xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/qp.c (revision 6c870213d6f3a25981c10728f46294a3bed1703f)
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 = 1000;
46 module_param(db_fc_threshold, int, 0644);
47 MODULE_PARM_DESC(db_fc_threshold,
48 		 "QP count/threshold that triggers"
49 		 " automatic db flow control mode (default = 1000)");
50 
51 int db_coalescing_threshold;
52 module_param(db_coalescing_threshold, int, 0644);
53 MODULE_PARM_DESC(db_coalescing_threshold,
54 		 "QP count/threshold that triggers"
55 		 " disabling db coalescing (default = 0)");
56 
57 static int max_fr_immd = T4_MAX_FR_IMMD;
58 module_param(max_fr_immd, int, 0644);
59 MODULE_PARM_DESC(max_fr_immd, "fastreg threshold for using DSGL instead of immedate");
60 
61 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
62 {
63 	unsigned long flag;
64 	spin_lock_irqsave(&qhp->lock, flag);
65 	qhp->attr.state = state;
66 	spin_unlock_irqrestore(&qhp->lock, flag);
67 }
68 
69 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
70 {
71 	c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
72 }
73 
74 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
75 {
76 	dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
77 			  pci_unmap_addr(sq, mapping));
78 }
79 
80 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
81 {
82 	if (t4_sq_onchip(sq))
83 		dealloc_oc_sq(rdev, sq);
84 	else
85 		dealloc_host_sq(rdev, sq);
86 }
87 
88 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
89 {
90 	if (!ocqp_support || !ocqp_supported(&rdev->lldi))
91 		return -ENOSYS;
92 	sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
93 	if (!sq->dma_addr)
94 		return -ENOMEM;
95 	sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
96 			rdev->lldi.vr->ocq.start;
97 	sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
98 					    rdev->lldi.vr->ocq.start);
99 	sq->flags |= T4_SQ_ONCHIP;
100 	return 0;
101 }
102 
103 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
104 {
105 	sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
106 				       &(sq->dma_addr), GFP_KERNEL);
107 	if (!sq->queue)
108 		return -ENOMEM;
109 	sq->phys_addr = virt_to_phys(sq->queue);
110 	pci_unmap_addr_set(sq, mapping, sq->dma_addr);
111 	return 0;
112 }
113 
114 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
115 {
116 	int ret = -ENOSYS;
117 	if (user)
118 		ret = alloc_oc_sq(rdev, sq);
119 	if (ret)
120 		ret = alloc_host_sq(rdev, sq);
121 	return ret;
122 }
123 
124 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
125 		      struct c4iw_dev_ucontext *uctx)
126 {
127 	/*
128 	 * uP clears EQ contexts when the connection exits rdma mode,
129 	 * so no need to post a RESET WR for these EQs.
130 	 */
131 	dma_free_coherent(&(rdev->lldi.pdev->dev),
132 			  wq->rq.memsize, wq->rq.queue,
133 			  dma_unmap_addr(&wq->rq, mapping));
134 	dealloc_sq(rdev, &wq->sq);
135 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
136 	kfree(wq->rq.sw_rq);
137 	kfree(wq->sq.sw_sq);
138 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
139 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
140 	return 0;
141 }
142 
143 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
144 		     struct t4_cq *rcq, struct t4_cq *scq,
145 		     struct c4iw_dev_ucontext *uctx)
146 {
147 	int user = (uctx != &rdev->uctx);
148 	struct fw_ri_res_wr *res_wr;
149 	struct fw_ri_res *res;
150 	int wr_len;
151 	struct c4iw_wr_wait wr_wait;
152 	struct sk_buff *skb;
153 	int ret = 0;
154 	int eqsize;
155 
156 	wq->sq.qid = c4iw_get_qpid(rdev, uctx);
157 	if (!wq->sq.qid)
158 		return -ENOMEM;
159 
160 	wq->rq.qid = c4iw_get_qpid(rdev, uctx);
161 	if (!wq->rq.qid) {
162 		ret = -ENOMEM;
163 		goto free_sq_qid;
164 	}
165 
166 	if (!user) {
167 		wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
168 				 GFP_KERNEL);
169 		if (!wq->sq.sw_sq) {
170 			ret = -ENOMEM;
171 			goto free_rq_qid;
172 		}
173 
174 		wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
175 				 GFP_KERNEL);
176 		if (!wq->rq.sw_rq) {
177 			ret = -ENOMEM;
178 			goto free_sw_sq;
179 		}
180 	}
181 
182 	/*
183 	 * RQT must be a power of 2.
184 	 */
185 	wq->rq.rqt_size = roundup_pow_of_two(wq->rq.size);
186 	wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
187 	if (!wq->rq.rqt_hwaddr) {
188 		ret = -ENOMEM;
189 		goto free_sw_rq;
190 	}
191 
192 	ret = alloc_sq(rdev, &wq->sq, user);
193 	if (ret)
194 		goto free_hwaddr;
195 	memset(wq->sq.queue, 0, wq->sq.memsize);
196 	dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
197 
198 	wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
199 					  wq->rq.memsize, &(wq->rq.dma_addr),
200 					  GFP_KERNEL);
201 	if (!wq->rq.queue) {
202 		ret = -ENOMEM;
203 		goto free_sq;
204 	}
205 	PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
206 		__func__, wq->sq.queue,
207 		(unsigned long long)virt_to_phys(wq->sq.queue),
208 		wq->rq.queue,
209 		(unsigned long long)virt_to_phys(wq->rq.queue));
210 	memset(wq->rq.queue, 0, wq->rq.memsize);
211 	dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
212 
213 	wq->db = rdev->lldi.db_reg;
214 	wq->gts = rdev->lldi.gts_reg;
215 	if (user) {
216 		wq->sq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
217 					(wq->sq.qid << rdev->qpshift);
218 		wq->sq.udb &= PAGE_MASK;
219 		wq->rq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
220 					(wq->rq.qid << rdev->qpshift);
221 		wq->rq.udb &= PAGE_MASK;
222 	}
223 	wq->rdev = rdev;
224 	wq->rq.msn = 1;
225 
226 	/* build fw_ri_res_wr */
227 	wr_len = sizeof *res_wr + 2 * sizeof *res;
228 
229 	skb = alloc_skb(wr_len, GFP_KERNEL);
230 	if (!skb) {
231 		ret = -ENOMEM;
232 		goto free_dma;
233 	}
234 	set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
235 
236 	res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
237 	memset(res_wr, 0, wr_len);
238 	res_wr->op_nres = cpu_to_be32(
239 			FW_WR_OP(FW_RI_RES_WR) |
240 			V_FW_RI_RES_WR_NRES(2) |
241 			FW_WR_COMPL(1));
242 	res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
243 	res_wr->cookie = (unsigned long) &wr_wait;
244 	res = res_wr->res;
245 	res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
246 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
247 
248 	/*
249 	 * eqsize is the number of 64B entries plus the status page size.
250 	 */
251 	eqsize = wq->sq.size * T4_SQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
252 
253 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
254 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
255 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
256 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
257 		(t4_sq_onchip(&wq->sq) ? F_FW_RI_RES_WR_ONCHIP : 0) |
258 		V_FW_RI_RES_WR_IQID(scq->cqid));
259 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
260 		V_FW_RI_RES_WR_DCAEN(0) |
261 		V_FW_RI_RES_WR_DCACPU(0) |
262 		V_FW_RI_RES_WR_FBMIN(2) |
263 		V_FW_RI_RES_WR_FBMAX(2) |
264 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
265 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
266 		V_FW_RI_RES_WR_EQSIZE(eqsize));
267 	res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
268 	res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
269 	res++;
270 	res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
271 	res->u.sqrq.op = FW_RI_RES_OP_WRITE;
272 
273 	/*
274 	 * eqsize is the number of 64B entries plus the status page size.
275 	 */
276 	eqsize = wq->rq.size * T4_RQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
277 	res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
278 		V_FW_RI_RES_WR_HOSTFCMODE(0) |	/* no host cidx updates */
279 		V_FW_RI_RES_WR_CPRIO(0) |	/* don't keep in chip cache */
280 		V_FW_RI_RES_WR_PCIECHN(0) |	/* set by uP at ri_init time */
281 		V_FW_RI_RES_WR_IQID(rcq->cqid));
282 	res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
283 		V_FW_RI_RES_WR_DCAEN(0) |
284 		V_FW_RI_RES_WR_DCACPU(0) |
285 		V_FW_RI_RES_WR_FBMIN(2) |
286 		V_FW_RI_RES_WR_FBMAX(2) |
287 		V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
288 		V_FW_RI_RES_WR_CIDXFTHRESH(0) |
289 		V_FW_RI_RES_WR_EQSIZE(eqsize));
290 	res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
291 	res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
292 
293 	c4iw_init_wr_wait(&wr_wait);
294 
295 	ret = c4iw_ofld_send(rdev, skb);
296 	if (ret)
297 		goto free_dma;
298 	ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
299 	if (ret)
300 		goto free_dma;
301 
302 	PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx\n",
303 	     __func__, wq->sq.qid, wq->rq.qid, wq->db,
304 	     (unsigned long long)wq->sq.udb, (unsigned long long)wq->rq.udb);
305 
306 	return 0;
307 free_dma:
308 	dma_free_coherent(&(rdev->lldi.pdev->dev),
309 			  wq->rq.memsize, wq->rq.queue,
310 			  dma_unmap_addr(&wq->rq, mapping));
311 free_sq:
312 	dealloc_sq(rdev, &wq->sq);
313 free_hwaddr:
314 	c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
315 free_sw_rq:
316 	kfree(wq->rq.sw_rq);
317 free_sw_sq:
318 	kfree(wq->sq.sw_sq);
319 free_rq_qid:
320 	c4iw_put_qpid(rdev, wq->rq.qid, uctx);
321 free_sq_qid:
322 	c4iw_put_qpid(rdev, wq->sq.qid, uctx);
323 	return ret;
324 }
325 
326 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
327 		      struct ib_send_wr *wr, int max, u32 *plenp)
328 {
329 	u8 *dstp, *srcp;
330 	u32 plen = 0;
331 	int i;
332 	int rem, len;
333 
334 	dstp = (u8 *)immdp->data;
335 	for (i = 0; i < wr->num_sge; i++) {
336 		if ((plen + wr->sg_list[i].length) > max)
337 			return -EMSGSIZE;
338 		srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
339 		plen += wr->sg_list[i].length;
340 		rem = wr->sg_list[i].length;
341 		while (rem) {
342 			if (dstp == (u8 *)&sq->queue[sq->size])
343 				dstp = (u8 *)sq->queue;
344 			if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
345 				len = rem;
346 			else
347 				len = (u8 *)&sq->queue[sq->size] - dstp;
348 			memcpy(dstp, srcp, len);
349 			dstp += len;
350 			srcp += len;
351 			rem -= len;
352 		}
353 	}
354 	len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
355 	if (len)
356 		memset(dstp, 0, len);
357 	immdp->op = FW_RI_DATA_IMMD;
358 	immdp->r1 = 0;
359 	immdp->r2 = 0;
360 	immdp->immdlen = cpu_to_be32(plen);
361 	*plenp = plen;
362 	return 0;
363 }
364 
365 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
366 		      struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
367 		      int num_sge, u32 *plenp)
368 
369 {
370 	int i;
371 	u32 plen = 0;
372 	__be64 *flitp = (__be64 *)isglp->sge;
373 
374 	for (i = 0; i < num_sge; i++) {
375 		if ((plen + sg_list[i].length) < plen)
376 			return -EMSGSIZE;
377 		plen += sg_list[i].length;
378 		*flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
379 				     sg_list[i].length);
380 		if (++flitp == queue_end)
381 			flitp = queue_start;
382 		*flitp = cpu_to_be64(sg_list[i].addr);
383 		if (++flitp == queue_end)
384 			flitp = queue_start;
385 	}
386 	*flitp = (__force __be64)0;
387 	isglp->op = FW_RI_DATA_ISGL;
388 	isglp->r1 = 0;
389 	isglp->nsge = cpu_to_be16(num_sge);
390 	isglp->r2 = 0;
391 	if (plenp)
392 		*plenp = plen;
393 	return 0;
394 }
395 
396 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
397 			   struct ib_send_wr *wr, u8 *len16)
398 {
399 	u32 plen;
400 	int size;
401 	int ret;
402 
403 	if (wr->num_sge > T4_MAX_SEND_SGE)
404 		return -EINVAL;
405 	switch (wr->opcode) {
406 	case IB_WR_SEND:
407 		if (wr->send_flags & IB_SEND_SOLICITED)
408 			wqe->send.sendop_pkd = cpu_to_be32(
409 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
410 		else
411 			wqe->send.sendop_pkd = cpu_to_be32(
412 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
413 		wqe->send.stag_inv = 0;
414 		break;
415 	case IB_WR_SEND_WITH_INV:
416 		if (wr->send_flags & IB_SEND_SOLICITED)
417 			wqe->send.sendop_pkd = cpu_to_be32(
418 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
419 		else
420 			wqe->send.sendop_pkd = cpu_to_be32(
421 				V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
422 		wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
423 		break;
424 
425 	default:
426 		return -EINVAL;
427 	}
428 
429 	plen = 0;
430 	if (wr->num_sge) {
431 		if (wr->send_flags & IB_SEND_INLINE) {
432 			ret = build_immd(sq, wqe->send.u.immd_src, wr,
433 					 T4_MAX_SEND_INLINE, &plen);
434 			if (ret)
435 				return ret;
436 			size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
437 			       plen;
438 		} else {
439 			ret = build_isgl((__be64 *)sq->queue,
440 					 (__be64 *)&sq->queue[sq->size],
441 					 wqe->send.u.isgl_src,
442 					 wr->sg_list, wr->num_sge, &plen);
443 			if (ret)
444 				return ret;
445 			size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
446 			       wr->num_sge * sizeof(struct fw_ri_sge);
447 		}
448 	} else {
449 		wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
450 		wqe->send.u.immd_src[0].r1 = 0;
451 		wqe->send.u.immd_src[0].r2 = 0;
452 		wqe->send.u.immd_src[0].immdlen = 0;
453 		size = sizeof wqe->send + sizeof(struct fw_ri_immd);
454 		plen = 0;
455 	}
456 	*len16 = DIV_ROUND_UP(size, 16);
457 	wqe->send.plen = cpu_to_be32(plen);
458 	return 0;
459 }
460 
461 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
462 			    struct ib_send_wr *wr, u8 *len16)
463 {
464 	u32 plen;
465 	int size;
466 	int ret;
467 
468 	if (wr->num_sge > T4_MAX_SEND_SGE)
469 		return -EINVAL;
470 	wqe->write.r2 = 0;
471 	wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
472 	wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
473 	if (wr->num_sge) {
474 		if (wr->send_flags & IB_SEND_INLINE) {
475 			ret = build_immd(sq, wqe->write.u.immd_src, wr,
476 					 T4_MAX_WRITE_INLINE, &plen);
477 			if (ret)
478 				return ret;
479 			size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
480 			       plen;
481 		} else {
482 			ret = build_isgl((__be64 *)sq->queue,
483 					 (__be64 *)&sq->queue[sq->size],
484 					 wqe->write.u.isgl_src,
485 					 wr->sg_list, wr->num_sge, &plen);
486 			if (ret)
487 				return ret;
488 			size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
489 			       wr->num_sge * sizeof(struct fw_ri_sge);
490 		}
491 	} else {
492 		wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
493 		wqe->write.u.immd_src[0].r1 = 0;
494 		wqe->write.u.immd_src[0].r2 = 0;
495 		wqe->write.u.immd_src[0].immdlen = 0;
496 		size = sizeof wqe->write + sizeof(struct fw_ri_immd);
497 		plen = 0;
498 	}
499 	*len16 = DIV_ROUND_UP(size, 16);
500 	wqe->write.plen = cpu_to_be32(plen);
501 	return 0;
502 }
503 
504 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
505 {
506 	if (wr->num_sge > 1)
507 		return -EINVAL;
508 	if (wr->num_sge) {
509 		wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
510 		wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
511 							>> 32));
512 		wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
513 		wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
514 		wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
515 		wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
516 							 >> 32));
517 		wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
518 	} else {
519 		wqe->read.stag_src = cpu_to_be32(2);
520 		wqe->read.to_src_hi = 0;
521 		wqe->read.to_src_lo = 0;
522 		wqe->read.stag_sink = cpu_to_be32(2);
523 		wqe->read.plen = 0;
524 		wqe->read.to_sink_hi = 0;
525 		wqe->read.to_sink_lo = 0;
526 	}
527 	wqe->read.r2 = 0;
528 	wqe->read.r5 = 0;
529 	*len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
530 	return 0;
531 }
532 
533 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
534 			   struct ib_recv_wr *wr, u8 *len16)
535 {
536 	int ret;
537 
538 	ret = build_isgl((__be64 *)qhp->wq.rq.queue,
539 			 (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
540 			 &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
541 	if (ret)
542 		return ret;
543 	*len16 = DIV_ROUND_UP(sizeof wqe->recv +
544 			      wr->num_sge * sizeof(struct fw_ri_sge), 16);
545 	return 0;
546 }
547 
548 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
549 			 struct ib_send_wr *wr, u8 *len16, u8 t5dev)
550 {
551 
552 	struct fw_ri_immd *imdp;
553 	__be64 *p;
554 	int i;
555 	int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
556 	int rem;
557 
558 	if (wr->wr.fast_reg.page_list_len > T4_MAX_FR_DEPTH)
559 		return -EINVAL;
560 
561 	wqe->fr.qpbinde_to_dcacpu = 0;
562 	wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
563 	wqe->fr.addr_type = FW_RI_VA_BASED_TO;
564 	wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
565 	wqe->fr.len_hi = 0;
566 	wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
567 	wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
568 	wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
569 	wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
570 					0xffffffff);
571 
572 	if (t5dev && use_dsgl && (pbllen > max_fr_immd)) {
573 		struct c4iw_fr_page_list *c4pl =
574 			to_c4iw_fr_page_list(wr->wr.fast_reg.page_list);
575 		struct fw_ri_dsgl *sglp;
576 
577 		for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
578 			wr->wr.fast_reg.page_list->page_list[i] = (__force u64)
579 				cpu_to_be64((u64)
580 				wr->wr.fast_reg.page_list->page_list[i]);
581 		}
582 
583 		sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
584 		sglp->op = FW_RI_DATA_DSGL;
585 		sglp->r1 = 0;
586 		sglp->nsge = cpu_to_be16(1);
587 		sglp->addr0 = cpu_to_be64(c4pl->dma_addr);
588 		sglp->len0 = cpu_to_be32(pbllen);
589 
590 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
591 	} else {
592 		imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
593 		imdp->op = FW_RI_DATA_IMMD;
594 		imdp->r1 = 0;
595 		imdp->r2 = 0;
596 		imdp->immdlen = cpu_to_be32(pbllen);
597 		p = (__be64 *)(imdp + 1);
598 		rem = pbllen;
599 		for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
600 			*p = cpu_to_be64(
601 				(u64)wr->wr.fast_reg.page_list->page_list[i]);
602 			rem -= sizeof(*p);
603 			if (++p == (__be64 *)&sq->queue[sq->size])
604 				p = (__be64 *)sq->queue;
605 		}
606 		BUG_ON(rem < 0);
607 		while (rem) {
608 			*p = 0;
609 			rem -= sizeof(*p);
610 			if (++p == (__be64 *)&sq->queue[sq->size])
611 				p = (__be64 *)sq->queue;
612 		}
613 		*len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
614 				      + pbllen, 16);
615 	}
616 	return 0;
617 }
618 
619 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
620 			  u8 *len16)
621 {
622 	wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
623 	wqe->inv.r2 = 0;
624 	*len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
625 	return 0;
626 }
627 
628 void c4iw_qp_add_ref(struct ib_qp *qp)
629 {
630 	PDBG("%s ib_qp %p\n", __func__, qp);
631 	atomic_inc(&(to_c4iw_qp(qp)->refcnt));
632 }
633 
634 void c4iw_qp_rem_ref(struct ib_qp *qp)
635 {
636 	PDBG("%s ib_qp %p\n", __func__, qp);
637 	if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
638 		wake_up(&(to_c4iw_qp(qp)->wait));
639 }
640 
641 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
642 {
643 	if (list_empty(entry))
644 		list_add_tail(entry, head);
645 }
646 
647 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
648 {
649 	unsigned long flags;
650 
651 	spin_lock_irqsave(&qhp->rhp->lock, flags);
652 	spin_lock(&qhp->lock);
653 	if (qhp->rhp->db_state == NORMAL) {
654 		t4_ring_sq_db(&qhp->wq, inc);
655 	} else {
656 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
657 		qhp->wq.sq.wq_pidx_inc += inc;
658 	}
659 	spin_unlock(&qhp->lock);
660 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
661 	return 0;
662 }
663 
664 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
665 {
666 	unsigned long flags;
667 
668 	spin_lock_irqsave(&qhp->rhp->lock, flags);
669 	spin_lock(&qhp->lock);
670 	if (qhp->rhp->db_state == NORMAL) {
671 		t4_ring_rq_db(&qhp->wq, inc);
672 	} else {
673 		add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
674 		qhp->wq.rq.wq_pidx_inc += inc;
675 	}
676 	spin_unlock(&qhp->lock);
677 	spin_unlock_irqrestore(&qhp->rhp->lock, flags);
678 	return 0;
679 }
680 
681 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
682 		   struct ib_send_wr **bad_wr)
683 {
684 	int err = 0;
685 	u8 len16 = 0;
686 	enum fw_wr_opcodes fw_opcode = 0;
687 	enum fw_ri_wr_flags fw_flags;
688 	struct c4iw_qp *qhp;
689 	union t4_wr *wqe;
690 	u32 num_wrs;
691 	struct t4_swsqe *swsqe;
692 	unsigned long flag;
693 	u16 idx = 0;
694 
695 	qhp = to_c4iw_qp(ibqp);
696 	spin_lock_irqsave(&qhp->lock, flag);
697 	if (t4_wq_in_error(&qhp->wq)) {
698 		spin_unlock_irqrestore(&qhp->lock, flag);
699 		return -EINVAL;
700 	}
701 	num_wrs = t4_sq_avail(&qhp->wq);
702 	if (num_wrs == 0) {
703 		spin_unlock_irqrestore(&qhp->lock, flag);
704 		return -ENOMEM;
705 	}
706 	while (wr) {
707 		if (num_wrs == 0) {
708 			err = -ENOMEM;
709 			*bad_wr = wr;
710 			break;
711 		}
712 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
713 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
714 
715 		fw_flags = 0;
716 		if (wr->send_flags & IB_SEND_SOLICITED)
717 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
718 		if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
719 			fw_flags |= FW_RI_COMPLETION_FLAG;
720 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
721 		switch (wr->opcode) {
722 		case IB_WR_SEND_WITH_INV:
723 		case IB_WR_SEND:
724 			if (wr->send_flags & IB_SEND_FENCE)
725 				fw_flags |= FW_RI_READ_FENCE_FLAG;
726 			fw_opcode = FW_RI_SEND_WR;
727 			if (wr->opcode == IB_WR_SEND)
728 				swsqe->opcode = FW_RI_SEND;
729 			else
730 				swsqe->opcode = FW_RI_SEND_WITH_INV;
731 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
732 			break;
733 		case IB_WR_RDMA_WRITE:
734 			fw_opcode = FW_RI_RDMA_WRITE_WR;
735 			swsqe->opcode = FW_RI_RDMA_WRITE;
736 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
737 			break;
738 		case IB_WR_RDMA_READ:
739 		case IB_WR_RDMA_READ_WITH_INV:
740 			fw_opcode = FW_RI_RDMA_READ_WR;
741 			swsqe->opcode = FW_RI_READ_REQ;
742 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
743 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
744 			else
745 				fw_flags = 0;
746 			err = build_rdma_read(wqe, wr, &len16);
747 			if (err)
748 				break;
749 			swsqe->read_len = wr->sg_list[0].length;
750 			if (!qhp->wq.sq.oldest_read)
751 				qhp->wq.sq.oldest_read = swsqe;
752 			break;
753 		case IB_WR_FAST_REG_MR:
754 			fw_opcode = FW_RI_FR_NSMR_WR;
755 			swsqe->opcode = FW_RI_FAST_REGISTER;
756 			err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16,
757 					    is_t5(
758 					    qhp->rhp->rdev.lldi.adapter_type) ?
759 					    1 : 0);
760 			break;
761 		case IB_WR_LOCAL_INV:
762 			if (wr->send_flags & IB_SEND_FENCE)
763 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
764 			fw_opcode = FW_RI_INV_LSTAG_WR;
765 			swsqe->opcode = FW_RI_LOCAL_INV;
766 			err = build_inv_stag(wqe, wr, &len16);
767 			break;
768 		default:
769 			PDBG("%s post of type=%d TBD!\n", __func__,
770 			     wr->opcode);
771 			err = -EINVAL;
772 		}
773 		if (err) {
774 			*bad_wr = wr;
775 			break;
776 		}
777 		swsqe->idx = qhp->wq.sq.pidx;
778 		swsqe->complete = 0;
779 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
780 				  qhp->sq_sig_all;
781 		swsqe->flushed = 0;
782 		swsqe->wr_id = wr->wr_id;
783 
784 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
785 
786 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
787 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
788 		     swsqe->opcode, swsqe->read_len);
789 		wr = wr->next;
790 		num_wrs--;
791 		t4_sq_produce(&qhp->wq, len16);
792 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
793 	}
794 	if (!qhp->rhp->rdev.status_page->db_off) {
795 		t4_ring_sq_db(&qhp->wq, idx);
796 		spin_unlock_irqrestore(&qhp->lock, flag);
797 	} else {
798 		spin_unlock_irqrestore(&qhp->lock, flag);
799 		ring_kernel_sq_db(qhp, idx);
800 	}
801 	return err;
802 }
803 
804 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
805 		      struct ib_recv_wr **bad_wr)
806 {
807 	int err = 0;
808 	struct c4iw_qp *qhp;
809 	union t4_recv_wr *wqe;
810 	u32 num_wrs;
811 	u8 len16 = 0;
812 	unsigned long flag;
813 	u16 idx = 0;
814 
815 	qhp = to_c4iw_qp(ibqp);
816 	spin_lock_irqsave(&qhp->lock, flag);
817 	if (t4_wq_in_error(&qhp->wq)) {
818 		spin_unlock_irqrestore(&qhp->lock, flag);
819 		return -EINVAL;
820 	}
821 	num_wrs = t4_rq_avail(&qhp->wq);
822 	if (num_wrs == 0) {
823 		spin_unlock_irqrestore(&qhp->lock, flag);
824 		return -ENOMEM;
825 	}
826 	while (wr) {
827 		if (wr->num_sge > T4_MAX_RECV_SGE) {
828 			err = -EINVAL;
829 			*bad_wr = wr;
830 			break;
831 		}
832 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
833 					   qhp->wq.rq.wq_pidx *
834 					   T4_EQ_ENTRY_SIZE);
835 		if (num_wrs)
836 			err = build_rdma_recv(qhp, wqe, wr, &len16);
837 		else
838 			err = -ENOMEM;
839 		if (err) {
840 			*bad_wr = wr;
841 			break;
842 		}
843 
844 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
845 
846 		wqe->recv.opcode = FW_RI_RECV_WR;
847 		wqe->recv.r1 = 0;
848 		wqe->recv.wrid = qhp->wq.rq.pidx;
849 		wqe->recv.r2[0] = 0;
850 		wqe->recv.r2[1] = 0;
851 		wqe->recv.r2[2] = 0;
852 		wqe->recv.len16 = len16;
853 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
854 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
855 		t4_rq_produce(&qhp->wq, len16);
856 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
857 		wr = wr->next;
858 		num_wrs--;
859 	}
860 	if (!qhp->rhp->rdev.status_page->db_off) {
861 		t4_ring_rq_db(&qhp->wq, idx);
862 		spin_unlock_irqrestore(&qhp->lock, flag);
863 	} else {
864 		spin_unlock_irqrestore(&qhp->lock, flag);
865 		ring_kernel_rq_db(qhp, idx);
866 	}
867 	return err;
868 }
869 
870 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
871 {
872 	return -ENOSYS;
873 }
874 
875 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
876 				    u8 *ecode)
877 {
878 	int status;
879 	int tagged;
880 	int opcode;
881 	int rqtype;
882 	int send_inv;
883 
884 	if (!err_cqe) {
885 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
886 		*ecode = 0;
887 		return;
888 	}
889 
890 	status = CQE_STATUS(err_cqe);
891 	opcode = CQE_OPCODE(err_cqe);
892 	rqtype = RQ_TYPE(err_cqe);
893 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
894 		   (opcode == FW_RI_SEND_WITH_SE_INV);
895 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
896 		 (rqtype && (opcode == FW_RI_READ_RESP));
897 
898 	switch (status) {
899 	case T4_ERR_STAG:
900 		if (send_inv) {
901 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
902 			*ecode = RDMAP_CANT_INV_STAG;
903 		} else {
904 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
905 			*ecode = RDMAP_INV_STAG;
906 		}
907 		break;
908 	case T4_ERR_PDID:
909 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
910 		if ((opcode == FW_RI_SEND_WITH_INV) ||
911 		    (opcode == FW_RI_SEND_WITH_SE_INV))
912 			*ecode = RDMAP_CANT_INV_STAG;
913 		else
914 			*ecode = RDMAP_STAG_NOT_ASSOC;
915 		break;
916 	case T4_ERR_QPID:
917 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
918 		*ecode = RDMAP_STAG_NOT_ASSOC;
919 		break;
920 	case T4_ERR_ACCESS:
921 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
922 		*ecode = RDMAP_ACC_VIOL;
923 		break;
924 	case T4_ERR_WRAP:
925 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
926 		*ecode = RDMAP_TO_WRAP;
927 		break;
928 	case T4_ERR_BOUND:
929 		if (tagged) {
930 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
931 			*ecode = DDPT_BASE_BOUNDS;
932 		} else {
933 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
934 			*ecode = RDMAP_BASE_BOUNDS;
935 		}
936 		break;
937 	case T4_ERR_INVALIDATE_SHARED_MR:
938 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
939 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
940 		*ecode = RDMAP_CANT_INV_STAG;
941 		break;
942 	case T4_ERR_ECC:
943 	case T4_ERR_ECC_PSTAG:
944 	case T4_ERR_INTERNAL_ERR:
945 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
946 		*ecode = 0;
947 		break;
948 	case T4_ERR_OUT_OF_RQE:
949 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
950 		*ecode = DDPU_INV_MSN_NOBUF;
951 		break;
952 	case T4_ERR_PBL_ADDR_BOUND:
953 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
954 		*ecode = DDPT_BASE_BOUNDS;
955 		break;
956 	case T4_ERR_CRC:
957 		*layer_type = LAYER_MPA|DDP_LLP;
958 		*ecode = MPA_CRC_ERR;
959 		break;
960 	case T4_ERR_MARKER:
961 		*layer_type = LAYER_MPA|DDP_LLP;
962 		*ecode = MPA_MARKER_ERR;
963 		break;
964 	case T4_ERR_PDU_LEN_ERR:
965 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
966 		*ecode = DDPU_MSG_TOOBIG;
967 		break;
968 	case T4_ERR_DDP_VERSION:
969 		if (tagged) {
970 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
971 			*ecode = DDPT_INV_VERS;
972 		} else {
973 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
974 			*ecode = DDPU_INV_VERS;
975 		}
976 		break;
977 	case T4_ERR_RDMA_VERSION:
978 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
979 		*ecode = RDMAP_INV_VERS;
980 		break;
981 	case T4_ERR_OPCODE:
982 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
983 		*ecode = RDMAP_INV_OPCODE;
984 		break;
985 	case T4_ERR_DDP_QUEUE_NUM:
986 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
987 		*ecode = DDPU_INV_QN;
988 		break;
989 	case T4_ERR_MSN:
990 	case T4_ERR_MSN_GAP:
991 	case T4_ERR_MSN_RANGE:
992 	case T4_ERR_IRD_OVERFLOW:
993 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
994 		*ecode = DDPU_INV_MSN_RANGE;
995 		break;
996 	case T4_ERR_TBIT:
997 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
998 		*ecode = 0;
999 		break;
1000 	case T4_ERR_MO:
1001 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1002 		*ecode = DDPU_INV_MO;
1003 		break;
1004 	default:
1005 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1006 		*ecode = 0;
1007 		break;
1008 	}
1009 }
1010 
1011 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1012 			   gfp_t gfp)
1013 {
1014 	struct fw_ri_wr *wqe;
1015 	struct sk_buff *skb;
1016 	struct terminate_message *term;
1017 
1018 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1019 	     qhp->ep->hwtid);
1020 
1021 	skb = alloc_skb(sizeof *wqe, gfp);
1022 	if (!skb)
1023 		return;
1024 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1025 
1026 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1027 	memset(wqe, 0, sizeof *wqe);
1028 	wqe->op_compl = cpu_to_be32(FW_WR_OP(FW_RI_INIT_WR));
1029 	wqe->flowid_len16 = cpu_to_be32(
1030 		FW_WR_FLOWID(qhp->ep->hwtid) |
1031 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1032 
1033 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1034 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1035 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
1036 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1037 		term->layer_etype = qhp->attr.layer_etype;
1038 		term->ecode = qhp->attr.ecode;
1039 	} else
1040 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1041 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
1042 }
1043 
1044 /*
1045  * Assumes qhp lock is held.
1046  */
1047 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1048 		       struct c4iw_cq *schp)
1049 {
1050 	int count;
1051 	int flushed;
1052 	unsigned long flag;
1053 
1054 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1055 
1056 	/* locking hierarchy: cq lock first, then qp lock. */
1057 	spin_lock_irqsave(&rchp->lock, flag);
1058 	spin_lock(&qhp->lock);
1059 
1060 	if (qhp->wq.flushed) {
1061 		spin_unlock(&qhp->lock);
1062 		spin_unlock_irqrestore(&rchp->lock, flag);
1063 		return;
1064 	}
1065 	qhp->wq.flushed = 1;
1066 
1067 	c4iw_flush_hw_cq(rchp);
1068 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1069 	flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1070 	spin_unlock(&qhp->lock);
1071 	spin_unlock_irqrestore(&rchp->lock, flag);
1072 	if (flushed) {
1073 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1074 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1075 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1076 	}
1077 
1078 	/* locking hierarchy: cq lock first, then qp lock. */
1079 	spin_lock_irqsave(&schp->lock, flag);
1080 	spin_lock(&qhp->lock);
1081 	if (schp != rchp)
1082 		c4iw_flush_hw_cq(schp);
1083 	flushed = c4iw_flush_sq(qhp);
1084 	spin_unlock(&qhp->lock);
1085 	spin_unlock_irqrestore(&schp->lock, flag);
1086 	if (flushed) {
1087 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
1088 		(*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
1089 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1090 	}
1091 }
1092 
1093 static void flush_qp(struct c4iw_qp *qhp)
1094 {
1095 	struct c4iw_cq *rchp, *schp;
1096 	unsigned long flag;
1097 
1098 	rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1099 	schp = to_c4iw_cq(qhp->ibqp.send_cq);
1100 
1101 	t4_set_wq_in_error(&qhp->wq);
1102 	if (qhp->ibqp.uobject) {
1103 		t4_set_cq_in_error(&rchp->cq);
1104 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1105 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1106 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1107 		if (schp != rchp) {
1108 			t4_set_cq_in_error(&schp->cq);
1109 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1110 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1111 					schp->ibcq.cq_context);
1112 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1113 		}
1114 		return;
1115 	}
1116 	__flush_qp(qhp, rchp, schp);
1117 }
1118 
1119 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1120 		     struct c4iw_ep *ep)
1121 {
1122 	struct fw_ri_wr *wqe;
1123 	int ret;
1124 	struct sk_buff *skb;
1125 
1126 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1127 	     ep->hwtid);
1128 
1129 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1130 	if (!skb)
1131 		return -ENOMEM;
1132 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1133 
1134 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1135 	memset(wqe, 0, sizeof *wqe);
1136 	wqe->op_compl = cpu_to_be32(
1137 		FW_WR_OP(FW_RI_INIT_WR) |
1138 		FW_WR_COMPL(1));
1139 	wqe->flowid_len16 = cpu_to_be32(
1140 		FW_WR_FLOWID(ep->hwtid) |
1141 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1142 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1143 
1144 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1145 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1146 	if (ret)
1147 		goto out;
1148 
1149 	ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1150 			     qhp->wq.sq.qid, __func__);
1151 out:
1152 	PDBG("%s ret %d\n", __func__, ret);
1153 	return ret;
1154 }
1155 
1156 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1157 {
1158 	PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1159 	memset(&init->u, 0, sizeof init->u);
1160 	switch (p2p_type) {
1161 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1162 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1163 		init->u.write.stag_sink = cpu_to_be32(1);
1164 		init->u.write.to_sink = cpu_to_be64(1);
1165 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1166 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1167 						   sizeof(struct fw_ri_immd),
1168 						   16);
1169 		break;
1170 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1171 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1172 		init->u.read.stag_src = cpu_to_be32(1);
1173 		init->u.read.to_src_lo = cpu_to_be32(1);
1174 		init->u.read.stag_sink = cpu_to_be32(1);
1175 		init->u.read.to_sink_lo = cpu_to_be32(1);
1176 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1177 		break;
1178 	}
1179 }
1180 
1181 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1182 {
1183 	struct fw_ri_wr *wqe;
1184 	int ret;
1185 	struct sk_buff *skb;
1186 
1187 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1188 	     qhp->ep->hwtid);
1189 
1190 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1191 	if (!skb)
1192 		return -ENOMEM;
1193 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1194 
1195 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1196 	memset(wqe, 0, sizeof *wqe);
1197 	wqe->op_compl = cpu_to_be32(
1198 		FW_WR_OP(FW_RI_INIT_WR) |
1199 		FW_WR_COMPL(1));
1200 	wqe->flowid_len16 = cpu_to_be32(
1201 		FW_WR_FLOWID(qhp->ep->hwtid) |
1202 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1203 
1204 	wqe->cookie = (unsigned long) &qhp->ep->com.wr_wait;
1205 
1206 	wqe->u.init.type = FW_RI_TYPE_INIT;
1207 	wqe->u.init.mpareqbit_p2ptype =
1208 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1209 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1210 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1211 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1212 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1213 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1214 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1215 	if (qhp->attr.mpa_attr.crc_enabled)
1216 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1217 
1218 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1219 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1220 			    FW_RI_QP_BIND_ENABLE;
1221 	if (!qhp->ibqp.uobject)
1222 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1223 				     FW_RI_QP_STAG0_ENABLE;
1224 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1225 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1226 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1227 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1228 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1229 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1230 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1231 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1232 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1233 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1234 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1235 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1236 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1237 					 rhp->rdev.lldi.vr->rq.start);
1238 	if (qhp->attr.mpa_attr.initiator)
1239 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1240 
1241 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1242 	if (ret)
1243 		goto out;
1244 
1245 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1246 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1247 out:
1248 	PDBG("%s ret %d\n", __func__, ret);
1249 	return ret;
1250 }
1251 
1252 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1253 		   enum c4iw_qp_attr_mask mask,
1254 		   struct c4iw_qp_attributes *attrs,
1255 		   int internal)
1256 {
1257 	int ret = 0;
1258 	struct c4iw_qp_attributes newattr = qhp->attr;
1259 	int disconnect = 0;
1260 	int terminate = 0;
1261 	int abort = 0;
1262 	int free = 0;
1263 	struct c4iw_ep *ep = NULL;
1264 
1265 	PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1266 	     qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1267 	     (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1268 
1269 	mutex_lock(&qhp->mutex);
1270 
1271 	/* Process attr changes if in IDLE */
1272 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1273 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1274 			ret = -EIO;
1275 			goto out;
1276 		}
1277 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1278 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1279 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1280 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1281 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1282 			newattr.enable_bind = attrs->enable_bind;
1283 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1284 			if (attrs->max_ord > c4iw_max_read_depth) {
1285 				ret = -EINVAL;
1286 				goto out;
1287 			}
1288 			newattr.max_ord = attrs->max_ord;
1289 		}
1290 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1291 			if (attrs->max_ird > c4iw_max_read_depth) {
1292 				ret = -EINVAL;
1293 				goto out;
1294 			}
1295 			newattr.max_ird = attrs->max_ird;
1296 		}
1297 		qhp->attr = newattr;
1298 	}
1299 
1300 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1301 		ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1302 		goto out;
1303 	}
1304 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1305 		ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1306 		goto out;
1307 	}
1308 
1309 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1310 		goto out;
1311 	if (qhp->attr.state == attrs->next_state)
1312 		goto out;
1313 
1314 	switch (qhp->attr.state) {
1315 	case C4IW_QP_STATE_IDLE:
1316 		switch (attrs->next_state) {
1317 		case C4IW_QP_STATE_RTS:
1318 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1319 				ret = -EINVAL;
1320 				goto out;
1321 			}
1322 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1323 				ret = -EINVAL;
1324 				goto out;
1325 			}
1326 			qhp->attr.mpa_attr = attrs->mpa_attr;
1327 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1328 			qhp->ep = qhp->attr.llp_stream_handle;
1329 			set_state(qhp, C4IW_QP_STATE_RTS);
1330 
1331 			/*
1332 			 * Ref the endpoint here and deref when we
1333 			 * disassociate the endpoint from the QP.  This
1334 			 * happens in CLOSING->IDLE transition or *->ERROR
1335 			 * transition.
1336 			 */
1337 			c4iw_get_ep(&qhp->ep->com);
1338 			ret = rdma_init(rhp, qhp);
1339 			if (ret)
1340 				goto err;
1341 			break;
1342 		case C4IW_QP_STATE_ERROR:
1343 			set_state(qhp, C4IW_QP_STATE_ERROR);
1344 			flush_qp(qhp);
1345 			break;
1346 		default:
1347 			ret = -EINVAL;
1348 			goto out;
1349 		}
1350 		break;
1351 	case C4IW_QP_STATE_RTS:
1352 		switch (attrs->next_state) {
1353 		case C4IW_QP_STATE_CLOSING:
1354 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1355 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1356 			ep = qhp->ep;
1357 			if (!internal) {
1358 				abort = 0;
1359 				disconnect = 1;
1360 				c4iw_get_ep(&qhp->ep->com);
1361 			}
1362 			t4_set_wq_in_error(&qhp->wq);
1363 			ret = rdma_fini(rhp, qhp, ep);
1364 			if (ret)
1365 				goto err;
1366 			break;
1367 		case C4IW_QP_STATE_TERMINATE:
1368 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1369 			qhp->attr.layer_etype = attrs->layer_etype;
1370 			qhp->attr.ecode = attrs->ecode;
1371 			t4_set_wq_in_error(&qhp->wq);
1372 			ep = qhp->ep;
1373 			disconnect = 1;
1374 			if (!internal)
1375 				terminate = 1;
1376 			else {
1377 				ret = rdma_fini(rhp, qhp, ep);
1378 				if (ret)
1379 					goto err;
1380 			}
1381 			c4iw_get_ep(&qhp->ep->com);
1382 			break;
1383 		case C4IW_QP_STATE_ERROR:
1384 			set_state(qhp, C4IW_QP_STATE_ERROR);
1385 			t4_set_wq_in_error(&qhp->wq);
1386 			if (!internal) {
1387 				abort = 1;
1388 				disconnect = 1;
1389 				ep = qhp->ep;
1390 				c4iw_get_ep(&qhp->ep->com);
1391 			}
1392 			goto err;
1393 			break;
1394 		default:
1395 			ret = -EINVAL;
1396 			goto out;
1397 		}
1398 		break;
1399 	case C4IW_QP_STATE_CLOSING:
1400 		if (!internal) {
1401 			ret = -EINVAL;
1402 			goto out;
1403 		}
1404 		switch (attrs->next_state) {
1405 		case C4IW_QP_STATE_IDLE:
1406 			flush_qp(qhp);
1407 			set_state(qhp, C4IW_QP_STATE_IDLE);
1408 			qhp->attr.llp_stream_handle = NULL;
1409 			c4iw_put_ep(&qhp->ep->com);
1410 			qhp->ep = NULL;
1411 			wake_up(&qhp->wait);
1412 			break;
1413 		case C4IW_QP_STATE_ERROR:
1414 			goto err;
1415 		default:
1416 			ret = -EINVAL;
1417 			goto err;
1418 		}
1419 		break;
1420 	case C4IW_QP_STATE_ERROR:
1421 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1422 			ret = -EINVAL;
1423 			goto out;
1424 		}
1425 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1426 			ret = -EINVAL;
1427 			goto out;
1428 		}
1429 		set_state(qhp, C4IW_QP_STATE_IDLE);
1430 		break;
1431 	case C4IW_QP_STATE_TERMINATE:
1432 		if (!internal) {
1433 			ret = -EINVAL;
1434 			goto out;
1435 		}
1436 		goto err;
1437 		break;
1438 	default:
1439 		printk(KERN_ERR "%s in a bad state %d\n",
1440 		       __func__, qhp->attr.state);
1441 		ret = -EINVAL;
1442 		goto err;
1443 		break;
1444 	}
1445 	goto out;
1446 err:
1447 	PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1448 	     qhp->wq.sq.qid);
1449 
1450 	/* disassociate the LLP connection */
1451 	qhp->attr.llp_stream_handle = NULL;
1452 	if (!ep)
1453 		ep = qhp->ep;
1454 	qhp->ep = NULL;
1455 	set_state(qhp, C4IW_QP_STATE_ERROR);
1456 	free = 1;
1457 	abort = 1;
1458 	wake_up(&qhp->wait);
1459 	BUG_ON(!ep);
1460 	flush_qp(qhp);
1461 out:
1462 	mutex_unlock(&qhp->mutex);
1463 
1464 	if (terminate)
1465 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1466 
1467 	/*
1468 	 * If disconnect is 1, then we need to initiate a disconnect
1469 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1470 	 * an abnormal close (RTS/CLOSING->ERROR).
1471 	 */
1472 	if (disconnect) {
1473 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1474 							 GFP_KERNEL);
1475 		c4iw_put_ep(&ep->com);
1476 	}
1477 
1478 	/*
1479 	 * If free is 1, then we've disassociated the EP from the QP
1480 	 * and we need to dereference the EP.
1481 	 */
1482 	if (free)
1483 		c4iw_put_ep(&ep->com);
1484 	PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1485 	return ret;
1486 }
1487 
1488 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1489 {
1490 	struct c4iw_dev *rhp;
1491 	struct c4iw_qp *qhp;
1492 	struct c4iw_qp_attributes attrs;
1493 	struct c4iw_ucontext *ucontext;
1494 
1495 	qhp = to_c4iw_qp(ib_qp);
1496 	rhp = qhp->rhp;
1497 
1498 	attrs.next_state = C4IW_QP_STATE_ERROR;
1499 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1500 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1501 	else
1502 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1503 	wait_event(qhp->wait, !qhp->ep);
1504 
1505 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1506 	atomic_dec(&qhp->refcnt);
1507 	wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1508 
1509 	spin_lock_irq(&rhp->lock);
1510 	if (!list_empty(&qhp->db_fc_entry))
1511 		list_del_init(&qhp->db_fc_entry);
1512 	spin_unlock_irq(&rhp->lock);
1513 
1514 	ucontext = ib_qp->uobject ?
1515 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1516 	destroy_qp(&rhp->rdev, &qhp->wq,
1517 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1518 
1519 	PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1520 	kfree(qhp);
1521 	return 0;
1522 }
1523 
1524 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1525 			     struct ib_udata *udata)
1526 {
1527 	struct c4iw_dev *rhp;
1528 	struct c4iw_qp *qhp;
1529 	struct c4iw_pd *php;
1530 	struct c4iw_cq *schp;
1531 	struct c4iw_cq *rchp;
1532 	struct c4iw_create_qp_resp uresp;
1533 	unsigned int sqsize, rqsize;
1534 	struct c4iw_ucontext *ucontext;
1535 	int ret;
1536 	struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1537 
1538 	PDBG("%s ib_pd %p\n", __func__, pd);
1539 
1540 	if (attrs->qp_type != IB_QPT_RC)
1541 		return ERR_PTR(-EINVAL);
1542 
1543 	php = to_c4iw_pd(pd);
1544 	rhp = php->rhp;
1545 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1546 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1547 	if (!schp || !rchp)
1548 		return ERR_PTR(-EINVAL);
1549 
1550 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1551 		return ERR_PTR(-EINVAL);
1552 
1553 	rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1554 	if (rqsize > T4_MAX_RQ_SIZE)
1555 		return ERR_PTR(-E2BIG);
1556 
1557 	sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1558 	if (sqsize > T4_MAX_SQ_SIZE)
1559 		return ERR_PTR(-E2BIG);
1560 
1561 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1562 
1563 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1564 	if (!qhp)
1565 		return ERR_PTR(-ENOMEM);
1566 	qhp->wq.sq.size = sqsize;
1567 	qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1568 	qhp->wq.sq.flush_cidx = -1;
1569 	qhp->wq.rq.size = rqsize;
1570 	qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1571 
1572 	if (ucontext) {
1573 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1574 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1575 	}
1576 
1577 	PDBG("%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu\n",
1578 	     __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1579 
1580 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1581 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1582 	if (ret)
1583 		goto err1;
1584 
1585 	attrs->cap.max_recv_wr = rqsize - 1;
1586 	attrs->cap.max_send_wr = sqsize - 1;
1587 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1588 
1589 	qhp->rhp = rhp;
1590 	qhp->attr.pd = php->pdid;
1591 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1592 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1593 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1594 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1595 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1596 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1597 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1598 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1599 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1600 	qhp->attr.enable_rdma_read = 1;
1601 	qhp->attr.enable_rdma_write = 1;
1602 	qhp->attr.enable_bind = 1;
1603 	qhp->attr.max_ord = 1;
1604 	qhp->attr.max_ird = 1;
1605 	qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1606 	spin_lock_init(&qhp->lock);
1607 	mutex_init(&qhp->mutex);
1608 	init_waitqueue_head(&qhp->wait);
1609 	atomic_set(&qhp->refcnt, 1);
1610 
1611 	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1612 	if (ret)
1613 		goto err2;
1614 
1615 	if (udata) {
1616 		mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1617 		if (!mm1) {
1618 			ret = -ENOMEM;
1619 			goto err3;
1620 		}
1621 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1622 		if (!mm2) {
1623 			ret = -ENOMEM;
1624 			goto err4;
1625 		}
1626 		mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1627 		if (!mm3) {
1628 			ret = -ENOMEM;
1629 			goto err5;
1630 		}
1631 		mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1632 		if (!mm4) {
1633 			ret = -ENOMEM;
1634 			goto err6;
1635 		}
1636 		if (t4_sq_onchip(&qhp->wq.sq)) {
1637 			mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1638 			if (!mm5) {
1639 				ret = -ENOMEM;
1640 				goto err7;
1641 			}
1642 			uresp.flags = C4IW_QPF_ONCHIP;
1643 		} else
1644 			uresp.flags = 0;
1645 		uresp.qid_mask = rhp->rdev.qpmask;
1646 		uresp.sqid = qhp->wq.sq.qid;
1647 		uresp.sq_size = qhp->wq.sq.size;
1648 		uresp.sq_memsize = qhp->wq.sq.memsize;
1649 		uresp.rqid = qhp->wq.rq.qid;
1650 		uresp.rq_size = qhp->wq.rq.size;
1651 		uresp.rq_memsize = qhp->wq.rq.memsize;
1652 		spin_lock(&ucontext->mmap_lock);
1653 		if (mm5) {
1654 			uresp.ma_sync_key = ucontext->key;
1655 			ucontext->key += PAGE_SIZE;
1656 		} else {
1657 			uresp.ma_sync_key =  0;
1658 		}
1659 		uresp.sq_key = ucontext->key;
1660 		ucontext->key += PAGE_SIZE;
1661 		uresp.rq_key = ucontext->key;
1662 		ucontext->key += PAGE_SIZE;
1663 		uresp.sq_db_gts_key = ucontext->key;
1664 		ucontext->key += PAGE_SIZE;
1665 		uresp.rq_db_gts_key = ucontext->key;
1666 		ucontext->key += PAGE_SIZE;
1667 		spin_unlock(&ucontext->mmap_lock);
1668 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1669 		if (ret)
1670 			goto err8;
1671 		mm1->key = uresp.sq_key;
1672 		mm1->addr = qhp->wq.sq.phys_addr;
1673 		mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1674 		insert_mmap(ucontext, mm1);
1675 		mm2->key = uresp.rq_key;
1676 		mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1677 		mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1678 		insert_mmap(ucontext, mm2);
1679 		mm3->key = uresp.sq_db_gts_key;
1680 		mm3->addr = qhp->wq.sq.udb;
1681 		mm3->len = PAGE_SIZE;
1682 		insert_mmap(ucontext, mm3);
1683 		mm4->key = uresp.rq_db_gts_key;
1684 		mm4->addr = qhp->wq.rq.udb;
1685 		mm4->len = PAGE_SIZE;
1686 		insert_mmap(ucontext, mm4);
1687 		if (mm5) {
1688 			mm5->key = uresp.ma_sync_key;
1689 			mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1690 				    + A_PCIE_MA_SYNC) & PAGE_MASK;
1691 			mm5->len = PAGE_SIZE;
1692 			insert_mmap(ucontext, mm5);
1693 		}
1694 	}
1695 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1696 	init_timer(&(qhp->timer));
1697 	INIT_LIST_HEAD(&qhp->db_fc_entry);
1698 	PDBG("%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x\n",
1699 	     __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1700 	     qhp->wq.sq.qid);
1701 	return &qhp->ibqp;
1702 err8:
1703 	kfree(mm5);
1704 err7:
1705 	kfree(mm4);
1706 err6:
1707 	kfree(mm3);
1708 err5:
1709 	kfree(mm2);
1710 err4:
1711 	kfree(mm1);
1712 err3:
1713 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1714 err2:
1715 	destroy_qp(&rhp->rdev, &qhp->wq,
1716 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1717 err1:
1718 	kfree(qhp);
1719 	return ERR_PTR(ret);
1720 }
1721 
1722 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1723 		      int attr_mask, struct ib_udata *udata)
1724 {
1725 	struct c4iw_dev *rhp;
1726 	struct c4iw_qp *qhp;
1727 	enum c4iw_qp_attr_mask mask = 0;
1728 	struct c4iw_qp_attributes attrs;
1729 
1730 	PDBG("%s ib_qp %p\n", __func__, ibqp);
1731 
1732 	/* iwarp does not support the RTR state */
1733 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1734 		attr_mask &= ~IB_QP_STATE;
1735 
1736 	/* Make sure we still have something left to do */
1737 	if (!attr_mask)
1738 		return 0;
1739 
1740 	memset(&attrs, 0, sizeof attrs);
1741 	qhp = to_c4iw_qp(ibqp);
1742 	rhp = qhp->rhp;
1743 
1744 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1745 	attrs.enable_rdma_read = (attr->qp_access_flags &
1746 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1747 	attrs.enable_rdma_write = (attr->qp_access_flags &
1748 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1749 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1750 
1751 
1752 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1753 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1754 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1755 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1756 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1757 
1758 	/*
1759 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1760 	 * ringing the queue db when we're in DB_FULL mode.
1761 	 */
1762 	attrs.sq_db_inc = attr->sq_psn;
1763 	attrs.rq_db_inc = attr->rq_psn;
1764 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1765 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1766 
1767 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1768 }
1769 
1770 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1771 {
1772 	PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1773 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1774 }
1775 
1776 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1777 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1778 {
1779 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1780 
1781 	memset(attr, 0, sizeof *attr);
1782 	memset(init_attr, 0, sizeof *init_attr);
1783 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1784 	return 0;
1785 }
1786