s_len = len;
118 		switch (wqe->wr.opcode) {
119 		case IB_WR_SEND:
120 		case IB_WR_SEND_WITH_IMM:
121 			if (len > pmtu) {
122 				qp->s_state = OP(SEND_FIRST);
123 				len = pmtu;
124 				break;
125 			}
126 			if (wqe->wr.opcode == IB_WR_SEND) {
127 				qp->s_state = OP(SEND_ONLY);
128 			} else {
129 				qp->s_state =
130 					OP(SEND_ONLY_WITH_IMMEDIATE);
131 				/* Immediate data comes after the BTH */
132 				ohdr->u.imm_data = wqe->wr.ex.imm_data;
133 				hwords += 1;
134 			}
135 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
136 				bth0 |= IB_BTH_SOLICITED;
137 			qp->s_wqe = wqe;
138 			if (++qp->s_cur >= qp->s_size)
139 				qp->s_cur = 0;
140 			break;
141 
142 		case IB_WR_RDMA_WRITE:
143 		case IB_WR_RDMA_WRITE_WITH_IMM:
144 			ohdr->u.rc.reth.vaddr =
145 				cpu_to_be64(wqe->rdma_wr.remote_addr);
146 			ohdr->u.rc.reth.rkey =
147 				cpu_to_be32(wqe->rdma_wr.rkey);
148 			ohdr->u.rc.reth.length = cpu_to_be32(len);
149 			hwords += sizeof(struct ib_reth) / 4;
150 			if (len > pmtu) {
151 				qp->s_state = OP(RDMA_WRITE_FIRST);
152 				len = pmtu;
153 				break;
154 			}
155 			if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
156 				qp->s_state = OP(RDMA_WRITE_ONLY);
157 			} else {
158 				qp->s_state =
159 					OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
160 				/* Immediate data comes after the RETH */
161 				ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
162 				hwords += 1;
163 				if (wqe->wr.send_flags & IB_SEND_SOLICITED)
164 					bth0 |= IB_BTH_SOLICITED;
165 			}
166 			qp->s_wqe = wqe;
167 			if (++qp->s_cur >= qp->s_size)
168 				qp->s_cur = 0;
169 			break;
170 
171 		default:
172 			goto bail;
173 		}
174 		break;
175 
176 	case OP(SEND_FIRST):
177 		qp->s_state = OP(SEND_MIDDLE);
178 		fallthrough;
179 	case OP(SEND_MIDDLE):
180 		len = qp->s_len;
181 		if (len > pmtu) {
182 			len = pmtu;
183 			middle = HFI1_CAP_IS_KSET(SDMA_AHG);
184 			break;
185 		}
186 		if (wqe->wr.opcode == IB_WR_SEND) {
187 			qp->s_state = OP(SEND_LAST);
188 		} else {
189 			qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
190 			/* Immediate data comes after the BTH */
191 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
192 			hwords += 1;
193 		}
194 		if (wqe->wr.send_flags & IB_SEND_SOLICITED)
195 			bth0 |= IB_BTH_SOLICITED;
196 		qp->s_wqe = wqe;
197 		if (++qp->s_cur >= qp->s_size)
198 			qp->s_cur = 0;
199 		break;
200 
201 	case OP(RDMA_WRITE_FIRST):
202 		qp->s_state = OP(RDMA_WRITE_MIDDLE);
203 		fallthrough;
204 	case OP(RDMA_WRITE_MIDDLE):
205 		len = qp->s_len;
206 		if (len > pmtu) {
207 			len = pmtu;
208 			middle = HFI1_CAP_IS_KSET(SDMA_AHG);
209 			break;
210 		}
211 		if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
212 			qp->s_state = OP(RDMA_WRITE_LAST);
213 		} else {
214 			qp->s_state =
215 				OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
216 			/* Immediate data comes after the BTH */
217 			ohdr->u.imm_data = wqe->wr.ex.imm_data;
218 			hwords += 1;
219 			if (wqe->wr.send_flags & IB_SEND_SOLICITED)
220 				bth0 |= IB_BTH_SOLICITED;
221 		}
222 		qp->s_wqe = wqe;
223 		if (++qp->s_cur >= qp->s_size)
224 			qp->s_cur = 0;
225 		break;
226 	}
227 	qp->s_len -= len;
228 	ps->s_txreq->hdr_dwords = hwords;
229 	ps->s_txreq->sde = priv->s_sde;
230 	ps->s_txreq->ss = &qp->s_sge;
231 	ps->s_txreq->s_cur_size = len;
232 	hfi1_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24),
233 			     qp->remote_qpn, mask_psn(qp->s_psn++),
234 			     middle, ps);
235 	return 1;
236 
237 done_free_tx:
238 	hfi1_put_txreq(ps->s_txreq);
239 	ps->s_txreq = NULL;
240 	return 1;
241 
242 bail:
243 	hfi1_put_txreq(ps->s_txreq);
244 
245 bail_no_tx:
246 	ps->s_txreq = NULL;
247 	qp->s_flags &= ~RVT_S_BUSY;
248 	return 0;
249 }
250 
251 /**
252  * hfi1_uc_rcv - handle an incoming UC packet
253  * @packet: the packet structure
254  *
255  * This is called from qp_rcv() to process an incoming UC packet
256  * for the given QP.
257  * Called at interrupt level.
258  */
hfi1_uc_rcv(struct hfi1_packet * packet)259 void hfi1_uc_rcv(struct hfi1_packet *packet)
260 {
261 	struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
262 	void *data = packet->payload;
263 	u32 tlen = packet->tlen;
264 	struct rvt_qp *qp = packet->qp;
265 	struct ib_other_headers *ohdr = packet->ohdr;
266 	u32 opcode = packet->opcode;
267 	u32 hdrsize = packet->hlen;
268 	u32 psn;
269 	u32 pad = packet->pad;
270 	struct ib_wc wc;
271 	u32 pmtu = qp->pmtu;
272 	struct ib_reth *reth;
273 	int ret;
274 	u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
275 
276 	if (hfi1_ruc_check_hdr(ibp, packet))
277 		return;
278 
279 	process_ecn(qp, packet);
280 
281 	psn = ib_bth_get_psn(ohdr);
282 	/* Compare the PSN verses the expected PSN. */
283 	if (unlikely(cmp_psn(psn, qp->r_psn) != 0)) {
284 		/*
285 		 * Handle a sequence error.
286 		 * Silently drop any current message.
287 		 */
288 		qp->r_psn = psn;
289 inv:
290 		if (qp->r_state == OP(SEND_FIRST) ||
291 		    qp->r_state == OP(SEND_MIDDLE)) {
292 			set_bit(RVT_R_REWIND_SGE, &qp->r_aflags);
293 			qp->r_sge.num_sge = 0;
294 		} else {
295 			rvt_put_ss(&qp->r_sge);
296 		}
297 		qp->r_state = OP(SEND_LAST);
298 		switch (opcode) {
299 		case OP(SEND_FIRST):
300 		case OP(SEND_ONLY):
301 		case OP(SEND_ONLY_WITH_IMMEDIATE):
302 			goto send_first;
303 
304 		case OP(RDMA_WRITE_FIRST):
305 		case OP(RDMA_WRITE_ONLY):
306 		case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
307 			goto rdma_first;
308 
309 		default:
310 			goto drop;
311 		}
312 	}
313 
314 	/* Check for opcode sequence errors. */
315 	switch (qp->r_state) {
316 	case OP(SEND_FIRST):
317 	case OP(SEND_MIDDLE):
318 		if (opcode == OP(SEND_MIDDLE) ||
319 		    opcode == OP(SEND_LAST) ||
320 		    opcode == OP(SEND_LAST_WITH_IMMEDIATE))
321 			break;
322 		goto inv;
323 
324 	case OP(RDMA_WRITE_FIRST):
325 	case OP(RDMA_WRITE_MIDDLE):
326 		if (opcode == OP(RDMA_WRITE_MIDDLE) ||
327 		    opcode == OP(RDMA_WRITE_LAST) ||
328 		    opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
329 			break;
330 		goto inv;
331 
332 	default:
333 		if (opcode == OP(SEND_FIRST) ||
334 		    opcode == OP(SEND_ONLY) ||
335 		    opcode == OP(SEND_ONLY_WITH_IMMEDIATE) ||
336 		    opcode == OP(RDMA_WRITE_FIRST) ||
337 		    opcode == OP(RDMA_WRITE_ONLY) ||
338 		    opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
339 			break;
340 		goto inv;
341 	}
342 
343 	if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
344 		rvt_comm_est(qp);
345 
346 	/* OK, process the packet. */
347 	switch (opcode) {
348 	case OP(SEND_FIRST):
349 	case OP(SEND_ONLY):
350 	case OP(SEND_ONLY_WITH_IMMEDIATE):
351 send_first:
352 		if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) {
353 			qp->r_sge = qp->s_rdma_read_sge;
354 		} else {
355 			ret = rvt_get_rwqe(qp, false);
356 			if (ret < 0)
357 				goto op_err;
358 			if (!ret)
359 				goto drop;
360 			/*
361 			 * qp->s_rdma_read_sge will be the owner
362 			 * of the mr references.
363 			 */
364 			qp->s_rdma_read_sge = qp->r_sge;
365 		}
366 		qp->r_rcv_len = 0;
367 		if (opcode == OP(SEND_ONLY))
368 			goto no_immediate_data;
369 		else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE))
370 			goto send_last_imm;
371 		fallthrough;
372 	case OP(SEND_MIDDLE):
373 		/* Check for invalid length PMTU or posted rwqe len. */
374 		/*
375 		 * There will be no padding for 9B packet but 16B packets
376 		 * will come in with some padding since we always add
377 		 * CRC and LT bytes which will need to be flit aligned
378 		 */
379 		if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
380 			goto rewind;
381 		qp->r_rcv_len += pmtu;
382 		if (unlikely(qp->r_rcv_len > qp->r_len))
383 			goto rewind;
384 		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, false, false);
385 		break;
386 
387 	case OP(SEND_LAST_WITH_IMMEDIATE):
388 send_last_imm:
389 		wc.ex.imm_data = ohdr->u.imm_data;
390 		wc.wc_flags = IB_WC_WITH_IMM;
391 		goto send_last;
392 	case OP(SEND_LAST):
393 no_immediate_data:
394 		wc.ex.imm_data = 0;
395 		wc.wc_flags = 0;
396 send_last:
397 		/* Check for invalid length. */
398 		/* LAST len should be >= 1 */
399 		if (unlikely(tlen < (hdrsize + extra_bytes)))
400 			goto rewind;
401 		/* Don't count the CRC. */
402 		tlen -= (hdrsize + extra_bytes);
403 		wc.byte_len = tlen + qp->r_rcv_len;
404 		if (unlikely(wc.byte_len > qp->r_len))
405 			goto rewind;
406 		wc.opcode = IB_WC_RECV;
407 		rvt_copy_sge(qp, &qp->r_sge, data, tlen, false, false);
408 		rvt_put_ss(&qp->s_rdma_read_sge);
409 last_imm:
410 		wc.wr_id = qp->r_wr_id;
411 		wc.status = IB_WC_SUCCESS;
412 		wc.qp = &qp->ibqp;
413 		wc.src_qp = qp->remote_qpn;
414 		wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
415 		/*
416 		 * It seems that IB mandates the presence of an SL in a
417 		 * work completion only for the UD transport (see section
418 		 * 11.4.2 of IBTA Vol. 1).
419 		 *
420 		 * However, the way the SL is chosen below is consistent
421 		 * with the way that IB/qib works and is trying avoid
422 		 * introducing incompatibilities.
423 		 *
424 		 * See also OPA Vol. 1, section 9.7.6, and table 9-17.
425 		 */
426 		wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
427 		/* zero fields that are N/A */
428 		wc.vendor_err = 0;
429 		wc.pkey_index = 0;
430 		wc.dlid_path_bits = 0;
431 		wc.port_num = 0;
432 		/* Signal completion event if the solicited bit is set. */
433 		rvt_recv_cq(qp, &wc, ib_bth_is_solicited(ohdr));
434 		break;
435 
436 	case OP(RDMA_WRITE_FIRST):
437 	case OP(RDMA_WRITE_ONLY):
438 	case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): /* consume RWQE */
439 rdma_first:
440 		if (unlikely(!(qp->qp_access_flags &
441 			       IB_ACCESS_REMOTE_WRITE))) {
442 			goto drop;
443 		}
444 		reth = &ohdr->u.rc.reth;
445 		qp->r_len = be32_to_cpu(reth->length);
446 		qp->r_rcv_len = 0;
447 		qp->r_sge.sg_list = NULL;
448 		if (qp->r_len != 0) {
449 			u32 rkey = be32_to_cpu(reth->rkey);
450 			u64 vaddr = be64_to_cpu(reth->vaddr);
451 			int ok;
452 
453 			/* Check rkey */
454 			ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len,
455 					 vaddr, rkey, IB_ACCESS_REMOTE_WRITE);
456 			if (unlikely(!ok))
457 				goto drop;
458 			qp->r_sge.num_sge = 1;
459 		} else {
460 			qp->r_sge.num_sge = 0;
461 			qp->r_sge.sge.mr = NULL;
462 			qp->r_sge.sge.vaddr = NULL;
463 			qp->r_sge.sge.length = 0;
464 			qp->r_sge.sge.sge_length = 0;
465 		}
466 		if (opcode == OP(RDMA_WRITE_ONLY)) {
467 			goto rdma_last;
468 		} else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) {
469 			wc.ex.imm_data = ohdr->u.rc.imm_data;
470 			goto rdma_last_imm;
471 		}
472 		fallthrough;
473 	case OP(RDMA_WRITE_MIDDLE):
474 		/* Check for invalid length PMTU or posted rwqe len. */
475 		if (unlikely(tlen != (hdrsize + pmtu + 4)))
476 			goto drop;
477 		qp->r_rcv_len += pmtu;
478 		if (unlikely(qp->r_rcv_len > qp->r_len))
479 			goto drop;
480 		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
481 		break;
482 
483 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
484 		wc.ex.imm_data = ohdr->u.imm_data;
485 rdma_last_imm:
486 		wc.wc_flags = IB_WC_WITH_IMM;
487 
488 		/* Check for invalid length. */
489 		/* LAST len should be >= 1 */
490 		if (unlikely(tlen < (hdrsize + pad + 4)))
491 			goto drop;
492 		/* Don't count the CRC. */
493 		tlen -= (hdrsize + extra_bytes);
494 		if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
495 			goto drop;
496 		if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) {
497 			rvt_put_ss(&qp->s_rdma_read_sge);
498 		} else {
499 			ret = rvt_get_rwqe(qp, true);
500 			if (ret < 0)
501 				goto op_err;
502 			if (!ret)
503 				goto drop;
504 		}
505 		wc.byte_len = qp->r_len;
506 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
507 		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
508 		rvt_put_ss(&qp->r_sge);
509 		goto last_imm;
510 
511 	case OP(RDMA_WRITE_LAST):
512 rdma_last:
513 		/* Check for invalid length. */
514 		/* LAST len should be >= 1 */
515 		if (unlikely(tlen < (hdrsize + pad + 4)))
516 			goto drop;
517 		/* Don't count the CRC. */
518 		tlen -= (hdrsize + extra_bytes);
519 		if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
520 			goto drop;
521 		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
522 		rvt_put_ss(&qp->r_sge);
523 		break;
524 
525 	default:
526 		/* Drop packet for unknown opcodes. */
527 		goto drop;
528 	}
529 	qp->r_psn++;
530 	qp->r_state = opcode;
531 	return;
532 
533 rewind:
534 	set_bit(RVT_R_REWIND_SGE, &qp->r_aflags);
535 	qp->r_sge.num_sge = 0;
536 drop:
537 	ibp->rvp.n_pkt_drops++;
538 	return;
539 
540 op_err:
541 	rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
542 }
543