xref: /linux/drivers/infiniband/sw/rxe/rxe_resp.c (revision a0efa2f362a69e47b9d8b48f770ef3a0249a7911)
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 	skb_queue_tail(&qp->req_pkts, skb);
53 	rxe_sched_task(&qp->recv_task);
54 }
55 
56 static inline enum resp_states get_req(struct rxe_qp *qp,
57 				       struct rxe_pkt_info **pkt_p)
58 {
59 	struct sk_buff *skb;
60 
61 	skb = skb_peek(&qp->req_pkts);
62 	if (!skb)
63 		return RESPST_EXIT;
64 
65 	*pkt_p = SKB_TO_PKT(skb);
66 
67 	return (qp->resp.res) ? RESPST_READ_REPLY : RESPST_CHK_PSN;
68 }
69 
70 static enum resp_states check_psn(struct rxe_qp *qp,
71 				  struct rxe_pkt_info *pkt)
72 {
73 	int diff = psn_compare(pkt->psn, qp->resp.psn);
74 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
75 
76 	switch (qp_type(qp)) {
77 	case IB_QPT_RC:
78 		if (diff > 0) {
79 			if (qp->resp.sent_psn_nak)
80 				return RESPST_CLEANUP;
81 
82 			qp->resp.sent_psn_nak = 1;
83 			rxe_counter_inc(rxe, RXE_CNT_OUT_OF_SEQ_REQ);
84 			return RESPST_ERR_PSN_OUT_OF_SEQ;
85 
86 		} else if (diff < 0) {
87 			rxe_counter_inc(rxe, RXE_CNT_DUP_REQ);
88 			return RESPST_DUPLICATE_REQUEST;
89 		}
90 
91 		if (qp->resp.sent_psn_nak)
92 			qp->resp.sent_psn_nak = 0;
93 
94 		break;
95 
96 	case IB_QPT_UC:
97 		if (qp->resp.drop_msg || diff != 0) {
98 			if (pkt->mask & RXE_START_MASK) {
99 				qp->resp.drop_msg = 0;
100 				return RESPST_CHK_OP_SEQ;
101 			}
102 
103 			qp->resp.drop_msg = 1;
104 			return RESPST_CLEANUP;
105 		}
106 		break;
107 	default:
108 		break;
109 	}
110 
111 	return RESPST_CHK_OP_SEQ;
112 }
113 
114 static enum resp_states check_op_seq(struct rxe_qp *qp,
115 				     struct rxe_pkt_info *pkt)
116 {
117 	switch (qp_type(qp)) {
118 	case IB_QPT_RC:
119 		switch (qp->resp.opcode) {
120 		case IB_OPCODE_RC_SEND_FIRST:
121 		case IB_OPCODE_RC_SEND_MIDDLE:
122 			switch (pkt->opcode) {
123 			case IB_OPCODE_RC_SEND_MIDDLE:
124 			case IB_OPCODE_RC_SEND_LAST:
125 			case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
126 			case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
127 				return RESPST_CHK_OP_VALID;
128 			default:
129 				return RESPST_ERR_MISSING_OPCODE_LAST_C;
130 			}
131 
132 		case IB_OPCODE_RC_RDMA_WRITE_FIRST:
133 		case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
134 			switch (pkt->opcode) {
135 			case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
136 			case IB_OPCODE_RC_RDMA_WRITE_LAST:
137 			case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
138 				return RESPST_CHK_OP_VALID;
139 			default:
140 				return RESPST_ERR_MISSING_OPCODE_LAST_C;
141 			}
142 
143 		default:
144 			switch (pkt->opcode) {
145 			case IB_OPCODE_RC_SEND_MIDDLE:
146 			case IB_OPCODE_RC_SEND_LAST:
147 			case IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE:
148 			case IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE:
149 			case IB_OPCODE_RC_RDMA_WRITE_MIDDLE:
150 			case IB_OPCODE_RC_RDMA_WRITE_LAST:
151 			case IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
152 				return RESPST_ERR_MISSING_OPCODE_FIRST;
153 			default:
154 				return RESPST_CHK_OP_VALID;
155 			}
156 		}
157 		break;
158 
159 	case IB_QPT_UC:
160 		switch (qp->resp.opcode) {
161 		case IB_OPCODE_UC_SEND_FIRST:
162 		case IB_OPCODE_UC_SEND_MIDDLE:
163 			switch (pkt->opcode) {
164 			case IB_OPCODE_UC_SEND_MIDDLE:
165 			case IB_OPCODE_UC_SEND_LAST:
166 			case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
167 				return RESPST_CHK_OP_VALID;
168 			default:
169 				return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
170 			}
171 
172 		case IB_OPCODE_UC_RDMA_WRITE_FIRST:
173 		case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
174 			switch (pkt->opcode) {
175 			case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
176 			case IB_OPCODE_UC_RDMA_WRITE_LAST:
177 			case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
178 				return RESPST_CHK_OP_VALID;
179 			default:
180 				return RESPST_ERR_MISSING_OPCODE_LAST_D1E;
181 			}
182 
183 		default:
184 			switch (pkt->opcode) {
185 			case IB_OPCODE_UC_SEND_MIDDLE:
186 			case IB_OPCODE_UC_SEND_LAST:
187 			case IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE:
188 			case IB_OPCODE_UC_RDMA_WRITE_MIDDLE:
189 			case IB_OPCODE_UC_RDMA_WRITE_LAST:
190 			case IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE:
191 				qp->resp.drop_msg = 1;
192 				return RESPST_CLEANUP;
193 			default:
194 				return RESPST_CHK_OP_VALID;
195 			}
196 		}
197 		break;
198 
199 	default:
200 		return RESPST_CHK_OP_VALID;
201 	}
202 }
203 
204 static bool check_qp_attr_access(struct rxe_qp *qp,
205 				 struct rxe_pkt_info *pkt)
206 {
207 	if (((pkt->mask & RXE_READ_MASK) &&
208 	     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_READ)) ||
209 	    ((pkt->mask & (RXE_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) &&
210 	     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) ||
211 	    ((pkt->mask & RXE_ATOMIC_MASK) &&
212 	     !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
213 		return false;
214 
215 	if (pkt->mask & RXE_FLUSH_MASK) {
216 		u32 flush_type = feth_plt(pkt);
217 
218 		if ((flush_type & IB_FLUSH_GLOBAL &&
219 		     !(qp->attr.qp_access_flags & IB_ACCESS_FLUSH_GLOBAL)) ||
220 		    (flush_type & IB_FLUSH_PERSISTENT &&
221 		     !(qp->attr.qp_access_flags & IB_ACCESS_FLUSH_PERSISTENT)))
222 			return false;
223 	}
224 
225 	return true;
226 }
227 
228 static enum resp_states check_op_valid(struct rxe_qp *qp,
229 				       struct rxe_pkt_info *pkt)
230 {
231 	switch (qp_type(qp)) {
232 	case IB_QPT_RC:
233 		if (!check_qp_attr_access(qp, pkt))
234 			return RESPST_ERR_UNSUPPORTED_OPCODE;
235 
236 		break;
237 
238 	case IB_QPT_UC:
239 		if ((pkt->mask & RXE_WRITE_MASK) &&
240 		    !(qp->attr.qp_access_flags & IB_ACCESS_REMOTE_WRITE)) {
241 			qp->resp.drop_msg = 1;
242 			return RESPST_CLEANUP;
243 		}
244 
245 		break;
246 
247 	case IB_QPT_UD:
248 	case IB_QPT_GSI:
249 		break;
250 
251 	default:
252 		WARN_ON_ONCE(1);
253 		break;
254 	}
255 
256 	return RESPST_CHK_RESOURCE;
257 }
258 
259 static enum resp_states get_srq_wqe(struct rxe_qp *qp)
260 {
261 	struct rxe_srq *srq = qp->srq;
262 	struct rxe_queue *q = srq->rq.queue;
263 	struct rxe_recv_wqe *wqe;
264 	struct ib_event ev;
265 	unsigned int count;
266 	size_t size;
267 	unsigned long flags;
268 
269 	if (srq->error)
270 		return RESPST_ERR_RNR;
271 
272 	spin_lock_irqsave(&srq->rq.consumer_lock, flags);
273 
274 	wqe = queue_head(q, QUEUE_TYPE_FROM_CLIENT);
275 	if (!wqe) {
276 		spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
277 		return RESPST_ERR_RNR;
278 	}
279 
280 	/* don't trust user space data */
281 	if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) {
282 		spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
283 		rxe_dbg_qp(qp, "invalid num_sge in SRQ entry\n");
284 		return RESPST_ERR_MALFORMED_WQE;
285 	}
286 	size = sizeof(*wqe) + wqe->dma.num_sge*sizeof(struct rxe_sge);
287 	memcpy(&qp->resp.srq_wqe, wqe, size);
288 
289 	qp->resp.wqe = &qp->resp.srq_wqe.wqe;
290 	queue_advance_consumer(q, QUEUE_TYPE_FROM_CLIENT);
291 	count = queue_count(q, QUEUE_TYPE_FROM_CLIENT);
292 
293 	if (srq->limit && srq->ibsrq.event_handler && (count < srq->limit)) {
294 		srq->limit = 0;
295 		goto event;
296 	}
297 
298 	spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
299 	return RESPST_CHK_LENGTH;
300 
301 event:
302 	spin_unlock_irqrestore(&srq->rq.consumer_lock, flags);
303 	ev.device = qp->ibqp.device;
304 	ev.element.srq = qp->ibqp.srq;
305 	ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
306 	srq->ibsrq.event_handler(&ev, srq->ibsrq.srq_context);
307 	return RESPST_CHK_LENGTH;
308 }
309 
310 static enum resp_states check_resource(struct rxe_qp *qp,
311 				       struct rxe_pkt_info *pkt)
312 {
313 	struct rxe_srq *srq = qp->srq;
314 
315 	if (pkt->mask & (RXE_READ_OR_ATOMIC_MASK | RXE_ATOMIC_WRITE_MASK)) {
316 		/* it is the requesters job to not send
317 		 * too many read/atomic ops, we just
318 		 * recycle the responder resource queue
319 		 */
320 		if (likely(qp->attr.max_dest_rd_atomic > 0))
321 			return RESPST_CHK_LENGTH;
322 		else
323 			return RESPST_ERR_TOO_MANY_RDMA_ATM_REQ;
324 	}
325 
326 	if (pkt->mask & RXE_RWR_MASK) {
327 		if (srq)
328 			return get_srq_wqe(qp);
329 
330 		qp->resp.wqe = queue_head(qp->rq.queue,
331 				QUEUE_TYPE_FROM_CLIENT);
332 		return (qp->resp.wqe) ? RESPST_CHK_LENGTH : RESPST_ERR_RNR;
333 	}
334 
335 	return RESPST_CHK_LENGTH;
336 }
337 
338 static enum resp_states rxe_resp_check_length(struct rxe_qp *qp,
339 					      struct rxe_pkt_info *pkt)
340 {
341 	/*
342 	 * See IBA C9-92
343 	 * For UD QPs we only check if the packet will fit in the
344 	 * receive buffer later. For RDMA operations additional
345 	 * length checks are performed in check_rkey.
346 	 */
347 	if ((qp_type(qp) == IB_QPT_GSI) || (qp_type(qp) == IB_QPT_UD)) {
348 		unsigned int payload = payload_size(pkt);
349 		unsigned int recv_buffer_len = 0;
350 		int i;
351 
352 		for (i = 0; i < qp->resp.wqe->dma.num_sge; i++)
353 			recv_buffer_len += qp->resp.wqe->dma.sge[i].length;
354 		if (payload + sizeof(union rdma_network_hdr) > recv_buffer_len) {
355 			rxe_dbg_qp(qp, "The receive buffer is too small for this UD packet.\n");
356 			return RESPST_ERR_LENGTH;
357 		}
358 	}
359 
360 	if (pkt->mask & RXE_PAYLOAD_MASK && ((qp_type(qp) == IB_QPT_RC) ||
361 					     (qp_type(qp) == IB_QPT_UC))) {
362 		unsigned int mtu = qp->mtu;
363 		unsigned int payload = payload_size(pkt);
364 
365 		if ((pkt->mask & RXE_START_MASK) &&
366 		    (pkt->mask & RXE_END_MASK)) {
367 			if (unlikely(payload > mtu)) {
368 				rxe_dbg_qp(qp, "only packet too long\n");
369 				return RESPST_ERR_LENGTH;
370 			}
371 		} else if ((pkt->mask & RXE_START_MASK) ||
372 			   (pkt->mask & RXE_MIDDLE_MASK)) {
373 			if (unlikely(payload != mtu)) {
374 				rxe_dbg_qp(qp, "first or middle packet not mtu\n");
375 				return RESPST_ERR_LENGTH;
376 			}
377 		} else if (pkt->mask & RXE_END_MASK) {
378 			if (unlikely((payload == 0) || (payload > mtu))) {
379 				rxe_dbg_qp(qp, "last packet zero or too long\n");
380 				return RESPST_ERR_LENGTH;
381 			}
382 		}
383 	}
384 
385 	/* See IBA C9-94 */
386 	if (pkt->mask & RXE_RETH_MASK) {
387 		if (reth_len(pkt) > (1U << 31)) {
388 			rxe_dbg_qp(qp, "dma length too long\n");
389 			return RESPST_ERR_LENGTH;
390 		}
391 	}
392 
393 	if (pkt->mask & RXE_RDMA_OP_MASK)
394 		return RESPST_CHK_RKEY;
395 	else
396 		return RESPST_EXECUTE;
397 }
398 
399 /* if the reth length field is zero we can assume nothing
400  * about the rkey value and should not validate or use it.
401  * Instead set qp->resp.rkey to 0 which is an invalid rkey
402  * value since the minimum index part is 1.
403  */
404 static void qp_resp_from_reth(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
405 {
406 	unsigned int length = reth_len(pkt);
407 
408 	qp->resp.va = reth_va(pkt);
409 	qp->resp.offset = 0;
410 	qp->resp.resid = length;
411 	qp->resp.length = length;
412 	if (pkt->mask & RXE_READ_OR_WRITE_MASK && length == 0)
413 		qp->resp.rkey = 0;
414 	else
415 		qp->resp.rkey = reth_rkey(pkt);
416 }
417 
418 static void qp_resp_from_atmeth(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
419 {
420 	qp->resp.va = atmeth_va(pkt);
421 	qp->resp.offset = 0;
422 	qp->resp.rkey = atmeth_rkey(pkt);
423 	qp->resp.resid = sizeof(u64);
424 }
425 
426 /* resolve the packet rkey to qp->resp.mr or set qp->resp.mr to NULL
427  * if an invalid rkey is received or the rdma length is zero. For middle
428  * or last packets use the stored value of mr.
429  */
430 static enum resp_states check_rkey(struct rxe_qp *qp,
431 				   struct rxe_pkt_info *pkt)
432 {
433 	struct rxe_mr *mr = NULL;
434 	struct rxe_mw *mw = NULL;
435 	u64 va;
436 	u32 rkey;
437 	u32 resid;
438 	u32 pktlen;
439 	int mtu = qp->mtu;
440 	enum resp_states state;
441 	int access = 0;
442 
443 	/* parse RETH or ATMETH header for first/only packets
444 	 * for va, length, rkey, etc. or use current value for
445 	 * middle/last packets.
446 	 */
447 	if (pkt->mask & (RXE_READ_OR_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) {
448 		if (pkt->mask & RXE_RETH_MASK)
449 			qp_resp_from_reth(qp, pkt);
450 
451 		access = (pkt->mask & RXE_READ_MASK) ? IB_ACCESS_REMOTE_READ
452 						     : IB_ACCESS_REMOTE_WRITE;
453 	} else if (pkt->mask & RXE_FLUSH_MASK) {
454 		u32 flush_type = feth_plt(pkt);
455 
456 		if (pkt->mask & RXE_RETH_MASK)
457 			qp_resp_from_reth(qp, pkt);
458 
459 		if (flush_type & IB_FLUSH_GLOBAL)
460 			access |= IB_ACCESS_FLUSH_GLOBAL;
461 		if (flush_type & IB_FLUSH_PERSISTENT)
462 			access |= IB_ACCESS_FLUSH_PERSISTENT;
463 	} else if (pkt->mask & RXE_ATOMIC_MASK) {
464 		qp_resp_from_atmeth(qp, pkt);
465 		access = IB_ACCESS_REMOTE_ATOMIC;
466 	} else {
467 		/* shouldn't happen */
468 		WARN_ON(1);
469 	}
470 
471 	/* A zero-byte read or write op is not required to
472 	 * set an addr or rkey. See C9-88
473 	 */
474 	if ((pkt->mask & RXE_READ_OR_WRITE_MASK) &&
475 	    (pkt->mask & RXE_RETH_MASK) && reth_len(pkt) == 0) {
476 		qp->resp.mr = NULL;
477 		return RESPST_EXECUTE;
478 	}
479 
480 	va	= qp->resp.va;
481 	rkey	= qp->resp.rkey;
482 	resid	= qp->resp.resid;
483 	pktlen	= payload_size(pkt);
484 
485 	if (rkey_is_mw(rkey)) {
486 		mw = rxe_lookup_mw(qp, access, rkey);
487 		if (!mw) {
488 			rxe_dbg_qp(qp, "no MW matches rkey %#x\n", rkey);
489 			state = RESPST_ERR_RKEY_VIOLATION;
490 			goto err;
491 		}
492 
493 		mr = mw->mr;
494 		if (!mr) {
495 			rxe_dbg_qp(qp, "MW doesn't have an MR\n");
496 			state = RESPST_ERR_RKEY_VIOLATION;
497 			goto err;
498 		}
499 
500 		if (mw->access & IB_ZERO_BASED)
501 			qp->resp.offset = mw->addr;
502 
503 		rxe_get(mr);
504 		rxe_put(mw);
505 		mw = NULL;
506 	} else {
507 		mr = lookup_mr(qp->pd, access, rkey, RXE_LOOKUP_REMOTE);
508 		if (!mr) {
509 			rxe_dbg_qp(qp, "no MR matches rkey %#x\n", rkey);
510 			state = RESPST_ERR_RKEY_VIOLATION;
511 			goto err;
512 		}
513 	}
514 
515 	if (pkt->mask & RXE_FLUSH_MASK) {
516 		/* FLUSH MR may not set va or resid
517 		 * no need to check range since we will flush whole mr
518 		 */
519 		if (feth_sel(pkt) == IB_FLUSH_MR)
520 			goto skip_check_range;
521 	}
522 
523 	if (mr_check_range(mr, va + qp->resp.offset, resid)) {
524 		state = RESPST_ERR_RKEY_VIOLATION;
525 		goto err;
526 	}
527 
528 skip_check_range:
529 	if (pkt->mask & (RXE_WRITE_MASK | RXE_ATOMIC_WRITE_MASK)) {
530 		if (resid > mtu) {
531 			if (pktlen != mtu || bth_pad(pkt)) {
532 				state = RESPST_ERR_LENGTH;
533 				goto err;
534 			}
535 		} else {
536 			if (pktlen != resid) {
537 				state = RESPST_ERR_LENGTH;
538 				goto err;
539 			}
540 			if ((bth_pad(pkt) != (0x3 & (-resid)))) {
541 				/* This case may not be exactly that
542 				 * but nothing else fits.
543 				 */
544 				state = RESPST_ERR_LENGTH;
545 				goto err;
546 			}
547 		}
548 	}
549 
550 	WARN_ON_ONCE(qp->resp.mr);
551 
552 	qp->resp.mr = mr;
553 	return RESPST_EXECUTE;
554 
555 err:
556 	qp->resp.mr = NULL;
557 	if (mr)
558 		rxe_put(mr);
559 	if (mw)
560 		rxe_put(mw);
561 
562 	return state;
563 }
564 
565 static enum resp_states send_data_in(struct rxe_qp *qp, void *data_addr,
566 				     int data_len)
567 {
568 	int err;
569 
570 	err = copy_data(qp->pd, IB_ACCESS_LOCAL_WRITE, &qp->resp.wqe->dma,
571 			data_addr, data_len, RXE_TO_MR_OBJ);
572 	if (unlikely(err))
573 		return (err == -ENOSPC) ? RESPST_ERR_LENGTH
574 					: RESPST_ERR_MALFORMED_WQE;
575 
576 	return RESPST_NONE;
577 }
578 
579 static enum resp_states write_data_in(struct rxe_qp *qp,
580 				      struct rxe_pkt_info *pkt)
581 {
582 	enum resp_states rc = RESPST_NONE;
583 	int	err;
584 	int data_len = payload_size(pkt);
585 
586 	err = rxe_mr_copy(qp->resp.mr, qp->resp.va + qp->resp.offset,
587 			  payload_addr(pkt), data_len, RXE_TO_MR_OBJ);
588 	if (err) {
589 		rc = RESPST_ERR_RKEY_VIOLATION;
590 		goto out;
591 	}
592 
593 	qp->resp.va += data_len;
594 	qp->resp.resid -= data_len;
595 
596 out:
597 	return rc;
598 }
599 
600 static struct resp_res *rxe_prepare_res(struct rxe_qp *qp,
601 					struct rxe_pkt_info *pkt,
602 					int type)
603 {
604 	struct resp_res *res;
605 	u32 pkts;
606 
607 	res = &qp->resp.resources[qp->resp.res_head];
608 	rxe_advance_resp_resource(qp);
609 	free_rd_atomic_resource(res);
610 
611 	res->type = type;
612 	res->replay = 0;
613 
614 	switch (type) {
615 	case RXE_READ_MASK:
616 		res->read.va = qp->resp.va + qp->resp.offset;
617 		res->read.va_org = qp->resp.va + qp->resp.offset;
618 		res->read.resid = qp->resp.resid;
619 		res->read.length = qp->resp.resid;
620 		res->read.rkey = qp->resp.rkey;
621 
622 		pkts = max_t(u32, (reth_len(pkt) + qp->mtu - 1)/qp->mtu, 1);
623 		res->first_psn = pkt->psn;
624 		res->cur_psn = pkt->psn;
625 		res->last_psn = (pkt->psn + pkts - 1) & BTH_PSN_MASK;
626 
627 		res->state = rdatm_res_state_new;
628 		break;
629 	case RXE_ATOMIC_MASK:
630 	case RXE_ATOMIC_WRITE_MASK:
631 		res->first_psn = pkt->psn;
632 		res->last_psn = pkt->psn;
633 		res->cur_psn = pkt->psn;
634 		break;
635 	case RXE_FLUSH_MASK:
636 		res->flush.va = qp->resp.va + qp->resp.offset;
637 		res->flush.length = qp->resp.length;
638 		res->flush.type = feth_plt(pkt);
639 		res->flush.level = feth_sel(pkt);
640 	}
641 
642 	return res;
643 }
644 
645 static enum resp_states process_flush(struct rxe_qp *qp,
646 				       struct rxe_pkt_info *pkt)
647 {
648 	u64 length, start;
649 	struct rxe_mr *mr = qp->resp.mr;
650 	struct resp_res *res = qp->resp.res;
651 
652 	/* oA19-14, oA19-15 */
653 	if (res && res->replay)
654 		return RESPST_ACKNOWLEDGE;
655 	else if (!res) {
656 		res = rxe_prepare_res(qp, pkt, RXE_FLUSH_MASK);
657 		qp->resp.res = res;
658 	}
659 
660 	if (res->flush.level == IB_FLUSH_RANGE) {
661 		start = res->flush.va;
662 		length = res->flush.length;
663 	} else { /* level == IB_FLUSH_MR */
664 		start = mr->ibmr.iova;
665 		length = mr->ibmr.length;
666 	}
667 
668 	if (res->flush.type & IB_FLUSH_PERSISTENT) {
669 		if (rxe_flush_pmem_iova(mr, start, length))
670 			return RESPST_ERR_RKEY_VIOLATION;
671 		/* Make data persistent. */
672 		wmb();
673 	} else if (res->flush.type & IB_FLUSH_GLOBAL) {
674 		/* Make data global visibility. */
675 		wmb();
676 	}
677 
678 	qp->resp.msn++;
679 
680 	/* next expected psn, read handles this separately */
681 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
682 	qp->resp.ack_psn = qp->resp.psn;
683 
684 	qp->resp.opcode = pkt->opcode;
685 	qp->resp.status = IB_WC_SUCCESS;
686 
687 	return RESPST_ACKNOWLEDGE;
688 }
689 
690 static enum resp_states atomic_reply(struct rxe_qp *qp,
691 				     struct rxe_pkt_info *pkt)
692 {
693 	struct rxe_mr *mr = qp->resp.mr;
694 	struct resp_res *res = qp->resp.res;
695 	int err;
696 
697 	if (!res) {
698 		res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_MASK);
699 		qp->resp.res = res;
700 	}
701 
702 	if (!res->replay) {
703 		u64 iova = qp->resp.va + qp->resp.offset;
704 
705 		err = rxe_mr_do_atomic_op(mr, iova, pkt->opcode,
706 					  atmeth_comp(pkt),
707 					  atmeth_swap_add(pkt),
708 					  &res->atomic.orig_val);
709 		if (err)
710 			return err;
711 
712 		qp->resp.msn++;
713 
714 		/* next expected psn, read handles this separately */
715 		qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
716 		qp->resp.ack_psn = qp->resp.psn;
717 
718 		qp->resp.opcode = pkt->opcode;
719 		qp->resp.status = IB_WC_SUCCESS;
720 	}
721 
722 	return RESPST_ACKNOWLEDGE;
723 }
724 
725 static enum resp_states atomic_write_reply(struct rxe_qp *qp,
726 					   struct rxe_pkt_info *pkt)
727 {
728 	struct resp_res *res = qp->resp.res;
729 	struct rxe_mr *mr;
730 	u64 value;
731 	u64 iova;
732 	int err;
733 
734 	if (!res) {
735 		res = rxe_prepare_res(qp, pkt, RXE_ATOMIC_WRITE_MASK);
736 		qp->resp.res = res;
737 	}
738 
739 	if (res->replay)
740 		return RESPST_ACKNOWLEDGE;
741 
742 	mr = qp->resp.mr;
743 	value = *(u64 *)payload_addr(pkt);
744 	iova = qp->resp.va + qp->resp.offset;
745 
746 	err = rxe_mr_do_atomic_write(mr, iova, value);
747 	if (err)
748 		return err;
749 
750 	qp->resp.resid = 0;
751 	qp->resp.msn++;
752 
753 	/* next expected psn, read handles this separately */
754 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
755 	qp->resp.ack_psn = qp->resp.psn;
756 
757 	qp->resp.opcode = pkt->opcode;
758 	qp->resp.status = IB_WC_SUCCESS;
759 
760 	return RESPST_ACKNOWLEDGE;
761 }
762 
763 static struct sk_buff *prepare_ack_packet(struct rxe_qp *qp,
764 					  struct rxe_pkt_info *ack,
765 					  int opcode,
766 					  int payload,
767 					  u32 psn,
768 					  u8 syndrome)
769 {
770 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
771 	struct sk_buff *skb;
772 	int paylen;
773 	int pad;
774 	int err;
775 
776 	/*
777 	 * allocate packet
778 	 */
779 	pad = (-payload) & 0x3;
780 	paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
781 
782 	skb = rxe_init_packet(rxe, &qp->pri_av, paylen, ack);
783 	if (!skb)
784 		return NULL;
785 
786 	ack->qp = qp;
787 	ack->opcode = opcode;
788 	ack->mask = rxe_opcode[opcode].mask;
789 	ack->paylen = paylen;
790 	ack->psn = psn;
791 
792 	bth_init(ack, opcode, 0, 0, pad, IB_DEFAULT_PKEY_FULL,
793 		 qp->attr.dest_qp_num, 0, psn);
794 
795 	if (ack->mask & RXE_AETH_MASK) {
796 		aeth_set_syn(ack, syndrome);
797 		aeth_set_msn(ack, qp->resp.msn);
798 	}
799 
800 	if (ack->mask & RXE_ATMACK_MASK)
801 		atmack_set_orig(ack, qp->resp.res->atomic.orig_val);
802 
803 	err = rxe_prepare(&qp->pri_av, ack, skb);
804 	if (err) {
805 		kfree_skb(skb);
806 		return NULL;
807 	}
808 
809 	return skb;
810 }
811 
812 /**
813  * rxe_recheck_mr - revalidate MR from rkey and get a reference
814  * @qp: the qp
815  * @rkey: the rkey
816  *
817  * This code allows the MR to be invalidated or deregistered or
818  * the MW if one was used to be invalidated or deallocated.
819  * It is assumed that the access permissions if originally good
820  * are OK and the mappings to be unchanged.
821  *
822  * TODO: If someone reregisters an MR to change its size or
823  * access permissions during the processing of an RDMA read
824  * we should kill the responder resource and complete the
825  * operation with an error.
826  *
827  * Return: mr on success else NULL
828  */
829 static struct rxe_mr *rxe_recheck_mr(struct rxe_qp *qp, u32 rkey)
830 {
831 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
832 	struct rxe_mr *mr;
833 	struct rxe_mw *mw;
834 
835 	if (rkey_is_mw(rkey)) {
836 		mw = rxe_pool_get_index(&rxe->mw_pool, rkey >> 8);
837 		if (!mw)
838 			return NULL;
839 
840 		mr = mw->mr;
841 		if (mw->rkey != rkey || mw->state != RXE_MW_STATE_VALID ||
842 		    !mr || mr->state != RXE_MR_STATE_VALID) {
843 			rxe_put(mw);
844 			return NULL;
845 		}
846 
847 		rxe_get(mr);
848 		rxe_put(mw);
849 
850 		return mr;
851 	}
852 
853 	mr = rxe_pool_get_index(&rxe->mr_pool, rkey >> 8);
854 	if (!mr)
855 		return NULL;
856 
857 	if (mr->rkey != rkey || mr->state != RXE_MR_STATE_VALID) {
858 		rxe_put(mr);
859 		return NULL;
860 	}
861 
862 	return mr;
863 }
864 
865 /* RDMA read response. If res is not NULL, then we have a current RDMA request
866  * being processed or replayed.
867  */
868 static enum resp_states read_reply(struct rxe_qp *qp,
869 				   struct rxe_pkt_info *req_pkt)
870 {
871 	struct rxe_pkt_info ack_pkt;
872 	struct sk_buff *skb;
873 	int mtu = qp->mtu;
874 	enum resp_states state;
875 	int payload;
876 	int opcode;
877 	int err;
878 	struct resp_res *res = qp->resp.res;
879 	struct rxe_mr *mr;
880 
881 	if (!res) {
882 		res = rxe_prepare_res(qp, req_pkt, RXE_READ_MASK);
883 		qp->resp.res = res;
884 	}
885 
886 	if (res->state == rdatm_res_state_new) {
887 		if (!res->replay || qp->resp.length == 0) {
888 			/* if length == 0 mr will be NULL (is ok)
889 			 * otherwise qp->resp.mr holds a ref on mr
890 			 * which we transfer to mr and drop below.
891 			 */
892 			mr = qp->resp.mr;
893 			qp->resp.mr = NULL;
894 		} else {
895 			mr = rxe_recheck_mr(qp, res->read.rkey);
896 			if (!mr)
897 				return RESPST_ERR_RKEY_VIOLATION;
898 		}
899 
900 		if (res->read.resid <= mtu)
901 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;
902 		else
903 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST;
904 	} else {
905 		/* re-lookup mr from rkey on all later packets.
906 		 * length will be non-zero. This can fail if someone
907 		 * modifies or destroys the mr since the first packet.
908 		 */
909 		mr = rxe_recheck_mr(qp, res->read.rkey);
910 		if (!mr)
911 			return RESPST_ERR_RKEY_VIOLATION;
912 
913 		if (res->read.resid > mtu)
914 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_MIDDLE;
915 		else
916 			opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_LAST;
917 	}
918 
919 	res->state = rdatm_res_state_next;
920 
921 	payload = min_t(int, res->read.resid, mtu);
922 
923 	skb = prepare_ack_packet(qp, &ack_pkt, opcode, payload,
924 				 res->cur_psn, AETH_ACK_UNLIMITED);
925 	if (!skb) {
926 		state = RESPST_ERR_RNR;
927 		goto err_out;
928 	}
929 
930 	err = rxe_mr_copy(mr, res->read.va, payload_addr(&ack_pkt),
931 			  payload, RXE_FROM_MR_OBJ);
932 	if (err) {
933 		kfree_skb(skb);
934 		state = RESPST_ERR_RKEY_VIOLATION;
935 		goto err_out;
936 	}
937 
938 	if (bth_pad(&ack_pkt)) {
939 		u8 *pad = payload_addr(&ack_pkt) + payload;
940 
941 		memset(pad, 0, bth_pad(&ack_pkt));
942 	}
943 
944 	/* rxe_xmit_packet always consumes the skb */
945 	err = rxe_xmit_packet(qp, &ack_pkt, skb);
946 	if (err) {
947 		state = RESPST_ERR_RNR;
948 		goto err_out;
949 	}
950 
951 	res->read.va += payload;
952 	res->read.resid -= payload;
953 	res->cur_psn = (res->cur_psn + 1) & BTH_PSN_MASK;
954 
955 	if (res->read.resid > 0) {
956 		state = RESPST_DONE;
957 	} else {
958 		qp->resp.res = NULL;
959 		if (!res->replay)
960 			qp->resp.opcode = -1;
961 		if (psn_compare(res->cur_psn, qp->resp.psn) >= 0)
962 			qp->resp.psn = res->cur_psn;
963 		state = RESPST_CLEANUP;
964 	}
965 
966 err_out:
967 	if (mr)
968 		rxe_put(mr);
969 	return state;
970 }
971 
972 static int invalidate_rkey(struct rxe_qp *qp, u32 rkey)
973 {
974 	if (rkey_is_mw(rkey))
975 		return rxe_invalidate_mw(qp, rkey);
976 	else
977 		return rxe_invalidate_mr(qp, rkey);
978 }
979 
980 /* Executes a new request. A retried request never reach that function (send
981  * and writes are discarded, and reads and atomics are retried elsewhere.
982  */
983 static enum resp_states execute(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
984 {
985 	enum resp_states err;
986 	struct sk_buff *skb = PKT_TO_SKB(pkt);
987 	union rdma_network_hdr hdr;
988 
989 	if (pkt->mask & RXE_SEND_MASK) {
990 		if (qp_type(qp) == IB_QPT_UD ||
991 		    qp_type(qp) == IB_QPT_GSI) {
992 			if (skb->protocol == htons(ETH_P_IP)) {
993 				memset(&hdr.reserved, 0,
994 						sizeof(hdr.reserved));
995 				memcpy(&hdr.roce4grh, ip_hdr(skb),
996 						sizeof(hdr.roce4grh));
997 				err = send_data_in(qp, &hdr, sizeof(hdr));
998 			} else {
999 				err = send_data_in(qp, ipv6_hdr(skb),
1000 						sizeof(hdr));
1001 			}
1002 			if (err)
1003 				return err;
1004 		}
1005 		err = send_data_in(qp, payload_addr(pkt), payload_size(pkt));
1006 		if (err)
1007 			return err;
1008 	} else if (pkt->mask & RXE_WRITE_MASK) {
1009 		err = write_data_in(qp, pkt);
1010 		if (err)
1011 			return err;
1012 	} else if (pkt->mask & RXE_READ_MASK) {
1013 		/* For RDMA Read we can increment the msn now. See C9-148. */
1014 		qp->resp.msn++;
1015 		return RESPST_READ_REPLY;
1016 	} else if (pkt->mask & RXE_ATOMIC_MASK) {
1017 		return RESPST_ATOMIC_REPLY;
1018 	} else if (pkt->mask & RXE_ATOMIC_WRITE_MASK) {
1019 		return RESPST_ATOMIC_WRITE_REPLY;
1020 	} else if (pkt->mask & RXE_FLUSH_MASK) {
1021 		return RESPST_PROCESS_FLUSH;
1022 	} else {
1023 		/* Unreachable */
1024 		WARN_ON_ONCE(1);
1025 	}
1026 
1027 	if (pkt->mask & RXE_IETH_MASK) {
1028 		u32 rkey = ieth_rkey(pkt);
1029 
1030 		err = invalidate_rkey(qp, rkey);
1031 		if (err)
1032 			return RESPST_ERR_INVALIDATE_RKEY;
1033 	}
1034 
1035 	if (pkt->mask & RXE_END_MASK)
1036 		/* We successfully processed this new request. */
1037 		qp->resp.msn++;
1038 
1039 	/* next expected psn, read handles this separately */
1040 	qp->resp.psn = (pkt->psn + 1) & BTH_PSN_MASK;
1041 	qp->resp.ack_psn = qp->resp.psn;
1042 
1043 	qp->resp.opcode = pkt->opcode;
1044 	qp->resp.status = IB_WC_SUCCESS;
1045 
1046 	if (pkt->mask & RXE_COMP_MASK)
1047 		return RESPST_COMPLETE;
1048 	else if (qp_type(qp) == IB_QPT_RC)
1049 		return RESPST_ACKNOWLEDGE;
1050 	else
1051 		return RESPST_CLEANUP;
1052 }
1053 
1054 static enum resp_states do_complete(struct rxe_qp *qp,
1055 				    struct rxe_pkt_info *pkt)
1056 {
1057 	struct rxe_cqe cqe;
1058 	struct ib_wc *wc = &cqe.ibwc;
1059 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
1060 	struct rxe_recv_wqe *wqe = qp->resp.wqe;
1061 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1062 	unsigned long flags;
1063 
1064 	if (!wqe)
1065 		goto finish;
1066 
1067 	memset(&cqe, 0, sizeof(cqe));
1068 
1069 	if (qp->rcq->is_user) {
1070 		uwc->status		= qp->resp.status;
1071 		uwc->qp_num		= qp->ibqp.qp_num;
1072 		uwc->wr_id		= wqe->wr_id;
1073 	} else {
1074 		wc->status		= qp->resp.status;
1075 		wc->qp			= &qp->ibqp;
1076 		wc->wr_id		= wqe->wr_id;
1077 	}
1078 
1079 	if (wc->status == IB_WC_SUCCESS) {
1080 		rxe_counter_inc(rxe, RXE_CNT_RDMA_RECV);
1081 		wc->opcode = (pkt->mask & RXE_IMMDT_MASK &&
1082 				pkt->mask & RXE_WRITE_MASK) ?
1083 					IB_WC_RECV_RDMA_WITH_IMM : IB_WC_RECV;
1084 		wc->byte_len = (pkt->mask & RXE_IMMDT_MASK &&
1085 				pkt->mask & RXE_WRITE_MASK) ?
1086 					qp->resp.length : wqe->dma.length - wqe->dma.resid;
1087 
1088 		/* fields after byte_len are different between kernel and user
1089 		 * space
1090 		 */
1091 		if (qp->rcq->is_user) {
1092 			uwc->wc_flags = IB_WC_GRH;
1093 
1094 			if (pkt->mask & RXE_IMMDT_MASK) {
1095 				uwc->wc_flags |= IB_WC_WITH_IMM;
1096 				uwc->ex.imm_data = immdt_imm(pkt);
1097 			}
1098 
1099 			if (pkt->mask & RXE_IETH_MASK) {
1100 				uwc->wc_flags |= IB_WC_WITH_INVALIDATE;
1101 				uwc->ex.invalidate_rkey = ieth_rkey(pkt);
1102 			}
1103 
1104 			if (pkt->mask & RXE_DETH_MASK)
1105 				uwc->src_qp = deth_sqp(pkt);
1106 
1107 			uwc->port_num		= qp->attr.port_num;
1108 		} else {
1109 			struct sk_buff *skb = PKT_TO_SKB(pkt);
1110 
1111 			wc->wc_flags = IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE;
1112 			if (skb->protocol == htons(ETH_P_IP))
1113 				wc->network_hdr_type = RDMA_NETWORK_IPV4;
1114 			else
1115 				wc->network_hdr_type = RDMA_NETWORK_IPV6;
1116 
1117 			if (is_vlan_dev(skb->dev)) {
1118 				wc->wc_flags |= IB_WC_WITH_VLAN;
1119 				wc->vlan_id = vlan_dev_vlan_id(skb->dev);
1120 			}
1121 
1122 			if (pkt->mask & RXE_IMMDT_MASK) {
1123 				wc->wc_flags |= IB_WC_WITH_IMM;
1124 				wc->ex.imm_data = immdt_imm(pkt);
1125 			}
1126 
1127 			if (pkt->mask & RXE_IETH_MASK) {
1128 				wc->wc_flags |= IB_WC_WITH_INVALIDATE;
1129 				wc->ex.invalidate_rkey = ieth_rkey(pkt);
1130 			}
1131 
1132 			if (pkt->mask & RXE_DETH_MASK)
1133 				wc->src_qp = deth_sqp(pkt);
1134 
1135 			wc->port_num		= qp->attr.port_num;
1136 		}
1137 	} else {
1138 		if (wc->status != IB_WC_WR_FLUSH_ERR)
1139 			rxe_err_qp(qp, "non-flush error status = %d\n",
1140 				wc->status);
1141 	}
1142 
1143 	/* have copy for srq and reference for !srq */
1144 	if (!qp->srq)
1145 		queue_advance_consumer(qp->rq.queue, QUEUE_TYPE_FROM_CLIENT);
1146 
1147 	qp->resp.wqe = NULL;
1148 
1149 	if (rxe_cq_post(qp->rcq, &cqe, pkt ? bth_se(pkt) : 1))
1150 		return RESPST_ERR_CQ_OVERFLOW;
1151 
1152 finish:
1153 	spin_lock_irqsave(&qp->state_lock, flags);
1154 	if (unlikely(qp_state(qp) == IB_QPS_ERR)) {
1155 		spin_unlock_irqrestore(&qp->state_lock, flags);
1156 		return RESPST_CHK_RESOURCE;
1157 	}
1158 	spin_unlock_irqrestore(&qp->state_lock, flags);
1159 
1160 	if (unlikely(!pkt))
1161 		return RESPST_DONE;
1162 	if (qp_type(qp) == IB_QPT_RC)
1163 		return RESPST_ACKNOWLEDGE;
1164 	else
1165 		return RESPST_CLEANUP;
1166 }
1167 
1168 
1169 static int send_common_ack(struct rxe_qp *qp, u8 syndrome, u32 psn,
1170 				  int opcode, const char *msg)
1171 {
1172 	int err;
1173 	struct rxe_pkt_info ack_pkt;
1174 	struct sk_buff *skb;
1175 
1176 	skb = prepare_ack_packet(qp, &ack_pkt, opcode, 0, psn, syndrome);
1177 	if (!skb)
1178 		return -ENOMEM;
1179 
1180 	err = rxe_xmit_packet(qp, &ack_pkt, skb);
1181 	if (err)
1182 		rxe_dbg_qp(qp, "Failed sending %s\n", msg);
1183 
1184 	return err;
1185 }
1186 
1187 static int send_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1188 {
1189 	return send_common_ack(qp, syndrome, psn,
1190 			IB_OPCODE_RC_ACKNOWLEDGE, "ACK");
1191 }
1192 
1193 static int send_atomic_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1194 {
1195 	int ret = send_common_ack(qp, syndrome, psn,
1196 			IB_OPCODE_RC_ATOMIC_ACKNOWLEDGE, "ATOMIC 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 int send_read_response_ack(struct rxe_qp *qp, u8 syndrome, u32 psn)
1206 {
1207 	int ret = send_common_ack(qp, syndrome, psn,
1208 			IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY,
1209 			"RDMA READ response of length zero ACK");
1210 
1211 	/* have to clear this since it is used to trigger
1212 	 * long read replies
1213 	 */
1214 	qp->resp.res = NULL;
1215 	return ret;
1216 }
1217 
1218 static enum resp_states acknowledge(struct rxe_qp *qp,
1219 				    struct rxe_pkt_info *pkt)
1220 {
1221 	if (qp_type(qp) != IB_QPT_RC)
1222 		return RESPST_CLEANUP;
1223 
1224 	if (qp->resp.aeth_syndrome != AETH_ACK_UNLIMITED)
1225 		send_ack(qp, qp->resp.aeth_syndrome, pkt->psn);
1226 	else if (pkt->mask & RXE_ATOMIC_MASK)
1227 		send_atomic_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1228 	else if (pkt->mask & (RXE_FLUSH_MASK | RXE_ATOMIC_WRITE_MASK))
1229 		send_read_response_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1230 	else if (bth_ack(pkt))
1231 		send_ack(qp, AETH_ACK_UNLIMITED, pkt->psn);
1232 
1233 	return RESPST_CLEANUP;
1234 }
1235 
1236 static enum resp_states cleanup(struct rxe_qp *qp,
1237 				struct rxe_pkt_info *pkt)
1238 {
1239 	struct sk_buff *skb;
1240 
1241 	if (pkt) {
1242 		skb = skb_dequeue(&qp->req_pkts);
1243 		rxe_put(qp);
1244 		kfree_skb(skb);
1245 		ib_device_put(qp->ibqp.device);
1246 	}
1247 
1248 	if (qp->resp.mr) {
1249 		rxe_put(qp->resp.mr);
1250 		qp->resp.mr = NULL;
1251 	}
1252 
1253 	return RESPST_DONE;
1254 }
1255 
1256 static struct resp_res *find_resource(struct rxe_qp *qp, u32 psn)
1257 {
1258 	int i;
1259 
1260 	for (i = 0; i < qp->attr.max_dest_rd_atomic; i++) {
1261 		struct resp_res *res = &qp->resp.resources[i];
1262 
1263 		if (res->type == 0)
1264 			continue;
1265 
1266 		if (psn_compare(psn, res->first_psn) >= 0 &&
1267 		    psn_compare(psn, res->last_psn) <= 0) {
1268 			return res;
1269 		}
1270 	}
1271 
1272 	return NULL;
1273 }
1274 
1275 static enum resp_states duplicate_request(struct rxe_qp *qp,
1276 					  struct rxe_pkt_info *pkt)
1277 {
1278 	enum resp_states rc;
1279 	u32 prev_psn = (qp->resp.ack_psn - 1) & BTH_PSN_MASK;
1280 
1281 	if (pkt->mask & RXE_SEND_MASK ||
1282 	    pkt->mask & RXE_WRITE_MASK) {
1283 		/* SEND. Ack again and cleanup. C9-105. */
1284 		send_ack(qp, AETH_ACK_UNLIMITED, prev_psn);
1285 		return RESPST_CLEANUP;
1286 	} else if (pkt->mask & RXE_FLUSH_MASK) {
1287 		struct resp_res *res;
1288 
1289 		/* Find the operation in our list of responder resources. */
1290 		res = find_resource(qp, pkt->psn);
1291 		if (res) {
1292 			res->replay = 1;
1293 			res->cur_psn = pkt->psn;
1294 			qp->resp.res = res;
1295 			rc = RESPST_PROCESS_FLUSH;
1296 			goto out;
1297 		}
1298 
1299 		/* Resource not found. Class D error. Drop the request. */
1300 		rc = RESPST_CLEANUP;
1301 		goto out;
1302 	} else if (pkt->mask & RXE_READ_MASK) {
1303 		struct resp_res *res;
1304 
1305 		res = find_resource(qp, pkt->psn);
1306 		if (!res) {
1307 			/* Resource not found. Class D error.  Drop the
1308 			 * request.
1309 			 */
1310 			rc = RESPST_CLEANUP;
1311 			goto out;
1312 		} else {
1313 			/* Ensure this new request is the same as the previous
1314 			 * one or a subset of it.
1315 			 */
1316 			u64 iova = reth_va(pkt);
1317 			u32 resid = reth_len(pkt);
1318 
1319 			if (iova < res->read.va_org ||
1320 			    resid > res->read.length ||
1321 			    (iova + resid) > (res->read.va_org +
1322 					      res->read.length)) {
1323 				rc = RESPST_CLEANUP;
1324 				goto out;
1325 			}
1326 
1327 			if (reth_rkey(pkt) != res->read.rkey) {
1328 				rc = RESPST_CLEANUP;
1329 				goto out;
1330 			}
1331 
1332 			res->cur_psn = pkt->psn;
1333 			res->state = (pkt->psn == res->first_psn) ?
1334 					rdatm_res_state_new :
1335 					rdatm_res_state_replay;
1336 			res->replay = 1;
1337 
1338 			/* Reset the resource, except length. */
1339 			res->read.va_org = iova;
1340 			res->read.va = iova;
1341 			res->read.resid = resid;
1342 
1343 			/* Replay the RDMA read reply. */
1344 			qp->resp.res = res;
1345 			rc = RESPST_READ_REPLY;
1346 			goto out;
1347 		}
1348 	} else {
1349 		struct resp_res *res;
1350 
1351 		/* Find the operation in our list of responder resources. */
1352 		res = find_resource(qp, pkt->psn);
1353 		if (res) {
1354 			res->replay = 1;
1355 			res->cur_psn = pkt->psn;
1356 			qp->resp.res = res;
1357 			rc = pkt->mask & RXE_ATOMIC_MASK ?
1358 					RESPST_ATOMIC_REPLY :
1359 					RESPST_ATOMIC_WRITE_REPLY;
1360 			goto out;
1361 		}
1362 
1363 		/* Resource not found. Class D error. Drop the request. */
1364 		rc = RESPST_CLEANUP;
1365 		goto out;
1366 	}
1367 out:
1368 	return rc;
1369 }
1370 
1371 /* Process a class A or C. Both are treated the same in this implementation. */
1372 static void do_class_ac_error(struct rxe_qp *qp, u8 syndrome,
1373 			      enum ib_wc_status status)
1374 {
1375 	qp->resp.aeth_syndrome	= syndrome;
1376 	qp->resp.status		= status;
1377 
1378 	/* indicate that we should go through the ERROR state */
1379 	qp->resp.goto_error	= 1;
1380 }
1381 
1382 static enum resp_states do_class_d1e_error(struct rxe_qp *qp)
1383 {
1384 	/* UC */
1385 	if (qp->srq) {
1386 		/* Class E */
1387 		qp->resp.drop_msg = 1;
1388 		if (qp->resp.wqe) {
1389 			qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1390 			return RESPST_COMPLETE;
1391 		} else {
1392 			return RESPST_CLEANUP;
1393 		}
1394 	} else {
1395 		/* Class D1. This packet may be the start of a
1396 		 * new message and could be valid. The previous
1397 		 * message is invalid and ignored. reset the
1398 		 * recv wr to its original state
1399 		 */
1400 		if (qp->resp.wqe) {
1401 			qp->resp.wqe->dma.resid = qp->resp.wqe->dma.length;
1402 			qp->resp.wqe->dma.cur_sge = 0;
1403 			qp->resp.wqe->dma.sge_offset = 0;
1404 			qp->resp.opcode = -1;
1405 		}
1406 
1407 		if (qp->resp.mr) {
1408 			rxe_put(qp->resp.mr);
1409 			qp->resp.mr = NULL;
1410 		}
1411 
1412 		return RESPST_CLEANUP;
1413 	}
1414 }
1415 
1416 /* drain incoming request packet queue */
1417 static void drain_req_pkts(struct rxe_qp *qp)
1418 {
1419 	struct sk_buff *skb;
1420 
1421 	while ((skb = skb_dequeue(&qp->req_pkts))) {
1422 		rxe_put(qp);
1423 		kfree_skb(skb);
1424 		ib_device_put(qp->ibqp.device);
1425 	}
1426 }
1427 
1428 /* complete receive wqe with flush error */
1429 static int flush_recv_wqe(struct rxe_qp *qp, struct rxe_recv_wqe *wqe)
1430 {
1431 	struct rxe_cqe cqe = {};
1432 	struct ib_wc *wc = &cqe.ibwc;
1433 	struct ib_uverbs_wc *uwc = &cqe.uibwc;
1434 	int err;
1435 
1436 	if (qp->rcq->is_user) {
1437 		uwc->wr_id = wqe->wr_id;
1438 		uwc->status = IB_WC_WR_FLUSH_ERR;
1439 		uwc->qp_num = qp_num(qp);
1440 	} else {
1441 		wc->wr_id = wqe->wr_id;
1442 		wc->status = IB_WC_WR_FLUSH_ERR;
1443 		wc->qp = &qp->ibqp;
1444 	}
1445 
1446 	err = rxe_cq_post(qp->rcq, &cqe, 0);
1447 	if (err)
1448 		rxe_dbg_cq(qp->rcq, "post cq failed err = %d\n", err);
1449 
1450 	return err;
1451 }
1452 
1453 /* drain and optionally complete the recive queue
1454  * if unable to complete a wqe stop completing and
1455  * just flush the remaining wqes
1456  */
1457 static void flush_recv_queue(struct rxe_qp *qp, bool notify)
1458 {
1459 	struct rxe_queue *q = qp->rq.queue;
1460 	struct rxe_recv_wqe *wqe;
1461 	int err;
1462 
1463 	if (qp->srq) {
1464 		if (notify && qp->ibqp.event_handler) {
1465 			struct ib_event ev;
1466 
1467 			ev.device = qp->ibqp.device;
1468 			ev.element.qp = &qp->ibqp;
1469 			ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1470 			qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1471 		}
1472 		return;
1473 	}
1474 
1475 	/* recv queue not created. nothing to do. */
1476 	if (!qp->rq.queue)
1477 		return;
1478 
1479 	while ((wqe = queue_head(q, q->type))) {
1480 		if (notify) {
1481 			err = flush_recv_wqe(qp, wqe);
1482 			if (err)
1483 				notify = 0;
1484 		}
1485 		queue_advance_consumer(q, q->type);
1486 	}
1487 
1488 	qp->resp.wqe = NULL;
1489 }
1490 
1491 int rxe_receiver(struct rxe_qp *qp)
1492 {
1493 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
1494 	enum resp_states state;
1495 	struct rxe_pkt_info *pkt = NULL;
1496 	int ret;
1497 	unsigned long flags;
1498 
1499 	spin_lock_irqsave(&qp->state_lock, flags);
1500 	if (!qp->valid || qp_state(qp) == IB_QPS_ERR ||
1501 			  qp_state(qp) == IB_QPS_RESET) {
1502 		bool notify = qp->valid && (qp_state(qp) == IB_QPS_ERR);
1503 
1504 		drain_req_pkts(qp);
1505 		flush_recv_queue(qp, notify);
1506 		spin_unlock_irqrestore(&qp->state_lock, flags);
1507 		goto exit;
1508 	}
1509 	spin_unlock_irqrestore(&qp->state_lock, flags);
1510 
1511 	qp->resp.aeth_syndrome = AETH_ACK_UNLIMITED;
1512 
1513 	state = RESPST_GET_REQ;
1514 
1515 	while (1) {
1516 		rxe_dbg_qp(qp, "state = %s\n", resp_state_name[state]);
1517 		switch (state) {
1518 		case RESPST_GET_REQ:
1519 			state = get_req(qp, &pkt);
1520 			break;
1521 		case RESPST_CHK_PSN:
1522 			state = check_psn(qp, pkt);
1523 			break;
1524 		case RESPST_CHK_OP_SEQ:
1525 			state = check_op_seq(qp, pkt);
1526 			break;
1527 		case RESPST_CHK_OP_VALID:
1528 			state = check_op_valid(qp, pkt);
1529 			break;
1530 		case RESPST_CHK_RESOURCE:
1531 			state = check_resource(qp, pkt);
1532 			break;
1533 		case RESPST_CHK_LENGTH:
1534 			state = rxe_resp_check_length(qp, pkt);
1535 			break;
1536 		case RESPST_CHK_RKEY:
1537 			state = check_rkey(qp, pkt);
1538 			break;
1539 		case RESPST_EXECUTE:
1540 			state = execute(qp, pkt);
1541 			break;
1542 		case RESPST_COMPLETE:
1543 			state = do_complete(qp, pkt);
1544 			break;
1545 		case RESPST_READ_REPLY:
1546 			state = read_reply(qp, pkt);
1547 			break;
1548 		case RESPST_ATOMIC_REPLY:
1549 			state = atomic_reply(qp, pkt);
1550 			break;
1551 		case RESPST_ATOMIC_WRITE_REPLY:
1552 			state = atomic_write_reply(qp, pkt);
1553 			break;
1554 		case RESPST_PROCESS_FLUSH:
1555 			state = process_flush(qp, pkt);
1556 			break;
1557 		case RESPST_ACKNOWLEDGE:
1558 			state = acknowledge(qp, pkt);
1559 			break;
1560 		case RESPST_CLEANUP:
1561 			state = cleanup(qp, pkt);
1562 			break;
1563 		case RESPST_DUPLICATE_REQUEST:
1564 			state = duplicate_request(qp, pkt);
1565 			break;
1566 		case RESPST_ERR_PSN_OUT_OF_SEQ:
1567 			/* RC only - Class B. Drop packet. */
1568 			send_ack(qp, AETH_NAK_PSN_SEQ_ERROR, qp->resp.psn);
1569 			state = RESPST_CLEANUP;
1570 			break;
1571 
1572 		case RESPST_ERR_TOO_MANY_RDMA_ATM_REQ:
1573 		case RESPST_ERR_MISSING_OPCODE_FIRST:
1574 		case RESPST_ERR_MISSING_OPCODE_LAST_C:
1575 		case RESPST_ERR_UNSUPPORTED_OPCODE:
1576 		case RESPST_ERR_MISALIGNED_ATOMIC:
1577 			/* RC Only - Class C. */
1578 			do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1579 					  IB_WC_REM_INV_REQ_ERR);
1580 			state = RESPST_COMPLETE;
1581 			break;
1582 
1583 		case RESPST_ERR_MISSING_OPCODE_LAST_D1E:
1584 			state = do_class_d1e_error(qp);
1585 			break;
1586 		case RESPST_ERR_RNR:
1587 			if (qp_type(qp) == IB_QPT_RC) {
1588 				rxe_counter_inc(rxe, RXE_CNT_SND_RNR);
1589 				/* RC - class B */
1590 				send_ack(qp, AETH_RNR_NAK |
1591 					 (~AETH_TYPE_MASK &
1592 					 qp->attr.min_rnr_timer),
1593 					 pkt->psn);
1594 			} else {
1595 				/* UD/UC - class D */
1596 				qp->resp.drop_msg = 1;
1597 			}
1598 			state = RESPST_CLEANUP;
1599 			break;
1600 
1601 		case RESPST_ERR_RKEY_VIOLATION:
1602 			if (qp_type(qp) == IB_QPT_RC) {
1603 				/* Class C */
1604 				do_class_ac_error(qp, AETH_NAK_REM_ACC_ERR,
1605 						  IB_WC_REM_ACCESS_ERR);
1606 				state = RESPST_COMPLETE;
1607 			} else {
1608 				qp->resp.drop_msg = 1;
1609 				if (qp->srq) {
1610 					/* UC/SRQ Class D */
1611 					qp->resp.status = IB_WC_REM_ACCESS_ERR;
1612 					state = RESPST_COMPLETE;
1613 				} else {
1614 					/* UC/non-SRQ Class E. */
1615 					state = RESPST_CLEANUP;
1616 				}
1617 			}
1618 			break;
1619 
1620 		case RESPST_ERR_INVALIDATE_RKEY:
1621 			/* RC - Class J. */
1622 			qp->resp.goto_error = 1;
1623 			qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1624 			state = RESPST_COMPLETE;
1625 			break;
1626 
1627 		case RESPST_ERR_LENGTH:
1628 			if (qp_type(qp) == IB_QPT_RC) {
1629 				/* Class C */
1630 				do_class_ac_error(qp, AETH_NAK_INVALID_REQ,
1631 						  IB_WC_REM_INV_REQ_ERR);
1632 				state = RESPST_COMPLETE;
1633 			} else if (qp->srq) {
1634 				/* UC/UD - class E */
1635 				qp->resp.status = IB_WC_REM_INV_REQ_ERR;
1636 				state = RESPST_COMPLETE;
1637 			} else {
1638 				/* UC/UD - class D */
1639 				qp->resp.drop_msg = 1;
1640 				state = RESPST_CLEANUP;
1641 			}
1642 			break;
1643 
1644 		case RESPST_ERR_MALFORMED_WQE:
1645 			/* All, Class A. */
1646 			do_class_ac_error(qp, AETH_NAK_REM_OP_ERR,
1647 					  IB_WC_LOC_QP_OP_ERR);
1648 			state = RESPST_COMPLETE;
1649 			break;
1650 
1651 		case RESPST_ERR_CQ_OVERFLOW:
1652 			/* All - Class G */
1653 			state = RESPST_ERROR;
1654 			break;
1655 
1656 		case RESPST_DONE:
1657 			if (qp->resp.goto_error) {
1658 				state = RESPST_ERROR;
1659 				break;
1660 			}
1661 
1662 			goto done;
1663 
1664 		case RESPST_EXIT:
1665 			if (qp->resp.goto_error) {
1666 				state = RESPST_ERROR;
1667 				break;
1668 			}
1669 
1670 			goto exit;
1671 
1672 		case RESPST_ERROR:
1673 			qp->resp.goto_error = 0;
1674 			rxe_dbg_qp(qp, "moved to error state\n");
1675 			rxe_qp_error(qp);
1676 			goto exit;
1677 
1678 		default:
1679 			WARN_ON_ONCE(1);
1680 		}
1681 	}
1682 
1683 	/* A non-zero return value will cause rxe_do_task to
1684 	 * exit its loop and end the work item. A zero return
1685 	 * will continue looping and return to rxe_responder
1686 	 */
1687 done:
1688 	ret = 0;
1689 	goto out;
1690 exit:
1691 	ret = -EAGAIN;
1692 out:
1693 	return ret;
1694 }
1695