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_put(mw);
493 		rxe_get(mr);
494 	} else {
495 		mr = lookup_mr(qp->pd, access, rkey, RXE_LOOKUP_REMOTE);
496 		if (!mr) {
497 			rxe_dbg_qp(qp, "no MR matches rkey %#x\n", rkey);
498 			state = RESPST_ERR_RKEY_VIOLATION;
499 			goto err;
500 		}
501 	}
502 
503 	if (pkt->mask & RXE_FLUSH_MASK) {
504 		/* FLUSH MR may not set va or resid
505 		 * no need to check range since we will flush whole mr
506 		 */
507 		if (feth_sel(pkt) == IB_FLUSH_MR)
508 			goto skip_check_range;
509 	}
510 
511 	if (mr_check_range(mr, va + qp->resp.offset, resid)) {
512 		state = RESPST_ERR_RKEY_VIOLATION;
513 		goto err;
514 	}
515 
516 skip_check_range:
517 	if (pkt->mask & (RXE_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) {
518 		if (resid > mtu) {
519 			if (pktlen != mtu || bth_pad(pkt)) {
520 				state = RESPST_ERR_LENGTH;
521 				goto err;
522 			}
523 		} else {
524 			if (pktlen != resid) {
525 				state = RESPST_ERR_LENGTH;
526 				goto err;
527 			}
528 			if ((bth_pad(pkt) != (0x3 & (-resid)))) {
529 				/* This case may not be exactly that
530 				 * but nothing else fits.
531 				 */
532 				state = RESPST_ERR_LENGTH;
533 				goto err;
534 			}
535 		}
536 	}
537 
538 	WARN_ON_ONCE(qp->resp.mr);
539 
540 	qp->resp.mr = mr;
541 	return RESPST_EXECUTE;
542 
543 err:
544 	qp->resp.mr = NULL;
545 	if (mr)
546 		rxe_put(mr);
547 	if (mw)
548 		rxe_put(mw);
549 
550 	return state;
551 }
552 
553 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
554 				     int data_len)
555 {
556 	int err;
557 
558 	err = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
559 			data_addr, data_len, RXE_TO_MR_OBJ);
560 	if (unlikely(err))
561 		return (err == -ENOSPC) ? RESPST_ERR_LENGTH
562 					: RESPST_ERR_MALFORMED_WQE;
563 
564 	return RESPST_NONE;
565 }
566 
567 static enum resp_states write_data_in(struct rxe_qp *qp,
568 				      struct rxe_pkt_info *pkt)
569 {
570 	enum resp_states rc = RESPST_NONE;
571 	int	err;
572 	int data_len = payload_size(pkt);
573 
574 	err = rxe_mr_copy(qp->resp.mr, qp->resp.va + qp->resp.offset,
575 			  payload_addr(pkt), data_len, RXE_TO_MR_OBJ);
576 	if (err) {
577 		rc = RESPST_ERR_RKEY_VIOLATION;
578 		goto out;
579 	}
580 
581 	qp->resp.va += data_len;
582 	qp->resp.resid -= data_len;
583 
584 out:
585 	return rc;
586 }
587 
588 static struct resp_res *rxe_prepare_res(struct rxe_qp *qp,
589 					struct rxe_pkt_info *pkt,
590 					int type)
591 {
592 	struct resp_res *res;
593 	u32 pkts;
594 
595 	res = &qp->resp.resources[qp->resp.res_head];
596 	rxe_advance_resp_resource(qp);
597 	free_rd_atomic_resource(res);
598 
599 	res->type = type;
600 	res->replay = 0;
601 
602 	switch (type) {
603 	case RXE_READ_MASK:
604 		res->read.va = qp->resp.va + qp->resp.offset;
605 		res->read.va_org = qp->resp.va + qp->resp.offset;
606 		res->read.resid = qp->resp.resid;
607 		res->read.length = qp->resp.resid;
608 		res->read.rkey = qp->resp.rkey;
609 
610 		pkts = max_t(u32, (reth_len(pkt) + qp->mtu - 1)/qp->mtu, 1);
611 		res->first_psn = pkt->psn;
612 		res->cur_psn = pkt->psn;
613 		res->last_psn = (pkt->psn + pkts - 1) & BTH_PSN_MASK;
614 
615 		res->state = rdatm_res_state_new;
616 		break;
617 	case RXE_ATOMIC_MASK:
618 	case RXE_ATOMIC_WRITE_MASK:
619 		res->first_psn = pkt->psn;
620 		res->last_psn = pkt->psn;
621 		res->cur_psn = pkt->psn;
622 		break;
623 	case RXE_FLUSH_MASK:
624 		res->flush.va = qp->resp.va + qp->resp.offset;
625 		res->flush.length = qp->resp.length;
626 		res->flush.type = feth_plt(pkt);
627 		res->flush.level = feth_sel(pkt);
628 	}
629 
630 	return res;
631 }
632 
633 static enum resp_states process_flush(struct rxe_qp *qp,
634 				       struct rxe_pkt_info *pkt)
635 {
636 	u64 length, start;
637 	struct rxe_mr *mr = qp->resp.mr;
638 	struct resp_res *res = qp->resp.res;
639 
640 	/* oA19-14, oA19-15 */
641 	if (res && res->replay)
642 		return RESPST_ACKNOWLEDGE;
643 	else if (!res) {
644 		res = rxe_prepare_res(qp, pkt, RXE_FLUSH_MASK);
645 		qp->resp.res = res;
646 	}
647 
648 	if (res->flush.level == IB_FLUSH_RANGE) {
649 		start = res->flush.va;
650 		length = res->flush.length;
651 	} else { /* level == IB_FLUSH_MR */
652 		start = mr->ibmr.iova;
653 		length = mr->ibmr.length;
654 	}
655 
656 	if (res->flush.type & IB_FLUSH_PERSISTENT) {
657 		if (rxe_flush_pmem_iova(mr, start, length))
658 			return RESPST_ERR_RKEY_VIOLATION;
659 		/* Make data persistent. */
660 		wmb();
661 	} else if (res->flush.type & IB_FLUSH_GLOBAL) {
662 		/* Make data global visibility. */
663 		wmb();
664 	}
665 
666 	qp->resp.msn++;
667 
668 	/* next expected psn, read handles this separately */
669 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
670 	qp->resp.ack_psn = qp->resp.psn;
671 
672 	qp->resp.opcode = pkt->opcode;
673 	qp->resp.status = IB_WC_SUCCESS;
674 
675 	return RESPST_ACKNOWLEDGE;
676 }
677 
678 static enum resp_states atomic_reply(struct rxe_qp *qp,
679 				     struct rxe_pkt_info *pkt)
680 {
681 	struct rxe_mr *mr = qp->resp.mr;
682 	struct resp_res *res = qp->resp.res;
683 	int err;
684 
685 	if (!res) {
686 		res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_MASK);
687 		qp->resp.res = res;
688 	}
689 
690 	if (!res->replay) {
691 		u64 iova = qp->resp.va + qp->resp.offset;
692 
693 		err = rxe_mr_do_atomic_op(mr, iova, pkt->opcode,
694 					  atmeth_comp(pkt),
695 					  atmeth_swap_add(pkt),
696 					  &res->atomic.orig_val);
697 		if (err)
698 			return err;
699 
700 		qp->resp.msn++;
701 
702 		/* next expected psn, read handles this separately */
703 		qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
704 		qp->resp.ack_psn = qp->resp.psn;
705 
706 		qp->resp.opcode = pkt->opcode;
707 		qp->resp.status = IB_WC_SUCCESS;
708 	}
709 
710 	return RESPST_ACKNOWLEDGE;
711 }
712 
713 static enum resp_states atomic_write_reply(struct rxe_qp *qp,
714 					   struct rxe_pkt_info *pkt)
715 {
716 	struct resp_res *res = qp->resp.res;
717 	struct rxe_mr *mr;
718 	u64 value;
719 	u64 iova;
720 	int err;
721 
722 	if (!res) {
723 		res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_WRITE_MASK);
724 		qp->resp.res = res;
725 	}
726 
727 	if (res->replay)
728 		return RESPST_ACKNOWLEDGE;
729 
730 	mr = qp->resp.mr;
731 	value = *(u64 *)payload_addr(pkt);
732 	iova = qp->resp.va + qp->resp.offset;
733 
734 	err = rxe_mr_do_atomic_write(mr, iova, value);
735 	if (err)
736 		return err;
737 
738 	qp->resp.resid = 0;
739 	qp->resp.msn++;
740 
741 	/* next expected psn, read handles this separately */
742 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
743 	qp->resp.ack_psn = qp->resp.psn;
744 
745 	qp->resp.opcode = pkt->opcode;
746 	qp->resp.status = IB_WC_SUCCESS;
747 
748 	return RESPST_ACKNOWLEDGE;
749 }
750 
751 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
752 					  struct rxe_pkt_info *ack,
753 					  int opcode,
754 					  int payload,
755 					  u32 psn,
756 					  u8 syndrome)
757 {
758 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
759 	struct sk_buff *skb;
760 	int paylen;
761 	int pad;
762 	int err;
763 
764 	/*
765 	 * allocate packet
766 	 */
767 	pad = (-payload) & 0x3;
768 	paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
769 
770 	skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack);
771 	if (!skb)
772 		return NULL;
773 
774 	ack->qp = qp;
775 	ack->opcode = opcode;
776 	ack->mask = rxe_opcode[opcode].mask;
777 	ack->paylen = paylen;
778 	ack->psn = psn;
779 
780 	bth_init(ack, opcode, 0, 0, pad, IB_DEFAULT_PKEY_FULL,
781 		 qp->attr.dest_qp_num, 0, psn);
782 
783 	if (ack->mask & RXE_AETH_MASK) {
784 		aeth_set_syn(ack, syndrome);
785 		aeth_set_msn(ack, qp->resp.msn);
786 	}
787 
788 	if (ack->mask & RXE_ATMACK_MASK)
789 		atmack_set_orig(ack, qp->resp.res->atomic.orig_val);
790 
791 	err = rxe_prepare(&qp->pri_av, ack, skb);
792 	if (err) {
793 		kfree_skb(skb);
794 		return NULL;
795 	}
796 
797 	return skb;
798 }
799 
800 /**
801  * rxe_recheck_mr - revalidate MR from rkey and get a reference
802  * @qp: the qp
803  * @rkey: the rkey
804  *
805  * This code allows the MR to be invalidated or deregistered or
806  * the MW if one was used to be invalidated or deallocated.
807  * It is assumed that the access permissions if originally good
808  * are OK and the mappings to be unchanged.
809  *
810  * TODO: If someone reregisters an MR to change its size or
811  * access permissions during the processing of an RDMA read
812  * we should kill the responder resource and complete the
813  * operation with an error.
814  *
815  * Return: mr on success else NULL
816  */
817 static struct rxe_mr *rxe_recheck_mr(struct rxe_qp *qp, u32 rkey)
818 {
819 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
820 	struct rxe_mr *mr;
821 	struct rxe_mw *mw;
822 
823 	if (rkey_is_mw(rkey)) {
824 		mw = rxe_pool_get_index(&rxe->mw_pool, rkey >> 8);
825 		if (!mw)
826 			return NULL;
827 
828 		mr = mw->mr;
829 		if (mw->rkey != rkey || mw->state != RXE_MW_STATE_VALID ||
830 		    !mr || mr->state != RXE_MR_STATE_VALID) {
831 			rxe_put(mw);
832 			return NULL;
833 		}
834 
835 		rxe_get(mr);
836 		rxe_put(mw);
837 
838 		return mr;
839 	}
840 
841 	mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
842 	if (!mr)
843 		return NULL;
844 
845 	if (mr->rkey != rkey || mr->state != RXE_MR_STATE_VALID) {
846 		rxe_put(mr);
847 		return NULL;
848 	}
849 
850 	return mr;
851 }
852 
853 /* RDMA read response. If res is not NULL, then we have a current RDMA request
854  * being processed or replayed.
855  */
856 static enum resp_states read_reply(struct rxe_qp *qp,
857 				   struct rxe_pkt_info *req_pkt)
858 {
859 	struct rxe_pkt_info ack_pkt;
860 	struct sk_buff *skb;
861 	int mtu = qp->mtu;
862 	enum resp_states state;
863 	int payload;
864 	int opcode;
865 	int err;
866 	struct resp_res *res = qp->resp.res;
867 	struct rxe_mr *mr;
868 
869 	if (!res) {
870 		res = rxe_prepare_res(qp, req_pkt, RXE_READ_MASK);
871 		qp->resp.res = res;
872 	}
873 
874 	if (res->state == rdatm_res_state_new) {
875 		if (!res->replay || qp->resp.length == 0) {
876 			/* if length == 0 mr will be NULL (is ok)
877 			 * otherwise qp->resp.mr holds a ref on mr
878 			 * which we transfer to mr and drop below.
879 			 */
880 			mr = qp->resp.mr;
881 			qp->resp.mr = NULL;
882 		} else {
883 			mr = rxe_recheck_mr(qp, res->read.rkey);
884 			if (!mr)
885 				return RESPST_ERR_RKEY_VIOLATION;
886 		}
887 
888 		if (res->read.resid <= mtu)
889 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
890 		else
891 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
892 	} else {
893 		/* re-lookup mr from rkey on all later packets.
894 		 * length will be non-zero. This can fail if someone
895 		 * modifies or destroys the mr since the first packet.
896 		 */
897 		mr = rxe_recheck_mr(qp, res->read.rkey);
898 		if (!mr)
899 			return RESPST_ERR_RKEY_VIOLATION;
900 
901 		if (res->read.resid > mtu)
902 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
903 		else
904 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
905 	}
906 
907 	res->state = rdatm_res_state_next;
908 
909 	payload = min_t(int, res->read.resid, mtu);
910 
911 	skb = prepare_ack_packet(qp, &ack_pkt, opcode, payload,
912 				 res->cur_psn, AETH_ACK_UNLIMITED);
913 	if (!skb) {
914 		state = RESPST_ERR_RNR;
915 		goto err_out;
916 	}
917 
918 	err = rxe_mr_copy(mr, res->read.va, payload_addr(&ack_pkt),
919 			  payload, RXE_FROM_MR_OBJ);
920 	if (err) {
921 		kfree_skb(skb);
922 		state = RESPST_ERR_RKEY_VIOLATION;
923 		goto err_out;
924 	}
925 
926 	if (bth_pad(&ack_pkt)) {
927 		u8 *pad = payload_addr(&ack_pkt) + payload;
928 
929 		memset(pad, 0, bth_pad(&ack_pkt));
930 	}
931 
932 	/* rxe_xmit_packet always consumes the skb */
933 	err = rxe_xmit_packet(qp, &ack_pkt, skb);
934 	if (err) {
935 		state = RESPST_ERR_RNR;
936 		goto err_out;
937 	}
938 
939 	res->read.va += payload;
940 	res->read.resid -= payload;
941 	res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
942 
943 	if (res->read.resid > 0) {
944 		state = RESPST_DONE;
945 	} else {
946 		qp->resp.res = NULL;
947 		if (!res->replay)
948 			qp->resp.opcode = -1;
949 		if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
950 			qp->resp.psn = res->cur_psn;
951 		state = RESPST_CLEANUP;
952 	}
953 
954 err_out:
955 	if (mr)
956 		rxe_put(mr);
957 	return state;
958 }
959 
960 static int invalidate_rkey(struct rxe_qp *qp, u32 rkey)
961 {
962 	if (rkey_is_mw(rkey))
963 		return rxe_invalidate_mw(qp, rkey);
964 	else
965 		return rxe_invalidate_mr(qp, rkey);
966 }
967 
968 /* Executes a new request. A retried request never reach that function (send
969  * and writes are discarded, and reads and atomics are retried elsewhere.
970  */
971 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
972 {
973 	enum resp_states err;
974 	struct sk_buff *skb = PKT_TO_SKB(pkt);
975 	union rdma_network_hdr hdr;
976 
977 	if (pkt->mask & RXE_SEND_MASK) {
978 		if (qp_type(qp) == IB_QPT_UD ||
979 		    qp_type(qp) == IB_QPT_GSI) {
980 			if (skb->protocol == htons(ETH_P_IP)) {
981 				memset(&hdr.reserved, 0,
982 						sizeof(hdr.reserved));
983 				memcpy(&hdr.roce4grh, ip_hdr(skb),
984 						sizeof(hdr.roce4grh));
985 				err = send_data_in(qp, &hdr, sizeof(hdr));
986 			} else {
987 				err = send_data_in(qp, ipv6_hdr(skb),
988 						sizeof(hdr));
989 			}
990 			if (err)
991 				return err;
992 		}
993 		err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
994 		if (err)
995 			return err;
996 	} else if (pkt->mask & RXE_WRITE_MASK) {
997 		err = write_data_in(qp, pkt);
998 		if (err)
999 			return err;
1000 	} else if (pkt->mask & RXE_READ_MASK) {
1001 		/* For RDMA Read we can increment the msn now. See C9-148. */
1002 		qp->resp.msn++;
1003 		return RESPST_READ_REPLY;
1004 	} else if (pkt->mask & RXE_ATOMIC_MASK) {
1005 		return RESPST_ATOMIC_REPLY;
1006 	} else if (pkt->mask & RXE_ATOMIC_WRITE_MASK) {
1007 		return RESPST_ATOMIC_WRITE_REPLY;
1008 	} else if (pkt->mask & RXE_FLUSH_MASK) {
1009 		return RESPST_PROCESS_FLUSH;
1010 	} else {
1011 		/* Unreachable */
1012 		WARN_ON_ONCE(1);
1013 	}
1014 
1015 	if (pkt->mask & RXE_IETH_MASK) {
1016 		u32 rkey = ieth_rkey(pkt);
1017 
1018 		err = invalidate_rkey(qp, rkey);
1019 		if (err)
1020 			return RESPST_ERR_INVALIDATE_RKEY;
1021 	}
1022 
1023 	if (pkt->mask & RXE_END_MASK)
1024 		/* We successfully processed this new request. */
1025 		qp->resp.msn++;
1026 
1027 	/* next expected psn, read handles this separately */
1028 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
1029 	qp->resp.ack_psn = qp->resp.psn;
1030 
1031 	qp->resp.opcode = pkt->opcode;
1032 	qp->resp.status = IB_WC_SUCCESS;
1033 
1034 	if (pkt->mask & RXE_COMP_MASK)
1035 		return RESPST_COMPLETE;
1036 	else if (qp_type(qp) == IB_QPT_RC)
1037 		return RESPST_ACKNOWLEDGE;
1038 	else
1039 		return RESPST_CLEANUP;
1040 }
1041 
1042 static enum resp_states do_complete(struct rxe_qp *qp,
1043 				    struct rxe_pkt_info *pkt)
1044 {
1045 	struct rxe_cqe cqe;
1046 	struct ib_wc *wc = &cqe.ibwc;
1047 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
1048 	struct rxe_recv_wqe *wqe = qp->resp.wqe;
1049 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1050 
1051 	if (!wqe)
1052 		goto finish;
1053 
1054 	memset(&cqe, 0, sizeof(cqe));
1055 
1056 	if (qp->rcq->is_user) {
1057 		uwc->status		= qp->resp.status;
1058 		uwc->qp_num		= qp->ibqp.qp_num;
1059 		uwc->wr_id		= wqe->wr_id;
1060 	} else {
1061 		wc->status		= qp->resp.status;
1062 		wc->qp			= &qp->ibqp;
1063 		wc->wr_id		= wqe->wr_id;
1064 	}
1065 
1066 	if (wc->status == IB_WC_SUCCESS) {
1067 		rxe_counter_inc(rxe, RXE_CNT_RDMA_RECV);
1068 		wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
1069 				pkt->mask & RXE_WRITE_MASK) ?
1070 					IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
1071 		wc->byte_len = (pkt->mask & RXE_IMMDT_MASK &&
1072 				pkt->mask & RXE_WRITE_MASK) ?
1073 					qp->resp.length : wqe->dma.length - wqe->dma.resid;
1074 
1075 		/* fields after byte_len are different between kernel and user
1076 		 * space
1077 		 */
1078 		if (qp->rcq->is_user) {
1079 			uwc->wc_flags = IB_WC_GRH;
1080 
1081 			if (pkt->mask & RXE_IMMDT_MASK) {
1082 				uwc->wc_flags |= IB_WC_WITH_IMM;
1083 				uwc->ex.imm_data = immdt_imm(pkt);
1084 			}
1085 
1086 			if (pkt->mask & RXE_IETH_MASK) {
1087 				uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
1088 				uwc->ex.invalidate_rkey = ieth_rkey(pkt);
1089 			}
1090 
1091 			if (pkt->mask & RXE_DETH_MASK)
1092 				uwc->src_qp = deth_sqp(pkt);
1093 
1094 			uwc->port_num		= qp->attr.port_num;
1095 		} else {
1096 			struct sk_buff *skb = PKT_TO_SKB(pkt);
1097 
1098 			wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
1099 			if (skb->protocol == htons(ETH_P_IP))
1100 				wc->network_hdr_type = RDMA_NETWORK_IPV4;
1101 			else
1102 				wc->network_hdr_type = RDMA_NETWORK_IPV6;
1103 
1104 			if (is_vlan_dev(skb->dev)) {
1105 				wc->wc_flags |= IB_WC_WITH_VLAN;
1106 				wc->vlan_id = vlan_dev_vlan_id(skb->dev);
1107 			}
1108 
1109 			if (pkt->mask & RXE_IMMDT_MASK) {
1110 				wc->wc_flags |= IB_WC_WITH_IMM;
1111 				wc->ex.imm_data = immdt_imm(pkt);
1112 			}
1113 
1114 			if (pkt->mask & RXE_IETH_MASK) {
1115 				wc->wc_flags |= IB_WC_WITH_INVALIDATE;
1116 				wc->ex.invalidate_rkey = ieth_rkey(pkt);
1117 			}
1118 
1119 			if (pkt->mask & RXE_DETH_MASK)
1120 				wc->src_qp = deth_sqp(pkt);
1121 
1122 			wc->port_num		= qp->attr.port_num;
1123 		}
1124 	} else {
1125 		if (wc->status != IB_WC_WR_FLUSH_ERR)
1126 			rxe_err_qp(qp, "non-flush error status = %d",
1127 				wc->status);
1128 	}
1129 
1130 	/* have copy for srq and reference for !srq */
1131 	if (!qp->srq)
1132 		queue_advance_consumer(qp->rq.queue, QUEUE_TYPE_FROM_CLIENT);
1133 
1134 	qp->resp.wqe = NULL;
1135 
1136 	if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
1137 		return RESPST_ERR_CQ_OVERFLOW;
1138 
1139 finish:
1140 	spin_lock_bh(&qp->state_lock);
1141 	if (unlikely(qp_state(qp) == IB_QPS_ERR)) {
1142 		spin_unlock_bh(&qp->state_lock);
1143 		return RESPST_CHK_RESOURCE;
1144 	}
1145 	spin_unlock_bh(&qp->state_lock);
1146 
1147 	if (unlikely(!pkt))
1148 		return RESPST_DONE;
1149 	if (qp_type(qp) == IB_QPT_RC)
1150 		return RESPST_ACKNOWLEDGE;
1151 	else
1152 		return RESPST_CLEANUP;
1153 }
1154 
1155 
1156 static int send_common_ack(struct rxe_qp *qp, u8 syndrome, u32 psn,
1157 				  int opcode, const char *msg)
1158 {
1159 	int err;
1160 	struct rxe_pkt_info ack_pkt;
1161 	struct sk_buff *skb;
1162 
1163 	skb = prepare_ack_packet(qp, &ack_pkt, opcode, 0, psn, syndrome);
1164 	if (!skb)
1165 		return -ENOMEM;
1166 
1167 	err = rxe_xmit_packet(qp, &ack_pkt, skb);
1168 	if (err)
1169 		rxe_dbg_qp(qp, "Failed sending %s\n", msg);
1170 
1171 	return err;
1172 }
1173 
1174 static int send_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1175 {
1176 	return send_common_ack(qp, syndrome, psn,
1177 			IB_OPCODE_RC_ACKNOWLEDGE, "ACK");
1178 }
1179 
1180 static int send_atomic_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1181 {
1182 	int ret = send_common_ack(qp, syndrome, psn,
1183 			IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, "ATOMIC ACK");
1184 
1185 	/* have to clear this since it is used to trigger
1186 	 * long read replies
1187 	 */
1188 	qp->resp.res = NULL;
1189 	return ret;
1190 }
1191 
1192 static int send_read_response_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1193 {
1194 	int ret = send_common_ack(qp, syndrome, psn,
1195 			IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY,
1196 			"RDMA READ response of length zero ACK");
1197 
1198 	/* have to clear this since it is used to trigger
1199 	 * long read replies
1200 	 */
1201 	qp->resp.res = NULL;
1202 	return ret;
1203 }
1204 
1205 static enum resp_states acknowledge(struct rxe_qp *qp,
1206 				    struct rxe_pkt_info *pkt)
1207 {
1208 	if (qp_type(qp) != IB_QPT_RC)
1209 		return RESPST_CLEANUP;
1210 
1211 	if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1212 		send_ack(qp, qp->resp.aeth_syndrome, pkt->psn);
1213 	else if (pkt->mask & RXE_ATOMIC_MASK)
1214 		send_atomic_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1215 	else if (pkt->mask & (RXE_FLUSH_MASK | RXE_ATOMIC_WRITE_MASK))
1216 		send_read_response_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1217 	else if (bth_ack(pkt))
1218 		send_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1219 
1220 	return RESPST_CLEANUP;
1221 }
1222 
1223 static enum resp_states cleanup(struct rxe_qp *qp,
1224 				struct rxe_pkt_info *pkt)
1225 {
1226 	struct sk_buff *skb;
1227 
1228 	if (pkt) {
1229 		skb = skb_dequeue(&qp->req_pkts);
1230 		rxe_put(qp);
1231 		kfree_skb(skb);
1232 		ib_device_put(qp->ibqp.device);
1233 	}
1234 
1235 	if (qp->resp.mr) {
1236 		rxe_put(qp->resp.mr);
1237 		qp->resp.mr = NULL;
1238 	}
1239 
1240 	return RESPST_DONE;
1241 }
1242 
1243 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1244 {
1245 	int i;
1246 
1247 	for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
1248 		struct resp_res *res = &qp->resp.resources[i];
1249 
1250 		if (res->type == 0)
1251 			continue;
1252 
1253 		if (psn_compare(psn, res->first_psn) >= 0 &&
1254 		    psn_compare(psn, res->last_psn) <= 0) {
1255 			return res;
1256 		}
1257 	}
1258 
1259 	return NULL;
1260 }
1261 
1262 static enum resp_states duplicate_request(struct rxe_qp *qp,
1263 					  struct rxe_pkt_info *pkt)
1264 {
1265 	enum resp_states rc;
1266 	u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK;
1267 
1268 	if (pkt->mask & RXE_SEND_MASK ||
1269 	    pkt->mask & RXE_WRITE_MASK) {
1270 		/* SEND. Ack again and cleanup. C9-105. */
1271 		send_ack(qp, AETH_ACK_UNLIMITED, prev_psn);
1272 		return RESPST_CLEANUP;
1273 	} else if (pkt->mask & RXE_FLUSH_MASK) {
1274 		struct resp_res *res;
1275 
1276 		/* Find the operation in our list of responder resources. */
1277 		res = find_resource(qp, pkt->psn);
1278 		if (res) {
1279 			res->replay = 1;
1280 			res->cur_psn = pkt->psn;
1281 			qp->resp.res = res;
1282 			rc = RESPST_PROCESS_FLUSH;
1283 			goto out;
1284 		}
1285 
1286 		/* Resource not found. Class D error. Drop the request. */
1287 		rc = RESPST_CLEANUP;
1288 		goto out;
1289 	} else if (pkt->mask & RXE_READ_MASK) {
1290 		struct resp_res *res;
1291 
1292 		res = find_resource(qp, pkt->psn);
1293 		if (!res) {
1294 			/* Resource not found. Class D error.  Drop the
1295 			 * request.
1296 			 */
1297 			rc = RESPST_CLEANUP;
1298 			goto out;
1299 		} else {
1300 			/* Ensure this new request is the same as the previous
1301 			 * one or a subset of it.
1302 			 */
1303 			u64 iova = reth_va(pkt);
1304 			u32 resid = reth_len(pkt);
1305 
1306 			if (iova < res->read.va_org ||
1307 			    resid > res->read.length ||
1308 			    (iova + resid) > (res->read.va_org +
1309 					      res->read.length)) {
1310 				rc = RESPST_CLEANUP;
1311 				goto out;
1312 			}
1313 
1314 			if (reth_rkey(pkt) != res->read.rkey) {
1315 				rc = RESPST_CLEANUP;
1316 				goto out;
1317 			}
1318 
1319 			res->cur_psn = pkt->psn;
1320 			res->state = (pkt->psn == res->first_psn) ?
1321 					rdatm_res_state_new :
1322 					rdatm_res_state_replay;
1323 			res->replay = 1;
1324 
1325 			/* Reset the resource, except length. */
1326 			res->read.va_org = iova;
1327 			res->read.va = iova;
1328 			res->read.resid = resid;
1329 
1330 			/* Replay the RDMA read reply. */
1331 			qp->resp.res = res;
1332 			rc = RESPST_READ_REPLY;
1333 			goto out;
1334 		}
1335 	} else {
1336 		struct resp_res *res;
1337 
1338 		/* Find the operation in our list of responder resources. */
1339 		res = find_resource(qp, pkt->psn);
1340 		if (res) {
1341 			res->replay = 1;
1342 			res->cur_psn = pkt->psn;
1343 			qp->resp.res = res;
1344 			rc = pkt->mask & RXE_ATOMIC_MASK ?
1345 					RESPST_ATOMIC_REPLY :
1346 					RESPST_ATOMIC_WRITE_REPLY;
1347 			goto out;
1348 		}
1349 
1350 		/* Resource not found. Class D error. Drop the request. */
1351 		rc = RESPST_CLEANUP;
1352 		goto out;
1353 	}
1354 out:
1355 	return rc;
1356 }
1357 
1358 /* Process a class A or C. Both are treated the same in this implementation. */
1359 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1360 			      enum ib_wc_status status)
1361 {
1362 	qp->resp.aeth_syndrome	= syndrome;
1363 	qp->resp.status		= status;
1364 
1365 	/* indicate that we should go through the ERROR state */
1366 	qp->resp.goto_error	= 1;
1367 }
1368 
1369 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1370 {
1371 	/* UC */
1372 	if (qp->srq) {
1373 		/* Class E */
1374 		qp->resp.drop_msg = 1;
1375 		if (qp->resp.wqe) {
1376 			qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1377 			return RESPST_COMPLETE;
1378 		} else {
1379 			return RESPST_CLEANUP;
1380 		}
1381 	} else {
1382 		/* Class D1. This packet may be the start of a
1383 		 * new message and could be valid. The previous
1384 		 * message is invalid and ignored. reset the
1385 		 * recv wr to its original state
1386 		 */
1387 		if (qp->resp.wqe) {
1388 			qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1389 			qp->resp.wqe->dma.cur_sge = 0;
1390 			qp->resp.wqe->dma.sge_offset = 0;
1391 			qp->resp.opcode = -1;
1392 		}
1393 
1394 		if (qp->resp.mr) {
1395 			rxe_put(qp->resp.mr);
1396 			qp->resp.mr = NULL;
1397 		}
1398 
1399 		return RESPST_CLEANUP;
1400 	}
1401 }
1402 
1403 /* drain incoming request packet queue */
1404 static void drain_req_pkts(struct rxe_qp *qp)
1405 {
1406 	struct sk_buff *skb;
1407 
1408 	while ((skb = skb_dequeue(&qp->req_pkts))) {
1409 		rxe_put(qp);
1410 		kfree_skb(skb);
1411 		ib_device_put(qp->ibqp.device);
1412 	}
1413 }
1414 
1415 /* complete receive wqe with flush error */
1416 static int flush_recv_wqe(struct rxe_qp *qp, struct rxe_recv_wqe *wqe)
1417 {
1418 	struct rxe_cqe cqe = {};
1419 	struct ib_wc *wc = &cqe.ibwc;
1420 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
1421 	int err;
1422 
1423 	if (qp->rcq->is_user) {
1424 		uwc->wr_id = wqe->wr_id;
1425 		uwc->status = IB_WC_WR_FLUSH_ERR;
1426 		uwc->qp_num = qp_num(qp);
1427 	} else {
1428 		wc->wr_id = wqe->wr_id;
1429 		wc->status = IB_WC_WR_FLUSH_ERR;
1430 		wc->qp = &qp->ibqp;
1431 	}
1432 
1433 	err = rxe_cq_post(qp->rcq, &cqe, 0);
1434 	if (err)
1435 		rxe_dbg_cq(qp->rcq, "post cq failed err = %d", err);
1436 
1437 	return err;
1438 }
1439 
1440 /* drain and optionally complete the recive queue
1441  * if unable to complete a wqe stop completing and
1442  * just flush the remaining wqes
1443  */
1444 static void flush_recv_queue(struct rxe_qp *qp, bool notify)
1445 {
1446 	struct rxe_queue *q = qp->rq.queue;
1447 	struct rxe_recv_wqe *wqe;
1448 	int err;
1449 
1450 	if (qp->srq)
1451 		return;
1452 
1453 	while ((wqe = queue_head(q, q->type))) {
1454 		if (notify) {
1455 			err = flush_recv_wqe(qp, wqe);
1456 			if (err)
1457 				notify = 0;
1458 		}
1459 		queue_advance_consumer(q, q->type);
1460 	}
1461 
1462 	qp->resp.wqe = NULL;
1463 }
1464 
1465 int rxe_responder(struct rxe_qp *qp)
1466 {
1467 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1468 	enum resp_states state;
1469 	struct rxe_pkt_info *pkt = NULL;
1470 	int ret;
1471 
1472 	spin_lock_bh(&qp->state_lock);
1473 	if (!qp->valid || qp_state(qp) == IB_QPS_ERR ||
1474 			  qp_state(qp) == IB_QPS_RESET) {
1475 		bool notify = qp->valid && (qp_state(qp) == IB_QPS_ERR);
1476 
1477 		drain_req_pkts(qp);
1478 		flush_recv_queue(qp, notify);
1479 		spin_unlock_bh(&qp->state_lock);
1480 		goto exit;
1481 	}
1482 	spin_unlock_bh(&qp->state_lock);
1483 
1484 	qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1485 
1486 	state = RESPST_GET_REQ;
1487 
1488 	while (1) {
1489 		rxe_dbg_qp(qp, "state = %s\n", resp_state_name[state]);
1490 		switch (state) {
1491 		case RESPST_GET_REQ:
1492 			state = get_req(qp, &pkt);
1493 			break;
1494 		case RESPST_CHK_PSN:
1495 			state = check_psn(qp, pkt);
1496 			break;
1497 		case RESPST_CHK_OP_SEQ:
1498 			state = check_op_seq(qp, pkt);
1499 			break;
1500 		case RESPST_CHK_OP_VALID:
1501 			state = check_op_valid(qp, pkt);
1502 			break;
1503 		case RESPST_CHK_RESOURCE:
1504 			state = check_resource(qp, pkt);
1505 			break;
1506 		case RESPST_CHK_LENGTH:
1507 			state = rxe_resp_check_length(qp, pkt);
1508 			break;
1509 		case RESPST_CHK_RKEY:
1510 			state = check_rkey(qp, pkt);
1511 			break;
1512 		case RESPST_EXECUTE:
1513 			state = execute(qp, pkt);
1514 			break;
1515 		case RESPST_COMPLETE:
1516 			state = do_complete(qp, pkt);
1517 			break;
1518 		case RESPST_READ_REPLY:
1519 			state = read_reply(qp, pkt);
1520 			break;
1521 		case RESPST_ATOMIC_REPLY:
1522 			state = atomic_reply(qp, pkt);
1523 			break;
1524 		case RESPST_ATOMIC_WRITE_REPLY:
1525 			state = atomic_write_reply(qp, pkt);
1526 			break;
1527 		case RESPST_PROCESS_FLUSH:
1528 			state = process_flush(qp, pkt);
1529 			break;
1530 		case RESPST_ACKNOWLEDGE:
1531 			state = acknowledge(qp, pkt);
1532 			break;
1533 		case RESPST_CLEANUP:
1534 			state = cleanup(qp, pkt);
1535 			break;
1536 		case RESPST_DUPLICATE_REQUEST:
1537 			state = duplicate_request(qp, pkt);
1538 			break;
1539 		case RESPST_ERR_PSN_OUT_OF_SEQ:
1540 			/* RC only - Class B. Drop packet. */
1541 			send_ack(qp, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1542 			state = RESPST_CLEANUP;
1543 			break;
1544 
1545 		case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1546 		case RESPST_ERR_MISSING_OPCODE_FIRST:
1547 		case RESPST_ERR_MISSING_OPCODE_LAST_C:
1548 		case RESPST_ERR_UNSUPPORTED_OPCODE:
1549 		case RESPST_ERR_MISALIGNED_ATOMIC:
1550 			/* RC Only - Class C. */
1551 			do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1552 					  IB_WC_REM_INV_REQ_ERR);
1553 			state = RESPST_COMPLETE;
1554 			break;
1555 
1556 		case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1557 			state = do_class_d1e_error(qp);
1558 			break;
1559 		case RESPST_ERR_RNR:
1560 			if (qp_type(qp) == IB_QPT_RC) {
1561 				rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1562 				/* RC - class B */
1563 				send_ack(qp, AETH_RNR_NAK |
1564 					 (~AETH_TYPE_MASK &
1565 					 qp->attr.min_rnr_timer),
1566 					 pkt->psn);
1567 			} else {
1568 				/* UD/UC - class D */
1569 				qp->resp.drop_msg = 1;
1570 			}
1571 			state = RESPST_CLEANUP;
1572 			break;
1573 
1574 		case RESPST_ERR_RKEY_VIOLATION:
1575 			if (qp_type(qp) == IB_QPT_RC) {
1576 				/* Class C */
1577 				do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1578 						  IB_WC_REM_ACCESS_ERR);
1579 				state = RESPST_COMPLETE;
1580 			} else {
1581 				qp->resp.drop_msg = 1;
1582 				if (qp->srq) {
1583 					/* UC/SRQ Class D */
1584 					qp->resp.status = IB_WC_REM_ACCESS_ERR;
1585 					state = RESPST_COMPLETE;
1586 				} else {
1587 					/* UC/non-SRQ Class E. */
1588 					state = RESPST_CLEANUP;
1589 				}
1590 			}
1591 			break;
1592 
1593 		case RESPST_ERR_INVALIDATE_RKEY:
1594 			/* RC - Class J. */
1595 			qp->resp.goto_error = 1;
1596 			qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1597 			state = RESPST_COMPLETE;
1598 			break;
1599 
1600 		case RESPST_ERR_LENGTH:
1601 			if (qp_type(qp) == IB_QPT_RC) {
1602 				/* Class C */
1603 				do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1604 						  IB_WC_REM_INV_REQ_ERR);
1605 				state = RESPST_COMPLETE;
1606 			} else if (qp->srq) {
1607 				/* UC/UD - class E */
1608 				qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1609 				state = RESPST_COMPLETE;
1610 			} else {
1611 				/* UC/UD - class D */
1612 				qp->resp.drop_msg = 1;
1613 				state = RESPST_CLEANUP;
1614 			}
1615 			break;
1616 
1617 		case RESPST_ERR_MALFORMED_WQE:
1618 			/* All, Class A. */
1619 			do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1620 					  IB_WC_LOC_QP_OP_ERR);
1621 			state = RESPST_COMPLETE;
1622 			break;
1623 
1624 		case RESPST_ERR_CQ_OVERFLOW:
1625 			/* All - Class G */
1626 			state = RESPST_ERROR;
1627 			break;
1628 
1629 		case RESPST_DONE:
1630 			if (qp->resp.goto_error) {
1631 				state = RESPST_ERROR;
1632 				break;
1633 			}
1634 
1635 			goto done;
1636 
1637 		case RESPST_EXIT:
1638 			if (qp->resp.goto_error) {
1639 				state = RESPST_ERROR;
1640 				break;
1641 			}
1642 
1643 			goto exit;
1644 
1645 		case RESPST_ERROR:
1646 			qp->resp.goto_error = 0;
1647 			rxe_dbg_qp(qp, "moved to error state\n");
1648 			rxe_qp_error(qp);
1649 			goto exit;
1650 
1651 		default:
1652 			WARN_ON_ONCE(1);
1653 		}
1654 	}
1655 
1656 	/* A non-zero return value will cause rxe_do_task to
1657 	 * exit its loop and end the tasklet. A zero return
1658 	 * will continue looping and return to rxe_responder
1659 	 */
1660 done:
1661 	ret = 0;
1662 	goto out;
1663 exit:
1664 	ret = -EAGAIN;
1665 out:
1666 	return ret;
1667 }
1668