xref: /linux/drivers/infiniband/hw/cxgb4/qp.c (revision 26b0d14106954ae46d2f4f7eec3481828a210f7d)
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 		goto err1;
142 
143 	if (!user) {
144 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
145 				 GFP_KERNEL);
146 		if (!wq->sq.sw_sq)
147 			goto err2;
148 
149 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
150 				 GFP_KERNEL);
151 		if (!wq->rq.sw_rq)
152 			goto err3;
153 	}
154 
155 	/*
156 	 * RQT must be a power of 2.
157 	 */
158 	wq->rq.rqt_size = roundup_pow_of_two(wq->rq.size);
159 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
160 	if (!wq->rq.rqt_hwaddr)
161 		goto err4;
162 
163 	if (user) {
164 		if (alloc_oc_sq(rdev, &wq->sq) && alloc_host_sq(rdev, &wq->sq))
165 			goto err5;
166 	} else
167 		if (alloc_host_sq(rdev, &wq->sq))
168 			goto err5;
169 	memset(wq->sq.queue, 0, wq->sq.memsize);
170 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
171 
172 	wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
173 					  wq->rq.memsize, &(wq->rq.dma_addr),
174 					  GFP_KERNEL);
175 	if (!wq->rq.queue)
176 		goto err6;
177 	PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
178 		__func__, wq->sq.queue,
179 		(unsigned long long)virt_to_phys(wq->sq.queue),
180 		wq->rq.queue,
181 		(unsigned long long)virt_to_phys(wq->rq.queue));
182 	memset(wq->rq.queue, 0, wq->rq.memsize);
183 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
184 
185 	wq->db = rdev->lldi.db_reg;
186 	wq->gts = rdev->lldi.gts_reg;
187 	if (user) {
188 		wq->sq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
189 					(wq->sq.qid << rdev->qpshift);
190 		wq->sq.udb &= PAGE_MASK;
191 		wq->rq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
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 	skb = alloc_skb(wr_len, GFP_KERNEL);
202 	if (!skb) {
203 		ret = -ENOMEM;
204 		goto err7;
205 	}
206 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
207 
208 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
209 	memset(res_wr, 0, wr_len);
210 	res_wr->op_nres = cpu_to_be32(
211 			FW_WR_OP(FW_RI_RES_WR) |
212 			V_FW_RI_RES_WR_NRES(2) |
213 			FW_WR_COMPL(1));
214 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
215 	res_wr->cookie = (unsigned long) &wr_wait;
216 	res = res_wr->res;
217 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
218 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
219 
220 	/*
221 	 * eqsize is the number of 64B entries plus the status page size.
222 	 */
223 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
224 
225 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
226 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
227 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
228 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
229 		(t4_sq_onchip(&wq->sq) ? F_FW_RI_RES_WR_ONCHIP : 0) |
230 		V_FW_RI_RES_WR_IQID(scq->cqid));
231 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
232 		V_FW_RI_RES_WR_DCAEN(0) |
233 		V_FW_RI_RES_WR_DCACPU(0) |
234 		V_FW_RI_RES_WR_FBMIN(2) |
235 		V_FW_RI_RES_WR_FBMAX(2) |
236 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
237 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
238 		V_FW_RI_RES_WR_EQSIZE(eqsize));
239 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
240 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
241 	res++;
242 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
243 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
244 
245 	/*
246 	 * eqsize is the number of 64B entries plus the status page size.
247 	 */
248 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
249 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
250 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
251 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
252 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
253 		V_FW_RI_RES_WR_IQID(rcq->cqid));
254 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
255 		V_FW_RI_RES_WR_DCAEN(0) |
256 		V_FW_RI_RES_WR_DCACPU(0) |
257 		V_FW_RI_RES_WR_FBMIN(2) |
258 		V_FW_RI_RES_WR_FBMAX(2) |
259 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
260 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
261 		V_FW_RI_RES_WR_EQSIZE(eqsize));
262 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
263 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
264 
265 	c4iw_init_wr_wait(&wr_wait);
266 
267 	ret = c4iw_ofld_send(rdev, skb);
268 	if (ret)
269 		goto err7;
270 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
271 	if (ret)
272 		goto err7;
273 
274 	PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx\n",
275 	     __func__, wq->sq.qid, wq->rq.qid, wq->db,
276 	     (unsigned long long)wq->sq.udb, (unsigned long long)wq->rq.udb);
277 
278 	return 0;
279 err7:
280 	dma_free_coherent(&(rdev->lldi.pdev->dev),
281 			  wq->rq.memsize, wq->rq.queue,
282 			  dma_unmap_addr(&wq->rq, mapping));
283 err6:
284 	dealloc_sq(rdev, &wq->sq);
285 err5:
286 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
287 err4:
288 	kfree(wq->rq.sw_rq);
289 err3:
290 	kfree(wq->sq.sw_sq);
291 err2:
292 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
293 err1:
294 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
295 	return -ENOMEM;
296 }
297 
298 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
299 		      struct ib_send_wr *wr, int max, u32 *plenp)
300 {
301 	u8 *dstp, *srcp;
302 	u32 plen = 0;
303 	int i;
304 	int rem, len;
305 
306 	dstp = (u8 *)immdp->data;
307 	for (i = 0; i < wr->num_sge; i++) {
308 		if ((plen + wr->sg_list[i].length) > max)
309 			return -EMSGSIZE;
310 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
311 		plen += wr->sg_list[i].length;
312 		rem = wr->sg_list[i].length;
313 		while (rem) {
314 			if (dstp == (u8 *)&sq->queue[sq->size])
315 				dstp = (u8 *)sq->queue;
316 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
317 				len = rem;
318 			else
319 				len = (u8 *)&sq->queue[sq->size] - dstp;
320 			memcpy(dstp, srcp, len);
321 			dstp += len;
322 			srcp += len;
323 			rem -= len;
324 		}
325 	}
326 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
327 	if (len)
328 		memset(dstp, 0, len);
329 	immdp->op = FW_RI_DATA_IMMD;
330 	immdp->r1 = 0;
331 	immdp->r2 = 0;
332 	immdp->immdlen = cpu_to_be32(plen);
333 	*plenp = plen;
334 	return 0;
335 }
336 
337 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
338 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
339 		      int num_sge, u32 *plenp)
340 
341 {
342 	int i;
343 	u32 plen = 0;
344 	__be64 *flitp = (__be64 *)isglp->sge;
345 
346 	for (i = 0; i < num_sge; i++) {
347 		if ((plen + sg_list[i].length) < plen)
348 			return -EMSGSIZE;
349 		plen += sg_list[i].length;
350 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
351 				     sg_list[i].length);
352 		if (++flitp == queue_end)
353 			flitp = queue_start;
354 		*flitp = cpu_to_be64(sg_list[i].addr);
355 		if (++flitp == queue_end)
356 			flitp = queue_start;
357 	}
358 	*flitp = (__force __be64)0;
359 	isglp->op = FW_RI_DATA_ISGL;
360 	isglp->r1 = 0;
361 	isglp->nsge = cpu_to_be16(num_sge);
362 	isglp->r2 = 0;
363 	if (plenp)
364 		*plenp = plen;
365 	return 0;
366 }
367 
368 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
369 			   struct ib_send_wr *wr, u8 *len16)
370 {
371 	u32 plen;
372 	int size;
373 	int ret;
374 
375 	if (wr->num_sge > T4_MAX_SEND_SGE)
376 		return -EINVAL;
377 	switch (wr->opcode) {
378 	case IB_WR_SEND:
379 		if (wr->send_flags & IB_SEND_SOLICITED)
380 			wqe->send.sendop_pkd = cpu_to_be32(
381 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
382 		else
383 			wqe->send.sendop_pkd = cpu_to_be32(
384 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
385 		wqe->send.stag_inv = 0;
386 		break;
387 	case IB_WR_SEND_WITH_INV:
388 		if (wr->send_flags & IB_SEND_SOLICITED)
389 			wqe->send.sendop_pkd = cpu_to_be32(
390 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
391 		else
392 			wqe->send.sendop_pkd = cpu_to_be32(
393 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
394 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
395 		break;
396 
397 	default:
398 		return -EINVAL;
399 	}
400 
401 	plen = 0;
402 	if (wr->num_sge) {
403 		if (wr->send_flags & IB_SEND_INLINE) {
404 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
405 					 T4_MAX_SEND_INLINE, &plen);
406 			if (ret)
407 				return ret;
408 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
409 			       plen;
410 		} else {
411 			ret = build_isgl((__be64 *)sq->queue,
412 					 (__be64 *)&sq->queue[sq->size],
413 					 wqe->send.u.isgl_src,
414 					 wr->sg_list, wr->num_sge, &plen);
415 			if (ret)
416 				return ret;
417 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
418 			       wr->num_sge * sizeof(struct fw_ri_sge);
419 		}
420 	} else {
421 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
422 		wqe->send.u.immd_src[0].r1 = 0;
423 		wqe->send.u.immd_src[0].r2 = 0;
424 		wqe->send.u.immd_src[0].immdlen = 0;
425 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
426 		plen = 0;
427 	}
428 	*len16 = DIV_ROUND_UP(size, 16);
429 	wqe->send.plen = cpu_to_be32(plen);
430 	return 0;
431 }
432 
433 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
434 			    struct ib_send_wr *wr, u8 *len16)
435 {
436 	u32 plen;
437 	int size;
438 	int ret;
439 
440 	if (wr->num_sge > T4_MAX_SEND_SGE)
441 		return -EINVAL;
442 	wqe->write.r2 = 0;
443 	wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
444 	wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
445 	if (wr->num_sge) {
446 		if (wr->send_flags & IB_SEND_INLINE) {
447 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
448 					 T4_MAX_WRITE_INLINE, &plen);
449 			if (ret)
450 				return ret;
451 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
452 			       plen;
453 		} else {
454 			ret = build_isgl((__be64 *)sq->queue,
455 					 (__be64 *)&sq->queue[sq->size],
456 					 wqe->write.u.isgl_src,
457 					 wr->sg_list, wr->num_sge, &plen);
458 			if (ret)
459 				return ret;
460 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
461 			       wr->num_sge * sizeof(struct fw_ri_sge);
462 		}
463 	} else {
464 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
465 		wqe->write.u.immd_src[0].r1 = 0;
466 		wqe->write.u.immd_src[0].r2 = 0;
467 		wqe->write.u.immd_src[0].immdlen = 0;
468 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
469 		plen = 0;
470 	}
471 	*len16 = DIV_ROUND_UP(size, 16);
472 	wqe->write.plen = cpu_to_be32(plen);
473 	return 0;
474 }
475 
476 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
477 {
478 	if (wr->num_sge > 1)
479 		return -EINVAL;
480 	if (wr->num_sge) {
481 		wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
482 		wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
483 							>> 32));
484 		wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
485 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
486 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
487 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
488 							 >> 32));
489 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
490 	} else {
491 		wqe->read.stag_src = cpu_to_be32(2);
492 		wqe->read.to_src_hi = 0;
493 		wqe->read.to_src_lo = 0;
494 		wqe->read.stag_sink = cpu_to_be32(2);
495 		wqe->read.plen = 0;
496 		wqe->read.to_sink_hi = 0;
497 		wqe->read.to_sink_lo = 0;
498 	}
499 	wqe->read.r2 = 0;
500 	wqe->read.r5 = 0;
501 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
502 	return 0;
503 }
504 
505 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
506 			   struct ib_recv_wr *wr, u8 *len16)
507 {
508 	int ret;
509 
510 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
511 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
512 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
513 	if (ret)
514 		return ret;
515 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
516 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
517 	return 0;
518 }
519 
520 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
521 			 struct ib_send_wr *wr, u8 *len16)
522 {
523 
524 	struct fw_ri_immd *imdp;
525 	__be64 *p;
526 	int i;
527 	int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
528 	int rem;
529 
530 	if (wr->wr.fast_reg.page_list_len > T4_MAX_FR_DEPTH)
531 		return -EINVAL;
532 
533 	wqe->fr.qpbinde_to_dcacpu = 0;
534 	wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
535 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
536 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
537 	wqe->fr.len_hi = 0;
538 	wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
539 	wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
540 	wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
541 	wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
542 					0xffffffff);
543 	WARN_ON(pbllen > T4_MAX_FR_IMMD);
544 	imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
545 	imdp->op = FW_RI_DATA_IMMD;
546 	imdp->r1 = 0;
547 	imdp->r2 = 0;
548 	imdp->immdlen = cpu_to_be32(pbllen);
549 	p = (__be64 *)(imdp + 1);
550 	rem = pbllen;
551 	for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
552 		*p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
553 		rem -= sizeof *p;
554 		if (++p == (__be64 *)&sq->queue[sq->size])
555 			p = (__be64 *)sq->queue;
556 	}
557 	BUG_ON(rem < 0);
558 	while (rem) {
559 		*p = 0;
560 		rem -= sizeof *p;
561 		if (++p == (__be64 *)&sq->queue[sq->size])
562 			p = (__be64 *)sq->queue;
563 	}
564 	*len16 = DIV_ROUND_UP(sizeof wqe->fr + sizeof *imdp + pbllen, 16);
565 	return 0;
566 }
567 
568 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
569 			  u8 *len16)
570 {
571 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
572 	wqe->inv.r2 = 0;
573 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
574 	return 0;
575 }
576 
577 void c4iw_qp_add_ref(struct ib_qp *qp)
578 {
579 	PDBG("%s ib_qp %p\n", __func__, qp);
580 	atomic_inc(&(to_c4iw_qp(qp)->refcnt));
581 }
582 
583 void c4iw_qp_rem_ref(struct ib_qp *qp)
584 {
585 	PDBG("%s ib_qp %p\n", __func__, qp);
586 	if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
587 		wake_up(&(to_c4iw_qp(qp)->wait));
588 }
589 
590 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
591 		   struct ib_send_wr **bad_wr)
592 {
593 	int err = 0;
594 	u8 len16 = 0;
595 	enum fw_wr_opcodes fw_opcode = 0;
596 	enum fw_ri_wr_flags fw_flags;
597 	struct c4iw_qp *qhp;
598 	union t4_wr *wqe;
599 	u32 num_wrs;
600 	struct t4_swsqe *swsqe;
601 	unsigned long flag;
602 	u16 idx = 0;
603 
604 	qhp = to_c4iw_qp(ibqp);
605 	spin_lock_irqsave(&qhp->lock, flag);
606 	if (t4_wq_in_error(&qhp->wq)) {
607 		spin_unlock_irqrestore(&qhp->lock, flag);
608 		return -EINVAL;
609 	}
610 	num_wrs = t4_sq_avail(&qhp->wq);
611 	if (num_wrs == 0) {
612 		spin_unlock_irqrestore(&qhp->lock, flag);
613 		return -ENOMEM;
614 	}
615 	while (wr) {
616 		if (num_wrs == 0) {
617 			err = -ENOMEM;
618 			*bad_wr = wr;
619 			break;
620 		}
621 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
622 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
623 
624 		fw_flags = 0;
625 		if (wr->send_flags & IB_SEND_SOLICITED)
626 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
627 		if (wr->send_flags & IB_SEND_SIGNALED)
628 			fw_flags |= FW_RI_COMPLETION_FLAG;
629 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
630 		switch (wr->opcode) {
631 		case IB_WR_SEND_WITH_INV:
632 		case IB_WR_SEND:
633 			if (wr->send_flags & IB_SEND_FENCE)
634 				fw_flags |= FW_RI_READ_FENCE_FLAG;
635 			fw_opcode = FW_RI_SEND_WR;
636 			if (wr->opcode == IB_WR_SEND)
637 				swsqe->opcode = FW_RI_SEND;
638 			else
639 				swsqe->opcode = FW_RI_SEND_WITH_INV;
640 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
641 			break;
642 		case IB_WR_RDMA_WRITE:
643 			fw_opcode = FW_RI_RDMA_WRITE_WR;
644 			swsqe->opcode = FW_RI_RDMA_WRITE;
645 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
646 			break;
647 		case IB_WR_RDMA_READ:
648 		case IB_WR_RDMA_READ_WITH_INV:
649 			fw_opcode = FW_RI_RDMA_READ_WR;
650 			swsqe->opcode = FW_RI_READ_REQ;
651 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
652 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
653 			else
654 				fw_flags = 0;
655 			err = build_rdma_read(wqe, wr, &len16);
656 			if (err)
657 				break;
658 			swsqe->read_len = wr->sg_list[0].length;
659 			if (!qhp->wq.sq.oldest_read)
660 				qhp->wq.sq.oldest_read = swsqe;
661 			break;
662 		case IB_WR_FAST_REG_MR:
663 			fw_opcode = FW_RI_FR_NSMR_WR;
664 			swsqe->opcode = FW_RI_FAST_REGISTER;
665 			err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16);
666 			break;
667 		case IB_WR_LOCAL_INV:
668 			if (wr->send_flags & IB_SEND_FENCE)
669 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
670 			fw_opcode = FW_RI_INV_LSTAG_WR;
671 			swsqe->opcode = FW_RI_LOCAL_INV;
672 			err = build_inv_stag(wqe, wr, &len16);
673 			break;
674 		default:
675 			PDBG("%s post of type=%d TBD!\n", __func__,
676 			     wr->opcode);
677 			err = -EINVAL;
678 		}
679 		if (err) {
680 			*bad_wr = wr;
681 			break;
682 		}
683 		swsqe->idx = qhp->wq.sq.pidx;
684 		swsqe->complete = 0;
685 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED);
686 		swsqe->wr_id = wr->wr_id;
687 
688 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
689 
690 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
691 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
692 		     swsqe->opcode, swsqe->read_len);
693 		wr = wr->next;
694 		num_wrs--;
695 		t4_sq_produce(&qhp->wq, len16);
696 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
697 	}
698 	if (t4_wq_db_enabled(&qhp->wq))
699 		t4_ring_sq_db(&qhp->wq, idx);
700 	spin_unlock_irqrestore(&qhp->lock, flag);
701 	return err;
702 }
703 
704 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
705 		      struct ib_recv_wr **bad_wr)
706 {
707 	int err = 0;
708 	struct c4iw_qp *qhp;
709 	union t4_recv_wr *wqe;
710 	u32 num_wrs;
711 	u8 len16 = 0;
712 	unsigned long flag;
713 	u16 idx = 0;
714 
715 	qhp = to_c4iw_qp(ibqp);
716 	spin_lock_irqsave(&qhp->lock, flag);
717 	if (t4_wq_in_error(&qhp->wq)) {
718 		spin_unlock_irqrestore(&qhp->lock, flag);
719 		return -EINVAL;
720 	}
721 	num_wrs = t4_rq_avail(&qhp->wq);
722 	if (num_wrs == 0) {
723 		spin_unlock_irqrestore(&qhp->lock, flag);
724 		return -ENOMEM;
725 	}
726 	while (wr) {
727 		if (wr->num_sge > T4_MAX_RECV_SGE) {
728 			err = -EINVAL;
729 			*bad_wr = wr;
730 			break;
731 		}
732 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
733 					   qhp->wq.rq.wq_pidx *
734 					   T4_EQ_ENTRY_SIZE);
735 		if (num_wrs)
736 			err = build_rdma_recv(qhp, wqe, wr, &len16);
737 		else
738 			err = -ENOMEM;
739 		if (err) {
740 			*bad_wr = wr;
741 			break;
742 		}
743 
744 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
745 
746 		wqe->recv.opcode = FW_RI_RECV_WR;
747 		wqe->recv.r1 = 0;
748 		wqe->recv.wrid = qhp->wq.rq.pidx;
749 		wqe->recv.r2[0] = 0;
750 		wqe->recv.r2[1] = 0;
751 		wqe->recv.r2[2] = 0;
752 		wqe->recv.len16 = len16;
753 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
754 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
755 		t4_rq_produce(&qhp->wq, len16);
756 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
757 		wr = wr->next;
758 		num_wrs--;
759 	}
760 	if (t4_wq_db_enabled(&qhp->wq))
761 		t4_ring_rq_db(&qhp->wq, idx);
762 	spin_unlock_irqrestore(&qhp->lock, flag);
763 	return err;
764 }
765 
766 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
767 {
768 	return -ENOSYS;
769 }
770 
771 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
772 				    u8 *ecode)
773 {
774 	int status;
775 	int tagged;
776 	int opcode;
777 	int rqtype;
778 	int send_inv;
779 
780 	if (!err_cqe) {
781 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
782 		*ecode = 0;
783 		return;
784 	}
785 
786 	status = CQE_STATUS(err_cqe);
787 	opcode = CQE_OPCODE(err_cqe);
788 	rqtype = RQ_TYPE(err_cqe);
789 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
790 		   (opcode == FW_RI_SEND_WITH_SE_INV);
791 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
792 		 (rqtype && (opcode == FW_RI_READ_RESP));
793 
794 	switch (status) {
795 	case T4_ERR_STAG:
796 		if (send_inv) {
797 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
798 			*ecode = RDMAP_CANT_INV_STAG;
799 		} else {
800 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
801 			*ecode = RDMAP_INV_STAG;
802 		}
803 		break;
804 	case T4_ERR_PDID:
805 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
806 		if ((opcode == FW_RI_SEND_WITH_INV) ||
807 		    (opcode == FW_RI_SEND_WITH_SE_INV))
808 			*ecode = RDMAP_CANT_INV_STAG;
809 		else
810 			*ecode = RDMAP_STAG_NOT_ASSOC;
811 		break;
812 	case T4_ERR_QPID:
813 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
814 		*ecode = RDMAP_STAG_NOT_ASSOC;
815 		break;
816 	case T4_ERR_ACCESS:
817 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
818 		*ecode = RDMAP_ACC_VIOL;
819 		break;
820 	case T4_ERR_WRAP:
821 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
822 		*ecode = RDMAP_TO_WRAP;
823 		break;
824 	case T4_ERR_BOUND:
825 		if (tagged) {
826 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
827 			*ecode = DDPT_BASE_BOUNDS;
828 		} else {
829 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
830 			*ecode = RDMAP_BASE_BOUNDS;
831 		}
832 		break;
833 	case T4_ERR_INVALIDATE_SHARED_MR:
834 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
835 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
836 		*ecode = RDMAP_CANT_INV_STAG;
837 		break;
838 	case T4_ERR_ECC:
839 	case T4_ERR_ECC_PSTAG:
840 	case T4_ERR_INTERNAL_ERR:
841 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
842 		*ecode = 0;
843 		break;
844 	case T4_ERR_OUT_OF_RQE:
845 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
846 		*ecode = DDPU_INV_MSN_NOBUF;
847 		break;
848 	case T4_ERR_PBL_ADDR_BOUND:
849 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
850 		*ecode = DDPT_BASE_BOUNDS;
851 		break;
852 	case T4_ERR_CRC:
853 		*layer_type = LAYER_MPA|DDP_LLP;
854 		*ecode = MPA_CRC_ERR;
855 		break;
856 	case T4_ERR_MARKER:
857 		*layer_type = LAYER_MPA|DDP_LLP;
858 		*ecode = MPA_MARKER_ERR;
859 		break;
860 	case T4_ERR_PDU_LEN_ERR:
861 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
862 		*ecode = DDPU_MSG_TOOBIG;
863 		break;
864 	case T4_ERR_DDP_VERSION:
865 		if (tagged) {
866 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
867 			*ecode = DDPT_INV_VERS;
868 		} else {
869 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
870 			*ecode = DDPU_INV_VERS;
871 		}
872 		break;
873 	case T4_ERR_RDMA_VERSION:
874 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
875 		*ecode = RDMAP_INV_VERS;
876 		break;
877 	case T4_ERR_OPCODE:
878 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
879 		*ecode = RDMAP_INV_OPCODE;
880 		break;
881 	case T4_ERR_DDP_QUEUE_NUM:
882 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
883 		*ecode = DDPU_INV_QN;
884 		break;
885 	case T4_ERR_MSN:
886 	case T4_ERR_MSN_GAP:
887 	case T4_ERR_MSN_RANGE:
888 	case T4_ERR_IRD_OVERFLOW:
889 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
890 		*ecode = DDPU_INV_MSN_RANGE;
891 		break;
892 	case T4_ERR_TBIT:
893 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
894 		*ecode = 0;
895 		break;
896 	case T4_ERR_MO:
897 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
898 		*ecode = DDPU_INV_MO;
899 		break;
900 	default:
901 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
902 		*ecode = 0;
903 		break;
904 	}
905 }
906 
907 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
908 			   gfp_t gfp)
909 {
910 	struct fw_ri_wr *wqe;
911 	struct sk_buff *skb;
912 	struct terminate_message *term;
913 
914 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
915 	     qhp->ep->hwtid);
916 
917 	skb = alloc_skb(sizeof *wqe, gfp);
918 	if (!skb)
919 		return;
920 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
921 
922 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
923 	memset(wqe, 0, sizeof *wqe);
924 	wqe->op_compl = cpu_to_be32(FW_WR_OP(FW_RI_INIT_WR));
925 	wqe->flowid_len16 = cpu_to_be32(
926 		FW_WR_FLOWID(qhp->ep->hwtid) |
927 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
928 
929 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
930 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
931 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
932 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
933 		term->layer_etype = qhp->attr.layer_etype;
934 		term->ecode = qhp->attr.ecode;
935 	} else
936 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
937 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
938 }
939 
940 /*
941  * Assumes qhp lock is held.
942  */
943 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
944 		       struct c4iw_cq *schp)
945 {
946 	int count;
947 	int flushed;
948 	unsigned long flag;
949 
950 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
951 
952 	/* locking hierarchy: cq lock first, then qp lock. */
953 	spin_lock_irqsave(&rchp->lock, flag);
954 	spin_lock(&qhp->lock);
955 	c4iw_flush_hw_cq(&rchp->cq);
956 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
957 	flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
958 	spin_unlock(&qhp->lock);
959 	spin_unlock_irqrestore(&rchp->lock, flag);
960 	if (flushed) {
961 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
962 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
963 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
964 	}
965 
966 	/* locking hierarchy: cq lock first, then qp lock. */
967 	spin_lock_irqsave(&schp->lock, flag);
968 	spin_lock(&qhp->lock);
969 	c4iw_flush_hw_cq(&schp->cq);
970 	c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
971 	flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
972 	spin_unlock(&qhp->lock);
973 	spin_unlock_irqrestore(&schp->lock, flag);
974 	if (flushed) {
975 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
976 		(*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
977 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
978 	}
979 }
980 
981 static void flush_qp(struct c4iw_qp *qhp)
982 {
983 	struct c4iw_cq *rchp, *schp;
984 	unsigned long flag;
985 
986 	rchp = get_chp(qhp->rhp, qhp->attr.rcq);
987 	schp = get_chp(qhp->rhp, qhp->attr.scq);
988 
989 	if (qhp->ibqp.uobject) {
990 		t4_set_wq_in_error(&qhp->wq);
991 		t4_set_cq_in_error(&rchp->cq);
992 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
993 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
994 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
995 		if (schp != rchp) {
996 			t4_set_cq_in_error(&schp->cq);
997 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
998 			(*schp->ibcq.comp_handler)(&schp->ibcq,
999 					schp->ibcq.cq_context);
1000 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1001 		}
1002 		return;
1003 	}
1004 	__flush_qp(qhp, rchp, schp);
1005 }
1006 
1007 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1008 		     struct c4iw_ep *ep)
1009 {
1010 	struct fw_ri_wr *wqe;
1011 	int ret;
1012 	struct sk_buff *skb;
1013 
1014 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1015 	     ep->hwtid);
1016 
1017 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1018 	if (!skb)
1019 		return -ENOMEM;
1020 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1021 
1022 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1023 	memset(wqe, 0, sizeof *wqe);
1024 	wqe->op_compl = cpu_to_be32(
1025 		FW_WR_OP(FW_RI_INIT_WR) |
1026 		FW_WR_COMPL(1));
1027 	wqe->flowid_len16 = cpu_to_be32(
1028 		FW_WR_FLOWID(ep->hwtid) |
1029 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1030 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1031 
1032 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1033 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1034 	if (ret)
1035 		goto out;
1036 
1037 	ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1038 			     qhp->wq.sq.qid, __func__);
1039 out:
1040 	PDBG("%s ret %d\n", __func__, ret);
1041 	return ret;
1042 }
1043 
1044 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1045 {
1046 	PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1047 	memset(&init->u, 0, sizeof init->u);
1048 	switch (p2p_type) {
1049 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1050 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1051 		init->u.write.stag_sink = cpu_to_be32(1);
1052 		init->u.write.to_sink = cpu_to_be64(1);
1053 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1054 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1055 						   sizeof(struct fw_ri_immd),
1056 						   16);
1057 		break;
1058 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1059 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1060 		init->u.read.stag_src = cpu_to_be32(1);
1061 		init->u.read.to_src_lo = cpu_to_be32(1);
1062 		init->u.read.stag_sink = cpu_to_be32(1);
1063 		init->u.read.to_sink_lo = cpu_to_be32(1);
1064 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1065 		break;
1066 	}
1067 }
1068 
1069 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1070 {
1071 	struct fw_ri_wr *wqe;
1072 	int ret;
1073 	struct sk_buff *skb;
1074 
1075 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1076 	     qhp->ep->hwtid);
1077 
1078 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1079 	if (!skb)
1080 		return -ENOMEM;
1081 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1082 
1083 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1084 	memset(wqe, 0, sizeof *wqe);
1085 	wqe->op_compl = cpu_to_be32(
1086 		FW_WR_OP(FW_RI_INIT_WR) |
1087 		FW_WR_COMPL(1));
1088 	wqe->flowid_len16 = cpu_to_be32(
1089 		FW_WR_FLOWID(qhp->ep->hwtid) |
1090 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1091 
1092 	wqe->cookie = (unsigned long) &qhp->ep->com.wr_wait;
1093 
1094 	wqe->u.init.type = FW_RI_TYPE_INIT;
1095 	wqe->u.init.mpareqbit_p2ptype =
1096 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1097 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1098 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1099 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1100 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1101 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1102 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1103 	if (qhp->attr.mpa_attr.crc_enabled)
1104 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1105 
1106 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1107 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1108 			    FW_RI_QP_BIND_ENABLE;
1109 	if (!qhp->ibqp.uobject)
1110 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1111 				     FW_RI_QP_STAG0_ENABLE;
1112 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1113 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1114 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1115 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1116 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1117 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1118 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1119 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1120 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1121 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1122 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1123 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1124 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1125 					 rhp->rdev.lldi.vr->rq.start);
1126 	if (qhp->attr.mpa_attr.initiator)
1127 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1128 
1129 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1130 	if (ret)
1131 		goto out;
1132 
1133 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1134 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1135 out:
1136 	PDBG("%s ret %d\n", __func__, ret);
1137 	return ret;
1138 }
1139 
1140 /*
1141  * Called by the library when the qp has user dbs disabled due to
1142  * a DB_FULL condition.  This function will single-thread all user
1143  * DB rings to avoid overflowing the hw db-fifo.
1144  */
1145 static int ring_kernel_db(struct c4iw_qp *qhp, u32 qid, u16 inc)
1146 {
1147 	int delay = db_delay_usecs;
1148 
1149 	mutex_lock(&qhp->rhp->db_mutex);
1150 	do {
1151 
1152 		/*
1153 		 * The interrupt threshold is dbfifo_int_thresh << 6. So
1154 		 * make sure we don't cross that and generate an interrupt.
1155 		 */
1156 		if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) <
1157 		    (qhp->rhp->rdev.lldi.dbfifo_int_thresh << 5)) {
1158 			writel(V_QID(qid) | V_PIDX(inc), qhp->wq.db);
1159 			break;
1160 		}
1161 		set_current_state(TASK_UNINTERRUPTIBLE);
1162 		schedule_timeout(usecs_to_jiffies(delay));
1163 		delay = min(delay << 1, 2000);
1164 	} while (1);
1165 	mutex_unlock(&qhp->rhp->db_mutex);
1166 	return 0;
1167 }
1168 
1169 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1170 		   enum c4iw_qp_attr_mask mask,
1171 		   struct c4iw_qp_attributes *attrs,
1172 		   int internal)
1173 {
1174 	int ret = 0;
1175 	struct c4iw_qp_attributes newattr = qhp->attr;
1176 	int disconnect = 0;
1177 	int terminate = 0;
1178 	int abort = 0;
1179 	int free = 0;
1180 	struct c4iw_ep *ep = NULL;
1181 
1182 	PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1183 	     qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1184 	     (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1185 
1186 	mutex_lock(&qhp->mutex);
1187 
1188 	/* Process attr changes if in IDLE */
1189 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1190 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1191 			ret = -EIO;
1192 			goto out;
1193 		}
1194 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1195 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1196 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1197 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1198 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1199 			newattr.enable_bind = attrs->enable_bind;
1200 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1201 			if (attrs->max_ord > c4iw_max_read_depth) {
1202 				ret = -EINVAL;
1203 				goto out;
1204 			}
1205 			newattr.max_ord = attrs->max_ord;
1206 		}
1207 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1208 			if (attrs->max_ird > c4iw_max_read_depth) {
1209 				ret = -EINVAL;
1210 				goto out;
1211 			}
1212 			newattr.max_ird = attrs->max_ird;
1213 		}
1214 		qhp->attr = newattr;
1215 	}
1216 
1217 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1218 		ret = ring_kernel_db(qhp, qhp->wq.sq.qid, attrs->sq_db_inc);
1219 		goto out;
1220 	}
1221 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1222 		ret = ring_kernel_db(qhp, qhp->wq.rq.qid, attrs->rq_db_inc);
1223 		goto out;
1224 	}
1225 
1226 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1227 		goto out;
1228 	if (qhp->attr.state == attrs->next_state)
1229 		goto out;
1230 
1231 	switch (qhp->attr.state) {
1232 	case C4IW_QP_STATE_IDLE:
1233 		switch (attrs->next_state) {
1234 		case C4IW_QP_STATE_RTS:
1235 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1236 				ret = -EINVAL;
1237 				goto out;
1238 			}
1239 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1240 				ret = -EINVAL;
1241 				goto out;
1242 			}
1243 			qhp->attr.mpa_attr = attrs->mpa_attr;
1244 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1245 			qhp->ep = qhp->attr.llp_stream_handle;
1246 			set_state(qhp, C4IW_QP_STATE_RTS);
1247 
1248 			/*
1249 			 * Ref the endpoint here and deref when we
1250 			 * disassociate the endpoint from the QP.  This
1251 			 * happens in CLOSING->IDLE transition or *->ERROR
1252 			 * transition.
1253 			 */
1254 			c4iw_get_ep(&qhp->ep->com);
1255 			ret = rdma_init(rhp, qhp);
1256 			if (ret)
1257 				goto err;
1258 			break;
1259 		case C4IW_QP_STATE_ERROR:
1260 			set_state(qhp, C4IW_QP_STATE_ERROR);
1261 			flush_qp(qhp);
1262 			break;
1263 		default:
1264 			ret = -EINVAL;
1265 			goto out;
1266 		}
1267 		break;
1268 	case C4IW_QP_STATE_RTS:
1269 		switch (attrs->next_state) {
1270 		case C4IW_QP_STATE_CLOSING:
1271 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1272 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1273 			ep = qhp->ep;
1274 			if (!internal) {
1275 				abort = 0;
1276 				disconnect = 1;
1277 				c4iw_get_ep(&qhp->ep->com);
1278 			}
1279 			if (qhp->ibqp.uobject)
1280 				t4_set_wq_in_error(&qhp->wq);
1281 			ret = rdma_fini(rhp, qhp, ep);
1282 			if (ret)
1283 				goto err;
1284 			break;
1285 		case C4IW_QP_STATE_TERMINATE:
1286 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1287 			qhp->attr.layer_etype = attrs->layer_etype;
1288 			qhp->attr.ecode = attrs->ecode;
1289 			if (qhp->ibqp.uobject)
1290 				t4_set_wq_in_error(&qhp->wq);
1291 			ep = qhp->ep;
1292 			if (!internal)
1293 				terminate = 1;
1294 			disconnect = 1;
1295 			c4iw_get_ep(&qhp->ep->com);
1296 			break;
1297 		case C4IW_QP_STATE_ERROR:
1298 			set_state(qhp, C4IW_QP_STATE_ERROR);
1299 			if (qhp->ibqp.uobject)
1300 				t4_set_wq_in_error(&qhp->wq);
1301 			if (!internal) {
1302 				abort = 1;
1303 				disconnect = 1;
1304 				ep = qhp->ep;
1305 				c4iw_get_ep(&qhp->ep->com);
1306 			}
1307 			goto err;
1308 			break;
1309 		default:
1310 			ret = -EINVAL;
1311 			goto out;
1312 		}
1313 		break;
1314 	case C4IW_QP_STATE_CLOSING:
1315 		if (!internal) {
1316 			ret = -EINVAL;
1317 			goto out;
1318 		}
1319 		switch (attrs->next_state) {
1320 		case C4IW_QP_STATE_IDLE:
1321 			flush_qp(qhp);
1322 			set_state(qhp, C4IW_QP_STATE_IDLE);
1323 			qhp->attr.llp_stream_handle = NULL;
1324 			c4iw_put_ep(&qhp->ep->com);
1325 			qhp->ep = NULL;
1326 			wake_up(&qhp->wait);
1327 			break;
1328 		case C4IW_QP_STATE_ERROR:
1329 			goto err;
1330 		default:
1331 			ret = -EINVAL;
1332 			goto err;
1333 		}
1334 		break;
1335 	case C4IW_QP_STATE_ERROR:
1336 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1337 			ret = -EINVAL;
1338 			goto out;
1339 		}
1340 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1341 			ret = -EINVAL;
1342 			goto out;
1343 		}
1344 		set_state(qhp, C4IW_QP_STATE_IDLE);
1345 		break;
1346 	case C4IW_QP_STATE_TERMINATE:
1347 		if (!internal) {
1348 			ret = -EINVAL;
1349 			goto out;
1350 		}
1351 		goto err;
1352 		break;
1353 	default:
1354 		printk(KERN_ERR "%s in a bad state %d\n",
1355 		       __func__, qhp->attr.state);
1356 		ret = -EINVAL;
1357 		goto err;
1358 		break;
1359 	}
1360 	goto out;
1361 err:
1362 	PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1363 	     qhp->wq.sq.qid);
1364 
1365 	/* disassociate the LLP connection */
1366 	qhp->attr.llp_stream_handle = NULL;
1367 	if (!ep)
1368 		ep = qhp->ep;
1369 	qhp->ep = NULL;
1370 	set_state(qhp, C4IW_QP_STATE_ERROR);
1371 	free = 1;
1372 	wake_up(&qhp->wait);
1373 	BUG_ON(!ep);
1374 	flush_qp(qhp);
1375 out:
1376 	mutex_unlock(&qhp->mutex);
1377 
1378 	if (terminate)
1379 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1380 
1381 	/*
1382 	 * If disconnect is 1, then we need to initiate a disconnect
1383 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1384 	 * an abnormal close (RTS/CLOSING->ERROR).
1385 	 */
1386 	if (disconnect) {
1387 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1388 							 GFP_KERNEL);
1389 		c4iw_put_ep(&ep->com);
1390 	}
1391 
1392 	/*
1393 	 * If free is 1, then we've disassociated the EP from the QP
1394 	 * and we need to dereference the EP.
1395 	 */
1396 	if (free)
1397 		c4iw_put_ep(&ep->com);
1398 	PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1399 	return ret;
1400 }
1401 
1402 static int enable_qp_db(int id, void *p, void *data)
1403 {
1404 	struct c4iw_qp *qp = p;
1405 
1406 	t4_enable_wq_db(&qp->wq);
1407 	return 0;
1408 }
1409 
1410 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1411 {
1412 	struct c4iw_dev *rhp;
1413 	struct c4iw_qp *qhp;
1414 	struct c4iw_qp_attributes attrs;
1415 	struct c4iw_ucontext *ucontext;
1416 
1417 	qhp = to_c4iw_qp(ib_qp);
1418 	rhp = qhp->rhp;
1419 
1420 	attrs.next_state = C4IW_QP_STATE_ERROR;
1421 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1422 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1423 	else
1424 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1425 	wait_event(qhp->wait, !qhp->ep);
1426 
1427 	spin_lock_irq(&rhp->lock);
1428 	remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1429 	rhp->qpcnt--;
1430 	BUG_ON(rhp->qpcnt < 0);
1431 	if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
1432 		rhp->rdev.stats.db_state_transitions++;
1433 		rhp->db_state = NORMAL;
1434 		idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
1435 	}
1436 	spin_unlock_irq(&rhp->lock);
1437 	atomic_dec(&qhp->refcnt);
1438 	wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1439 
1440 	ucontext = ib_qp->uobject ?
1441 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1442 	destroy_qp(&rhp->rdev, &qhp->wq,
1443 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1444 
1445 	PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1446 	kfree(qhp);
1447 	return 0;
1448 }
1449 
1450 static int disable_qp_db(int id, void *p, void *data)
1451 {
1452 	struct c4iw_qp *qp = p;
1453 
1454 	t4_disable_wq_db(&qp->wq);
1455 	return 0;
1456 }
1457 
1458 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1459 			     struct ib_udata *udata)
1460 {
1461 	struct c4iw_dev *rhp;
1462 	struct c4iw_qp *qhp;
1463 	struct c4iw_pd *php;
1464 	struct c4iw_cq *schp;
1465 	struct c4iw_cq *rchp;
1466 	struct c4iw_create_qp_resp uresp;
1467 	int sqsize, rqsize;
1468 	struct c4iw_ucontext *ucontext;
1469 	int ret;
1470 	struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1471 
1472 	PDBG("%s ib_pd %p\n", __func__, pd);
1473 
1474 	if (attrs->qp_type != IB_QPT_RC)
1475 		return ERR_PTR(-EINVAL);
1476 
1477 	php = to_c4iw_pd(pd);
1478 	rhp = php->rhp;
1479 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1480 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1481 	if (!schp || !rchp)
1482 		return ERR_PTR(-EINVAL);
1483 
1484 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1485 		return ERR_PTR(-EINVAL);
1486 
1487 	rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1488 	if (rqsize > T4_MAX_RQ_SIZE)
1489 		return ERR_PTR(-E2BIG);
1490 
1491 	sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1492 	if (sqsize > T4_MAX_SQ_SIZE)
1493 		return ERR_PTR(-E2BIG);
1494 
1495 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1496 
1497 
1498 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1499 	if (!qhp)
1500 		return ERR_PTR(-ENOMEM);
1501 	qhp->wq.sq.size = sqsize;
1502 	qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1503 	qhp->wq.rq.size = rqsize;
1504 	qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1505 
1506 	if (ucontext) {
1507 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1508 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1509 	}
1510 
1511 	PDBG("%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu\n",
1512 	     __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1513 
1514 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1515 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1516 	if (ret)
1517 		goto err1;
1518 
1519 	attrs->cap.max_recv_wr = rqsize - 1;
1520 	attrs->cap.max_send_wr = sqsize - 1;
1521 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1522 
1523 	qhp->rhp = rhp;
1524 	qhp->attr.pd = php->pdid;
1525 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1526 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1527 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1528 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1529 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1530 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1531 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1532 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1533 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1534 	qhp->attr.enable_rdma_read = 1;
1535 	qhp->attr.enable_rdma_write = 1;
1536 	qhp->attr.enable_bind = 1;
1537 	qhp->attr.max_ord = 1;
1538 	qhp->attr.max_ird = 1;
1539 	spin_lock_init(&qhp->lock);
1540 	mutex_init(&qhp->mutex);
1541 	init_waitqueue_head(&qhp->wait);
1542 	atomic_set(&qhp->refcnt, 1);
1543 
1544 	spin_lock_irq(&rhp->lock);
1545 	if (rhp->db_state != NORMAL)
1546 		t4_disable_wq_db(&qhp->wq);
1547 	if (++rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
1548 		rhp->rdev.stats.db_state_transitions++;
1549 		rhp->db_state = FLOW_CONTROL;
1550 		idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
1551 	}
1552 	ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1553 	spin_unlock_irq(&rhp->lock);
1554 	if (ret)
1555 		goto err2;
1556 
1557 	if (udata) {
1558 		mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1559 		if (!mm1) {
1560 			ret = -ENOMEM;
1561 			goto err3;
1562 		}
1563 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1564 		if (!mm2) {
1565 			ret = -ENOMEM;
1566 			goto err4;
1567 		}
1568 		mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1569 		if (!mm3) {
1570 			ret = -ENOMEM;
1571 			goto err5;
1572 		}
1573 		mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1574 		if (!mm4) {
1575 			ret = -ENOMEM;
1576 			goto err6;
1577 		}
1578 		if (t4_sq_onchip(&qhp->wq.sq)) {
1579 			mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1580 			if (!mm5) {
1581 				ret = -ENOMEM;
1582 				goto err7;
1583 			}
1584 			uresp.flags = C4IW_QPF_ONCHIP;
1585 		} else
1586 			uresp.flags = 0;
1587 		uresp.qid_mask = rhp->rdev.qpmask;
1588 		uresp.sqid = qhp->wq.sq.qid;
1589 		uresp.sq_size = qhp->wq.sq.size;
1590 		uresp.sq_memsize = qhp->wq.sq.memsize;
1591 		uresp.rqid = qhp->wq.rq.qid;
1592 		uresp.rq_size = qhp->wq.rq.size;
1593 		uresp.rq_memsize = qhp->wq.rq.memsize;
1594 		spin_lock(&ucontext->mmap_lock);
1595 		if (mm5) {
1596 			uresp.ma_sync_key = ucontext->key;
1597 			ucontext->key += PAGE_SIZE;
1598 		}
1599 		uresp.sq_key = ucontext->key;
1600 		ucontext->key += PAGE_SIZE;
1601 		uresp.rq_key = ucontext->key;
1602 		ucontext->key += PAGE_SIZE;
1603 		uresp.sq_db_gts_key = ucontext->key;
1604 		ucontext->key += PAGE_SIZE;
1605 		uresp.rq_db_gts_key = ucontext->key;
1606 		ucontext->key += PAGE_SIZE;
1607 		spin_unlock(&ucontext->mmap_lock);
1608 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1609 		if (ret)
1610 			goto err8;
1611 		mm1->key = uresp.sq_key;
1612 		mm1->addr = qhp->wq.sq.phys_addr;
1613 		mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1614 		insert_mmap(ucontext, mm1);
1615 		mm2->key = uresp.rq_key;
1616 		mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1617 		mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1618 		insert_mmap(ucontext, mm2);
1619 		mm3->key = uresp.sq_db_gts_key;
1620 		mm3->addr = qhp->wq.sq.udb;
1621 		mm3->len = PAGE_SIZE;
1622 		insert_mmap(ucontext, mm3);
1623 		mm4->key = uresp.rq_db_gts_key;
1624 		mm4->addr = qhp->wq.rq.udb;
1625 		mm4->len = PAGE_SIZE;
1626 		insert_mmap(ucontext, mm4);
1627 		if (mm5) {
1628 			mm5->key = uresp.ma_sync_key;
1629 			mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1630 				    + A_PCIE_MA_SYNC) & PAGE_MASK;
1631 			mm5->len = PAGE_SIZE;
1632 			insert_mmap(ucontext, mm5);
1633 		}
1634 	}
1635 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1636 	init_timer(&(qhp->timer));
1637 	PDBG("%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x\n",
1638 	     __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1639 	     qhp->wq.sq.qid);
1640 	return &qhp->ibqp;
1641 err8:
1642 	kfree(mm5);
1643 err7:
1644 	kfree(mm4);
1645 err6:
1646 	kfree(mm3);
1647 err5:
1648 	kfree(mm2);
1649 err4:
1650 	kfree(mm1);
1651 err3:
1652 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1653 err2:
1654 	destroy_qp(&rhp->rdev, &qhp->wq,
1655 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1656 err1:
1657 	kfree(qhp);
1658 	return ERR_PTR(ret);
1659 }
1660 
1661 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1662 		      int attr_mask, struct ib_udata *udata)
1663 {
1664 	struct c4iw_dev *rhp;
1665 	struct c4iw_qp *qhp;
1666 	enum c4iw_qp_attr_mask mask = 0;
1667 	struct c4iw_qp_attributes attrs;
1668 
1669 	PDBG("%s ib_qp %p\n", __func__, ibqp);
1670 
1671 	/* iwarp does not support the RTR state */
1672 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1673 		attr_mask &= ~IB_QP_STATE;
1674 
1675 	/* Make sure we still have something left to do */
1676 	if (!attr_mask)
1677 		return 0;
1678 
1679 	memset(&attrs, 0, sizeof attrs);
1680 	qhp = to_c4iw_qp(ibqp);
1681 	rhp = qhp->rhp;
1682 
1683 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1684 	attrs.enable_rdma_read = (attr->qp_access_flags &
1685 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1686 	attrs.enable_rdma_write = (attr->qp_access_flags &
1687 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1688 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1689 
1690 
1691 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1692 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1693 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1694 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1695 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1696 
1697 	/*
1698 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1699 	 * ringing the queue db when we're in DB_FULL mode.
1700 	 */
1701 	attrs.sq_db_inc = attr->sq_psn;
1702 	attrs.rq_db_inc = attr->rq_psn;
1703 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1704 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1705 
1706 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1707 }
1708 
1709 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1710 {
1711 	PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1712 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1713 }
1714 
1715 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1716 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1717 {
1718 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1719 
1720 	memset(attr, 0, sizeof *attr);
1721 	memset(init_attr, 0, sizeof *init_attr);
1722 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1723 	return 0;
1724 }
1725