xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/qp.c (revision d7d962ead0b6e5e8a39202d0590022082bf5bfb6)
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(chip_id(sc) <= CHELSIO_T5 ?
270 		    X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
271 		V_FW_RI_RES_WR_FBMAX(3) |
272 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
273 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
274 		V_FW_RI_RES_WR_EQSIZE(eqsize));
275 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
276 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
277 	res++;
278 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
279 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
280 
281 	/* eqsize is the number of 64B entries plus the status page size. */
282 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
283 			rdev->hw_queue.t4_eq_status_entries;
284 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
285 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
286 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
287 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
288 		V_FW_RI_RES_WR_IQID(rcq->cqid));
289 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
290 		V_FW_RI_RES_WR_DCAEN(0) |
291 		V_FW_RI_RES_WR_DCACPU(0) |
292 		V_FW_RI_RES_WR_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
293 		    X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
294 		V_FW_RI_RES_WR_FBMAX(3) |
295 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
296 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
297 		V_FW_RI_RES_WR_EQSIZE(eqsize));
298 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
299 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
300 
301 	c4iw_init_wr_wait(&wr_wait);
302 
303 	t4_wrq_tx(sc, wr);
304 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid,
305 			NULL, __func__);
306 	if (ret)
307 		goto free_rq_dma;
308 
309 	CTR5(KTR_IW_CXGBE,
310 	    "%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx",
311 	    __func__, wq->sq.qid, wq->rq.qid,
312 	    (unsigned long long)wq->sq.bar2_va,
313 	    (unsigned long long)wq->rq.bar2_va);
314 
315 	return 0;
316 free_rq_dma:
317 	dma_free_coherent(rhp->ibdev.dma_device,
318 			  wq->rq.memsize, wq->rq.queue,
319 			  dma_unmap_addr(&wq->rq, mapping));
320 free_sq_dma:
321 	dma_free_coherent(rhp->ibdev.dma_device,
322 			  wq->sq.memsize, wq->sq.queue,
323 			  dma_unmap_addr(&wq->sq, mapping));
324 free_hwaddr:
325 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
326 free_sw_rq:
327 	kfree(wq->rq.sw_rq);
328 free_sw_sq:
329 	kfree(wq->sq.sw_sq);
330 free_rq_qid:
331 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
332 free_sq_qid:
333 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
334 	return ret;
335 }
336 
337 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
338 		      const struct ib_send_wr *wr, int max, u32 *plenp)
339 {
340 	u8 *dstp, *srcp;
341 	u32 plen = 0;
342 	int i;
343 	int rem, len;
344 
345 	dstp = (u8 *)immdp->data;
346 	for (i = 0; i < wr->num_sge; i++) {
347 		if ((plen + wr->sg_list[i].length) > max)
348 			return -EMSGSIZE;
349 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
350 		plen += wr->sg_list[i].length;
351 		rem = wr->sg_list[i].length;
352 		while (rem) {
353 			if (dstp == (u8 *)&sq->queue[sq->size])
354 				dstp = (u8 *)sq->queue;
355 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
356 				len = rem;
357 			else
358 				len = (u8 *)&sq->queue[sq->size] - dstp;
359 			memcpy(dstp, srcp, len);
360 			dstp += len;
361 			srcp += len;
362 			rem -= len;
363 		}
364 	}
365 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
366 	if (len)
367 		memset(dstp, 0, len);
368 	immdp->op = FW_RI_DATA_IMMD;
369 	immdp->r1 = 0;
370 	immdp->r2 = 0;
371 	immdp->immdlen = cpu_to_be32(plen);
372 	*plenp = plen;
373 	return 0;
374 }
375 
376 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
377 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
378 		      int num_sge, u32 *plenp)
379 
380 {
381 	int i;
382 	u32 plen = 0;
383 	__be64 *flitp = (__be64 *)isglp->sge;
384 
385 	for (i = 0; i < num_sge; i++) {
386 		if ((plen + sg_list[i].length) < plen)
387 			return -EMSGSIZE;
388 		plen += sg_list[i].length;
389 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
390 				     sg_list[i].length);
391 		if (++flitp == queue_end)
392 			flitp = queue_start;
393 		*flitp = cpu_to_be64(sg_list[i].addr);
394 		if (++flitp == queue_end)
395 			flitp = queue_start;
396 	}
397 	*flitp = (__force __be64)0;
398 	isglp->op = FW_RI_DATA_ISGL;
399 	isglp->r1 = 0;
400 	isglp->nsge = cpu_to_be16(num_sge);
401 	isglp->r2 = 0;
402 	if (plenp)
403 		*plenp = plen;
404 	return 0;
405 }
406 
407 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
408 			   const struct ib_send_wr *wr, u8 *len16)
409 {
410 	u32 plen;
411 	int size;
412 	int ret;
413 
414 	if (wr->num_sge > T4_MAX_SEND_SGE)
415 		return -EINVAL;
416 	switch (wr->opcode) {
417 	case IB_WR_SEND:
418 		if (wr->send_flags & IB_SEND_SOLICITED)
419 			wqe->send.sendop_pkd = cpu_to_be32(
420 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
421 		else
422 			wqe->send.sendop_pkd = cpu_to_be32(
423 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
424 		wqe->send.stag_inv = 0;
425 		break;
426 	case IB_WR_SEND_WITH_INV:
427 		if (wr->send_flags & IB_SEND_SOLICITED)
428 			wqe->send.sendop_pkd = cpu_to_be32(
429 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
430 		else
431 			wqe->send.sendop_pkd = cpu_to_be32(
432 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
433 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
434 		break;
435 
436 	default:
437 		return -EINVAL;
438 	}
439 	wqe->send.r3 = 0;
440 	wqe->send.r4 = 0;
441 
442 	plen = 0;
443 	if (wr->num_sge) {
444 		if (wr->send_flags & IB_SEND_INLINE) {
445 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
446 					 T4_MAX_SEND_INLINE, &plen);
447 			if (ret)
448 				return ret;
449 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
450 			       plen;
451 		} else {
452 			ret = build_isgl((__be64 *)sq->queue,
453 					 (__be64 *)&sq->queue[sq->size],
454 					 wqe->send.u.isgl_src,
455 					 wr->sg_list, wr->num_sge, &plen);
456 			if (ret)
457 				return ret;
458 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
459 			       wr->num_sge * sizeof(struct fw_ri_sge);
460 		}
461 	} else {
462 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
463 		wqe->send.u.immd_src[0].r1 = 0;
464 		wqe->send.u.immd_src[0].r2 = 0;
465 		wqe->send.u.immd_src[0].immdlen = 0;
466 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
467 		plen = 0;
468 	}
469 	*len16 = DIV_ROUND_UP(size, 16);
470 	wqe->send.plen = cpu_to_be32(plen);
471 	return 0;
472 }
473 
474 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
475 			    const struct ib_send_wr *wr, u8 *len16)
476 {
477 	u32 plen;
478 	int size;
479 	int ret;
480 
481 	if (wr->num_sge > T4_MAX_SEND_SGE)
482 		return -EINVAL;
483 	wqe->write.immd_data = 0;
484 	wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
485 	wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
486 	if (wr->num_sge) {
487 		if (wr->send_flags & IB_SEND_INLINE) {
488 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
489 					 T4_MAX_WRITE_INLINE, &plen);
490 			if (ret)
491 				return ret;
492 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
493 			       plen;
494 		} else {
495 			ret = build_isgl((__be64 *)sq->queue,
496 					 (__be64 *)&sq->queue[sq->size],
497 					 wqe->write.u.isgl_src,
498 					 wr->sg_list, wr->num_sge, &plen);
499 			if (ret)
500 				return ret;
501 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
502 			       wr->num_sge * sizeof(struct fw_ri_sge);
503 		}
504 	} else {
505 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
506 		wqe->write.u.immd_src[0].r1 = 0;
507 		wqe->write.u.immd_src[0].r2 = 0;
508 		wqe->write.u.immd_src[0].immdlen = 0;
509 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
510 		plen = 0;
511 	}
512 	*len16 = DIV_ROUND_UP(size, 16);
513 	wqe->write.plen = cpu_to_be32(plen);
514 	return 0;
515 }
516 
517 static int build_rdma_read(union t4_wr *wqe, const struct ib_send_wr *wr, u8 *len16)
518 {
519 	if (wr->num_sge > 1)
520 		return -EINVAL;
521 	if (wr->num_sge && wr->sg_list[0].length) {
522 		wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
523 		wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
524 							>> 32));
525 		wqe->read.to_src_lo =
526 			cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
527 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
528 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
529 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
530 							 >> 32));
531 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
532 	} else {
533 		wqe->read.stag_src = cpu_to_be32(2);
534 		wqe->read.to_src_hi = 0;
535 		wqe->read.to_src_lo = 0;
536 		wqe->read.stag_sink = cpu_to_be32(2);
537 		wqe->read.plen = 0;
538 		wqe->read.to_sink_hi = 0;
539 		wqe->read.to_sink_lo = 0;
540 	}
541 	wqe->read.r2 = 0;
542 	wqe->read.r5 = 0;
543 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
544 	return 0;
545 }
546 
547 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
548 			   const struct ib_recv_wr *wr, u8 *len16)
549 {
550 	int ret;
551 
552 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
553 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
554 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
555 	if (ret)
556 		return ret;
557 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
558 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
559 	return 0;
560 }
561 
562 static int build_inv_stag(union t4_wr *wqe, const struct ib_send_wr *wr,
563 			  u8 *len16)
564 {
565 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
566 	wqe->inv.r2 = 0;
567 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
568 	return 0;
569 }
570 
571 static void free_qp_work(struct work_struct *work)
572 {
573 	struct c4iw_ucontext *ucontext;
574 	struct c4iw_qp *qhp;
575 	struct c4iw_dev *rhp;
576 
577 	qhp = container_of(work, struct c4iw_qp, free_work);
578 	ucontext = qhp->ucontext;
579 	rhp = qhp->rhp;
580 
581 	CTR3(KTR_IW_CXGBE, "%s qhp %p ucontext %p", __func__,
582 			qhp, ucontext);
583 	destroy_qp(&rhp->rdev, &qhp->wq,
584 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
585 
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, const 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, const 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 int build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr,
671 		const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16)
672 {
673 	__be64 *p = (__be64 *)fr->pbl;
674 
675 	if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE)
676 		return -EINVAL;
677 
678 	fr->r2 = cpu_to_be32(0);
679 	fr->stag = cpu_to_be32(mhp->ibmr.rkey);
680 
681 	fr->tpte.valid_to_pdid = cpu_to_be32(F_FW_RI_TPTE_VALID |
682 			V_FW_RI_TPTE_STAGKEY((mhp->ibmr.rkey & M_FW_RI_TPTE_STAGKEY)) |
683 			V_FW_RI_TPTE_STAGSTATE(1) |
684 			V_FW_RI_TPTE_STAGTYPE(FW_RI_STAG_NSMR) |
685 			V_FW_RI_TPTE_PDID(mhp->attr.pdid));
686 	fr->tpte.locread_to_qpid = cpu_to_be32(
687 			V_FW_RI_TPTE_PERM(c4iw_ib_to_tpt_access(wr->access)) |
688 			V_FW_RI_TPTE_ADDRTYPE(FW_RI_VA_BASED_TO) |
689 			V_FW_RI_TPTE_PS(ilog2(wr->mr->page_size) - 12));
690 	fr->tpte.nosnoop_pbladdr = cpu_to_be32(V_FW_RI_TPTE_PBLADDR(
691 			      PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3));
692 	fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0);
693 	fr->tpte.len_hi = cpu_to_be32(mhp->ibmr.length >> 32);
694 	fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff);
695 	fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
696 	fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
697 
698 	p[0] = cpu_to_be64((u64)mhp->mpl[0]);
699 	p[1] = cpu_to_be64((u64)mhp->mpl[1]);
700 
701 	*len16 = DIV_ROUND_UP(sizeof(*fr), 16);
702 	return 0;
703 }
704 
705 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
706 		const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16,
707 		bool dsgl_supported)
708 {
709 	struct fw_ri_immd *imdp;
710 	__be64 *p;
711 	int i;
712 	int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
713 	int rem;
714 
715 	if (mhp->mpl_len > t4_max_fr_depth(&mhp->rhp->rdev, use_dsgl))
716 		return -EINVAL;
717 	if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE)
718 		return -EINVAL;
719 
720 	wqe->fr.qpbinde_to_dcacpu = 0;
721 	wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
722 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
723 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
724 	wqe->fr.len_hi = cpu_to_be32(mhp->ibmr.length >> 32);
725 	wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff);
726 	wqe->fr.stag = cpu_to_be32(wr->key);
727 	wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
728 	wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff);
729 
730 	if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) {
731 		struct fw_ri_dsgl *sglp;
732 
733 		for (i = 0; i < mhp->mpl_len; i++)
734 			mhp->mpl[i] =
735 				     (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
736 
737 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
738 		sglp->op = FW_RI_DATA_DSGL;
739 		sglp->r1 = 0;
740 		sglp->nsge = cpu_to_be16(1);
741 		sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
742 		sglp->len0 = cpu_to_be32(pbllen);
743 
744 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
745 	} else {
746 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
747 		imdp->op = FW_RI_DATA_IMMD;
748 		imdp->r1 = 0;
749 		imdp->r2 = 0;
750 		imdp->immdlen = cpu_to_be32(pbllen);
751 		p = (__be64 *)(imdp + 1);
752 		rem = pbllen;
753 		for (i = 0; i < mhp->mpl_len; i++) {
754 			*p = cpu_to_be64((u64)mhp->mpl[i]);
755 			rem -= sizeof(*p);
756 			if (++p == (__be64 *)&sq->queue[sq->size])
757 				p = (__be64 *)sq->queue;
758 		}
759 		BUG_ON(rem < 0);
760 		while (rem) {
761 			*p = 0;
762 			rem -= sizeof(*p);
763 			if (++p == (__be64 *)&sq->queue[sq->size])
764 				p = (__be64 *)sq->queue;
765 		}
766 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
767 				+ pbllen, 16);
768 	}
769 
770 	return 0;
771 }
772 
773 int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
774 		   const struct ib_send_wr **bad_wr)
775 {
776 	int err = 0;
777 	u8 len16 = 0;
778 	enum fw_wr_opcodes fw_opcode = 0;
779 	enum fw_ri_wr_flags fw_flags;
780 	struct c4iw_qp *qhp;
781 	union t4_wr *wqe = NULL;
782 	u32 num_wrs;
783 	struct t4_swsqe *swsqe;
784 	unsigned long flag;
785 	u16 idx = 0;
786 	struct c4iw_rdev *rdev;
787 
788 	qhp = to_c4iw_qp(ibqp);
789 	rdev = &qhp->rhp->rdev;
790 	spin_lock_irqsave(&qhp->lock, flag);
791 	if (t4_wq_in_error(&qhp->wq)) {
792 		spin_unlock_irqrestore(&qhp->lock, flag);
793 		complete_sq_drain_wr(qhp, wr);
794 		return err;
795 	}
796 	num_wrs = t4_sq_avail(&qhp->wq);
797 	if (num_wrs == 0) {
798 		spin_unlock_irqrestore(&qhp->lock, flag);
799 		*bad_wr = wr;
800 		return -ENOMEM;
801 	}
802 	while (wr) {
803 		if (num_wrs == 0) {
804 			err = -ENOMEM;
805 			*bad_wr = wr;
806 			break;
807 		}
808 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
809 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
810 
811 		fw_flags = 0;
812 		if (wr->send_flags & IB_SEND_SOLICITED)
813 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
814 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
815 			fw_flags |= FW_RI_COMPLETION_FLAG;
816 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
817 		switch (wr->opcode) {
818 		case IB_WR_SEND_WITH_INV:
819 		case IB_WR_SEND:
820 			if (wr->send_flags & IB_SEND_FENCE)
821 				fw_flags |= FW_RI_READ_FENCE_FLAG;
822 			fw_opcode = FW_RI_SEND_WR;
823 			if (wr->opcode == IB_WR_SEND)
824 				swsqe->opcode = FW_RI_SEND;
825 			else
826 				swsqe->opcode = FW_RI_SEND_WITH_INV;
827 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
828 			break;
829 		case IB_WR_RDMA_WRITE:
830 			fw_opcode = FW_RI_RDMA_WRITE_WR;
831 			swsqe->opcode = FW_RI_RDMA_WRITE;
832 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
833 			break;
834 		case IB_WR_RDMA_READ:
835 		case IB_WR_RDMA_READ_WITH_INV:
836 			fw_opcode = FW_RI_RDMA_READ_WR;
837 			swsqe->opcode = FW_RI_READ_REQ;
838 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) {
839 				c4iw_invalidate_mr(qhp->rhp,
840 						   wr->sg_list[0].lkey);
841 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
842 			} else {
843 				fw_flags = 0;
844 			}
845 			err = build_rdma_read(wqe, wr, &len16);
846 			if (err)
847 				break;
848 			swsqe->read_len = wr->sg_list[0].length;
849 			if (!qhp->wq.sq.oldest_read)
850 				qhp->wq.sq.oldest_read = swsqe;
851 			break;
852 		case IB_WR_REG_MR: {
853 			struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr);
854 
855 			swsqe->opcode = FW_RI_FAST_REGISTER;
856 			if (rdev->adap->params.fr_nsmr_tpte_wr_support &&
857 					!mhp->attr.state && mhp->mpl_len <= 2) {
858 				fw_opcode = FW_RI_FR_NSMR_TPTE_WR;
859 				err = build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr),
860 						mhp, &len16);
861 			} else {
862 				fw_opcode = FW_RI_FR_NSMR_WR;
863 				err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr),
864 					mhp, &len16,
865 					rdev->adap->params.ulptx_memwrite_dsgl);
866 			}
867 			if (err)
868 				break;
869 			mhp->attr.state = 1;
870 			break;
871 		}
872 		case IB_WR_LOCAL_INV:
873 			if (wr->send_flags & IB_SEND_FENCE)
874 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
875 			fw_opcode = FW_RI_INV_LSTAG_WR;
876 			swsqe->opcode = FW_RI_LOCAL_INV;
877 			err = build_inv_stag(wqe, wr, &len16);
878 			c4iw_invalidate_mr(qhp->rhp, wr->ex.invalidate_rkey);
879 			break;
880 		default:
881 			CTR2(KTR_IW_CXGBE, "%s post of type =%d TBD!", __func__,
882 			     wr->opcode);
883 			err = -EINVAL;
884 		}
885 		if (err) {
886 			*bad_wr = wr;
887 			break;
888 		}
889 		swsqe->idx = qhp->wq.sq.pidx;
890 		swsqe->complete = 0;
891 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
892 					qhp->sq_sig_all;
893 		swsqe->flushed = 0;
894 		swsqe->wr_id = wr->wr_id;
895 
896 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
897 
898 		CTR5(KTR_IW_CXGBE,
899 		    "%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u",
900 		    __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
901 		    swsqe->opcode, swsqe->read_len);
902 		wr = wr->next;
903 		num_wrs--;
904 		t4_sq_produce(&qhp->wq, len16);
905 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
906 	}
907 
908 	t4_ring_sq_db(&qhp->wq, idx, wqe, rdev->adap->iwt.wc_en);
909 	spin_unlock_irqrestore(&qhp->lock, flag);
910 	return err;
911 }
912 
913 int c4iw_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
914 		      const struct ib_recv_wr **bad_wr)
915 {
916 	int err = 0;
917 	struct c4iw_qp *qhp;
918 	union t4_recv_wr *wqe = NULL;
919 	u32 num_wrs;
920 	u8 len16 = 0;
921 	unsigned long flag;
922 	u16 idx = 0;
923 
924 	qhp = to_c4iw_qp(ibqp);
925 	spin_lock_irqsave(&qhp->lock, flag);
926 	if (t4_wq_in_error(&qhp->wq)) {
927 		spin_unlock_irqrestore(&qhp->lock, flag);
928 		complete_rq_drain_wr(qhp, wr);
929 		return err;
930 	}
931 	num_wrs = t4_rq_avail(&qhp->wq);
932 	if (num_wrs == 0) {
933 		spin_unlock_irqrestore(&qhp->lock, flag);
934 		*bad_wr = wr;
935 		return -ENOMEM;
936 	}
937 	while (wr) {
938 		if (wr->num_sge > T4_MAX_RECV_SGE) {
939 			err = -EINVAL;
940 			*bad_wr = wr;
941 			break;
942 		}
943 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
944 					   qhp->wq.rq.wq_pidx *
945 					   T4_EQ_ENTRY_SIZE);
946 		if (num_wrs)
947 			err = build_rdma_recv(qhp, wqe, wr, &len16);
948 		else
949 			err = -ENOMEM;
950 		if (err) {
951 			*bad_wr = wr;
952 			break;
953 		}
954 
955 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
956 
957 		wqe->recv.opcode = FW_RI_RECV_WR;
958 		wqe->recv.r1 = 0;
959 		wqe->recv.wrid = qhp->wq.rq.pidx;
960 		wqe->recv.r2[0] = 0;
961 		wqe->recv.r2[1] = 0;
962 		wqe->recv.r2[2] = 0;
963 		wqe->recv.len16 = len16;
964 		CTR3(KTR_IW_CXGBE, "%s cookie 0x%llx pidx %u", __func__,
965 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
966 		t4_rq_produce(&qhp->wq, len16);
967 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
968 		wr = wr->next;
969 		num_wrs--;
970 	}
971 
972 	t4_ring_rq_db(&qhp->wq, idx, wqe, qhp->rhp->rdev.adap->iwt.wc_en);
973 	spin_unlock_irqrestore(&qhp->lock, flag);
974 	return err;
975 }
976 
977 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
978 				    u8 *ecode)
979 {
980 	int status;
981 	int tagged;
982 	int opcode;
983 	int rqtype;
984 	int send_inv;
985 
986 	if (!err_cqe) {
987 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
988 		*ecode = 0;
989 		return;
990 	}
991 
992 	status = CQE_STATUS(err_cqe);
993 	opcode = CQE_OPCODE(err_cqe);
994 	rqtype = RQ_TYPE(err_cqe);
995 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
996 		   (opcode == FW_RI_SEND_WITH_SE_INV);
997 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
998 		 (rqtype && (opcode == FW_RI_READ_RESP));
999 
1000 	switch (status) {
1001 	case T4_ERR_STAG:
1002 		if (send_inv) {
1003 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1004 			*ecode = RDMAP_CANT_INV_STAG;
1005 		} else {
1006 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1007 			*ecode = RDMAP_INV_STAG;
1008 		}
1009 		break;
1010 	case T4_ERR_PDID:
1011 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1012 		if ((opcode == FW_RI_SEND_WITH_INV) ||
1013 		    (opcode == FW_RI_SEND_WITH_SE_INV))
1014 			*ecode = RDMAP_CANT_INV_STAG;
1015 		else
1016 			*ecode = RDMAP_STAG_NOT_ASSOC;
1017 		break;
1018 	case T4_ERR_QPID:
1019 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1020 		*ecode = RDMAP_STAG_NOT_ASSOC;
1021 		break;
1022 	case T4_ERR_ACCESS:
1023 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1024 		*ecode = RDMAP_ACC_VIOL;
1025 		break;
1026 	case T4_ERR_WRAP:
1027 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1028 		*ecode = RDMAP_TO_WRAP;
1029 		break;
1030 	case T4_ERR_BOUND:
1031 		if (tagged) {
1032 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1033 			*ecode = DDPT_BASE_BOUNDS;
1034 		} else {
1035 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1036 			*ecode = RDMAP_BASE_BOUNDS;
1037 		}
1038 		break;
1039 	case T4_ERR_INVALIDATE_SHARED_MR:
1040 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1041 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1042 		*ecode = RDMAP_CANT_INV_STAG;
1043 		break;
1044 	case T4_ERR_ECC:
1045 	case T4_ERR_ECC_PSTAG:
1046 	case T4_ERR_INTERNAL_ERR:
1047 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1048 		*ecode = 0;
1049 		break;
1050 	case T4_ERR_OUT_OF_RQE:
1051 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1052 		*ecode = DDPU_INV_MSN_NOBUF;
1053 		break;
1054 	case T4_ERR_PBL_ADDR_BOUND:
1055 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1056 		*ecode = DDPT_BASE_BOUNDS;
1057 		break;
1058 	case T4_ERR_CRC:
1059 		*layer_type = LAYER_MPA|DDP_LLP;
1060 		*ecode = MPA_CRC_ERR;
1061 		break;
1062 	case T4_ERR_MARKER:
1063 		*layer_type = LAYER_MPA|DDP_LLP;
1064 		*ecode = MPA_MARKER_ERR;
1065 		break;
1066 	case T4_ERR_PDU_LEN_ERR:
1067 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1068 		*ecode = DDPU_MSG_TOOBIG;
1069 		break;
1070 	case T4_ERR_DDP_VERSION:
1071 		if (tagged) {
1072 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1073 			*ecode = DDPT_INV_VERS;
1074 		} else {
1075 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1076 			*ecode = DDPU_INV_VERS;
1077 		}
1078 		break;
1079 	case T4_ERR_RDMA_VERSION:
1080 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1081 		*ecode = RDMAP_INV_VERS;
1082 		break;
1083 	case T4_ERR_OPCODE:
1084 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1085 		*ecode = RDMAP_INV_OPCODE;
1086 		break;
1087 	case T4_ERR_DDP_QUEUE_NUM:
1088 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1089 		*ecode = DDPU_INV_QN;
1090 		break;
1091 	case T4_ERR_MSN:
1092 	case T4_ERR_MSN_GAP:
1093 	case T4_ERR_MSN_RANGE:
1094 	case T4_ERR_IRD_OVERFLOW:
1095 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1096 		*ecode = DDPU_INV_MSN_RANGE;
1097 		break;
1098 	case T4_ERR_TBIT:
1099 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1100 		*ecode = 0;
1101 		break;
1102 	case T4_ERR_MO:
1103 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1104 		*ecode = DDPU_INV_MO;
1105 		break;
1106 	default:
1107 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1108 		*ecode = 0;
1109 		break;
1110 	}
1111 }
1112 
1113 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1114 			   gfp_t gfp)
1115 {
1116 	int ret;
1117 	struct fw_ri_wr *wqe;
1118 	struct terminate_message *term;
1119 	struct wrqe *wr;
1120 	struct socket *so = qhp->ep->com.so;
1121         struct inpcb *inp = sotoinpcb(so);
1122         struct tcpcb *tp = intotcpcb(inp);
1123         struct toepcb *toep = tp->t_toe;
1124 
1125 	CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1126 	    qhp->wq.sq.qid, qhp->ep->hwtid);
1127 
1128 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1129 	if (wr == NULL)
1130 		return;
1131         wqe = wrtod(wr);
1132 
1133 	memset(wqe, 0, sizeof *wqe);
1134 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR));
1135 	wqe->flowid_len16 = cpu_to_be32(
1136 		V_FW_WR_FLOWID(qhp->ep->hwtid) |
1137 		V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1138 
1139 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1140 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1141 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1142 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1143 		term->layer_etype = qhp->attr.layer_etype;
1144 		term->ecode = qhp->attr.ecode;
1145 	} else
1146 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1147 	ret = creds(toep, inp, sizeof(*wqe));
1148 	if (ret) {
1149 		free_wrqe(wr);
1150 		return;
1151 	}
1152 	t4_wrq_tx(qhp->rhp->rdev.adap, wr);
1153 }
1154 
1155 /* Assumes qhp lock is held. */
1156 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1157 		       struct c4iw_cq *schp)
1158 {
1159 	int count;
1160 	int rq_flushed, sq_flushed;
1161 	unsigned long flag;
1162 
1163 	CTR4(KTR_IW_CXGBE, "%s qhp %p rchp %p schp %p", __func__, qhp, rchp,
1164 	    schp);
1165 
1166 	/* locking hierarchy: cq lock first, then qp lock. */
1167 	spin_lock_irqsave(&rchp->lock, flag);
1168 	spin_lock(&qhp->lock);
1169 
1170 	if (qhp->wq.flushed) {
1171 		spin_unlock(&qhp->lock);
1172 		spin_unlock_irqrestore(&rchp->lock, flag);
1173 		return;
1174 	}
1175 	qhp->wq.flushed = 1;
1176 
1177 	c4iw_flush_hw_cq(rchp);
1178 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1179 	rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1180 	spin_unlock(&qhp->lock);
1181 	spin_unlock_irqrestore(&rchp->lock, flag);
1182 
1183 	/* locking hierarchy: cq lock first, then qp lock. */
1184 	spin_lock_irqsave(&schp->lock, flag);
1185 	spin_lock(&qhp->lock);
1186 	if (schp != rchp)
1187 		c4iw_flush_hw_cq(schp);
1188 	sq_flushed = c4iw_flush_sq(qhp);
1189 	spin_unlock(&qhp->lock);
1190 	spin_unlock_irqrestore(&schp->lock, flag);
1191 
1192 	if (schp == rchp) {
1193 		if (t4_clear_cq_armed(&rchp->cq) &&
1194 		    (rq_flushed || sq_flushed)) {
1195 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1196 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1197 						   rchp->ibcq.cq_context);
1198 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1199 		}
1200 	} else {
1201 		if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1202 			spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1203 			(*rchp->ibcq.comp_handler)(&rchp->ibcq,
1204 						   rchp->ibcq.cq_context);
1205 			spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1206 		}
1207 		if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1208 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1209 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1210 						   schp->ibcq.cq_context);
1211 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1212 		}
1213 	}
1214 }
1215 
1216 static void flush_qp(struct c4iw_qp *qhp)
1217 {
1218 	struct c4iw_cq *rchp, *schp;
1219 	unsigned long flag;
1220 
1221 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1222 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1223 
1224 	t4_set_wq_in_error(&qhp->wq);
1225 	if (qhp->ibqp.uobject) {
1226 		t4_set_cq_in_error(&rchp->cq);
1227 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1228 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1229 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1230 		if (schp != rchp) {
1231 			t4_set_cq_in_error(&schp->cq);
1232 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1233 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1234 					schp->ibcq.cq_context);
1235 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1236 		}
1237 		return;
1238 	}
1239 	__flush_qp(qhp, rchp, schp);
1240 }
1241 
1242 static int
1243 rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp, struct c4iw_ep *ep)
1244 {
1245 	struct c4iw_rdev *rdev = &rhp->rdev;
1246 	struct adapter *sc = rdev->adap;
1247 	struct fw_ri_wr *wqe;
1248 	int ret;
1249 	struct wrqe *wr;
1250 	struct socket *so = ep->com.so;
1251         struct inpcb *inp = sotoinpcb(so);
1252         struct tcpcb *tp = intotcpcb(inp);
1253         struct toepcb *toep = tp->t_toe;
1254 
1255 	KASSERT(rhp == qhp->rhp && ep == qhp->ep, ("%s: EDOOFUS", __func__));
1256 
1257 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1258 	    qhp->wq.sq.qid, ep, ep->hwtid);
1259 
1260 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1261 	if (wr == NULL)
1262 		return (0);
1263 	wqe = wrtod(wr);
1264 
1265 	memset(wqe, 0, sizeof *wqe);
1266 
1267 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL);
1268 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1269 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1270 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1271 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1272 
1273 	c4iw_init_wr_wait(&ep->com.wr_wait);
1274 
1275 	ret = creds(toep, inp, sizeof(*wqe));
1276 	if (ret) {
1277 		free_wrqe(wr);
1278 		return ret;
1279 	}
1280 	t4_wrq_tx(sc, wr);
1281 
1282 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1283 			qhp->wq.sq.qid, ep->com.so, __func__);
1284 	return ret;
1285 }
1286 
1287 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1288 {
1289 	CTR2(KTR_IW_CXGBE, "%s p2p_type = %d", __func__, p2p_type);
1290 	memset(&init->u, 0, sizeof init->u);
1291 	switch (p2p_type) {
1292 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1293 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1294 		init->u.write.stag_sink = cpu_to_be32(1);
1295 		init->u.write.to_sink = cpu_to_be64(1);
1296 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1297 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1298 						   sizeof(struct fw_ri_immd),
1299 						   16);
1300 		break;
1301 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1302 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1303 		init->u.read.stag_src = cpu_to_be32(1);
1304 		init->u.read.to_src_lo = cpu_to_be32(1);
1305 		init->u.read.stag_sink = cpu_to_be32(1);
1306 		init->u.read.to_sink_lo = cpu_to_be32(1);
1307 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1308 		break;
1309 	}
1310 }
1311 
1312 static int
1313 creds(struct toepcb *toep, struct inpcb *inp, size_t wrsize)
1314 {
1315 	struct ofld_tx_sdesc *txsd;
1316 
1317 	CTR3(KTR_IW_CXGBE, "%s:creB  %p %u", __func__, toep , wrsize);
1318 	INP_WLOCK(inp);
1319 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) {
1320 		INP_WUNLOCK(inp);
1321 		return (EINVAL);
1322 	}
1323 	txsd = &toep->txsd[toep->txsd_pidx];
1324 	txsd->tx_credits = howmany(wrsize, 16);
1325 	txsd->plen = 0;
1326 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
1327 			("%s: not enough credits (%d)", __func__, toep->tx_credits));
1328 	toep->tx_credits -= txsd->tx_credits;
1329 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1330 		toep->txsd_pidx = 0;
1331 	toep->txsd_avail--;
1332 	INP_WUNLOCK(inp);
1333 	CTR5(KTR_IW_CXGBE, "%s:creE  %p %u %u %u", __func__, toep ,
1334 	    txsd->tx_credits, toep->tx_credits, toep->txsd_pidx);
1335 	return (0);
1336 }
1337 
1338 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1339 {
1340 	struct fw_ri_wr *wqe;
1341 	int ret;
1342 	struct wrqe *wr;
1343 	struct c4iw_ep *ep = qhp->ep;
1344 	struct c4iw_rdev *rdev = &qhp->rhp->rdev;
1345 	struct adapter *sc = rdev->adap;
1346 	struct socket *so = ep->com.so;
1347         struct inpcb *inp = sotoinpcb(so);
1348         struct tcpcb *tp = intotcpcb(inp);
1349         struct toepcb *toep = tp->t_toe;
1350 
1351 	CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp,
1352 	    qhp->wq.sq.qid, ep, ep->hwtid);
1353 
1354 	wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq);
1355 	if (wr == NULL)
1356 		return (0);
1357 	wqe = wrtod(wr);
1358 	ret = alloc_ird(rhp, qhp->attr.max_ird);
1359 	if (ret) {
1360 		qhp->attr.max_ird = 0;
1361 		free_wrqe(wr);
1362 		return ret;
1363 	}
1364 
1365 	memset(wqe, 0, sizeof *wqe);
1366 
1367 	wqe->op_compl = cpu_to_be32(
1368 		V_FW_WR_OP(FW_RI_WR) |
1369 		F_FW_WR_COMPL);
1370 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1371 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1372 
1373 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1374 
1375 	wqe->u.init.type = FW_RI_TYPE_INIT;
1376 	wqe->u.init.mpareqbit_p2ptype =
1377 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1378 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1379 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1380 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1381 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1382 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1383 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1384 	if (qhp->attr.mpa_attr.crc_enabled)
1385 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1386 
1387 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1388 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1389 			    FW_RI_QP_BIND_ENABLE;
1390 	if (!qhp->ibqp.uobject)
1391 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1392 				     FW_RI_QP_STAG0_ENABLE;
1393 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1394 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1395 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1396 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1397 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1398 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1399 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1400 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1401 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1402 	wqe->u.init.iss = cpu_to_be32(ep->snd_seq);
1403 	wqe->u.init.irs = cpu_to_be32(ep->rcv_seq);
1404 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1405 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1406 	    sc->vres.rq.start);
1407 	if (qhp->attr.mpa_attr.initiator)
1408 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1409 
1410 	c4iw_init_wr_wait(&ep->com.wr_wait);
1411 
1412 	ret = creds(toep, inp, sizeof(*wqe));
1413 	if (ret) {
1414 		free_wrqe(wr);
1415 		free_ird(rhp, qhp->attr.max_ird);
1416 		return ret;
1417 	}
1418 	t4_wrq_tx(sc, wr);
1419 
1420 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1421 			qhp->wq.sq.qid, ep->com.so, __func__);
1422 
1423 	toep->params.ulp_mode = ULP_MODE_RDMA;
1424 	free_ird(rhp, qhp->attr.max_ird);
1425 
1426 	return ret;
1427 }
1428 
1429 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1430 		   enum c4iw_qp_attr_mask mask,
1431 		   struct c4iw_qp_attributes *attrs,
1432 		   int internal)
1433 {
1434 	int ret = 0;
1435 	struct c4iw_qp_attributes newattr = qhp->attr;
1436 	int disconnect = 0;
1437 	int terminate = 0;
1438 	int abort = 0;
1439 	int free = 0;
1440 	struct c4iw_ep *ep = NULL;
1441 
1442 	CTR5(KTR_IW_CXGBE, "%s qhp %p sqid 0x%x rqid 0x%x ep %p", __func__, qhp,
1443 	    qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep);
1444 	CTR3(KTR_IW_CXGBE, "%s state %d -> %d", __func__, qhp->attr.state,
1445 	    (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1446 
1447 	mutex_lock(&qhp->mutex);
1448 
1449 	/* Process attr changes if in IDLE */
1450 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1451 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1452 			ret = -EIO;
1453 			goto out;
1454 		}
1455 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1456 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1457 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1458 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1459 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1460 			newattr.enable_bind = attrs->enable_bind;
1461 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1462 			if (attrs->max_ord > c4iw_max_read_depth) {
1463 				ret = -EINVAL;
1464 				goto out;
1465 			}
1466 			newattr.max_ord = attrs->max_ord;
1467 		}
1468 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1469 			if (attrs->max_ird > cur_max_read_depth(rhp)) {
1470 				ret = -EINVAL;
1471 				goto out;
1472 			}
1473 			newattr.max_ird = attrs->max_ird;
1474 		}
1475 		qhp->attr = newattr;
1476 	}
1477 
1478 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1479 		goto out;
1480 	if (qhp->attr.state == attrs->next_state)
1481 		goto out;
1482 
1483 	/* Return EINPROGRESS if QP is already in transition state.
1484 	 * Eg: CLOSING->IDLE transition or *->ERROR transition.
1485 	 * This can happen while connection is switching(due to rdma_fini)
1486 	 * from iWARP/RDDP to TOE mode and any inflight RDMA RX data will
1487 	 * reach TOE driver -> TCP stack -> iWARP driver. In this way
1488 	 * iWARP driver keep receiving inflight RDMA RX data until socket
1489 	 * is closed or aborted. And if iWARP CM is in FPDU sate, then
1490 	 * it tries to put QP in TERM state and disconnects endpoint.
1491 	 * But as QP is already in transition state, this event is ignored.
1492 	 */
1493 	if ((qhp->attr.state >= C4IW_QP_STATE_ERROR) &&
1494 		(attrs->next_state == C4IW_QP_STATE_TERMINATE)) {
1495 		ret = -EINPROGRESS;
1496 		goto out;
1497 	}
1498 
1499 	switch (qhp->attr.state) {
1500 	case C4IW_QP_STATE_IDLE:
1501 		switch (attrs->next_state) {
1502 		case C4IW_QP_STATE_RTS:
1503 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1504 				ret = -EINVAL;
1505 				goto out;
1506 			}
1507 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1508 				ret = -EINVAL;
1509 				goto out;
1510 			}
1511 			qhp->attr.mpa_attr = attrs->mpa_attr;
1512 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1513 			qhp->ep = qhp->attr.llp_stream_handle;
1514 			set_state(qhp, C4IW_QP_STATE_RTS);
1515 
1516 			/*
1517 			 * Ref the endpoint here and deref when we
1518 			 * disassociate the endpoint from the QP.  This
1519 			 * happens in CLOSING->IDLE transition or *->ERROR
1520 			 * transition.
1521 			 */
1522 			c4iw_get_ep(&qhp->ep->com);
1523 			ret = rdma_init(rhp, qhp);
1524 			if (ret)
1525 				goto err;
1526 			break;
1527 		case C4IW_QP_STATE_ERROR:
1528 			set_state(qhp, C4IW_QP_STATE_ERROR);
1529 			flush_qp(qhp);
1530 			break;
1531 		default:
1532 			ret = -EINVAL;
1533 			goto out;
1534 		}
1535 		break;
1536 	case C4IW_QP_STATE_RTS:
1537 		switch (attrs->next_state) {
1538 		case C4IW_QP_STATE_CLOSING:
1539 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1540 			t4_set_wq_in_error(&qhp->wq);
1541 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1542 			ep = qhp->ep;
1543 			if (!internal) {
1544 				abort = 0;
1545 				disconnect = 1;
1546 				c4iw_get_ep(&qhp->ep->com);
1547 			}
1548 			ret = rdma_fini(rhp, qhp, ep);
1549 			if (ret)
1550 				goto err;
1551 			break;
1552 		case C4IW_QP_STATE_TERMINATE:
1553 			t4_set_wq_in_error(&qhp->wq);
1554 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1555 			qhp->attr.layer_etype = attrs->layer_etype;
1556 			qhp->attr.ecode = attrs->ecode;
1557 			ep = qhp->ep;
1558 			if (!internal) {
1559 				c4iw_get_ep(&qhp->ep->com);
1560 				terminate = 1;
1561 				disconnect = 1;
1562 			} else {
1563 				terminate = qhp->attr.send_term;
1564 				ret = rdma_fini(rhp, qhp, ep);
1565 				if (ret)
1566 					goto err;
1567 			}
1568 			break;
1569 		case C4IW_QP_STATE_ERROR:
1570 			t4_set_wq_in_error(&qhp->wq);
1571 			set_state(qhp, C4IW_QP_STATE_ERROR);
1572 			if (!internal) {
1573 				abort = 1;
1574 				disconnect = 1;
1575 				ep = qhp->ep;
1576 				c4iw_get_ep(&qhp->ep->com);
1577 			}
1578 			goto err;
1579 			break;
1580 		default:
1581 			ret = -EINVAL;
1582 			goto out;
1583 		}
1584 		break;
1585 	case C4IW_QP_STATE_CLOSING:
1586 
1587 		/*
1588 		 * Allow kernel users to move to ERROR for qp draining.
1589 		 */
1590 		if (!internal && (qhp->ibqp.uobject || attrs->next_state !=
1591 				  C4IW_QP_STATE_ERROR)) {
1592 			ret = -EINVAL;
1593 			goto out;
1594 		}
1595 		switch (attrs->next_state) {
1596 		case C4IW_QP_STATE_IDLE:
1597 			flush_qp(qhp);
1598 			set_state(qhp, C4IW_QP_STATE_IDLE);
1599 			qhp->attr.llp_stream_handle = NULL;
1600 			c4iw_put_ep(&qhp->ep->com);
1601 			qhp->ep = NULL;
1602 			wake_up(&qhp->wait);
1603 			break;
1604 		case C4IW_QP_STATE_ERROR:
1605 			goto err;
1606 		default:
1607 			ret = -EINVAL;
1608 			goto err;
1609 		}
1610 		break;
1611 	case C4IW_QP_STATE_ERROR:
1612 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1613 			ret = -EINVAL;
1614 			goto out;
1615 		}
1616 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1617 			ret = -EINVAL;
1618 			goto out;
1619 		}
1620 		set_state(qhp, C4IW_QP_STATE_IDLE);
1621 		break;
1622 	case C4IW_QP_STATE_TERMINATE:
1623 		if (!internal) {
1624 			ret = -EINVAL;
1625 			goto out;
1626 		}
1627 		goto err;
1628 		break;
1629 	default:
1630 		printf("%s in a bad state %d\n",
1631 		       __func__, qhp->attr.state);
1632 		ret = -EINVAL;
1633 		goto err;
1634 		break;
1635 	}
1636 	goto out;
1637 err:
1638 	CTR3(KTR_IW_CXGBE, "%s disassociating ep %p qpid 0x%x", __func__,
1639 	    qhp->ep, qhp->wq.sq.qid);
1640 
1641 	/* disassociate the LLP connection */
1642 	qhp->attr.llp_stream_handle = NULL;
1643 	if (!ep)
1644 		ep = qhp->ep;
1645 	qhp->ep = NULL;
1646 	set_state(qhp, C4IW_QP_STATE_ERROR);
1647 	free = 1;
1648 	abort = 1;
1649 	BUG_ON(!ep);
1650 	flush_qp(qhp);
1651 	wake_up(&qhp->wait);
1652 out:
1653 	mutex_unlock(&qhp->mutex);
1654 
1655 	if (terminate)
1656 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1657 
1658 	/*
1659 	 * If disconnect is 1, then we need to initiate a disconnect
1660 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1661 	 * an abnormal close (RTS/CLOSING->ERROR).
1662 	 */
1663 	if (disconnect) {
1664 		__c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1665 							 GFP_KERNEL);
1666 		c4iw_put_ep(&ep->com);
1667 	}
1668 
1669 	/*
1670 	 * If free is 1, then we've disassociated the EP from the QP
1671 	 * and we need to dereference the EP.
1672 	 */
1673 	if (free)
1674 		c4iw_put_ep(&ep->com);
1675 	CTR2(KTR_IW_CXGBE, "%s exit state %d", __func__, qhp->attr.state);
1676 	return ret;
1677 }
1678 
1679 int c4iw_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
1680 {
1681 	struct c4iw_dev *rhp;
1682 	struct c4iw_qp *qhp;
1683 	struct c4iw_qp_attributes attrs;
1684 
1685 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ib_qp);
1686 	qhp = to_c4iw_qp(ib_qp);
1687 	rhp = qhp->rhp;
1688 
1689 	attrs.next_state = C4IW_QP_STATE_ERROR;
1690 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1691 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1692 	else
1693 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1694 	wait_event(qhp->wait, !qhp->ep);
1695 
1696 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1697 
1698 	free_ird(rhp, qhp->attr.max_ird);
1699 	c4iw_qp_rem_ref(ib_qp);
1700 
1701 	CTR3(KTR_IW_CXGBE, "%s ib_qp %p qpid 0x%0x", __func__, ib_qp,
1702 	    qhp->wq.sq.qid);
1703 	return 0;
1704 }
1705 
1706 struct ib_qp *
1707 c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1708     struct ib_udata *udata)
1709 {
1710 	struct c4iw_dev *rhp;
1711 	struct c4iw_qp *qhp;
1712 	struct c4iw_pd *php;
1713 	struct c4iw_cq *schp;
1714 	struct c4iw_cq *rchp;
1715 	struct c4iw_create_qp_resp uresp;
1716 	unsigned int sqsize, rqsize;
1717 	struct c4iw_ucontext *ucontext;
1718 	int ret;
1719 	struct c4iw_mm_entry *sq_key_mm = NULL, *rq_key_mm = NULL;
1720 	struct c4iw_mm_entry *sq_db_key_mm = NULL, *rq_db_key_mm = NULL;
1721 
1722 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
1723 
1724 	if (attrs->qp_type != IB_QPT_RC)
1725 		return ERR_PTR(-EINVAL);
1726 
1727 	php = to_c4iw_pd(pd);
1728 	rhp = php->rhp;
1729 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1730 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1731 	if (!schp || !rchp)
1732 		return ERR_PTR(-EINVAL);
1733 
1734 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1735 		return ERR_PTR(-EINVAL);
1736 
1737 	if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1738 		return ERR_PTR(-E2BIG);
1739 	rqsize = attrs->cap.max_recv_wr + 1;
1740 	if (rqsize < 8)
1741 		rqsize = 8;
1742 
1743 	if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1744 		return ERR_PTR(-E2BIG);
1745 	sqsize = attrs->cap.max_send_wr + 1;
1746 	if (sqsize < 8)
1747 		sqsize = 8;
1748 
1749 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1750 
1751 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1752 	if (!qhp)
1753 		return ERR_PTR(-ENOMEM);
1754 	qhp->wq.sq.size = sqsize;
1755 	qhp->wq.sq.memsize =
1756 		(sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1757 		sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1758 	qhp->wq.sq.flush_cidx = -1;
1759 	qhp->wq.rq.size = rqsize;
1760 	qhp->wq.rq.memsize =
1761 		(rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1762 		sizeof(*qhp->wq.rq.queue);
1763 
1764 	if (ucontext) {
1765 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1766 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1767 	}
1768 
1769 	CTR5(KTR_IW_CXGBE, "%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu",
1770 	    __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1771 
1772 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1773 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1774 	if (ret)
1775 		goto err1;
1776 
1777 	attrs->cap.max_recv_wr = rqsize - 1;
1778 	attrs->cap.max_send_wr = sqsize - 1;
1779 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1780 
1781 	qhp->rhp = rhp;
1782 	qhp->attr.pd = php->pdid;
1783 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1784 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1785 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1786 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1787 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1788 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1789 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1790 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1791 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1792 	qhp->attr.enable_rdma_read = 1;
1793 	qhp->attr.enable_rdma_write = 1;
1794 	qhp->attr.enable_bind = 1;
1795 	qhp->attr.max_ord = 0;
1796 	qhp->attr.max_ird = 0;
1797 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1798 	spin_lock_init(&qhp->lock);
1799 	mutex_init(&qhp->mutex);
1800 	init_waitqueue_head(&qhp->wait);
1801 	kref_init(&qhp->kref);
1802 	INIT_WORK(&qhp->free_work, free_qp_work);
1803 
1804 	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1805 	if (ret)
1806 		goto err2;
1807 
1808 	if (udata) {
1809 		sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL);
1810 		if (!sq_key_mm) {
1811 			ret = -ENOMEM;
1812 			goto err3;
1813 		}
1814 		rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL);
1815 		if (!rq_key_mm) {
1816 			ret = -ENOMEM;
1817 			goto err4;
1818 		}
1819 		sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL);
1820 		if (!sq_db_key_mm) {
1821 			ret = -ENOMEM;
1822 			goto err5;
1823 		}
1824 		rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL);
1825 		if (!rq_db_key_mm) {
1826 			ret = -ENOMEM;
1827 			goto err6;
1828 		}
1829 		uresp.flags = 0;
1830 		uresp.qid_mask = rhp->rdev.qpmask;
1831 		uresp.sqid = qhp->wq.sq.qid;
1832 		uresp.sq_size = qhp->wq.sq.size;
1833 		uresp.sq_memsize = qhp->wq.sq.memsize;
1834 		uresp.rqid = qhp->wq.rq.qid;
1835 		uresp.rq_size = qhp->wq.rq.size;
1836 		uresp.rq_memsize = qhp->wq.rq.memsize;
1837 		spin_lock(&ucontext->mmap_lock);
1838 		uresp.ma_sync_key =  0;
1839 		uresp.sq_key = ucontext->key;
1840 		ucontext->key += PAGE_SIZE;
1841 		uresp.rq_key = ucontext->key;
1842 		ucontext->key += PAGE_SIZE;
1843 		uresp.sq_db_gts_key = ucontext->key;
1844 		ucontext->key += PAGE_SIZE;
1845 		uresp.rq_db_gts_key = ucontext->key;
1846 		ucontext->key += PAGE_SIZE;
1847 		spin_unlock(&ucontext->mmap_lock);
1848 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1849 		if (ret)
1850 			goto err7;
1851 		sq_key_mm->key = uresp.sq_key;
1852 		sq_key_mm->addr = qhp->wq.sq.phys_addr;
1853 		sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1854 		CTR4(KTR_IW_CXGBE, "%s sq_key_mm %x, %x, %d", __func__,
1855 				sq_key_mm->key, sq_key_mm->addr,
1856 				sq_key_mm->len);
1857 		insert_mmap(ucontext, sq_key_mm);
1858 		rq_key_mm->key = uresp.rq_key;
1859 		rq_key_mm->addr = qhp->wq.rq.phys_addr;
1860 		rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1861 		CTR4(KTR_IW_CXGBE, "%s rq_key_mm %x, %x, %d", __func__,
1862 				rq_key_mm->key, rq_key_mm->addr,
1863 				rq_key_mm->len);
1864 		insert_mmap(ucontext, rq_key_mm);
1865 		sq_db_key_mm->key = uresp.sq_db_gts_key;
1866 		sq_db_key_mm->addr = (u64)qhp->wq.sq.bar2_pa;
1867 		sq_db_key_mm->len = PAGE_SIZE;
1868 		CTR4(KTR_IW_CXGBE, "%s sq_db_key_mm %x, %x, %d", __func__,
1869 				sq_db_key_mm->key, sq_db_key_mm->addr,
1870 				sq_db_key_mm->len);
1871 		insert_mmap(ucontext, sq_db_key_mm);
1872 		rq_db_key_mm->key = uresp.rq_db_gts_key;
1873 		rq_db_key_mm->addr = (u64)qhp->wq.rq.bar2_pa;
1874 		rq_db_key_mm->len = PAGE_SIZE;
1875 		CTR4(KTR_IW_CXGBE, "%s rq_db_key_mm %x, %x, %d", __func__,
1876 				rq_db_key_mm->key, rq_db_key_mm->addr,
1877 				rq_db_key_mm->len);
1878 		insert_mmap(ucontext, rq_db_key_mm);
1879 
1880 		qhp->ucontext = ucontext;
1881 	}
1882 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1883 	init_timer(&(qhp->timer));
1884 
1885 	CTR5(KTR_IW_CXGBE, "%s sq id %u size %u memsize %zu num_entries %u",
1886 		 __func__, qhp->wq.sq.qid,
1887 		 qhp->wq.sq.size, qhp->wq.sq.memsize, attrs->cap.max_send_wr);
1888 	CTR5(KTR_IW_CXGBE, "%s rq id %u size %u memsize %zu num_entries %u",
1889 		 __func__, qhp->wq.rq.qid,
1890 		 qhp->wq.rq.size, qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1891 	return &qhp->ibqp;
1892 err7:
1893 	kfree(rq_db_key_mm);
1894 err6:
1895 	kfree(sq_db_key_mm);
1896 err5:
1897 	kfree(rq_key_mm);
1898 err4:
1899 	kfree(sq_key_mm);
1900 err3:
1901 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1902 err2:
1903 	destroy_qp(&rhp->rdev, &qhp->wq,
1904 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1905 err1:
1906 	kfree(qhp);
1907 	return ERR_PTR(ret);
1908 }
1909 
1910 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1911 		      int attr_mask, struct ib_udata *udata)
1912 {
1913 	struct c4iw_dev *rhp;
1914 	struct c4iw_qp *qhp;
1915 	enum c4iw_qp_attr_mask mask = 0;
1916 	struct c4iw_qp_attributes attrs;
1917 
1918 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ibqp);
1919 
1920 	/* iwarp does not support the RTR state */
1921 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1922 		attr_mask &= ~IB_QP_STATE;
1923 
1924 	/* Make sure we still have something left to do */
1925 	if (!attr_mask)
1926 		return 0;
1927 
1928 	memset(&attrs, 0, sizeof attrs);
1929 	qhp = to_c4iw_qp(ibqp);
1930 	rhp = qhp->rhp;
1931 
1932 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1933 	attrs.enable_rdma_read = (attr->qp_access_flags &
1934 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1935 	attrs.enable_rdma_write = (attr->qp_access_flags &
1936 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1937 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1938 
1939 
1940 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1941 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1942 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1943 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1944 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1945 
1946 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1947 }
1948 
1949 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1950 {
1951 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p qpn 0x%x", __func__, dev, qpn);
1952 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1953 }
1954 
1955 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1956 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1957 {
1958 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1959 
1960 	memset(attr, 0, sizeof *attr);
1961 	memset(init_attr, 0, sizeof *init_attr);
1962 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1963 	init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1964 	init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1965 	init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1966 	init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1967 	init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1968 	init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1969 	return 0;
1970 }
1971 #endif
1972