xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/qp.c (revision 7937bfbc0ca53fe7cdd0d54414f9296e273a518e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2009-2013 Chelsio, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 #include <sys/cdefs.h>
35 #include "opt_inet.h"
36 
37 #ifdef TCP_OFFLOAD
38 #include <sys/types.h>
39 #include <sys/malloc.h>
40 #include <sys/socket.h>
41 #include <sys/socketvar.h>
42 #include <sys/sockio.h>
43 #include <sys/taskqueue.h>
44 #include <netinet/in.h>
45 #include <net/route.h>
46 
47 #include <netinet/in_systm.h>
48 #include <netinet/in_pcb.h>
49 #include <netinet/ip.h>
50 #include <netinet/ip_var.h>
51 #include <netinet/tcp_var.h>
52 #include <netinet/tcp.h>
53 #include <netinet/tcpip.h>
54 
55 #include <netinet/toecore.h>
56 
57 struct sge_iq;
58 struct rss_header;
59 struct cpl_set_tcb_rpl;
60 #include <linux/types.h>
61 #include "offload.h"
62 #include "tom/t4_tom.h"
63 
64 #include "iw_cxgbe.h"
65 #include "user.h"
66 
67 static int creds(struct toepcb *toep, struct inpcb *inp, size_t wrsize);
68 static int max_fr_immd = T4_MAX_FR_IMMD;//SYSCTL parameter later...
69 
70 static int alloc_ird(struct c4iw_dev *dev, u32 ird)
71 {
72 	int ret = 0;
73 
74 	spin_lock_irq(&dev->lock);
75 	if (ird <= dev->avail_ird)
76 		dev->avail_ird -= ird;
77 	else
78 		ret = -ENOMEM;
79 	spin_unlock_irq(&dev->lock);
80 
81 	if (ret)
82 		log(LOG_WARNING, "%s: device IRD resources exhausted\n",
83 			device_get_nameunit(dev->rdev.adap->dev));
84 
85 	return ret;
86 }
87 
88 static void free_ird(struct c4iw_dev *dev, int ird)
89 {
90 	spin_lock_irq(&dev->lock);
91 	dev->avail_ird += ird;
92 	spin_unlock_irq(&dev->lock);
93 }
94 
95 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
96 {
97 	unsigned long flag;
98 	spin_lock_irqsave(&qhp->lock, flag);
99 	qhp->attr.state = state;
100 	spin_unlock_irqrestore(&qhp->lock, flag);
101 }
102 
103 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
104 		      struct c4iw_dev_ucontext *uctx)
105 {
106 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
107 	/*
108 	 * uP clears EQ contexts when the connection exits rdma mode,
109 	 * so no need to post a RESET WR for these EQs.
110 	 */
111 	dma_free_coherent(rhp->ibdev.dma_device,
112 			wq->rq.memsize, wq->rq.queue,
113 			dma_unmap_addr(&wq->rq, mapping));
114 	dma_free_coherent(rhp->ibdev.dma_device,
115 			wq->sq.memsize, wq->sq.queue,
116 			dma_unmap_addr(&wq->sq, mapping));
117 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
118 	kfree(wq->rq.sw_rq);
119 	kfree(wq->sq.sw_sq);
120 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
121 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
122 	return 0;
123 }
124 
125 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
126 		     struct t4_cq *rcq, struct t4_cq *scq,
127 		     struct c4iw_dev_ucontext *uctx)
128 {
129 	struct adapter *sc = rdev->adap;
130 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
131 	int user = (uctx != &rdev->uctx);
132 	struct fw_ri_res_wr *res_wr;
133 	struct fw_ri_res *res;
134 	int wr_len;
135 	struct c4iw_wr_wait wr_wait;
136 	int ret = 0;
137 	int eqsize;
138 	struct wrqe *wr;
139 	u64 sq_bar2_qoffset = 0, rq_bar2_qoffset = 0;
140 
141 	if (__predict_false(c4iw_stopped(rdev)))
142 		return -EIO;
143 
144 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
145 	if (!wq->sq.qid)
146 		return -ENOMEM;
147 
148 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
149 	if (!wq->rq.qid) {
150 		ret = -ENOMEM;
151 		goto free_sq_qid;
152 	}
153 
154 	if (!user) {
155 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
156 				 GFP_KERNEL);
157 		if (!wq->sq.sw_sq) {
158 			ret = -ENOMEM;
159 			goto free_rq_qid;
160 		}
161 
162 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
163 				 GFP_KERNEL);
164 		if (!wq->rq.sw_rq) {
165 			ret = -ENOMEM;
166 			goto free_sw_sq;
167 		}
168 	}
169 
170 	/*
171 	 * RQT must be a power of 2 and at least 16 deep.
172 	 */
173 	wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
174 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
175 	if (!wq->rq.rqt_hwaddr) {
176 		ret = -ENOMEM;
177 		goto free_sw_rq;
178 	}
179 
180 	/*QP memory, allocate DMAable memory for Send & Receive Queues */
181 	wq->sq.queue = dma_alloc_coherent(rhp->ibdev.dma_device, wq->sq.memsize,
182 				       &(wq->sq.dma_addr), GFP_KERNEL);
183 	if (!wq->sq.queue) {
184 		ret = -ENOMEM;
185 		goto free_hwaddr;
186 	}
187 	wq->sq.phys_addr = vtophys(wq->sq.queue);
188 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
189 	memset(wq->sq.queue, 0, wq->sq.memsize);
190 
191 	wq->rq.queue = dma_alloc_coherent(rhp->ibdev.dma_device,
192 			wq->rq.memsize, &(wq->rq.dma_addr), GFP_KERNEL);
193 	if (!wq->rq.queue) {
194 		ret = -ENOMEM;
195 		goto free_sq_dma;
196 	}
197 	wq->rq.phys_addr = vtophys(wq->rq.queue);
198 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
199 	memset(wq->rq.queue, 0, wq->rq.memsize);
200 
201 	CTR5(KTR_IW_CXGBE,
202 	    "%s QP sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx",
203 	    __func__,
204 	    wq->sq.queue, (unsigned long long)wq->sq.phys_addr,
205 	    wq->rq.queue, (unsigned long long)wq->rq.phys_addr);
206 
207 	/* Doorbell/WC regions, determine the BAR2 queue offset and qid. */
208 	t4_bar2_sge_qregs(rdev->adap, wq->sq.qid, T4_BAR2_QTYPE_EGRESS, user,
209 			&sq_bar2_qoffset, &wq->sq.bar2_qid);
210 	t4_bar2_sge_qregs(rdev->adap, wq->rq.qid, T4_BAR2_QTYPE_EGRESS, user,
211 			&rq_bar2_qoffset, &wq->rq.bar2_qid);
212 
213 	if (user) {
214 		/* Compute BAR2 DB/WC physical address(page-aligned) for
215 		 * Userspace mapping.
216 		 */
217 		wq->sq.bar2_pa = (rdev->bar2_pa + sq_bar2_qoffset) & PAGE_MASK;
218 		wq->rq.bar2_pa = (rdev->bar2_pa + rq_bar2_qoffset) & PAGE_MASK;
219 		CTR3(KTR_IW_CXGBE,
220 			"%s BAR2 DB/WC sq base pa 0x%llx rq base pa 0x%llx",
221 			__func__, (unsigned long long)wq->sq.bar2_pa,
222 			(unsigned long long)wq->rq.bar2_pa);
223 	} else {
224 		/* Compute BAR2 DB/WC virtual address to access in kernel. */
225 		wq->sq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
226 				sq_bar2_qoffset);
227 		wq->rq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
228 				rq_bar2_qoffset);
229 		CTR3(KTR_IW_CXGBE, "%s BAR2 DB/WC sq base va %p rq base va %p",
230 			__func__, (unsigned long long)wq->sq.bar2_va,
231 			(unsigned long long)wq->rq.bar2_va);
232 	}
233 
234 	wq->rdev = rdev;
235 	wq->rq.msn = 1;
236 
237 	/* build fw_ri_res_wr */
238 	wr_len = sizeof *res_wr + 2 * sizeof *res;
239 
240 	wr = alloc_wrqe(wr_len, &sc->sge.ctrlq[0]);
241 	if (wr == NULL) {
242 		ret = -ENOMEM;
243 		goto free_rq_dma;
244 	}
245         res_wr = wrtod(wr);
246 
247 	memset(res_wr, 0, wr_len);
248 	res_wr->op_nres = cpu_to_be32(
249 			V_FW_WR_OP(FW_RI_RES_WR) |
250 			V_FW_RI_RES_WR_NRES(2) |
251 			F_FW_WR_COMPL);
252 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
253 	res_wr->cookie = (unsigned long) &wr_wait;
254 	res = res_wr->res;
255 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
256 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
257 
258 	/* eqsize is the number of 64B entries plus the status page size. */
259 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
260 			rdev->hw_queue.t4_eq_status_entries;
261 
262 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
263 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
264 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
265 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
266 		V_FW_RI_RES_WR_IQID(scq->cqid));
267 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
268 		V_FW_RI_RES_WR_DCAEN(0) |
269 		V_FW_RI_RES_WR_DCACPU(0) |
270 		V_FW_RI_RES_WR_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
271 		    X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
272 		V_FW_RI_RES_WR_FBMAX(3) |
273 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
274 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
275 		V_FW_RI_RES_WR_EQSIZE(eqsize));
276 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
277 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
278 	res++;
279 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
280 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
281 
282 	/* eqsize is the number of 64B entries plus the status page size. */
283 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
284 			rdev->hw_queue.t4_eq_status_entries;
285 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
286 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
287 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
288 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
289 		V_FW_RI_RES_WR_IQID(rcq->cqid));
290 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
291 		V_FW_RI_RES_WR_DCAEN(0) |
292 		V_FW_RI_RES_WR_DCACPU(0) |
293 		V_FW_RI_RES_WR_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
294 		    X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
295 		V_FW_RI_RES_WR_FBMAX(3) |
296 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
297 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
298 		V_FW_RI_RES_WR_EQSIZE(eqsize));
299 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
300 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
301 
302 	c4iw_init_wr_wait(&wr_wait);
303 
304 	t4_wrq_tx(sc, wr);
305 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid,
306 			NULL, __func__);
307 	if (ret)
308 		goto free_rq_dma;
309 
310 	CTR5(KTR_IW_CXGBE,
311 	    "%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx",
312 	    __func__, wq->sq.qid, wq->rq.qid,
313 	    (unsigned long long)wq->sq.bar2_va,
314 	    (unsigned long long)wq->rq.bar2_va);
315 
316 	return 0;
317 free_rq_dma:
318 	dma_free_coherent(rhp->ibdev.dma_device,
319 			  wq->rq.memsize, wq->rq.queue,
320 			  dma_unmap_addr(&wq->rq, mapping));
321 free_sq_dma:
322 	dma_free_coherent(rhp->ibdev.dma_device,
323 			  wq->sq.memsize, wq->sq.queue,
324 			  dma_unmap_addr(&wq->sq, mapping));
325 free_hwaddr:
326 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
327 free_sw_rq:
328 	kfree(wq->rq.sw_rq);
329 free_sw_sq:
330 	kfree(wq->sq.sw_sq);
331 free_rq_qid:
332 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
333 free_sq_qid:
334 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
335 	return ret;
336 }
337 
338 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
339 		      const struct ib_send_wr *wr, int max, u32 *plenp)
340 {
341 	u8 *dstp, *srcp;
342 	u32 plen = 0;
343 	int i;
344 	int rem, len;
345 
346 	dstp = (u8 *)immdp->data;
347 	for (i = 0; i < wr->num_sge; i++) {
348 		if ((plen + wr->sg_list[i].length) > max)
349 			return -EMSGSIZE;
350 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
351 		plen += wr->sg_list[i].length;
352 		rem = wr->sg_list[i].length;
353 		while (rem) {
354 			if (dstp == (u8 *)&sq->queue[sq->size])
355 				dstp = (u8 *)sq->queue;
356 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
357 				len = rem;
358 			else
359 				len = (u8 *)&sq->queue[sq->size] - dstp;
360 			memcpy(dstp, srcp, len);
361 			dstp += len;
362 			srcp += len;
363 			rem -= len;
364 		}
365 	}
366 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
367 	if (len)
368 		memset(dstp, 0, len);
369 	immdp->op = FW_RI_DATA_IMMD;
370 	immdp->r1 = 0;
371 	immdp->r2 = 0;
372 	immdp->immdlen = cpu_to_be32(plen);
373 	*plenp = plen;
374 	return 0;
375 }
376 
377 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
378 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
379 		      int num_sge, u32 *plenp)
380 
381 {
382 	int i;
383 	u32 plen = 0;
384 	__be64 *flitp = (__be64 *)isglp->sge;
385 
386 	for (i = 0; i < num_sge; i++) {
387 		if ((plen + sg_list[i].length) < plen)
388 			return -EMSGSIZE;
389 		plen += sg_list[i].length;
390 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
391 				     sg_list[i].length);
392 		if (++flitp == queue_end)
393 			flitp = queue_start;
394 		*flitp = cpu_to_be64(sg_list[i].addr);
395 		if (++flitp == queue_end)
396 			flitp = queue_start;
397 	}
398 	*flitp = (__force __be64)0;
399 	isglp->op = FW_RI_DATA_ISGL;
400 	isglp->r1 = 0;
401 	isglp->nsge = cpu_to_be16(num_sge);
402 	isglp->r2 = 0;
403 	if (plenp)
404 		*plenp = plen;
405 	return 0;
406 }
407 
408 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
409 			   const struct ib_send_wr *wr, u8 *len16)
410 {
411 	u32 plen;
412 	int size;
413 	int ret;
414 
415 	if (wr->num_sge > T4_MAX_SEND_SGE)
416 		return -EINVAL;
417 	switch (wr->opcode) {
418 	case IB_WR_SEND:
419 		if (wr->send_flags & IB_SEND_SOLICITED)
420 			wqe->send.sendop_pkd = cpu_to_be32(
421 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
422 		else
423 			wqe->send.sendop_pkd = cpu_to_be32(
424 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
425 		wqe->send.stag_inv = 0;
426 		break;
427 	case IB_WR_SEND_WITH_INV:
428 		if (wr->send_flags & IB_SEND_SOLICITED)
429 			wqe->send.sendop_pkd = cpu_to_be32(
430 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
431 		else
432 			wqe->send.sendop_pkd = cpu_to_be32(
433 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
434 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
435 		break;
436 
437 	default:
438 		return -EINVAL;
439 	}
440 	wqe->send.r3 = 0;
441 	wqe->send.r4 = 0;
442 
443 	plen = 0;
444 	if (wr->num_sge) {
445 		if (wr->send_flags & IB_SEND_INLINE) {
446 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
447 					 T4_MAX_SEND_INLINE, &plen);
448 			if (ret)
449 				return ret;
450 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
451 			       plen;
452 		} else {
453 			ret = build_isgl((__be64 *)sq->queue,
454 					 (__be64 *)&sq->queue[sq->size],
455 					 wqe->send.u.isgl_src,
456 					 wr->sg_list, wr->num_sge, &plen);
457 			if (ret)
458 				return ret;
459 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
460 			       wr->num_sge * sizeof(struct fw_ri_sge);
461 		}
462 	} else {
463 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
464 		wqe->send.u.immd_src[0].r1 = 0;
465 		wqe->send.u.immd_src[0].r2 = 0;
466 		wqe->send.u.immd_src[0].immdlen = 0;
467 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
468 		plen = 0;
469 	}
470 	*len16 = DIV_ROUND_UP(size, 16);
471 	wqe->send.plen = cpu_to_be32(plen);
472 	return 0;
473 }
474 
475 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
476 			    const struct ib_send_wr *wr, u8 *len16)
477 {
478 	u32 plen;
479 	int size;
480 	int ret;
481 
482 	if (wr->num_sge > T4_MAX_SEND_SGE)
483 		return -EINVAL;
484 	wqe->write.immd_data = 0;
485 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
486 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
487 	if (wr->num_sge) {
488 		if (wr->send_flags & IB_SEND_INLINE) {
489 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
490 					 T4_MAX_WRITE_INLINE, &plen);
491 			if (ret)
492 				return ret;
493 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
494 			       plen;
495 		} else {
496 			ret = build_isgl((__be64 *)sq->queue,
497 					 (__be64 *)&sq->queue[sq->size],
498 					 wqe->write.u.isgl_src,
499 					 wr->sg_list, wr->num_sge, &plen);
500 			if (ret)
501 				return ret;
502 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
503 			       wr->num_sge * sizeof(struct fw_ri_sge);
504 		}
505 	} else {
506 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
507 		wqe->write.u.immd_src[0].r1 = 0;
508 		wqe->write.u.immd_src[0].r2 = 0;
509 		wqe->write.u.immd_src[0].immdlen = 0;
510 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
511 		plen = 0;
512 	}
513 	*len16 = DIV_ROUND_UP(size, 16);
514 	wqe->write.plen = cpu_to_be32(plen);
515 	return 0;
516 }
517 
518 static int build_rdma_read(union t4_wr *wqe, const struct ib_send_wr *wr, u8 *len16)
519 {
520 	if (wr->num_sge > 1)
521 		return -EINVAL;
522 	if (wr->num_sge && wr->sg_list[0].length) {
523 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
524 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
525 							>> 32));
526 		wqe->read.to_src_lo =
527 			cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
528 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
529 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
530 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
531 							 >> 32));
532 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
533 	} else {
534 		wqe->read.stag_src = cpu_to_be32(2);
535 		wqe->read.to_src_hi = 0;
536 		wqe->read.to_src_lo = 0;
537 		wqe->read.stag_sink = cpu_to_be32(2);
538 		wqe->read.plen = 0;
539 		wqe->read.to_sink_hi = 0;
540 		wqe->read.to_sink_lo = 0;
541 	}
542 	wqe->read.r2 = 0;
543 	wqe->read.r5 = 0;
544 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
545 	return 0;
546 }
547 
548 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
549 			   const struct ib_recv_wr *wr, u8 *len16)
550 {
551 	int ret;
552 
553 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
554 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
555 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
556 	if (ret)
557 		return ret;
558 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
559 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
560 	return 0;
561 }
562 
563 static int build_inv_stag(union t4_wr *wqe, const struct ib_send_wr *wr,
564 			  u8 *len16)
565 {
566 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
567 	wqe->inv.r2 = 0;
568 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
569 	return 0;
570 }
571 
572 static void free_qp_work(struct work_struct *work)
573 {
574 	struct c4iw_ucontext *ucontext;
575 	struct c4iw_qp *qhp;
576 	struct c4iw_dev *rhp;
577 
578 	qhp = container_of(work, struct c4iw_qp, free_work);
579 	ucontext = qhp->ucontext;
580 	rhp = qhp->rhp;
581 
582 	CTR3(KTR_IW_CXGBE, "%s qhp %p ucontext %p", __func__,
583 			qhp, ucontext);
584 	destroy_qp(&rhp->rdev, &qhp->wq,
585 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
586 
587 	kfree(qhp);
588 }
589 
590 static void queue_qp_free(struct kref *kref)
591 {
592 	struct c4iw_qp *qhp;
593 
594 	qhp = container_of(kref, struct c4iw_qp, kref);
595 	CTR2(KTR_IW_CXGBE, "%s qhp %p", __func__, qhp);
596 	queue_work(qhp->rhp->rdev.free_workq, &qhp->free_work);
597 }
598 
599 void c4iw_qp_add_ref(struct ib_qp *qp)
600 {
601 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
602 	kref_get(&to_c4iw_qp(qp)->kref);
603 }
604 
605 void c4iw_qp_rem_ref(struct ib_qp *qp)
606 {
607 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
608 	kref_put(&to_c4iw_qp(qp)->kref, queue_qp_free);
609 }
610 
611 static void complete_sq_drain_wr(struct c4iw_qp *qhp, const struct ib_send_wr *wr)
612 {
613 	struct t4_cqe cqe = {};
614 	struct c4iw_cq *schp;
615 	unsigned long flag;
616 	struct t4_cq *cq;
617 
618 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
619 	cq = &schp->cq;
620 
621 	PDBG("%s drain sq id %u\n", __func__, qhp->wq.sq.qid);
622 	cqe.u.drain_cookie = wr->wr_id;
623 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
624 				 V_CQE_OPCODE(C4IW_DRAIN_OPCODE) |
625 				 V_CQE_TYPE(1) |
626 				 V_CQE_SWCQE(1) |
627 				 V_CQE_QPID(qhp->wq.sq.qid));
628 
629 	spin_lock_irqsave(&schp->lock, flag);
630 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
631 	cq->sw_queue[cq->sw_pidx] = cqe;
632 	t4_swcq_produce(cq);
633 	spin_unlock_irqrestore(&schp->lock, flag);
634 
635 	spin_lock_irqsave(&schp->comp_handler_lock, flag);
636 	(*schp->ibcq.comp_handler)(&schp->ibcq,
637 				   schp->ibcq.cq_context);
638 	spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
639 }
640 
641 static void complete_rq_drain_wr(struct c4iw_qp *qhp, const struct ib_recv_wr *wr)
642 {
643 	struct t4_cqe cqe = {};
644 	struct c4iw_cq *rchp;
645 	unsigned long flag;
646 	struct t4_cq *cq;
647 
648 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
649 	cq = &rchp->cq;
650 
651 	PDBG("%s drain rq id %u\n", __func__, qhp->wq.sq.qid);
652 	cqe.u.drain_cookie = wr->wr_id;
653 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
654 				 V_CQE_OPCODE(C4IW_DRAIN_OPCODE) |
655 				 V_CQE_TYPE(0) |
656 				 V_CQE_SWCQE(1) |
657 				 V_CQE_QPID(qhp->wq.sq.qid));
658 
659 	spin_lock_irqsave(&rchp->lock, flag);
660 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
661 	cq->sw_queue[cq->sw_pidx] = cqe;
662 	t4_swcq_produce(cq);
663 	spin_unlock_irqrestore(&rchp->lock, flag);
664 
665 	spin_lock_irqsave(&rchp->comp_handler_lock, flag);
666 	(*rchp->ibcq.comp_handler)(&rchp->ibcq,
667 				   rchp->ibcq.cq_context);
668 	spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
669 }
670 
671 static int build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
672 		const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16)
673 {
674 	__be64 *p = (__be64 *)fr->pbl;
675 
676 	if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE)
677 		return -EINVAL;
678 
679 	fr->r2 = cpu_to_be32(0);
680 	fr->stag = cpu_to_be32(mhp->ibmr.rkey);
681 
682 	fr->tpte.valid_to_pdid = cpu_to_be32(F_FW_RI_TPTE_VALID |
683 			V_FW_RI_TPTE_STAGKEY((mhp->ibmr.rkey & M_FW_RI_TPTE_STAGKEY)) |
684 			V_FW_RI_TPTE_STAGSTATE(1) |
685 			V_FW_RI_TPTE_STAGTYPE(FW_RI_STAG_NSMR) |
686 			V_FW_RI_TPTE_PDID(mhp->attr.pdid));
687 	fr->tpte.locread_to_qpid = cpu_to_be32(
688 			V_FW_RI_TPTE_PERM(c4iw_ib_to_tpt_access(wr->access)) |
689 			V_FW_RI_TPTE_ADDRTYPE(FW_RI_VA_BASED_TO) |
690 			V_FW_RI_TPTE_PS(ilog2(wr->mr->page_size) - 12));
691 	fr->tpte.nosnoop_pbladdr = cpu_to_be32(V_FW_RI_TPTE_PBLADDR(
692 			      PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
693 	fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
694 	fr->tpte.len_hi = cpu_to_be32(mhp->ibmr.length >> 32);
695 	fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff);
696 	fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
697 	fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
698 
699 	p[0] = cpu_to_be64((u64)mhp->mpl[0]);
700 	p[1] = cpu_to_be64((u64)mhp->mpl[1]);
701 
702 	*len16 = DIV_ROUND_UP(sizeof(*fr), 16);
703 	return 0;
704 }
705 
706 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
707 		const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16,
708 		bool dsgl_supported)
709 {
710 	struct fw_ri_immd *imdp;
711 	__be64 *p;
712 	int i;
713 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
714 	int rem;
715 
716 	if (mhp->mpl_len > t4_max_fr_depth(&mhp->rhp->rdev, use_dsgl))
717 		return -EINVAL;
718 	if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE)
719 		return -EINVAL;
720 
721 	wqe->fr.qpbinde_to_dcacpu = 0;
722 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
723 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
724 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
725 	wqe->fr.len_hi = cpu_to_be32(mhp->ibmr.length >> 32);
726 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff);
727 	wqe->fr.stag = cpu_to_be32(wr->key);
728 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
729 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
730 
731 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
732 		struct fw_ri_dsgl *sglp;
733 
734 		for (i = 0; i < mhp->mpl_len; i++)
735 			mhp->mpl[i] =
736 				     (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
737 
738 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
739 		sglp->op = FW_RI_DATA_DSGL;
740 		sglp->r1 = 0;
741 		sglp->nsge = cpu_to_be16(1);
742 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
743 		sglp->len0 = cpu_to_be32(pbllen);
744 
745 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
746 	} else {
747 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
748 		imdp->op = FW_RI_DATA_IMMD;
749 		imdp->r1 = 0;
750 		imdp->r2 = 0;
751 		imdp->immdlen = cpu_to_be32(pbllen);
752 		p = (__be64 *)(imdp + 1);
753 		rem = pbllen;
754 		for (i = 0; i < mhp->mpl_len; i++) {
755 			*p = cpu_to_be64((u64)mhp->mpl[i]);
756 			rem -= sizeof(*p);
757 			if (++p == (__be64 *)&sq->queue[sq->size])
758 				p = (__be64 *)sq->queue;
759 		}
760 		BUG_ON(rem < 0);
761 		while (rem) {
762 			*p = 0;
763 			rem -= sizeof(*p);
764 			if (++p == (__be64 *)&sq->queue[sq->size])
765 				p = (__be64 *)sq->queue;
766 		}
767 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
768 				+ pbllen, 16);
769 	}
770 
771 	return 0;
772 }
773 
774 int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
775 		   const struct ib_send_wr **bad_wr)
776 {
777 	int err = 0;
778 	u8 len16 = 0;
779 	enum fw_wr_opcodes fw_opcode = 0;
780 	enum fw_ri_wr_flags fw_flags;
781 	struct c4iw_qp *qhp;
782 	union t4_wr *wqe = NULL;
783 	u32 num_wrs;
784 	struct t4_swsqe *swsqe;
785 	unsigned long flag;
786 	u16 idx = 0;
787 	struct c4iw_rdev *rdev;
788 
789 	qhp = to_c4iw_qp(ibqp);
790 	rdev = &qhp->rhp->rdev;
791 	if (__predict_false(c4iw_stopped(rdev)))
792 		return -EIO;
793 	spin_lock_irqsave(&qhp->lock, flag);
794 	if (t4_wq_in_error(&qhp->wq)) {
795 		spin_unlock_irqrestore(&qhp->lock, flag);
796 		complete_sq_drain_wr(qhp, wr);
797 		return err;
798 	}
799 	num_wrs = t4_sq_avail(&qhp->wq);
800 	if (num_wrs == 0) {
801 		spin_unlock_irqrestore(&qhp->lock, flag);
802 		*bad_wr = wr;
803 		return -ENOMEM;
804 	}
805 	while (wr) {
806 		if (num_wrs == 0) {
807 			err = -ENOMEM;
808 			*bad_wr = wr;
809 			break;
810 		}
811 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
812 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
813 
814 		fw_flags = 0;
815 		if (wr->send_flags & IB_SEND_SOLICITED)
816 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
817 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
818 			fw_flags |= FW_RI_COMPLETION_FLAG;
819 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
820 		switch (wr->opcode) {
821 		case IB_WR_SEND_WITH_INV:
822 		case IB_WR_SEND:
823 			if (wr->send_flags & IB_SEND_FENCE)
824 				fw_flags |= FW_RI_READ_FENCE_FLAG;
825 			fw_opcode = FW_RI_SEND_WR;
826 			if (wr->opcode == IB_WR_SEND)
827 				swsqe->opcode = FW_RI_SEND;
828 			else
829 				swsqe->opcode = FW_RI_SEND_WITH_INV;
830 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
831 			break;
832 		case IB_WR_RDMA_WRITE:
833 			fw_opcode = FW_RI_RDMA_WRITE_WR;
834 			swsqe->opcode = FW_RI_RDMA_WRITE;
835 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
836 			break;
837 		case IB_WR_RDMA_READ:
838 		case IB_WR_RDMA_READ_WITH_INV:
839 			fw_opcode = FW_RI_RDMA_READ_WR;
840 			swsqe->opcode = FW_RI_READ_REQ;
841 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
842 				c4iw_invalidate_mr(qhp->rhp,
843 						   wr->sg_list[0].lkey);
844 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
845 			} else {
846 				fw_flags = 0;
847 			}
848 			err = build_rdma_read(wqe, wr, &len16);
849 			if (err)
850 				break;
851 			swsqe->read_len = wr->sg_list[0].length;
852 			if (!qhp->wq.sq.oldest_read)
853 				qhp->wq.sq.oldest_read = swsqe;
854 			break;
855 		case IB_WR_REG_MR: {
856 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
857 
858 			swsqe->opcode = FW_RI_FAST_REGISTER;
859 			if (rdev->adap->params.fr_nsmr_tpte_wr_support &&
860 					!mhp->attr.state && mhp->mpl_len <= 2) {
861 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
862 				err = build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
863 						mhp, &len16);
864 			} else {
865 				fw_opcode = FW_RI_FR_NSMR_WR;
866 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
867 					mhp, &len16,
868 					rdev->adap->params.ulptx_memwrite_dsgl);
869 			}
870 			if (err)
871 				break;
872 			mhp->attr.state = 1;
873 			break;
874 		}
875 		case IB_WR_LOCAL_INV:
876 			if (wr->send_flags & IB_SEND_FENCE)
877 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
878 			fw_opcode = FW_RI_INV_LSTAG_WR;
879 			swsqe->opcode = FW_RI_LOCAL_INV;
880 			err = build_inv_stag(wqe, wr, &len16);
881 			c4iw_invalidate_mr(qhp->rhp, wr->ex.invalidate_rkey);
882 			break;
883 		default:
884 			CTR2(KTR_IW_CXGBE, "%s post of type =%d TBD!", __func__,
885 			     wr->opcode);
886 			err = -EINVAL;
887 		}
888 		if (err) {
889 			*bad_wr = wr;
890 			break;
891 		}
892 		swsqe->idx = qhp->wq.sq.pidx;
893 		swsqe->complete = 0;
894 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
895 					qhp->sq_sig_all;
896 		swsqe->flushed = 0;
897 		swsqe->wr_id = wr->wr_id;
898 
899 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
900 
901 		CTR5(KTR_IW_CXGBE,
902 		    "%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u",
903 		    __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
904 		    swsqe->opcode, swsqe->read_len);
905 		wr = wr->next;
906 		num_wrs--;
907 		t4_sq_produce(&qhp->wq, len16);
908 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
909 	}
910 
911 	t4_ring_sq_db(&qhp->wq, idx, wqe, rdev->adap->iwt.wc_en);
912 	spin_unlock_irqrestore(&qhp->lock, flag);
913 	return err;
914 }
915 
916 int c4iw_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
917 		      const struct ib_recv_wr **bad_wr)
918 {
919 	int err = 0;
920 	struct c4iw_qp *qhp;
921 	union t4_recv_wr *wqe = NULL;
922 	u32 num_wrs;
923 	u8 len16 = 0;
924 	unsigned long flag;
925 	u16 idx = 0;
926 
927 	qhp = to_c4iw_qp(ibqp);
928 	if (__predict_false(c4iw_stopped(&qhp->rhp->rdev)))
929 		return -EIO;
930 	spin_lock_irqsave(&qhp->lock, flag);
931 	if (t4_wq_in_error(&qhp->wq)) {
932 		spin_unlock_irqrestore(&qhp->lock, flag);
933 		complete_rq_drain_wr(qhp, wr);
934 		return err;
935 	}
936 	num_wrs = t4_rq_avail(&qhp->wq);
937 	if (num_wrs == 0) {
938 		spin_unlock_irqrestore(&qhp->lock, flag);
939 		*bad_wr = wr;
940 		return -ENOMEM;
941 	}
942 	while (wr) {
943 		if (wr->num_sge > T4_MAX_RECV_SGE) {
944 			err = -EINVAL;
945 			*bad_wr = wr;
946 			break;
947 		}
948 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
949 					   qhp->wq.rq.wq_pidx *
950 					   T4_EQ_ENTRY_SIZE);
951 		if (num_wrs)
952 			err = build_rdma_recv(qhp, wqe, wr, &len16);
953 		else
954 			err = -ENOMEM;
955 		if (err) {
956 			*bad_wr = wr;
957 			break;
958 		}
959 
960 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
961 
962 		wqe->recv.opcode = FW_RI_RECV_WR;
963 		wqe->recv.r1 = 0;
964 		wqe->recv.wrid = qhp->wq.rq.pidx;
965 		wqe->recv.r2[0] = 0;
966 		wqe->recv.r2[1] = 0;
967 		wqe->recv.r2[2] = 0;
968 		wqe->recv.len16 = len16;
969 		CTR3(KTR_IW_CXGBE, "%s cookie 0x%llx pidx %u", __func__,
970 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
971 		t4_rq_produce(&qhp->wq, len16);
972 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
973 		wr = wr->next;
974 		num_wrs--;
975 	}
976 
977 	t4_ring_rq_db(&qhp->wq, idx, wqe, qhp->rhp->rdev.adap->iwt.wc_en);
978 	spin_unlock_irqrestore(&qhp->lock, flag);
979 	return err;
980 }
981 
982 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
983 				    u8 *ecode)
984 {
985 	int status;
986 	int tagged;
987 	int opcode;
988 	int rqtype;
989 	int send_inv;
990 
991 	if (!err_cqe) {
992 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
993 		*ecode = 0;
994 		return;
995 	}
996 
997 	status = CQE_STATUS(err_cqe);
998 	opcode = CQE_OPCODE(err_cqe);
999 	rqtype = RQ_TYPE(err_cqe);
1000 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
1001 		   (opcode == FW_RI_SEND_WITH_SE_INV);
1002 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
1003 		 (rqtype && (opcode == FW_RI_READ_RESP));
1004 
1005 	switch (status) {
1006 	case T4_ERR_STAG:
1007 		if (send_inv) {
1008 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1009 			*ecode = RDMAP_CANT_INV_STAG;
1010 		} else {
1011 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1012 			*ecode = RDMAP_INV_STAG;
1013 		}
1014 		break;
1015 	case T4_ERR_PDID:
1016 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1017 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1018 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1019 			*ecode = RDMAP_CANT_INV_STAG;
1020 		else
1021 			*ecode = RDMAP_STAG_NOT_ASSOC;
1022 		break;
1023 	case T4_ERR_QPID:
1024 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1025 		*ecode = RDMAP_STAG_NOT_ASSOC;
1026 		break;
1027 	case T4_ERR_ACCESS:
1028 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1029 		*ecode = RDMAP_ACC_VIOL;
1030 		break;
1031 	case T4_ERR_WRAP:
1032 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1033 		*ecode = RDMAP_TO_WRAP;
1034 		break;
1035 	case T4_ERR_BOUND:
1036 		if (tagged) {
1037 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1038 			*ecode = DDPT_BASE_BOUNDS;
1039 		} else {
1040 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1041 			*ecode = RDMAP_BASE_BOUNDS;
1042 		}
1043 		break;
1044 	case T4_ERR_INVALIDATE_SHARED_MR:
1045 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1046 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1047 		*ecode = RDMAP_CANT_INV_STAG;
1048 		break;
1049 	case T4_ERR_ECC:
1050 	case T4_ERR_ECC_PSTAG:
1051 	case T4_ERR_INTERNAL_ERR:
1052 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1053 		*ecode = 0;
1054 		break;
1055 	case T4_ERR_OUT_OF_RQE:
1056 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1057 		*ecode = DDPU_INV_MSN_NOBUF;
1058 		break;
1059 	case T4_ERR_PBL_ADDR_BOUND:
1060 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1061 		*ecode = DDPT_BASE_BOUNDS;
1062 		break;
1063 	case T4_ERR_CRC:
1064 		*layer_type = LAYER_MPA|DDP_LLP;
1065 		*ecode = MPA_CRC_ERR;
1066 		break;
1067 	case T4_ERR_MARKER:
1068 		*layer_type = LAYER_MPA|DDP_LLP;
1069 		*ecode = MPA_MARKER_ERR;
1070 		break;
1071 	case T4_ERR_PDU_LEN_ERR:
1072 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1073 		*ecode = DDPU_MSG_TOOBIG;
1074 		break;
1075 	case T4_ERR_DDP_VERSION:
1076 		if (tagged) {
1077 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1078 			*ecode = DDPT_INV_VERS;
1079 		} else {
1080 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1081 			*ecode = DDPU_INV_VERS;
1082 		}
1083 		break;
1084 	case T4_ERR_RDMA_VERSION:
1085 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1086 		*ecode = RDMAP_INV_VERS;
1087 		break;
1088 	case T4_ERR_OPCODE:
1089 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1090 		*ecode = RDMAP_INV_OPCODE;
1091 		break;
1092 	case T4_ERR_DDP_QUEUE_NUM:
1093 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1094 		*ecode = DDPU_INV_QN;
1095 		break;
1096 	case T4_ERR_MSN:
1097 	case T4_ERR_MSN_GAP:
1098 	case T4_ERR_MSN_RANGE:
1099 	case T4_ERR_IRD_OVERFLOW:
1100 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1101 		*ecode = DDPU_INV_MSN_RANGE;
1102 		break;
1103 	case T4_ERR_TBIT:
1104 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1105 		*ecode = 0;
1106 		break;
1107 	case T4_ERR_MO:
1108 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1109 		*ecode = DDPU_INV_MO;
1110 		break;
1111 	default:
1112 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1113 		*ecode = 0;
1114 		break;
1115 	}
1116 }
1117 
1118 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1119 			   gfp_t gfp)
1120 {
1121 	int ret;
1122 	struct fw_ri_wr *wqe;
1123 	struct terminate_message *term;
1124 	struct wrqe *wr;
1125 	struct socket *so = qhp->ep->com.so;
1126         struct inpcb *inp = sotoinpcb(so);
1127         struct tcpcb *tp = intotcpcb(inp);
1128         struct toepcb *toep = tp->t_toe;
1129 
1130 	CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1131 	    qhp->wq.sq.qid, qhp->ep->hwtid);
1132 
1133 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1134 	if (wr == NULL)
1135 		return;
1136         wqe = wrtod(wr);
1137 
1138 	memset(wqe, 0, sizeof *wqe);
1139 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR));
1140 	wqe->flowid_len16 = cpu_to_be32(
1141 		V_FW_WR_FLOWID(qhp->ep->hwtid) |
1142 		V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1143 
1144 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1145 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1146 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1147 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1148 		term->layer_etype = qhp->attr.layer_etype;
1149 		term->ecode = qhp->attr.ecode;
1150 	} else
1151 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1152 	ret = creds(toep, inp, sizeof(*wqe));
1153 	if (ret) {
1154 		free_wrqe(wr);
1155 		return;
1156 	}
1157 	t4_wrq_tx(qhp->rhp->rdev.adap, wr);
1158 }
1159 
1160 /* Assumes qhp lock is held. */
1161 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1162 		       struct c4iw_cq *schp)
1163 {
1164 	int count;
1165 	int rq_flushed, sq_flushed;
1166 	unsigned long flag;
1167 
1168 	CTR4(KTR_IW_CXGBE, "%s qhp %p rchp %p schp %p", __func__, qhp, rchp,
1169 	    schp);
1170 
1171 	/* locking hierarchy: cq lock first, then qp lock. */
1172 	spin_lock_irqsave(&rchp->lock, flag);
1173 	spin_lock(&qhp->lock);
1174 
1175 	if (qhp->wq.flushed) {
1176 		spin_unlock(&qhp->lock);
1177 		spin_unlock_irqrestore(&rchp->lock, flag);
1178 		return;
1179 	}
1180 	qhp->wq.flushed = 1;
1181 
1182 	c4iw_flush_hw_cq(rchp);
1183 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1184 	rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1185 	spin_unlock(&qhp->lock);
1186 	spin_unlock_irqrestore(&rchp->lock, flag);
1187 
1188 	/* locking hierarchy: cq lock first, then qp lock. */
1189 	spin_lock_irqsave(&schp->lock, flag);
1190 	spin_lock(&qhp->lock);
1191 	if (schp != rchp)
1192 		c4iw_flush_hw_cq(schp);
1193 	sq_flushed = c4iw_flush_sq(qhp);
1194 	spin_unlock(&qhp->lock);
1195 	spin_unlock_irqrestore(&schp->lock, flag);
1196 
1197 	if (schp == rchp) {
1198 		if (t4_clear_cq_armed(&rchp->cq) &&
1199 		    (rq_flushed || sq_flushed)) {
1200 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1201 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1202 						   rchp->ibcq.cq_context);
1203 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1204 		}
1205 	} else {
1206 		if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1207 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1208 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1209 						   rchp->ibcq.cq_context);
1210 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1211 		}
1212 		if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1213 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1214 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1215 						   schp->ibcq.cq_context);
1216 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1217 		}
1218 	}
1219 }
1220 
1221 static void flush_qp(struct c4iw_qp *qhp)
1222 {
1223 	struct c4iw_cq *rchp, *schp;
1224 	unsigned long flag;
1225 
1226 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1227 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1228 
1229 	t4_set_wq_in_error(&qhp->wq);
1230 	if (qhp->ibqp.uobject) {
1231 		t4_set_cq_in_error(&rchp->cq);
1232 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1233 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1234 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1235 		if (schp != rchp) {
1236 			t4_set_cq_in_error(&schp->cq);
1237 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1238 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1239 					schp->ibcq.cq_context);
1240 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1241 		}
1242 		return;
1243 	}
1244 	__flush_qp(qhp, rchp, schp);
1245 }
1246 
1247 static int
1248 rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp, struct c4iw_ep *ep)
1249 {
1250 	struct c4iw_rdev *rdev = &rhp->rdev;
1251 	struct adapter *sc = rdev->adap;
1252 	struct fw_ri_wr *wqe;
1253 	int ret;
1254 	struct wrqe *wr;
1255 	struct socket *so = ep->com.so;
1256         struct inpcb *inp = sotoinpcb(so);
1257         struct tcpcb *tp = intotcpcb(inp);
1258         struct toepcb *toep = tp->t_toe;
1259 
1260 	KASSERT(rhp == qhp->rhp && ep == qhp->ep, ("%s: EDOOFUS", __func__));
1261 
1262 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1263 	    qhp->wq.sq.qid, ep, ep->hwtid);
1264 
1265 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1266 	if (wr == NULL)
1267 		return (0);
1268 	wqe = wrtod(wr);
1269 
1270 	memset(wqe, 0, sizeof *wqe);
1271 
1272 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL);
1273 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1274 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1275 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1276 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1277 
1278 	c4iw_init_wr_wait(&ep->com.wr_wait);
1279 
1280 	ret = creds(toep, inp, sizeof(*wqe));
1281 	if (ret) {
1282 		free_wrqe(wr);
1283 		return ret;
1284 	}
1285 	t4_wrq_tx(sc, wr);
1286 
1287 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1288 			qhp->wq.sq.qid, ep->com.so, __func__);
1289 	return ret;
1290 }
1291 
1292 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1293 {
1294 	CTR2(KTR_IW_CXGBE, "%s p2p_type = %d", __func__, p2p_type);
1295 	memset(&init->u, 0, sizeof init->u);
1296 	switch (p2p_type) {
1297 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1298 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1299 		init->u.write.stag_sink = cpu_to_be32(1);
1300 		init->u.write.to_sink = cpu_to_be64(1);
1301 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1302 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1303 						   sizeof(struct fw_ri_immd),
1304 						   16);
1305 		break;
1306 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1307 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1308 		init->u.read.stag_src = cpu_to_be32(1);
1309 		init->u.read.to_src_lo = cpu_to_be32(1);
1310 		init->u.read.stag_sink = cpu_to_be32(1);
1311 		init->u.read.to_sink_lo = cpu_to_be32(1);
1312 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1313 		break;
1314 	}
1315 }
1316 
1317 static int
1318 creds(struct toepcb *toep, struct inpcb *inp, size_t wrsize)
1319 {
1320 	struct ofld_tx_sdesc *txsd;
1321 
1322 	CTR3(KTR_IW_CXGBE, "%s:creB  %p %u", __func__, toep , wrsize);
1323 	INP_WLOCK(inp);
1324 	if ((inp->inp_flags & INP_DROPPED) != 0) {
1325 		INP_WUNLOCK(inp);
1326 		return (EINVAL);
1327 	}
1328 	txsd = &toep->txsd[toep->txsd_pidx];
1329 	txsd->tx_credits = howmany(wrsize, 16);
1330 	txsd->plen = 0;
1331 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
1332 			("%s: not enough credits (%d)", __func__, toep->tx_credits));
1333 	toep->tx_credits -= txsd->tx_credits;
1334 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1335 		toep->txsd_pidx = 0;
1336 	toep->txsd_avail--;
1337 	INP_WUNLOCK(inp);
1338 	CTR5(KTR_IW_CXGBE, "%s:creE  %p %u %u %u", __func__, toep ,
1339 	    txsd->tx_credits, toep->tx_credits, toep->txsd_pidx);
1340 	return (0);
1341 }
1342 
1343 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1344 {
1345 	struct fw_ri_wr *wqe;
1346 	int ret;
1347 	struct wrqe *wr;
1348 	struct c4iw_ep *ep = qhp->ep;
1349 	struct c4iw_rdev *rdev = &qhp->rhp->rdev;
1350 	struct adapter *sc = rdev->adap;
1351 	struct socket *so = ep->com.so;
1352         struct inpcb *inp = sotoinpcb(so);
1353         struct tcpcb *tp = intotcpcb(inp);
1354         struct toepcb *toep = tp->t_toe;
1355 
1356 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1357 	    qhp->wq.sq.qid, ep, ep->hwtid);
1358 
1359 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1360 	if (wr == NULL)
1361 		return (0);
1362 	wqe = wrtod(wr);
1363 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1364 	if (ret) {
1365 		qhp->attr.max_ird = 0;
1366 		free_wrqe(wr);
1367 		return ret;
1368 	}
1369 
1370 	memset(wqe, 0, sizeof *wqe);
1371 
1372 	wqe->op_compl = cpu_to_be32(
1373 		V_FW_WR_OP(FW_RI_WR) |
1374 		F_FW_WR_COMPL);
1375 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1376 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1377 
1378 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1379 
1380 	wqe->u.init.type = FW_RI_TYPE_INIT;
1381 	wqe->u.init.mpareqbit_p2ptype =
1382 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1383 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1384 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1385 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1386 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1387 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1388 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1389 	if (qhp->attr.mpa_attr.crc_enabled)
1390 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1391 
1392 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1393 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1394 			    FW_RI_QP_BIND_ENABLE;
1395 	if (!qhp->ibqp.uobject)
1396 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1397 				     FW_RI_QP_STAG0_ENABLE;
1398 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1399 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1400 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1401 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1402 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1403 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1404 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1405 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1406 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1407 	wqe->u.init.iss = cpu_to_be32(ep->snd_seq);
1408 	wqe->u.init.irs = cpu_to_be32(ep->rcv_seq);
1409 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1410 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1411 	    sc->vres.rq.start);
1412 	if (qhp->attr.mpa_attr.initiator)
1413 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1414 
1415 	c4iw_init_wr_wait(&ep->com.wr_wait);
1416 
1417 	ret = creds(toep, inp, sizeof(*wqe));
1418 	if (ret) {
1419 		free_wrqe(wr);
1420 		free_ird(rhp, qhp->attr.max_ird);
1421 		return ret;
1422 	}
1423 	t4_wrq_tx(sc, wr);
1424 
1425 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1426 			qhp->wq.sq.qid, ep->com.so, __func__);
1427 
1428 	toep->params.ulp_mode = ULP_MODE_RDMA;
1429 	free_ird(rhp, qhp->attr.max_ird);
1430 
1431 	return ret;
1432 }
1433 
1434 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1435 		   enum c4iw_qp_attr_mask mask,
1436 		   struct c4iw_qp_attributes *attrs,
1437 		   int internal)
1438 {
1439 	int ret = 0;
1440 	struct c4iw_qp_attributes newattr = qhp->attr;
1441 	int disconnect = 0;
1442 	int terminate = 0;
1443 	int abort = 0;
1444 	int free = 0;
1445 	struct c4iw_ep *ep = NULL;
1446 
1447 	CTR5(KTR_IW_CXGBE, "%s qhp %p sqid 0x%x rqid 0x%x ep %p", __func__, qhp,
1448 	    qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep);
1449 	CTR3(KTR_IW_CXGBE, "%s state %d -> %d", __func__, qhp->attr.state,
1450 	    (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1451 
1452 	mutex_lock(&qhp->mutex);
1453 
1454 	/* Process attr changes if in IDLE */
1455 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1456 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1457 			ret = -EIO;
1458 			goto out;
1459 		}
1460 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1461 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1462 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1463 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1464 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1465 			newattr.enable_bind = attrs->enable_bind;
1466 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1467 			if (attrs->max_ord > c4iw_max_read_depth) {
1468 				ret = -EINVAL;
1469 				goto out;
1470 			}
1471 			newattr.max_ord = attrs->max_ord;
1472 		}
1473 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1474 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1475 				ret = -EINVAL;
1476 				goto out;
1477 			}
1478 			newattr.max_ird = attrs->max_ird;
1479 		}
1480 		qhp->attr = newattr;
1481 	}
1482 
1483 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1484 		goto out;
1485 	if (qhp->attr.state == attrs->next_state)
1486 		goto out;
1487 
1488 	/* Return EINPROGRESS if QP is already in transition state.
1489 	 * Eg: CLOSING->IDLE transition or *->ERROR transition.
1490 	 * This can happen while connection is switching(due to rdma_fini)
1491 	 * from iWARP/RDDP to TOE mode and any inflight RDMA RX data will
1492 	 * reach TOE driver -> TCP stack -> iWARP driver. In this way
1493 	 * iWARP driver keep receiving inflight RDMA RX data until socket
1494 	 * is closed or aborted. And if iWARP CM is in FPDU sate, then
1495 	 * it tries to put QP in TERM state and disconnects endpoint.
1496 	 * But as QP is already in transition state, this event is ignored.
1497 	 */
1498 	if ((qhp->attr.state >= C4IW_QP_STATE_ERROR) &&
1499 		(attrs->next_state == C4IW_QP_STATE_TERMINATE)) {
1500 		ret = -EINPROGRESS;
1501 		goto out;
1502 	}
1503 
1504 	switch (qhp->attr.state) {
1505 	case C4IW_QP_STATE_IDLE:
1506 		switch (attrs->next_state) {
1507 		case C4IW_QP_STATE_RTS:
1508 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1509 				ret = -EINVAL;
1510 				goto out;
1511 			}
1512 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1513 				ret = -EINVAL;
1514 				goto out;
1515 			}
1516 			qhp->attr.mpa_attr = attrs->mpa_attr;
1517 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1518 			qhp->ep = qhp->attr.llp_stream_handle;
1519 			set_state(qhp, C4IW_QP_STATE_RTS);
1520 
1521 			/*
1522 			 * Ref the endpoint here and deref when we
1523 			 * disassociate the endpoint from the QP.  This
1524 			 * happens in CLOSING->IDLE transition or *->ERROR
1525 			 * transition.
1526 			 */
1527 			c4iw_get_ep(&qhp->ep->com);
1528 			ret = rdma_init(rhp, qhp);
1529 			if (ret)
1530 				goto err;
1531 			break;
1532 		case C4IW_QP_STATE_ERROR:
1533 			set_state(qhp, C4IW_QP_STATE_ERROR);
1534 			flush_qp(qhp);
1535 			break;
1536 		default:
1537 			ret = -EINVAL;
1538 			goto out;
1539 		}
1540 		break;
1541 	case C4IW_QP_STATE_RTS:
1542 		switch (attrs->next_state) {
1543 		case C4IW_QP_STATE_CLOSING:
1544 			BUG_ON(kref_read(&qhp->ep->com.kref) < 2);
1545 			t4_set_wq_in_error(&qhp->wq);
1546 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1547 			ep = qhp->ep;
1548 			if (!internal) {
1549 				abort = 0;
1550 				disconnect = 1;
1551 				c4iw_get_ep(&qhp->ep->com);
1552 			}
1553 			ret = rdma_fini(rhp, qhp, ep);
1554 			if (ret)
1555 				goto err;
1556 			break;
1557 		case C4IW_QP_STATE_TERMINATE:
1558 			t4_set_wq_in_error(&qhp->wq);
1559 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1560 			qhp->attr.layer_etype = attrs->layer_etype;
1561 			qhp->attr.ecode = attrs->ecode;
1562 			ep = qhp->ep;
1563 			if (!internal) {
1564 				c4iw_get_ep(&qhp->ep->com);
1565 				terminate = 1;
1566 				disconnect = 1;
1567 			} else {
1568 				terminate = qhp->attr.send_term;
1569 				ret = rdma_fini(rhp, qhp, ep);
1570 				if (ret)
1571 					goto err;
1572 			}
1573 			break;
1574 		case C4IW_QP_STATE_ERROR:
1575 			t4_set_wq_in_error(&qhp->wq);
1576 			set_state(qhp, C4IW_QP_STATE_ERROR);
1577 			if (!internal) {
1578 				abort = 1;
1579 				disconnect = 1;
1580 				ep = qhp->ep;
1581 				c4iw_get_ep(&qhp->ep->com);
1582 			}
1583 			goto err;
1584 			break;
1585 		default:
1586 			ret = -EINVAL;
1587 			goto out;
1588 		}
1589 		break;
1590 	case C4IW_QP_STATE_CLOSING:
1591 
1592 		/*
1593 		 * Allow kernel users to move to ERROR for qp draining.
1594 		 */
1595 		if (!internal && (qhp->ibqp.uobject || attrs->next_state !=
1596 				  C4IW_QP_STATE_ERROR)) {
1597 			ret = -EINVAL;
1598 			goto out;
1599 		}
1600 		switch (attrs->next_state) {
1601 		case C4IW_QP_STATE_IDLE:
1602 			flush_qp(qhp);
1603 			set_state(qhp, C4IW_QP_STATE_IDLE);
1604 			qhp->attr.llp_stream_handle = NULL;
1605 			c4iw_put_ep(&qhp->ep->com);
1606 			qhp->ep = NULL;
1607 			wake_up(&qhp->wait);
1608 			break;
1609 		case C4IW_QP_STATE_ERROR:
1610 			goto err;
1611 		default:
1612 			ret = -EINVAL;
1613 			goto err;
1614 		}
1615 		break;
1616 	case C4IW_QP_STATE_ERROR:
1617 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1618 			ret = -EINVAL;
1619 			goto out;
1620 		}
1621 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1622 			ret = -EINVAL;
1623 			goto out;
1624 		}
1625 		set_state(qhp, C4IW_QP_STATE_IDLE);
1626 		break;
1627 	case C4IW_QP_STATE_TERMINATE:
1628 		if (!internal) {
1629 			ret = -EINVAL;
1630 			goto out;
1631 		}
1632 		goto err;
1633 		break;
1634 	default:
1635 		printf("%s in a bad state %d\n",
1636 		       __func__, qhp->attr.state);
1637 		ret = -EINVAL;
1638 		goto err;
1639 		break;
1640 	}
1641 	goto out;
1642 err:
1643 	CTR3(KTR_IW_CXGBE, "%s disassociating ep %p qpid 0x%x", __func__,
1644 	    qhp->ep, qhp->wq.sq.qid);
1645 
1646 	/* disassociate the LLP connection */
1647 	qhp->attr.llp_stream_handle = NULL;
1648 	if (!ep)
1649 		ep = qhp->ep;
1650 	qhp->ep = NULL;
1651 	set_state(qhp, C4IW_QP_STATE_ERROR);
1652 	free = 1;
1653 	abort = 1;
1654 	BUG_ON(!ep);
1655 	flush_qp(qhp);
1656 	wake_up(&qhp->wait);
1657 out:
1658 	mutex_unlock(&qhp->mutex);
1659 
1660 	if (terminate)
1661 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1662 
1663 	/*
1664 	 * If disconnect is 1, then we need to initiate a disconnect
1665 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1666 	 * an abnormal close (RTS/CLOSING->ERROR).
1667 	 */
1668 	if (disconnect) {
1669 		__c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1670 							 GFP_KERNEL);
1671 		c4iw_put_ep(&ep->com);
1672 	}
1673 
1674 	/*
1675 	 * If free is 1, then we've disassociated the EP from the QP
1676 	 * and we need to dereference the EP.
1677 	 */
1678 	if (free)
1679 		c4iw_put_ep(&ep->com);
1680 	CTR2(KTR_IW_CXGBE, "%s exit state %d", __func__, qhp->attr.state);
1681 	return ret;
1682 }
1683 
1684 int c4iw_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
1685 {
1686 	struct c4iw_dev *rhp;
1687 	struct c4iw_qp *qhp;
1688 	struct c4iw_qp_attributes attrs;
1689 
1690 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ib_qp);
1691 	qhp = to_c4iw_qp(ib_qp);
1692 	rhp = qhp->rhp;
1693 
1694 	attrs.next_state = C4IW_QP_STATE_ERROR;
1695 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1696 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1697 	else
1698 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1699 	wait_event(qhp->wait, !qhp->ep);
1700 
1701 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1702 
1703 	free_ird(rhp, qhp->attr.max_ird);
1704 	c4iw_qp_rem_ref(ib_qp);
1705 
1706 	CTR3(KTR_IW_CXGBE, "%s ib_qp %p qpid 0x%0x", __func__, ib_qp,
1707 	    qhp->wq.sq.qid);
1708 	return 0;
1709 }
1710 
1711 struct ib_qp *
1712 c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1713     struct ib_udata *udata)
1714 {
1715 	struct c4iw_dev *rhp;
1716 	struct c4iw_qp *qhp;
1717 	struct c4iw_pd *php;
1718 	struct c4iw_cq *schp;
1719 	struct c4iw_cq *rchp;
1720 	struct c4iw_create_qp_resp uresp;
1721 	unsigned int sqsize, rqsize;
1722 	struct c4iw_ucontext *ucontext;
1723 	int ret;
1724 	struct c4iw_mm_entry *sq_key_mm = NULL, *rq_key_mm = NULL;
1725 	struct c4iw_mm_entry *sq_db_key_mm = NULL, *rq_db_key_mm = NULL;
1726 
1727 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
1728 
1729 	if (attrs->qp_type != IB_QPT_RC)
1730 		return ERR_PTR(-EINVAL);
1731 
1732 	php = to_c4iw_pd(pd);
1733 	rhp = php->rhp;
1734 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1735 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1736 	if (!schp || !rchp)
1737 		return ERR_PTR(-EINVAL);
1738 
1739 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1740 		return ERR_PTR(-EINVAL);
1741 
1742 	if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1743 		return ERR_PTR(-E2BIG);
1744 	rqsize = attrs->cap.max_recv_wr + 1;
1745 	if (rqsize < 8)
1746 		rqsize = 8;
1747 
1748 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1749 		return ERR_PTR(-E2BIG);
1750 	sqsize = attrs->cap.max_send_wr + 1;
1751 	if (sqsize < 8)
1752 		sqsize = 8;
1753 
1754 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1755 
1756 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1757 	if (!qhp)
1758 		return ERR_PTR(-ENOMEM);
1759 	qhp->wq.sq.size = sqsize;
1760 	qhp->wq.sq.memsize =
1761 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1762 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1763 	qhp->wq.sq.flush_cidx = -1;
1764 	qhp->wq.rq.size = rqsize;
1765 	qhp->wq.rq.memsize =
1766 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1767 		sizeof(*qhp->wq.rq.queue);
1768 
1769 	if (ucontext) {
1770 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1771 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1772 	}
1773 
1774 	CTR5(KTR_IW_CXGBE, "%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu",
1775 	    __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1776 
1777 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1778 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1779 	if (ret)
1780 		goto err1;
1781 
1782 	attrs->cap.max_recv_wr = rqsize - 1;
1783 	attrs->cap.max_send_wr = sqsize - 1;
1784 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1785 
1786 	qhp->rhp = rhp;
1787 	qhp->attr.pd = php->pdid;
1788 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1789 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1790 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1791 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1792 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1793 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1794 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1795 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1796 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1797 	qhp->attr.enable_rdma_read = 1;
1798 	qhp->attr.enable_rdma_write = 1;
1799 	qhp->attr.enable_bind = 1;
1800 	qhp->attr.max_ord = 0;
1801 	qhp->attr.max_ird = 0;
1802 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1803 	spin_lock_init(&qhp->lock);
1804 	mutex_init(&qhp->mutex);
1805 	init_waitqueue_head(&qhp->wait);
1806 	kref_init(&qhp->kref);
1807 	INIT_WORK(&qhp->free_work, free_qp_work);
1808 
1809 	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1810 	if (ret)
1811 		goto err2;
1812 
1813 	if (udata) {
1814 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
1815 		if (!sq_key_mm) {
1816 			ret = -ENOMEM;
1817 			goto err3;
1818 		}
1819 		rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
1820 		if (!rq_key_mm) {
1821 			ret = -ENOMEM;
1822 			goto err4;
1823 		}
1824 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
1825 		if (!sq_db_key_mm) {
1826 			ret = -ENOMEM;
1827 			goto err5;
1828 		}
1829 		rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
1830 		if (!rq_db_key_mm) {
1831 			ret = -ENOMEM;
1832 			goto err6;
1833 		}
1834 		uresp.flags = 0;
1835 		uresp.qid_mask = rhp->rdev.qpmask;
1836 		uresp.sqid = qhp->wq.sq.qid;
1837 		uresp.sq_size = qhp->wq.sq.size;
1838 		uresp.sq_memsize = qhp->wq.sq.memsize;
1839 		uresp.rqid = qhp->wq.rq.qid;
1840 		uresp.rq_size = qhp->wq.rq.size;
1841 		uresp.rq_memsize = qhp->wq.rq.memsize;
1842 		spin_lock(&ucontext->mmap_lock);
1843 		uresp.ma_sync_key =  0;
1844 		uresp.sq_key = ucontext->key;
1845 		ucontext->key += PAGE_SIZE;
1846 		uresp.rq_key = ucontext->key;
1847 		ucontext->key += PAGE_SIZE;
1848 		uresp.sq_db_gts_key = ucontext->key;
1849 		ucontext->key += PAGE_SIZE;
1850 		uresp.rq_db_gts_key = ucontext->key;
1851 		ucontext->key += PAGE_SIZE;
1852 		spin_unlock(&ucontext->mmap_lock);
1853 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1854 		if (ret)
1855 			goto err7;
1856 		sq_key_mm->key = uresp.sq_key;
1857 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
1858 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1859 		CTR4(KTR_IW_CXGBE, "%s sq_key_mm %x, %x, %d", __func__,
1860 				sq_key_mm->key, sq_key_mm->addr,
1861 				sq_key_mm->len);
1862 		insert_mmap(ucontext, sq_key_mm);
1863 		rq_key_mm->key = uresp.rq_key;
1864 		rq_key_mm->addr = qhp->wq.rq.phys_addr;
1865 		rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1866 		CTR4(KTR_IW_CXGBE, "%s rq_key_mm %x, %x, %d", __func__,
1867 				rq_key_mm->key, rq_key_mm->addr,
1868 				rq_key_mm->len);
1869 		insert_mmap(ucontext, rq_key_mm);
1870 		sq_db_key_mm->key = uresp.sq_db_gts_key;
1871 		sq_db_key_mm->addr = (u64)qhp->wq.sq.bar2_pa;
1872 		sq_db_key_mm->len = PAGE_SIZE;
1873 		CTR4(KTR_IW_CXGBE, "%s sq_db_key_mm %x, %x, %d", __func__,
1874 				sq_db_key_mm->key, sq_db_key_mm->addr,
1875 				sq_db_key_mm->len);
1876 		insert_mmap(ucontext, sq_db_key_mm);
1877 		rq_db_key_mm->key = uresp.rq_db_gts_key;
1878 		rq_db_key_mm->addr = (u64)qhp->wq.rq.bar2_pa;
1879 		rq_db_key_mm->len = PAGE_SIZE;
1880 		CTR4(KTR_IW_CXGBE, "%s rq_db_key_mm %x, %x, %d", __func__,
1881 				rq_db_key_mm->key, rq_db_key_mm->addr,
1882 				rq_db_key_mm->len);
1883 		insert_mmap(ucontext, rq_db_key_mm);
1884 
1885 		qhp->ucontext = ucontext;
1886 	}
1887 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1888 	init_timer(&(qhp->timer));
1889 
1890 	CTR5(KTR_IW_CXGBE, "%s sq id %u size %u memsize %zu num_entries %u",
1891 		 __func__, qhp->wq.sq.qid,
1892 		 qhp->wq.sq.size, qhp->wq.sq.memsize, attrs->cap.max_send_wr);
1893 	CTR5(KTR_IW_CXGBE, "%s rq id %u size %u memsize %zu num_entries %u",
1894 		 __func__, qhp->wq.rq.qid,
1895 		 qhp->wq.rq.size, qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1896 	return &qhp->ibqp;
1897 err7:
1898 	kfree(rq_db_key_mm);
1899 err6:
1900 	kfree(sq_db_key_mm);
1901 err5:
1902 	kfree(rq_key_mm);
1903 err4:
1904 	kfree(sq_key_mm);
1905 err3:
1906 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1907 err2:
1908 	destroy_qp(&rhp->rdev, &qhp->wq,
1909 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1910 err1:
1911 	kfree(qhp);
1912 	return ERR_PTR(ret);
1913 }
1914 
1915 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1916 		      int attr_mask, struct ib_udata *udata)
1917 {
1918 	struct c4iw_dev *rhp;
1919 	struct c4iw_qp *qhp;
1920 	enum c4iw_qp_attr_mask mask = 0;
1921 	struct c4iw_qp_attributes attrs;
1922 
1923 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ibqp);
1924 
1925 	/* iwarp does not support the RTR state */
1926 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1927 		attr_mask &= ~IB_QP_STATE;
1928 
1929 	/* Make sure we still have something left to do */
1930 	if (!attr_mask)
1931 		return 0;
1932 
1933 	memset(&attrs, 0, sizeof attrs);
1934 	qhp = to_c4iw_qp(ibqp);
1935 	rhp = qhp->rhp;
1936 
1937 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1938 	attrs.enable_rdma_read = (attr->qp_access_flags &
1939 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1940 	attrs.enable_rdma_write = (attr->qp_access_flags &
1941 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1942 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1943 
1944 
1945 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1946 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1947 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1948 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1949 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1950 
1951 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1952 }
1953 
1954 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1955 {
1956 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p qpn 0x%x", __func__, dev, qpn);
1957 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1958 }
1959 
1960 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1961 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1962 {
1963 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1964 
1965 	memset(attr, 0, sizeof *attr);
1966 	memset(init_attr, 0, sizeof *init_attr);
1967 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1968 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1969 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1970 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1971 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1972 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1973 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1974 	return 0;
1975 }
1976 #endif
1977