xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/qp.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
36 
37 #include "opt_inet.h"
38 
39 #ifdef TCP_OFFLOAD
40 #include <sys/types.h>
41 #include <sys/malloc.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/sockio.h>
45 #include <sys/taskqueue.h>
46 #include <netinet/in.h>
47 #include <net/route.h>
48 
49 #include <netinet/in_systm.h>
50 #include <netinet/in_pcb.h>
51 #include <netinet/ip.h>
52 #include <netinet/ip_var.h>
53 #include <netinet/tcp_var.h>
54 #include <netinet/tcp.h>
55 #include <netinet/tcpip.h>
56 
57 #include <netinet/toecore.h>
58 
59 struct sge_iq;
60 struct rss_header;
61 struct cpl_set_tcb_rpl;
62 #include <linux/types.h>
63 #include "offload.h"
64 #include "tom/t4_tom.h"
65 
66 #include "iw_cxgbe.h"
67 #include "user.h"
68 
69 static int creds(struct toepcb *toep, struct inpcb *inp, size_t wrsize);
70 static int max_fr_immd = T4_MAX_FR_IMMD;//SYSCTL parameter later...
71 
72 static int alloc_ird(struct c4iw_dev *dev, u32 ird)
73 {
74 	int ret = 0;
75 
76 	spin_lock_irq(&dev->lock);
77 	if (ird <= dev->avail_ird)
78 		dev->avail_ird -= ird;
79 	else
80 		ret = -ENOMEM;
81 	spin_unlock_irq(&dev->lock);
82 
83 	if (ret)
84 		log(LOG_WARNING, "%s: device IRD resources exhausted\n",
85 			device_get_nameunit(dev->rdev.adap->dev));
86 
87 	return ret;
88 }
89 
90 static void free_ird(struct c4iw_dev *dev, int ird)
91 {
92 	spin_lock_irq(&dev->lock);
93 	dev->avail_ird += ird;
94 	spin_unlock_irq(&dev->lock);
95 }
96 
97 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
98 {
99 	unsigned long flag;
100 	spin_lock_irqsave(&qhp->lock, flag);
101 	qhp->attr.state = state;
102 	spin_unlock_irqrestore(&qhp->lock, flag);
103 }
104 
105 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
106 		      struct c4iw_dev_ucontext *uctx)
107 {
108 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
109 	/*
110 	 * uP clears EQ contexts when the connection exits rdma mode,
111 	 * so no need to post a RESET WR for these EQs.
112 	 */
113 	dma_free_coherent(rhp->ibdev.dma_device,
114 			wq->rq.memsize, wq->rq.queue,
115 			dma_unmap_addr(&wq->rq, mapping));
116 	dma_free_coherent(rhp->ibdev.dma_device,
117 			wq->sq.memsize, wq->sq.queue,
118 			dma_unmap_addr(&wq->sq, mapping));
119 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
120 	kfree(wq->rq.sw_rq);
121 	kfree(wq->sq.sw_sq);
122 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
123 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
124 	return 0;
125 }
126 
127 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
128 		     struct t4_cq *rcq, struct t4_cq *scq,
129 		     struct c4iw_dev_ucontext *uctx)
130 {
131 	struct adapter *sc = rdev->adap;
132 	struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev);
133 	int user = (uctx != &rdev->uctx);
134 	struct fw_ri_res_wr *res_wr;
135 	struct fw_ri_res *res;
136 	int wr_len;
137 	struct c4iw_wr_wait wr_wait;
138 	int ret = 0;
139 	int eqsize;
140 	struct wrqe *wr;
141 	u64 sq_bar2_qoffset = 0, rq_bar2_qoffset = 0;
142 
143 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
144 	if (!wq->sq.qid)
145 		return -ENOMEM;
146 
147 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
148 	if (!wq->rq.qid) {
149 		ret = -ENOMEM;
150 		goto free_sq_qid;
151 	}
152 
153 	if (!user) {
154 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
155 				 GFP_KERNEL);
156 		if (!wq->sq.sw_sq) {
157 			ret = -ENOMEM;
158 			goto free_rq_qid;
159 		}
160 
161 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
162 				 GFP_KERNEL);
163 		if (!wq->rq.sw_rq) {
164 			ret = -ENOMEM;
165 			goto free_sw_sq;
166 		}
167 	}
168 
169 	/*
170 	 * RQT must be a power of 2 and at least 16 deep.
171 	 */
172 	wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
173 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
174 	if (!wq->rq.rqt_hwaddr) {
175 		ret = -ENOMEM;
176 		goto free_sw_rq;
177 	}
178 
179 	/*QP memory, allocate DMAable memory for Send & Receive Queues */
180 	wq->sq.queue = dma_alloc_coherent(rhp->ibdev.dma_device, wq->sq.memsize,
181 				       &(wq->sq.dma_addr), GFP_KERNEL);
182 	if (!wq->sq.queue) {
183 		ret = -ENOMEM;
184 		goto free_hwaddr;
185 	}
186 	wq->sq.phys_addr = vtophys(wq->sq.queue);
187 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
188 	memset(wq->sq.queue, 0, wq->sq.memsize);
189 
190 	wq->rq.queue = dma_alloc_coherent(rhp->ibdev.dma_device,
191 			wq->rq.memsize, &(wq->rq.dma_addr), GFP_KERNEL);
192 	if (!wq->rq.queue) {
193 		ret = -ENOMEM;
194 		goto free_sq_dma;
195 	}
196 	wq->rq.phys_addr = vtophys(wq->rq.queue);
197 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
198 	memset(wq->rq.queue, 0, wq->rq.memsize);
199 
200 	CTR5(KTR_IW_CXGBE,
201 	    "%s QP sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx",
202 	    __func__,
203 	    wq->sq.queue, (unsigned long long)wq->sq.phys_addr,
204 	    wq->rq.queue, (unsigned long long)wq->rq.phys_addr);
205 
206 	/* Doorbell/WC regions, determine the BAR2 queue offset and qid. */
207 	t4_bar2_sge_qregs(rdev->adap, wq->sq.qid, T4_BAR2_QTYPE_EGRESS, user,
208 			&sq_bar2_qoffset, &wq->sq.bar2_qid);
209 	t4_bar2_sge_qregs(rdev->adap, wq->rq.qid, T4_BAR2_QTYPE_EGRESS, user,
210 			&rq_bar2_qoffset, &wq->rq.bar2_qid);
211 
212 	if (user) {
213 		/* Compute BAR2 DB/WC physical address(page-aligned) for
214 		 * Userspace mapping.
215 		 */
216 		wq->sq.bar2_pa = (rdev->bar2_pa + sq_bar2_qoffset) & PAGE_MASK;
217 		wq->rq.bar2_pa = (rdev->bar2_pa + rq_bar2_qoffset) & PAGE_MASK;
218 		CTR3(KTR_IW_CXGBE,
219 			"%s BAR2 DB/WC sq base pa 0x%llx rq base pa 0x%llx",
220 			__func__, (unsigned long long)wq->sq.bar2_pa,
221 			(unsigned long long)wq->rq.bar2_pa);
222 	} else {
223 		/* Compute BAR2 DB/WC virtual address to access in kernel. */
224 		wq->sq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
225 				sq_bar2_qoffset);
226 		wq->rq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva +
227 				rq_bar2_qoffset);
228 		CTR3(KTR_IW_CXGBE, "%s BAR2 DB/WC sq base va %p rq base va %p",
229 			__func__, (unsigned long long)wq->sq.bar2_va,
230 			(unsigned long long)wq->rq.bar2_va);
231 	}
232 
233 	wq->rdev = rdev;
234 	wq->rq.msn = 1;
235 
236 	/* build fw_ri_res_wr */
237 	wr_len = sizeof *res_wr + 2 * sizeof *res;
238 
239 	wr = alloc_wrqe(wr_len, &sc->sge.ctrlq[0]);
240 	if (wr == NULL) {
241 		ret = -ENOMEM;
242 		goto free_rq_dma;
243 	}
244         res_wr = wrtod(wr);
245 
246 	memset(res_wr, 0, wr_len);
247 	res_wr->op_nres = cpu_to_be32(
248 			V_FW_WR_OP(FW_RI_RES_WR) |
249 			V_FW_RI_RES_WR_NRES(2) |
250 			F_FW_WR_COMPL);
251 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
252 	res_wr->cookie = (unsigned long) &wr_wait;
253 	res = res_wr->res;
254 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
255 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
256 
257 	/* eqsize is the number of 64B entries plus the status page size. */
258 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
259 			rdev->hw_queue.t4_eq_status_entries;
260 
261 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
262 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
263 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
264 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
265 		V_FW_RI_RES_WR_IQID(scq->cqid));
266 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
267 		V_FW_RI_RES_WR_DCAEN(0) |
268 		V_FW_RI_RES_WR_DCACPU(0) |
269 		V_FW_RI_RES_WR_FBMIN(2) |
270 		V_FW_RI_RES_WR_FBMAX(3) |
271 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
272 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
273 		V_FW_RI_RES_WR_EQSIZE(eqsize));
274 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
275 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
276 	res++;
277 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
278 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
279 
280 	/* eqsize is the number of 64B entries plus the status page size. */
281 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
282 			rdev->hw_queue.t4_eq_status_entries;
283 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
284 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
285 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
286 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
287 		V_FW_RI_RES_WR_IQID(rcq->cqid));
288 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
289 		V_FW_RI_RES_WR_DCAEN(0) |
290 		V_FW_RI_RES_WR_DCACPU(0) |
291 		V_FW_RI_RES_WR_FBMIN(2) |
292 		V_FW_RI_RES_WR_FBMAX(3) |
293 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
294 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
295 		V_FW_RI_RES_WR_EQSIZE(eqsize));
296 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
297 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
298 
299 	c4iw_init_wr_wait(&wr_wait);
300 
301 	t4_wrq_tx(sc, wr);
302 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid,
303 			NULL, __func__);
304 	if (ret)
305 		goto free_rq_dma;
306 
307 	CTR5(KTR_IW_CXGBE,
308 	    "%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx",
309 	    __func__, wq->sq.qid, wq->rq.qid,
310 	    (unsigned long long)wq->sq.bar2_va,
311 	    (unsigned long long)wq->rq.bar2_va);
312 
313 	return 0;
314 free_rq_dma:
315 	dma_free_coherent(rhp->ibdev.dma_device,
316 			  wq->rq.memsize, wq->rq.queue,
317 			  dma_unmap_addr(&wq->rq, mapping));
318 free_sq_dma:
319 	dma_free_coherent(rhp->ibdev.dma_device,
320 			  wq->sq.memsize, wq->sq.queue,
321 			  dma_unmap_addr(&wq->sq, mapping));
322 free_hwaddr:
323 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
324 free_sw_rq:
325 	kfree(wq->rq.sw_rq);
326 free_sw_sq:
327 	kfree(wq->sq.sw_sq);
328 free_rq_qid:
329 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
330 free_sq_qid:
331 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
332 	return ret;
333 }
334 
335 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
336 		      struct ib_send_wr *wr, int max, u32 *plenp)
337 {
338 	u8 *dstp, *srcp;
339 	u32 plen = 0;
340 	int i;
341 	int rem, len;
342 
343 	dstp = (u8 *)immdp->data;
344 	for (i = 0; i < wr->num_sge; i++) {
345 		if ((plen + wr->sg_list[i].length) > max)
346 			return -EMSGSIZE;
347 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
348 		plen += wr->sg_list[i].length;
349 		rem = wr->sg_list[i].length;
350 		while (rem) {
351 			if (dstp == (u8 *)&sq->queue[sq->size])
352 				dstp = (u8 *)sq->queue;
353 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
354 				len = rem;
355 			else
356 				len = (u8 *)&sq->queue[sq->size] - dstp;
357 			memcpy(dstp, srcp, len);
358 			dstp += len;
359 			srcp += len;
360 			rem -= len;
361 		}
362 	}
363 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
364 	if (len)
365 		memset(dstp, 0, len);
366 	immdp->op = FW_RI_DATA_IMMD;
367 	immdp->r1 = 0;
368 	immdp->r2 = 0;
369 	immdp->immdlen = cpu_to_be32(plen);
370 	*plenp = plen;
371 	return 0;
372 }
373 
374 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
375 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
376 		      int num_sge, u32 *plenp)
377 
378 {
379 	int i;
380 	u32 plen = 0;
381 	__be64 *flitp = (__be64 *)isglp->sge;
382 
383 	for (i = 0; i < num_sge; i++) {
384 		if ((plen + sg_list[i].length) < plen)
385 			return -EMSGSIZE;
386 		plen += sg_list[i].length;
387 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
388 				     sg_list[i].length);
389 		if (++flitp == queue_end)
390 			flitp = queue_start;
391 		*flitp = cpu_to_be64(sg_list[i].addr);
392 		if (++flitp == queue_end)
393 			flitp = queue_start;
394 	}
395 	*flitp = (__force __be64)0;
396 	isglp->op = FW_RI_DATA_ISGL;
397 	isglp->r1 = 0;
398 	isglp->nsge = cpu_to_be16(num_sge);
399 	isglp->r2 = 0;
400 	if (plenp)
401 		*plenp = plen;
402 	return 0;
403 }
404 
405 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
406 			   struct ib_send_wr *wr, u8 *len16)
407 {
408 	u32 plen;
409 	int size;
410 	int ret;
411 
412 	if (wr->num_sge > T4_MAX_SEND_SGE)
413 		return -EINVAL;
414 	switch (wr->opcode) {
415 	case IB_WR_SEND:
416 		if (wr->send_flags & IB_SEND_SOLICITED)
417 			wqe->send.sendop_pkd = cpu_to_be32(
418 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
419 		else
420 			wqe->send.sendop_pkd = cpu_to_be32(
421 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
422 		wqe->send.stag_inv = 0;
423 		break;
424 	case IB_WR_SEND_WITH_INV:
425 		if (wr->send_flags & IB_SEND_SOLICITED)
426 			wqe->send.sendop_pkd = cpu_to_be32(
427 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
428 		else
429 			wqe->send.sendop_pkd = cpu_to_be32(
430 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
431 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
432 		break;
433 
434 	default:
435 		return -EINVAL;
436 	}
437 	wqe->send.r3 = 0;
438 	wqe->send.r4 = 0;
439 
440 	plen = 0;
441 	if (wr->num_sge) {
442 		if (wr->send_flags & IB_SEND_INLINE) {
443 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
444 					 T4_MAX_SEND_INLINE, &plen);
445 			if (ret)
446 				return ret;
447 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
448 			       plen;
449 		} else {
450 			ret = build_isgl((__be64 *)sq->queue,
451 					 (__be64 *)&sq->queue[sq->size],
452 					 wqe->send.u.isgl_src,
453 					 wr->sg_list, wr->num_sge, &plen);
454 			if (ret)
455 				return ret;
456 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
457 			       wr->num_sge * sizeof(struct fw_ri_sge);
458 		}
459 	} else {
460 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
461 		wqe->send.u.immd_src[0].r1 = 0;
462 		wqe->send.u.immd_src[0].r2 = 0;
463 		wqe->send.u.immd_src[0].immdlen = 0;
464 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
465 		plen = 0;
466 	}
467 	*len16 = DIV_ROUND_UP(size, 16);
468 	wqe->send.plen = cpu_to_be32(plen);
469 	return 0;
470 }
471 
472 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
473 			    struct ib_send_wr *wr, u8 *len16)
474 {
475 	u32 plen;
476 	int size;
477 	int ret;
478 
479 	if (wr->num_sge > T4_MAX_SEND_SGE)
480 		return -EINVAL;
481 	wqe->write.immd_data = 0;
482 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
483 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
484 	if (wr->num_sge) {
485 		if (wr->send_flags & IB_SEND_INLINE) {
486 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
487 					 T4_MAX_WRITE_INLINE, &plen);
488 			if (ret)
489 				return ret;
490 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
491 			       plen;
492 		} else {
493 			ret = build_isgl((__be64 *)sq->queue,
494 					 (__be64 *)&sq->queue[sq->size],
495 					 wqe->write.u.isgl_src,
496 					 wr->sg_list, wr->num_sge, &plen);
497 			if (ret)
498 				return ret;
499 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
500 			       wr->num_sge * sizeof(struct fw_ri_sge);
501 		}
502 	} else {
503 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
504 		wqe->write.u.immd_src[0].r1 = 0;
505 		wqe->write.u.immd_src[0].r2 = 0;
506 		wqe->write.u.immd_src[0].immdlen = 0;
507 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
508 		plen = 0;
509 	}
510 	*len16 = DIV_ROUND_UP(size, 16);
511 	wqe->write.plen = cpu_to_be32(plen);
512 	return 0;
513 }
514 
515 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
516 {
517 	if (wr->num_sge > 1)
518 		return -EINVAL;
519 	if (wr->num_sge && wr->sg_list[0].length) {
520 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
521 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
522 							>> 32));
523 		wqe->read.to_src_lo =
524 			cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
525 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
526 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
527 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
528 							 >> 32));
529 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
530 	} else {
531 		wqe->read.stag_src = cpu_to_be32(2);
532 		wqe->read.to_src_hi = 0;
533 		wqe->read.to_src_lo = 0;
534 		wqe->read.stag_sink = cpu_to_be32(2);
535 		wqe->read.plen = 0;
536 		wqe->read.to_sink_hi = 0;
537 		wqe->read.to_sink_lo = 0;
538 	}
539 	wqe->read.r2 = 0;
540 	wqe->read.r5 = 0;
541 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
542 	return 0;
543 }
544 
545 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
546 			   struct ib_recv_wr *wr, u8 *len16)
547 {
548 	int ret;
549 
550 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
551 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
552 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
553 	if (ret)
554 		return ret;
555 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
556 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
557 	return 0;
558 }
559 
560 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
561 			  u8 *len16)
562 {
563 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
564 	wqe->inv.r2 = 0;
565 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
566 	return 0;
567 }
568 
569 static void free_qp_work(struct work_struct *work)
570 {
571 	struct c4iw_ucontext *ucontext;
572 	struct c4iw_qp *qhp;
573 	struct c4iw_dev *rhp;
574 
575 	qhp = container_of(work, struct c4iw_qp, free_work);
576 	ucontext = qhp->ucontext;
577 	rhp = qhp->rhp;
578 
579 	CTR3(KTR_IW_CXGBE, "%s qhp %p ucontext %p", __func__,
580 			qhp, ucontext);
581 	destroy_qp(&rhp->rdev, &qhp->wq,
582 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
583 
584 	if (ucontext)
585 		c4iw_put_ucontext(ucontext);
586 	kfree(qhp);
587 }
588 
589 static void queue_qp_free(struct kref *kref)
590 {
591 	struct c4iw_qp *qhp;
592 
593 	qhp = container_of(kref, struct c4iw_qp, kref);
594 	CTR2(KTR_IW_CXGBE, "%s qhp %p", __func__, qhp);
595 	queue_work(qhp->rhp->rdev.free_workq, &qhp->free_work);
596 }
597 
598 void c4iw_qp_add_ref(struct ib_qp *qp)
599 {
600 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
601 	kref_get(&to_c4iw_qp(qp)->kref);
602 }
603 
604 void c4iw_qp_rem_ref(struct ib_qp *qp)
605 {
606 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
607 	kref_put(&to_c4iw_qp(qp)->kref, queue_qp_free);
608 }
609 
610 static void complete_sq_drain_wr(struct c4iw_qp *qhp, struct ib_send_wr *wr)
611 {
612 	struct t4_cqe cqe = {};
613 	struct c4iw_cq *schp;
614 	unsigned long flag;
615 	struct t4_cq *cq;
616 
617 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
618 	cq = &schp->cq;
619 
620 	PDBG("%s drain sq id %u\n", __func__, qhp->wq.sq.qid);
621 	cqe.u.drain_cookie = wr->wr_id;
622 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
623 				 V_CQE_OPCODE(C4IW_DRAIN_OPCODE) |
624 				 V_CQE_TYPE(1) |
625 				 V_CQE_SWCQE(1) |
626 				 V_CQE_QPID(qhp->wq.sq.qid));
627 
628 	spin_lock_irqsave(&schp->lock, flag);
629 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
630 	cq->sw_queue[cq->sw_pidx] = cqe;
631 	t4_swcq_produce(cq);
632 	spin_unlock_irqrestore(&schp->lock, flag);
633 
634 	spin_lock_irqsave(&schp->comp_handler_lock, flag);
635 	(*schp->ibcq.comp_handler)(&schp->ibcq,
636 				   schp->ibcq.cq_context);
637 	spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
638 }
639 
640 static void complete_rq_drain_wr(struct c4iw_qp *qhp, struct ib_recv_wr *wr)
641 {
642 	struct t4_cqe cqe = {};
643 	struct c4iw_cq *rchp;
644 	unsigned long flag;
645 	struct t4_cq *cq;
646 
647 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
648 	cq = &rchp->cq;
649 
650 	PDBG("%s drain rq id %u\n", __func__, qhp->wq.sq.qid);
651 	cqe.u.drain_cookie = wr->wr_id;
652 	cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) |
653 				 V_CQE_OPCODE(C4IW_DRAIN_OPCODE) |
654 				 V_CQE_TYPE(0) |
655 				 V_CQE_SWCQE(1) |
656 				 V_CQE_QPID(qhp->wq.sq.qid));
657 
658 	spin_lock_irqsave(&rchp->lock, flag);
659 	cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen));
660 	cq->sw_queue[cq->sw_pidx] = cqe;
661 	t4_swcq_produce(cq);
662 	spin_unlock_irqrestore(&rchp->lock, flag);
663 
664 	spin_lock_irqsave(&rchp->comp_handler_lock, flag);
665 	(*rchp->ibcq.comp_handler)(&rchp->ibcq,
666 				   rchp->ibcq.cq_context);
667 	spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
668 }
669 
670 static void build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
671 		struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16)
672 {
673 	__be64 *p = (__be64 *)fr->pbl;
674 
675 	fr->r2 = cpu_to_be32(0);
676 	fr->stag = cpu_to_be32(mhp->ibmr.rkey);
677 
678 	fr->tpte.valid_to_pdid = cpu_to_be32(F_FW_RI_TPTE_VALID |
679 			V_FW_RI_TPTE_STAGKEY((mhp->ibmr.rkey & M_FW_RI_TPTE_STAGKEY)) |
680 			V_FW_RI_TPTE_STAGSTATE(1) |
681 			V_FW_RI_TPTE_STAGTYPE(FW_RI_STAG_NSMR) |
682 			V_FW_RI_TPTE_PDID(mhp->attr.pdid));
683 	fr->tpte.locread_to_qpid = cpu_to_be32(
684 			V_FW_RI_TPTE_PERM(c4iw_ib_to_tpt_access(wr->access)) |
685 			V_FW_RI_TPTE_ADDRTYPE(FW_RI_VA_BASED_TO) |
686 			V_FW_RI_TPTE_PS(ilog2(wr->mr->page_size) - 12));
687 	fr->tpte.nosnoop_pbladdr = cpu_to_be32(V_FW_RI_TPTE_PBLADDR(
688 			      PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
689 	fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
690 	fr->tpte.len_hi = cpu_to_be32(0);
691 	fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length);
692 	fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
693 	fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
694 
695 	p[0] = cpu_to_be64((u64)mhp->mpl[0]);
696 	p[1] = cpu_to_be64((u64)mhp->mpl[1]);
697 
698 	*len16 = DIV_ROUND_UP(sizeof(*fr), 16);
699 }
700 
701 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
702 		struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16,
703 		bool dsgl_supported)
704 {
705 	struct fw_ri_immd *imdp;
706 	__be64 *p;
707 	int i;
708 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
709 	int rem;
710 
711 	if (mhp->mpl_len > t4_max_fr_depth(use_dsgl && dsgl_supported))
712 		return -EINVAL;
713 
714 	wqe->fr.qpbinde_to_dcacpu = 0;
715 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
716 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
717 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
718 	wqe->fr.len_hi = 0;
719 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
720 	wqe->fr.stag = cpu_to_be32(wr->key);
721 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
722 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
723 			0xffffffff);
724 
725 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
726 		struct fw_ri_dsgl *sglp;
727 
728 		for (i = 0; i < mhp->mpl_len; i++)
729 			mhp->mpl[i] =
730 				     (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
731 
732 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
733 		sglp->op = FW_RI_DATA_DSGL;
734 		sglp->r1 = 0;
735 		sglp->nsge = cpu_to_be16(1);
736 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
737 		sglp->len0 = cpu_to_be32(pbllen);
738 
739 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
740 	} else {
741 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
742 		imdp->op = FW_RI_DATA_IMMD;
743 		imdp->r1 = 0;
744 		imdp->r2 = 0;
745 		imdp->immdlen = cpu_to_be32(pbllen);
746 		p = (__be64 *)(imdp + 1);
747 		rem = pbllen;
748 		for (i = 0; i < mhp->mpl_len; i++) {
749 			*p = cpu_to_be64((u64)mhp->mpl[i]);
750 			rem -= sizeof(*p);
751 			if (++p == (__be64 *)&sq->queue[sq->size])
752 				p = (__be64 *)sq->queue;
753 		}
754 		BUG_ON(rem < 0);
755 		while (rem) {
756 			*p = 0;
757 			rem -= sizeof(*p);
758 			if (++p == (__be64 *)&sq->queue[sq->size])
759 				p = (__be64 *)sq->queue;
760 		}
761 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
762 				+ pbllen, 16);
763 	}
764 
765 	return 0;
766 }
767 
768 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
769 		   struct ib_send_wr **bad_wr)
770 {
771 	int err = 0;
772 	u8 len16 = 0;
773 	enum fw_wr_opcodes fw_opcode = 0;
774 	enum fw_ri_wr_flags fw_flags;
775 	struct c4iw_qp *qhp;
776 	union t4_wr *wqe = NULL;
777 	u32 num_wrs;
778 	struct t4_swsqe *swsqe;
779 	unsigned long flag;
780 	u16 idx = 0;
781 	struct c4iw_rdev *rdev;
782 
783 	qhp = to_c4iw_qp(ibqp);
784 	rdev = &qhp->rhp->rdev;
785 	spin_lock_irqsave(&qhp->lock, flag);
786 	if (t4_wq_in_error(&qhp->wq)) {
787 		spin_unlock_irqrestore(&qhp->lock, flag);
788 		complete_sq_drain_wr(qhp, wr);
789 		return err;
790 	}
791 	num_wrs = t4_sq_avail(&qhp->wq);
792 	if (num_wrs == 0) {
793 		spin_unlock_irqrestore(&qhp->lock, flag);
794 		*bad_wr = wr;
795 		return -ENOMEM;
796 	}
797 	while (wr) {
798 		if (num_wrs == 0) {
799 			err = -ENOMEM;
800 			*bad_wr = wr;
801 			break;
802 		}
803 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
804 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
805 
806 		fw_flags = 0;
807 		if (wr->send_flags & IB_SEND_SOLICITED)
808 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
809 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
810 			fw_flags |= FW_RI_COMPLETION_FLAG;
811 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
812 		switch (wr->opcode) {
813 		case IB_WR_SEND_WITH_INV:
814 		case IB_WR_SEND:
815 			if (wr->send_flags & IB_SEND_FENCE)
816 				fw_flags |= FW_RI_READ_FENCE_FLAG;
817 			fw_opcode = FW_RI_SEND_WR;
818 			if (wr->opcode == IB_WR_SEND)
819 				swsqe->opcode = FW_RI_SEND;
820 			else
821 				swsqe->opcode = FW_RI_SEND_WITH_INV;
822 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
823 			break;
824 		case IB_WR_RDMA_WRITE:
825 			fw_opcode = FW_RI_RDMA_WRITE_WR;
826 			swsqe->opcode = FW_RI_RDMA_WRITE;
827 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
828 			break;
829 		case IB_WR_RDMA_READ:
830 		case IB_WR_RDMA_READ_WITH_INV:
831 			fw_opcode = FW_RI_RDMA_READ_WR;
832 			swsqe->opcode = FW_RI_READ_REQ;
833 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
834 				c4iw_invalidate_mr(qhp->rhp,
835 						   wr->sg_list[0].lkey);
836 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
837 			} else {
838 				fw_flags = 0;
839 			}
840 			err = build_rdma_read(wqe, wr, &len16);
841 			if (err)
842 				break;
843 			swsqe->read_len = wr->sg_list[0].length;
844 			if (!qhp->wq.sq.oldest_read)
845 				qhp->wq.sq.oldest_read = swsqe;
846 			break;
847 		case IB_WR_REG_MR: {
848 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
849 
850 			swsqe->opcode = FW_RI_FAST_REGISTER;
851 			if (rdev->adap->params.fr_nsmr_tpte_wr_support &&
852 					!mhp->attr.state && mhp->mpl_len <= 2) {
853 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
854 				build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
855 						mhp, &len16);
856 			} else {
857 				fw_opcode = FW_RI_FR_NSMR_WR;
858 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
859 					mhp, &len16,
860 					rdev->adap->params.ulptx_memwrite_dsgl);
861 				if (err)
862 					break;
863 			}
864 			mhp->attr.state = 1;
865 			break;
866 		}
867 		case IB_WR_LOCAL_INV:
868 			if (wr->send_flags & IB_SEND_FENCE)
869 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
870 			fw_opcode = FW_RI_INV_LSTAG_WR;
871 			swsqe->opcode = FW_RI_LOCAL_INV;
872 			err = build_inv_stag(wqe, wr, &len16);
873 			c4iw_invalidate_mr(qhp->rhp, wr->ex.invalidate_rkey);
874 			break;
875 		default:
876 			CTR2(KTR_IW_CXGBE, "%s post of type =%d TBD!", __func__,
877 			     wr->opcode);
878 			err = -EINVAL;
879 		}
880 		if (err) {
881 			*bad_wr = wr;
882 			break;
883 		}
884 		swsqe->idx = qhp->wq.sq.pidx;
885 		swsqe->complete = 0;
886 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
887 					qhp->sq_sig_all;
888 		swsqe->flushed = 0;
889 		swsqe->wr_id = wr->wr_id;
890 
891 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
892 
893 		CTR5(KTR_IW_CXGBE,
894 		    "%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u",
895 		    __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
896 		    swsqe->opcode, swsqe->read_len);
897 		wr = wr->next;
898 		num_wrs--;
899 		t4_sq_produce(&qhp->wq, len16);
900 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
901 	}
902 
903 	t4_ring_sq_db(&qhp->wq, idx, wqe, rdev->adap->iwt.wc_en);
904 	spin_unlock_irqrestore(&qhp->lock, flag);
905 	return err;
906 }
907 
908 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
909 		      struct ib_recv_wr **bad_wr)
910 {
911 	int err = 0;
912 	struct c4iw_qp *qhp;
913 	union t4_recv_wr *wqe = NULL;
914 	u32 num_wrs;
915 	u8 len16 = 0;
916 	unsigned long flag;
917 	u16 idx = 0;
918 
919 	qhp = to_c4iw_qp(ibqp);
920 	spin_lock_irqsave(&qhp->lock, flag);
921 	if (t4_wq_in_error(&qhp->wq)) {
922 		spin_unlock_irqrestore(&qhp->lock, flag);
923 		complete_rq_drain_wr(qhp, wr);
924 		return err;
925 	}
926 	num_wrs = t4_rq_avail(&qhp->wq);
927 	if (num_wrs == 0) {
928 		spin_unlock_irqrestore(&qhp->lock, flag);
929 		*bad_wr = wr;
930 		return -ENOMEM;
931 	}
932 	while (wr) {
933 		if (wr->num_sge > T4_MAX_RECV_SGE) {
934 			err = -EINVAL;
935 			*bad_wr = wr;
936 			break;
937 		}
938 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
939 					   qhp->wq.rq.wq_pidx *
940 					   T4_EQ_ENTRY_SIZE);
941 		if (num_wrs)
942 			err = build_rdma_recv(qhp, wqe, wr, &len16);
943 		else
944 			err = -ENOMEM;
945 		if (err) {
946 			*bad_wr = wr;
947 			break;
948 		}
949 
950 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
951 
952 		wqe->recv.opcode = FW_RI_RECV_WR;
953 		wqe->recv.r1 = 0;
954 		wqe->recv.wrid = qhp->wq.rq.pidx;
955 		wqe->recv.r2[0] = 0;
956 		wqe->recv.r2[1] = 0;
957 		wqe->recv.r2[2] = 0;
958 		wqe->recv.len16 = len16;
959 		CTR3(KTR_IW_CXGBE, "%s cookie 0x%llx pidx %u", __func__,
960 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
961 		t4_rq_produce(&qhp->wq, len16);
962 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
963 		wr = wr->next;
964 		num_wrs--;
965 	}
966 
967 	t4_ring_rq_db(&qhp->wq, idx, wqe, qhp->rhp->rdev.adap->iwt.wc_en);
968 	spin_unlock_irqrestore(&qhp->lock, flag);
969 	return err;
970 }
971 
972 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
973 				    u8 *ecode)
974 {
975 	int status;
976 	int tagged;
977 	int opcode;
978 	int rqtype;
979 	int send_inv;
980 
981 	if (!err_cqe) {
982 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
983 		*ecode = 0;
984 		return;
985 	}
986 
987 	status = CQE_STATUS(err_cqe);
988 	opcode = CQE_OPCODE(err_cqe);
989 	rqtype = RQ_TYPE(err_cqe);
990 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
991 		   (opcode == FW_RI_SEND_WITH_SE_INV);
992 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
993 		 (rqtype && (opcode == FW_RI_READ_RESP));
994 
995 	switch (status) {
996 	case T4_ERR_STAG:
997 		if (send_inv) {
998 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
999 			*ecode = RDMAP_CANT_INV_STAG;
1000 		} else {
1001 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1002 			*ecode = RDMAP_INV_STAG;
1003 		}
1004 		break;
1005 	case T4_ERR_PDID:
1006 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1007 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1008 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1009 			*ecode = RDMAP_CANT_INV_STAG;
1010 		else
1011 			*ecode = RDMAP_STAG_NOT_ASSOC;
1012 		break;
1013 	case T4_ERR_QPID:
1014 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1015 		*ecode = RDMAP_STAG_NOT_ASSOC;
1016 		break;
1017 	case T4_ERR_ACCESS:
1018 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1019 		*ecode = RDMAP_ACC_VIOL;
1020 		break;
1021 	case T4_ERR_WRAP:
1022 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1023 		*ecode = RDMAP_TO_WRAP;
1024 		break;
1025 	case T4_ERR_BOUND:
1026 		if (tagged) {
1027 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1028 			*ecode = DDPT_BASE_BOUNDS;
1029 		} else {
1030 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1031 			*ecode = RDMAP_BASE_BOUNDS;
1032 		}
1033 		break;
1034 	case T4_ERR_INVALIDATE_SHARED_MR:
1035 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1036 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1037 		*ecode = RDMAP_CANT_INV_STAG;
1038 		break;
1039 	case T4_ERR_ECC:
1040 	case T4_ERR_ECC_PSTAG:
1041 	case T4_ERR_INTERNAL_ERR:
1042 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1043 		*ecode = 0;
1044 		break;
1045 	case T4_ERR_OUT_OF_RQE:
1046 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1047 		*ecode = DDPU_INV_MSN_NOBUF;
1048 		break;
1049 	case T4_ERR_PBL_ADDR_BOUND:
1050 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1051 		*ecode = DDPT_BASE_BOUNDS;
1052 		break;
1053 	case T4_ERR_CRC:
1054 		*layer_type = LAYER_MPA|DDP_LLP;
1055 		*ecode = MPA_CRC_ERR;
1056 		break;
1057 	case T4_ERR_MARKER:
1058 		*layer_type = LAYER_MPA|DDP_LLP;
1059 		*ecode = MPA_MARKER_ERR;
1060 		break;
1061 	case T4_ERR_PDU_LEN_ERR:
1062 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1063 		*ecode = DDPU_MSG_TOOBIG;
1064 		break;
1065 	case T4_ERR_DDP_VERSION:
1066 		if (tagged) {
1067 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1068 			*ecode = DDPT_INV_VERS;
1069 		} else {
1070 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1071 			*ecode = DDPU_INV_VERS;
1072 		}
1073 		break;
1074 	case T4_ERR_RDMA_VERSION:
1075 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1076 		*ecode = RDMAP_INV_VERS;
1077 		break;
1078 	case T4_ERR_OPCODE:
1079 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1080 		*ecode = RDMAP_INV_OPCODE;
1081 		break;
1082 	case T4_ERR_DDP_QUEUE_NUM:
1083 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1084 		*ecode = DDPU_INV_QN;
1085 		break;
1086 	case T4_ERR_MSN:
1087 	case T4_ERR_MSN_GAP:
1088 	case T4_ERR_MSN_RANGE:
1089 	case T4_ERR_IRD_OVERFLOW:
1090 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1091 		*ecode = DDPU_INV_MSN_RANGE;
1092 		break;
1093 	case T4_ERR_TBIT:
1094 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1095 		*ecode = 0;
1096 		break;
1097 	case T4_ERR_MO:
1098 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1099 		*ecode = DDPU_INV_MO;
1100 		break;
1101 	default:
1102 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1103 		*ecode = 0;
1104 		break;
1105 	}
1106 }
1107 
1108 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1109 			   gfp_t gfp)
1110 {
1111 	int ret;
1112 	struct fw_ri_wr *wqe;
1113 	struct terminate_message *term;
1114 	struct wrqe *wr;
1115 	struct socket *so = qhp->ep->com.so;
1116         struct inpcb *inp = sotoinpcb(so);
1117         struct tcpcb *tp = intotcpcb(inp);
1118         struct toepcb *toep = tp->t_toe;
1119 
1120 	CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1121 	    qhp->wq.sq.qid, qhp->ep->hwtid);
1122 
1123 	wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
1124 	if (wr == NULL)
1125 		return;
1126         wqe = wrtod(wr);
1127 
1128 	memset(wqe, 0, sizeof *wqe);
1129 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR));
1130 	wqe->flowid_len16 = cpu_to_be32(
1131 		V_FW_WR_FLOWID(qhp->ep->hwtid) |
1132 		V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1133 
1134 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1135 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1136 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1137 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1138 		term->layer_etype = qhp->attr.layer_etype;
1139 		term->ecode = qhp->attr.ecode;
1140 	} else
1141 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1142 	ret = creds(toep, inp, sizeof(*wqe));
1143 	if (ret) {
1144 		free_wrqe(wr);
1145 		return;
1146 	}
1147 	t4_wrq_tx(qhp->rhp->rdev.adap, wr);
1148 }
1149 
1150 /* Assumes qhp lock is held. */
1151 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1152 		       struct c4iw_cq *schp)
1153 {
1154 	int count;
1155 	int rq_flushed, sq_flushed;
1156 	unsigned long flag;
1157 
1158 	CTR4(KTR_IW_CXGBE, "%s qhp %p rchp %p schp %p", __func__, qhp, rchp,
1159 	    schp);
1160 
1161 	/* locking hierarchy: cq lock first, then qp lock. */
1162 	spin_lock_irqsave(&rchp->lock, flag);
1163 	spin_lock(&qhp->lock);
1164 
1165 	if (qhp->wq.flushed) {
1166 		spin_unlock(&qhp->lock);
1167 		spin_unlock_irqrestore(&rchp->lock, flag);
1168 		return;
1169 	}
1170 	qhp->wq.flushed = 1;
1171 
1172 	c4iw_flush_hw_cq(rchp);
1173 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1174 	rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1175 	spin_unlock(&qhp->lock);
1176 	spin_unlock_irqrestore(&rchp->lock, flag);
1177 
1178 	/* locking hierarchy: cq lock first, then qp lock. */
1179 	spin_lock_irqsave(&schp->lock, flag);
1180 	spin_lock(&qhp->lock);
1181 	if (schp != rchp)
1182 		c4iw_flush_hw_cq(schp);
1183 	sq_flushed = c4iw_flush_sq(qhp);
1184 	spin_unlock(&qhp->lock);
1185 	spin_unlock_irqrestore(&schp->lock, flag);
1186 
1187 	if (schp == rchp) {
1188 		if (t4_clear_cq_armed(&rchp->cq) &&
1189 		    (rq_flushed || sq_flushed)) {
1190 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1191 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1192 						   rchp->ibcq.cq_context);
1193 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1194 		}
1195 	} else {
1196 		if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1197 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1198 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1199 						   rchp->ibcq.cq_context);
1200 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1201 		}
1202 		if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1203 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1204 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1205 						   schp->ibcq.cq_context);
1206 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1207 		}
1208 	}
1209 }
1210 
1211 static void flush_qp(struct c4iw_qp *qhp)
1212 {
1213 	struct c4iw_cq *rchp, *schp;
1214 	unsigned long flag;
1215 
1216 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1217 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1218 
1219 	t4_set_wq_in_error(&qhp->wq);
1220 	if (qhp->ibqp.uobject) {
1221 		t4_set_cq_in_error(&rchp->cq);
1222 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1223 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1224 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1225 		if (schp != rchp) {
1226 			t4_set_cq_in_error(&schp->cq);
1227 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1228 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1229 					schp->ibcq.cq_context);
1230 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1231 		}
1232 		return;
1233 	}
1234 	__flush_qp(qhp, rchp, schp);
1235 }
1236 
1237 static int
1238 rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp, struct c4iw_ep *ep)
1239 {
1240 	struct c4iw_rdev *rdev = &rhp->rdev;
1241 	struct adapter *sc = rdev->adap;
1242 	struct fw_ri_wr *wqe;
1243 	int ret;
1244 	struct wrqe *wr;
1245 	struct socket *so = ep->com.so;
1246         struct inpcb *inp = sotoinpcb(so);
1247         struct tcpcb *tp = intotcpcb(inp);
1248         struct toepcb *toep = tp->t_toe;
1249 
1250 	KASSERT(rhp == qhp->rhp && ep == qhp->ep, ("%s: EDOOFUS", __func__));
1251 
1252 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1253 	    qhp->wq.sq.qid, ep, ep->hwtid);
1254 
1255 	wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
1256 	if (wr == NULL)
1257 		return (0);
1258 	wqe = wrtod(wr);
1259 
1260 	memset(wqe, 0, sizeof *wqe);
1261 
1262 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL);
1263 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1264 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1265 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1266 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1267 
1268 	c4iw_init_wr_wait(&ep->com.wr_wait);
1269 
1270 	ret = creds(toep, inp, sizeof(*wqe));
1271 	if (ret) {
1272 		free_wrqe(wr);
1273 		return ret;
1274 	}
1275 	t4_wrq_tx(sc, wr);
1276 
1277 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1278 			qhp->wq.sq.qid, ep->com.so, __func__);
1279 	return ret;
1280 }
1281 
1282 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1283 {
1284 	CTR2(KTR_IW_CXGBE, "%s p2p_type = %d", __func__, p2p_type);
1285 	memset(&init->u, 0, sizeof init->u);
1286 	switch (p2p_type) {
1287 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1288 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1289 		init->u.write.stag_sink = cpu_to_be32(1);
1290 		init->u.write.to_sink = cpu_to_be64(1);
1291 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1292 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1293 						   sizeof(struct fw_ri_immd),
1294 						   16);
1295 		break;
1296 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1297 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1298 		init->u.read.stag_src = cpu_to_be32(1);
1299 		init->u.read.to_src_lo = cpu_to_be32(1);
1300 		init->u.read.stag_sink = cpu_to_be32(1);
1301 		init->u.read.to_sink_lo = cpu_to_be32(1);
1302 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1303 		break;
1304 	}
1305 }
1306 
1307 static int
1308 creds(struct toepcb *toep, struct inpcb *inp, size_t wrsize)
1309 {
1310 	struct ofld_tx_sdesc *txsd;
1311 
1312 	CTR3(KTR_IW_CXGBE, "%s:creB  %p %u", __func__, toep , wrsize);
1313 	INP_WLOCK(inp);
1314 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) {
1315 		INP_WUNLOCK(inp);
1316 		return (EINVAL);
1317 	}
1318 	txsd = &toep->txsd[toep->txsd_pidx];
1319 	txsd->tx_credits = howmany(wrsize, 16);
1320 	txsd->plen = 0;
1321 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
1322 			("%s: not enough credits (%d)", __func__, toep->tx_credits));
1323 	toep->tx_credits -= txsd->tx_credits;
1324 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1325 		toep->txsd_pidx = 0;
1326 	toep->txsd_avail--;
1327 	INP_WUNLOCK(inp);
1328 	CTR5(KTR_IW_CXGBE, "%s:creE  %p %u %u %u", __func__, toep ,
1329 	    txsd->tx_credits, toep->tx_credits, toep->txsd_pidx);
1330 	return (0);
1331 }
1332 
1333 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1334 {
1335 	struct fw_ri_wr *wqe;
1336 	int ret;
1337 	struct wrqe *wr;
1338 	struct c4iw_ep *ep = qhp->ep;
1339 	struct c4iw_rdev *rdev = &qhp->rhp->rdev;
1340 	struct adapter *sc = rdev->adap;
1341 	struct socket *so = ep->com.so;
1342         struct inpcb *inp = sotoinpcb(so);
1343         struct tcpcb *tp = intotcpcb(inp);
1344         struct toepcb *toep = tp->t_toe;
1345 
1346 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1347 	    qhp->wq.sq.qid, ep, ep->hwtid);
1348 
1349 	wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
1350 	if (wr == NULL)
1351 		return (0);
1352 	wqe = wrtod(wr);
1353 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1354 	if (ret) {
1355 		qhp->attr.max_ird = 0;
1356 		free_wrqe(wr);
1357 		return ret;
1358 	}
1359 
1360 	memset(wqe, 0, sizeof *wqe);
1361 
1362 	wqe->op_compl = cpu_to_be32(
1363 		V_FW_WR_OP(FW_RI_WR) |
1364 		F_FW_WR_COMPL);
1365 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1366 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1367 
1368 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1369 
1370 	wqe->u.init.type = FW_RI_TYPE_INIT;
1371 	wqe->u.init.mpareqbit_p2ptype =
1372 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1373 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1374 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1375 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1376 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1377 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1378 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1379 	if (qhp->attr.mpa_attr.crc_enabled)
1380 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1381 
1382 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1383 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1384 			    FW_RI_QP_BIND_ENABLE;
1385 	if (!qhp->ibqp.uobject)
1386 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1387 				     FW_RI_QP_STAG0_ENABLE;
1388 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1389 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1390 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1391 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1392 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1393 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1394 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1395 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1396 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1397 	wqe->u.init.iss = cpu_to_be32(ep->snd_seq);
1398 	wqe->u.init.irs = cpu_to_be32(ep->rcv_seq);
1399 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1400 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1401 	    sc->vres.rq.start);
1402 	if (qhp->attr.mpa_attr.initiator)
1403 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1404 
1405 	c4iw_init_wr_wait(&ep->com.wr_wait);
1406 
1407 	ret = creds(toep, inp, sizeof(*wqe));
1408 	if (ret) {
1409 		free_wrqe(wr);
1410 		free_ird(rhp, qhp->attr.max_ird);
1411 		return ret;
1412 	}
1413 	t4_wrq_tx(sc, wr);
1414 
1415 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1416 			qhp->wq.sq.qid, ep->com.so, __func__);
1417 
1418 	toep->ulp_mode = ULP_MODE_RDMA;
1419 	free_ird(rhp, qhp->attr.max_ird);
1420 
1421 	return ret;
1422 }
1423 
1424 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1425 		   enum c4iw_qp_attr_mask mask,
1426 		   struct c4iw_qp_attributes *attrs,
1427 		   int internal)
1428 {
1429 	int ret = 0;
1430 	struct c4iw_qp_attributes newattr = qhp->attr;
1431 	int disconnect = 0;
1432 	int terminate = 0;
1433 	int abort = 0;
1434 	int free = 0;
1435 	struct c4iw_ep *ep = NULL;
1436 
1437 	CTR5(KTR_IW_CXGBE, "%s qhp %p sqid 0x%x rqid 0x%x ep %p", __func__, qhp,
1438 	    qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep);
1439 	CTR3(KTR_IW_CXGBE, "%s state %d -> %d", __func__, qhp->attr.state,
1440 	    (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1441 
1442 	mutex_lock(&qhp->mutex);
1443 
1444 	/* Process attr changes if in IDLE */
1445 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1446 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1447 			ret = -EIO;
1448 			goto out;
1449 		}
1450 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1451 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1452 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1453 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1454 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1455 			newattr.enable_bind = attrs->enable_bind;
1456 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1457 			if (attrs->max_ord > c4iw_max_read_depth) {
1458 				ret = -EINVAL;
1459 				goto out;
1460 			}
1461 			newattr.max_ord = attrs->max_ord;
1462 		}
1463 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1464 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1465 				ret = -EINVAL;
1466 				goto out;
1467 			}
1468 			newattr.max_ird = attrs->max_ird;
1469 		}
1470 		qhp->attr = newattr;
1471 	}
1472 
1473 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1474 		goto out;
1475 	if (qhp->attr.state == attrs->next_state)
1476 		goto out;
1477 
1478 	/* Return EINPROGRESS if QP is already in transition state.
1479 	 * Eg: CLOSING->IDLE transition or *->ERROR transition.
1480 	 * This can happen while connection is switching(due to rdma_fini)
1481 	 * from iWARP/RDDP to TOE mode and any inflight RDMA RX data will
1482 	 * reach TOE driver -> TCP stack -> iWARP driver. In this way
1483 	 * iWARP driver keep receiving inflight RDMA RX data until socket
1484 	 * is closed or aborted. And if iWARP CM is in FPDU sate, then
1485 	 * it tries to put QP in TERM state and disconnects endpoint.
1486 	 * But as QP is already in transition state, this event is ignored.
1487 	 */
1488 	if ((qhp->attr.state >= C4IW_QP_STATE_ERROR) &&
1489 		(attrs->next_state == C4IW_QP_STATE_TERMINATE)) {
1490 		ret = -EINPROGRESS;
1491 		goto out;
1492 	}
1493 
1494 	switch (qhp->attr.state) {
1495 	case C4IW_QP_STATE_IDLE:
1496 		switch (attrs->next_state) {
1497 		case C4IW_QP_STATE_RTS:
1498 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1499 				ret = -EINVAL;
1500 				goto out;
1501 			}
1502 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1503 				ret = -EINVAL;
1504 				goto out;
1505 			}
1506 			qhp->attr.mpa_attr = attrs->mpa_attr;
1507 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1508 			qhp->ep = qhp->attr.llp_stream_handle;
1509 			set_state(qhp, C4IW_QP_STATE_RTS);
1510 
1511 			/*
1512 			 * Ref the endpoint here and deref when we
1513 			 * disassociate the endpoint from the QP.  This
1514 			 * happens in CLOSING->IDLE transition or *->ERROR
1515 			 * transition.
1516 			 */
1517 			c4iw_get_ep(&qhp->ep->com);
1518 			ret = rdma_init(rhp, qhp);
1519 			if (ret)
1520 				goto err;
1521 			break;
1522 		case C4IW_QP_STATE_ERROR:
1523 			set_state(qhp, C4IW_QP_STATE_ERROR);
1524 			flush_qp(qhp);
1525 			break;
1526 		default:
1527 			ret = -EINVAL;
1528 			goto out;
1529 		}
1530 		break;
1531 	case C4IW_QP_STATE_RTS:
1532 		switch (attrs->next_state) {
1533 		case C4IW_QP_STATE_CLOSING:
1534 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1535 			t4_set_wq_in_error(&qhp->wq);
1536 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1537 			ep = qhp->ep;
1538 			if (!internal) {
1539 				abort = 0;
1540 				disconnect = 1;
1541 				c4iw_get_ep(&qhp->ep->com);
1542 			}
1543 			ret = rdma_fini(rhp, qhp, ep);
1544 			if (ret)
1545 				goto err;
1546 			break;
1547 		case C4IW_QP_STATE_TERMINATE:
1548 			t4_set_wq_in_error(&qhp->wq);
1549 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1550 			qhp->attr.layer_etype = attrs->layer_etype;
1551 			qhp->attr.ecode = attrs->ecode;
1552 			ep = qhp->ep;
1553 			if (!internal) {
1554 				c4iw_get_ep(&qhp->ep->com);
1555 				terminate = 1;
1556 				disconnect = 1;
1557 			} else {
1558 				terminate = qhp->attr.send_term;
1559 				ret = rdma_fini(rhp, qhp, ep);
1560 				if (ret)
1561 					goto err;
1562 			}
1563 			break;
1564 		case C4IW_QP_STATE_ERROR:
1565 			t4_set_wq_in_error(&qhp->wq);
1566 			set_state(qhp, C4IW_QP_STATE_ERROR);
1567 			if (!internal) {
1568 				abort = 1;
1569 				disconnect = 1;
1570 				ep = qhp->ep;
1571 				c4iw_get_ep(&qhp->ep->com);
1572 			}
1573 			goto err;
1574 			break;
1575 		default:
1576 			ret = -EINVAL;
1577 			goto out;
1578 		}
1579 		break;
1580 	case C4IW_QP_STATE_CLOSING:
1581 
1582 		/*
1583 		 * Allow kernel users to move to ERROR for qp draining.
1584 		 */
1585 		if (!internal && (qhp->ibqp.uobject || attrs->next_state !=
1586 				  C4IW_QP_STATE_ERROR)) {
1587 			ret = -EINVAL;
1588 			goto out;
1589 		}
1590 		switch (attrs->next_state) {
1591 		case C4IW_QP_STATE_IDLE:
1592 			flush_qp(qhp);
1593 			set_state(qhp, C4IW_QP_STATE_IDLE);
1594 			qhp->attr.llp_stream_handle = NULL;
1595 			c4iw_put_ep(&qhp->ep->com);
1596 			qhp->ep = NULL;
1597 			wake_up(&qhp->wait);
1598 			break;
1599 		case C4IW_QP_STATE_ERROR:
1600 			goto err;
1601 		default:
1602 			ret = -EINVAL;
1603 			goto err;
1604 		}
1605 		break;
1606 	case C4IW_QP_STATE_ERROR:
1607 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1608 			ret = -EINVAL;
1609 			goto out;
1610 		}
1611 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1612 			ret = -EINVAL;
1613 			goto out;
1614 		}
1615 		set_state(qhp, C4IW_QP_STATE_IDLE);
1616 		break;
1617 	case C4IW_QP_STATE_TERMINATE:
1618 		if (!internal) {
1619 			ret = -EINVAL;
1620 			goto out;
1621 		}
1622 		goto err;
1623 		break;
1624 	default:
1625 		printf("%s in a bad state %d\n",
1626 		       __func__, qhp->attr.state);
1627 		ret = -EINVAL;
1628 		goto err;
1629 		break;
1630 	}
1631 	goto out;
1632 err:
1633 	CTR3(KTR_IW_CXGBE, "%s disassociating ep %p qpid 0x%x", __func__,
1634 	    qhp->ep, qhp->wq.sq.qid);
1635 
1636 	/* disassociate the LLP connection */
1637 	qhp->attr.llp_stream_handle = NULL;
1638 	if (!ep)
1639 		ep = qhp->ep;
1640 	qhp->ep = NULL;
1641 	set_state(qhp, C4IW_QP_STATE_ERROR);
1642 	free = 1;
1643 	abort = 1;
1644 	BUG_ON(!ep);
1645 	flush_qp(qhp);
1646 	wake_up(&qhp->wait);
1647 out:
1648 	mutex_unlock(&qhp->mutex);
1649 
1650 	if (terminate)
1651 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1652 
1653 	/*
1654 	 * If disconnect is 1, then we need to initiate a disconnect
1655 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1656 	 * an abnormal close (RTS/CLOSING->ERROR).
1657 	 */
1658 	if (disconnect) {
1659 		__c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1660 							 GFP_KERNEL);
1661 		c4iw_put_ep(&ep->com);
1662 	}
1663 
1664 	/*
1665 	 * If free is 1, then we've disassociated the EP from the QP
1666 	 * and we need to dereference the EP.
1667 	 */
1668 	if (free)
1669 		c4iw_put_ep(&ep->com);
1670 	CTR2(KTR_IW_CXGBE, "%s exit state %d", __func__, qhp->attr.state);
1671 	return ret;
1672 }
1673 
1674 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1675 {
1676 	struct c4iw_dev *rhp;
1677 	struct c4iw_qp *qhp;
1678 	struct c4iw_qp_attributes attrs;
1679 
1680 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ib_qp);
1681 	qhp = to_c4iw_qp(ib_qp);
1682 	rhp = qhp->rhp;
1683 
1684 	attrs.next_state = C4IW_QP_STATE_ERROR;
1685 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1686 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1687 	else
1688 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1689 	wait_event(qhp->wait, !qhp->ep);
1690 
1691 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1692 
1693 	free_ird(rhp, qhp->attr.max_ird);
1694 	c4iw_qp_rem_ref(ib_qp);
1695 
1696 	CTR3(KTR_IW_CXGBE, "%s ib_qp %p qpid 0x%0x", __func__, ib_qp,
1697 	    qhp->wq.sq.qid);
1698 	return 0;
1699 }
1700 
1701 struct ib_qp *
1702 c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1703     struct ib_udata *udata)
1704 {
1705 	struct c4iw_dev *rhp;
1706 	struct c4iw_qp *qhp;
1707 	struct c4iw_pd *php;
1708 	struct c4iw_cq *schp;
1709 	struct c4iw_cq *rchp;
1710 	struct c4iw_create_qp_resp uresp;
1711 	unsigned int sqsize, rqsize;
1712 	struct c4iw_ucontext *ucontext;
1713 	int ret;
1714 	struct c4iw_mm_entry *sq_key_mm = NULL, *rq_key_mm = NULL;
1715 	struct c4iw_mm_entry *sq_db_key_mm = NULL, *rq_db_key_mm = NULL;
1716 
1717 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
1718 
1719 	if (attrs->qp_type != IB_QPT_RC)
1720 		return ERR_PTR(-EINVAL);
1721 
1722 	php = to_c4iw_pd(pd);
1723 	rhp = php->rhp;
1724 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1725 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1726 	if (!schp || !rchp)
1727 		return ERR_PTR(-EINVAL);
1728 
1729 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1730 		return ERR_PTR(-EINVAL);
1731 
1732 	if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1733 		return ERR_PTR(-E2BIG);
1734 	rqsize = attrs->cap.max_recv_wr + 1;
1735 	if (rqsize < 8)
1736 		rqsize = 8;
1737 
1738 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1739 		return ERR_PTR(-E2BIG);
1740 	sqsize = attrs->cap.max_send_wr + 1;
1741 	if (sqsize < 8)
1742 		sqsize = 8;
1743 
1744 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1745 
1746 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1747 	if (!qhp)
1748 		return ERR_PTR(-ENOMEM);
1749 	qhp->wq.sq.size = sqsize;
1750 	qhp->wq.sq.memsize =
1751 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1752 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1753 	qhp->wq.sq.flush_cidx = -1;
1754 	qhp->wq.rq.size = rqsize;
1755 	qhp->wq.rq.memsize =
1756 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1757 		sizeof(*qhp->wq.rq.queue);
1758 
1759 	if (ucontext) {
1760 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1761 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1762 	}
1763 
1764 	CTR5(KTR_IW_CXGBE, "%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu",
1765 	    __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1766 
1767 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1768 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1769 	if (ret)
1770 		goto err1;
1771 
1772 	attrs->cap.max_recv_wr = rqsize - 1;
1773 	attrs->cap.max_send_wr = sqsize - 1;
1774 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1775 
1776 	qhp->rhp = rhp;
1777 	qhp->attr.pd = php->pdid;
1778 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1779 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1780 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1781 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1782 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1783 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1784 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1785 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1786 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1787 	qhp->attr.enable_rdma_read = 1;
1788 	qhp->attr.enable_rdma_write = 1;
1789 	qhp->attr.enable_bind = 1;
1790 	qhp->attr.max_ord = 0;
1791 	qhp->attr.max_ird = 0;
1792 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1793 	spin_lock_init(&qhp->lock);
1794 	mutex_init(&qhp->mutex);
1795 	init_waitqueue_head(&qhp->wait);
1796 	kref_init(&qhp->kref);
1797 	INIT_WORK(&qhp->free_work, free_qp_work);
1798 
1799 	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1800 	if (ret)
1801 		goto err2;
1802 
1803 	if (udata) {
1804 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
1805 		if (!sq_key_mm) {
1806 			ret = -ENOMEM;
1807 			goto err3;
1808 		}
1809 		rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
1810 		if (!rq_key_mm) {
1811 			ret = -ENOMEM;
1812 			goto err4;
1813 		}
1814 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
1815 		if (!sq_db_key_mm) {
1816 			ret = -ENOMEM;
1817 			goto err5;
1818 		}
1819 		rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
1820 		if (!rq_db_key_mm) {
1821 			ret = -ENOMEM;
1822 			goto err6;
1823 		}
1824 		uresp.flags = 0;
1825 		uresp.qid_mask = rhp->rdev.qpmask;
1826 		uresp.sqid = qhp->wq.sq.qid;
1827 		uresp.sq_size = qhp->wq.sq.size;
1828 		uresp.sq_memsize = qhp->wq.sq.memsize;
1829 		uresp.rqid = qhp->wq.rq.qid;
1830 		uresp.rq_size = qhp->wq.rq.size;
1831 		uresp.rq_memsize = qhp->wq.rq.memsize;
1832 		spin_lock(&ucontext->mmap_lock);
1833 		uresp.ma_sync_key =  0;
1834 		uresp.sq_key = ucontext->key;
1835 		ucontext->key += PAGE_SIZE;
1836 		uresp.rq_key = ucontext->key;
1837 		ucontext->key += PAGE_SIZE;
1838 		uresp.sq_db_gts_key = ucontext->key;
1839 		ucontext->key += PAGE_SIZE;
1840 		uresp.rq_db_gts_key = ucontext->key;
1841 		ucontext->key += PAGE_SIZE;
1842 		spin_unlock(&ucontext->mmap_lock);
1843 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1844 		if (ret)
1845 			goto err7;
1846 		sq_key_mm->key = uresp.sq_key;
1847 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
1848 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1849 		CTR4(KTR_IW_CXGBE, "%s sq_key_mm %x, %x, %d", __func__,
1850 				sq_key_mm->key, sq_key_mm->addr,
1851 				sq_key_mm->len);
1852 		insert_mmap(ucontext, sq_key_mm);
1853 		rq_key_mm->key = uresp.rq_key;
1854 		rq_key_mm->addr = qhp->wq.rq.phys_addr;
1855 		rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1856 		CTR4(KTR_IW_CXGBE, "%s rq_key_mm %x, %x, %d", __func__,
1857 				rq_key_mm->key, rq_key_mm->addr,
1858 				rq_key_mm->len);
1859 		insert_mmap(ucontext, rq_key_mm);
1860 		sq_db_key_mm->key = uresp.sq_db_gts_key;
1861 		sq_db_key_mm->addr = (u64)qhp->wq.sq.bar2_pa;
1862 		sq_db_key_mm->len = PAGE_SIZE;
1863 		CTR4(KTR_IW_CXGBE, "%s sq_db_key_mm %x, %x, %d", __func__,
1864 				sq_db_key_mm->key, sq_db_key_mm->addr,
1865 				sq_db_key_mm->len);
1866 		insert_mmap(ucontext, sq_db_key_mm);
1867 		rq_db_key_mm->key = uresp.rq_db_gts_key;
1868 		rq_db_key_mm->addr = (u64)qhp->wq.rq.bar2_pa;
1869 		rq_db_key_mm->len = PAGE_SIZE;
1870 		CTR4(KTR_IW_CXGBE, "%s rq_db_key_mm %x, %x, %d", __func__,
1871 				rq_db_key_mm->key, rq_db_key_mm->addr,
1872 				rq_db_key_mm->len);
1873 		insert_mmap(ucontext, rq_db_key_mm);
1874 
1875 		c4iw_get_ucontext(ucontext);
1876 		qhp->ucontext = ucontext;
1877 	}
1878 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1879 	init_timer(&(qhp->timer));
1880 
1881 	CTR5(KTR_IW_CXGBE, "%s sq id %u size %u memsize %zu num_entries %u",
1882 		 __func__, qhp->wq.sq.qid,
1883 		 qhp->wq.sq.size, qhp->wq.sq.memsize, attrs->cap.max_send_wr);
1884 	CTR5(KTR_IW_CXGBE, "%s rq id %u size %u memsize %zu num_entries %u",
1885 		 __func__, qhp->wq.rq.qid,
1886 		 qhp->wq.rq.size, qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1887 	return &qhp->ibqp;
1888 err7:
1889 	kfree(rq_db_key_mm);
1890 err6:
1891 	kfree(sq_db_key_mm);
1892 err5:
1893 	kfree(rq_key_mm);
1894 err4:
1895 	kfree(sq_key_mm);
1896 err3:
1897 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1898 err2:
1899 	destroy_qp(&rhp->rdev, &qhp->wq,
1900 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1901 err1:
1902 	kfree(qhp);
1903 	return ERR_PTR(ret);
1904 }
1905 
1906 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1907 		      int attr_mask, struct ib_udata *udata)
1908 {
1909 	struct c4iw_dev *rhp;
1910 	struct c4iw_qp *qhp;
1911 	enum c4iw_qp_attr_mask mask = 0;
1912 	struct c4iw_qp_attributes attrs;
1913 
1914 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ibqp);
1915 
1916 	/* iwarp does not support the RTR state */
1917 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1918 		attr_mask &= ~IB_QP_STATE;
1919 
1920 	/* Make sure we still have something left to do */
1921 	if (!attr_mask)
1922 		return 0;
1923 
1924 	memset(&attrs, 0, sizeof attrs);
1925 	qhp = to_c4iw_qp(ibqp);
1926 	rhp = qhp->rhp;
1927 
1928 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1929 	attrs.enable_rdma_read = (attr->qp_access_flags &
1930 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1931 	attrs.enable_rdma_write = (attr->qp_access_flags &
1932 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1933 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1934 
1935 
1936 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1937 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1938 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1939 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1940 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1941 
1942 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1943 }
1944 
1945 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1946 {
1947 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p qpn 0x%x", __func__, dev, qpn);
1948 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1949 }
1950 
1951 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1952 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1953 {
1954 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1955 
1956 	memset(attr, 0, sizeof *attr);
1957 	memset(init_attr, 0, sizeof *init_attr);
1958 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1959 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1960 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1961 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1962 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1963 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1964 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1965 	return 0;
1966 }
1967 #endif
1968