xref: /freebsd/sys/dev/cxgbe/iw_cxgbe/qp.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 /*
2  * Copyright (c) 2009-2013 Chelsio, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 
37 #ifdef TCP_OFFLOAD
38 #include <sys/types.h>
39 #include <sys/malloc.h>
40 #include <sys/socket.h>
41 #include <sys/socketvar.h>
42 #include <sys/sockio.h>
43 #include <sys/taskqueue.h>
44 #include <netinet/in.h>
45 #include <net/route.h>
46 
47 #include <netinet/in_systm.h>
48 #include <netinet/in_pcb.h>
49 #include <netinet/ip.h>
50 #include <netinet/ip_var.h>
51 #include <netinet/tcp_var.h>
52 #include <netinet/tcp.h>
53 #include <netinet/tcpip.h>
54 
55 #include <netinet/toecore.h>
56 
57 struct sge_iq;
58 struct rss_header;
59 #include <linux/types.h>
60 #include "offload.h"
61 #include "tom/t4_tom.h"
62 
63 #include "iw_cxgbe.h"
64 #include "user.h"
65 
66 extern int db_delay_usecs;
67 extern int db_fc_threshold;
68 static void creds(struct toepcb *toep, size_t wrsize);
69 
70 
71 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
72 {
73 	unsigned long flag;
74 	spin_lock_irqsave(&qhp->lock, flag);
75 	qhp->attr.state = state;
76 	spin_unlock_irqrestore(&qhp->lock, flag);
77 }
78 
79 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
80 {
81 
82 	contigfree(sq->queue, sq->memsize, M_DEVBUF);
83 }
84 
85 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
86 {
87 
88 	dealloc_host_sq(rdev, sq);
89 }
90 
91 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
92 {
93 	sq->queue = contigmalloc(sq->memsize, M_DEVBUF, M_NOWAIT, 0ul, ~0ul,
94 	    4096, 0);
95 
96 	if (sq->queue)
97 		sq->dma_addr = vtophys(sq->queue);
98 	else
99 		return -ENOMEM;
100 	sq->phys_addr = vtophys(sq->queue);
101 	pci_unmap_addr_set(sq, mapping, sq->dma_addr);
102 	CTR4(KTR_IW_CXGBE, "%s sq %p dma_addr %p phys_addr %p", __func__,
103 	    sq->queue, sq->dma_addr, sq->phys_addr);
104 	return 0;
105 }
106 
107 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
108 		      struct c4iw_dev_ucontext *uctx)
109 {
110 	/*
111 	 * uP clears EQ contexts when the connection exits rdma mode,
112 	 * so no need to post a RESET WR for these EQs.
113 	 */
114 	contigfree(wq->rq.queue, wq->rq.memsize, M_DEVBUF);
115 	dealloc_sq(rdev, &wq->sq);
116 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
117 	kfree(wq->rq.sw_rq);
118 	kfree(wq->sq.sw_sq);
119 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
120 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
121 	return 0;
122 }
123 
124 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
125 		     struct t4_cq *rcq, struct t4_cq *scq,
126 		     struct c4iw_dev_ucontext *uctx)
127 {
128 	struct adapter *sc = rdev->adap;
129 	int user = (uctx != &rdev->uctx);
130 	struct fw_ri_res_wr *res_wr;
131 	struct fw_ri_res *res;
132 	int wr_len;
133 	struct c4iw_wr_wait wr_wait;
134 	int ret;
135 	int eqsize;
136 	struct wrqe *wr;
137 
138 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
139 	if (!wq->sq.qid)
140 		return -ENOMEM;
141 
142 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
143 	if (!wq->rq.qid)
144 		goto err1;
145 
146 	if (!user) {
147 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
148 				 GFP_KERNEL);
149 		if (!wq->sq.sw_sq)
150 			goto err2;
151 
152 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
153 				 GFP_KERNEL);
154 		if (!wq->rq.sw_rq)
155 			goto err3;
156 	}
157 
158 	/* RQT must be a power of 2. */
159 	wq->rq.rqt_size = roundup_pow_of_two(wq->rq.size);
160 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
161 	if (!wq->rq.rqt_hwaddr)
162 		goto err4;
163 
164 	if (alloc_host_sq(rdev, &wq->sq))
165 		goto err5;
166 
167 	memset(wq->sq.queue, 0, wq->sq.memsize);
168 	pci_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
169 
170 	wq->rq.queue = contigmalloc(wq->rq.memsize,
171             M_DEVBUF, M_NOWAIT, 0ul, ~0ul, 4096, 0);
172         if (wq->rq.queue)
173                 wq->rq.dma_addr = vtophys(wq->rq.queue);
174         else
175                 goto err6;
176 	CTR5(KTR_IW_CXGBE,
177 	    "%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx", __func__,
178 	    wq->sq.queue, (unsigned long long)vtophys(wq->sq.queue),
179 	    wq->rq.queue, (unsigned long long)vtophys(wq->rq.queue));
180 	memset(wq->rq.queue, 0, wq->rq.memsize);
181 	pci_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
182 
183 	wq->db = (void *)((unsigned long)rman_get_virtual(sc->regs_res) +
184 	    MYPF_REG(SGE_PF_KDOORBELL));
185 	wq->gts = (void *)((unsigned long)rman_get_virtual(rdev->adap->regs_res)
186 			   + MYPF_REG(SGE_PF_GTS));
187 	if (user) {
188 		wq->sq.udb = (u64)((char*)rman_get_virtual(rdev->adap->udbs_res) +
189 						(wq->sq.qid << rdev->qpshift));
190 		wq->sq.udb &= PAGE_MASK;
191 		wq->rq.udb = (u64)((char*)rman_get_virtual(rdev->adap->udbs_res) +
192 						(wq->rq.qid << rdev->qpshift));
193 		wq->rq.udb &= PAGE_MASK;
194 	}
195 	wq->rdev = rdev;
196 	wq->rq.msn = 1;
197 
198 	/* build fw_ri_res_wr */
199 	wr_len = sizeof *res_wr + 2 * sizeof *res;
200 
201 	wr = alloc_wrqe(wr_len, &sc->sge.mgmtq);
202         if (wr == NULL)
203 		return (0);
204         res_wr = wrtod(wr);
205 
206 	memset(res_wr, 0, wr_len);
207 	res_wr->op_nres = cpu_to_be32(
208 			V_FW_WR_OP(FW_RI_RES_WR) |
209 			V_FW_RI_RES_WR_NRES(2) |
210 			F_FW_WR_COMPL);
211 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
212 	res_wr->cookie = (unsigned long) &wr_wait;
213 	res = res_wr->res;
214 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
215 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
216 
217 	/* eqsize is the number of 64B entries plus the status page size. */
218 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS + spg_creds;
219 
220 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
221 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
222 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
223 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
224 		V_FW_RI_RES_WR_IQID(scq->cqid));
225 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
226 		V_FW_RI_RES_WR_DCAEN(0) |
227 		V_FW_RI_RES_WR_DCACPU(0) |
228 		V_FW_RI_RES_WR_FBMIN(2) |
229 		V_FW_RI_RES_WR_FBMAX(2) |
230 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
231 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
232 		V_FW_RI_RES_WR_EQSIZE(eqsize));
233 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
234 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
235 	res++;
236 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
237 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
238 
239 	/* eqsize is the number of 64B entries plus the status page size. */
240 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS + spg_creds ;
241 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
242 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
243 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
244 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
245 		V_FW_RI_RES_WR_IQID(rcq->cqid));
246 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
247 		V_FW_RI_RES_WR_DCAEN(0) |
248 		V_FW_RI_RES_WR_DCACPU(0) |
249 		V_FW_RI_RES_WR_FBMIN(2) |
250 		V_FW_RI_RES_WR_FBMAX(2) |
251 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
252 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
253 		V_FW_RI_RES_WR_EQSIZE(eqsize));
254 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
255 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
256 
257 	c4iw_init_wr_wait(&wr_wait);
258 
259 	t4_wrq_tx(sc, wr);
260 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
261 	if (ret)
262 		goto err7;
263 
264 	CTR6(KTR_IW_CXGBE,
265 	    "%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx",
266 	    __func__, wq->sq.qid, wq->rq.qid, wq->db,
267 	    (unsigned long long)wq->sq.udb, (unsigned long long)wq->rq.udb);
268 
269 	return 0;
270 err7:
271 	contigfree(wq->rq.queue, wq->rq.memsize, M_DEVBUF);
272 err6:
273 	dealloc_sq(rdev, &wq->sq);
274 err5:
275 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
276 err4:
277 	kfree(wq->rq.sw_rq);
278 err3:
279 	kfree(wq->sq.sw_sq);
280 err2:
281 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
282 err1:
283 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
284 	return -ENOMEM;
285 }
286 
287 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
288 		      struct ib_send_wr *wr, int max, u32 *plenp)
289 {
290 	u8 *dstp, *srcp;
291 	u32 plen = 0;
292 	int i;
293 	int rem, len;
294 
295 	dstp = (u8 *)immdp->data;
296 	for (i = 0; i < wr->num_sge; i++) {
297 		if ((plen + wr->sg_list[i].length) > max)
298 			return -EMSGSIZE;
299 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
300 		plen += wr->sg_list[i].length;
301 		rem = wr->sg_list[i].length;
302 		while (rem) {
303 			if (dstp == (u8 *)&sq->queue[sq->size])
304 				dstp = (u8 *)sq->queue;
305 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
306 				len = rem;
307 			else
308 				len = (u8 *)&sq->queue[sq->size] - dstp;
309 			memcpy(dstp, srcp, len);
310 			dstp += len;
311 			srcp += len;
312 			rem -= len;
313 		}
314 	}
315 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
316 	if (len)
317 		memset(dstp, 0, len);
318 	immdp->op = FW_RI_DATA_IMMD;
319 	immdp->r1 = 0;
320 	immdp->r2 = 0;
321 	immdp->immdlen = cpu_to_be32(plen);
322 	*plenp = plen;
323 	return 0;
324 }
325 
326 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
327 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
328 		      int num_sge, u32 *plenp)
329 
330 {
331 	int i;
332 	u32 plen = 0;
333 	__be64 *flitp = (__be64 *)isglp->sge;
334 
335 	for (i = 0; i < num_sge; i++) {
336 		if ((plen + sg_list[i].length) < plen)
337 			return -EMSGSIZE;
338 		plen += sg_list[i].length;
339 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
340 				     sg_list[i].length);
341 		if (++flitp == queue_end)
342 			flitp = queue_start;
343 		*flitp = cpu_to_be64(sg_list[i].addr);
344 		if (++flitp == queue_end)
345 			flitp = queue_start;
346 	}
347 	*flitp = (__force __be64)0;
348 	isglp->op = FW_RI_DATA_ISGL;
349 	isglp->r1 = 0;
350 	isglp->nsge = cpu_to_be16(num_sge);
351 	isglp->r2 = 0;
352 	if (plenp)
353 		*plenp = plen;
354 	return 0;
355 }
356 
357 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
358 			   struct ib_send_wr *wr, u8 *len16)
359 {
360 	u32 plen;
361 	int size;
362 	int ret;
363 
364 	if (wr->num_sge > T4_MAX_SEND_SGE)
365 		return -EINVAL;
366 	switch (wr->opcode) {
367 	case IB_WR_SEND:
368 		if (wr->send_flags & IB_SEND_SOLICITED)
369 			wqe->send.sendop_pkd = cpu_to_be32(
370 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
371 		else
372 			wqe->send.sendop_pkd = cpu_to_be32(
373 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
374 		wqe->send.stag_inv = 0;
375 		break;
376 	case IB_WR_SEND_WITH_INV:
377 		if (wr->send_flags & IB_SEND_SOLICITED)
378 			wqe->send.sendop_pkd = cpu_to_be32(
379 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
380 		else
381 			wqe->send.sendop_pkd = cpu_to_be32(
382 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
383 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
384 		break;
385 
386 	default:
387 		return -EINVAL;
388 	}
389 
390 	plen = 0;
391 	if (wr->num_sge) {
392 		if (wr->send_flags & IB_SEND_INLINE) {
393 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
394 					 T4_MAX_SEND_INLINE, &plen);
395 			if (ret)
396 				return ret;
397 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
398 			       plen;
399 		} else {
400 			ret = build_isgl((__be64 *)sq->queue,
401 					 (__be64 *)&sq->queue[sq->size],
402 					 wqe->send.u.isgl_src,
403 					 wr->sg_list, wr->num_sge, &plen);
404 			if (ret)
405 				return ret;
406 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
407 			       wr->num_sge * sizeof(struct fw_ri_sge);
408 		}
409 	} else {
410 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
411 		wqe->send.u.immd_src[0].r1 = 0;
412 		wqe->send.u.immd_src[0].r2 = 0;
413 		wqe->send.u.immd_src[0].immdlen = 0;
414 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
415 		plen = 0;
416 	}
417 	*len16 = DIV_ROUND_UP(size, 16);
418 	wqe->send.plen = cpu_to_be32(plen);
419 	return 0;
420 }
421 
422 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
423 			    struct ib_send_wr *wr, u8 *len16)
424 {
425 	u32 plen;
426 	int size;
427 	int ret;
428 
429 	if (wr->num_sge > T4_MAX_SEND_SGE)
430 		return -EINVAL;
431 	wqe->write.r2 = 0;
432 	wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
433 	wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
434 	if (wr->num_sge) {
435 		if (wr->send_flags & IB_SEND_INLINE) {
436 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
437 					 T4_MAX_WRITE_INLINE, &plen);
438 			if (ret)
439 				return ret;
440 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
441 			       plen;
442 		} else {
443 			ret = build_isgl((__be64 *)sq->queue,
444 					 (__be64 *)&sq->queue[sq->size],
445 					 wqe->write.u.isgl_src,
446 					 wr->sg_list, wr->num_sge, &plen);
447 			if (ret)
448 				return ret;
449 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
450 			       wr->num_sge * sizeof(struct fw_ri_sge);
451 		}
452 	} else {
453 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
454 		wqe->write.u.immd_src[0].r1 = 0;
455 		wqe->write.u.immd_src[0].r2 = 0;
456 		wqe->write.u.immd_src[0].immdlen = 0;
457 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
458 		plen = 0;
459 	}
460 	*len16 = DIV_ROUND_UP(size, 16);
461 	wqe->write.plen = cpu_to_be32(plen);
462 	return 0;
463 }
464 
465 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
466 {
467 	if (wr->num_sge > 1)
468 		return -EINVAL;
469 	if (wr->num_sge) {
470 		wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
471 		wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
472 							>> 32));
473 		wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
474 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
475 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
476 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
477 							 >> 32));
478 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
479 	} else {
480 		wqe->read.stag_src = cpu_to_be32(2);
481 		wqe->read.to_src_hi = 0;
482 		wqe->read.to_src_lo = 0;
483 		wqe->read.stag_sink = cpu_to_be32(2);
484 		wqe->read.plen = 0;
485 		wqe->read.to_sink_hi = 0;
486 		wqe->read.to_sink_lo = 0;
487 	}
488 	wqe->read.r2 = 0;
489 	wqe->read.r5 = 0;
490 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
491 	return 0;
492 }
493 
494 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
495 			   struct ib_recv_wr *wr, u8 *len16)
496 {
497 	int ret;
498 
499 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
500 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
501 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
502 	if (ret)
503 		return ret;
504 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
505 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
506 	return 0;
507 }
508 
509 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
510 			 struct ib_send_wr *wr, u8 *len16)
511 {
512 
513 	struct fw_ri_immd *imdp;
514 	__be64 *p;
515 	int i;
516 	int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
517 	int rem;
518 
519 	if (wr->wr.fast_reg.page_list_len > T4_MAX_FR_DEPTH)
520 		return -EINVAL;
521 
522 	wqe->fr.qpbinde_to_dcacpu = 0;
523 	wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
524 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
525 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
526 	wqe->fr.len_hi = 0;
527 	wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
528 	wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
529 	wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
530 	wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
531 					0xffffffff);
532 	WARN_ON(pbllen > T4_MAX_FR_IMMD);
533 	imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
534 	imdp->op = FW_RI_DATA_IMMD;
535 	imdp->r1 = 0;
536 	imdp->r2 = 0;
537 	imdp->immdlen = cpu_to_be32(pbllen);
538 	p = (__be64 *)(imdp + 1);
539 	rem = pbllen;
540 	for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
541 		*p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
542 		rem -= sizeof *p;
543 		if (++p == (__be64 *)&sq->queue[sq->size])
544 			p = (__be64 *)sq->queue;
545 	}
546 	BUG_ON(rem < 0);
547 	while (rem) {
548 		*p = 0;
549 		rem -= sizeof *p;
550 		if (++p == (__be64 *)&sq->queue[sq->size])
551 			p = (__be64 *)sq->queue;
552 	}
553 	*len16 = DIV_ROUND_UP(sizeof wqe->fr + sizeof *imdp + pbllen, 16);
554 	return 0;
555 }
556 
557 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
558 			  u8 *len16)
559 {
560 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
561 	wqe->inv.r2 = 0;
562 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
563 	return 0;
564 }
565 
566 void c4iw_qp_add_ref(struct ib_qp *qp)
567 {
568 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
569 	atomic_inc(&(to_c4iw_qp(qp)->refcnt));
570 }
571 
572 void c4iw_qp_rem_ref(struct ib_qp *qp)
573 {
574 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp);
575 	if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
576 		wake_up(&(to_c4iw_qp(qp)->wait));
577 }
578 
579 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
580 		   struct ib_send_wr **bad_wr)
581 {
582 	int err = 0;
583 	u8 len16 = 0;
584 	enum fw_wr_opcodes fw_opcode = 0;
585 	enum fw_ri_wr_flags fw_flags;
586 	struct c4iw_qp *qhp;
587 	union t4_wr *wqe;
588 	u32 num_wrs;
589 	struct t4_swsqe *swsqe;
590 	unsigned long flag;
591 	u16 idx = 0;
592 
593 	qhp = to_c4iw_qp(ibqp);
594 	spin_lock_irqsave(&qhp->lock, flag);
595 	if (t4_wq_in_error(&qhp->wq)) {
596 		spin_unlock_irqrestore(&qhp->lock, flag);
597 		return -EINVAL;
598 	}
599 	num_wrs = t4_sq_avail(&qhp->wq);
600 	if (num_wrs == 0) {
601 		spin_unlock_irqrestore(&qhp->lock, flag);
602 		return -ENOMEM;
603 	}
604 	while (wr) {
605 		if (num_wrs == 0) {
606 			err = -ENOMEM;
607 			*bad_wr = wr;
608 			break;
609 		}
610 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
611 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
612 
613 		fw_flags = 0;
614 		if (wr->send_flags & IB_SEND_SOLICITED)
615 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
616 		if (wr->send_flags & IB_SEND_SIGNALED)
617 			fw_flags |= FW_RI_COMPLETION_FLAG;
618 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
619 		switch (wr->opcode) {
620 		case IB_WR_SEND_WITH_INV:
621 		case IB_WR_SEND:
622 			if (wr->send_flags & IB_SEND_FENCE)
623 				fw_flags |= FW_RI_READ_FENCE_FLAG;
624 			fw_opcode = FW_RI_SEND_WR;
625 			if (wr->opcode == IB_WR_SEND)
626 				swsqe->opcode = FW_RI_SEND;
627 			else
628 				swsqe->opcode = FW_RI_SEND_WITH_INV;
629 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
630 			break;
631 		case IB_WR_RDMA_WRITE:
632 			fw_opcode = FW_RI_RDMA_WRITE_WR;
633 			swsqe->opcode = FW_RI_RDMA_WRITE;
634 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
635 			break;
636 		case IB_WR_RDMA_READ:
637 		case IB_WR_RDMA_READ_WITH_INV:
638 			fw_opcode = FW_RI_RDMA_READ_WR;
639 			swsqe->opcode = FW_RI_READ_REQ;
640 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
641 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
642 			else
643 				fw_flags = 0;
644 			err = build_rdma_read(wqe, wr, &len16);
645 			if (err)
646 				break;
647 			swsqe->read_len = wr->sg_list[0].length;
648 			if (!qhp->wq.sq.oldest_read)
649 				qhp->wq.sq.oldest_read = swsqe;
650 			break;
651 		case IB_WR_FAST_REG_MR:
652 			fw_opcode = FW_RI_FR_NSMR_WR;
653 			swsqe->opcode = FW_RI_FAST_REGISTER;
654 			err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16);
655 			break;
656 		case IB_WR_LOCAL_INV:
657 			if (wr->send_flags & IB_SEND_FENCE)
658 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
659 			fw_opcode = FW_RI_INV_LSTAG_WR;
660 			swsqe->opcode = FW_RI_LOCAL_INV;
661 			err = build_inv_stag(wqe, wr, &len16);
662 			break;
663 		default:
664 			CTR2(KTR_IW_CXGBE, "%s post of type =%d TBD!", __func__,
665 			     wr->opcode);
666 			err = -EINVAL;
667 		}
668 		if (err) {
669 			*bad_wr = wr;
670 			break;
671 		}
672 		swsqe->idx = qhp->wq.sq.pidx;
673 		swsqe->complete = 0;
674 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED);
675 		swsqe->wr_id = wr->wr_id;
676 
677 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
678 
679 		CTR5(KTR_IW_CXGBE,
680 		    "%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u",
681 		    __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
682 		    swsqe->opcode, swsqe->read_len);
683 		wr = wr->next;
684 		num_wrs--;
685 		t4_sq_produce(&qhp->wq, len16);
686 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
687 	}
688 	if (t4_wq_db_enabled(&qhp->wq))
689 		t4_ring_sq_db(&qhp->wq, idx);
690 	spin_unlock_irqrestore(&qhp->lock, flag);
691 	return err;
692 }
693 
694 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
695 		      struct ib_recv_wr **bad_wr)
696 {
697 	int err = 0;
698 	struct c4iw_qp *qhp;
699 	union t4_recv_wr *wqe;
700 	u32 num_wrs;
701 	u8 len16 = 0;
702 	unsigned long flag;
703 	u16 idx = 0;
704 
705 	qhp = to_c4iw_qp(ibqp);
706 	spin_lock_irqsave(&qhp->lock, flag);
707 	if (t4_wq_in_error(&qhp->wq)) {
708 		spin_unlock_irqrestore(&qhp->lock, flag);
709 		return -EINVAL;
710 	}
711 	num_wrs = t4_rq_avail(&qhp->wq);
712 	if (num_wrs == 0) {
713 		spin_unlock_irqrestore(&qhp->lock, flag);
714 		return -ENOMEM;
715 	}
716 	while (wr) {
717 		if (wr->num_sge > T4_MAX_RECV_SGE) {
718 			err = -EINVAL;
719 			*bad_wr = wr;
720 			break;
721 		}
722 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
723 					   qhp->wq.rq.wq_pidx *
724 					   T4_EQ_ENTRY_SIZE);
725 		if (num_wrs)
726 			err = build_rdma_recv(qhp, wqe, wr, &len16);
727 		else
728 			err = -ENOMEM;
729 		if (err) {
730 			*bad_wr = wr;
731 			break;
732 		}
733 
734 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
735 
736 		wqe->recv.opcode = FW_RI_RECV_WR;
737 		wqe->recv.r1 = 0;
738 		wqe->recv.wrid = qhp->wq.rq.pidx;
739 		wqe->recv.r2[0] = 0;
740 		wqe->recv.r2[1] = 0;
741 		wqe->recv.r2[2] = 0;
742 		wqe->recv.len16 = len16;
743 		CTR3(KTR_IW_CXGBE, "%s cookie 0x%llx pidx %u", __func__,
744 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
745 		t4_rq_produce(&qhp->wq, len16);
746 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
747 		wr = wr->next;
748 		num_wrs--;
749 	}
750 	if (t4_wq_db_enabled(&qhp->wq))
751 		t4_ring_rq_db(&qhp->wq, idx);
752 	spin_unlock_irqrestore(&qhp->lock, flag);
753 	return err;
754 }
755 
756 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
757 {
758 	return -ENOSYS;
759 }
760 
761 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
762 				    u8 *ecode)
763 {
764 	int status;
765 	int tagged;
766 	int opcode;
767 	int rqtype;
768 	int send_inv;
769 
770 	if (!err_cqe) {
771 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
772 		*ecode = 0;
773 		return;
774 	}
775 
776 	status = CQE_STATUS(err_cqe);
777 	opcode = CQE_OPCODE(err_cqe);
778 	rqtype = RQ_TYPE(err_cqe);
779 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
780 		   (opcode == FW_RI_SEND_WITH_SE_INV);
781 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
782 		 (rqtype && (opcode == FW_RI_READ_RESP));
783 
784 	switch (status) {
785 	case T4_ERR_STAG:
786 		if (send_inv) {
787 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
788 			*ecode = RDMAP_CANT_INV_STAG;
789 		} else {
790 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
791 			*ecode = RDMAP_INV_STAG;
792 		}
793 		break;
794 	case T4_ERR_PDID:
795 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
796 		if ((opcode == FW_RI_SEND_WITH_INV) ||
797 		    (opcode == FW_RI_SEND_WITH_SE_INV))
798 			*ecode = RDMAP_CANT_INV_STAG;
799 		else
800 			*ecode = RDMAP_STAG_NOT_ASSOC;
801 		break;
802 	case T4_ERR_QPID:
803 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
804 		*ecode = RDMAP_STAG_NOT_ASSOC;
805 		break;
806 	case T4_ERR_ACCESS:
807 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
808 		*ecode = RDMAP_ACC_VIOL;
809 		break;
810 	case T4_ERR_WRAP:
811 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
812 		*ecode = RDMAP_TO_WRAP;
813 		break;
814 	case T4_ERR_BOUND:
815 		if (tagged) {
816 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
817 			*ecode = DDPT_BASE_BOUNDS;
818 		} else {
819 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
820 			*ecode = RDMAP_BASE_BOUNDS;
821 		}
822 		break;
823 	case T4_ERR_INVALIDATE_SHARED_MR:
824 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
825 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
826 		*ecode = RDMAP_CANT_INV_STAG;
827 		break;
828 	case T4_ERR_ECC:
829 	case T4_ERR_ECC_PSTAG:
830 	case T4_ERR_INTERNAL_ERR:
831 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
832 		*ecode = 0;
833 		break;
834 	case T4_ERR_OUT_OF_RQE:
835 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
836 		*ecode = DDPU_INV_MSN_NOBUF;
837 		break;
838 	case T4_ERR_PBL_ADDR_BOUND:
839 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
840 		*ecode = DDPT_BASE_BOUNDS;
841 		break;
842 	case T4_ERR_CRC:
843 		*layer_type = LAYER_MPA|DDP_LLP;
844 		*ecode = MPA_CRC_ERR;
845 		break;
846 	case T4_ERR_MARKER:
847 		*layer_type = LAYER_MPA|DDP_LLP;
848 		*ecode = MPA_MARKER_ERR;
849 		break;
850 	case T4_ERR_PDU_LEN_ERR:
851 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
852 		*ecode = DDPU_MSG_TOOBIG;
853 		break;
854 	case T4_ERR_DDP_VERSION:
855 		if (tagged) {
856 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
857 			*ecode = DDPT_INV_VERS;
858 		} else {
859 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
860 			*ecode = DDPU_INV_VERS;
861 		}
862 		break;
863 	case T4_ERR_RDMA_VERSION:
864 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
865 		*ecode = RDMAP_INV_VERS;
866 		break;
867 	case T4_ERR_OPCODE:
868 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
869 		*ecode = RDMAP_INV_OPCODE;
870 		break;
871 	case T4_ERR_DDP_QUEUE_NUM:
872 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
873 		*ecode = DDPU_INV_QN;
874 		break;
875 	case T4_ERR_MSN:
876 	case T4_ERR_MSN_GAP:
877 	case T4_ERR_MSN_RANGE:
878 	case T4_ERR_IRD_OVERFLOW:
879 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
880 		*ecode = DDPU_INV_MSN_RANGE;
881 		break;
882 	case T4_ERR_TBIT:
883 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
884 		*ecode = 0;
885 		break;
886 	case T4_ERR_MO:
887 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
888 		*ecode = DDPU_INV_MO;
889 		break;
890 	default:
891 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
892 		*ecode = 0;
893 		break;
894 	}
895 }
896 
897 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
898 			   gfp_t gfp)
899 {
900 	struct fw_ri_wr *wqe;
901 	struct terminate_message *term;
902 	struct wrqe *wr;
903 	struct socket *so = qhp->ep->com.so;
904         struct inpcb *inp = sotoinpcb(so);
905         struct tcpcb *tp = intotcpcb(inp);
906         struct toepcb *toep = tp->t_toe;
907 
908 	CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
909 	    qhp->wq.sq.qid, qhp->ep->hwtid);
910 
911 	wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
912 	if (wr == NULL)
913 		return;
914         wqe = wrtod(wr);
915 
916 	memset(wqe, 0, sizeof *wqe);
917 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR));
918 	wqe->flowid_len16 = cpu_to_be32(
919 		V_FW_WR_FLOWID(qhp->ep->hwtid) |
920 		V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
921 
922 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
923 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
924 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
925 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
926 		term->layer_etype = qhp->attr.layer_etype;
927 		term->ecode = qhp->attr.ecode;
928 	} else
929 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
930         creds(toep, sizeof(*wqe));
931 	t4_wrq_tx(qhp->rhp->rdev.adap, wr);
932 }
933 
934 /* Assumes qhp lock is held. */
935 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
936 		       struct c4iw_cq *schp)
937 {
938 	int count;
939 	int flushed;
940 	unsigned long flag;
941 
942 	CTR4(KTR_IW_CXGBE, "%s qhp %p rchp %p schp %p", __func__, qhp, rchp,
943 	    schp);
944 
945 	/* locking hierarchy: cq lock first, then qp lock. */
946 	spin_lock_irqsave(&rchp->lock, flag);
947 	spin_lock(&qhp->lock);
948 	c4iw_flush_hw_cq(&rchp->cq);
949 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
950 	flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
951 	spin_unlock(&qhp->lock);
952 	spin_unlock_irqrestore(&rchp->lock, flag);
953 	if (flushed) {
954 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
955 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
956 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
957 	}
958 
959 	/* locking hierarchy: cq lock first, then qp lock. */
960 	spin_lock_irqsave(&schp->lock, flag);
961 	spin_lock(&qhp->lock);
962 	c4iw_flush_hw_cq(&schp->cq);
963 	c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
964 	flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
965 	spin_unlock(&qhp->lock);
966 	spin_unlock_irqrestore(&schp->lock, flag);
967 	if (flushed) {
968 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
969 		(*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
970 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
971 	}
972 }
973 
974 static void flush_qp(struct c4iw_qp *qhp)
975 {
976 	struct c4iw_cq *rchp, *schp;
977 	unsigned long flag;
978 
979 	rchp = get_chp(qhp->rhp, qhp->attr.rcq);
980 	schp = get_chp(qhp->rhp, qhp->attr.scq);
981 
982 	if (qhp->ibqp.uobject) {
983 		t4_set_wq_in_error(&qhp->wq);
984 		t4_set_cq_in_error(&rchp->cq);
985 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
986 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
987 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
988 		if (schp != rchp) {
989 			t4_set_cq_in_error(&schp->cq);
990 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
991 			(*schp->ibcq.comp_handler)(&schp->ibcq,
992 					schp->ibcq.cq_context);
993 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
994 		}
995 		return;
996 	}
997 	__flush_qp(qhp, rchp, schp);
998 }
999 
1000 static int
1001 rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp, struct c4iw_ep *ep)
1002 {
1003 	struct c4iw_rdev *rdev = &rhp->rdev;
1004 	struct adapter *sc = rdev->adap;
1005 	struct fw_ri_wr *wqe;
1006 	int ret;
1007 	struct wrqe *wr;
1008 	struct socket *so = ep->com.so;
1009         struct inpcb *inp = sotoinpcb(so);
1010         struct tcpcb *tp = intotcpcb(inp);
1011         struct toepcb *toep = tp->t_toe;
1012 
1013 	KASSERT(rhp == qhp->rhp && ep == qhp->ep, ("%s: EDOOFUS", __func__));
1014 
1015 	CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1016 	    qhp->wq.sq.qid, ep->hwtid);
1017 
1018 	wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
1019 	if (wr == NULL)
1020 		return (0);
1021 	wqe = wrtod(wr);
1022 
1023 	memset(wqe, 0, sizeof *wqe);
1024 
1025 	wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL);
1026 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1027 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1028 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1029 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1030 
1031 	c4iw_init_wr_wait(&ep->com.wr_wait);
1032 
1033         creds(toep, sizeof(*wqe));
1034 	t4_wrq_tx(sc, wr);
1035 
1036 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1037 	    qhp->wq.sq.qid, __func__);
1038 	return ret;
1039 }
1040 
1041 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1042 {
1043 	CTR2(KTR_IW_CXGBE, "%s p2p_type = %d", __func__, p2p_type);
1044 	memset(&init->u, 0, sizeof init->u);
1045 	switch (p2p_type) {
1046 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1047 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1048 		init->u.write.stag_sink = cpu_to_be32(1);
1049 		init->u.write.to_sink = cpu_to_be64(1);
1050 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1051 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1052 						   sizeof(struct fw_ri_immd),
1053 						   16);
1054 		break;
1055 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1056 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1057 		init->u.read.stag_src = cpu_to_be32(1);
1058 		init->u.read.to_src_lo = cpu_to_be32(1);
1059 		init->u.read.stag_sink = cpu_to_be32(1);
1060 		init->u.read.to_sink_lo = cpu_to_be32(1);
1061 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1062 		break;
1063 	}
1064 }
1065 
1066 static void
1067 creds(struct toepcb *toep, size_t wrsize)
1068 {
1069 	struct ofld_tx_sdesc *txsd;
1070 
1071 	CTR3(KTR_IW_CXGBE, "%s:creB  %p %u", __func__, toep , wrsize);
1072 	INP_WLOCK(toep->inp);
1073 	txsd = &toep->txsd[toep->txsd_pidx];
1074 	txsd->tx_credits = howmany(wrsize, 16);
1075 	txsd->plen = 0;
1076 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
1077 			("%s: not enough credits (%d)", __func__, toep->tx_credits));
1078 	toep->tx_credits -= txsd->tx_credits;
1079 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1080 		toep->txsd_pidx = 0;
1081 	toep->txsd_avail--;
1082 	INP_WUNLOCK(toep->inp);
1083 	CTR5(KTR_IW_CXGBE, "%s:creE  %p %u %u %u", __func__, toep ,
1084 	    txsd->tx_credits, toep->tx_credits, toep->txsd_pidx);
1085 }
1086 
1087 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1088 {
1089 	struct fw_ri_wr *wqe;
1090 	int ret;
1091 	struct wrqe *wr;
1092 	struct c4iw_ep *ep = qhp->ep;
1093 	struct c4iw_rdev *rdev = &qhp->rhp->rdev;
1094 	struct adapter *sc = rdev->adap;
1095 	struct socket *so = ep->com.so;
1096         struct inpcb *inp = sotoinpcb(so);
1097         struct tcpcb *tp = intotcpcb(inp);
1098         struct toepcb *toep = tp->t_toe;
1099 
1100 	CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp,
1101 	    qhp->wq.sq.qid, ep->hwtid);
1102 
1103 	wr = alloc_wrqe(sizeof(*wqe), toep->ofld_txq);
1104 	if (wr == NULL)
1105 		return (0);
1106 	wqe = wrtod(wr);
1107 
1108 	memset(wqe, 0, sizeof *wqe);
1109 
1110 	wqe->op_compl = cpu_to_be32(
1111 		V_FW_WR_OP(FW_RI_WR) |
1112 		F_FW_WR_COMPL);
1113 	wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) |
1114 	    V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1115 
1116 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1117 
1118 	wqe->u.init.type = FW_RI_TYPE_INIT;
1119 	wqe->u.init.mpareqbit_p2ptype =
1120 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1121 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1122 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1123 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1124 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1125 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1126 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1127 	if (qhp->attr.mpa_attr.crc_enabled)
1128 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1129 
1130 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1131 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1132 			    FW_RI_QP_BIND_ENABLE;
1133 	if (!qhp->ibqp.uobject)
1134 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1135 				     FW_RI_QP_STAG0_ENABLE;
1136 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1137 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1138 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1139 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1140 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1141 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1142 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1143 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1144 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1145 	wqe->u.init.iss = cpu_to_be32(ep->snd_seq);
1146 	wqe->u.init.irs = cpu_to_be32(ep->rcv_seq);
1147 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1148 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1149 	    sc->vres.rq.start);
1150 	if (qhp->attr.mpa_attr.initiator)
1151 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1152 
1153 	c4iw_init_wr_wait(&ep->com.wr_wait);
1154 
1155 	creds(toep, sizeof(*wqe));
1156 	t4_wrq_tx(sc, wr);
1157 
1158 	ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid,
1159 	    qhp->wq.sq.qid, __func__);
1160 
1161 	toep->ulp_mode = ULP_MODE_RDMA;
1162 
1163 	return ret;
1164 }
1165 
1166 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1167 		   enum c4iw_qp_attr_mask mask,
1168 		   struct c4iw_qp_attributes *attrs,
1169 		   int internal)
1170 {
1171 	int ret = 0;
1172 	struct c4iw_qp_attributes newattr = qhp->attr;
1173 	int disconnect = 0;
1174 	int terminate = 0;
1175 	int abort = 0;
1176 	int free = 0;
1177 	struct c4iw_ep *ep = NULL;
1178 
1179 	CTR5(KTR_IW_CXGBE, "%s qhp %p sqid 0x%x rqid 0x%x ep %p", __func__, qhp,
1180 	    qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep);
1181 	CTR3(KTR_IW_CXGBE, "%s state %d -> %d", __func__, qhp->attr.state,
1182 	    (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1183 
1184 	mutex_lock(&qhp->mutex);
1185 
1186 	/* Process attr changes if in IDLE */
1187 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1188 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1189 			ret = -EIO;
1190 			goto out;
1191 		}
1192 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1193 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1194 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1195 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1196 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1197 			newattr.enable_bind = attrs->enable_bind;
1198 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1199 			if (attrs->max_ord > c4iw_max_read_depth) {
1200 				ret = -EINVAL;
1201 				goto out;
1202 			}
1203 			newattr.max_ord = attrs->max_ord;
1204 		}
1205 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1206 			if (attrs->max_ird > c4iw_max_read_depth) {
1207 				ret = -EINVAL;
1208 				goto out;
1209 			}
1210 			newattr.max_ird = attrs->max_ird;
1211 		}
1212 		qhp->attr = newattr;
1213 	}
1214 
1215 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1216 		goto out;
1217 	if (qhp->attr.state == attrs->next_state)
1218 		goto out;
1219 
1220 	switch (qhp->attr.state) {
1221 	case C4IW_QP_STATE_IDLE:
1222 		switch (attrs->next_state) {
1223 		case C4IW_QP_STATE_RTS:
1224 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1225 				ret = -EINVAL;
1226 				goto out;
1227 			}
1228 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1229 				ret = -EINVAL;
1230 				goto out;
1231 			}
1232 			qhp->attr.mpa_attr = attrs->mpa_attr;
1233 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1234 			qhp->ep = qhp->attr.llp_stream_handle;
1235 			set_state(qhp, C4IW_QP_STATE_RTS);
1236 
1237 			/*
1238 			 * Ref the endpoint here and deref when we
1239 			 * disassociate the endpoint from the QP.  This
1240 			 * happens in CLOSING->IDLE transition or *->ERROR
1241 			 * transition.
1242 			 */
1243 			c4iw_get_ep(&qhp->ep->com);
1244 			ret = rdma_init(rhp, qhp);
1245 			if (ret)
1246 				goto err;
1247 			break;
1248 		case C4IW_QP_STATE_ERROR:
1249 			set_state(qhp, C4IW_QP_STATE_ERROR);
1250 			flush_qp(qhp);
1251 			break;
1252 		default:
1253 			ret = -EINVAL;
1254 			goto out;
1255 		}
1256 		break;
1257 	case C4IW_QP_STATE_RTS:
1258 		switch (attrs->next_state) {
1259 		case C4IW_QP_STATE_CLOSING:
1260 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1261 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1262 			ep = qhp->ep;
1263 			if (!internal) {
1264 				abort = 0;
1265 				disconnect = 1;
1266 				c4iw_get_ep(&qhp->ep->com);
1267 			}
1268 			if (qhp->ibqp.uobject)
1269 				t4_set_wq_in_error(&qhp->wq);
1270 			ret = rdma_fini(rhp, qhp, ep);
1271 			if (ret)
1272 				goto err;
1273 			break;
1274 		case C4IW_QP_STATE_TERMINATE:
1275 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1276 			qhp->attr.layer_etype = attrs->layer_etype;
1277 			qhp->attr.ecode = attrs->ecode;
1278 			if (qhp->ibqp.uobject)
1279 				t4_set_wq_in_error(&qhp->wq);
1280 			ep = qhp->ep;
1281 			if (!internal)
1282 				terminate = 1;
1283 			disconnect = 1;
1284 			c4iw_get_ep(&qhp->ep->com);
1285 			break;
1286 		case C4IW_QP_STATE_ERROR:
1287 			set_state(qhp, C4IW_QP_STATE_ERROR);
1288 			if (qhp->ibqp.uobject)
1289 				t4_set_wq_in_error(&qhp->wq);
1290 			if (!internal) {
1291 				abort = 1;
1292 				disconnect = 1;
1293 				ep = qhp->ep;
1294 				c4iw_get_ep(&qhp->ep->com);
1295 			}
1296 			goto err;
1297 			break;
1298 		default:
1299 			ret = -EINVAL;
1300 			goto out;
1301 		}
1302 		break;
1303 	case C4IW_QP_STATE_CLOSING:
1304 		if (!internal) {
1305 			ret = -EINVAL;
1306 			goto out;
1307 		}
1308 		switch (attrs->next_state) {
1309 		case C4IW_QP_STATE_IDLE:
1310 			flush_qp(qhp);
1311 			set_state(qhp, C4IW_QP_STATE_IDLE);
1312 			qhp->attr.llp_stream_handle = NULL;
1313 			c4iw_put_ep(&qhp->ep->com);
1314 			qhp->ep = NULL;
1315 			wake_up(&qhp->wait);
1316 			break;
1317 		case C4IW_QP_STATE_ERROR:
1318 			goto err;
1319 		default:
1320 			ret = -EINVAL;
1321 			goto err;
1322 		}
1323 		break;
1324 	case C4IW_QP_STATE_ERROR:
1325 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1326 			ret = -EINVAL;
1327 			goto out;
1328 		}
1329 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1330 			ret = -EINVAL;
1331 			goto out;
1332 		}
1333 		set_state(qhp, C4IW_QP_STATE_IDLE);
1334 		break;
1335 	case C4IW_QP_STATE_TERMINATE:
1336 		if (!internal) {
1337 			ret = -EINVAL;
1338 			goto out;
1339 		}
1340 		goto err;
1341 		break;
1342 	default:
1343 		printf("%s in a bad state %d\n",
1344 		       __func__, qhp->attr.state);
1345 		ret = -EINVAL;
1346 		goto err;
1347 		break;
1348 	}
1349 	goto out;
1350 err:
1351 	CTR3(KTR_IW_CXGBE, "%s disassociating ep %p qpid 0x%x", __func__,
1352 	    qhp->ep, qhp->wq.sq.qid);
1353 
1354 	/* disassociate the LLP connection */
1355 	qhp->attr.llp_stream_handle = NULL;
1356 	if (!ep)
1357 		ep = qhp->ep;
1358 	qhp->ep = NULL;
1359 	set_state(qhp, C4IW_QP_STATE_ERROR);
1360 	free = 1;
1361 	BUG_ON(!ep);
1362 	flush_qp(qhp);
1363 	wake_up(&qhp->wait);
1364 out:
1365 	mutex_unlock(&qhp->mutex);
1366 
1367 	if (terminate)
1368 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1369 
1370 	/*
1371 	 * If disconnect is 1, then we need to initiate a disconnect
1372 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1373 	 * an abnormal close (RTS/CLOSING->ERROR).
1374 	 */
1375 	if (disconnect) {
1376 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1377 							 GFP_KERNEL);
1378 		c4iw_put_ep(&ep->com);
1379 	}
1380 
1381 	/*
1382 	 * If free is 1, then we've disassociated the EP from the QP
1383 	 * and we need to dereference the EP.
1384 	 */
1385 	if (free)
1386 		c4iw_put_ep(&ep->com);
1387 	CTR2(KTR_IW_CXGBE, "%s exit state %d", __func__, qhp->attr.state);
1388 	return ret;
1389 }
1390 
1391 static int enable_qp_db(int id, void *p, void *data)
1392 {
1393 	struct c4iw_qp *qp = p;
1394 
1395 	t4_enable_wq_db(&qp->wq);
1396 	return 0;
1397 }
1398 
1399 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1400 {
1401 	struct c4iw_dev *rhp;
1402 	struct c4iw_qp *qhp;
1403 	struct c4iw_qp_attributes attrs;
1404 	struct c4iw_ucontext *ucontext;
1405 
1406 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ib_qp);
1407 	qhp = to_c4iw_qp(ib_qp);
1408 	rhp = qhp->rhp;
1409 
1410 	attrs.next_state = C4IW_QP_STATE_ERROR;
1411 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1412 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1413 	else
1414 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1415 	wait_event(qhp->wait, !qhp->ep);
1416 
1417 	spin_lock_irq(&rhp->lock);
1418 	remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1419 	rhp->qpcnt--;
1420 	BUG_ON(rhp->qpcnt < 0);
1421 	if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
1422 		rhp->rdev.stats.db_state_transitions++;
1423 		rhp->db_state = NORMAL;
1424 		idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
1425 	}
1426 	spin_unlock_irq(&rhp->lock);
1427 	atomic_dec(&qhp->refcnt);
1428 	wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1429 
1430 	ucontext = ib_qp->uobject ?
1431 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1432 	destroy_qp(&rhp->rdev, &qhp->wq,
1433 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1434 
1435 	CTR3(KTR_IW_CXGBE, "%s ib_qp %p qpid 0x%0x", __func__, ib_qp,
1436 	    qhp->wq.sq.qid);
1437 	kfree(qhp);
1438 	return 0;
1439 }
1440 
1441 static int disable_qp_db(int id, void *p, void *data)
1442 {
1443 	struct c4iw_qp *qp = p;
1444 
1445 	t4_disable_wq_db(&qp->wq);
1446 	return 0;
1447 }
1448 
1449 struct ib_qp *
1450 c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1451     struct ib_udata *udata)
1452 {
1453 	struct c4iw_dev *rhp;
1454 	struct c4iw_qp *qhp;
1455 	struct c4iw_pd *php;
1456 	struct c4iw_cq *schp;
1457 	struct c4iw_cq *rchp;
1458 	struct c4iw_create_qp_resp uresp;
1459 	int sqsize, rqsize;
1460 	struct c4iw_ucontext *ucontext;
1461 	int ret;
1462 	struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4;
1463 
1464 	CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd);
1465 
1466 	if (attrs->qp_type != IB_QPT_RC)
1467 		return ERR_PTR(-EINVAL);
1468 
1469 	php = to_c4iw_pd(pd);
1470 	rhp = php->rhp;
1471 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1472 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1473 	if (!schp || !rchp)
1474 		return ERR_PTR(-EINVAL);
1475 
1476 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1477 		return ERR_PTR(-EINVAL);
1478 
1479 	rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1480 	if (rqsize > T4_MAX_RQ_SIZE)
1481 		return ERR_PTR(-E2BIG);
1482 
1483 	sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1484 	if (sqsize > T4_MAX_SQ_SIZE)
1485 		return ERR_PTR(-E2BIG);
1486 
1487 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1488 
1489 
1490 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1491 	if (!qhp)
1492 		return ERR_PTR(-ENOMEM);
1493 	qhp->wq.sq.size = sqsize;
1494 	qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1495 	qhp->wq.rq.size = rqsize;
1496 	qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1497 
1498 	if (ucontext) {
1499 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1500 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1501 	}
1502 
1503 	CTR5(KTR_IW_CXGBE, "%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu",
1504 	    __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1505 
1506 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1507 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1508 	if (ret)
1509 		goto err1;
1510 
1511 	attrs->cap.max_recv_wr = rqsize - 1;
1512 	attrs->cap.max_send_wr = sqsize - 1;
1513 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1514 
1515 	qhp->rhp = rhp;
1516 	qhp->attr.pd = php->pdid;
1517 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1518 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1519 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1520 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1521 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1522 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1523 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1524 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1525 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1526 	qhp->attr.enable_rdma_read = 1;
1527 	qhp->attr.enable_rdma_write = 1;
1528 	qhp->attr.enable_bind = 1;
1529 	qhp->attr.max_ord = 1;
1530 	qhp->attr.max_ird = 1;
1531 	spin_lock_init(&qhp->lock);
1532 	mutex_init(&qhp->mutex);
1533 	init_waitqueue_head(&qhp->wait);
1534 	atomic_set(&qhp->refcnt, 1);
1535 
1536 	spin_lock_irq(&rhp->lock);
1537 	if (rhp->db_state != NORMAL)
1538 		t4_disable_wq_db(&qhp->wq);
1539 	if (++rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
1540 		rhp->rdev.stats.db_state_transitions++;
1541 		rhp->db_state = FLOW_CONTROL;
1542 		idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
1543 	}
1544 	ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1545 	spin_unlock_irq(&rhp->lock);
1546 	if (ret)
1547 		goto err2;
1548 
1549 	if (udata) {
1550 		mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1551 		if (!mm1) {
1552 			ret = -ENOMEM;
1553 			goto err3;
1554 		}
1555 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1556 		if (!mm2) {
1557 			ret = -ENOMEM;
1558 			goto err4;
1559 		}
1560 		mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1561 		if (!mm3) {
1562 			ret = -ENOMEM;
1563 			goto err5;
1564 		}
1565 		mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1566 		if (!mm4) {
1567 			ret = -ENOMEM;
1568 			goto err6;
1569 		}
1570 		uresp.flags = 0;
1571 		uresp.qid_mask = rhp->rdev.qpmask;
1572 		uresp.sqid = qhp->wq.sq.qid;
1573 		uresp.sq_size = qhp->wq.sq.size;
1574 		uresp.sq_memsize = qhp->wq.sq.memsize;
1575 		uresp.rqid = qhp->wq.rq.qid;
1576 		uresp.rq_size = qhp->wq.rq.size;
1577 		uresp.rq_memsize = qhp->wq.rq.memsize;
1578 		spin_lock(&ucontext->mmap_lock);
1579 		uresp.sq_key = ucontext->key;
1580 		ucontext->key += PAGE_SIZE;
1581 		uresp.rq_key = ucontext->key;
1582 		ucontext->key += PAGE_SIZE;
1583 		uresp.sq_db_gts_key = ucontext->key;
1584 		ucontext->key += PAGE_SIZE;
1585 		uresp.rq_db_gts_key = ucontext->key;
1586 		ucontext->key += PAGE_SIZE;
1587 		spin_unlock(&ucontext->mmap_lock);
1588 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1589 		if (ret)
1590 			goto err7;
1591 		mm1->key = uresp.sq_key;
1592 		mm1->addr = qhp->wq.sq.phys_addr;
1593 		mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1594 		CTR4(KTR_IW_CXGBE, "%s mm1 %x, %x, %d", __func__, mm1->key,
1595 		    mm1->addr, mm1->len);
1596 		insert_mmap(ucontext, mm1);
1597 		mm2->key = uresp.rq_key;
1598 		mm2->addr = vtophys(qhp->wq.rq.queue);
1599 		mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1600 		CTR4(KTR_IW_CXGBE, "%s mm2 %x, %x, %d", __func__, mm2->key,
1601 		    mm2->addr, mm2->len);
1602 		insert_mmap(ucontext, mm2);
1603 		mm3->key = uresp.sq_db_gts_key;
1604 		mm3->addr = qhp->wq.sq.udb;
1605 		mm3->len = PAGE_SIZE;
1606 		CTR4(KTR_IW_CXGBE, "%s mm3 %x, %x, %d", __func__, mm3->key,
1607 		    mm3->addr, mm3->len);
1608 		insert_mmap(ucontext, mm3);
1609 		mm4->key = uresp.rq_db_gts_key;
1610 		mm4->addr = qhp->wq.rq.udb;
1611 		mm4->len = PAGE_SIZE;
1612 		CTR4(KTR_IW_CXGBE, "%s mm4 %x, %x, %d", __func__, mm4->key,
1613 		    mm4->addr, mm4->len);
1614 		insert_mmap(ucontext, mm4);
1615 	}
1616 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1617 	init_timer(&(qhp->timer));
1618 	CTR5(KTR_IW_CXGBE,
1619 	    "%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x",
1620 	    __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1621 	    qhp->wq.sq.qid);
1622 	return &qhp->ibqp;
1623 err7:
1624 	kfree(mm4);
1625 err6:
1626 	kfree(mm3);
1627 err5:
1628 	kfree(mm2);
1629 err4:
1630 	kfree(mm1);
1631 err3:
1632 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1633 err2:
1634 	destroy_qp(&rhp->rdev, &qhp->wq,
1635 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1636 err1:
1637 	kfree(qhp);
1638 	return ERR_PTR(ret);
1639 }
1640 
1641 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1642 		      int attr_mask, struct ib_udata *udata)
1643 {
1644 	struct c4iw_dev *rhp;
1645 	struct c4iw_qp *qhp;
1646 	enum c4iw_qp_attr_mask mask = 0;
1647 	struct c4iw_qp_attributes attrs;
1648 
1649 	CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ibqp);
1650 
1651 	/* iwarp does not support the RTR state */
1652 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1653 		attr_mask &= ~IB_QP_STATE;
1654 
1655 	/* Make sure we still have something left to do */
1656 	if (!attr_mask)
1657 		return 0;
1658 
1659 	memset(&attrs, 0, sizeof attrs);
1660 	qhp = to_c4iw_qp(ibqp);
1661 	rhp = qhp->rhp;
1662 
1663 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1664 	attrs.enable_rdma_read = (attr->qp_access_flags &
1665 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1666 	attrs.enable_rdma_write = (attr->qp_access_flags &
1667 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1668 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1669 
1670 
1671 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1672 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1673 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1674 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1675 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1676 
1677 	/*
1678 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1679 	 * ringing the queue db when we're in DB_FULL mode.
1680 	 */
1681 	attrs.sq_db_inc = attr->sq_psn;
1682 	attrs.rq_db_inc = attr->rq_psn;
1683 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1684 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1685 
1686 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1687 }
1688 
1689 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1690 {
1691 	CTR3(KTR_IW_CXGBE, "%s ib_dev %p qpn 0x%x", __func__, dev, qpn);
1692 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1693 }
1694 
1695 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1696 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1697 {
1698 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1699 
1700 	memset(attr, 0, sizeof *attr);
1701 	memset(init_attr, 0, sizeof *init_attr);
1702 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1703 	return 0;
1704 }
1705 #endif
1706