1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5  */
6 
7 #include <linux/skbuff.h>
8 
9 #include "rxe.h"
10 #include "rxe_loc.h"
11 #include "rxe_queue.h"
12 
13 static char *resp_state_name[] = {
14 	[RESPST_NONE]				= "NONE",
15 	[RESPST_GET_REQ]			= "GET_REQ",
16 	[RESPST_CHK_PSN]			= "CHK_PSN",
17 	[RESPST_CHK_OP_SEQ]			= "CHK_OP_SEQ",
18 	[RESPST_CHK_OP_VALID]			= "CHK_OP_VALID",
19 	[RESPST_CHK_RESOURCE]			= "CHK_RESOURCE",
20 	[RESPST_CHK_LENGTH]			= "CHK_LENGTH",
21 	[RESPST_CHK_RKEY]			= "CHK_RKEY",
22 	[RESPST_EXECUTE]			= "EXECUTE",
23 	[RESPST_READ_REPLY]			= "READ_REPLY",
24 	[RESPST_ATOMIC_REPLY]			= "ATOMIC_REPLY",
25 	[RESPST_ATOMIC_WRITE_REPLY]		= "ATOMIC_WRITE_REPLY",
26 	[RESPST_PROCESS_FLUSH]			= "PROCESS_FLUSH",
27 	[RESPST_COMPLETE]			= "COMPLETE",
28 	[RESPST_ACKNOWLEDGE]			= "ACKNOWLEDGE",
29 	[RESPST_CLEANUP]			= "CLEANUP",
30 	[RESPST_DUPLICATE_REQUEST]		= "DUPLICATE_REQUEST",
31 	[RESPST_ERR_MALFORMED_WQE]		= "ERR_MALFORMED_WQE",
32 	[RESPST_ERR_UNSUPPORTED_OPCODE]		= "ERR_UNSUPPORTED_OPCODE",
33 	[RESPST_ERR_MISALIGNED_ATOMIC]		= "ERR_MISALIGNED_ATOMIC",
34 	[RESPST_ERR_PSN_OUT_OF_SEQ]		= "ERR_PSN_OUT_OF_SEQ",
35 	[RESPST_ERR_MISSING_OPCODE_FIRST]	= "ERR_MISSING_OPCODE_FIRST",
36 	[RESPST_ERR_MISSING_OPCODE_LAST_C]	= "ERR_MISSING_OPCODE_LAST_C",
37 	[RESPST_ERR_MISSING_OPCODE_LAST_D1E]	= "ERR_MISSING_OPCODE_LAST_D1E",
38 	[RESPST_ERR_TOO_MANY_RDMA_ATM_REQ]	= "ERR_TOO_MANY_RDMA_ATM_REQ",
39 	[RESPST_ERR_RNR]			= "ERR_RNR",
40 	[RESPST_ERR_RKEY_VIOLATION]		= "ERR_RKEY_VIOLATION",
41 	[RESPST_ERR_INVALIDATE_RKEY]		= "ERR_INVALIDATE_RKEY_VIOLATION",
42 	[RESPST_ERR_LENGTH]			= "ERR_LENGTH",
43 	[RESPST_ERR_CQ_OVERFLOW]		= "ERR_CQ_OVERFLOW",
44 	[RESPST_ERROR]				= "ERROR",
45 	[RESPST_DONE]				= "DONE",
46 	[RESPST_EXIT]				= "EXIT",
47 };
48 
49 /* rxe_recv calls here to add a request packet to the input queue */
50 void rxe_resp_queue_pkt(struct rxe_qp *qp, struct sk_buff *skb)
51 {
52 	int must_sched;
53 	struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
54 
55 	skb_queue_tail(&qp->req_pkts, skb);
56 
57 	must_sched = (pkt->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST) ||
58 			(skb_queue_len(&qp->req_pkts) > 1);
59 
60 	if (must_sched)
61 		rxe_sched_task(&qp->resp.task);
62 	else
63 		rxe_run_task(&qp->resp.task);
64 }
65 
66 static inline enum resp_states get_req(struct rxe_qp *qp,
67 				       struct rxe_pkt_info **pkt_p)
68 {
69 	struct sk_buff *skb;
70 
71 	skb = skb_peek(&qp->req_pkts);
72 	if (!skb)
73 		return RESPST_EXIT;
74 
75 	*pkt_p = SKB_TO_PKT(skb);
76 
77 	return (qp->resp.res) ? RESPST_READ_REPLY : RESPST_CHK_PSN;
78 }
79 
80 static enum resp_states check_psn(struct rxe_qp *qp,
81 				  struct rxe_pkt_info *pkt)
82 {
83 	int diff = psn_compare(pkt->psn, qp->resp.psn);
84 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
85 
86 	switch (qp_type(qp)) {
87 	case IB_QPT_RC:
88 		if (diff > 0) {
89 			if (qp->resp.sent_psn_nak)
90 				return RESPST_CLEANUP;
91 
92 			qp->resp.sent_psn_nak = 1;
93 			rxe_counter_inc(rxe, RXE_CNT_OUT_OF_SEQ_REQ);
94 			return RESPST_ERR_PSN_OUT_OF_SEQ;
95 
96 		} else if (diff < 0) {
97 			rxe_counter_inc(rxe, RXE_CNT_DUP_REQ);
98 			return RESPST_DUPLICATE_REQUEST;
99 		}
100 
101 		if (qp->resp.sent_psn_nak)
102 			qp->resp.sent_psn_nak = 0;
103 
104 		break;
105 
106 	case IB_QPT_UC:
107 		if (qp->resp.drop_msg || diff != 0) {
108 			if (pkt->mask & RXE_START_MASK) {
109 				qp->resp.drop_msg = 0;
110 				return RESPST_CHK_OP_SEQ;
111 			}
112 
113 			qp->resp.drop_msg = 1;
114 			return RESPST_CLEANUP;
115 		}
116 		break;
117 	default:
118 		break;
119 	}
120 
121 	return RESPST_CHK_OP_SEQ;
122 }
123 
124 static enum resp_states check_op_seq(struct rxe_qp *qp,
125 				     struct rxe_pkt_info *pkt)
126 {
127 	switch (qp_type(qp)) {
128 	case IB_QPT_RC:
129 		switch (qp->resp.opcode) {
130 		case IB_OPCODE_RC_SEND_FIRST:
131 		case IB_OPCODE_RC_SEND_MIDDLE:
132 			switch (pkt->opcode) {
133 			case IB_OPCODE_RC_SEND_MIDDLE:
134 			case IB_OPCODE_RC_SEND_LAST:
135 			case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
136 			case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
137 				return RESPST_CHK_OP_VALID;
138 			default:
139 				return RESPST_ERR_MISSING_OPCODE_LAST_C;
140 			}
141 
142 		case IB_OPCODE_RC_RDMA_WRITE_FIRST:
143 		case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
144 			switch (pkt->opcode) {
145 			case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
146 			case IB_OPCODE_RC_RDMA_WRITE_LAST:
147 			case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
148 				return RESPST_CHK_OP_VALID;
149 			default:
150 				return RESPST_ERR_MISSING_OPCODE_LAST_C;
151 			}
152 
153 		default:
154 			switch (pkt->opcode) {
155 			case IB_OPCODE_RC_SEND_MIDDLE:
156 			case IB_OPCODE_RC_SEND_LAST:
157 			case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
158 			case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
159 			case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
160 			case IB_OPCODE_RC_RDMA_WRITE_LAST:
161 			case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
162 				return RESPST_ERR_MISSING_OPCODE_FIRST;
163 			default:
164 				return RESPST_CHK_OP_VALID;
165 			}
166 		}
167 		break;
168 
169 	case IB_QPT_UC:
170 		switch (qp->resp.opcode) {
171 		case IB_OPCODE_UC_SEND_FIRST:
172 		case IB_OPCODE_UC_SEND_MIDDLE:
173 			switch (pkt->opcode) {
174 			case IB_OPCODE_UC_SEND_MIDDLE:
175 			case IB_OPCODE_UC_SEND_LAST:
176 			case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
177 				return RESPST_CHK_OP_VALID;
178 			default:
179 				return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
180 			}
181 
182 		case IB_OPCODE_UC_RDMA_WRITE_FIRST:
183 		case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
184 			switch (pkt->opcode) {
185 			case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
186 			case IB_OPCODE_UC_RDMA_WRITE_LAST:
187 			case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
188 				return RESPST_CHK_OP_VALID;
189 			default:
190 				return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
191 			}
192 
193 		default:
194 			switch (pkt->opcode) {
195 			case IB_OPCODE_UC_SEND_MIDDLE:
196 			case IB_OPCODE_UC_SEND_LAST:
197 			case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
198 			case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
199 			case IB_OPCODE_UC_RDMA_WRITE_LAST:
200 			case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
201 				qp->resp.drop_msg = 1;
202 				return RESPST_CLEANUP;
203 			default:
204 				return RESPST_CHK_OP_VALID;
205 			}
206 		}
207 		break;
208 
209 	default:
210 		return RESPST_CHK_OP_VALID;
211 	}
212 }
213 
214 static bool check_qp_attr_access(struct rxe_qp *qp,
215 				 struct rxe_pkt_info *pkt)
216 {
217 	if (((pkt->mask & RXE_READ_MASK) &&
218 	     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) ||
219 	    ((pkt->mask & (RXE_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) &&
220 	     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) ||
221 	    ((pkt->mask & RXE_ATOMIC_MASK) &&
222 	     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
223 		return false;
224 
225 	if (pkt->mask & RXE_FLUSH_MASK) {
226 		u32 flush_type = feth_plt(pkt);
227 
228 		if ((flush_type & IB_FLUSH_GLOBAL &&
229 		     !(qp->attr.qp_access_flags & IB_ACCESS_FLUSH_GLOBAL)) ||
230 		    (flush_type & IB_FLUSH_PERSISTENT &&
231 		     !(qp->attr.qp_access_flags & IB_ACCESS_FLUSH_PERSISTENT)))
232 			return false;
233 	}
234 
235 	return true;
236 }
237 
238 static enum resp_states check_op_valid(struct rxe_qp *qp,
239 				       struct rxe_pkt_info *pkt)
240 {
241 	switch (qp_type(qp)) {
242 	case IB_QPT_RC:
243 		if (!check_qp_attr_access(qp, pkt))
244 			return RESPST_ERR_UNSUPPORTED_OPCODE;
245 
246 		break;
247 
248 	case IB_QPT_UC:
249 		if ((pkt->mask & RXE_WRITE_MASK) &&
250 		    !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) {
251 			qp->resp.drop_msg = 1;
252 			return RESPST_CLEANUP;
253 		}
254 
255 		break;
256 
257 	case IB_QPT_UD:
258 	case IB_QPT_GSI:
259 		break;
260 
261 	default:
262 		WARN_ON_ONCE(1);
263 		break;
264 	}
265 
266 	return RESPST_CHK_RESOURCE;
267 }
268 
269 static enum resp_states get_srq_wqe(struct rxe_qp *qp)
270 {
271 	struct rxe_srq *srq = qp->srq;
272 	struct rxe_queue *q = srq->rq.queue;
273 	struct rxe_recv_wqe *wqe;
274 	struct ib_event ev;
275 	unsigned int count;
276 	size_t size;
277 	unsigned long flags;
278 
279 	if (srq->error)
280 		return RESPST_ERR_RNR;
281 
282 	spin_lock_irqsave(&srq->rq.consumer_lock, flags);
283 
284 	wqe = queue_head(q, QUEUE_TYPE_FROM_CLIENT);
285 	if (!wqe) {
286 		spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
287 		return RESPST_ERR_RNR;
288 	}
289 
290 	/* don't trust user space data */
291 	if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) {
292 		spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
293 		rxe_dbg_qp(qp, "invalid num_sge in SRQ entry\n");
294 		return RESPST_ERR_MALFORMED_WQE;
295 	}
296 	size = sizeof(*wqe) + wqe->dma.num_sge*sizeof(struct rxe_sge);
297 	memcpy(&qp->resp.srq_wqe, wqe, size);
298 
299 	qp->resp.wqe = &qp->resp.srq_wqe.wqe;
300 	queue_advance_consumer(q, QUEUE_TYPE_FROM_CLIENT);
301 	count = queue_count(q, QUEUE_TYPE_FROM_CLIENT);
302 
303 	if (srq->limit && srq->ibsrq.event_handler && (count < srq->limit)) {
304 		srq->limit = 0;
305 		goto event;
306 	}
307 
308 	spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
309 	return RESPST_CHK_LENGTH;
310 
311 event:
312 	spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
313 	ev.device = qp->ibqp.device;
314 	ev.element.srq = qp->ibqp.srq;
315 	ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
316 	srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context);
317 	return RESPST_CHK_LENGTH;
318 }
319 
320 static enum resp_states check_resource(struct rxe_qp *qp,
321 				       struct rxe_pkt_info *pkt)
322 {
323 	struct rxe_srq *srq = qp->srq;
324 
325 	if (pkt->mask & (RXE_READ_OR_ATOMIC_MASK | RXE_ATOMIC_WRITE_MASK)) {
326 		/* it is the requesters job to not send
327 		 * too many read/atomic ops, we just
328 		 * recycle the responder resource queue
329 		 */
330 		if (likely(qp->attr.max_dest_rd_atomic > 0))
331 			return RESPST_CHK_LENGTH;
332 		else
333 			return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ;
334 	}
335 
336 	if (pkt->mask & RXE_RWR_MASK) {
337 		if (srq)
338 			return get_srq_wqe(qp);
339 
340 		qp->resp.wqe = queue_head(qp->rq.queue,
341 				QUEUE_TYPE_FROM_CLIENT);
342 		return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR;
343 	}
344 
345 	return RESPST_CHK_LENGTH;
346 }
347 
348 static enum resp_states rxe_resp_check_length(struct rxe_qp *qp,
349 					      struct rxe_pkt_info *pkt)
350 {
351 	/*
352 	 * See IBA C9-92
353 	 * For UD QPs we only check if the packet will fit in the
354 	 * receive buffer later. For rmda operations additional
355 	 * length checks are performed in check_rkey.
356 	 */
357 	if (pkt->mask & RXE_PAYLOAD_MASK && ((qp_type(qp) == IB_QPT_RC) ||
358 					     (qp_type(qp) == IB_QPT_UC))) {
359 		unsigned int mtu = qp->mtu;
360 		unsigned int payload = payload_size(pkt);
361 
362 		if ((pkt->mask & RXE_START_MASK) &&
363 		    (pkt->mask & RXE_END_MASK)) {
364 			if (unlikely(payload > mtu)) {
365 				rxe_dbg_qp(qp, "only packet too long");
366 				return RESPST_ERR_LENGTH;
367 			}
368 		} else if ((pkt->mask & RXE_START_MASK) ||
369 			   (pkt->mask & RXE_MIDDLE_MASK)) {
370 			if (unlikely(payload != mtu)) {
371 				rxe_dbg_qp(qp, "first or middle packet not mtu");
372 				return RESPST_ERR_LENGTH;
373 			}
374 		} else if (pkt->mask & RXE_END_MASK) {
375 			if (unlikely((payload == 0) || (payload > mtu))) {
376 				rxe_dbg_qp(qp, "last packet zero or too long");
377 				return RESPST_ERR_LENGTH;
378 			}
379 		}
380 	}
381 
382 	/* See IBA C9-94 */
383 	if (pkt->mask & RXE_RETH_MASK) {
384 		if (reth_len(pkt) > (1U << 31)) {
385 			rxe_dbg_qp(qp, "dma length too long");
386 			return RESPST_ERR_LENGTH;
387 		}
388 	}
389 
390 	return RESPST_CHK_RKEY;
391 }
392 
393 /* if the reth length field is zero we can assume nothing
394  * about the rkey value and should not validate or use it.
395  * Instead set qp->resp.rkey to 0 which is an invalid rkey
396  * value since the minimum index part is 1.
397  */
398 static void qp_resp_from_reth(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
399 {
400 	unsigned int length = reth_len(pkt);
401 
402 	qp->resp.va = reth_va(pkt);
403 	qp->resp.offset = 0;
404 	qp->resp.resid = length;
405 	qp->resp.length = length;
406 	if (pkt->mask & RXE_READ_OR_WRITE_MASK && length == 0)
407 		qp->resp.rkey = 0;
408 	else
409 		qp->resp.rkey = reth_rkey(pkt);
410 }
411 
412 static void qp_resp_from_atmeth(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
413 {
414 	qp->resp.va = atmeth_va(pkt);
415 	qp->resp.offset = 0;
416 	qp->resp.rkey = atmeth_rkey(pkt);
417 	qp->resp.resid = sizeof(u64);
418 }
419 
420 /* resolve the packet rkey to qp->resp.mr or set qp->resp.mr to NULL
421  * if an invalid rkey is received or the rdma length is zero. For middle
422  * or last packets use the stored value of mr.
423  */
424 static enum resp_states check_rkey(struct rxe_qp *qp,
425 				   struct rxe_pkt_info *pkt)
426 {
427 	struct rxe_mr *mr = NULL;
428 	struct rxe_mw *mw = NULL;
429 	u64 va;
430 	u32 rkey;
431 	u32 resid;
432 	u32 pktlen;
433 	int mtu = qp->mtu;
434 	enum resp_states state;
435 	int access = 0;
436 
437 	if (pkt->mask & (RXE_READ_OR_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) {
438 		if (pkt->mask & RXE_RETH_MASK)
439 			qp_resp_from_reth(qp, pkt);
440 
441 		access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
442 						     : IB_ACCESS_REMOTE_WRITE;
443 	} else if (pkt->mask & RXE_FLUSH_MASK) {
444 		u32 flush_type = feth_plt(pkt);
445 
446 		if (pkt->mask & RXE_RETH_MASK)
447 			qp_resp_from_reth(qp, pkt);
448 
449 		if (flush_type & IB_FLUSH_GLOBAL)
450 			access |= IB_ACCESS_FLUSH_GLOBAL;
451 		if (flush_type & IB_FLUSH_PERSISTENT)
452 			access |= IB_ACCESS_FLUSH_PERSISTENT;
453 	} else if (pkt->mask & RXE_ATOMIC_MASK) {
454 		qp_resp_from_atmeth(qp, pkt);
455 		access = IB_ACCESS_REMOTE_ATOMIC;
456 	} else {
457 		return RESPST_EXECUTE;
458 	}
459 
460 	/* A zero-byte read or write op is not required to
461 	 * set an addr or rkey. See C9-88
462 	 */
463 	if ((pkt->mask & RXE_READ_OR_WRITE_MASK) &&
464 	    (pkt->mask & RXE_RETH_MASK) && reth_len(pkt) == 0) {
465 		qp->resp.mr = NULL;
466 		return RESPST_EXECUTE;
467 	}
468 
469 	va	= qp->resp.va;
470 	rkey	= qp->resp.rkey;
471 	resid	= qp->resp.resid;
472 	pktlen	= payload_size(pkt);
473 
474 	if (rkey_is_mw(rkey)) {
475 		mw = rxe_lookup_mw(qp, access, rkey);
476 		if (!mw) {
477 			rxe_dbg_qp(qp, "no MW matches rkey %#x\n", rkey);
478 			state = RESPST_ERR_RKEY_VIOLATION;
479 			goto err;
480 		}
481 
482 		mr = mw->mr;
483 		if (!mr) {
484 			rxe_dbg_qp(qp, "MW doesn't have an MR\n");
485 			state = RESPST_ERR_RKEY_VIOLATION;
486 			goto err;
487 		}
488 
489 		if (mw->access & IB_ZERO_BASED)
490 			qp->resp.offset = mw->addr;
491 
492 		rxe_get(mr);
493 		rxe_put(mw);
494 		mw = NULL;
495 	} else {
496 		mr = lookup_mr(qp->pd, access, rkey, RXE_LOOKUP_REMOTE);
497 		if (!mr) {
498 			rxe_dbg_qp(qp, "no MR matches rkey %#x\n", rkey);
499 			state = RESPST_ERR_RKEY_VIOLATION;
500 			goto err;
501 		}
502 	}
503 
504 	if (pkt->mask & RXE_FLUSH_MASK) {
505 		/* FLUSH MR may not set va or resid
506 		 * no need to check range since we will flush whole mr
507 		 */
508 		if (feth_sel(pkt) == IB_FLUSH_MR)
509 			goto skip_check_range;
510 	}
511 
512 	if (mr_check_range(mr, va + qp->resp.offset, resid)) {
513 		state = RESPST_ERR_RKEY_VIOLATION;
514 		goto err;
515 	}
516 
517 skip_check_range:
518 	if (pkt->mask & (RXE_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) {
519 		if (resid > mtu) {
520 			if (pktlen != mtu || bth_pad(pkt)) {
521 				state = RESPST_ERR_LENGTH;
522 				goto err;
523 			}
524 		} else {
525 			if (pktlen != resid) {
526 				state = RESPST_ERR_LENGTH;
527 				goto err;
528 			}
529 			if ((bth_pad(pkt) != (0x3 & (-resid)))) {
530 				/* This case may not be exactly that
531 				 * but nothing else fits.
532 				 */
533 				state = RESPST_ERR_LENGTH;
534 				goto err;
535 			}
536 		}
537 	}
538 
539 	WARN_ON_ONCE(qp->resp.mr);
540 
541 	qp->resp.mr = mr;
542 	return RESPST_EXECUTE;
543 
544 err:
545 	qp->resp.mr = NULL;
546 	if (mr)
547 		rxe_put(mr);
548 	if (mw)
549 		rxe_put(mw);
550 
551 	return state;
552 }
553 
554 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
555 				     int data_len)
556 {
557 	int err;
558 
559 	err = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
560 			data_addr, data_len, RXE_TO_MR_OBJ);
561 	if (unlikely(err))
562 		return (err == -ENOSPC) ? RESPST_ERR_LENGTH
563 					: RESPST_ERR_MALFORMED_WQE;
564 
565 	return RESPST_NONE;
566 }
567 
568 static enum resp_states write_data_in(struct rxe_qp *qp,
569 				      struct rxe_pkt_info *pkt)
570 {
571 	enum resp_states rc = RESPST_NONE;
572 	int	err;
573 	int data_len = payload_size(pkt);
574 
575 	err = rxe_mr_copy(qp->resp.mr, qp->resp.va + qp->resp.offset,
576 			  payload_addr(pkt), data_len, RXE_TO_MR_OBJ);
577 	if (err) {
578 		rc = RESPST_ERR_RKEY_VIOLATION;
579 		goto out;
580 	}
581 
582 	qp->resp.va += data_len;
583 	qp->resp.resid -= data_len;
584 
585 out:
586 	return rc;
587 }
588 
589 static struct resp_res *rxe_prepare_res(struct rxe_qp *qp,
590 					struct rxe_pkt_info *pkt,
591 					int type)
592 {
593 	struct resp_res *res;
594 	u32 pkts;
595 
596 	res = &qp->resp.resources[qp->resp.res_head];
597 	rxe_advance_resp_resource(qp);
598 	free_rd_atomic_resource(res);
599 
600 	res->type = type;
601 	res->replay = 0;
602 
603 	switch (type) {
604 	case RXE_READ_MASK:
605 		res->read.va = qp->resp.va + qp->resp.offset;
606 		res->read.va_org = qp->resp.va + qp->resp.offset;
607 		res->read.resid = qp->resp.resid;
608 		res->read.length = qp->resp.resid;
609 		res->read.rkey = qp->resp.rkey;
610 
611 		pkts = max_t(u32, (reth_len(pkt) + qp->mtu - 1)/qp->mtu, 1);
612 		res->first_psn = pkt->psn;
613 		res->cur_psn = pkt->psn;
614 		res->last_psn = (pkt->psn + pkts - 1) & BTH_PSN_MASK;
615 
616 		res->state = rdatm_res_state_new;
617 		break;
618 	case RXE_ATOMIC_MASK:
619 	case RXE_ATOMIC_WRITE_MASK:
620 		res->first_psn = pkt->psn;
621 		res->last_psn = pkt->psn;
622 		res->cur_psn = pkt->psn;
623 		break;
624 	case RXE_FLUSH_MASK:
625 		res->flush.va = qp->resp.va + qp->resp.offset;
626 		res->flush.length = qp->resp.length;
627 		res->flush.type = feth_plt(pkt);
628 		res->flush.level = feth_sel(pkt);
629 	}
630 
631 	return res;
632 }
633 
634 static enum resp_states process_flush(struct rxe_qp *qp,
635 				       struct rxe_pkt_info *pkt)
636 {
637 	u64 length, start;
638 	struct rxe_mr *mr = qp->resp.mr;
639 	struct resp_res *res = qp->resp.res;
640 
641 	/* oA19-14, oA19-15 */
642 	if (res && res->replay)
643 		return RESPST_ACKNOWLEDGE;
644 	else if (!res) {
645 		res = rxe_prepare_res(qp, pkt, RXE_FLUSH_MASK);
646 		qp->resp.res = res;
647 	}
648 
649 	if (res->flush.level == IB_FLUSH_RANGE) {
650 		start = res->flush.va;
651 		length = res->flush.length;
652 	} else { /* level == IB_FLUSH_MR */
653 		start = mr->ibmr.iova;
654 		length = mr->ibmr.length;
655 	}
656 
657 	if (res->flush.type & IB_FLUSH_PERSISTENT) {
658 		if (rxe_flush_pmem_iova(mr, start, length))
659 			return RESPST_ERR_RKEY_VIOLATION;
660 		/* Make data persistent. */
661 		wmb();
662 	} else if (res->flush.type & IB_FLUSH_GLOBAL) {
663 		/* Make data global visibility. */
664 		wmb();
665 	}
666 
667 	qp->resp.msn++;
668 
669 	/* next expected psn, read handles this separately */
670 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
671 	qp->resp.ack_psn = qp->resp.psn;
672 
673 	qp->resp.opcode = pkt->opcode;
674 	qp->resp.status = IB_WC_SUCCESS;
675 
676 	return RESPST_ACKNOWLEDGE;
677 }
678 
679 static enum resp_states atomic_reply(struct rxe_qp *qp,
680 				     struct rxe_pkt_info *pkt)
681 {
682 	struct rxe_mr *mr = qp->resp.mr;
683 	struct resp_res *res = qp->resp.res;
684 	int err;
685 
686 	if (!res) {
687 		res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_MASK);
688 		qp->resp.res = res;
689 	}
690 
691 	if (!res->replay) {
692 		u64 iova = qp->resp.va + qp->resp.offset;
693 
694 		err = rxe_mr_do_atomic_op(mr, iova, pkt->opcode,
695 					  atmeth_comp(pkt),
696 					  atmeth_swap_add(pkt),
697 					  &res->atomic.orig_val);
698 		if (err)
699 			return err;
700 
701 		qp->resp.msn++;
702 
703 		/* next expected psn, read handles this separately */
704 		qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
705 		qp->resp.ack_psn = qp->resp.psn;
706 
707 		qp->resp.opcode = pkt->opcode;
708 		qp->resp.status = IB_WC_SUCCESS;
709 	}
710 
711 	return RESPST_ACKNOWLEDGE;
712 }
713 
714 static enum resp_states atomic_write_reply(struct rxe_qp *qp,
715 					   struct rxe_pkt_info *pkt)
716 {
717 	struct resp_res *res = qp->resp.res;
718 	struct rxe_mr *mr;
719 	u64 value;
720 	u64 iova;
721 	int err;
722 
723 	if (!res) {
724 		res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_WRITE_MASK);
725 		qp->resp.res = res;
726 	}
727 
728 	if (res->replay)
729 		return RESPST_ACKNOWLEDGE;
730 
731 	mr = qp->resp.mr;
732 	value = *(u64 *)payload_addr(pkt);
733 	iova = qp->resp.va + qp->resp.offset;
734 
735 	err = rxe_mr_do_atomic_write(mr, iova, value);
736 	if (err)
737 		return err;
738 
739 	qp->resp.resid = 0;
740 	qp->resp.msn++;
741 
742 	/* next expected psn, read handles this separately */
743 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
744 	qp->resp.ack_psn = qp->resp.psn;
745 
746 	qp->resp.opcode = pkt->opcode;
747 	qp->resp.status = IB_WC_SUCCESS;
748 
749 	return RESPST_ACKNOWLEDGE;
750 }
751 
752 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
753 					  struct rxe_pkt_info *ack,
754 					  int opcode,
755 					  int payload,
756 					  u32 psn,
757 					  u8 syndrome)
758 {
759 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
760 	struct sk_buff *skb;
761 	int paylen;
762 	int pad;
763 	int err;
764 
765 	/*
766 	 * allocate packet
767 	 */
768 	pad = (-payload) & 0x3;
769 	paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
770 
771 	skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack);
772 	if (!skb)
773 		return NULL;
774 
775 	ack->qp = qp;
776 	ack->opcode = opcode;
777 	ack->mask = rxe_opcode[opcode].mask;
778 	ack->paylen = paylen;
779 	ack->psn = psn;
780 
781 	bth_init(ack, opcode, 0, 0, pad, IB_DEFAULT_PKEY_FULL,
782 		 qp->attr.dest_qp_num, 0, psn);
783 
784 	if (ack->mask & RXE_AETH_MASK) {
785 		aeth_set_syn(ack, syndrome);
786 		aeth_set_msn(ack, qp->resp.msn);
787 	}
788 
789 	if (ack->mask & RXE_ATMACK_MASK)
790 		atmack_set_orig(ack, qp->resp.res->atomic.orig_val);
791 
792 	err = rxe_prepare(&qp->pri_av, ack, skb);
793 	if (err) {
794 		kfree_skb(skb);
795 		return NULL;
796 	}
797 
798 	return skb;
799 }
800 
801 /**
802  * rxe_recheck_mr - revalidate MR from rkey and get a reference
803  * @qp: the qp
804  * @rkey: the rkey
805  *
806  * This code allows the MR to be invalidated or deregistered or
807  * the MW if one was used to be invalidated or deallocated.
808  * It is assumed that the access permissions if originally good
809  * are OK and the mappings to be unchanged.
810  *
811  * TODO: If someone reregisters an MR to change its size or
812  * access permissions during the processing of an RDMA read
813  * we should kill the responder resource and complete the
814  * operation with an error.
815  *
816  * Return: mr on success else NULL
817  */
818 static struct rxe_mr *rxe_recheck_mr(struct rxe_qp *qp, u32 rkey)
819 {
820 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
821 	struct rxe_mr *mr;
822 	struct rxe_mw *mw;
823 
824 	if (rkey_is_mw(rkey)) {
825 		mw = rxe_pool_get_index(&rxe->mw_pool, rkey >> 8);
826 		if (!mw)
827 			return NULL;
828 
829 		mr = mw->mr;
830 		if (mw->rkey != rkey || mw->state != RXE_MW_STATE_VALID ||
831 		    !mr || mr->state != RXE_MR_STATE_VALID) {
832 			rxe_put(mw);
833 			return NULL;
834 		}
835 
836 		rxe_get(mr);
837 		rxe_put(mw);
838 
839 		return mr;
840 	}
841 
842 	mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
843 	if (!mr)
844 		return NULL;
845 
846 	if (mr->rkey != rkey || mr->state != RXE_MR_STATE_VALID) {
847 		rxe_put(mr);
848 		return NULL;
849 	}
850 
851 	return mr;
852 }
853 
854 /* RDMA read response. If res is not NULL, then we have a current RDMA request
855  * being processed or replayed.
856  */
857 static enum resp_states read_reply(struct rxe_qp *qp,
858 				   struct rxe_pkt_info *req_pkt)
859 {
860 	struct rxe_pkt_info ack_pkt;
861 	struct sk_buff *skb;
862 	int mtu = qp->mtu;
863 	enum resp_states state;
864 	int payload;
865 	int opcode;
866 	int err;
867 	struct resp_res *res = qp->resp.res;
868 	struct rxe_mr *mr;
869 
870 	if (!res) {
871 		res = rxe_prepare_res(qp, req_pkt, RXE_READ_MASK);
872 		qp->resp.res = res;
873 	}
874 
875 	if (res->state == rdatm_res_state_new) {
876 		if (!res->replay || qp->resp.length == 0) {
877 			/* if length == 0 mr will be NULL (is ok)
878 			 * otherwise qp->resp.mr holds a ref on mr
879 			 * which we transfer to mr and drop below.
880 			 */
881 			mr = qp->resp.mr;
882 			qp->resp.mr = NULL;
883 		} else {
884 			mr = rxe_recheck_mr(qp, res->read.rkey);
885 			if (!mr)
886 				return RESPST_ERR_RKEY_VIOLATION;
887 		}
888 
889 		if (res->read.resid <= mtu)
890 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
891 		else
892 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
893 	} else {
894 		/* re-lookup mr from rkey on all later packets.
895 		 * length will be non-zero. This can fail if someone
896 		 * modifies or destroys the mr since the first packet.
897 		 */
898 		mr = rxe_recheck_mr(qp, res->read.rkey);
899 		if (!mr)
900 			return RESPST_ERR_RKEY_VIOLATION;
901 
902 		if (res->read.resid > mtu)
903 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
904 		else
905 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
906 	}
907 
908 	res->state = rdatm_res_state_next;
909 
910 	payload = min_t(int, res->read.resid, mtu);
911 
912 	skb = prepare_ack_packet(qp, &ack_pkt, opcode, payload,
913 				 res->cur_psn, AETH_ACK_UNLIMITED);
914 	if (!skb) {
915 		state = RESPST_ERR_RNR;
916 		goto err_out;
917 	}
918 
919 	err = rxe_mr_copy(mr, res->read.va, payload_addr(&ack_pkt),
920 			  payload, RXE_FROM_MR_OBJ);
921 	if (err) {
922 		kfree_skb(skb);
923 		state = RESPST_ERR_RKEY_VIOLATION;
924 		goto err_out;
925 	}
926 
927 	if (bth_pad(&ack_pkt)) {
928 		u8 *pad = payload_addr(&ack_pkt) + payload;
929 
930 		memset(pad, 0, bth_pad(&ack_pkt));
931 	}
932 
933 	/* rxe_xmit_packet always consumes the skb */
934 	err = rxe_xmit_packet(qp, &ack_pkt, skb);
935 	if (err) {
936 		state = RESPST_ERR_RNR;
937 		goto err_out;
938 	}
939 
940 	res->read.va += payload;
941 	res->read.resid -= payload;
942 	res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
943 
944 	if (res->read.resid > 0) {
945 		state = RESPST_DONE;
946 	} else {
947 		qp->resp.res = NULL;
948 		if (!res->replay)
949 			qp->resp.opcode = -1;
950 		if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
951 			qp->resp.psn = res->cur_psn;
952 		state = RESPST_CLEANUP;
953 	}
954 
955 err_out:
956 	if (mr)
957 		rxe_put(mr);
958 	return state;
959 }
960 
961 static int invalidate_rkey(struct rxe_qp *qp, u32 rkey)
962 {
963 	if (rkey_is_mw(rkey))
964 		return rxe_invalidate_mw(qp, rkey);
965 	else
966 		return rxe_invalidate_mr(qp, rkey);
967 }
968 
969 /* Executes a new request. A retried request never reach that function (send
970  * and writes are discarded, and reads and atomics are retried elsewhere.
971  */
972 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
973 {
974 	enum resp_states err;
975 	struct sk_buff *skb = PKT_TO_SKB(pkt);
976 	union rdma_network_hdr hdr;
977 
978 	if (pkt->mask & RXE_SEND_MASK) {
979 		if (qp_type(qp) == IB_QPT_UD ||
980 		    qp_type(qp) == IB_QPT_GSI) {
981 			if (skb->protocol == htons(ETH_P_IP)) {
982 				memset(&hdr.reserved, 0,
983 						sizeof(hdr.reserved));
984 				memcpy(&hdr.roce4grh, ip_hdr(skb),
985 						sizeof(hdr.roce4grh));
986 				err = send_data_in(qp, &hdr, sizeof(hdr));
987 			} else {
988 				err = send_data_in(qp, ipv6_hdr(skb),
989 						sizeof(hdr));
990 			}
991 			if (err)
992 				return err;
993 		}
994 		err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
995 		if (err)
996 			return err;
997 	} else if (pkt->mask & RXE_WRITE_MASK) {
998 		err = write_data_in(qp, pkt);
999 		if (err)
1000 			return err;
1001 	} else if (pkt->mask & RXE_READ_MASK) {
1002 		/* For RDMA Read we can increment the msn now. See C9-148. */
1003 		qp->resp.msn++;
1004 		return RESPST_READ_REPLY;
1005 	} else if (pkt->mask & RXE_ATOMIC_MASK) {
1006 		return RESPST_ATOMIC_REPLY;
1007 	} else if (pkt->mask & RXE_ATOMIC_WRITE_MASK) {
1008 		return RESPST_ATOMIC_WRITE_REPLY;
1009 	} else if (pkt->mask & RXE_FLUSH_MASK) {
1010 		return RESPST_PROCESS_FLUSH;
1011 	} else {
1012 		/* Unreachable */
1013 		WARN_ON_ONCE(1);
1014 	}
1015 
1016 	if (pkt->mask & RXE_IETH_MASK) {
1017 		u32 rkey = ieth_rkey(pkt);
1018 
1019 		err = invalidate_rkey(qp, rkey);
1020 		if (err)
1021 			return RESPST_ERR_INVALIDATE_RKEY;
1022 	}
1023 
1024 	if (pkt->mask & RXE_END_MASK)
1025 		/* We successfully processed this new request. */
1026 		qp->resp.msn++;
1027 
1028 	/* next expected psn, read handles this separately */
1029 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
1030 	qp->resp.ack_psn = qp->resp.psn;
1031 
1032 	qp->resp.opcode = pkt->opcode;
1033 	qp->resp.status = IB_WC_SUCCESS;
1034 
1035 	if (pkt->mask & RXE_COMP_MASK)
1036 		return RESPST_COMPLETE;
1037 	else if (qp_type(qp) == IB_QPT_RC)
1038 		return RESPST_ACKNOWLEDGE;
1039 	else
1040 		return RESPST_CLEANUP;
1041 }
1042 
1043 static enum resp_states do_complete(struct rxe_qp *qp,
1044 				    struct rxe_pkt_info *pkt)
1045 {
1046 	struct rxe_cqe cqe;
1047 	struct ib_wc *wc = &cqe.ibwc;
1048 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
1049 	struct rxe_recv_wqe *wqe = qp->resp.wqe;
1050 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1051 	unsigned long flags;
1052 
1053 	if (!wqe)
1054 		goto finish;
1055 
1056 	memset(&cqe, 0, sizeof(cqe));
1057 
1058 	if (qp->rcq->is_user) {
1059 		uwc->status		= qp->resp.status;
1060 		uwc->qp_num		= qp->ibqp.qp_num;
1061 		uwc->wr_id		= wqe->wr_id;
1062 	} else {
1063 		wc->status		= qp->resp.status;
1064 		wc->qp			= &qp->ibqp;
1065 		wc->wr_id		= wqe->wr_id;
1066 	}
1067 
1068 	if (wc->status == IB_WC_SUCCESS) {
1069 		rxe_counter_inc(rxe, RXE_CNT_RDMA_RECV);
1070 		wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
1071 				pkt->mask & RXE_WRITE_MASK) ?
1072 					IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
1073 		wc->byte_len = (pkt->mask & RXE_IMMDT_MASK &&
1074 				pkt->mask & RXE_WRITE_MASK) ?
1075 					qp->resp.length : wqe->dma.length - wqe->dma.resid;
1076 
1077 		/* fields after byte_len are different between kernel and user
1078 		 * space
1079 		 */
1080 		if (qp->rcq->is_user) {
1081 			uwc->wc_flags = IB_WC_GRH;
1082 
1083 			if (pkt->mask & RXE_IMMDT_MASK) {
1084 				uwc->wc_flags |= IB_WC_WITH_IMM;
1085 				uwc->ex.imm_data = immdt_imm(pkt);
1086 			}
1087 
1088 			if (pkt->mask & RXE_IETH_MASK) {
1089 				uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
1090 				uwc->ex.invalidate_rkey = ieth_rkey(pkt);
1091 			}
1092 
1093 			if (pkt->mask & RXE_DETH_MASK)
1094 				uwc->src_qp = deth_sqp(pkt);
1095 
1096 			uwc->port_num		= qp->attr.port_num;
1097 		} else {
1098 			struct sk_buff *skb = PKT_TO_SKB(pkt);
1099 
1100 			wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
1101 			if (skb->protocol == htons(ETH_P_IP))
1102 				wc->network_hdr_type = RDMA_NETWORK_IPV4;
1103 			else
1104 				wc->network_hdr_type = RDMA_NETWORK_IPV6;
1105 
1106 			if (is_vlan_dev(skb->dev)) {
1107 				wc->wc_flags |= IB_WC_WITH_VLAN;
1108 				wc->vlan_id = vlan_dev_vlan_id(skb->dev);
1109 			}
1110 
1111 			if (pkt->mask & RXE_IMMDT_MASK) {
1112 				wc->wc_flags |= IB_WC_WITH_IMM;
1113 				wc->ex.imm_data = immdt_imm(pkt);
1114 			}
1115 
1116 			if (pkt->mask & RXE_IETH_MASK) {
1117 				wc->wc_flags |= IB_WC_WITH_INVALIDATE;
1118 				wc->ex.invalidate_rkey = ieth_rkey(pkt);
1119 			}
1120 
1121 			if (pkt->mask & RXE_DETH_MASK)
1122 				wc->src_qp = deth_sqp(pkt);
1123 
1124 			wc->port_num		= qp->attr.port_num;
1125 		}
1126 	} else {
1127 		if (wc->status != IB_WC_WR_FLUSH_ERR)
1128 			rxe_err_qp(qp, "non-flush error status = %d",
1129 				wc->status);
1130 	}
1131 
1132 	/* have copy for srq and reference for !srq */
1133 	if (!qp->srq)
1134 		queue_advance_consumer(qp->rq.queue, QUEUE_TYPE_FROM_CLIENT);
1135 
1136 	qp->resp.wqe = NULL;
1137 
1138 	if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
1139 		return RESPST_ERR_CQ_OVERFLOW;
1140 
1141 finish:
1142 	spin_lock_irqsave(&qp->state_lock, flags);
1143 	if (unlikely(qp_state(qp) == IB_QPS_ERR)) {
1144 		spin_unlock_irqrestore(&qp->state_lock, flags);
1145 		return RESPST_CHK_RESOURCE;
1146 	}
1147 	spin_unlock_irqrestore(&qp->state_lock, flags);
1148 
1149 	if (unlikely(!pkt))
1150 		return RESPST_DONE;
1151 	if (qp_type(qp) == IB_QPT_RC)
1152 		return RESPST_ACKNOWLEDGE;
1153 	else
1154 		return RESPST_CLEANUP;
1155 }
1156 
1157 
1158 static int send_common_ack(struct rxe_qp *qp, u8 syndrome, u32 psn,
1159 				  int opcode, const char *msg)
1160 {
1161 	int err;
1162 	struct rxe_pkt_info ack_pkt;
1163 	struct sk_buff *skb;
1164 
1165 	skb = prepare_ack_packet(qp, &ack_pkt, opcode, 0, psn, syndrome);
1166 	if (!skb)
1167 		return -ENOMEM;
1168 
1169 	err = rxe_xmit_packet(qp, &ack_pkt, skb);
1170 	if (err)
1171 		rxe_dbg_qp(qp, "Failed sending %s\n", msg);
1172 
1173 	return err;
1174 }
1175 
1176 static int send_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1177 {
1178 	return send_common_ack(qp, syndrome, psn,
1179 			IB_OPCODE_RC_ACKNOWLEDGE, "ACK");
1180 }
1181 
1182 static int send_atomic_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1183 {
1184 	int ret = send_common_ack(qp, syndrome, psn,
1185 			IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, "ATOMIC ACK");
1186 
1187 	/* have to clear this since it is used to trigger
1188 	 * long read replies
1189 	 */
1190 	qp->resp.res = NULL;
1191 	return ret;
1192 }
1193 
1194 static int send_read_response_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1195 {
1196 	int ret = send_common_ack(qp, syndrome, psn,
1197 			IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY,
1198 			"RDMA READ response of length zero ACK");
1199 
1200 	/* have to clear this since it is used to trigger
1201 	 * long read replies
1202 	 */
1203 	qp->resp.res = NULL;
1204 	return ret;
1205 }
1206 
1207 static enum resp_states acknowledge(struct rxe_qp *qp,
1208 				    struct rxe_pkt_info *pkt)
1209 {
1210 	if (qp_type(qp) != IB_QPT_RC)
1211 		return RESPST_CLEANUP;
1212 
1213 	if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1214 		send_ack(qp, qp->resp.aeth_syndrome, pkt->psn);
1215 	else if (pkt->mask & RXE_ATOMIC_MASK)
1216 		send_atomic_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1217 	else if (pkt->mask & (RXE_FLUSH_MASK | RXE_ATOMIC_WRITE_MASK))
1218 		send_read_response_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1219 	else if (bth_ack(pkt))
1220 		send_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1221 
1222 	return RESPST_CLEANUP;
1223 }
1224 
1225 static enum resp_states cleanup(struct rxe_qp *qp,
1226 				struct rxe_pkt_info *pkt)
1227 {
1228 	struct sk_buff *skb;
1229 
1230 	if (pkt) {
1231 		skb = skb_dequeue(&qp->req_pkts);
1232 		rxe_put(qp);
1233 		kfree_skb(skb);
1234 		ib_device_put(qp->ibqp.device);
1235 	}
1236 
1237 	if (qp->resp.mr) {
1238 		rxe_put(qp->resp.mr);
1239 		qp->resp.mr = NULL;
1240 	}
1241 
1242 	return RESPST_DONE;
1243 }
1244 
1245 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1246 {
1247 	int i;
1248 
1249 	for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
1250 		struct resp_res *res = &qp->resp.resources[i];
1251 
1252 		if (res->type == 0)
1253 			continue;
1254 
1255 		if (psn_compare(psn, res->first_psn) >= 0 &&
1256 		    psn_compare(psn, res->last_psn) <= 0) {
1257 			return res;
1258 		}
1259 	}
1260 
1261 	return NULL;
1262 }
1263 
1264 static enum resp_states duplicate_request(struct rxe_qp *qp,
1265 					  struct rxe_pkt_info *pkt)
1266 {
1267 	enum resp_states rc;
1268 	u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK;
1269 
1270 	if (pkt->mask & RXE_SEND_MASK ||
1271 	    pkt->mask & RXE_WRITE_MASK) {
1272 		/* SEND. Ack again and cleanup. C9-105. */
1273 		send_ack(qp, AETH_ACK_UNLIMITED, prev_psn);
1274 		return RESPST_CLEANUP;
1275 	} else if (pkt->mask & RXE_FLUSH_MASK) {
1276 		struct resp_res *res;
1277 
1278 		/* Find the operation in our list of responder resources. */
1279 		res = find_resource(qp, pkt->psn);
1280 		if (res) {
1281 			res->replay = 1;
1282 			res->cur_psn = pkt->psn;
1283 			qp->resp.res = res;
1284 			rc = RESPST_PROCESS_FLUSH;
1285 			goto out;
1286 		}
1287 
1288 		/* Resource not found. Class D error. Drop the request. */
1289 		rc = RESPST_CLEANUP;
1290 		goto out;
1291 	} else if (pkt->mask & RXE_READ_MASK) {
1292 		struct resp_res *res;
1293 
1294 		res = find_resource(qp, pkt->psn);
1295 		if (!res) {
1296 			/* Resource not found. Class D error.  Drop the
1297 			 * request.
1298 			 */
1299 			rc = RESPST_CLEANUP;
1300 			goto out;
1301 		} else {
1302 			/* Ensure this new request is the same as the previous
1303 			 * one or a subset of it.
1304 			 */
1305 			u64 iova = reth_va(pkt);
1306 			u32 resid = reth_len(pkt);
1307 
1308 			if (iova < res->read.va_org ||
1309 			    resid > res->read.length ||
1310 			    (iova + resid) > (res->read.va_org +
1311 					      res->read.length)) {
1312 				rc = RESPST_CLEANUP;
1313 				goto out;
1314 			}
1315 
1316 			if (reth_rkey(pkt) != res->read.rkey) {
1317 				rc = RESPST_CLEANUP;
1318 				goto out;
1319 			}
1320 
1321 			res->cur_psn = pkt->psn;
1322 			res->state = (pkt->psn == res->first_psn) ?
1323 					rdatm_res_state_new :
1324 					rdatm_res_state_replay;
1325 			res->replay = 1;
1326 
1327 			/* Reset the resource, except length. */
1328 			res->read.va_org = iova;
1329 			res->read.va = iova;
1330 			res->read.resid = resid;
1331 
1332 			/* Replay the RDMA read reply. */
1333 			qp->resp.res = res;
1334 			rc = RESPST_READ_REPLY;
1335 			goto out;
1336 		}
1337 	} else {
1338 		struct resp_res *res;
1339 
1340 		/* Find the operation in our list of responder resources. */
1341 		res = find_resource(qp, pkt->psn);
1342 		if (res) {
1343 			res->replay = 1;
1344 			res->cur_psn = pkt->psn;
1345 			qp->resp.res = res;
1346 			rc = pkt->mask & RXE_ATOMIC_MASK ?
1347 					RESPST_ATOMIC_REPLY :
1348 					RESPST_ATOMIC_WRITE_REPLY;
1349 			goto out;
1350 		}
1351 
1352 		/* Resource not found. Class D error. Drop the request. */
1353 		rc = RESPST_CLEANUP;
1354 		goto out;
1355 	}
1356 out:
1357 	return rc;
1358 }
1359 
1360 /* Process a class A or C. Both are treated the same in this implementation. */
1361 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1362 			      enum ib_wc_status status)
1363 {
1364 	qp->resp.aeth_syndrome	= syndrome;
1365 	qp->resp.status		= status;
1366 
1367 	/* indicate that we should go through the ERROR state */
1368 	qp->resp.goto_error	= 1;
1369 }
1370 
1371 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1372 {
1373 	/* UC */
1374 	if (qp->srq) {
1375 		/* Class E */
1376 		qp->resp.drop_msg = 1;
1377 		if (qp->resp.wqe) {
1378 			qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1379 			return RESPST_COMPLETE;
1380 		} else {
1381 			return RESPST_CLEANUP;
1382 		}
1383 	} else {
1384 		/* Class D1. This packet may be the start of a
1385 		 * new message and could be valid. The previous
1386 		 * message is invalid and ignored. reset the
1387 		 * recv wr to its original state
1388 		 */
1389 		if (qp->resp.wqe) {
1390 			qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1391 			qp->resp.wqe->dma.cur_sge = 0;
1392 			qp->resp.wqe->dma.sge_offset = 0;
1393 			qp->resp.opcode = -1;
1394 		}
1395 
1396 		if (qp->resp.mr) {
1397 			rxe_put(qp->resp.mr);
1398 			qp->resp.mr = NULL;
1399 		}
1400 
1401 		return RESPST_CLEANUP;
1402 	}
1403 }
1404 
1405 /* drain incoming request packet queue */
1406 static void drain_req_pkts(struct rxe_qp *qp)
1407 {
1408 	struct sk_buff *skb;
1409 
1410 	while ((skb = skb_dequeue(&qp->req_pkts))) {
1411 		rxe_put(qp);
1412 		kfree_skb(skb);
1413 		ib_device_put(qp->ibqp.device);
1414 	}
1415 }
1416 
1417 /* complete receive wqe with flush error */
1418 static int flush_recv_wqe(struct rxe_qp *qp, struct rxe_recv_wqe *wqe)
1419 {
1420 	struct rxe_cqe cqe = {};
1421 	struct ib_wc *wc = &cqe.ibwc;
1422 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
1423 	int err;
1424 
1425 	if (qp->rcq->is_user) {
1426 		uwc->wr_id = wqe->wr_id;
1427 		uwc->status = IB_WC_WR_FLUSH_ERR;
1428 		uwc->qp_num = qp_num(qp);
1429 	} else {
1430 		wc->wr_id = wqe->wr_id;
1431 		wc->status = IB_WC_WR_FLUSH_ERR;
1432 		wc->qp = &qp->ibqp;
1433 	}
1434 
1435 	err = rxe_cq_post(qp->rcq, &cqe, 0);
1436 	if (err)
1437 		rxe_dbg_cq(qp->rcq, "post cq failed err = %d", err);
1438 
1439 	return err;
1440 }
1441 
1442 /* drain and optionally complete the recive queue
1443  * if unable to complete a wqe stop completing and
1444  * just flush the remaining wqes
1445  */
1446 static void flush_recv_queue(struct rxe_qp *qp, bool notify)
1447 {
1448 	struct rxe_queue *q = qp->rq.queue;
1449 	struct rxe_recv_wqe *wqe;
1450 	int err;
1451 
1452 	if (qp->srq)
1453 		return;
1454 
1455 	while ((wqe = queue_head(q, q->type))) {
1456 		if (notify) {
1457 			err = flush_recv_wqe(qp, wqe);
1458 			if (err)
1459 				notify = 0;
1460 		}
1461 		queue_advance_consumer(q, q->type);
1462 	}
1463 
1464 	qp->resp.wqe = NULL;
1465 }
1466 
1467 int rxe_responder(struct rxe_qp *qp)
1468 {
1469 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1470 	enum resp_states state;
1471 	struct rxe_pkt_info *pkt = NULL;
1472 	int ret;
1473 	unsigned long flags;
1474 
1475 	spin_lock_irqsave(&qp->state_lock, flags);
1476 	if (!qp->valid || qp_state(qp) == IB_QPS_ERR ||
1477 			  qp_state(qp) == IB_QPS_RESET) {
1478 		bool notify = qp->valid && (qp_state(qp) == IB_QPS_ERR);
1479 
1480 		drain_req_pkts(qp);
1481 		flush_recv_queue(qp, notify);
1482 		spin_unlock_irqrestore(&qp->state_lock, flags);
1483 		goto exit;
1484 	}
1485 	spin_unlock_irqrestore(&qp->state_lock, flags);
1486 
1487 	qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1488 
1489 	state = RESPST_GET_REQ;
1490 
1491 	while (1) {
1492 		rxe_dbg_qp(qp, "state = %s\n", resp_state_name[state]);
1493 		switch (state) {
1494 		case RESPST_GET_REQ:
1495 			state = get_req(qp, &pkt);
1496 			break;
1497 		case RESPST_CHK_PSN:
1498 			state = check_psn(qp, pkt);
1499 			break;
1500 		case RESPST_CHK_OP_SEQ:
1501 			state = check_op_seq(qp, pkt);
1502 			break;
1503 		case RESPST_CHK_OP_VALID:
1504 			state = check_op_valid(qp, pkt);
1505 			break;
1506 		case RESPST_CHK_RESOURCE:
1507 			state = check_resource(qp, pkt);
1508 			break;
1509 		case RESPST_CHK_LENGTH:
1510 			state = rxe_resp_check_length(qp, pkt);
1511 			break;
1512 		case RESPST_CHK_RKEY:
1513 			state = check_rkey(qp, pkt);
1514 			break;
1515 		case RESPST_EXECUTE:
1516 			state = execute(qp, pkt);
1517 			break;
1518 		case RESPST_COMPLETE:
1519 			state = do_complete(qp, pkt);
1520 			break;
1521 		case RESPST_READ_REPLY:
1522 			state = read_reply(qp, pkt);
1523 			break;
1524 		case RESPST_ATOMIC_REPLY:
1525 			state = atomic_reply(qp, pkt);
1526 			break;
1527 		case RESPST_ATOMIC_WRITE_REPLY:
1528 			state = atomic_write_reply(qp, pkt);
1529 			break;
1530 		case RESPST_PROCESS_FLUSH:
1531 			state = process_flush(qp, pkt);
1532 			break;
1533 		case RESPST_ACKNOWLEDGE:
1534 			state = acknowledge(qp, pkt);
1535 			break;
1536 		case RESPST_CLEANUP:
1537 			state = cleanup(qp, pkt);
1538 			break;
1539 		case RESPST_DUPLICATE_REQUEST:
1540 			state = duplicate_request(qp, pkt);
1541 			break;
1542 		case RESPST_ERR_PSN_OUT_OF_SEQ:
1543 			/* RC only - Class B. Drop packet. */
1544 			send_ack(qp, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1545 			state = RESPST_CLEANUP;
1546 			break;
1547 
1548 		case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1549 		case RESPST_ERR_MISSING_OPCODE_FIRST:
1550 		case RESPST_ERR_MISSING_OPCODE_LAST_C:
1551 		case RESPST_ERR_UNSUPPORTED_OPCODE:
1552 		case RESPST_ERR_MISALIGNED_ATOMIC:
1553 			/* RC Only - Class C. */
1554 			do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1555 					  IB_WC_REM_INV_REQ_ERR);
1556 			state = RESPST_COMPLETE;
1557 			break;
1558 
1559 		case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1560 			state = do_class_d1e_error(qp);
1561 			break;
1562 		case RESPST_ERR_RNR:
1563 			if (qp_type(qp) == IB_QPT_RC) {
1564 				rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1565 				/* RC - class B */
1566 				send_ack(qp, AETH_RNR_NAK |
1567 					 (~AETH_TYPE_MASK &
1568 					 qp->attr.min_rnr_timer),
1569 					 pkt->psn);
1570 			} else {
1571 				/* UD/UC - class D */
1572 				qp->resp.drop_msg = 1;
1573 			}
1574 			state = RESPST_CLEANUP;
1575 			break;
1576 
1577 		case RESPST_ERR_RKEY_VIOLATION:
1578 			if (qp_type(qp) == IB_QPT_RC) {
1579 				/* Class C */
1580 				do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1581 						  IB_WC_REM_ACCESS_ERR);
1582 				state = RESPST_COMPLETE;
1583 			} else {
1584 				qp->resp.drop_msg = 1;
1585 				if (qp->srq) {
1586 					/* UC/SRQ Class D */
1587 					qp->resp.status = IB_WC_REM_ACCESS_ERR;
1588 					state = RESPST_COMPLETE;
1589 				} else {
1590 					/* UC/non-SRQ Class E. */
1591 					state = RESPST_CLEANUP;
1592 				}
1593 			}
1594 			break;
1595 
1596 		case RESPST_ERR_INVALIDATE_RKEY:
1597 			/* RC - Class J. */
1598 			qp->resp.goto_error = 1;
1599 			qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1600 			state = RESPST_COMPLETE;
1601 			break;
1602 
1603 		case RESPST_ERR_LENGTH:
1604 			if (qp_type(qp) == IB_QPT_RC) {
1605 				/* Class C */
1606 				do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1607 						  IB_WC_REM_INV_REQ_ERR);
1608 				state = RESPST_COMPLETE;
1609 			} else if (qp->srq) {
1610 				/* UC/UD - class E */
1611 				qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1612 				state = RESPST_COMPLETE;
1613 			} else {
1614 				/* UC/UD - class D */
1615 				qp->resp.drop_msg = 1;
1616 				state = RESPST_CLEANUP;
1617 			}
1618 			break;
1619 
1620 		case RESPST_ERR_MALFORMED_WQE:
1621 			/* All, Class A. */
1622 			do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1623 					  IB_WC_LOC_QP_OP_ERR);
1624 			state = RESPST_COMPLETE;
1625 			break;
1626 
1627 		case RESPST_ERR_CQ_OVERFLOW:
1628 			/* All - Class G */
1629 			state = RESPST_ERROR;
1630 			break;
1631 
1632 		case RESPST_DONE:
1633 			if (qp->resp.goto_error) {
1634 				state = RESPST_ERROR;
1635 				break;
1636 			}
1637 
1638 			goto done;
1639 
1640 		case RESPST_EXIT:
1641 			if (qp->resp.goto_error) {
1642 				state = RESPST_ERROR;
1643 				break;
1644 			}
1645 
1646 			goto exit;
1647 
1648 		case RESPST_ERROR:
1649 			qp->resp.goto_error = 0;
1650 			rxe_dbg_qp(qp, "moved to error state\n");
1651 			rxe_qp_error(qp);
1652 			goto exit;
1653 
1654 		default:
1655 			WARN_ON_ONCE(1);
1656 		}
1657 	}
1658 
1659 	/* A non-zero return value will cause rxe_do_task to
1660 	 * exit its loop and end the tasklet. A zero return
1661 	 * will continue looping and return to rxe_responder
1662 	 */
1663 done:
1664 	ret = 0;
1665 	goto out;
1666 exit:
1667 	ret = -EAGAIN;
1668 out:
1669 	return ret;
1670 }
1671