xref: /openbmc/linux/drivers/infiniband/sw/siw/siw_qp.c (revision 63705da3)
1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
2 
3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */
4 /* Copyright (c) 2008-2019, IBM Corporation */
5 
6 #include <linux/errno.h>
7 #include <linux/types.h>
8 #include <linux/net.h>
9 #include <linux/scatterlist.h>
10 #include <linux/llist.h>
11 #include <asm/barrier.h>
12 #include <net/tcp.h>
13 
14 #include "siw.h"
15 #include "siw_verbs.h"
16 #include "siw_mem.h"
17 
18 static char siw_qp_state_to_string[SIW_QP_STATE_COUNT][sizeof "TERMINATE"] = {
19 	[SIW_QP_STATE_IDLE] = "IDLE",
20 	[SIW_QP_STATE_RTR] = "RTR",
21 	[SIW_QP_STATE_RTS] = "RTS",
22 	[SIW_QP_STATE_CLOSING] = "CLOSING",
23 	[SIW_QP_STATE_TERMINATE] = "TERMINATE",
24 	[SIW_QP_STATE_ERROR] = "ERROR"
25 };
26 
27 /*
28  * iWARP (RDMAP, DDP and MPA) parameters as well as Softiwarp settings on a
29  * per-RDMAP message basis. Please keep order of initializer. All MPA len
30  * is initialized to minimum packet size.
31  */
32 struct iwarp_msg_info iwarp_pktinfo[RDMAP_TERMINATE + 1] = {
33 	{ /* RDMAP_RDMA_WRITE */
34 	  .hdr_len = sizeof(struct iwarp_rdma_write),
35 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_rdma_write) - 2),
36 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_TAGGED | DDP_FLAG_LAST |
37 				 cpu_to_be16(DDP_VERSION << 8) |
38 				 cpu_to_be16(RDMAP_VERSION << 6) |
39 				 cpu_to_be16(RDMAP_RDMA_WRITE),
40 	  .rx_data = siw_proc_write },
41 	{ /* RDMAP_RDMA_READ_REQ */
42 	  .hdr_len = sizeof(struct iwarp_rdma_rreq),
43 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_rdma_rreq) - 2),
44 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_LAST | cpu_to_be16(DDP_VERSION << 8) |
45 				 cpu_to_be16(RDMAP_VERSION << 6) |
46 				 cpu_to_be16(RDMAP_RDMA_READ_REQ),
47 	  .rx_data = siw_proc_rreq },
48 	{ /* RDMAP_RDMA_READ_RESP */
49 	  .hdr_len = sizeof(struct iwarp_rdma_rresp),
50 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_rdma_rresp) - 2),
51 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_TAGGED | DDP_FLAG_LAST |
52 				 cpu_to_be16(DDP_VERSION << 8) |
53 				 cpu_to_be16(RDMAP_VERSION << 6) |
54 				 cpu_to_be16(RDMAP_RDMA_READ_RESP),
55 	  .rx_data = siw_proc_rresp },
56 	{ /* RDMAP_SEND */
57 	  .hdr_len = sizeof(struct iwarp_send),
58 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_send) - 2),
59 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_LAST | cpu_to_be16(DDP_VERSION << 8) |
60 				 cpu_to_be16(RDMAP_VERSION << 6) |
61 				 cpu_to_be16(RDMAP_SEND),
62 	  .rx_data = siw_proc_send },
63 	{ /* RDMAP_SEND_INVAL */
64 	  .hdr_len = sizeof(struct iwarp_send_inv),
65 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_send_inv) - 2),
66 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_LAST | cpu_to_be16(DDP_VERSION << 8) |
67 				 cpu_to_be16(RDMAP_VERSION << 6) |
68 				 cpu_to_be16(RDMAP_SEND_INVAL),
69 	  .rx_data = siw_proc_send },
70 	{ /* RDMAP_SEND_SE */
71 	  .hdr_len = sizeof(struct iwarp_send),
72 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_send) - 2),
73 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_LAST | cpu_to_be16(DDP_VERSION << 8) |
74 				 cpu_to_be16(RDMAP_VERSION << 6) |
75 				 cpu_to_be16(RDMAP_SEND_SE),
76 	  .rx_data = siw_proc_send },
77 	{ /* RDMAP_SEND_SE_INVAL */
78 	  .hdr_len = sizeof(struct iwarp_send_inv),
79 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_send_inv) - 2),
80 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_LAST | cpu_to_be16(DDP_VERSION << 8) |
81 				 cpu_to_be16(RDMAP_VERSION << 6) |
82 				 cpu_to_be16(RDMAP_SEND_SE_INVAL),
83 	  .rx_data = siw_proc_send },
84 	{ /* RDMAP_TERMINATE */
85 	  .hdr_len = sizeof(struct iwarp_terminate),
86 	  .ctrl.mpa_len = htons(sizeof(struct iwarp_terminate) - 2),
87 	  .ctrl.ddp_rdmap_ctrl = DDP_FLAG_LAST | cpu_to_be16(DDP_VERSION << 8) |
88 				 cpu_to_be16(RDMAP_VERSION << 6) |
89 				 cpu_to_be16(RDMAP_TERMINATE),
90 	  .rx_data = siw_proc_terminate }
91 };
92 
93 void siw_qp_llp_data_ready(struct sock *sk)
94 {
95 	struct siw_qp *qp;
96 
97 	read_lock(&sk->sk_callback_lock);
98 
99 	if (unlikely(!sk->sk_user_data || !sk_to_qp(sk)))
100 		goto done;
101 
102 	qp = sk_to_qp(sk);
103 
104 	if (likely(!qp->rx_stream.rx_suspend &&
105 		   down_read_trylock(&qp->state_lock))) {
106 		read_descriptor_t rd_desc = { .arg.data = qp, .count = 1 };
107 
108 		if (likely(qp->attrs.state == SIW_QP_STATE_RTS))
109 			/*
110 			 * Implements data receive operation during
111 			 * socket callback. TCP gracefully catches
112 			 * the case where there is nothing to receive
113 			 * (not calling siw_tcp_rx_data() then).
114 			 */
115 			tcp_read_sock(sk, &rd_desc, siw_tcp_rx_data);
116 
117 		up_read(&qp->state_lock);
118 	} else {
119 		siw_dbg_qp(qp, "unable to process RX, suspend: %d\n",
120 			   qp->rx_stream.rx_suspend);
121 	}
122 done:
123 	read_unlock(&sk->sk_callback_lock);
124 }
125 
126 void siw_qp_llp_close(struct siw_qp *qp)
127 {
128 	siw_dbg_qp(qp, "enter llp close, state = %s\n",
129 		   siw_qp_state_to_string[qp->attrs.state]);
130 
131 	down_write(&qp->state_lock);
132 
133 	qp->rx_stream.rx_suspend = 1;
134 	qp->tx_ctx.tx_suspend = 1;
135 	qp->attrs.sk = NULL;
136 
137 	switch (qp->attrs.state) {
138 	case SIW_QP_STATE_RTS:
139 	case SIW_QP_STATE_RTR:
140 	case SIW_QP_STATE_IDLE:
141 	case SIW_QP_STATE_TERMINATE:
142 		qp->attrs.state = SIW_QP_STATE_ERROR;
143 		break;
144 	/*
145 	 * SIW_QP_STATE_CLOSING:
146 	 *
147 	 * This is a forced close. shall the QP be moved to
148 	 * ERROR or IDLE ?
149 	 */
150 	case SIW_QP_STATE_CLOSING:
151 		if (tx_wqe(qp)->wr_status == SIW_WR_IDLE)
152 			qp->attrs.state = SIW_QP_STATE_ERROR;
153 		else
154 			qp->attrs.state = SIW_QP_STATE_IDLE;
155 		break;
156 
157 	default:
158 		siw_dbg_qp(qp, "llp close: no state transition needed: %s\n",
159 			   siw_qp_state_to_string[qp->attrs.state]);
160 		break;
161 	}
162 	siw_sq_flush(qp);
163 	siw_rq_flush(qp);
164 
165 	/*
166 	 * Dereference closing CEP
167 	 */
168 	if (qp->cep) {
169 		siw_cep_put(qp->cep);
170 		qp->cep = NULL;
171 	}
172 
173 	up_write(&qp->state_lock);
174 
175 	siw_dbg_qp(qp, "llp close exit: state %s\n",
176 		   siw_qp_state_to_string[qp->attrs.state]);
177 }
178 
179 /*
180  * socket callback routine informing about newly available send space.
181  * Function schedules SQ work for processing SQ items.
182  */
183 void siw_qp_llp_write_space(struct sock *sk)
184 {
185 	struct siw_cep *cep;
186 
187 	read_lock(&sk->sk_callback_lock);
188 
189 	cep  = sk_to_cep(sk);
190 	if (cep) {
191 		cep->sk_write_space(sk);
192 
193 		if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
194 			(void)siw_sq_start(cep->qp);
195 	}
196 
197 	read_unlock(&sk->sk_callback_lock);
198 }
199 
200 static int siw_qp_readq_init(struct siw_qp *qp, int irq_size, int orq_size)
201 {
202 	if (irq_size) {
203 		irq_size = roundup_pow_of_two(irq_size);
204 		qp->irq = vzalloc(irq_size * sizeof(struct siw_sqe));
205 		if (!qp->irq) {
206 			qp->attrs.irq_size = 0;
207 			return -ENOMEM;
208 		}
209 	}
210 	if (orq_size) {
211 		orq_size = roundup_pow_of_two(orq_size);
212 		qp->orq = vzalloc(orq_size * sizeof(struct siw_sqe));
213 		if (!qp->orq) {
214 			qp->attrs.orq_size = 0;
215 			qp->attrs.irq_size = 0;
216 			vfree(qp->irq);
217 			return -ENOMEM;
218 		}
219 	}
220 	qp->attrs.irq_size = irq_size;
221 	qp->attrs.orq_size = orq_size;
222 	siw_dbg_qp(qp, "ORD %d, IRD %d\n", orq_size, irq_size);
223 	return 0;
224 }
225 
226 static int siw_qp_enable_crc(struct siw_qp *qp)
227 {
228 	struct siw_rx_stream *c_rx = &qp->rx_stream;
229 	struct siw_iwarp_tx *c_tx = &qp->tx_ctx;
230 	int size;
231 
232 	if (siw_crypto_shash == NULL)
233 		return -ENOENT;
234 
235 	size = crypto_shash_descsize(siw_crypto_shash) +
236 		sizeof(struct shash_desc);
237 
238 	c_tx->mpa_crc_hd = kzalloc(size, GFP_KERNEL);
239 	c_rx->mpa_crc_hd = kzalloc(size, GFP_KERNEL);
240 	if (!c_tx->mpa_crc_hd || !c_rx->mpa_crc_hd) {
241 		kfree(c_tx->mpa_crc_hd);
242 		kfree(c_rx->mpa_crc_hd);
243 		c_tx->mpa_crc_hd = NULL;
244 		c_rx->mpa_crc_hd = NULL;
245 		return -ENOMEM;
246 	}
247 	c_tx->mpa_crc_hd->tfm = siw_crypto_shash;
248 	c_rx->mpa_crc_hd->tfm = siw_crypto_shash;
249 
250 	return 0;
251 }
252 
253 /*
254  * Send a non signalled READ or WRITE to peer side as negotiated
255  * with MPAv2 P2P setup protocol. The work request is only created
256  * as a current active WR and does not consume Send Queue space.
257  *
258  * Caller must hold QP state lock.
259  */
260 int siw_qp_mpa_rts(struct siw_qp *qp, enum mpa_v2_ctrl ctrl)
261 {
262 	struct siw_wqe *wqe = tx_wqe(qp);
263 	unsigned long flags;
264 	int rv = 0;
265 
266 	spin_lock_irqsave(&qp->sq_lock, flags);
267 
268 	if (unlikely(wqe->wr_status != SIW_WR_IDLE)) {
269 		spin_unlock_irqrestore(&qp->sq_lock, flags);
270 		return -EIO;
271 	}
272 	memset(wqe->mem, 0, sizeof(*wqe->mem) * SIW_MAX_SGE);
273 
274 	wqe->wr_status = SIW_WR_QUEUED;
275 	wqe->sqe.flags = 0;
276 	wqe->sqe.num_sge = 1;
277 	wqe->sqe.sge[0].length = 0;
278 	wqe->sqe.sge[0].laddr = 0;
279 	wqe->sqe.sge[0].lkey = 0;
280 	/*
281 	 * While it must not be checked for inbound zero length
282 	 * READ/WRITE, some HW may treat STag 0 special.
283 	 */
284 	wqe->sqe.rkey = 1;
285 	wqe->sqe.raddr = 0;
286 	wqe->processed = 0;
287 
288 	if (ctrl & MPA_V2_RDMA_WRITE_RTR)
289 		wqe->sqe.opcode = SIW_OP_WRITE;
290 	else if (ctrl & MPA_V2_RDMA_READ_RTR) {
291 		struct siw_sqe *rreq = NULL;
292 
293 		wqe->sqe.opcode = SIW_OP_READ;
294 
295 		spin_lock(&qp->orq_lock);
296 
297 		if (qp->attrs.orq_size)
298 			rreq = orq_get_free(qp);
299 		if (rreq) {
300 			siw_read_to_orq(rreq, &wqe->sqe);
301 			qp->orq_put++;
302 		} else
303 			rv = -EIO;
304 
305 		spin_unlock(&qp->orq_lock);
306 	} else
307 		rv = -EINVAL;
308 
309 	if (rv)
310 		wqe->wr_status = SIW_WR_IDLE;
311 
312 	spin_unlock_irqrestore(&qp->sq_lock, flags);
313 
314 	if (!rv)
315 		rv = siw_sq_start(qp);
316 
317 	return rv;
318 }
319 
320 /*
321  * Map memory access error to DDP tagged error
322  */
323 enum ddp_ecode siw_tagged_error(enum siw_access_state state)
324 {
325 	switch (state) {
326 	case E_STAG_INVALID:
327 		return DDP_ECODE_T_INVALID_STAG;
328 	case E_BASE_BOUNDS:
329 		return DDP_ECODE_T_BASE_BOUNDS;
330 	case E_PD_MISMATCH:
331 		return DDP_ECODE_T_STAG_NOT_ASSOC;
332 	case E_ACCESS_PERM:
333 		/*
334 		 * RFC 5041 (DDP) lacks an ecode for insufficient access
335 		 * permissions. 'Invalid STag' seem to be the closest
336 		 * match though.
337 		 */
338 		return DDP_ECODE_T_INVALID_STAG;
339 	default:
340 		WARN_ON(1);
341 		return DDP_ECODE_T_INVALID_STAG;
342 	}
343 }
344 
345 /*
346  * Map memory access error to RDMAP protection error
347  */
348 enum rdmap_ecode siw_rdmap_error(enum siw_access_state state)
349 {
350 	switch (state) {
351 	case E_STAG_INVALID:
352 		return RDMAP_ECODE_INVALID_STAG;
353 	case E_BASE_BOUNDS:
354 		return RDMAP_ECODE_BASE_BOUNDS;
355 	case E_PD_MISMATCH:
356 		return RDMAP_ECODE_STAG_NOT_ASSOC;
357 	case E_ACCESS_PERM:
358 		return RDMAP_ECODE_ACCESS_RIGHTS;
359 	default:
360 		return RDMAP_ECODE_UNSPECIFIED;
361 	}
362 }
363 
364 void siw_init_terminate(struct siw_qp *qp, enum term_elayer layer, u8 etype,
365 			u8 ecode, int in_tx)
366 {
367 	if (!qp->term_info.valid) {
368 		memset(&qp->term_info, 0, sizeof(qp->term_info));
369 		qp->term_info.layer = layer;
370 		qp->term_info.etype = etype;
371 		qp->term_info.ecode = ecode;
372 		qp->term_info.in_tx = in_tx;
373 		qp->term_info.valid = 1;
374 	}
375 	siw_dbg_qp(qp, "init TERM: layer %d, type %d, code %d, in tx %s\n",
376 		   layer, etype, ecode, in_tx ? "yes" : "no");
377 }
378 
379 /*
380  * Send a TERMINATE message, as defined in RFC's 5040/5041/5044/6581.
381  * Sending TERMINATE messages is best effort - such messages
382  * can only be send if the QP is still connected and it does
383  * not have another outbound message in-progress, i.e. the
384  * TERMINATE message must not interfer with an incomplete current
385  * transmit operation.
386  */
387 void siw_send_terminate(struct siw_qp *qp)
388 {
389 	struct kvec iov[3];
390 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
391 	struct iwarp_terminate *term = NULL;
392 	union iwarp_hdr *err_hdr = NULL;
393 	struct socket *s = qp->attrs.sk;
394 	struct siw_rx_stream *srx = &qp->rx_stream;
395 	union iwarp_hdr *rx_hdr = &srx->hdr;
396 	u32 crc = 0;
397 	int num_frags, len_terminate, rv;
398 
399 	if (!qp->term_info.valid)
400 		return;
401 
402 	qp->term_info.valid = 0;
403 
404 	if (tx_wqe(qp)->wr_status == SIW_WR_INPROGRESS) {
405 		siw_dbg_qp(qp, "cannot send TERMINATE: op %d in progress\n",
406 			   tx_type(tx_wqe(qp)));
407 		return;
408 	}
409 	if (!s && qp->cep)
410 		/* QP not yet in RTS. Take socket from connection end point */
411 		s = qp->cep->sock;
412 
413 	if (!s) {
414 		siw_dbg_qp(qp, "cannot send TERMINATE: not connected\n");
415 		return;
416 	}
417 
418 	term = kzalloc(sizeof(*term), GFP_KERNEL);
419 	if (!term)
420 		return;
421 
422 	term->ddp_qn = cpu_to_be32(RDMAP_UNTAGGED_QN_TERMINATE);
423 	term->ddp_mo = 0;
424 	term->ddp_msn = cpu_to_be32(1);
425 
426 	iov[0].iov_base = term;
427 	iov[0].iov_len = sizeof(*term);
428 
429 	if ((qp->term_info.layer == TERM_ERROR_LAYER_DDP) ||
430 	    ((qp->term_info.layer == TERM_ERROR_LAYER_RDMAP) &&
431 	     (qp->term_info.etype != RDMAP_ETYPE_CATASTROPHIC))) {
432 		err_hdr = kzalloc(sizeof(*err_hdr), GFP_KERNEL);
433 		if (!err_hdr) {
434 			kfree(term);
435 			return;
436 		}
437 	}
438 	memcpy(&term->ctrl, &iwarp_pktinfo[RDMAP_TERMINATE].ctrl,
439 	       sizeof(struct iwarp_ctrl));
440 
441 	__rdmap_term_set_layer(term, qp->term_info.layer);
442 	__rdmap_term_set_etype(term, qp->term_info.etype);
443 	__rdmap_term_set_ecode(term, qp->term_info.ecode);
444 
445 	switch (qp->term_info.layer) {
446 	case TERM_ERROR_LAYER_RDMAP:
447 		if (qp->term_info.etype == RDMAP_ETYPE_CATASTROPHIC)
448 			/* No additional DDP/RDMAP header to be included */
449 			break;
450 
451 		if (qp->term_info.etype == RDMAP_ETYPE_REMOTE_PROTECTION) {
452 			/*
453 			 * Complete RDMAP frame will get attached, and
454 			 * DDP segment length is valid
455 			 */
456 			term->flag_m = 1;
457 			term->flag_d = 1;
458 			term->flag_r = 1;
459 
460 			if (qp->term_info.in_tx) {
461 				struct iwarp_rdma_rreq *rreq;
462 				struct siw_wqe *wqe = tx_wqe(qp);
463 
464 				/* Inbound RREQ error, detected during
465 				 * RRESP creation. Take state from
466 				 * current TX work queue element to
467 				 * reconstruct peers RREQ.
468 				 */
469 				rreq = (struct iwarp_rdma_rreq *)err_hdr;
470 
471 				memcpy(&rreq->ctrl,
472 				       &iwarp_pktinfo[RDMAP_RDMA_READ_REQ].ctrl,
473 				       sizeof(struct iwarp_ctrl));
474 
475 				rreq->rsvd = 0;
476 				rreq->ddp_qn =
477 					htonl(RDMAP_UNTAGGED_QN_RDMA_READ);
478 
479 				/* Provide RREQ's MSN as kept aside */
480 				rreq->ddp_msn = htonl(wqe->sqe.sge[0].length);
481 
482 				rreq->ddp_mo = htonl(wqe->processed);
483 				rreq->sink_stag = htonl(wqe->sqe.rkey);
484 				rreq->sink_to = cpu_to_be64(wqe->sqe.raddr);
485 				rreq->read_size = htonl(wqe->sqe.sge[0].length);
486 				rreq->source_stag = htonl(wqe->sqe.sge[0].lkey);
487 				rreq->source_to =
488 					cpu_to_be64(wqe->sqe.sge[0].laddr);
489 
490 				iov[1].iov_base = rreq;
491 				iov[1].iov_len = sizeof(*rreq);
492 
493 				rx_hdr = (union iwarp_hdr *)rreq;
494 			} else {
495 				/* Take RDMAP/DDP information from
496 				 * current (failed) inbound frame.
497 				 */
498 				iov[1].iov_base = rx_hdr;
499 
500 				if (__rdmap_get_opcode(&rx_hdr->ctrl) ==
501 				    RDMAP_RDMA_READ_REQ)
502 					iov[1].iov_len =
503 						sizeof(struct iwarp_rdma_rreq);
504 				else /* SEND type */
505 					iov[1].iov_len =
506 						sizeof(struct iwarp_send);
507 			}
508 		} else {
509 			/* Do not report DDP hdr information if packet
510 			 * layout is unknown
511 			 */
512 			if ((qp->term_info.ecode == RDMAP_ECODE_VERSION) ||
513 			    (qp->term_info.ecode == RDMAP_ECODE_OPCODE))
514 				break;
515 
516 			iov[1].iov_base = rx_hdr;
517 
518 			/* Only DDP frame will get attached */
519 			if (rx_hdr->ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED)
520 				iov[1].iov_len =
521 					sizeof(struct iwarp_rdma_write);
522 			else
523 				iov[1].iov_len = sizeof(struct iwarp_send);
524 
525 			term->flag_m = 1;
526 			term->flag_d = 1;
527 		}
528 		term->ctrl.mpa_len = cpu_to_be16(iov[1].iov_len);
529 		break;
530 
531 	case TERM_ERROR_LAYER_DDP:
532 		/* Report error encountered while DDP processing.
533 		 * This can only happen as a result of inbound
534 		 * DDP processing
535 		 */
536 
537 		/* Do not report DDP hdr information if packet
538 		 * layout is unknown
539 		 */
540 		if (((qp->term_info.etype == DDP_ETYPE_TAGGED_BUF) &&
541 		     (qp->term_info.ecode == DDP_ECODE_T_VERSION)) ||
542 		    ((qp->term_info.etype == DDP_ETYPE_UNTAGGED_BUF) &&
543 		     (qp->term_info.ecode == DDP_ECODE_UT_VERSION)))
544 			break;
545 
546 		iov[1].iov_base = rx_hdr;
547 
548 		if (rx_hdr->ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED)
549 			iov[1].iov_len = sizeof(struct iwarp_ctrl_tagged);
550 		else
551 			iov[1].iov_len = sizeof(struct iwarp_ctrl_untagged);
552 
553 		term->flag_m = 1;
554 		term->flag_d = 1;
555 		break;
556 
557 	default:
558 		break;
559 	}
560 	if (term->flag_m || term->flag_d || term->flag_r) {
561 		iov[2].iov_base = &crc;
562 		iov[2].iov_len = sizeof(crc);
563 		len_terminate = sizeof(*term) + iov[1].iov_len + MPA_CRC_SIZE;
564 		num_frags = 3;
565 	} else {
566 		iov[1].iov_base = &crc;
567 		iov[1].iov_len = sizeof(crc);
568 		len_terminate = sizeof(*term) + MPA_CRC_SIZE;
569 		num_frags = 2;
570 	}
571 
572 	/* Adjust DDP Segment Length parameter, if valid */
573 	if (term->flag_m) {
574 		u32 real_ddp_len = be16_to_cpu(rx_hdr->ctrl.mpa_len);
575 		enum rdma_opcode op = __rdmap_get_opcode(&rx_hdr->ctrl);
576 
577 		real_ddp_len -= iwarp_pktinfo[op].hdr_len - MPA_HDR_SIZE;
578 		rx_hdr->ctrl.mpa_len = cpu_to_be16(real_ddp_len);
579 	}
580 
581 	term->ctrl.mpa_len =
582 		cpu_to_be16(len_terminate - (MPA_HDR_SIZE + MPA_CRC_SIZE));
583 	if (qp->tx_ctx.mpa_crc_hd) {
584 		crypto_shash_init(qp->tx_ctx.mpa_crc_hd);
585 		if (crypto_shash_update(qp->tx_ctx.mpa_crc_hd,
586 					(u8 *)iov[0].iov_base,
587 					iov[0].iov_len))
588 			goto out;
589 
590 		if (num_frags == 3) {
591 			if (crypto_shash_update(qp->tx_ctx.mpa_crc_hd,
592 						(u8 *)iov[1].iov_base,
593 						iov[1].iov_len))
594 				goto out;
595 		}
596 		crypto_shash_final(qp->tx_ctx.mpa_crc_hd, (u8 *)&crc);
597 	}
598 
599 	rv = kernel_sendmsg(s, &msg, iov, num_frags, len_terminate);
600 	siw_dbg_qp(qp, "sent TERM: %s, layer %d, type %d, code %d (%d bytes)\n",
601 		   rv == len_terminate ? "success" : "failure",
602 		   __rdmap_term_layer(term), __rdmap_term_etype(term),
603 		   __rdmap_term_ecode(term), rv);
604 out:
605 	kfree(term);
606 	kfree(err_hdr);
607 }
608 
609 /*
610  * Handle all attrs other than state
611  */
612 static void siw_qp_modify_nonstate(struct siw_qp *qp,
613 				   struct siw_qp_attrs *attrs,
614 				   enum siw_qp_attr_mask mask)
615 {
616 	if (mask & SIW_QP_ATTR_ACCESS_FLAGS) {
617 		if (attrs->flags & SIW_RDMA_BIND_ENABLED)
618 			qp->attrs.flags |= SIW_RDMA_BIND_ENABLED;
619 		else
620 			qp->attrs.flags &= ~SIW_RDMA_BIND_ENABLED;
621 
622 		if (attrs->flags & SIW_RDMA_WRITE_ENABLED)
623 			qp->attrs.flags |= SIW_RDMA_WRITE_ENABLED;
624 		else
625 			qp->attrs.flags &= ~SIW_RDMA_WRITE_ENABLED;
626 
627 		if (attrs->flags & SIW_RDMA_READ_ENABLED)
628 			qp->attrs.flags |= SIW_RDMA_READ_ENABLED;
629 		else
630 			qp->attrs.flags &= ~SIW_RDMA_READ_ENABLED;
631 	}
632 }
633 
634 static int siw_qp_nextstate_from_idle(struct siw_qp *qp,
635 				      struct siw_qp_attrs *attrs,
636 				      enum siw_qp_attr_mask mask)
637 {
638 	int rv = 0;
639 
640 	switch (attrs->state) {
641 	case SIW_QP_STATE_RTS:
642 		if (attrs->flags & SIW_MPA_CRC) {
643 			rv = siw_qp_enable_crc(qp);
644 			if (rv)
645 				break;
646 		}
647 		if (!(mask & SIW_QP_ATTR_LLP_HANDLE)) {
648 			siw_dbg_qp(qp, "no socket\n");
649 			rv = -EINVAL;
650 			break;
651 		}
652 		if (!(mask & SIW_QP_ATTR_MPA)) {
653 			siw_dbg_qp(qp, "no MPA\n");
654 			rv = -EINVAL;
655 			break;
656 		}
657 		/*
658 		 * Initialize iWARP TX state
659 		 */
660 		qp->tx_ctx.ddp_msn[RDMAP_UNTAGGED_QN_SEND] = 0;
661 		qp->tx_ctx.ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ] = 0;
662 		qp->tx_ctx.ddp_msn[RDMAP_UNTAGGED_QN_TERMINATE] = 0;
663 
664 		/*
665 		 * Initialize iWARP RX state
666 		 */
667 		qp->rx_stream.ddp_msn[RDMAP_UNTAGGED_QN_SEND] = 1;
668 		qp->rx_stream.ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ] = 1;
669 		qp->rx_stream.ddp_msn[RDMAP_UNTAGGED_QN_TERMINATE] = 1;
670 
671 		/*
672 		 * init IRD free queue, caller has already checked
673 		 * limits.
674 		 */
675 		rv = siw_qp_readq_init(qp, attrs->irq_size,
676 				       attrs->orq_size);
677 		if (rv)
678 			break;
679 
680 		qp->attrs.sk = attrs->sk;
681 		qp->attrs.state = SIW_QP_STATE_RTS;
682 
683 		siw_dbg_qp(qp, "enter RTS: crc=%s, ord=%u, ird=%u\n",
684 			   attrs->flags & SIW_MPA_CRC ? "y" : "n",
685 			   qp->attrs.orq_size, qp->attrs.irq_size);
686 		break;
687 
688 	case SIW_QP_STATE_ERROR:
689 		siw_rq_flush(qp);
690 		qp->attrs.state = SIW_QP_STATE_ERROR;
691 		if (qp->cep) {
692 			siw_cep_put(qp->cep);
693 			qp->cep = NULL;
694 		}
695 		break;
696 
697 	default:
698 		break;
699 	}
700 	return rv;
701 }
702 
703 static int siw_qp_nextstate_from_rts(struct siw_qp *qp,
704 				     struct siw_qp_attrs *attrs)
705 {
706 	int drop_conn = 0;
707 
708 	switch (attrs->state) {
709 	case SIW_QP_STATE_CLOSING:
710 		/*
711 		 * Verbs: move to IDLE if SQ and ORQ are empty.
712 		 * Move to ERROR otherwise. But first of all we must
713 		 * close the connection. So we keep CLOSING or ERROR
714 		 * as a transient state, schedule connection drop work
715 		 * and wait for the socket state change upcall to
716 		 * come back closed.
717 		 */
718 		if (tx_wqe(qp)->wr_status == SIW_WR_IDLE) {
719 			qp->attrs.state = SIW_QP_STATE_CLOSING;
720 		} else {
721 			qp->attrs.state = SIW_QP_STATE_ERROR;
722 			siw_sq_flush(qp);
723 		}
724 		siw_rq_flush(qp);
725 
726 		drop_conn = 1;
727 		break;
728 
729 	case SIW_QP_STATE_TERMINATE:
730 		qp->attrs.state = SIW_QP_STATE_TERMINATE;
731 
732 		siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP,
733 				   RDMAP_ETYPE_CATASTROPHIC,
734 				   RDMAP_ECODE_UNSPECIFIED, 1);
735 		drop_conn = 1;
736 		break;
737 
738 	case SIW_QP_STATE_ERROR:
739 		/*
740 		 * This is an emergency close.
741 		 *
742 		 * Any in progress transmit operation will get
743 		 * cancelled.
744 		 * This will likely result in a protocol failure,
745 		 * if a TX operation is in transit. The caller
746 		 * could unconditional wait to give the current
747 		 * operation a chance to complete.
748 		 * Esp., how to handle the non-empty IRQ case?
749 		 * The peer was asking for data transfer at a valid
750 		 * point in time.
751 		 */
752 		siw_sq_flush(qp);
753 		siw_rq_flush(qp);
754 		qp->attrs.state = SIW_QP_STATE_ERROR;
755 		drop_conn = 1;
756 		break;
757 
758 	default:
759 		break;
760 	}
761 	return drop_conn;
762 }
763 
764 static void siw_qp_nextstate_from_term(struct siw_qp *qp,
765 				       struct siw_qp_attrs *attrs)
766 {
767 	switch (attrs->state) {
768 	case SIW_QP_STATE_ERROR:
769 		siw_rq_flush(qp);
770 		qp->attrs.state = SIW_QP_STATE_ERROR;
771 
772 		if (tx_wqe(qp)->wr_status != SIW_WR_IDLE)
773 			siw_sq_flush(qp);
774 		break;
775 
776 	default:
777 		break;
778 	}
779 }
780 
781 static int siw_qp_nextstate_from_close(struct siw_qp *qp,
782 				       struct siw_qp_attrs *attrs)
783 {
784 	int rv = 0;
785 
786 	switch (attrs->state) {
787 	case SIW_QP_STATE_IDLE:
788 		WARN_ON(tx_wqe(qp)->wr_status != SIW_WR_IDLE);
789 		qp->attrs.state = SIW_QP_STATE_IDLE;
790 		break;
791 
792 	case SIW_QP_STATE_CLOSING:
793 		/*
794 		 * The LLP may already moved the QP to closing
795 		 * due to graceful peer close init
796 		 */
797 		break;
798 
799 	case SIW_QP_STATE_ERROR:
800 		/*
801 		 * QP was moved to CLOSING by LLP event
802 		 * not yet seen by user.
803 		 */
804 		qp->attrs.state = SIW_QP_STATE_ERROR;
805 
806 		if (tx_wqe(qp)->wr_status != SIW_WR_IDLE)
807 			siw_sq_flush(qp);
808 
809 		siw_rq_flush(qp);
810 		break;
811 
812 	default:
813 		siw_dbg_qp(qp, "state transition undefined: %s => %s\n",
814 			   siw_qp_state_to_string[qp->attrs.state],
815 			   siw_qp_state_to_string[attrs->state]);
816 
817 		rv = -ECONNABORTED;
818 	}
819 	return rv;
820 }
821 
822 /*
823  * Caller must hold qp->state_lock
824  */
825 int siw_qp_modify(struct siw_qp *qp, struct siw_qp_attrs *attrs,
826 		  enum siw_qp_attr_mask mask)
827 {
828 	int drop_conn = 0, rv = 0;
829 
830 	if (!mask)
831 		return 0;
832 
833 	siw_dbg_qp(qp, "state: %s => %s\n",
834 		   siw_qp_state_to_string[qp->attrs.state],
835 		   siw_qp_state_to_string[attrs->state]);
836 
837 	if (mask != SIW_QP_ATTR_STATE)
838 		siw_qp_modify_nonstate(qp, attrs, mask);
839 
840 	if (!(mask & SIW_QP_ATTR_STATE))
841 		return 0;
842 
843 	switch (qp->attrs.state) {
844 	case SIW_QP_STATE_IDLE:
845 	case SIW_QP_STATE_RTR:
846 		rv = siw_qp_nextstate_from_idle(qp, attrs, mask);
847 		break;
848 
849 	case SIW_QP_STATE_RTS:
850 		drop_conn = siw_qp_nextstate_from_rts(qp, attrs);
851 		break;
852 
853 	case SIW_QP_STATE_TERMINATE:
854 		siw_qp_nextstate_from_term(qp, attrs);
855 		break;
856 
857 	case SIW_QP_STATE_CLOSING:
858 		siw_qp_nextstate_from_close(qp, attrs);
859 		break;
860 	default:
861 		break;
862 	}
863 	if (drop_conn)
864 		siw_qp_cm_drop(qp, 0);
865 
866 	return rv;
867 }
868 
869 void siw_read_to_orq(struct siw_sqe *rreq, struct siw_sqe *sqe)
870 {
871 	rreq->id = sqe->id;
872 	rreq->opcode = sqe->opcode;
873 	rreq->sge[0].laddr = sqe->sge[0].laddr;
874 	rreq->sge[0].length = sqe->sge[0].length;
875 	rreq->sge[0].lkey = sqe->sge[0].lkey;
876 	rreq->sge[1].lkey = sqe->sge[1].lkey;
877 	rreq->flags = sqe->flags | SIW_WQE_VALID;
878 	rreq->num_sge = 1;
879 }
880 
881 static int siw_activate_tx_from_sq(struct siw_qp *qp)
882 {
883 	struct siw_sqe *sqe;
884 	struct siw_wqe *wqe = tx_wqe(qp);
885 	int rv = 1;
886 
887 	sqe = sq_get_next(qp);
888 	if (!sqe)
889 		return 0;
890 
891 	memset(wqe->mem, 0, sizeof(*wqe->mem) * SIW_MAX_SGE);
892 	wqe->wr_status = SIW_WR_QUEUED;
893 
894 	/* First copy SQE to kernel private memory */
895 	memcpy(&wqe->sqe, sqe, sizeof(*sqe));
896 
897 	if (wqe->sqe.opcode >= SIW_NUM_OPCODES) {
898 		rv = -EINVAL;
899 		goto out;
900 	}
901 	if (wqe->sqe.flags & SIW_WQE_INLINE) {
902 		if (wqe->sqe.opcode != SIW_OP_SEND &&
903 		    wqe->sqe.opcode != SIW_OP_WRITE) {
904 			rv = -EINVAL;
905 			goto out;
906 		}
907 		if (wqe->sqe.sge[0].length > SIW_MAX_INLINE) {
908 			rv = -EINVAL;
909 			goto out;
910 		}
911 		wqe->sqe.sge[0].laddr = (uintptr_t)&wqe->sqe.sge[1];
912 		wqe->sqe.sge[0].lkey = 0;
913 		wqe->sqe.num_sge = 1;
914 	}
915 	if (wqe->sqe.flags & SIW_WQE_READ_FENCE) {
916 		/* A READ cannot be fenced */
917 		if (unlikely(wqe->sqe.opcode == SIW_OP_READ ||
918 			     wqe->sqe.opcode ==
919 				     SIW_OP_READ_LOCAL_INV)) {
920 			siw_dbg_qp(qp, "cannot fence read\n");
921 			rv = -EINVAL;
922 			goto out;
923 		}
924 		spin_lock(&qp->orq_lock);
925 
926 		if (qp->attrs.orq_size && !siw_orq_empty(qp)) {
927 			qp->tx_ctx.orq_fence = 1;
928 			rv = 0;
929 		}
930 		spin_unlock(&qp->orq_lock);
931 
932 	} else if (wqe->sqe.opcode == SIW_OP_READ ||
933 		   wqe->sqe.opcode == SIW_OP_READ_LOCAL_INV) {
934 		struct siw_sqe *rreq;
935 
936 		if (unlikely(!qp->attrs.orq_size)) {
937 			/* We negotiated not to send READ req's */
938 			rv = -EINVAL;
939 			goto out;
940 		}
941 		wqe->sqe.num_sge = 1;
942 
943 		spin_lock(&qp->orq_lock);
944 
945 		rreq = orq_get_free(qp);
946 		if (rreq) {
947 			/*
948 			 * Make an immediate copy in ORQ to be ready
949 			 * to process loopback READ reply
950 			 */
951 			siw_read_to_orq(rreq, &wqe->sqe);
952 			qp->orq_put++;
953 		} else {
954 			qp->tx_ctx.orq_fence = 1;
955 			rv = 0;
956 		}
957 		spin_unlock(&qp->orq_lock);
958 	}
959 
960 	/* Clear SQE, can be re-used by application */
961 	smp_store_mb(sqe->flags, 0);
962 	qp->sq_get++;
963 out:
964 	if (unlikely(rv < 0)) {
965 		siw_dbg_qp(qp, "error %d\n", rv);
966 		wqe->wr_status = SIW_WR_IDLE;
967 	}
968 	return rv;
969 }
970 
971 /*
972  * Must be called with SQ locked.
973  * To avoid complete SQ starvation by constant inbound READ requests,
974  * the active IRQ will not be served after qp->irq_burst, if the
975  * SQ has pending work.
976  */
977 int siw_activate_tx(struct siw_qp *qp)
978 {
979 	struct siw_sqe *irqe;
980 	struct siw_wqe *wqe = tx_wqe(qp);
981 
982 	if (!qp->attrs.irq_size)
983 		return siw_activate_tx_from_sq(qp);
984 
985 	irqe = &qp->irq[qp->irq_get % qp->attrs.irq_size];
986 
987 	if (!(irqe->flags & SIW_WQE_VALID))
988 		return siw_activate_tx_from_sq(qp);
989 
990 	/*
991 	 * Avoid local WQE processing starvation in case
992 	 * of constant inbound READ request stream
993 	 */
994 	if (sq_get_next(qp) && ++qp->irq_burst >= SIW_IRQ_MAXBURST_SQ_ACTIVE) {
995 		qp->irq_burst = 0;
996 		return siw_activate_tx_from_sq(qp);
997 	}
998 	memset(wqe->mem, 0, sizeof(*wqe->mem) * SIW_MAX_SGE);
999 	wqe->wr_status = SIW_WR_QUEUED;
1000 
1001 	/* start READ RESPONSE */
1002 	wqe->sqe.opcode = SIW_OP_READ_RESPONSE;
1003 	wqe->sqe.flags = 0;
1004 	if (irqe->num_sge) {
1005 		wqe->sqe.num_sge = 1;
1006 		wqe->sqe.sge[0].length = irqe->sge[0].length;
1007 		wqe->sqe.sge[0].laddr = irqe->sge[0].laddr;
1008 		wqe->sqe.sge[0].lkey = irqe->sge[0].lkey;
1009 	} else {
1010 		wqe->sqe.num_sge = 0;
1011 	}
1012 
1013 	/* Retain original RREQ's message sequence number for
1014 	 * potential error reporting cases.
1015 	 */
1016 	wqe->sqe.sge[1].length = irqe->sge[1].length;
1017 
1018 	wqe->sqe.rkey = irqe->rkey;
1019 	wqe->sqe.raddr = irqe->raddr;
1020 
1021 	wqe->processed = 0;
1022 	qp->irq_get++;
1023 
1024 	/* mark current IRQ entry free */
1025 	smp_store_mb(irqe->flags, 0);
1026 
1027 	return 1;
1028 }
1029 
1030 /*
1031  * Check if current CQ state qualifies for calling CQ completion
1032  * handler. Must be called with CQ lock held.
1033  */
1034 static bool siw_cq_notify_now(struct siw_cq *cq, u32 flags)
1035 {
1036 	u32 cq_notify;
1037 
1038 	if (!cq->base_cq.comp_handler)
1039 		return false;
1040 
1041 	/* Read application shared notification state */
1042 	cq_notify = READ_ONCE(cq->notify->flags);
1043 
1044 	if ((cq_notify & SIW_NOTIFY_NEXT_COMPLETION) ||
1045 	    ((cq_notify & SIW_NOTIFY_SOLICITED) &&
1046 	     (flags & SIW_WQE_SOLICITED))) {
1047 		/*
1048 		 * CQ notification is one-shot: Since the
1049 		 * current CQE causes user notification,
1050 		 * the CQ gets dis-aremd and must be re-aremd
1051 		 * by the user for a new notification.
1052 		 */
1053 		WRITE_ONCE(cq->notify->flags, SIW_NOTIFY_NOT);
1054 
1055 		return true;
1056 	}
1057 	return false;
1058 }
1059 
1060 int siw_sqe_complete(struct siw_qp *qp, struct siw_sqe *sqe, u32 bytes,
1061 		     enum siw_wc_status status)
1062 {
1063 	struct siw_cq *cq = qp->scq;
1064 	int rv = 0;
1065 
1066 	if (cq) {
1067 		u32 sqe_flags = sqe->flags;
1068 		struct siw_cqe *cqe;
1069 		u32 idx;
1070 		unsigned long flags;
1071 
1072 		spin_lock_irqsave(&cq->lock, flags);
1073 
1074 		idx = cq->cq_put % cq->num_cqe;
1075 		cqe = &cq->queue[idx];
1076 
1077 		if (!READ_ONCE(cqe->flags)) {
1078 			bool notify;
1079 
1080 			cqe->id = sqe->id;
1081 			cqe->opcode = sqe->opcode;
1082 			cqe->status = status;
1083 			cqe->imm_data = 0;
1084 			cqe->bytes = bytes;
1085 
1086 			if (rdma_is_kernel_res(&cq->base_cq.res))
1087 				cqe->base_qp = &qp->base_qp;
1088 			else
1089 				cqe->qp_id = qp_id(qp);
1090 
1091 			/* mark CQE valid for application */
1092 			WRITE_ONCE(cqe->flags, SIW_WQE_VALID);
1093 			/* recycle SQE */
1094 			smp_store_mb(sqe->flags, 0);
1095 
1096 			cq->cq_put++;
1097 			notify = siw_cq_notify_now(cq, sqe_flags);
1098 
1099 			spin_unlock_irqrestore(&cq->lock, flags);
1100 
1101 			if (notify) {
1102 				siw_dbg_cq(cq, "Call completion handler\n");
1103 				cq->base_cq.comp_handler(&cq->base_cq,
1104 						cq->base_cq.cq_context);
1105 			}
1106 		} else {
1107 			spin_unlock_irqrestore(&cq->lock, flags);
1108 			rv = -ENOMEM;
1109 			siw_cq_event(cq, IB_EVENT_CQ_ERR);
1110 		}
1111 	} else {
1112 		/* recycle SQE */
1113 		smp_store_mb(sqe->flags, 0);
1114 	}
1115 	return rv;
1116 }
1117 
1118 int siw_rqe_complete(struct siw_qp *qp, struct siw_rqe *rqe, u32 bytes,
1119 		     u32 inval_stag, enum siw_wc_status status)
1120 {
1121 	struct siw_cq *cq = qp->rcq;
1122 	int rv = 0;
1123 
1124 	if (cq) {
1125 		struct siw_cqe *cqe;
1126 		u32 idx;
1127 		unsigned long flags;
1128 
1129 		spin_lock_irqsave(&cq->lock, flags);
1130 
1131 		idx = cq->cq_put % cq->num_cqe;
1132 		cqe = &cq->queue[idx];
1133 
1134 		if (!READ_ONCE(cqe->flags)) {
1135 			bool notify;
1136 			u8 cqe_flags = SIW_WQE_VALID;
1137 
1138 			cqe->id = rqe->id;
1139 			cqe->opcode = SIW_OP_RECEIVE;
1140 			cqe->status = status;
1141 			cqe->imm_data = 0;
1142 			cqe->bytes = bytes;
1143 
1144 			if (rdma_is_kernel_res(&cq->base_cq.res)) {
1145 				cqe->base_qp = &qp->base_qp;
1146 				if (inval_stag) {
1147 					cqe_flags |= SIW_WQE_REM_INVAL;
1148 					cqe->inval_stag = inval_stag;
1149 				}
1150 			} else {
1151 				cqe->qp_id = qp_id(qp);
1152 			}
1153 			/* mark CQE valid for application */
1154 			WRITE_ONCE(cqe->flags, cqe_flags);
1155 			/* recycle RQE */
1156 			smp_store_mb(rqe->flags, 0);
1157 
1158 			cq->cq_put++;
1159 			notify = siw_cq_notify_now(cq, SIW_WQE_SIGNALLED);
1160 
1161 			spin_unlock_irqrestore(&cq->lock, flags);
1162 
1163 			if (notify) {
1164 				siw_dbg_cq(cq, "Call completion handler\n");
1165 				cq->base_cq.comp_handler(&cq->base_cq,
1166 						cq->base_cq.cq_context);
1167 			}
1168 		} else {
1169 			spin_unlock_irqrestore(&cq->lock, flags);
1170 			rv = -ENOMEM;
1171 			siw_cq_event(cq, IB_EVENT_CQ_ERR);
1172 		}
1173 	} else {
1174 		/* recycle RQE */
1175 		smp_store_mb(rqe->flags, 0);
1176 	}
1177 	return rv;
1178 }
1179 
1180 /*
1181  * siw_sq_flush()
1182  *
1183  * Flush SQ and ORRQ entries to CQ.
1184  *
1185  * Must be called with QP state write lock held.
1186  * Therefore, SQ and ORQ lock must not be taken.
1187  */
1188 void siw_sq_flush(struct siw_qp *qp)
1189 {
1190 	struct siw_sqe *sqe;
1191 	struct siw_wqe *wqe = tx_wqe(qp);
1192 	int async_event = 0;
1193 
1194 	/*
1195 	 * Start with completing any work currently on the ORQ
1196 	 */
1197 	while (qp->attrs.orq_size) {
1198 		sqe = &qp->orq[qp->orq_get % qp->attrs.orq_size];
1199 		if (!READ_ONCE(sqe->flags))
1200 			break;
1201 
1202 		if (siw_sqe_complete(qp, sqe, 0, SIW_WC_WR_FLUSH_ERR) != 0)
1203 			break;
1204 
1205 		WRITE_ONCE(sqe->flags, 0);
1206 		qp->orq_get++;
1207 	}
1208 	/*
1209 	 * Flush an in-progress WQE if present
1210 	 */
1211 	if (wqe->wr_status != SIW_WR_IDLE) {
1212 		siw_dbg_qp(qp, "flush current SQE, type %d, status %d\n",
1213 			   tx_type(wqe), wqe->wr_status);
1214 
1215 		siw_wqe_put_mem(wqe, tx_type(wqe));
1216 
1217 		if (tx_type(wqe) != SIW_OP_READ_RESPONSE &&
1218 		    ((tx_type(wqe) != SIW_OP_READ &&
1219 		      tx_type(wqe) != SIW_OP_READ_LOCAL_INV) ||
1220 		     wqe->wr_status == SIW_WR_QUEUED))
1221 			/*
1222 			 * An in-progress Read Request is already in
1223 			 * the ORQ
1224 			 */
1225 			siw_sqe_complete(qp, &wqe->sqe, wqe->bytes,
1226 					 SIW_WC_WR_FLUSH_ERR);
1227 
1228 		wqe->wr_status = SIW_WR_IDLE;
1229 	}
1230 	/*
1231 	 * Flush the Send Queue
1232 	 */
1233 	while (qp->attrs.sq_size) {
1234 		sqe = &qp->sendq[qp->sq_get % qp->attrs.sq_size];
1235 		if (!READ_ONCE(sqe->flags))
1236 			break;
1237 
1238 		async_event = 1;
1239 		if (siw_sqe_complete(qp, sqe, 0, SIW_WC_WR_FLUSH_ERR) != 0)
1240 			/*
1241 			 * Shall IB_EVENT_SQ_DRAINED be supressed if work
1242 			 * completion fails?
1243 			 */
1244 			break;
1245 
1246 		WRITE_ONCE(sqe->flags, 0);
1247 		qp->sq_get++;
1248 	}
1249 	if (async_event)
1250 		siw_qp_event(qp, IB_EVENT_SQ_DRAINED);
1251 }
1252 
1253 /*
1254  * siw_rq_flush()
1255  *
1256  * Flush recv queue entries to CQ. Also
1257  * takes care of pending active tagged and untagged
1258  * inbound transfers, which have target memory
1259  * referenced.
1260  *
1261  * Must be called with QP state write lock held.
1262  * Therefore, RQ lock must not be taken.
1263  */
1264 void siw_rq_flush(struct siw_qp *qp)
1265 {
1266 	struct siw_wqe *wqe = &qp->rx_untagged.wqe_active;
1267 
1268 	/*
1269 	 * Flush an in-progress untagged operation if present
1270 	 */
1271 	if (wqe->wr_status != SIW_WR_IDLE) {
1272 		siw_dbg_qp(qp, "flush current rqe, type %d, status %d\n",
1273 			   rx_type(wqe), wqe->wr_status);
1274 
1275 		siw_wqe_put_mem(wqe, rx_type(wqe));
1276 
1277 		if (rx_type(wqe) == SIW_OP_RECEIVE) {
1278 			siw_rqe_complete(qp, &wqe->rqe, wqe->bytes,
1279 					 0, SIW_WC_WR_FLUSH_ERR);
1280 		} else if (rx_type(wqe) != SIW_OP_READ &&
1281 			   rx_type(wqe) != SIW_OP_READ_RESPONSE &&
1282 			   rx_type(wqe) != SIW_OP_WRITE) {
1283 			siw_sqe_complete(qp, &wqe->sqe, 0, SIW_WC_WR_FLUSH_ERR);
1284 		}
1285 		wqe->wr_status = SIW_WR_IDLE;
1286 	}
1287 	wqe = &qp->rx_tagged.wqe_active;
1288 
1289 	if (wqe->wr_status != SIW_WR_IDLE) {
1290 		siw_wqe_put_mem(wqe, rx_type(wqe));
1291 		wqe->wr_status = SIW_WR_IDLE;
1292 	}
1293 	/*
1294 	 * Flush the Receive Queue
1295 	 */
1296 	while (qp->attrs.rq_size) {
1297 		struct siw_rqe *rqe =
1298 			&qp->recvq[qp->rq_get % qp->attrs.rq_size];
1299 
1300 		if (!READ_ONCE(rqe->flags))
1301 			break;
1302 
1303 		if (siw_rqe_complete(qp, rqe, 0, 0, SIW_WC_WR_FLUSH_ERR) != 0)
1304 			break;
1305 
1306 		WRITE_ONCE(rqe->flags, 0);
1307 		qp->rq_get++;
1308 	}
1309 }
1310 
1311 int siw_qp_add(struct siw_device *sdev, struct siw_qp *qp)
1312 {
1313 	int rv = xa_alloc(&sdev->qp_xa, &qp->base_qp.qp_num, qp, xa_limit_32b,
1314 			  GFP_KERNEL);
1315 
1316 	if (!rv) {
1317 		kref_init(&qp->ref);
1318 		qp->sdev = sdev;
1319 		siw_dbg_qp(qp, "new QP\n");
1320 	}
1321 	return rv;
1322 }
1323 
1324 void siw_free_qp(struct kref *ref)
1325 {
1326 	struct siw_qp *found, *qp = container_of(ref, struct siw_qp, ref);
1327 	struct siw_device *sdev = qp->sdev;
1328 	unsigned long flags;
1329 
1330 	if (qp->cep)
1331 		siw_cep_put(qp->cep);
1332 
1333 	found = xa_erase(&sdev->qp_xa, qp_id(qp));
1334 	WARN_ON(found != qp);
1335 	spin_lock_irqsave(&sdev->lock, flags);
1336 	list_del(&qp->devq);
1337 	spin_unlock_irqrestore(&sdev->lock, flags);
1338 
1339 	vfree(qp->sendq);
1340 	vfree(qp->recvq);
1341 	vfree(qp->irq);
1342 	vfree(qp->orq);
1343 
1344 	siw_put_tx_cpu(qp->tx_cpu);
1345 
1346 	atomic_dec(&sdev->num_qp);
1347 }
1348