xref: /openbmc/linux/drivers/infiniband/hw/cxgb4/qp.c (revision 80ecbd24)
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 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
642 		   struct ib_send_wr **bad_wr)
643 {
644 	int err = 0;
645 	u8 len16 = 0;
646 	enum fw_wr_opcodes fw_opcode = 0;
647 	enum fw_ri_wr_flags fw_flags;
648 	struct c4iw_qp *qhp;
649 	union t4_wr *wqe;
650 	u32 num_wrs;
651 	struct t4_swsqe *swsqe;
652 	unsigned long flag;
653 	u16 idx = 0;
654 
655 	qhp = to_c4iw_qp(ibqp);
656 	spin_lock_irqsave(&qhp->lock, flag);
657 	if (t4_wq_in_error(&qhp->wq)) {
658 		spin_unlock_irqrestore(&qhp->lock, flag);
659 		return -EINVAL;
660 	}
661 	num_wrs = t4_sq_avail(&qhp->wq);
662 	if (num_wrs == 0) {
663 		spin_unlock_irqrestore(&qhp->lock, flag);
664 		return -ENOMEM;
665 	}
666 	while (wr) {
667 		if (num_wrs == 0) {
668 			err = -ENOMEM;
669 			*bad_wr = wr;
670 			break;
671 		}
672 		wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
673 		      qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
674 
675 		fw_flags = 0;
676 		if (wr->send_flags & IB_SEND_SOLICITED)
677 			fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
678 		if (wr->send_flags & IB_SEND_SIGNALED)
679 			fw_flags |= FW_RI_COMPLETION_FLAG;
680 		swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
681 		switch (wr->opcode) {
682 		case IB_WR_SEND_WITH_INV:
683 		case IB_WR_SEND:
684 			if (wr->send_flags & IB_SEND_FENCE)
685 				fw_flags |= FW_RI_READ_FENCE_FLAG;
686 			fw_opcode = FW_RI_SEND_WR;
687 			if (wr->opcode == IB_WR_SEND)
688 				swsqe->opcode = FW_RI_SEND;
689 			else
690 				swsqe->opcode = FW_RI_SEND_WITH_INV;
691 			err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
692 			break;
693 		case IB_WR_RDMA_WRITE:
694 			fw_opcode = FW_RI_RDMA_WRITE_WR;
695 			swsqe->opcode = FW_RI_RDMA_WRITE;
696 			err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
697 			break;
698 		case IB_WR_RDMA_READ:
699 		case IB_WR_RDMA_READ_WITH_INV:
700 			fw_opcode = FW_RI_RDMA_READ_WR;
701 			swsqe->opcode = FW_RI_READ_REQ;
702 			if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
703 				fw_flags = FW_RI_RDMA_READ_INVALIDATE;
704 			else
705 				fw_flags = 0;
706 			err = build_rdma_read(wqe, wr, &len16);
707 			if (err)
708 				break;
709 			swsqe->read_len = wr->sg_list[0].length;
710 			if (!qhp->wq.sq.oldest_read)
711 				qhp->wq.sq.oldest_read = swsqe;
712 			break;
713 		case IB_WR_FAST_REG_MR:
714 			fw_opcode = FW_RI_FR_NSMR_WR;
715 			swsqe->opcode = FW_RI_FAST_REGISTER;
716 			err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16,
717 					    is_t5(
718 					    qhp->rhp->rdev.lldi.adapter_type) ?
719 					    1 : 0);
720 			break;
721 		case IB_WR_LOCAL_INV:
722 			if (wr->send_flags & IB_SEND_FENCE)
723 				fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
724 			fw_opcode = FW_RI_INV_LSTAG_WR;
725 			swsqe->opcode = FW_RI_LOCAL_INV;
726 			err = build_inv_stag(wqe, wr, &len16);
727 			break;
728 		default:
729 			PDBG("%s post of type=%d TBD!\n", __func__,
730 			     wr->opcode);
731 			err = -EINVAL;
732 		}
733 		if (err) {
734 			*bad_wr = wr;
735 			break;
736 		}
737 		swsqe->idx = qhp->wq.sq.pidx;
738 		swsqe->complete = 0;
739 		swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED);
740 		swsqe->wr_id = wr->wr_id;
741 
742 		init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
743 
744 		PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
745 		     __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
746 		     swsqe->opcode, swsqe->read_len);
747 		wr = wr->next;
748 		num_wrs--;
749 		t4_sq_produce(&qhp->wq, len16);
750 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
751 	}
752 	if (t4_wq_db_enabled(&qhp->wq))
753 		t4_ring_sq_db(&qhp->wq, idx);
754 	spin_unlock_irqrestore(&qhp->lock, flag);
755 	return err;
756 }
757 
758 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
759 		      struct ib_recv_wr **bad_wr)
760 {
761 	int err = 0;
762 	struct c4iw_qp *qhp;
763 	union t4_recv_wr *wqe;
764 	u32 num_wrs;
765 	u8 len16 = 0;
766 	unsigned long flag;
767 	u16 idx = 0;
768 
769 	qhp = to_c4iw_qp(ibqp);
770 	spin_lock_irqsave(&qhp->lock, flag);
771 	if (t4_wq_in_error(&qhp->wq)) {
772 		spin_unlock_irqrestore(&qhp->lock, flag);
773 		return -EINVAL;
774 	}
775 	num_wrs = t4_rq_avail(&qhp->wq);
776 	if (num_wrs == 0) {
777 		spin_unlock_irqrestore(&qhp->lock, flag);
778 		return -ENOMEM;
779 	}
780 	while (wr) {
781 		if (wr->num_sge > T4_MAX_RECV_SGE) {
782 			err = -EINVAL;
783 			*bad_wr = wr;
784 			break;
785 		}
786 		wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
787 					   qhp->wq.rq.wq_pidx *
788 					   T4_EQ_ENTRY_SIZE);
789 		if (num_wrs)
790 			err = build_rdma_recv(qhp, wqe, wr, &len16);
791 		else
792 			err = -ENOMEM;
793 		if (err) {
794 			*bad_wr = wr;
795 			break;
796 		}
797 
798 		qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
799 
800 		wqe->recv.opcode = FW_RI_RECV_WR;
801 		wqe->recv.r1 = 0;
802 		wqe->recv.wrid = qhp->wq.rq.pidx;
803 		wqe->recv.r2[0] = 0;
804 		wqe->recv.r2[1] = 0;
805 		wqe->recv.r2[2] = 0;
806 		wqe->recv.len16 = len16;
807 		PDBG("%s cookie 0x%llx pidx %u\n", __func__,
808 		     (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
809 		t4_rq_produce(&qhp->wq, len16);
810 		idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
811 		wr = wr->next;
812 		num_wrs--;
813 	}
814 	if (t4_wq_db_enabled(&qhp->wq))
815 		t4_ring_rq_db(&qhp->wq, idx);
816 	spin_unlock_irqrestore(&qhp->lock, flag);
817 	return err;
818 }
819 
820 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
821 {
822 	return -ENOSYS;
823 }
824 
825 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
826 				    u8 *ecode)
827 {
828 	int status;
829 	int tagged;
830 	int opcode;
831 	int rqtype;
832 	int send_inv;
833 
834 	if (!err_cqe) {
835 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
836 		*ecode = 0;
837 		return;
838 	}
839 
840 	status = CQE_STATUS(err_cqe);
841 	opcode = CQE_OPCODE(err_cqe);
842 	rqtype = RQ_TYPE(err_cqe);
843 	send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
844 		   (opcode == FW_RI_SEND_WITH_SE_INV);
845 	tagged = (opcode == FW_RI_RDMA_WRITE) ||
846 		 (rqtype && (opcode == FW_RI_READ_RESP));
847 
848 	switch (status) {
849 	case T4_ERR_STAG:
850 		if (send_inv) {
851 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
852 			*ecode = RDMAP_CANT_INV_STAG;
853 		} else {
854 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
855 			*ecode = RDMAP_INV_STAG;
856 		}
857 		break;
858 	case T4_ERR_PDID:
859 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
860 		if ((opcode == FW_RI_SEND_WITH_INV) ||
861 		    (opcode == FW_RI_SEND_WITH_SE_INV))
862 			*ecode = RDMAP_CANT_INV_STAG;
863 		else
864 			*ecode = RDMAP_STAG_NOT_ASSOC;
865 		break;
866 	case T4_ERR_QPID:
867 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
868 		*ecode = RDMAP_STAG_NOT_ASSOC;
869 		break;
870 	case T4_ERR_ACCESS:
871 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
872 		*ecode = RDMAP_ACC_VIOL;
873 		break;
874 	case T4_ERR_WRAP:
875 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
876 		*ecode = RDMAP_TO_WRAP;
877 		break;
878 	case T4_ERR_BOUND:
879 		if (tagged) {
880 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
881 			*ecode = DDPT_BASE_BOUNDS;
882 		} else {
883 			*layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
884 			*ecode = RDMAP_BASE_BOUNDS;
885 		}
886 		break;
887 	case T4_ERR_INVALIDATE_SHARED_MR:
888 	case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
889 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
890 		*ecode = RDMAP_CANT_INV_STAG;
891 		break;
892 	case T4_ERR_ECC:
893 	case T4_ERR_ECC_PSTAG:
894 	case T4_ERR_INTERNAL_ERR:
895 		*layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
896 		*ecode = 0;
897 		break;
898 	case T4_ERR_OUT_OF_RQE:
899 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
900 		*ecode = DDPU_INV_MSN_NOBUF;
901 		break;
902 	case T4_ERR_PBL_ADDR_BOUND:
903 		*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
904 		*ecode = DDPT_BASE_BOUNDS;
905 		break;
906 	case T4_ERR_CRC:
907 		*layer_type = LAYER_MPA|DDP_LLP;
908 		*ecode = MPA_CRC_ERR;
909 		break;
910 	case T4_ERR_MARKER:
911 		*layer_type = LAYER_MPA|DDP_LLP;
912 		*ecode = MPA_MARKER_ERR;
913 		break;
914 	case T4_ERR_PDU_LEN_ERR:
915 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
916 		*ecode = DDPU_MSG_TOOBIG;
917 		break;
918 	case T4_ERR_DDP_VERSION:
919 		if (tagged) {
920 			*layer_type = LAYER_DDP|DDP_TAGGED_ERR;
921 			*ecode = DDPT_INV_VERS;
922 		} else {
923 			*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
924 			*ecode = DDPU_INV_VERS;
925 		}
926 		break;
927 	case T4_ERR_RDMA_VERSION:
928 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
929 		*ecode = RDMAP_INV_VERS;
930 		break;
931 	case T4_ERR_OPCODE:
932 		*layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
933 		*ecode = RDMAP_INV_OPCODE;
934 		break;
935 	case T4_ERR_DDP_QUEUE_NUM:
936 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
937 		*ecode = DDPU_INV_QN;
938 		break;
939 	case T4_ERR_MSN:
940 	case T4_ERR_MSN_GAP:
941 	case T4_ERR_MSN_RANGE:
942 	case T4_ERR_IRD_OVERFLOW:
943 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
944 		*ecode = DDPU_INV_MSN_RANGE;
945 		break;
946 	case T4_ERR_TBIT:
947 		*layer_type = LAYER_DDP|DDP_LOCAL_CATA;
948 		*ecode = 0;
949 		break;
950 	case T4_ERR_MO:
951 		*layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
952 		*ecode = DDPU_INV_MO;
953 		break;
954 	default:
955 		*layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
956 		*ecode = 0;
957 		break;
958 	}
959 }
960 
961 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
962 			   gfp_t gfp)
963 {
964 	struct fw_ri_wr *wqe;
965 	struct sk_buff *skb;
966 	struct terminate_message *term;
967 
968 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
969 	     qhp->ep->hwtid);
970 
971 	skb = alloc_skb(sizeof *wqe, gfp);
972 	if (!skb)
973 		return;
974 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
975 
976 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
977 	memset(wqe, 0, sizeof *wqe);
978 	wqe->op_compl = cpu_to_be32(FW_WR_OP(FW_RI_INIT_WR));
979 	wqe->flowid_len16 = cpu_to_be32(
980 		FW_WR_FLOWID(qhp->ep->hwtid) |
981 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
982 
983 	wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
984 	wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
985 	term = (struct terminate_message *)wqe->u.terminate.termmsg;
986 	if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
987 		term->layer_etype = qhp->attr.layer_etype;
988 		term->ecode = qhp->attr.ecode;
989 	} else
990 		build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
991 	c4iw_ofld_send(&qhp->rhp->rdev, skb);
992 }
993 
994 /*
995  * Assumes qhp lock is held.
996  */
997 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
998 		       struct c4iw_cq *schp)
999 {
1000 	int count;
1001 	int flushed;
1002 	unsigned long flag;
1003 
1004 	PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1005 
1006 	/* locking hierarchy: cq lock first, then qp lock. */
1007 	spin_lock_irqsave(&rchp->lock, flag);
1008 	spin_lock(&qhp->lock);
1009 	c4iw_flush_hw_cq(&rchp->cq);
1010 	c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1011 	flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1012 	spin_unlock(&qhp->lock);
1013 	spin_unlock_irqrestore(&rchp->lock, flag);
1014 	if (flushed) {
1015 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1016 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1017 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1018 	}
1019 
1020 	/* locking hierarchy: cq lock first, then qp lock. */
1021 	spin_lock_irqsave(&schp->lock, flag);
1022 	spin_lock(&qhp->lock);
1023 	c4iw_flush_hw_cq(&schp->cq);
1024 	c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
1025 	flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
1026 	spin_unlock(&qhp->lock);
1027 	spin_unlock_irqrestore(&schp->lock, flag);
1028 	if (flushed) {
1029 		spin_lock_irqsave(&schp->comp_handler_lock, flag);
1030 		(*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
1031 		spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1032 	}
1033 }
1034 
1035 static void flush_qp(struct c4iw_qp *qhp)
1036 {
1037 	struct c4iw_cq *rchp, *schp;
1038 	unsigned long flag;
1039 
1040 	rchp = get_chp(qhp->rhp, qhp->attr.rcq);
1041 	schp = get_chp(qhp->rhp, qhp->attr.scq);
1042 
1043 	if (qhp->ibqp.uobject) {
1044 		t4_set_wq_in_error(&qhp->wq);
1045 		t4_set_cq_in_error(&rchp->cq);
1046 		spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1047 		(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1048 		spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1049 		if (schp != rchp) {
1050 			t4_set_cq_in_error(&schp->cq);
1051 			spin_lock_irqsave(&schp->comp_handler_lock, flag);
1052 			(*schp->ibcq.comp_handler)(&schp->ibcq,
1053 					schp->ibcq.cq_context);
1054 			spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1055 		}
1056 		return;
1057 	}
1058 	__flush_qp(qhp, rchp, schp);
1059 }
1060 
1061 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1062 		     struct c4iw_ep *ep)
1063 {
1064 	struct fw_ri_wr *wqe;
1065 	int ret;
1066 	struct sk_buff *skb;
1067 
1068 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1069 	     ep->hwtid);
1070 
1071 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1072 	if (!skb)
1073 		return -ENOMEM;
1074 	set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1075 
1076 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1077 	memset(wqe, 0, sizeof *wqe);
1078 	wqe->op_compl = cpu_to_be32(
1079 		FW_WR_OP(FW_RI_INIT_WR) |
1080 		FW_WR_COMPL(1));
1081 	wqe->flowid_len16 = cpu_to_be32(
1082 		FW_WR_FLOWID(ep->hwtid) |
1083 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1084 	wqe->cookie = (unsigned long) &ep->com.wr_wait;
1085 
1086 	wqe->u.fini.type = FW_RI_TYPE_FINI;
1087 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1088 	if (ret)
1089 		goto out;
1090 
1091 	ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1092 			     qhp->wq.sq.qid, __func__);
1093 out:
1094 	PDBG("%s ret %d\n", __func__, ret);
1095 	return ret;
1096 }
1097 
1098 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1099 {
1100 	PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1101 	memset(&init->u, 0, sizeof init->u);
1102 	switch (p2p_type) {
1103 	case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1104 		init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1105 		init->u.write.stag_sink = cpu_to_be32(1);
1106 		init->u.write.to_sink = cpu_to_be64(1);
1107 		init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1108 		init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1109 						   sizeof(struct fw_ri_immd),
1110 						   16);
1111 		break;
1112 	case FW_RI_INIT_P2PTYPE_READ_REQ:
1113 		init->u.write.opcode = FW_RI_RDMA_READ_WR;
1114 		init->u.read.stag_src = cpu_to_be32(1);
1115 		init->u.read.to_src_lo = cpu_to_be32(1);
1116 		init->u.read.stag_sink = cpu_to_be32(1);
1117 		init->u.read.to_sink_lo = cpu_to_be32(1);
1118 		init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1119 		break;
1120 	}
1121 }
1122 
1123 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1124 {
1125 	struct fw_ri_wr *wqe;
1126 	int ret;
1127 	struct sk_buff *skb;
1128 
1129 	PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1130 	     qhp->ep->hwtid);
1131 
1132 	skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1133 	if (!skb)
1134 		return -ENOMEM;
1135 	set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1136 
1137 	wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1138 	memset(wqe, 0, sizeof *wqe);
1139 	wqe->op_compl = cpu_to_be32(
1140 		FW_WR_OP(FW_RI_INIT_WR) |
1141 		FW_WR_COMPL(1));
1142 	wqe->flowid_len16 = cpu_to_be32(
1143 		FW_WR_FLOWID(qhp->ep->hwtid) |
1144 		FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1145 
1146 	wqe->cookie = (unsigned long) &qhp->ep->com.wr_wait;
1147 
1148 	wqe->u.init.type = FW_RI_TYPE_INIT;
1149 	wqe->u.init.mpareqbit_p2ptype =
1150 		V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1151 		V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1152 	wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1153 	if (qhp->attr.mpa_attr.recv_marker_enabled)
1154 		wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1155 	if (qhp->attr.mpa_attr.xmit_marker_enabled)
1156 		wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1157 	if (qhp->attr.mpa_attr.crc_enabled)
1158 		wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1159 
1160 	wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1161 			    FW_RI_QP_RDMA_WRITE_ENABLE |
1162 			    FW_RI_QP_BIND_ENABLE;
1163 	if (!qhp->ibqp.uobject)
1164 		wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1165 				     FW_RI_QP_STAG0_ENABLE;
1166 	wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1167 	wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1168 	wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1169 	wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1170 	wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1171 	wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1172 	wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1173 	wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1174 	wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1175 	wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1176 	wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1177 	wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1178 	wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1179 					 rhp->rdev.lldi.vr->rq.start);
1180 	if (qhp->attr.mpa_attr.initiator)
1181 		build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1182 
1183 	ret = c4iw_ofld_send(&rhp->rdev, skb);
1184 	if (ret)
1185 		goto out;
1186 
1187 	ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1188 				  qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1189 out:
1190 	PDBG("%s ret %d\n", __func__, ret);
1191 	return ret;
1192 }
1193 
1194 /*
1195  * Called by the library when the qp has user dbs disabled due to
1196  * a DB_FULL condition.  This function will single-thread all user
1197  * DB rings to avoid overflowing the hw db-fifo.
1198  */
1199 static int ring_kernel_db(struct c4iw_qp *qhp, u32 qid, u16 inc)
1200 {
1201 	int delay = db_delay_usecs;
1202 
1203 	mutex_lock(&qhp->rhp->db_mutex);
1204 	do {
1205 
1206 		/*
1207 		 * The interrupt threshold is dbfifo_int_thresh << 6. So
1208 		 * make sure we don't cross that and generate an interrupt.
1209 		 */
1210 		if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) <
1211 		    (qhp->rhp->rdev.lldi.dbfifo_int_thresh << 5)) {
1212 			writel(QID(qid) | PIDX(inc), qhp->wq.db);
1213 			break;
1214 		}
1215 		set_current_state(TASK_UNINTERRUPTIBLE);
1216 		schedule_timeout(usecs_to_jiffies(delay));
1217 		delay = min(delay << 1, 2000);
1218 	} while (1);
1219 	mutex_unlock(&qhp->rhp->db_mutex);
1220 	return 0;
1221 }
1222 
1223 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1224 		   enum c4iw_qp_attr_mask mask,
1225 		   struct c4iw_qp_attributes *attrs,
1226 		   int internal)
1227 {
1228 	int ret = 0;
1229 	struct c4iw_qp_attributes newattr = qhp->attr;
1230 	int disconnect = 0;
1231 	int terminate = 0;
1232 	int abort = 0;
1233 	int free = 0;
1234 	struct c4iw_ep *ep = NULL;
1235 
1236 	PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1237 	     qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1238 	     (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1239 
1240 	mutex_lock(&qhp->mutex);
1241 
1242 	/* Process attr changes if in IDLE */
1243 	if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1244 		if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1245 			ret = -EIO;
1246 			goto out;
1247 		}
1248 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1249 			newattr.enable_rdma_read = attrs->enable_rdma_read;
1250 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1251 			newattr.enable_rdma_write = attrs->enable_rdma_write;
1252 		if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1253 			newattr.enable_bind = attrs->enable_bind;
1254 		if (mask & C4IW_QP_ATTR_MAX_ORD) {
1255 			if (attrs->max_ord > c4iw_max_read_depth) {
1256 				ret = -EINVAL;
1257 				goto out;
1258 			}
1259 			newattr.max_ord = attrs->max_ord;
1260 		}
1261 		if (mask & C4IW_QP_ATTR_MAX_IRD) {
1262 			if (attrs->max_ird > c4iw_max_read_depth) {
1263 				ret = -EINVAL;
1264 				goto out;
1265 			}
1266 			newattr.max_ird = attrs->max_ird;
1267 		}
1268 		qhp->attr = newattr;
1269 	}
1270 
1271 	if (mask & C4IW_QP_ATTR_SQ_DB) {
1272 		ret = ring_kernel_db(qhp, qhp->wq.sq.qid, attrs->sq_db_inc);
1273 		goto out;
1274 	}
1275 	if (mask & C4IW_QP_ATTR_RQ_DB) {
1276 		ret = ring_kernel_db(qhp, qhp->wq.rq.qid, attrs->rq_db_inc);
1277 		goto out;
1278 	}
1279 
1280 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1281 		goto out;
1282 	if (qhp->attr.state == attrs->next_state)
1283 		goto out;
1284 
1285 	switch (qhp->attr.state) {
1286 	case C4IW_QP_STATE_IDLE:
1287 		switch (attrs->next_state) {
1288 		case C4IW_QP_STATE_RTS:
1289 			if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1290 				ret = -EINVAL;
1291 				goto out;
1292 			}
1293 			if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1294 				ret = -EINVAL;
1295 				goto out;
1296 			}
1297 			qhp->attr.mpa_attr = attrs->mpa_attr;
1298 			qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1299 			qhp->ep = qhp->attr.llp_stream_handle;
1300 			set_state(qhp, C4IW_QP_STATE_RTS);
1301 
1302 			/*
1303 			 * Ref the endpoint here and deref when we
1304 			 * disassociate the endpoint from the QP.  This
1305 			 * happens in CLOSING->IDLE transition or *->ERROR
1306 			 * transition.
1307 			 */
1308 			c4iw_get_ep(&qhp->ep->com);
1309 			ret = rdma_init(rhp, qhp);
1310 			if (ret)
1311 				goto err;
1312 			break;
1313 		case C4IW_QP_STATE_ERROR:
1314 			set_state(qhp, C4IW_QP_STATE_ERROR);
1315 			flush_qp(qhp);
1316 			break;
1317 		default:
1318 			ret = -EINVAL;
1319 			goto out;
1320 		}
1321 		break;
1322 	case C4IW_QP_STATE_RTS:
1323 		switch (attrs->next_state) {
1324 		case C4IW_QP_STATE_CLOSING:
1325 			BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1326 			set_state(qhp, C4IW_QP_STATE_CLOSING);
1327 			ep = qhp->ep;
1328 			if (!internal) {
1329 				abort = 0;
1330 				disconnect = 1;
1331 				c4iw_get_ep(&qhp->ep->com);
1332 			}
1333 			if (qhp->ibqp.uobject)
1334 				t4_set_wq_in_error(&qhp->wq);
1335 			ret = rdma_fini(rhp, qhp, ep);
1336 			if (ret)
1337 				goto err;
1338 			break;
1339 		case C4IW_QP_STATE_TERMINATE:
1340 			set_state(qhp, C4IW_QP_STATE_TERMINATE);
1341 			qhp->attr.layer_etype = attrs->layer_etype;
1342 			qhp->attr.ecode = attrs->ecode;
1343 			if (qhp->ibqp.uobject)
1344 				t4_set_wq_in_error(&qhp->wq);
1345 			ep = qhp->ep;
1346 			if (!internal)
1347 				terminate = 1;
1348 			disconnect = 1;
1349 			c4iw_get_ep(&qhp->ep->com);
1350 			break;
1351 		case C4IW_QP_STATE_ERROR:
1352 			set_state(qhp, C4IW_QP_STATE_ERROR);
1353 			if (qhp->ibqp.uobject)
1354 				t4_set_wq_in_error(&qhp->wq);
1355 			if (!internal) {
1356 				abort = 1;
1357 				disconnect = 1;
1358 				ep = qhp->ep;
1359 				c4iw_get_ep(&qhp->ep->com);
1360 			}
1361 			goto err;
1362 			break;
1363 		default:
1364 			ret = -EINVAL;
1365 			goto out;
1366 		}
1367 		break;
1368 	case C4IW_QP_STATE_CLOSING:
1369 		if (!internal) {
1370 			ret = -EINVAL;
1371 			goto out;
1372 		}
1373 		switch (attrs->next_state) {
1374 		case C4IW_QP_STATE_IDLE:
1375 			flush_qp(qhp);
1376 			set_state(qhp, C4IW_QP_STATE_IDLE);
1377 			qhp->attr.llp_stream_handle = NULL;
1378 			c4iw_put_ep(&qhp->ep->com);
1379 			qhp->ep = NULL;
1380 			wake_up(&qhp->wait);
1381 			break;
1382 		case C4IW_QP_STATE_ERROR:
1383 			goto err;
1384 		default:
1385 			ret = -EINVAL;
1386 			goto err;
1387 		}
1388 		break;
1389 	case C4IW_QP_STATE_ERROR:
1390 		if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1391 			ret = -EINVAL;
1392 			goto out;
1393 		}
1394 		if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1395 			ret = -EINVAL;
1396 			goto out;
1397 		}
1398 		set_state(qhp, C4IW_QP_STATE_IDLE);
1399 		break;
1400 	case C4IW_QP_STATE_TERMINATE:
1401 		if (!internal) {
1402 			ret = -EINVAL;
1403 			goto out;
1404 		}
1405 		goto err;
1406 		break;
1407 	default:
1408 		printk(KERN_ERR "%s in a bad state %d\n",
1409 		       __func__, qhp->attr.state);
1410 		ret = -EINVAL;
1411 		goto err;
1412 		break;
1413 	}
1414 	goto out;
1415 err:
1416 	PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1417 	     qhp->wq.sq.qid);
1418 
1419 	/* disassociate the LLP connection */
1420 	qhp->attr.llp_stream_handle = NULL;
1421 	if (!ep)
1422 		ep = qhp->ep;
1423 	qhp->ep = NULL;
1424 	set_state(qhp, C4IW_QP_STATE_ERROR);
1425 	free = 1;
1426 	abort = 1;
1427 	wake_up(&qhp->wait);
1428 	BUG_ON(!ep);
1429 	flush_qp(qhp);
1430 out:
1431 	mutex_unlock(&qhp->mutex);
1432 
1433 	if (terminate)
1434 		post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1435 
1436 	/*
1437 	 * If disconnect is 1, then we need to initiate a disconnect
1438 	 * on the EP.  This can be a normal close (RTS->CLOSING) or
1439 	 * an abnormal close (RTS/CLOSING->ERROR).
1440 	 */
1441 	if (disconnect) {
1442 		c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1443 							 GFP_KERNEL);
1444 		c4iw_put_ep(&ep->com);
1445 	}
1446 
1447 	/*
1448 	 * If free is 1, then we've disassociated the EP from the QP
1449 	 * and we need to dereference the EP.
1450 	 */
1451 	if (free)
1452 		c4iw_put_ep(&ep->com);
1453 	PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1454 	return ret;
1455 }
1456 
1457 static int enable_qp_db(int id, void *p, void *data)
1458 {
1459 	struct c4iw_qp *qp = p;
1460 
1461 	t4_enable_wq_db(&qp->wq);
1462 	return 0;
1463 }
1464 
1465 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1466 {
1467 	struct c4iw_dev *rhp;
1468 	struct c4iw_qp *qhp;
1469 	struct c4iw_qp_attributes attrs;
1470 	struct c4iw_ucontext *ucontext;
1471 
1472 	qhp = to_c4iw_qp(ib_qp);
1473 	rhp = qhp->rhp;
1474 
1475 	attrs.next_state = C4IW_QP_STATE_ERROR;
1476 	if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1477 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1478 	else
1479 		c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1480 	wait_event(qhp->wait, !qhp->ep);
1481 
1482 	spin_lock_irq(&rhp->lock);
1483 	remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1484 	rhp->qpcnt--;
1485 	BUG_ON(rhp->qpcnt < 0);
1486 	if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
1487 		rhp->rdev.stats.db_state_transitions++;
1488 		rhp->db_state = NORMAL;
1489 		idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
1490 	}
1491 	if (db_coalescing_threshold >= 0)
1492 		if (rhp->qpcnt <= db_coalescing_threshold)
1493 			cxgb4_enable_db_coalescing(rhp->rdev.lldi.ports[0]);
1494 	spin_unlock_irq(&rhp->lock);
1495 	atomic_dec(&qhp->refcnt);
1496 	wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1497 
1498 	ucontext = ib_qp->uobject ?
1499 		   to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1500 	destroy_qp(&rhp->rdev, &qhp->wq,
1501 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1502 
1503 	PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1504 	kfree(qhp);
1505 	return 0;
1506 }
1507 
1508 static int disable_qp_db(int id, void *p, void *data)
1509 {
1510 	struct c4iw_qp *qp = p;
1511 
1512 	t4_disable_wq_db(&qp->wq);
1513 	return 0;
1514 }
1515 
1516 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1517 			     struct ib_udata *udata)
1518 {
1519 	struct c4iw_dev *rhp;
1520 	struct c4iw_qp *qhp;
1521 	struct c4iw_pd *php;
1522 	struct c4iw_cq *schp;
1523 	struct c4iw_cq *rchp;
1524 	struct c4iw_create_qp_resp uresp;
1525 	int sqsize, rqsize;
1526 	struct c4iw_ucontext *ucontext;
1527 	int ret;
1528 	struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1529 
1530 	PDBG("%s ib_pd %p\n", __func__, pd);
1531 
1532 	if (attrs->qp_type != IB_QPT_RC)
1533 		return ERR_PTR(-EINVAL);
1534 
1535 	php = to_c4iw_pd(pd);
1536 	rhp = php->rhp;
1537 	schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1538 	rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1539 	if (!schp || !rchp)
1540 		return ERR_PTR(-EINVAL);
1541 
1542 	if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1543 		return ERR_PTR(-EINVAL);
1544 
1545 	rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1546 	if (rqsize > T4_MAX_RQ_SIZE)
1547 		return ERR_PTR(-E2BIG);
1548 
1549 	sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1550 	if (sqsize > T4_MAX_SQ_SIZE)
1551 		return ERR_PTR(-E2BIG);
1552 
1553 	ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1554 
1555 
1556 	qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1557 	if (!qhp)
1558 		return ERR_PTR(-ENOMEM);
1559 	qhp->wq.sq.size = sqsize;
1560 	qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1561 	qhp->wq.rq.size = rqsize;
1562 	qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1563 
1564 	if (ucontext) {
1565 		qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1566 		qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1567 	}
1568 
1569 	PDBG("%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu\n",
1570 	     __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1571 
1572 	ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1573 			ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1574 	if (ret)
1575 		goto err1;
1576 
1577 	attrs->cap.max_recv_wr = rqsize - 1;
1578 	attrs->cap.max_send_wr = sqsize - 1;
1579 	attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1580 
1581 	qhp->rhp = rhp;
1582 	qhp->attr.pd = php->pdid;
1583 	qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1584 	qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1585 	qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1586 	qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1587 	qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1588 	qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1589 	qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1590 	qhp->attr.state = C4IW_QP_STATE_IDLE;
1591 	qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1592 	qhp->attr.enable_rdma_read = 1;
1593 	qhp->attr.enable_rdma_write = 1;
1594 	qhp->attr.enable_bind = 1;
1595 	qhp->attr.max_ord = 1;
1596 	qhp->attr.max_ird = 1;
1597 	spin_lock_init(&qhp->lock);
1598 	mutex_init(&qhp->mutex);
1599 	init_waitqueue_head(&qhp->wait);
1600 	atomic_set(&qhp->refcnt, 1);
1601 
1602 	spin_lock_irq(&rhp->lock);
1603 	if (rhp->db_state != NORMAL)
1604 		t4_disable_wq_db(&qhp->wq);
1605 	rhp->qpcnt++;
1606 	if (rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
1607 		rhp->rdev.stats.db_state_transitions++;
1608 		rhp->db_state = FLOW_CONTROL;
1609 		idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
1610 	}
1611 	if (db_coalescing_threshold >= 0)
1612 		if (rhp->qpcnt > db_coalescing_threshold)
1613 			cxgb4_disable_db_coalescing(rhp->rdev.lldi.ports[0]);
1614 	ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1615 	spin_unlock_irq(&rhp->lock);
1616 	if (ret)
1617 		goto err2;
1618 
1619 	if (udata) {
1620 		mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1621 		if (!mm1) {
1622 			ret = -ENOMEM;
1623 			goto err3;
1624 		}
1625 		mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1626 		if (!mm2) {
1627 			ret = -ENOMEM;
1628 			goto err4;
1629 		}
1630 		mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1631 		if (!mm3) {
1632 			ret = -ENOMEM;
1633 			goto err5;
1634 		}
1635 		mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1636 		if (!mm4) {
1637 			ret = -ENOMEM;
1638 			goto err6;
1639 		}
1640 		if (t4_sq_onchip(&qhp->wq.sq)) {
1641 			mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1642 			if (!mm5) {
1643 				ret = -ENOMEM;
1644 				goto err7;
1645 			}
1646 			uresp.flags = C4IW_QPF_ONCHIP;
1647 		} else
1648 			uresp.flags = 0;
1649 		uresp.qid_mask = rhp->rdev.qpmask;
1650 		uresp.sqid = qhp->wq.sq.qid;
1651 		uresp.sq_size = qhp->wq.sq.size;
1652 		uresp.sq_memsize = qhp->wq.sq.memsize;
1653 		uresp.rqid = qhp->wq.rq.qid;
1654 		uresp.rq_size = qhp->wq.rq.size;
1655 		uresp.rq_memsize = qhp->wq.rq.memsize;
1656 		spin_lock(&ucontext->mmap_lock);
1657 		if (mm5) {
1658 			uresp.ma_sync_key = ucontext->key;
1659 			ucontext->key += PAGE_SIZE;
1660 		} else {
1661 			uresp.ma_sync_key =  0;
1662 		}
1663 		uresp.sq_key = ucontext->key;
1664 		ucontext->key += PAGE_SIZE;
1665 		uresp.rq_key = ucontext->key;
1666 		ucontext->key += PAGE_SIZE;
1667 		uresp.sq_db_gts_key = ucontext->key;
1668 		ucontext->key += PAGE_SIZE;
1669 		uresp.rq_db_gts_key = ucontext->key;
1670 		ucontext->key += PAGE_SIZE;
1671 		spin_unlock(&ucontext->mmap_lock);
1672 		ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1673 		if (ret)
1674 			goto err8;
1675 		mm1->key = uresp.sq_key;
1676 		mm1->addr = qhp->wq.sq.phys_addr;
1677 		mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1678 		insert_mmap(ucontext, mm1);
1679 		mm2->key = uresp.rq_key;
1680 		mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1681 		mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1682 		insert_mmap(ucontext, mm2);
1683 		mm3->key = uresp.sq_db_gts_key;
1684 		mm3->addr = qhp->wq.sq.udb;
1685 		mm3->len = PAGE_SIZE;
1686 		insert_mmap(ucontext, mm3);
1687 		mm4->key = uresp.rq_db_gts_key;
1688 		mm4->addr = qhp->wq.rq.udb;
1689 		mm4->len = PAGE_SIZE;
1690 		insert_mmap(ucontext, mm4);
1691 		if (mm5) {
1692 			mm5->key = uresp.ma_sync_key;
1693 			mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1694 				    + A_PCIE_MA_SYNC) & PAGE_MASK;
1695 			mm5->len = PAGE_SIZE;
1696 			insert_mmap(ucontext, mm5);
1697 		}
1698 	}
1699 	qhp->ibqp.qp_num = qhp->wq.sq.qid;
1700 	init_timer(&(qhp->timer));
1701 	PDBG("%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x\n",
1702 	     __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1703 	     qhp->wq.sq.qid);
1704 	return &qhp->ibqp;
1705 err8:
1706 	kfree(mm5);
1707 err7:
1708 	kfree(mm4);
1709 err6:
1710 	kfree(mm3);
1711 err5:
1712 	kfree(mm2);
1713 err4:
1714 	kfree(mm1);
1715 err3:
1716 	remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1717 err2:
1718 	destroy_qp(&rhp->rdev, &qhp->wq,
1719 		   ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1720 err1:
1721 	kfree(qhp);
1722 	return ERR_PTR(ret);
1723 }
1724 
1725 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1726 		      int attr_mask, struct ib_udata *udata)
1727 {
1728 	struct c4iw_dev *rhp;
1729 	struct c4iw_qp *qhp;
1730 	enum c4iw_qp_attr_mask mask = 0;
1731 	struct c4iw_qp_attributes attrs;
1732 
1733 	PDBG("%s ib_qp %p\n", __func__, ibqp);
1734 
1735 	/* iwarp does not support the RTR state */
1736 	if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1737 		attr_mask &= ~IB_QP_STATE;
1738 
1739 	/* Make sure we still have something left to do */
1740 	if (!attr_mask)
1741 		return 0;
1742 
1743 	memset(&attrs, 0, sizeof attrs);
1744 	qhp = to_c4iw_qp(ibqp);
1745 	rhp = qhp->rhp;
1746 
1747 	attrs.next_state = c4iw_convert_state(attr->qp_state);
1748 	attrs.enable_rdma_read = (attr->qp_access_flags &
1749 			       IB_ACCESS_REMOTE_READ) ?  1 : 0;
1750 	attrs.enable_rdma_write = (attr->qp_access_flags &
1751 				IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1752 	attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1753 
1754 
1755 	mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1756 	mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1757 			(C4IW_QP_ATTR_ENABLE_RDMA_READ |
1758 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1759 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1760 
1761 	/*
1762 	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1763 	 * ringing the queue db when we're in DB_FULL mode.
1764 	 */
1765 	attrs.sq_db_inc = attr->sq_psn;
1766 	attrs.rq_db_inc = attr->rq_psn;
1767 	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1768 	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1769 
1770 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1771 }
1772 
1773 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1774 {
1775 	PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1776 	return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1777 }
1778 
1779 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1780 		     int attr_mask, struct ib_qp_init_attr *init_attr)
1781 {
1782 	struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1783 
1784 	memset(attr, 0, sizeof *attr);
1785 	memset(init_attr, 0, sizeof *init_attr);
1786 	attr->qp_state = to_ib_qp_state(qhp->attr.state);
1787 	return 0;
1788 }
1789