xref: /linux/drivers/infiniband/hw/cxgb4/qp.c (revision 0d456bad36d42d16022be045c8a53ddbb59ee478)
1 /*
2  * Copyright (c) 2009-2010 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 
33 #include <linux/module.h>
34 
35 #include "iw_cxgb4.h"
36 
37 static int db_delay_usecs = 1;
38 module_param(db_delay_usecs, int, 0644);
39 MODULE_PARM_DESC(db_delay_usecs, "Usecs to delay awaiting db fifo to drain");
40 
41 static int ocqp_support = 1;
42 module_param(ocqp_support, int, 0644);
43 MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
44 
45 int db_fc_threshold = 2000;
46 module_param(db_fc_threshold, int, 0644);
47 MODULE_PARM_DESC(db_fc_threshold, "QP count/threshold that triggers automatic "
48 		 "db flow control mode (default = 2000)");
49 
50 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
51 {
52 	unsigned long flag;
53 	spin_lock_irqsave(&qhp->lock, flag);
54 	qhp->attr.state = state;
55 	spin_unlock_irqrestore(&qhp->lock, flag);
56 }
57 
58 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
59 {
60 	c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
61 }
62 
63 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
64 {
65 	dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
66 			  pci_unmap_addr(sq, mapping));
67 }
68 
69 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
70 {
71 	if (t4_sq_onchip(sq))
72 		dealloc_oc_sq(rdev, sq);
73 	else
74 		dealloc_host_sq(rdev, sq);
75 }
76 
77 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
78 {
79 	if (!ocqp_support || !t4_ocqp_supported())
80 		return -ENOSYS;
81 	sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
82 	if (!sq->dma_addr)
83 		return -ENOMEM;
84 	sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
85 			rdev->lldi.vr->ocq.start;
86 	sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
87 					    rdev->lldi.vr->ocq.start);
88 	sq->flags |= T4_SQ_ONCHIP;
89 	return 0;
90 }
91 
92 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
93 {
94 	sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
95 				       &(sq->dma_addr), GFP_KERNEL);
96 	if (!sq->queue)
97 		return -ENOMEM;
98 	sq->phys_addr = virt_to_phys(sq->queue);
99 	pci_unmap_addr_set(sq, mapping, sq->dma_addr);
100 	return 0;
101 }
102 
103 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
104 		      struct c4iw_dev_ucontext *uctx)
105 {
106 	/*
107 	 * uP clears EQ contexts when the connection exits rdma mode,
108 	 * so no need to post a RESET WR for these EQs.
109 	 */
110 	dma_free_coherent(&(rdev->lldi.pdev->dev),
111 			  wq->rq.memsize, wq->rq.queue,
112 			  dma_unmap_addr(&wq->rq, mapping));
113 	dealloc_sq(rdev, &wq->sq);
114 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
115 	kfree(wq->rq.sw_rq);
116 	kfree(wq->sq.sw_sq);
117 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
118 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
119 	return 0;
120 }
121 
122 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
123 		     struct t4_cq *rcq, struct t4_cq *scq,
124 		     struct c4iw_dev_ucontext *uctx)
125 {
126 	int user = (uctx != &rdev->uctx);
127 	struct fw_ri_res_wr *res_wr;
128 	struct fw_ri_res *res;
129 	int wr_len;
130 	struct c4iw_wr_wait wr_wait;
131 	struct sk_buff *skb;
132 	int ret;
133 	int eqsize;
134 
135 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
136 	if (!wq->sq.qid)
137 		return -ENOMEM;
138 
139 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
140 	if (!wq->rq.qid) {
141 		ret = -ENOMEM;
142 		goto free_sq_qid;
143 	}
144 
145 	if (!user) {
146 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
147 				 GFP_KERNEL);
148 		if (!wq->sq.sw_sq) {
149 			ret = -ENOMEM;
150 			goto free_rq_qid;
151 		}
152 
153 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
154 				 GFP_KERNEL);
155 		if (!wq->rq.sw_rq) {
156 			ret = -ENOMEM;
157 			goto free_sw_sq;
158 		}
159 	}
160 
161 	/*
162 	 * RQT must be a power of 2.
163 	 */
164 	wq->rq.rqt_size = roundup_pow_of_two(wq->rq.size);
165 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
166 	if (!wq->rq.rqt_hwaddr) {
167 		ret = -ENOMEM;
168 		goto free_sw_rq;
169 	}
170 
171 	if (user) {
172 		ret = alloc_oc_sq(rdev, &wq->sq);
173 		if (ret)
174 			goto free_hwaddr;
175 
176 		ret = alloc_host_sq(rdev, &wq->sq);
177 		if (ret)
178 			goto free_sq;
179 	} else
180 		ret = alloc_host_sq(rdev, &wq->sq);
181 		if (ret)
182 			goto free_hwaddr;
183 	memset(wq->sq.queue, 0, wq->sq.memsize);
184 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
185 
186 	wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
187 					  wq->rq.memsize, &(wq->rq.dma_addr),
188 					  GFP_KERNEL);
189 	if (!wq->rq.queue)
190 		goto free_sq;
191 	PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
192 		__func__, wq->sq.queue,
193 		(unsigned long long)virt_to_phys(wq->sq.queue),
194 		wq->rq.queue,
195 		(unsigned long long)virt_to_phys(wq->rq.queue));
196 	memset(wq->rq.queue, 0, wq->rq.memsize);
197 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
198 
199 	wq->db = rdev->lldi.db_reg;
200 	wq->gts = rdev->lldi.gts_reg;
201 	if (user) {
202 		wq->sq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
203 					(wq->sq.qid << rdev->qpshift);
204 		wq->sq.udb &= PAGE_MASK;
205 		wq->rq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
206 					(wq->rq.qid << rdev->qpshift);
207 		wq->rq.udb &= PAGE_MASK;
208 	}
209 	wq->rdev = rdev;
210 	wq->rq.msn = 1;
211 
212 	/* build fw_ri_res_wr */
213 	wr_len = sizeof *res_wr + 2 * sizeof *res;
214 
215 	skb = alloc_skb(wr_len, GFP_KERNEL);
216 	if (!skb) {
217 		ret = -ENOMEM;
218 		goto free_dma;
219 	}
220 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
221 
222 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
223 	memset(res_wr, 0, wr_len);
224 	res_wr->op_nres = cpu_to_be32(
225 			FW_WR_OP(FW_RI_RES_WR) |
226 			V_FW_RI_RES_WR_NRES(2) |
227 			FW_WR_COMPL(1));
228 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
229 	res_wr->cookie = (unsigned long) &wr_wait;
230 	res = res_wr->res;
231 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
232 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
233 
234 	/*
235 	 * eqsize is the number of 64B entries plus the status page size.
236 	 */
237 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
238 
239 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
240 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
241 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
242 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
243 		(t4_sq_onchip(&wq->sq) ? F_FW_RI_RES_WR_ONCHIP : 0) |
244 		V_FW_RI_RES_WR_IQID(scq->cqid));
245 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
246 		V_FW_RI_RES_WR_DCAEN(0) |
247 		V_FW_RI_RES_WR_DCACPU(0) |
248 		V_FW_RI_RES_WR_FBMIN(2) |
249 		V_FW_RI_RES_WR_FBMAX(2) |
250 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
251 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
252 		V_FW_RI_RES_WR_EQSIZE(eqsize));
253 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
254 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
255 	res++;
256 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
257 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
258 
259 	/*
260 	 * eqsize is the number of 64B entries plus the status page size.
261 	 */
262 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
263 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
264 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
265 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
266 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
267 		V_FW_RI_RES_WR_IQID(rcq->cqid));
268 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
269 		V_FW_RI_RES_WR_DCAEN(0) |
270 		V_FW_RI_RES_WR_DCACPU(0) |
271 		V_FW_RI_RES_WR_FBMIN(2) |
272 		V_FW_RI_RES_WR_FBMAX(2) |
273 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
274 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
275 		V_FW_RI_RES_WR_EQSIZE(eqsize));
276 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
277 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
278 
279 	c4iw_init_wr_wait(&wr_wait);
280 
281 	ret = c4iw_ofld_send(rdev, skb);
282 	if (ret)
283 		goto free_dma;
284 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
285 	if (ret)
286 		goto free_dma;
287 
288 	PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx\n",
289 	     __func__, wq->sq.qid, wq->rq.qid, wq->db,
290 	     (unsigned long long)wq->sq.udb, (unsigned long long)wq->rq.udb);
291 
292 	return 0;
293 free_dma:
294 	dma_free_coherent(&(rdev->lldi.pdev->dev),
295 			  wq->rq.memsize, wq->rq.queue,
296 			  dma_unmap_addr(&wq->rq, mapping));
297 free_sq:
298 	dealloc_sq(rdev, &wq->sq);
299 free_hwaddr:
300 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
301 free_sw_rq:
302 	kfree(wq->rq.sw_rq);
303 free_sw_sq:
304 	kfree(wq->sq.sw_sq);
305 free_rq_qid:
306 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
307 free_sq_qid:
308 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
309 	return ret;
310 }
311 
312 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
313 		      struct ib_send_wr *wr, int max, u32 *plenp)
314 {
315 	u8 *dstp, *srcp;
316 	u32 plen = 0;
317 	int i;
318 	int rem, len;
319 
320 	dstp = (u8 *)immdp->data;
321 	for (i = 0; i < wr->num_sge; i++) {
322 		if ((plen + wr->sg_list[i].length) > max)
323 			return -EMSGSIZE;
324 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
325 		plen += wr->sg_list[i].length;
326 		rem = wr->sg_list[i].length;
327 		while (rem) {
328 			if (dstp == (u8 *)&sq->queue[sq->size])
329 				dstp = (u8 *)sq->queue;
330 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
331 				len = rem;
332 			else
333 				len = (u8 *)&sq->queue[sq->size] - dstp;
334 			memcpy(dstp, srcp, len);
335 			dstp += len;
336 			srcp += len;
337 			rem -= len;
338 		}
339 	}
340 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
341 	if (len)
342 		memset(dstp, 0, len);
343 	immdp->op = FW_RI_DATA_IMMD;
344 	immdp->r1 = 0;
345 	immdp->r2 = 0;
346 	immdp->immdlen = cpu_to_be32(plen);
347 	*plenp = plen;
348 	return 0;
349 }
350 
351 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
352 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
353 		      int num_sge, u32 *plenp)
354 
355 {
356 	int i;
357 	u32 plen = 0;
358 	__be64 *flitp = (__be64 *)isglp->sge;
359 
360 	for (i = 0; i < num_sge; i++) {
361 		if ((plen + sg_list[i].length) < plen)
362 			return -EMSGSIZE;
363 		plen += sg_list[i].length;
364 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
365 				     sg_list[i].length);
366 		if (++flitp == queue_end)
367 			flitp = queue_start;
368 		*flitp = cpu_to_be64(sg_list[i].addr);
369 		if (++flitp == queue_end)
370 			flitp = queue_start;
371 	}
372 	*flitp = (__force __be64)0;
373 	isglp->op = FW_RI_DATA_ISGL;
374 	isglp->r1 = 0;
375 	isglp->nsge = cpu_to_be16(num_sge);
376 	isglp->r2 = 0;
377 	if (plenp)
378 		*plenp = plen;
379 	return 0;
380 }
381 
382 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
383 			   struct ib_send_wr *wr, u8 *len16)
384 {
385 	u32 plen;
386 	int size;
387 	int ret;
388 
389 	if (wr->num_sge > T4_MAX_SEND_SGE)
390 		return -EINVAL;
391 	switch (wr->opcode) {
392 	case IB_WR_SEND:
393 		if (wr->send_flags & IB_SEND_SOLICITED)
394 			wqe->send.sendop_pkd = cpu_to_be32(
395 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
396 		else
397 			wqe->send.sendop_pkd = cpu_to_be32(
398 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
399 		wqe->send.stag_inv = 0;
400 		break;
401 	case IB_WR_SEND_WITH_INV:
402 		if (wr->send_flags & IB_SEND_SOLICITED)
403 			wqe->send.sendop_pkd = cpu_to_be32(
404 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
405 		else
406 			wqe->send.sendop_pkd = cpu_to_be32(
407 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
408 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
409 		break;
410 
411 	default:
412 		return -EINVAL;
413 	}
414 
415 	plen = 0;
416 	if (wr->num_sge) {
417 		if (wr->send_flags & IB_SEND_INLINE) {
418 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
419 					 T4_MAX_SEND_INLINE, &plen);
420 			if (ret)
421 				return ret;
422 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
423 			       plen;
424 		} else {
425 			ret = build_isgl((__be64 *)sq->queue,
426 					 (__be64 *)&sq->queue[sq->size],
427 					 wqe->send.u.isgl_src,
428 					 wr->sg_list, wr->num_sge, &plen);
429 			if (ret)
430 				return ret;
431 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
432 			       wr->num_sge * sizeof(struct fw_ri_sge);
433 		}
434 	} else {
435 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
436 		wqe->send.u.immd_src[0].r1 = 0;
437 		wqe->send.u.immd_src[0].r2 = 0;
438 		wqe->send.u.immd_src[0].immdlen = 0;
439 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
440 		plen = 0;
441 	}
442 	*len16 = DIV_ROUND_UP(size, 16);
443 	wqe->send.plen = cpu_to_be32(plen);
444 	return 0;
445 }
446 
447 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
448 			    struct ib_send_wr *wr, u8 *len16)
449 {
450 	u32 plen;
451 	int size;
452 	int ret;
453 
454 	if (wr->num_sge > T4_MAX_SEND_SGE)
455 		return -EINVAL;
456 	wqe->write.r2 = 0;
457 	wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
458 	wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
459 	if (wr->num_sge) {
460 		if (wr->send_flags & IB_SEND_INLINE) {
461 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
462 					 T4_MAX_WRITE_INLINE, &plen);
463 			if (ret)
464 				return ret;
465 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
466 			       plen;
467 		} else {
468 			ret = build_isgl((__be64 *)sq->queue,
469 					 (__be64 *)&sq->queue[sq->size],
470 					 wqe->write.u.isgl_src,
471 					 wr->sg_list, wr->num_sge, &plen);
472 			if (ret)
473 				return ret;
474 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
475 			       wr->num_sge * sizeof(struct fw_ri_sge);
476 		}
477 	} else {
478 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
479 		wqe->write.u.immd_src[0].r1 = 0;
480 		wqe->write.u.immd_src[0].r2 = 0;
481 		wqe->write.u.immd_src[0].immdlen = 0;
482 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
483 		plen = 0;
484 	}
485 	*len16 = DIV_ROUND_UP(size, 16);
486 	wqe->write.plen = cpu_to_be32(plen);
487 	return 0;
488 }
489 
490 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
491 {
492 	if (wr->num_sge > 1)
493 		return -EINVAL;
494 	if (wr->num_sge) {
495 		wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
496 		wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
497 							>> 32));
498 		wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
499 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
500 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
501 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
502 							 >> 32));
503 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
504 	} else {
505 		wqe->read.stag_src = cpu_to_be32(2);
506 		wqe->read.to_src_hi = 0;
507 		wqe->read.to_src_lo = 0;
508 		wqe->read.stag_sink = cpu_to_be32(2);
509 		wqe->read.plen = 0;
510 		wqe->read.to_sink_hi = 0;
511 		wqe->read.to_sink_lo = 0;
512 	}
513 	wqe->read.r2 = 0;
514 	wqe->read.r5 = 0;
515 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
516 	return 0;
517 }
518 
519 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
520 			   struct ib_recv_wr *wr, u8 *len16)
521 {
522 	int ret;
523 
524 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
525 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
526 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
527 	if (ret)
528 		return ret;
529 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
530 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
531 	return 0;
532 }
533 
534 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
535 			 struct ib_send_wr *wr, u8 *len16)
536 {
537 
538 	struct fw_ri_immd *imdp;
539 	__be64 *p;
540 	int i;
541 	int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
542 	int rem;
543 
544 	if (wr->wr.fast_reg.page_list_len > T4_MAX_FR_DEPTH)
545 		return -EINVAL;
546 
547 	wqe->fr.qpbinde_to_dcacpu = 0;
548 	wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
549 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
550 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
551 	wqe->fr.len_hi = 0;
552 	wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
553 	wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
554 	wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
555 	wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
556 					0xffffffff);
557 	WARN_ON(pbllen > T4_MAX_FR_IMMD);
558 	imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
559 	imdp->op = FW_RI_DATA_IMMD;
560 	imdp->r1 = 0;
561 	imdp->r2 = 0;
562 	imdp->immdlen = cpu_to_be32(pbllen);
563 	p = (__be64 *)(imdp + 1);
564 	rem = pbllen;
565 	for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
566 		*p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
567 		rem -= sizeof *p;
568 		if (++p == (__be64 *)&sq->queue[sq->size])
569 			p = (__be64 *)sq->queue;
570 	}
571 	BUG_ON(rem < 0);
572 	while (rem) {
573 		*p = 0;
574 		rem -= sizeof *p;
575 		if (++p == (__be64 *)&sq->queue[sq->size])
576 			p = (__be64 *)sq->queue;
577 	}
578 	*len16 = DIV_ROUND_UP(sizeof wqe->fr + sizeof *imdp + pbllen, 16);
579 	return 0;
580 }
581 
582 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
583 			  u8 *len16)
584 {
585 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
586 	wqe->inv.r2 = 0;
587 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
588 	return 0;
589 }
590 
591 void c4iw_qp_add_ref(struct ib_qp *qp)
592 {
593 	PDBG("%s ib_qp %p\n", __func__, qp);
594 	atomic_inc(&(to_c4iw_qp(qp)->refcnt));
595 }
596 
597 void c4iw_qp_rem_ref(struct ib_qp *qp)
598 {
599 	PDBG("%s ib_qp %p\n", __func__, qp);
600 	if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
601 		wake_up(&(to_c4iw_qp(qp)->wait));
602 }
603 
604 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
605 		   struct ib_send_wr **bad_wr)
606 {
607 	int err = 0;
608 	u8 len16 = 0;
609 	enum fw_wr_opcodes fw_opcode = 0;
610 	enum fw_ri_wr_flags fw_flags;
611 	struct c4iw_qp *qhp;
612 	union t4_wr *wqe;
613 	u32 num_wrs;
614 	struct t4_swsqe *swsqe;
615 	unsigned long flag;
616 	u16 idx = 0;
617 
618 	qhp = to_c4iw_qp(ibqp);
619 	spin_lock_irqsave(&qhp->lock, flag);
620 	if (t4_wq_in_error(&qhp->wq)) {
621 		spin_unlock_irqrestore(&qhp->lock, flag);
622 		return -EINVAL;
623 	}
624 	num_wrs = t4_sq_avail(&qhp->wq);
625 	if (num_wrs == 0) {
626 		spin_unlock_irqrestore(&qhp->lock, flag);
627 		return -ENOMEM;
628 	}
629 	while (wr) {
630 		if (num_wrs == 0) {
631 			err = -ENOMEM;
632 			*bad_wr = wr;
633 			break;
634 		}
635 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
636 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
637 
638 		fw_flags = 0;
639 		if (wr->send_flags & IB_SEND_SOLICITED)
640 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
641 		if (wr->send_flags & IB_SEND_SIGNALED)
642 			fw_flags |= FW_RI_COMPLETION_FLAG;
643 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
644 		switch (wr->opcode) {
645 		case IB_WR_SEND_WITH_INV:
646 		case IB_WR_SEND:
647 			if (wr->send_flags & IB_SEND_FENCE)
648 				fw_flags |= FW_RI_READ_FENCE_FLAG;
649 			fw_opcode = FW_RI_SEND_WR;
650 			if (wr->opcode == IB_WR_SEND)
651 				swsqe->opcode = FW_RI_SEND;
652 			else
653 				swsqe->opcode = FW_RI_SEND_WITH_INV;
654 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
655 			break;
656 		case IB_WR_RDMA_WRITE:
657 			fw_opcode = FW_RI_RDMA_WRITE_WR;
658 			swsqe->opcode = FW_RI_RDMA_WRITE;
659 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
660 			break;
661 		case IB_WR_RDMA_READ:
662 		case IB_WR_RDMA_READ_WITH_INV:
663 			fw_opcode = FW_RI_RDMA_READ_WR;
664 			swsqe->opcode = FW_RI_READ_REQ;
665 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
666 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
667 			else
668 				fw_flags = 0;
669 			err = build_rdma_read(wqe, wr, &len16);
670 			if (err)
671 				break;
672 			swsqe->read_len = wr->sg_list[0].length;
673 			if (!qhp->wq.sq.oldest_read)
674 				qhp->wq.sq.oldest_read = swsqe;
675 			break;
676 		case IB_WR_FAST_REG_MR:
677 			fw_opcode = FW_RI_FR_NSMR_WR;
678 			swsqe->opcode = FW_RI_FAST_REGISTER;
679 			err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16);
680 			break;
681 		case IB_WR_LOCAL_INV:
682 			if (wr->send_flags & IB_SEND_FENCE)
683 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
684 			fw_opcode = FW_RI_INV_LSTAG_WR;
685 			swsqe->opcode = FW_RI_LOCAL_INV;
686 			err = build_inv_stag(wqe, wr, &len16);
687 			break;
688 		default:
689 			PDBG("%s post of type=%d TBD!\n", __func__,
690 			     wr->opcode);
691 			err = -EINVAL;
692 		}
693 		if (err) {
694 			*bad_wr = wr;
695 			break;
696 		}
697 		swsqe->idx = qhp->wq.sq.pidx;
698 		swsqe->complete = 0;
699 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED);
700 		swsqe->wr_id = wr->wr_id;
701 
702 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
703 
704 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
705 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
706 		     swsqe->opcode, swsqe->read_len);
707 		wr = wr->next;
708 		num_wrs--;
709 		t4_sq_produce(&qhp->wq, len16);
710 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
711 	}
712 	if (t4_wq_db_enabled(&qhp->wq))
713 		t4_ring_sq_db(&qhp->wq, idx);
714 	spin_unlock_irqrestore(&qhp->lock, flag);
715 	return err;
716 }
717 
718 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
719 		      struct ib_recv_wr **bad_wr)
720 {
721 	int err = 0;
722 	struct c4iw_qp *qhp;
723 	union t4_recv_wr *wqe;
724 	u32 num_wrs;
725 	u8 len16 = 0;
726 	unsigned long flag;
727 	u16 idx = 0;
728 
729 	qhp = to_c4iw_qp(ibqp);
730 	spin_lock_irqsave(&qhp->lock, flag);
731 	if (t4_wq_in_error(&qhp->wq)) {
732 		spin_unlock_irqrestore(&qhp->lock, flag);
733 		return -EINVAL;
734 	}
735 	num_wrs = t4_rq_avail(&qhp->wq);
736 	if (num_wrs == 0) {
737 		spin_unlock_irqrestore(&qhp->lock, flag);
738 		return -ENOMEM;
739 	}
740 	while (wr) {
741 		if (wr->num_sge > T4_MAX_RECV_SGE) {
742 			err = -EINVAL;
743 			*bad_wr = wr;
744 			break;
745 		}
746 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
747 					   qhp->wq.rq.wq_pidx *
748 					   T4_EQ_ENTRY_SIZE);
749 		if (num_wrs)
750 			err = build_rdma_recv(qhp, wqe, wr, &len16);
751 		else
752 			err = -ENOMEM;
753 		if (err) {
754 			*bad_wr = wr;
755 			break;
756 		}
757 
758 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
759 
760 		wqe->recv.opcode = FW_RI_RECV_WR;
761 		wqe->recv.r1 = 0;
762 		wqe->recv.wrid = qhp->wq.rq.pidx;
763 		wqe->recv.r2[0] = 0;
764 		wqe->recv.r2[1] = 0;
765 		wqe->recv.r2[2] = 0;
766 		wqe->recv.len16 = len16;
767 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
768 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
769 		t4_rq_produce(&qhp->wq, len16);
770 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
771 		wr = wr->next;
772 		num_wrs--;
773 	}
774 	if (t4_wq_db_enabled(&qhp->wq))
775 		t4_ring_rq_db(&qhp->wq, idx);
776 	spin_unlock_irqrestore(&qhp->lock, flag);
777 	return err;
778 }
779 
780 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
781 {
782 	return -ENOSYS;
783 }
784 
785 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
786 				    u8 *ecode)
787 {
788 	int status;
789 	int tagged;
790 	int opcode;
791 	int rqtype;
792 	int send_inv;
793 
794 	if (!err_cqe) {
795 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
796 		*ecode = 0;
797 		return;
798 	}
799 
800 	status = CQE_STATUS(err_cqe);
801 	opcode = CQE_OPCODE(err_cqe);
802 	rqtype = RQ_TYPE(err_cqe);
803 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
804 		   (opcode == FW_RI_SEND_WITH_SE_INV);
805 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
806 		 (rqtype && (opcode == FW_RI_READ_RESP));
807 
808 	switch (status) {
809 	case T4_ERR_STAG:
810 		if (send_inv) {
811 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
812 			*ecode = RDMAP_CANT_INV_STAG;
813 		} else {
814 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
815 			*ecode = RDMAP_INV_STAG;
816 		}
817 		break;
818 	case T4_ERR_PDID:
819 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
820 		if ((opcode == FW_RI_SEND_WITH_INV) ||
821 		    (opcode == FW_RI_SEND_WITH_SE_INV))
822 			*ecode = RDMAP_CANT_INV_STAG;
823 		else
824 			*ecode = RDMAP_STAG_NOT_ASSOC;
825 		break;
826 	case T4_ERR_QPID:
827 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
828 		*ecode = RDMAP_STAG_NOT_ASSOC;
829 		break;
830 	case T4_ERR_ACCESS:
831 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
832 		*ecode = RDMAP_ACC_VIOL;
833 		break;
834 	case T4_ERR_WRAP:
835 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
836 		*ecode = RDMAP_TO_WRAP;
837 		break;
838 	case T4_ERR_BOUND:
839 		if (tagged) {
840 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
841 			*ecode = DDPT_BASE_BOUNDS;
842 		} else {
843 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
844 			*ecode = RDMAP_BASE_BOUNDS;
845 		}
846 		break;
847 	case T4_ERR_INVALIDATE_SHARED_MR:
848 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
849 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
850 		*ecode = RDMAP_CANT_INV_STAG;
851 		break;
852 	case T4_ERR_ECC:
853 	case T4_ERR_ECC_PSTAG:
854 	case T4_ERR_INTERNAL_ERR:
855 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
856 		*ecode = 0;
857 		break;
858 	case T4_ERR_OUT_OF_RQE:
859 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
860 		*ecode = DDPU_INV_MSN_NOBUF;
861 		break;
862 	case T4_ERR_PBL_ADDR_BOUND:
863 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
864 		*ecode = DDPT_BASE_BOUNDS;
865 		break;
866 	case T4_ERR_CRC:
867 		*layer_type = LAYER_MPA|DDP_LLP;
868 		*ecode = MPA_CRC_ERR;
869 		break;
870 	case T4_ERR_MARKER:
871 		*layer_type = LAYER_MPA|DDP_LLP;
872 		*ecode = MPA_MARKER_ERR;
873 		break;
874 	case T4_ERR_PDU_LEN_ERR:
875 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
876 		*ecode = DDPU_MSG_TOOBIG;
877 		break;
878 	case T4_ERR_DDP_VERSION:
879 		if (tagged) {
880 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
881 			*ecode = DDPT_INV_VERS;
882 		} else {
883 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
884 			*ecode = DDPU_INV_VERS;
885 		}
886 		break;
887 	case T4_ERR_RDMA_VERSION:
888 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
889 		*ecode = RDMAP_INV_VERS;
890 		break;
891 	case T4_ERR_OPCODE:
892 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
893 		*ecode = RDMAP_INV_OPCODE;
894 		break;
895 	case T4_ERR_DDP_QUEUE_NUM:
896 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
897 		*ecode = DDPU_INV_QN;
898 		break;
899 	case T4_ERR_MSN:
900 	case T4_ERR_MSN_GAP:
901 	case T4_ERR_MSN_RANGE:
902 	case T4_ERR_IRD_OVERFLOW:
903 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
904 		*ecode = DDPU_INV_MSN_RANGE;
905 		break;
906 	case T4_ERR_TBIT:
907 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
908 		*ecode = 0;
909 		break;
910 	case T4_ERR_MO:
911 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
912 		*ecode = DDPU_INV_MO;
913 		break;
914 	default:
915 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
916 		*ecode = 0;
917 		break;
918 	}
919 }
920 
921 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
922 			   gfp_t gfp)
923 {
924 	struct fw_ri_wr *wqe;
925 	struct sk_buff *skb;
926 	struct terminate_message *term;
927 
928 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
929 	     qhp->ep->hwtid);
930 
931 	skb = alloc_skb(sizeof *wqe, gfp);
932 	if (!skb)
933 		return;
934 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
935 
936 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
937 	memset(wqe, 0, sizeof *wqe);
938 	wqe->op_compl = cpu_to_be32(FW_WR_OP(FW_RI_INIT_WR));
939 	wqe->flowid_len16 = cpu_to_be32(
940 		FW_WR_FLOWID(qhp->ep->hwtid) |
941 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
942 
943 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
944 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
945 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
946 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
947 		term->layer_etype = qhp->attr.layer_etype;
948 		term->ecode = qhp->attr.ecode;
949 	} else
950 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
951 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
952 }
953 
954 /*
955  * Assumes qhp lock is held.
956  */
957 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
958 		       struct c4iw_cq *schp)
959 {
960 	int count;
961 	int flushed;
962 	unsigned long flag;
963 
964 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
965 
966 	/* locking hierarchy: cq lock first, then qp lock. */
967 	spin_lock_irqsave(&rchp->lock, flag);
968 	spin_lock(&qhp->lock);
969 	c4iw_flush_hw_cq(&rchp->cq);
970 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
971 	flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
972 	spin_unlock(&qhp->lock);
973 	spin_unlock_irqrestore(&rchp->lock, flag);
974 	if (flushed) {
975 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
976 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
977 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
978 	}
979 
980 	/* locking hierarchy: cq lock first, then qp lock. */
981 	spin_lock_irqsave(&schp->lock, flag);
982 	spin_lock(&qhp->lock);
983 	c4iw_flush_hw_cq(&schp->cq);
984 	c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
985 	flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
986 	spin_unlock(&qhp->lock);
987 	spin_unlock_irqrestore(&schp->lock, flag);
988 	if (flushed) {
989 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
990 		(*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
991 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
992 	}
993 }
994 
995 static void flush_qp(struct c4iw_qp *qhp)
996 {
997 	struct c4iw_cq *rchp, *schp;
998 	unsigned long flag;
999 
1000 	rchp = get_chp(qhp->rhp, qhp->attr.rcq);
1001 	schp = get_chp(qhp->rhp, qhp->attr.scq);
1002 
1003 	if (qhp->ibqp.uobject) {
1004 		t4_set_wq_in_error(&qhp->wq);
1005 		t4_set_cq_in_error(&rchp->cq);
1006 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1007 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1008 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1009 		if (schp != rchp) {
1010 			t4_set_cq_in_error(&schp->cq);
1011 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1012 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1013 					schp->ibcq.cq_context);
1014 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1015 		}
1016 		return;
1017 	}
1018 	__flush_qp(qhp, rchp, schp);
1019 }
1020 
1021 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1022 		     struct c4iw_ep *ep)
1023 {
1024 	struct fw_ri_wr *wqe;
1025 	int ret;
1026 	struct sk_buff *skb;
1027 
1028 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1029 	     ep->hwtid);
1030 
1031 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1032 	if (!skb)
1033 		return -ENOMEM;
1034 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1035 
1036 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1037 	memset(wqe, 0, sizeof *wqe);
1038 	wqe->op_compl = cpu_to_be32(
1039 		FW_WR_OP(FW_RI_INIT_WR) |
1040 		FW_WR_COMPL(1));
1041 	wqe->flowid_len16 = cpu_to_be32(
1042 		FW_WR_FLOWID(ep->hwtid) |
1043 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1044 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1045 
1046 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1047 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1048 	if (ret)
1049 		goto out;
1050 
1051 	ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1052 			     qhp->wq.sq.qid, __func__);
1053 out:
1054 	PDBG("%s ret %d\n", __func__, ret);
1055 	return ret;
1056 }
1057 
1058 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1059 {
1060 	PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1061 	memset(&init->u, 0, sizeof init->u);
1062 	switch (p2p_type) {
1063 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1064 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1065 		init->u.write.stag_sink = cpu_to_be32(1);
1066 		init->u.write.to_sink = cpu_to_be64(1);
1067 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1068 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1069 						   sizeof(struct fw_ri_immd),
1070 						   16);
1071 		break;
1072 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1073 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1074 		init->u.read.stag_src = cpu_to_be32(1);
1075 		init->u.read.to_src_lo = cpu_to_be32(1);
1076 		init->u.read.stag_sink = cpu_to_be32(1);
1077 		init->u.read.to_sink_lo = cpu_to_be32(1);
1078 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1079 		break;
1080 	}
1081 }
1082 
1083 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1084 {
1085 	struct fw_ri_wr *wqe;
1086 	int ret;
1087 	struct sk_buff *skb;
1088 
1089 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1090 	     qhp->ep->hwtid);
1091 
1092 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1093 	if (!skb)
1094 		return -ENOMEM;
1095 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1096 
1097 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1098 	memset(wqe, 0, sizeof *wqe);
1099 	wqe->op_compl = cpu_to_be32(
1100 		FW_WR_OP(FW_RI_INIT_WR) |
1101 		FW_WR_COMPL(1));
1102 	wqe->flowid_len16 = cpu_to_be32(
1103 		FW_WR_FLOWID(qhp->ep->hwtid) |
1104 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1105 
1106 	wqe->cookie = (unsigned long) &qhp->ep->com.wr_wait;
1107 
1108 	wqe->u.init.type = FW_RI_TYPE_INIT;
1109 	wqe->u.init.mpareqbit_p2ptype =
1110 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1111 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1112 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1113 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1114 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1115 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1116 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1117 	if (qhp->attr.mpa_attr.crc_enabled)
1118 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1119 
1120 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1121 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1122 			    FW_RI_QP_BIND_ENABLE;
1123 	if (!qhp->ibqp.uobject)
1124 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1125 				     FW_RI_QP_STAG0_ENABLE;
1126 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1127 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1128 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1129 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1130 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1131 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1132 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1133 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1134 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1135 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1136 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1137 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1138 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1139 					 rhp->rdev.lldi.vr->rq.start);
1140 	if (qhp->attr.mpa_attr.initiator)
1141 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1142 
1143 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1144 	if (ret)
1145 		goto out;
1146 
1147 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1148 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1149 out:
1150 	PDBG("%s ret %d\n", __func__, ret);
1151 	return ret;
1152 }
1153 
1154 /*
1155  * Called by the library when the qp has user dbs disabled due to
1156  * a DB_FULL condition.  This function will single-thread all user
1157  * DB rings to avoid overflowing the hw db-fifo.
1158  */
1159 static int ring_kernel_db(struct c4iw_qp *qhp, u32 qid, u16 inc)
1160 {
1161 	int delay = db_delay_usecs;
1162 
1163 	mutex_lock(&qhp->rhp->db_mutex);
1164 	do {
1165 
1166 		/*
1167 		 * The interrupt threshold is dbfifo_int_thresh << 6. So
1168 		 * make sure we don't cross that and generate an interrupt.
1169 		 */
1170 		if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) <
1171 		    (qhp->rhp->rdev.lldi.dbfifo_int_thresh << 5)) {
1172 			writel(QID(qid) | PIDX(inc), qhp->wq.db);
1173 			break;
1174 		}
1175 		set_current_state(TASK_UNINTERRUPTIBLE);
1176 		schedule_timeout(usecs_to_jiffies(delay));
1177 		delay = min(delay << 1, 2000);
1178 	} while (1);
1179 	mutex_unlock(&qhp->rhp->db_mutex);
1180 	return 0;
1181 }
1182 
1183 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1184 		   enum c4iw_qp_attr_mask mask,
1185 		   struct c4iw_qp_attributes *attrs,
1186 		   int internal)
1187 {
1188 	int ret = 0;
1189 	struct c4iw_qp_attributes newattr = qhp->attr;
1190 	int disconnect = 0;
1191 	int terminate = 0;
1192 	int abort = 0;
1193 	int free = 0;
1194 	struct c4iw_ep *ep = NULL;
1195 
1196 	PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1197 	     qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1198 	     (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1199 
1200 	mutex_lock(&qhp->mutex);
1201 
1202 	/* Process attr changes if in IDLE */
1203 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1204 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1205 			ret = -EIO;
1206 			goto out;
1207 		}
1208 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1209 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1210 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1211 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1212 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1213 			newattr.enable_bind = attrs->enable_bind;
1214 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1215 			if (attrs->max_ord > c4iw_max_read_depth) {
1216 				ret = -EINVAL;
1217 				goto out;
1218 			}
1219 			newattr.max_ord = attrs->max_ord;
1220 		}
1221 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1222 			if (attrs->max_ird > c4iw_max_read_depth) {
1223 				ret = -EINVAL;
1224 				goto out;
1225 			}
1226 			newattr.max_ird = attrs->max_ird;
1227 		}
1228 		qhp->attr = newattr;
1229 	}
1230 
1231 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1232 		ret = ring_kernel_db(qhp, qhp->wq.sq.qid, attrs->sq_db_inc);
1233 		goto out;
1234 	}
1235 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1236 		ret = ring_kernel_db(qhp, qhp->wq.rq.qid, attrs->rq_db_inc);
1237 		goto out;
1238 	}
1239 
1240 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1241 		goto out;
1242 	if (qhp->attr.state == attrs->next_state)
1243 		goto out;
1244 
1245 	switch (qhp->attr.state) {
1246 	case C4IW_QP_STATE_IDLE:
1247 		switch (attrs->next_state) {
1248 		case C4IW_QP_STATE_RTS:
1249 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1250 				ret = -EINVAL;
1251 				goto out;
1252 			}
1253 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1254 				ret = -EINVAL;
1255 				goto out;
1256 			}
1257 			qhp->attr.mpa_attr = attrs->mpa_attr;
1258 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1259 			qhp->ep = qhp->attr.llp_stream_handle;
1260 			set_state(qhp, C4IW_QP_STATE_RTS);
1261 
1262 			/*
1263 			 * Ref the endpoint here and deref when we
1264 			 * disassociate the endpoint from the QP.  This
1265 			 * happens in CLOSING->IDLE transition or *->ERROR
1266 			 * transition.
1267 			 */
1268 			c4iw_get_ep(&qhp->ep->com);
1269 			ret = rdma_init(rhp, qhp);
1270 			if (ret)
1271 				goto err;
1272 			break;
1273 		case C4IW_QP_STATE_ERROR:
1274 			set_state(qhp, C4IW_QP_STATE_ERROR);
1275 			flush_qp(qhp);
1276 			break;
1277 		default:
1278 			ret = -EINVAL;
1279 			goto out;
1280 		}
1281 		break;
1282 	case C4IW_QP_STATE_RTS:
1283 		switch (attrs->next_state) {
1284 		case C4IW_QP_STATE_CLOSING:
1285 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1286 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1287 			ep = qhp->ep;
1288 			if (!internal) {
1289 				abort = 0;
1290 				disconnect = 1;
1291 				c4iw_get_ep(&qhp->ep->com);
1292 			}
1293 			if (qhp->ibqp.uobject)
1294 				t4_set_wq_in_error(&qhp->wq);
1295 			ret = rdma_fini(rhp, qhp, ep);
1296 			if (ret)
1297 				goto err;
1298 			break;
1299 		case C4IW_QP_STATE_TERMINATE:
1300 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1301 			qhp->attr.layer_etype = attrs->layer_etype;
1302 			qhp->attr.ecode = attrs->ecode;
1303 			if (qhp->ibqp.uobject)
1304 				t4_set_wq_in_error(&qhp->wq);
1305 			ep = qhp->ep;
1306 			if (!internal)
1307 				terminate = 1;
1308 			disconnect = 1;
1309 			c4iw_get_ep(&qhp->ep->com);
1310 			break;
1311 		case C4IW_QP_STATE_ERROR:
1312 			set_state(qhp, C4IW_QP_STATE_ERROR);
1313 			if (qhp->ibqp.uobject)
1314 				t4_set_wq_in_error(&qhp->wq);
1315 			if (!internal) {
1316 				abort = 1;
1317 				disconnect = 1;
1318 				ep = qhp->ep;
1319 				c4iw_get_ep(&qhp->ep->com);
1320 			}
1321 			goto err;
1322 			break;
1323 		default:
1324 			ret = -EINVAL;
1325 			goto out;
1326 		}
1327 		break;
1328 	case C4IW_QP_STATE_CLOSING:
1329 		if (!internal) {
1330 			ret = -EINVAL;
1331 			goto out;
1332 		}
1333 		switch (attrs->next_state) {
1334 		case C4IW_QP_STATE_IDLE:
1335 			flush_qp(qhp);
1336 			set_state(qhp, C4IW_QP_STATE_IDLE);
1337 			qhp->attr.llp_stream_handle = NULL;
1338 			c4iw_put_ep(&qhp->ep->com);
1339 			qhp->ep = NULL;
1340 			wake_up(&qhp->wait);
1341 			break;
1342 		case C4IW_QP_STATE_ERROR:
1343 			goto err;
1344 		default:
1345 			ret = -EINVAL;
1346 			goto err;
1347 		}
1348 		break;
1349 	case C4IW_QP_STATE_ERROR:
1350 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1351 			ret = -EINVAL;
1352 			goto out;
1353 		}
1354 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1355 			ret = -EINVAL;
1356 			goto out;
1357 		}
1358 		set_state(qhp, C4IW_QP_STATE_IDLE);
1359 		break;
1360 	case C4IW_QP_STATE_TERMINATE:
1361 		if (!internal) {
1362 			ret = -EINVAL;
1363 			goto out;
1364 		}
1365 		goto err;
1366 		break;
1367 	default:
1368 		printk(KERN_ERR "%s in a bad state %d\n",
1369 		       __func__, qhp->attr.state);
1370 		ret = -EINVAL;
1371 		goto err;
1372 		break;
1373 	}
1374 	goto out;
1375 err:
1376 	PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1377 	     qhp->wq.sq.qid);
1378 
1379 	/* disassociate the LLP connection */
1380 	qhp->attr.llp_stream_handle = NULL;
1381 	if (!ep)
1382 		ep = qhp->ep;
1383 	qhp->ep = NULL;
1384 	set_state(qhp, C4IW_QP_STATE_ERROR);
1385 	free = 1;
1386 	wake_up(&qhp->wait);
1387 	BUG_ON(!ep);
1388 	flush_qp(qhp);
1389 out:
1390 	mutex_unlock(&qhp->mutex);
1391 
1392 	if (terminate)
1393 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1394 
1395 	/*
1396 	 * If disconnect is 1, then we need to initiate a disconnect
1397 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1398 	 * an abnormal close (RTS/CLOSING->ERROR).
1399 	 */
1400 	if (disconnect) {
1401 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1402 							 GFP_KERNEL);
1403 		c4iw_put_ep(&ep->com);
1404 	}
1405 
1406 	/*
1407 	 * If free is 1, then we've disassociated the EP from the QP
1408 	 * and we need to dereference the EP.
1409 	 */
1410 	if (free)
1411 		c4iw_put_ep(&ep->com);
1412 	PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1413 	return ret;
1414 }
1415 
1416 static int enable_qp_db(int id, void *p, void *data)
1417 {
1418 	struct c4iw_qp *qp = p;
1419 
1420 	t4_enable_wq_db(&qp->wq);
1421 	return 0;
1422 }
1423 
1424 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1425 {
1426 	struct c4iw_dev *rhp;
1427 	struct c4iw_qp *qhp;
1428 	struct c4iw_qp_attributes attrs;
1429 	struct c4iw_ucontext *ucontext;
1430 
1431 	qhp = to_c4iw_qp(ib_qp);
1432 	rhp = qhp->rhp;
1433 
1434 	attrs.next_state = C4IW_QP_STATE_ERROR;
1435 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1436 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1437 	else
1438 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1439 	wait_event(qhp->wait, !qhp->ep);
1440 
1441 	spin_lock_irq(&rhp->lock);
1442 	remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1443 	rhp->qpcnt--;
1444 	BUG_ON(rhp->qpcnt < 0);
1445 	if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
1446 		rhp->rdev.stats.db_state_transitions++;
1447 		rhp->db_state = NORMAL;
1448 		idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
1449 	}
1450 	spin_unlock_irq(&rhp->lock);
1451 	atomic_dec(&qhp->refcnt);
1452 	wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1453 
1454 	ucontext = ib_qp->uobject ?
1455 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1456 	destroy_qp(&rhp->rdev, &qhp->wq,
1457 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1458 
1459 	PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1460 	kfree(qhp);
1461 	return 0;
1462 }
1463 
1464 static int disable_qp_db(int id, void *p, void *data)
1465 {
1466 	struct c4iw_qp *qp = p;
1467 
1468 	t4_disable_wq_db(&qp->wq);
1469 	return 0;
1470 }
1471 
1472 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1473 			     struct ib_udata *udata)
1474 {
1475 	struct c4iw_dev *rhp;
1476 	struct c4iw_qp *qhp;
1477 	struct c4iw_pd *php;
1478 	struct c4iw_cq *schp;
1479 	struct c4iw_cq *rchp;
1480 	struct c4iw_create_qp_resp uresp;
1481 	int sqsize, rqsize;
1482 	struct c4iw_ucontext *ucontext;
1483 	int ret;
1484 	struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1485 
1486 	PDBG("%s ib_pd %p\n", __func__, pd);
1487 
1488 	if (attrs->qp_type != IB_QPT_RC)
1489 		return ERR_PTR(-EINVAL);
1490 
1491 	php = to_c4iw_pd(pd);
1492 	rhp = php->rhp;
1493 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1494 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1495 	if (!schp || !rchp)
1496 		return ERR_PTR(-EINVAL);
1497 
1498 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1499 		return ERR_PTR(-EINVAL);
1500 
1501 	rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1502 	if (rqsize > T4_MAX_RQ_SIZE)
1503 		return ERR_PTR(-E2BIG);
1504 
1505 	sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1506 	if (sqsize > T4_MAX_SQ_SIZE)
1507 		return ERR_PTR(-E2BIG);
1508 
1509 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1510 
1511 
1512 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1513 	if (!qhp)
1514 		return ERR_PTR(-ENOMEM);
1515 	qhp->wq.sq.size = sqsize;
1516 	qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1517 	qhp->wq.rq.size = rqsize;
1518 	qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1519 
1520 	if (ucontext) {
1521 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1522 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1523 	}
1524 
1525 	PDBG("%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu\n",
1526 	     __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1527 
1528 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1529 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1530 	if (ret)
1531 		goto err1;
1532 
1533 	attrs->cap.max_recv_wr = rqsize - 1;
1534 	attrs->cap.max_send_wr = sqsize - 1;
1535 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1536 
1537 	qhp->rhp = rhp;
1538 	qhp->attr.pd = php->pdid;
1539 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1540 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1541 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1542 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1543 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1544 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1545 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1546 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1547 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1548 	qhp->attr.enable_rdma_read = 1;
1549 	qhp->attr.enable_rdma_write = 1;
1550 	qhp->attr.enable_bind = 1;
1551 	qhp->attr.max_ord = 1;
1552 	qhp->attr.max_ird = 1;
1553 	spin_lock_init(&qhp->lock);
1554 	mutex_init(&qhp->mutex);
1555 	init_waitqueue_head(&qhp->wait);
1556 	atomic_set(&qhp->refcnt, 1);
1557 
1558 	spin_lock_irq(&rhp->lock);
1559 	if (rhp->db_state != NORMAL)
1560 		t4_disable_wq_db(&qhp->wq);
1561 	if (++rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
1562 		rhp->rdev.stats.db_state_transitions++;
1563 		rhp->db_state = FLOW_CONTROL;
1564 		idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
1565 	}
1566 	ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1567 	spin_unlock_irq(&rhp->lock);
1568 	if (ret)
1569 		goto err2;
1570 
1571 	if (udata) {
1572 		mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1573 		if (!mm1) {
1574 			ret = -ENOMEM;
1575 			goto err3;
1576 		}
1577 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1578 		if (!mm2) {
1579 			ret = -ENOMEM;
1580 			goto err4;
1581 		}
1582 		mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1583 		if (!mm3) {
1584 			ret = -ENOMEM;
1585 			goto err5;
1586 		}
1587 		mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1588 		if (!mm4) {
1589 			ret = -ENOMEM;
1590 			goto err6;
1591 		}
1592 		if (t4_sq_onchip(&qhp->wq.sq)) {
1593 			mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1594 			if (!mm5) {
1595 				ret = -ENOMEM;
1596 				goto err7;
1597 			}
1598 			uresp.flags = C4IW_QPF_ONCHIP;
1599 		} else
1600 			uresp.flags = 0;
1601 		uresp.qid_mask = rhp->rdev.qpmask;
1602 		uresp.sqid = qhp->wq.sq.qid;
1603 		uresp.sq_size = qhp->wq.sq.size;
1604 		uresp.sq_memsize = qhp->wq.sq.memsize;
1605 		uresp.rqid = qhp->wq.rq.qid;
1606 		uresp.rq_size = qhp->wq.rq.size;
1607 		uresp.rq_memsize = qhp->wq.rq.memsize;
1608 		spin_lock(&ucontext->mmap_lock);
1609 		if (mm5) {
1610 			uresp.ma_sync_key = ucontext->key;
1611 			ucontext->key += PAGE_SIZE;
1612 		}
1613 		uresp.sq_key = ucontext->key;
1614 		ucontext->key += PAGE_SIZE;
1615 		uresp.rq_key = ucontext->key;
1616 		ucontext->key += PAGE_SIZE;
1617 		uresp.sq_db_gts_key = ucontext->key;
1618 		ucontext->key += PAGE_SIZE;
1619 		uresp.rq_db_gts_key = ucontext->key;
1620 		ucontext->key += PAGE_SIZE;
1621 		spin_unlock(&ucontext->mmap_lock);
1622 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1623 		if (ret)
1624 			goto err8;
1625 		mm1->key = uresp.sq_key;
1626 		mm1->addr = qhp->wq.sq.phys_addr;
1627 		mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1628 		insert_mmap(ucontext, mm1);
1629 		mm2->key = uresp.rq_key;
1630 		mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1631 		mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1632 		insert_mmap(ucontext, mm2);
1633 		mm3->key = uresp.sq_db_gts_key;
1634 		mm3->addr = qhp->wq.sq.udb;
1635 		mm3->len = PAGE_SIZE;
1636 		insert_mmap(ucontext, mm3);
1637 		mm4->key = uresp.rq_db_gts_key;
1638 		mm4->addr = qhp->wq.rq.udb;
1639 		mm4->len = PAGE_SIZE;
1640 		insert_mmap(ucontext, mm4);
1641 		if (mm5) {
1642 			mm5->key = uresp.ma_sync_key;
1643 			mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1644 				    + A_PCIE_MA_SYNC) & PAGE_MASK;
1645 			mm5->len = PAGE_SIZE;
1646 			insert_mmap(ucontext, mm5);
1647 		}
1648 	}
1649 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1650 	init_timer(&(qhp->timer));
1651 	PDBG("%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x\n",
1652 	     __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1653 	     qhp->wq.sq.qid);
1654 	return &qhp->ibqp;
1655 err8:
1656 	kfree(mm5);
1657 err7:
1658 	kfree(mm4);
1659 err6:
1660 	kfree(mm3);
1661 err5:
1662 	kfree(mm2);
1663 err4:
1664 	kfree(mm1);
1665 err3:
1666 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1667 err2:
1668 	destroy_qp(&rhp->rdev, &qhp->wq,
1669 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1670 err1:
1671 	kfree(qhp);
1672 	return ERR_PTR(ret);
1673 }
1674 
1675 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1676 		      int attr_mask, struct ib_udata *udata)
1677 {
1678 	struct c4iw_dev *rhp;
1679 	struct c4iw_qp *qhp;
1680 	enum c4iw_qp_attr_mask mask = 0;
1681 	struct c4iw_qp_attributes attrs;
1682 
1683 	PDBG("%s ib_qp %p\n", __func__, ibqp);
1684 
1685 	/* iwarp does not support the RTR state */
1686 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1687 		attr_mask &= ~IB_QP_STATE;
1688 
1689 	/* Make sure we still have something left to do */
1690 	if (!attr_mask)
1691 		return 0;
1692 
1693 	memset(&attrs, 0, sizeof attrs);
1694 	qhp = to_c4iw_qp(ibqp);
1695 	rhp = qhp->rhp;
1696 
1697 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1698 	attrs.enable_rdma_read = (attr->qp_access_flags &
1699 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1700 	attrs.enable_rdma_write = (attr->qp_access_flags &
1701 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1702 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1703 
1704 
1705 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1706 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1707 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1708 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1709 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1710 
1711 	/*
1712 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1713 	 * ringing the queue db when we're in DB_FULL mode.
1714 	 */
1715 	attrs.sq_db_inc = attr->sq_psn;
1716 	attrs.rq_db_inc = attr->rq_psn;
1717 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1718 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1719 
1720 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1721 }
1722 
1723 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1724 {
1725 	PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1726 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1727 }
1728 
1729 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1730 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1731 {
1732 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1733 
1734 	memset(attr, 0, sizeof *attr);
1735 	memset(init_attr, 0, sizeof *init_attr);
1736 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1737 	return 0;
1738 }
1739