xref: /freebsd/sys/dev/qlnx/qlnxr/qlnxr_verbs.c (revision d0b2dbfa0ecf2bbc9709efc5e20baf8e4b44bbbf)
1 /*
2  * Copyright (c) 2018-2019 Cavium, Inc.
3  * All rights reserved.
4  *
5  *  Redistribution and use in source and binary forms, with or without
6  *  modification, are permitted provided that the following conditions
7  *  are met:
8  *
9  *  1. Redistributions of source code must retain the above copyright
10  *     notice, this list of conditions and the following disclaimer.
11  *  2. Redistributions in binary form must reproduce the above copyright
12  *     notice, this list of conditions and the following disclaimer in the
13  *     documentation and/or other materials provided with the distribution.
14  *
15  *  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  *  AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  *  IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  *  ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  *  LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  *  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  *  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  *  INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  *  CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  *  ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  *  POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 /*
29  * File: qlnxr_verbs.c
30  */
31 #include <sys/cdefs.h>
32 #include "qlnxr_def.h"
33 #include "rdma_common.h"
34 #include "qlnxr_roce.h"
35 #include "qlnxr_cm.h"
36 
37 #define HILO_U64(hi, lo)		((((u64)(hi)) << 32) + (lo))
38 
39 #define TYPEPTR_ADDR_SET(type_ptr, field, vaddr)			\
40 	do {								\
41 		(type_ptr)->field.hi = cpu_to_le32(upper_32_bits(vaddr));\
42 		(type_ptr)->field.lo = cpu_to_le32(lower_32_bits(vaddr));\
43 	} while (0)
44 
45 #define RQ_SGE_SET(sge, vaddr, vlength, vflags)			\
46 	do {							\
47 		TYPEPTR_ADDR_SET(sge, addr, vaddr);		\
48 		(sge)->length = cpu_to_le32(vlength);		\
49 		(sge)->flags = cpu_to_le32(vflags);		\
50 	} while (0)
51 
52 #define SRQ_HDR_SET(hdr, vwr_id, num_sge)			\
53 	do {							\
54 		TYPEPTR_ADDR_SET(hdr, wr_id, vwr_id);		\
55 		(hdr)->num_sges = num_sge;			\
56 	} while (0)
57 
58 #define SRQ_SGE_SET(sge, vaddr, vlength, vlkey)			\
59 	do {							\
60 		TYPEPTR_ADDR_SET(sge, addr, vaddr);		\
61 		(sge)->length = cpu_to_le32(vlength);		\
62 		(sge)->l_key = cpu_to_le32(vlkey);		\
63 	} while (0)
64 
65 #define NIPQUAD(addr) \
66 	((unsigned char *)&addr)[0], \
67 	((unsigned char *)&addr)[1], \
68 	((unsigned char *)&addr)[2], \
69 	((unsigned char *)&addr)[3]
70 
71 static int
72 qlnxr_check_srq_params(struct qlnxr_dev *dev,
73 	struct ib_srq_init_attr *attrs);
74 
75 static int
76 qlnxr_init_srq_user_params(struct ib_ucontext *ib_ctx,
77 	struct qlnxr_srq *srq,
78 	struct qlnxr_create_srq_ureq *ureq,
79 	int access, int dmasync);
80 
81 static int
82 qlnxr_alloc_srq_kernel_params(struct qlnxr_srq *srq,
83 	struct qlnxr_dev *dev,
84 	struct ib_srq_init_attr *init_attr);
85 
86 static int
87 qlnxr_copy_srq_uresp(struct qlnxr_dev *dev,
88 	struct qlnxr_srq *srq,
89 	struct ib_udata *udata);
90 
91 static void
92 qlnxr_free_srq_user_params(struct qlnxr_srq *srq);
93 
94 static void
95 qlnxr_free_srq_kernel_params(struct qlnxr_srq *srq);
96 
97 static u32
98 qlnxr_srq_elem_left(struct qlnxr_srq_hwq_info *hw_srq);
99 
100 int
101 qlnxr_iw_query_gid(struct ib_device *ibdev, u8 port, int index,
102 	union ib_gid *sgid)
103 {
104 	struct qlnxr_dev	*dev;
105 	qlnx_host_t		*ha;
106 
107 	dev = get_qlnxr_dev(ibdev);
108 	ha = dev->ha;
109 
110 	QL_DPRINT12(ha, "enter\n");
111 
112 	memset(sgid->raw, 0, sizeof(sgid->raw));
113 
114 	memcpy(sgid->raw, dev->ha->primary_mac, sizeof (dev->ha->primary_mac));
115 
116 	QL_DPRINT12(ha, "exit\n");
117 
118 	return 0;
119 }
120 
121 int
122 qlnxr_query_gid(struct ib_device *ibdev, u8 port, int index,
123 	union ib_gid *sgid)
124 {
125 	struct qlnxr_dev	*dev;
126 	qlnx_host_t		*ha;
127 
128 	dev = get_qlnxr_dev(ibdev);
129 	ha = dev->ha;
130 	QL_DPRINT12(ha, "enter index: %d\n", index);
131 #if 0
132 	int ret = 0;
133 	/* @@@: if DEFINE_ROCE_GID_TABLE to be used here */
134 	//if (!rdma_cap_roce_gid_table(ibdev, port)) {
135 	if (!(rdma_protocol_roce(ibdev, port) &&
136 		ibdev->add_gid && ibdev->del_gid)) {
137 		QL_DPRINT11(ha, "acquire gid failed\n");
138 		return -ENODEV;
139 	}
140 
141 	ret = ib_get_cached_gid(ibdev, port, index, sgid, NULL);
142 	if (ret == -EAGAIN) {
143 		memcpy(sgid, &zgid, sizeof(*sgid));
144 		return 0;
145 	}
146 #endif
147 	if ((index >= QLNXR_MAX_SGID) || (index < 0)) {
148 		QL_DPRINT12(ha, "invalid gid index %d\n", index);
149 		memset(sgid, 0, sizeof(*sgid));
150 		return -EINVAL;
151 	}
152 	memcpy(sgid, &dev->sgid_tbl[index], sizeof(*sgid));
153 
154 	QL_DPRINT12(ha, "exit : %p\n", sgid);
155 
156 	return 0;
157 }
158 
159 int
160 qlnxr_create_srq(struct ib_srq *ibsrq,
161 		 struct ib_srq_init_attr *init_attr,
162 		 struct ib_udata *udata)
163 {
164 	struct qlnxr_dev	*dev;
165 	qlnx_host_t		*ha;
166 	struct ecore_rdma_destroy_srq_in_params destroy_in_params;
167 	struct ecore_rdma_create_srq_out_params out_params;
168 	struct ecore_rdma_create_srq_in_params in_params;
169 	u64 pbl_base_addr, phy_prod_pair_addr;
170 	struct qlnxr_srq_hwq_info *hw_srq;
171 	struct qlnxr_ucontext *ctx;
172 	struct qlnxr_create_srq_ureq ureq;
173 	u32 page_cnt, page_size;
174 	struct qlnxr_srq *srq = get_qlnxr_srq(ibsrq);
175 	int ret = 0;
176 
177 	dev = get_qlnxr_dev(ibsrq->device);
178 	ha = dev->ha;
179 
180 	QL_DPRINT12(ha, "enter\n");
181 
182 	ret = qlnxr_check_srq_params(dev, init_attr);
183 
184 	srq->dev = dev;
185 	hw_srq = &srq->hw_srq;
186 	spin_lock_init(&srq->lock);
187 	memset(&in_params, 0, sizeof(in_params));
188 
189 	if (udata) {
190 		ctx = rdma_udata_to_drv_context(
191 		    udata, struct qlnxr_ucontext, ibucontext);
192 
193 		memset(&ureq, 0, sizeof(ureq));
194 		if (ib_copy_from_udata(&ureq, udata, min(sizeof(ureq),
195 			udata->inlen))) {
196 			QL_DPRINT11(ha, "problem"
197 				" copying data from user space\n");
198 			goto err0;
199 		}
200 
201 		ret = qlnxr_init_srq_user_params(&ctx->ibucontext, srq, &ureq, 0, 0);
202 		if (ret)
203 			goto err0;
204 
205 		page_cnt = srq->usrq.pbl_info.num_pbes;
206 		pbl_base_addr = srq->usrq.pbl_tbl->pa;
207 		phy_prod_pair_addr = hw_srq->phy_prod_pair_addr;
208 		// @@@ : if DEFINE_IB_UMEM_PAGE_SHIFT
209 		// page_size = BIT(srq->usrq.umem->page_shift);
210 		// else
211 		page_size = srq->usrq.umem->page_size;
212 	} else {
213 		struct ecore_chain *pbl;
214 		ret = qlnxr_alloc_srq_kernel_params(srq, dev, init_attr);
215 		if (ret)
216 			goto err0;
217 		pbl = &hw_srq->pbl;
218 
219 		page_cnt = ecore_chain_get_page_cnt(pbl);
220 		pbl_base_addr = ecore_chain_get_pbl_phys(pbl);
221 		phy_prod_pair_addr = hw_srq->phy_prod_pair_addr;
222 		page_size = pbl->elem_per_page << 4;
223 	}
224 
225 	in_params.pd_id = get_qlnxr_pd(ibsrq->pd)->pd_id;
226 	in_params.pbl_base_addr = pbl_base_addr;
227 	in_params.prod_pair_addr = phy_prod_pair_addr;
228 	in_params.num_pages = page_cnt;
229 	in_params.page_size = page_size;
230 
231 	ret = ecore_rdma_create_srq(dev->rdma_ctx, &in_params, &out_params);
232 	if (ret)
233 		goto err1;
234 
235 	srq->srq_id = out_params.srq_id;
236 
237 	if (udata) {
238 		ret = qlnxr_copy_srq_uresp(dev, srq, udata);
239 		if (ret)
240 			goto err2;
241 	}
242 
243 	QL_DPRINT12(ha, "created srq with srq_id = 0x%0x\n", srq->srq_id);
244 	return (0);
245 err2:
246 	memset(&in_params, 0, sizeof(in_params));
247 	destroy_in_params.srq_id = srq->srq_id;
248 	ecore_rdma_destroy_srq(dev->rdma_ctx, &destroy_in_params);
249 
250 err1:
251 	if (udata)
252 		qlnxr_free_srq_user_params(srq);
253 	else
254 		qlnxr_free_srq_kernel_params(srq);
255 
256 err0:
257 	return (-EFAULT);
258 }
259 
260 void
261 qlnxr_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
262 {
263 	struct qlnxr_dev	*dev;
264 	struct qlnxr_srq	*srq;
265 	qlnx_host_t		*ha;
266 	struct ecore_rdma_destroy_srq_in_params in_params;
267 
268 	srq = get_qlnxr_srq(ibsrq);
269 	dev = srq->dev;
270 	ha = dev->ha;
271 
272 	memset(&in_params, 0, sizeof(in_params));
273 	in_params.srq_id = srq->srq_id;
274 
275 	ecore_rdma_destroy_srq(dev->rdma_ctx, &in_params);
276 
277 	if (ibsrq->pd->uobject && ibsrq->pd->uobject->context)
278 		qlnxr_free_srq_user_params(srq);
279 	else
280 		qlnxr_free_srq_kernel_params(srq);
281 
282 	QL_DPRINT12(ha, "destroyed srq_id=0x%0x\n", srq->srq_id);
283 }
284 
285 int
286 qlnxr_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
287 	enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
288 {
289 	struct qlnxr_dev	*dev;
290 	struct qlnxr_srq	*srq;
291 	qlnx_host_t		*ha;
292 	struct ecore_rdma_modify_srq_in_params in_params;
293 	int ret = 0;
294 
295 	srq = get_qlnxr_srq(ibsrq);
296 	dev = srq->dev;
297 	ha = dev->ha;
298 
299 	QL_DPRINT12(ha, "enter\n");
300 	if (attr_mask & IB_SRQ_MAX_WR) {
301 		QL_DPRINT12(ha, "invalid attribute mask=0x%x"
302 			" specified for %p\n", attr_mask, srq);
303 		return -EINVAL;
304 	}
305 
306 	if (attr_mask & IB_SRQ_LIMIT) {
307 		if (attr->srq_limit >= srq->hw_srq.max_wr) {
308 			QL_DPRINT12(ha, "invalid srq_limit=0x%x"
309 				" (max_srq_limit = 0x%x)\n",
310 			       attr->srq_limit, srq->hw_srq.max_wr);
311 			return -EINVAL;
312 		}
313 		memset(&in_params, 0, sizeof(in_params));
314 		in_params.srq_id = srq->srq_id;
315 		in_params.wqe_limit = attr->srq_limit;
316 		ret = ecore_rdma_modify_srq(dev->rdma_ctx, &in_params);
317 		if (ret)
318 			return ret;
319 	}
320 
321 	QL_DPRINT12(ha, "modified srq with srq_id = 0x%0x\n", srq->srq_id);
322 	return 0;
323 }
324 
325 int
326 qlnxr_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
327 {
328 	struct qlnxr_dev	*dev;
329 	struct qlnxr_srq	*srq;
330 	qlnx_host_t		*ha;
331 	struct ecore_rdma_device *qattr;
332 	srq = get_qlnxr_srq(ibsrq);
333 	dev = srq->dev;
334 	ha = dev->ha;
335 	//qattr = &dev->attr;
336 	qattr = ecore_rdma_query_device(dev->rdma_ctx);
337 	QL_DPRINT12(ha, "enter\n");
338 
339 	if (!dev->rdma_ctx) {
340 		QL_DPRINT12(ha, "called with invalid params"
341 			" rdma_ctx is NULL\n");
342 		return -EINVAL;
343 	}
344 
345 	srq_attr->srq_limit = qattr->max_srq;
346 	srq_attr->max_wr = qattr->max_srq_wr;
347 	srq_attr->max_sge = qattr->max_sge;
348 
349 	QL_DPRINT12(ha, "exit\n");
350 	return 0;
351 }
352 
353 /* Increment srq wr producer by one */
354 static
355 void qlnxr_inc_srq_wr_prod (struct qlnxr_srq_hwq_info *info)
356 {
357 	info->wr_prod_cnt++;
358 }
359 
360 /* Increment srq wr consumer by one */
361 static
362 void qlnxr_inc_srq_wr_cons(struct qlnxr_srq_hwq_info *info)
363 {
364         info->wr_cons_cnt++;
365 }
366 
367 /* get_port_immutable verb is not available in FreeBSD */
368 #if 0
369 int
370 qlnxr_roce_port_immutable(struct ib_device *ibdev, u8 port_num,
371 	struct ib_port_immutable *immutable)
372 {
373 	struct qlnxr_dev                *dev;
374 	qlnx_host_t                     *ha;
375 	dev = get_qlnxr_dev(ibdev);
376 	ha = dev->ha;
377 
378 	QL_DPRINT12(ha, "entered but not implemented!!!\n");
379 }
380 #endif
381 
382 int
383 qlnxr_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
384 	const struct ib_recv_wr **bad_wr)
385 {
386 	struct qlnxr_dev	*dev;
387 	struct qlnxr_srq	*srq;
388 	qlnx_host_t		*ha;
389 	struct qlnxr_srq_hwq_info *hw_srq;
390 	struct ecore_chain *pbl;
391 	unsigned long flags;
392 	int status = 0;
393 	u32 num_sge, offset;
394 
395 	srq = get_qlnxr_srq(ibsrq);
396 	dev = srq->dev;
397 	ha = dev->ha;
398 	hw_srq = &srq->hw_srq;
399 
400 	QL_DPRINT12(ha, "enter\n");
401 	spin_lock_irqsave(&srq->lock, flags);
402 
403 	pbl = &srq->hw_srq.pbl;
404 	while (wr) {
405 		struct rdma_srq_wqe_header *hdr;
406 		int i;
407 
408 		if (!qlnxr_srq_elem_left(hw_srq) ||
409 		    wr->num_sge > srq->hw_srq.max_sges) {
410 			QL_DPRINT11(ha, "WR cannot be posted"
411 			    " (%d, %d) || (%d > %d)\n",
412 			    hw_srq->wr_prod_cnt, hw_srq->wr_cons_cnt,
413 			    wr->num_sge, srq->hw_srq.max_sges);
414 			status = -ENOMEM;
415 			*bad_wr = wr;
416 			break;
417 		}
418 
419 		hdr = ecore_chain_produce(pbl);
420 		num_sge = wr->num_sge;
421 		/* Set number of sge and WR id in header */
422 		SRQ_HDR_SET(hdr, wr->wr_id, num_sge);
423 
424                 /* PBL is maintained in case of WR granularity.
425                  * So increment WR producer in case we post a WR.
426                  */
427 		qlnxr_inc_srq_wr_prod(hw_srq);
428 		hw_srq->wqe_prod++;
429 		hw_srq->sge_prod++;
430 
431 		QL_DPRINT12(ha, "SRQ WR : SGEs: %d with wr_id[%d] = %llx\n",
432 			wr->num_sge, hw_srq->wqe_prod, wr->wr_id);
433 
434 		for (i = 0; i < wr->num_sge; i++) {
435 			struct rdma_srq_sge *srq_sge =
436 			    ecore_chain_produce(pbl);
437 			/* Set SGE length, lkey and address */
438 			SRQ_SGE_SET(srq_sge, wr->sg_list[i].addr,
439 				wr->sg_list[i].length, wr->sg_list[i].lkey);
440 
441 			QL_DPRINT12(ha, "[%d]: len %d, key %x, addr %x:%x\n",
442 				i, srq_sge->length, srq_sge->l_key,
443 				srq_sge->addr.hi, srq_sge->addr.lo);
444 			hw_srq->sge_prod++;
445 		}
446 		wmb();
447 		/*
448 		 * SRQ prod is 8 bytes. Need to update SGE prod in index
449 		 * in first 4 bytes and need to update WQE prod in next
450 		 * 4 bytes.
451 		 */
452 		*(srq->hw_srq.virt_prod_pair_addr) = hw_srq->sge_prod;
453 		offset = offsetof(struct rdma_srq_producers, wqe_prod);
454 		*((u8 *)srq->hw_srq.virt_prod_pair_addr + offset) =
455 			hw_srq->wqe_prod;
456 		/* Flush prod after updating it */
457 		wmb();
458 		wr = wr->next;
459 	}
460 
461 	QL_DPRINT12(ha, "Elements in SRQ: %d\n",
462 		ecore_chain_get_elem_left(pbl));
463 
464 	spin_unlock_irqrestore(&srq->lock, flags);
465 	QL_DPRINT12(ha, "exit\n");
466 	return status;
467 }
468 
469 int
470 qlnxr_query_device(struct ib_device *ibdev, struct ib_device_attr *attr,
471 	struct ib_udata *udata)
472 {
473 	struct qlnxr_dev		*dev;
474 	struct ecore_rdma_device	*qattr;
475 	qlnx_host_t			*ha;
476 
477 	dev = get_qlnxr_dev(ibdev);
478 	ha = dev->ha;
479 
480 	QL_DPRINT12(ha, "enter\n");
481 
482 	if (udata->inlen || udata->outlen)
483 		return -EINVAL;
484 
485 	if (dev->rdma_ctx == NULL) {
486 		return -EINVAL;
487 	}
488 
489 	qattr = ecore_rdma_query_device(dev->rdma_ctx);
490 
491 	memset(attr, 0, sizeof *attr);
492 
493 	attr->fw_ver = qattr->fw_ver;
494 	attr->sys_image_guid = qattr->sys_image_guid;
495 	attr->max_mr_size = qattr->max_mr_size;
496 	attr->page_size_cap = qattr->page_size_caps;
497 	attr->vendor_id = qattr->vendor_id;
498 	attr->vendor_part_id = qattr->vendor_part_id;
499 	attr->hw_ver = qattr->hw_ver;
500 	attr->max_qp = qattr->max_qp;
501 	attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
502 					IB_DEVICE_RC_RNR_NAK_GEN |
503 					IB_DEVICE_LOCAL_DMA_LKEY |
504 					IB_DEVICE_MEM_MGT_EXTENSIONS;
505 
506 	attr->max_sge = qattr->max_sge;
507 	attr->max_sge_rd = qattr->max_sge;
508 	attr->max_cq = qattr->max_cq;
509 	attr->max_cqe = qattr->max_cqe;
510 	attr->max_mr = qattr->max_mr;
511 	attr->max_mw = qattr->max_mw;
512 	attr->max_pd = qattr->max_pd;
513 	attr->atomic_cap = dev->atomic_cap;
514 	attr->max_fmr = qattr->max_fmr;
515 	attr->max_map_per_fmr = 16; /* TBD: FMR */
516 
517 	/* There is an implicit assumption in some of the ib_xxx apps that the
518 	 * qp_rd_atom is smaller than the qp_init_rd_atom. Specifically, in
519 	 * communication the qp_rd_atom is passed to the other side and used as
520 	 * init_rd_atom without check device capabilities for init_rd_atom.
521 	 * for this reason, we set the qp_rd_atom to be the minimum between the
522 	 * two...There is an additional assumption in mlx4 driver that the
523 	 * values are power of two, fls is performed on the value - 1, which
524 	 * in fact gives a larger power of two for values which are not a power
525 	 * of two. This should be fixed in mlx4 driver, but until then ->
526 	 * we provide a value that is a power of two in our code.
527 	 */
528 	attr->max_qp_init_rd_atom =
529 		1 << (fls(qattr->max_qp_req_rd_atomic_resc) - 1);
530 	attr->max_qp_rd_atom =
531 		min(1 << (fls(qattr->max_qp_resp_rd_atomic_resc) - 1),
532 		    attr->max_qp_init_rd_atom);
533 
534 	attr->max_srq = qattr->max_srq;
535 	attr->max_srq_sge = qattr->max_srq_sge;
536 	attr->max_srq_wr = qattr->max_srq_wr;
537 
538 	/* TODO: R&D to more properly configure the following */
539 	attr->local_ca_ack_delay = qattr->dev_ack_delay;
540 	attr->max_fast_reg_page_list_len = qattr->max_mr/8;
541 	attr->max_pkeys = QLNXR_ROCE_PKEY_MAX;
542 	attr->max_ah = qattr->max_ah;
543 
544 	QL_DPRINT12(ha, "exit\n");
545 	return 0;
546 }
547 
548 static inline void
549 get_link_speed_and_width(int speed, uint8_t *ib_speed, uint8_t *ib_width)
550 {
551 	switch (speed) {
552 	case 1000:
553 		*ib_speed = IB_SPEED_SDR;
554 		*ib_width = IB_WIDTH_1X;
555 		break;
556 	case 10000:
557 		*ib_speed = IB_SPEED_QDR;
558 		*ib_width = IB_WIDTH_1X;
559 		break;
560 
561 	case 20000:
562 		*ib_speed = IB_SPEED_DDR;
563 		*ib_width = IB_WIDTH_4X;
564 		break;
565 
566 	case 25000:
567 		*ib_speed = IB_SPEED_EDR;
568 		*ib_width = IB_WIDTH_1X;
569 		break;
570 
571 	case 40000:
572 		*ib_speed = IB_SPEED_QDR;
573 		*ib_width = IB_WIDTH_4X;
574 		break;
575 
576 	case 50000:
577 		*ib_speed = IB_SPEED_QDR;
578 		*ib_width = IB_WIDTH_4X; // TODO doesn't add up to 50...
579 		break;
580 
581 	case 100000:
582 		*ib_speed = IB_SPEED_EDR;
583 		*ib_width = IB_WIDTH_4X;
584 		break;
585 
586 	default:
587 		/* Unsupported */
588 		*ib_speed = IB_SPEED_SDR;
589 		*ib_width = IB_WIDTH_1X;
590 	}
591 	return;
592 }
593 
594 int
595 qlnxr_query_port(struct ib_device *ibdev, uint8_t port,
596 	struct ib_port_attr *attr)
597 {
598 	struct qlnxr_dev	*dev;
599 	struct ecore_rdma_port	*rdma_port;
600 	qlnx_host_t		*ha;
601 
602 	dev = get_qlnxr_dev(ibdev);
603 	ha = dev->ha;
604 
605 	QL_DPRINT12(ha, "enter\n");
606 
607 	if (port > 1) {
608 		QL_DPRINT12(ha, "port [%d] > 1 \n", port);
609 		return -EINVAL;
610 	}
611 
612 	if (dev->rdma_ctx == NULL) {
613 		QL_DPRINT12(ha, "rdma_ctx == NULL\n");
614 		return -EINVAL;
615 	}
616 
617 	rdma_port = ecore_rdma_query_port(dev->rdma_ctx);
618 	memset(attr, 0, sizeof *attr);
619 
620 	if (rdma_port->port_state == ECORE_RDMA_PORT_UP) {
621 		attr->state = IB_PORT_ACTIVE;
622 		attr->phys_state = 5;
623 	} else {
624 		attr->state = IB_PORT_DOWN;
625 		attr->phys_state = 3;
626 	}
627 
628 	attr->max_mtu = IB_MTU_4096;
629 	attr->active_mtu = iboe_get_mtu(if_getmtu(dev->ha->ifp));
630 	attr->lid = 0;
631 	attr->lmc = 0;
632 	attr->sm_lid = 0;
633 	attr->sm_sl = 0;
634 	attr->port_cap_flags = 0;
635 
636 	if (QLNX_IS_IWARP(dev)) {
637 		attr->gid_tbl_len = 1;
638 		attr->pkey_tbl_len = 1;
639 	} else {
640 		attr->gid_tbl_len = QLNXR_MAX_SGID;
641 		attr->pkey_tbl_len = QLNXR_ROCE_PKEY_TABLE_LEN;
642 	}
643 
644 	attr->bad_pkey_cntr = rdma_port->pkey_bad_counter;
645 	attr->qkey_viol_cntr = 0;
646 
647 	get_link_speed_and_width(rdma_port->link_speed,
648 				 &attr->active_speed, &attr->active_width);
649 
650 	attr->max_msg_sz = rdma_port->max_msg_size;
651 	attr->max_vl_num = 4; /* TODO -> figure this one out... */
652 
653 	QL_DPRINT12(ha, "state = %d phys_state = %d "
654 		" link_speed = %d active_speed = %d active_width = %d"
655 		" attr->gid_tbl_len = %d attr->pkey_tbl_len = %d"
656 		" max_msg_sz = 0x%x max_vl_num = 0x%x \n",
657 		attr->state, attr->phys_state,
658 		rdma_port->link_speed, attr->active_speed,
659 		attr->active_width, attr->gid_tbl_len, attr->pkey_tbl_len,
660 		attr->max_msg_sz, attr->max_vl_num);
661 
662 	QL_DPRINT12(ha, "exit\n");
663 	return 0;
664 }
665 
666 int
667 qlnxr_modify_port(struct ib_device *ibdev, uint8_t port, int mask,
668 	struct ib_port_modify *props)
669 {
670 	struct qlnxr_dev	*dev;
671 	qlnx_host_t		*ha;
672 
673 	dev = get_qlnxr_dev(ibdev);
674 	ha = dev->ha;
675 
676 	QL_DPRINT12(ha, "enter\n");
677 
678 	if (port > 1) {
679 		QL_DPRINT12(ha, "port (%d) > 1\n", port);
680 		return -EINVAL;
681 	}
682 
683 	QL_DPRINT12(ha, "exit\n");
684 	return 0;
685 }
686 
687 enum rdma_link_layer
688 qlnxr_link_layer(struct ib_device *ibdev, uint8_t port_num)
689 {
690 	struct qlnxr_dev	*dev;
691 	qlnx_host_t		*ha;
692 
693 	dev = get_qlnxr_dev(ibdev);
694 	ha = dev->ha;
695 
696 	QL_DPRINT12(ha, "ibdev = %p port_num = 0x%x\n", ibdev, port_num);
697 
698         return IB_LINK_LAYER_ETHERNET;
699 }
700 
701 int
702 qlnxr_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
703 {
704 	struct ib_device *ibdev = ibpd->device;
705 	struct qlnxr_pd		*pd = get_qlnxr_pd(ibpd);
706 	u16			pd_id;
707 	int			rc;
708 	struct qlnxr_dev	*dev;
709 	qlnx_host_t		*ha;
710 
711 	dev = get_qlnxr_dev(ibdev);
712 	ha = dev->ha;
713 
714 	QL_DPRINT12(ha, "ibdev = %p udata = %p enter\n", ibdev, udata);
715 
716 	if (dev->rdma_ctx == NULL) {
717 		QL_DPRINT11(ha, "dev->rdma_ctx = NULL\n");
718 		rc = -1;
719 		goto err;
720 	}
721 
722 	rc = ecore_rdma_alloc_pd(dev->rdma_ctx, &pd_id);
723 	if (rc)	{
724 		QL_DPRINT11(ha, "ecore_rdma_alloc_pd failed\n");
725 		goto err;
726 	}
727 
728 	pd->pd_id = pd_id;
729 
730 	if (udata) {
731 		rc = ib_copy_to_udata(udata, &pd->pd_id, sizeof(pd->pd_id));
732 		if (rc) {
733 			QL_DPRINT11(ha, "ib_copy_to_udata failed\n");
734 			ecore_rdma_free_pd(dev->rdma_ctx, pd_id);
735 			goto err;
736 		}
737 
738 		pd->uctx = rdma_udata_to_drv_context(
739 		    udata, struct qlnxr_ucontext, ibucontext);
740 		pd->uctx->pd = pd;
741 	}
742 
743 	atomic_add_rel_32(&dev->pd_count, 1);
744 	QL_DPRINT12(ha, "exit [pd, pd_id, pd_count] = [%p, 0x%x, %d]\n",
745 		pd, pd_id, dev->pd_count);
746 
747 	return (0);
748 
749 err:
750 	QL_DPRINT12(ha, "exit -1\n");
751 	return (rc);
752 }
753 
754 void
755 qlnxr_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
756 {
757 	struct qlnxr_pd		*pd;
758 	struct qlnxr_dev	*dev;
759 	qlnx_host_t		*ha;
760 
761 	pd = get_qlnxr_pd(ibpd);
762 	dev = get_qlnxr_dev((ibpd->device));
763 	ha = dev->ha;
764 
765 	QL_DPRINT12(ha, "enter\n");
766 
767 	if (pd == NULL) {
768 		QL_DPRINT11(ha, "pd = NULL\n");
769 	} else {
770 		ecore_rdma_free_pd(dev->rdma_ctx, pd->pd_id);
771 		atomic_subtract_rel_32(&dev->pd_count, 1);
772 		QL_DPRINT12(ha, "exit [pd, pd_id, pd_count] = [%p, 0x%x, %d]\n",
773 			pd, pd->pd_id, dev->pd_count);
774 	}
775 
776 	QL_DPRINT12(ha, "exit\n");
777 }
778 
779 #define ROCE_WQE_ELEM_SIZE	sizeof(struct rdma_sq_sge)
780 #define	RDMA_MAX_SGE_PER_SRQ	(4) /* Should be part of HSI */
781 /* Should be part of HSI */
782 #define RDMA_MAX_SRQ_WQE_SIZE	(RDMA_MAX_SGE_PER_SRQ + 1) /* +1 for header */
783 #define DB_ADDR_SHIFT(addr)		((addr) << DB_PWM_ADDR_OFFSET_SHIFT)
784 
785 static void qlnxr_cleanup_user(struct qlnxr_dev *, struct qlnxr_qp *);
786 static void qlnxr_cleanup_kernel(struct qlnxr_dev *, struct qlnxr_qp *);
787 
788 int
789 qlnxr_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
790 {
791 	struct qlnxr_dev	*dev;
792 	qlnx_host_t		*ha;
793 
794 	dev = get_qlnxr_dev(ibdev);
795 	ha = dev->ha;
796 
797 	QL_DPRINT12(ha, "enter index = 0x%x\n", index);
798 
799 	if (index > QLNXR_ROCE_PKEY_TABLE_LEN)
800 		return -EINVAL;
801 
802 	*pkey = QLNXR_ROCE_PKEY_DEFAULT;
803 
804 	QL_DPRINT12(ha, "exit\n");
805 	return 0;
806 }
807 
808 static inline bool
809 qlnxr_get_vlan_id_qp(qlnx_host_t *ha, struct ib_qp_attr *attr, int attr_mask,
810        u16 *vlan_id)
811 {
812 	bool ret = false;
813 
814 	u16 tmp_vlan_id;
815 
816 	union ib_gid *dgid;
817 
818 	QL_DPRINT12(ha, "enter \n");
819 
820 	*vlan_id = 0;
821 
822 	dgid = &attr->ah_attr.grh.dgid;
823 	tmp_vlan_id = (dgid->raw[11] << 8) | dgid->raw[12];
824 
825 	if (!(tmp_vlan_id & ~EVL_VLID_MASK)) {
826 		*vlan_id = tmp_vlan_id;
827 		ret = true;
828 	}
829 
830 	QL_DPRINT12(ha, "exit vlan_id = 0x%x ret = %d \n", *vlan_id, ret);
831 
832 	return (ret);
833 }
834 
835 static inline void
836 get_gid_info(struct ib_qp *ibqp, struct ib_qp_attr *attr,
837 	int attr_mask,
838 	struct qlnxr_dev *dev,
839 	struct qlnxr_qp *qp,
840 	struct ecore_rdma_modify_qp_in_params *qp_params)
841 {
842 	int		i;
843 	qlnx_host_t	*ha;
844 
845 	ha = dev->ha;
846 
847 	QL_DPRINT12(ha, "enter\n");
848 
849 	memcpy(&qp_params->sgid.bytes[0],
850 	       &dev->sgid_tbl[qp->sgid_idx].raw[0],
851 	       sizeof(qp_params->sgid.bytes));
852 	memcpy(&qp_params->dgid.bytes[0],
853 	       &attr->ah_attr.grh.dgid.raw[0],
854 	       sizeof(qp_params->dgid));
855 
856 	qlnxr_get_vlan_id_qp(ha, attr, attr_mask, &qp_params->vlan_id);
857 
858 	for (i = 0; i < (sizeof(qp_params->sgid.dwords)/sizeof(uint32_t)); i++) {
859 		qp_params->sgid.dwords[i] = ntohl(qp_params->sgid.dwords[i]);
860 		qp_params->dgid.dwords[i] = ntohl(qp_params->dgid.dwords[i]);
861 	}
862 
863 	QL_DPRINT12(ha, "exit\n");
864 	return;
865 }
866 
867 static int
868 qlnxr_add_mmap(struct qlnxr_ucontext *uctx, u64 phy_addr, unsigned long len)
869 {
870 	struct qlnxr_mm	*mm;
871 	qlnx_host_t	*ha;
872 
873 	ha = uctx->dev->ha;
874 
875 	QL_DPRINT12(ha, "enter\n");
876 
877 	mm = kzalloc(sizeof(*mm), GFP_KERNEL);
878 	if (mm == NULL) {
879 		QL_DPRINT11(ha, "mm = NULL\n");
880 		return -ENOMEM;
881 	}
882 
883 	mm->key.phy_addr = phy_addr;
884 
885 	/* This function might be called with a length which is not a multiple
886 	 * of PAGE_SIZE, while the mapping is PAGE_SIZE grained and the kernel
887 	 * forces this granularity by increasing the requested size if needed.
888 	 * When qedr_mmap is called, it will search the list with the updated
889 	 * length as a key. To prevent search failures, the length is rounded up
890 	 * in advance to PAGE_SIZE.
891 	 */
892 	mm->key.len = roundup(len, PAGE_SIZE);
893 	INIT_LIST_HEAD(&mm->entry);
894 
895 	mutex_lock(&uctx->mm_list_lock);
896 	list_add(&mm->entry, &uctx->mm_head);
897 	mutex_unlock(&uctx->mm_list_lock);
898 
899 	QL_DPRINT12(ha, "added (addr=0x%llx,len=0x%lx) for ctx=%p\n",
900 		(unsigned long long)mm->key.phy_addr,
901 		(unsigned long)mm->key.len, uctx);
902 
903 	return 0;
904 }
905 
906 static bool
907 qlnxr_search_mmap(struct qlnxr_ucontext *uctx, u64 phy_addr, unsigned long len)
908 {
909 	bool		found = false;
910 	struct qlnxr_mm	*mm;
911 	qlnx_host_t	*ha;
912 
913 	ha = uctx->dev->ha;
914 
915 	QL_DPRINT12(ha, "enter\n");
916 
917 	mutex_lock(&uctx->mm_list_lock);
918 	list_for_each_entry(mm, &uctx->mm_head, entry) {
919 		if (len != mm->key.len || phy_addr != mm->key.phy_addr)
920 			continue;
921 
922 		found = true;
923 		break;
924 	}
925 	mutex_unlock(&uctx->mm_list_lock);
926 
927 	QL_DPRINT12(ha,
928 		"searched for (addr=0x%llx,len=0x%lx) for ctx=%p, found=%d\n",
929 		mm->key.phy_addr, mm->key.len, uctx, found);
930 
931 	return found;
932 }
933 
934 int
935 qlnxr_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
936 {
937         int rc;
938         struct qlnxr_ucontext *ctx = get_qlnxr_ucontext(uctx);
939         struct qlnxr_alloc_ucontext_resp uresp;
940         struct qlnxr_dev *dev = get_qlnxr_dev(uctx->device);
941         qlnx_host_t *ha = dev->ha;
942         struct ecore_rdma_add_user_out_params oparams;
943 
944         if (!udata)
945                 return -EFAULT;
946 
947 	rc = ecore_rdma_add_user(dev->rdma_ctx, &oparams);
948 	if (rc) {
949 		QL_DPRINT12(ha,
950 			"Failed to allocate a DPI for a new RoCE application "
951 			",rc = %d. To overcome this, consider to increase "
952 			"the number of DPIs, increase the doorbell BAR size "
953 			"or just close unnecessary RoCE applications. In "
954 			"order to increase the number of DPIs consult the "
955 			"README\n", rc);
956 		goto err;
957 	}
958 
959 	ctx->dpi = oparams.dpi;
960 	ctx->dpi_addr = oparams.dpi_addr;
961 	ctx->dpi_phys_addr = oparams.dpi_phys_addr;
962 	ctx->dpi_size = oparams.dpi_size;
963 	INIT_LIST_HEAD(&ctx->mm_head);
964 	mutex_init(&ctx->mm_list_lock);
965 
966 	memset(&uresp, 0, sizeof(uresp));
967 	uresp.dpm_enabled = offsetof(struct qlnxr_alloc_ucontext_resp, dpm_enabled)
968 				< udata->outlen ? dev->user_dpm_enabled : 0; //TODO: figure this out
969 	uresp.wids_enabled = offsetof(struct qlnxr_alloc_ucontext_resp, wids_enabled)
970 				< udata->outlen ? 1 : 0; //TODO: figure this out
971 	uresp.wid_count = offsetof(struct qlnxr_alloc_ucontext_resp, wid_count)
972 				< udata->outlen ? oparams.wid_count : 0; //TODO: figure this out
973         uresp.db_pa = ctx->dpi_phys_addr;
974         uresp.db_size = ctx->dpi_size;
975         uresp.max_send_wr = dev->attr.max_sqe;
976         uresp.max_recv_wr = dev->attr.max_rqe;
977         uresp.max_srq_wr = dev->attr.max_srq_wr;
978         uresp.sges_per_send_wr = QLNXR_MAX_SQE_ELEMENTS_PER_SQE;
979         uresp.sges_per_recv_wr = QLNXR_MAX_RQE_ELEMENTS_PER_RQE;
980         uresp.sges_per_srq_wr = dev->attr.max_srq_sge;
981         uresp.max_cqes = QLNXR_MAX_CQES;
982 
983 	rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
984 	if (rc)
985 		goto err;
986 
987 	ctx->dev = dev;
988 
989 	rc = qlnxr_add_mmap(ctx, ctx->dpi_phys_addr, ctx->dpi_size);
990 	if (rc)
991 		goto err;
992 	QL_DPRINT12(ha, "Allocated user context %p\n",
993 		&ctx->ibucontext);
994 
995 	return (0);
996 err:
997 	return (rc);
998 }
999 
1000 void
1001 qlnxr_dealloc_ucontext(struct ib_ucontext *ibctx)
1002 {
1003         struct qlnxr_ucontext *uctx = get_qlnxr_ucontext(ibctx);
1004         struct qlnxr_dev *dev = uctx->dev;
1005         qlnx_host_t *ha = dev->ha;
1006         struct qlnxr_mm *mm, *tmp;
1007 
1008         QL_DPRINT12(ha, "Deallocating user context %p\n",
1009                         uctx);
1010 
1011         if (dev) {
1012                 ecore_rdma_remove_user(uctx->dev->rdma_ctx, uctx->dpi);
1013         }
1014 
1015         list_for_each_entry_safe(mm, tmp, &uctx->mm_head, entry) {
1016                 QL_DPRINT12(ha, "deleted addr= 0x%llx, len = 0x%lx for"
1017                                 " ctx=%p\n",
1018                                 mm->key.phy_addr, mm->key.len, uctx);
1019                 list_del(&mm->entry);
1020                 kfree(mm);
1021         }
1022 }
1023 
1024 int
1025 qlnxr_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
1026 {
1027 	struct qlnxr_ucontext	*ucontext = get_qlnxr_ucontext(context);
1028 	struct qlnxr_dev	*dev = get_qlnxr_dev((context->device));
1029 	unsigned long		vm_page = vma->vm_pgoff << PAGE_SHIFT;
1030 	u64 			unmapped_db;
1031 	unsigned long 		len = (vma->vm_end - vma->vm_start);
1032 	int 			rc = 0;
1033 	bool 			found;
1034 	qlnx_host_t		*ha;
1035 
1036 	ha = dev->ha;
1037 
1038 	unmapped_db = dev->db_phys_addr + (ucontext->dpi * ucontext->dpi_size);
1039 
1040 	QL_DPRINT12(ha, "qedr_mmap enter vm_page=0x%lx"
1041 		" vm_pgoff=0x%lx unmapped_db=0x%llx db_size=%x, len=%lx\n",
1042 		vm_page, vma->vm_pgoff, unmapped_db,
1043 		dev->db_size, len);
1044 
1045 	if ((vma->vm_start & (PAGE_SIZE - 1)) || (len & (PAGE_SIZE - 1))) {
1046 		QL_DPRINT11(ha, "Vma_start not page aligned "
1047 			"vm_start = %ld vma_end = %ld\n", vma->vm_start,
1048 			vma->vm_end);
1049 		return -EINVAL;
1050 	}
1051 
1052 	found = qlnxr_search_mmap(ucontext, vm_page, len);
1053 	if (!found) {
1054 		QL_DPRINT11(ha, "Vma_pgoff not found in mapped array = %ld\n",
1055 			vma->vm_pgoff);
1056 		return -EINVAL;
1057 	}
1058 
1059 	QL_DPRINT12(ha, "Mapping doorbell bar\n");
1060 
1061 	if ((vm_page < unmapped_db) ||
1062 		((vm_page + len) > (unmapped_db + ucontext->dpi_size))) {
1063 		QL_DPRINT11(ha, "failed pages are outside of dpi;"
1064 			"page address=0x%lx, unmapped_db=0x%lx, dpi_size=0x%x\n",
1065 			vm_page, unmapped_db, ucontext->dpi_size);
1066 		return -EINVAL;
1067 	}
1068 
1069 	if (vma->vm_flags & VM_READ) {
1070 		QL_DPRINT11(ha, "failed mmap, cannot map doorbell bar for read\n");
1071 		return -EINVAL;
1072 	}
1073 
1074 	vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
1075 	rc = io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff, len,
1076 			vma->vm_page_prot);
1077 
1078 	QL_DPRINT12(ha, "exit [%d]\n", rc);
1079 	return rc;
1080 }
1081 
1082 struct ib_mr *
1083 qlnxr_get_dma_mr(struct ib_pd *ibpd, int acc)
1084 {
1085 	struct qlnxr_mr		*mr;
1086 	struct qlnxr_dev	*dev = get_qlnxr_dev((ibpd->device));
1087 	struct qlnxr_pd		*pd = get_qlnxr_pd(ibpd);
1088 	int			rc;
1089 	qlnx_host_t		*ha;
1090 
1091 	ha = dev->ha;
1092 
1093 	QL_DPRINT12(ha, "enter\n");
1094 
1095 	if (acc & IB_ACCESS_MW_BIND) {
1096 		QL_DPRINT12(ha, "Unsupported access flags received for dma mr\n");
1097 	}
1098 
1099 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1100 	if (!mr) {
1101 		rc = -ENOMEM;
1102 		QL_DPRINT12(ha, "kzalloc(mr) failed %d\n", rc);
1103 		goto err0;
1104 	}
1105 
1106 	mr->type = QLNXR_MR_DMA;
1107 
1108 	rc = ecore_rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
1109 	if (rc) {
1110 		QL_DPRINT12(ha, "ecore_rdma_alloc_tid failed %d\n", rc);
1111 		goto err1;
1112 	}
1113 
1114 	/* index only, 18 bit long, lkey = itid << 8 | key */
1115 	mr->hw_mr.tid_type = ECORE_RDMA_TID_REGISTERED_MR;
1116 	mr->hw_mr.pd = pd->pd_id;
1117 	mr->hw_mr.local_read = 1;
1118 	mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
1119 	mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
1120 	mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1121 	mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
1122 	mr->hw_mr.dma_mr = true;
1123 
1124 	rc = ecore_rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
1125 	if (rc) {
1126 		QL_DPRINT12(ha, "ecore_rdma_register_tid failed %d\n", rc);
1127 		goto err2;
1128 	}
1129 
1130 	mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
1131 
1132 	if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
1133 		mr->hw_mr.remote_atomic) {
1134 		mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
1135 	}
1136 
1137 	QL_DPRINT12(ha, "lkey = %x\n", mr->ibmr.lkey);
1138 
1139 	return &mr->ibmr;
1140 
1141 err2:
1142 	ecore_rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
1143 err1:
1144 	kfree(mr);
1145 err0:
1146 	QL_DPRINT12(ha, "exit [%d]\n", rc);
1147 
1148 	return ERR_PTR(rc);
1149 }
1150 
1151 static void
1152 qlnxr_free_pbl(struct qlnxr_dev *dev, struct qlnxr_pbl_info *pbl_info,
1153 	struct qlnxr_pbl *pbl)
1154 {
1155 	int		i;
1156 	qlnx_host_t	*ha;
1157 
1158 	ha = dev->ha;
1159 
1160 	QL_DPRINT12(ha, "enter\n");
1161 
1162 	for (i = 0; i < pbl_info->num_pbls; i++) {
1163 		if (!pbl[i].va)
1164 			continue;
1165 		qlnx_dma_free_coherent(&dev->ha->cdev, pbl[i].va, pbl[i].pa,
1166 			pbl_info->pbl_size);
1167 	}
1168 	kfree(pbl);
1169 
1170 	QL_DPRINT12(ha, "exit\n");
1171 	return;
1172 }
1173 
1174 #define MIN_FW_PBL_PAGE_SIZE (4*1024)
1175 #define MAX_FW_PBL_PAGE_SIZE (64*1024)
1176 
1177 #define NUM_PBES_ON_PAGE(_page_size) (_page_size / sizeof(u64))
1178 #define MAX_PBES_ON_PAGE NUM_PBES_ON_PAGE(MAX_FW_PBL_PAGE_SIZE)
1179 #define MAX_PBES_TWO_LAYER (MAX_PBES_ON_PAGE*MAX_PBES_ON_PAGE)
1180 
1181 static struct qlnxr_pbl *
1182 qlnxr_alloc_pbl_tbl(struct qlnxr_dev *dev,
1183 	struct qlnxr_pbl_info *pbl_info, gfp_t flags)
1184 {
1185 	void			*va;
1186 	dma_addr_t		pa;
1187 	dma_addr_t		*pbl_main_tbl;
1188 	struct qlnxr_pbl	*pbl_table;
1189 	int			i;
1190 	qlnx_host_t		*ha;
1191 
1192 	ha = dev->ha;
1193 
1194 	QL_DPRINT12(ha, "enter\n");
1195 
1196 	pbl_table = kzalloc(sizeof(*pbl_table) * pbl_info->num_pbls, flags);
1197 
1198 	if (!pbl_table) {
1199 		QL_DPRINT12(ha, "pbl_table = NULL\n");
1200 		return NULL;
1201 	}
1202 
1203 	for (i = 0; i < pbl_info->num_pbls; i++) {
1204 		va = qlnx_dma_alloc_coherent(&dev->ha->cdev, &pa, pbl_info->pbl_size);
1205 		if (!va) {
1206 			QL_DPRINT11(ha, "Failed to allocate pbl#%d\n", i);
1207 			goto err;
1208 		}
1209 		memset(va, 0, pbl_info->pbl_size);
1210 		pbl_table[i].va = va;
1211 		pbl_table[i].pa = pa;
1212 	}
1213 
1214 	/* Two-Layer PBLs, if we have more than one pbl we need to initialize
1215 	 * the first one with physical pointers to all of the rest
1216 	 */
1217 	pbl_main_tbl = (dma_addr_t *)pbl_table[0].va;
1218 	for (i = 0; i < pbl_info->num_pbls - 1; i++)
1219 		pbl_main_tbl[i] = pbl_table[i + 1].pa;
1220 
1221 	QL_DPRINT12(ha, "exit\n");
1222 	return pbl_table;
1223 
1224 err:
1225 	qlnxr_free_pbl(dev, pbl_info, pbl_table);
1226 
1227 	QL_DPRINT12(ha, "exit with error\n");
1228 	return NULL;
1229 }
1230 
1231 static int
1232 qlnxr_prepare_pbl_tbl(struct qlnxr_dev *dev,
1233 	struct qlnxr_pbl_info *pbl_info,
1234 	u32 num_pbes,
1235 	int two_layer_capable)
1236 {
1237 	u32		pbl_capacity;
1238 	u32		pbl_size;
1239 	u32		num_pbls;
1240 	qlnx_host_t	*ha;
1241 
1242 	ha = dev->ha;
1243 
1244 	QL_DPRINT12(ha, "enter\n");
1245 
1246 	if ((num_pbes > MAX_PBES_ON_PAGE) && two_layer_capable) {
1247 		if (num_pbes > MAX_PBES_TWO_LAYER) {
1248 			QL_DPRINT11(ha, "prepare pbl table: too many pages %d\n",
1249 				num_pbes);
1250 			return -EINVAL;
1251 		}
1252 
1253 		/* calculate required pbl page size */
1254 		pbl_size = MIN_FW_PBL_PAGE_SIZE;
1255 		pbl_capacity = NUM_PBES_ON_PAGE(pbl_size) *
1256 			NUM_PBES_ON_PAGE(pbl_size);
1257 
1258 		while (pbl_capacity < num_pbes) {
1259 			pbl_size *= 2;
1260 			pbl_capacity = pbl_size / sizeof(u64);
1261 			pbl_capacity = pbl_capacity * pbl_capacity;
1262 		}
1263 
1264 		num_pbls = DIV_ROUND_UP(num_pbes, NUM_PBES_ON_PAGE(pbl_size));
1265 		num_pbls++; /* One for the layer0 ( points to the pbls) */
1266 		pbl_info->two_layered = true;
1267 	} else {
1268 		/* One layered PBL */
1269 		num_pbls = 1;
1270 		pbl_size = max_t(u32, MIN_FW_PBL_PAGE_SIZE, \
1271 				roundup_pow_of_two((num_pbes * sizeof(u64))));
1272 		pbl_info->two_layered = false;
1273 	}
1274 
1275 	pbl_info->num_pbls = num_pbls;
1276 	pbl_info->pbl_size = pbl_size;
1277 	pbl_info->num_pbes = num_pbes;
1278 
1279 	QL_DPRINT12(ha, "prepare pbl table: num_pbes=%d, num_pbls=%d pbl_size=%d\n",
1280 		pbl_info->num_pbes, pbl_info->num_pbls, pbl_info->pbl_size);
1281 
1282 	return 0;
1283 }
1284 
1285 static void
1286 qlnxr_populate_pbls(struct qlnxr_dev *dev, struct ib_umem *umem,
1287 	struct qlnxr_pbl *pbl, struct qlnxr_pbl_info *pbl_info)
1288 {
1289 	struct regpair		*pbe;
1290 	struct qlnxr_pbl	*pbl_tbl;
1291 	struct scatterlist	*sg;
1292 	int			shift, pg_cnt, pages, pbe_cnt, total_num_pbes = 0;
1293 	qlnx_host_t		*ha;
1294         int                     entry;
1295 
1296 	ha = dev->ha;
1297 
1298 	QL_DPRINT12(ha, "enter\n");
1299 
1300 	if (!pbl_info) {
1301 		QL_DPRINT11(ha, "PBL_INFO not initialized\n");
1302 		return;
1303 	}
1304 
1305 	if (!pbl_info->num_pbes) {
1306 		QL_DPRINT11(ha, "pbl_info->num_pbes == 0\n");
1307 		return;
1308 	}
1309 
1310 	/* If we have a two layered pbl, the first pbl points to the rest
1311 	 * of the pbls and the first entry lays on the second pbl in the table
1312 	 */
1313 	if (pbl_info->two_layered)
1314 		pbl_tbl = &pbl[1];
1315 	else
1316 		pbl_tbl = pbl;
1317 
1318 	pbe = (struct regpair *)pbl_tbl->va;
1319 	if (!pbe) {
1320 		QL_DPRINT12(ha, "pbe is NULL\n");
1321 		return;
1322 	}
1323 
1324 	pbe_cnt = 0;
1325 
1326 	shift = ilog2(umem->page_size);
1327 
1328 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
1329 		pages = sg_dma_len(sg) >> shift;
1330 		for (pg_cnt = 0; pg_cnt < pages; pg_cnt++) {
1331 			/* store the page address in pbe */
1332 			pbe->lo =
1333 			    cpu_to_le32(sg_dma_address(sg) +
1334 					(umem->page_size * pg_cnt));
1335 			pbe->hi =
1336 			    cpu_to_le32(upper_32_bits
1337 					((sg_dma_address(sg) +
1338 					  umem->page_size * pg_cnt)));
1339 
1340 			QL_DPRINT12(ha,
1341 				"Populate pbl table:"
1342 				" pbe->addr=0x%x:0x%x "
1343 				" pbe_cnt = %d total_num_pbes=%d"
1344 				" pbe=%p\n", pbe->lo, pbe->hi, pbe_cnt,
1345 				total_num_pbes, pbe);
1346 
1347 			pbe_cnt ++;
1348 			total_num_pbes ++;
1349 			pbe++;
1350 
1351 			if (total_num_pbes == pbl_info->num_pbes)
1352 				return;
1353 
1354 			/* if the given pbl is full storing the pbes,
1355 			 * move to next pbl.
1356 			 */
1357 			if (pbe_cnt ==
1358 				(pbl_info->pbl_size / sizeof(u64))) {
1359 				pbl_tbl++;
1360 				pbe = (struct regpair *)pbl_tbl->va;
1361 				pbe_cnt = 0;
1362 			}
1363 		}
1364 	}
1365 	QL_DPRINT12(ha, "exit\n");
1366 	return;
1367 }
1368 
1369 static void
1370 free_mr_info(struct qlnxr_dev *dev, struct mr_info *info)
1371 {
1372 	struct qlnxr_pbl *pbl, *tmp;
1373 	qlnx_host_t		*ha;
1374 
1375 	ha = dev->ha;
1376 
1377 	QL_DPRINT12(ha, "enter\n");
1378 
1379 	if (info->pbl_table)
1380 		list_add_tail(&info->pbl_table->list_entry,
1381 			      &info->free_pbl_list);
1382 
1383 	if (!list_empty(&info->inuse_pbl_list))
1384 		list_splice(&info->inuse_pbl_list, &info->free_pbl_list);
1385 
1386 	list_for_each_entry_safe(pbl, tmp, &info->free_pbl_list, list_entry) {
1387 		list_del(&pbl->list_entry);
1388 		qlnxr_free_pbl(dev, &info->pbl_info, pbl);
1389 	}
1390 	QL_DPRINT12(ha, "exit\n");
1391 
1392 	return;
1393 }
1394 
1395 static int
1396 qlnxr_init_mr_info(struct qlnxr_dev *dev, struct mr_info *info,
1397 	size_t page_list_len, bool two_layered)
1398 {
1399 	int			rc;
1400 	struct qlnxr_pbl	*tmp;
1401 	qlnx_host_t		*ha;
1402 
1403 	ha = dev->ha;
1404 
1405 	QL_DPRINT12(ha, "enter\n");
1406 
1407 	INIT_LIST_HEAD(&info->free_pbl_list);
1408 	INIT_LIST_HEAD(&info->inuse_pbl_list);
1409 
1410 	rc = qlnxr_prepare_pbl_tbl(dev, &info->pbl_info,
1411 				  page_list_len, two_layered);
1412 	if (rc) {
1413 		QL_DPRINT11(ha, "qlnxr_prepare_pbl_tbl [%d]\n", rc);
1414 		goto done;
1415 	}
1416 
1417 	info->pbl_table = qlnxr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
1418 
1419 	if (!info->pbl_table) {
1420 		rc = -ENOMEM;
1421 		QL_DPRINT11(ha, "qlnxr_alloc_pbl_tbl returned NULL\n");
1422 		goto done;
1423 	}
1424 
1425 	QL_DPRINT12(ha, "pbl_table_pa = %pa\n", &info->pbl_table->pa);
1426 
1427 	/* in usual case we use 2 PBLs, so we add one to free
1428 	 * list and allocating another one
1429 	 */
1430 	tmp = qlnxr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
1431 
1432 	if (!tmp) {
1433 		QL_DPRINT11(ha, "Extra PBL is not allocated\n");
1434 		goto done; /* it's OK if second allocation fails, so rc = 0*/
1435 	}
1436 
1437 	list_add_tail(&tmp->list_entry, &info->free_pbl_list);
1438 
1439 	QL_DPRINT12(ha, "extra pbl_table_pa = %pa\n", &tmp->pa);
1440 
1441 done:
1442 	if (rc)
1443 		free_mr_info(dev, info);
1444 
1445 	QL_DPRINT12(ha, "exit [%d]\n", rc);
1446 
1447 	return rc;
1448 }
1449 
1450 struct ib_mr *
1451 qlnxr_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
1452 	u64 usr_addr, int acc, struct ib_udata *udata)
1453 {
1454 	int		rc = -ENOMEM;
1455 	struct qlnxr_dev *dev = get_qlnxr_dev((ibpd->device));
1456 	struct qlnxr_mr *mr;
1457 	struct qlnxr_pd *pd;
1458 	qlnx_host_t	*ha;
1459 
1460 	ha = dev->ha;
1461 
1462 	QL_DPRINT12(ha, "enter\n");
1463 
1464 	pd = get_qlnxr_pd(ibpd);
1465 
1466 	QL_DPRINT12(ha, "qedr_register user mr pd = %d"
1467 		" start = %lld, len = %lld, usr_addr = %lld, acc = %d\n",
1468 		pd->pd_id, start, len, usr_addr, acc);
1469 
1470 	if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE)) {
1471 		QL_DPRINT11(ha,
1472 			"(acc & IB_ACCESS_REMOTE_WRITE &&"
1473 			" !(acc & IB_ACCESS_LOCAL_WRITE))\n");
1474 		return ERR_PTR(-EINVAL);
1475 	}
1476 
1477 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1478 	if (!mr) {
1479 		QL_DPRINT11(ha, "kzalloc(mr) failed\n");
1480 		return ERR_PTR(rc);
1481 	}
1482 
1483 	mr->type = QLNXR_MR_USER;
1484 
1485 	mr->umem = ib_umem_get(ibpd->uobject->context, start, len, acc, 0);
1486 	if (IS_ERR(mr->umem)) {
1487 		rc = -EFAULT;
1488 		QL_DPRINT11(ha, "ib_umem_get failed [%p]\n", mr->umem);
1489 		goto err0;
1490 	}
1491 
1492 	rc = qlnxr_init_mr_info(dev, &mr->info, ib_umem_page_count(mr->umem), 1);
1493 	if (rc) {
1494 		QL_DPRINT11(ha,
1495 			"qlnxr_init_mr_info failed [%d]\n", rc);
1496 		goto err1;
1497 	}
1498 
1499 	qlnxr_populate_pbls(dev, mr->umem, mr->info.pbl_table,
1500 			   &mr->info.pbl_info);
1501 
1502 	rc = ecore_rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
1503 
1504 	if (rc) {
1505 		QL_DPRINT11(ha, "roce alloc tid returned an error %d\n", rc);
1506 		goto err1;
1507 	}
1508 
1509 	/* index only, 18 bit long, lkey = itid << 8 | key */
1510 	mr->hw_mr.tid_type = ECORE_RDMA_TID_REGISTERED_MR;
1511 	mr->hw_mr.key = 0;
1512 	mr->hw_mr.pd = pd->pd_id;
1513 	mr->hw_mr.local_read = 1;
1514 	mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
1515 	mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
1516 	mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1517 	mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
1518 	mr->hw_mr.mw_bind = false; /* TBD MW BIND */
1519 	mr->hw_mr.pbl_ptr = mr->info.pbl_table[0].pa;
1520 	mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
1521 	mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
1522 	mr->hw_mr.page_size_log = ilog2(mr->umem->page_size); /* for the MR pages */
1523 
1524 	mr->hw_mr.fbo = ib_umem_offset(mr->umem);
1525 	mr->hw_mr.length = len;
1526 	mr->hw_mr.vaddr = usr_addr;
1527 	mr->hw_mr.zbva = false; /* TBD figure when this should be true */
1528 	mr->hw_mr.phy_mr = false; /* Fast MR - True, Regular Register False */
1529 	mr->hw_mr.dma_mr = false;
1530 
1531 	rc = ecore_rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
1532 	if (rc) {
1533 		QL_DPRINT11(ha, "roce register tid returned an error %d\n", rc);
1534 		goto err2;
1535 	}
1536 
1537 	mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
1538 	if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
1539 		mr->hw_mr.remote_atomic)
1540 		mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
1541 
1542 	QL_DPRINT12(ha, "register user mr lkey: %x\n", mr->ibmr.lkey);
1543 
1544 	return (&mr->ibmr);
1545 
1546 err2:
1547 	ecore_rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
1548 err1:
1549 	qlnxr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
1550 err0:
1551 	kfree(mr);
1552 
1553 	QL_DPRINT12(ha, "exit [%d]\n", rc);
1554 	return (ERR_PTR(rc));
1555 }
1556 
1557 int
1558 qlnxr_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
1559 {
1560 	struct qlnxr_mr	*mr = get_qlnxr_mr(ib_mr);
1561 	struct qlnxr_dev *dev = get_qlnxr_dev((ib_mr->device));
1562 	int		rc = 0;
1563 	qlnx_host_t	*ha;
1564 
1565 	ha = dev->ha;
1566 
1567 	QL_DPRINT12(ha, "enter\n");
1568 
1569 	if ((mr->type != QLNXR_MR_DMA) && (mr->type != QLNXR_MR_FRMR))
1570 		qlnxr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
1571 
1572 	/* it could be user registered memory. */
1573 	if (mr->umem)
1574 		ib_umem_release(mr->umem);
1575 
1576 	kfree(mr->pages);
1577 
1578 	kfree(mr);
1579 
1580 	QL_DPRINT12(ha, "exit\n");
1581 	return rc;
1582 }
1583 
1584 static int
1585 qlnxr_copy_cq_uresp(struct qlnxr_dev *dev,
1586 	struct qlnxr_cq *cq, struct ib_udata *udata)
1587 {
1588 	struct qlnxr_create_cq_uresp	uresp;
1589 	int				rc;
1590 	qlnx_host_t			*ha;
1591 
1592 	ha = dev->ha;
1593 
1594 	QL_DPRINT12(ha, "enter\n");
1595 
1596 	memset(&uresp, 0, sizeof(uresp));
1597 
1598 	uresp.db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT);
1599 	uresp.icid = cq->icid;
1600 
1601 	rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1602 
1603 	if (rc) {
1604 		QL_DPRINT12(ha, "ib_copy_to_udata error cqid=0x%x[%d]\n",
1605 			cq->icid, rc);
1606 	}
1607 
1608 	QL_DPRINT12(ha, "exit [%d]\n", rc);
1609 	return rc;
1610 }
1611 
1612 static void
1613 consume_cqe(struct qlnxr_cq *cq)
1614 {
1615 
1616 	if (cq->latest_cqe == cq->toggle_cqe)
1617 		cq->pbl_toggle ^= RDMA_RESIZE_CQ_RAMROD_DATA_TOGGLE_BIT_MASK;
1618 
1619 	cq->latest_cqe = ecore_chain_consume(&cq->pbl);
1620 }
1621 
1622 static inline int
1623 qlnxr_align_cq_entries(int entries)
1624 {
1625 	u64 size, aligned_size;
1626 
1627 	/* We allocate an extra entry that we don't report to the FW.
1628 	 * Why?
1629 	 * The CQE size is 32 bytes but the FW writes in chunks of 64 bytes
1630 	 * (for performance purposes). Allocating an extra entry and telling
1631 	 * the FW we have less prevents overwriting the first entry in case of
1632 	 * a wrap i.e. when the FW writes the last entry and the application
1633 	 * hasn't read the first one.
1634 	 */
1635 	size = (entries + 1) * QLNXR_CQE_SIZE;
1636 
1637 	/* We align to PAGE_SIZE.
1638 	 * Why?
1639 	 * Since the CQ is going to be mapped and the mapping is anyhow in whole
1640 	 * kernel pages we benefit from the possibly extra CQEs.
1641 	 */
1642 	aligned_size = ALIGN(size, PAGE_SIZE);
1643 
1644 	/* note: for CQs created in user space the result of this function
1645 	 * should match the size mapped in user space
1646 	 */
1647 	return (aligned_size / QLNXR_CQE_SIZE);
1648 }
1649 
1650 static inline int
1651 qlnxr_init_user_queue(struct ib_ucontext *ib_ctx, struct qlnxr_dev *dev,
1652 	struct qlnxr_userq *q, u64 buf_addr, size_t buf_len,
1653 	int access, int dmasync, int alloc_and_init)
1654 {
1655 	int		page_cnt;
1656 	int		rc;
1657 	qlnx_host_t	*ha;
1658 
1659 	ha = dev->ha;
1660 
1661 	QL_DPRINT12(ha, "enter\n");
1662 
1663 	q->buf_addr = buf_addr;
1664 	q->buf_len = buf_len;
1665 
1666 	QL_DPRINT12(ha, "buf_addr : %llx, buf_len : %x, access : %x"
1667 	      " dmasync : %x\n", q->buf_addr, q->buf_len,
1668 		access, dmasync);
1669 
1670 	q->umem = ib_umem_get(ib_ctx, q->buf_addr, q->buf_len, access, dmasync);
1671 
1672 	if (IS_ERR(q->umem)) {
1673 		QL_DPRINT11(ha, "ib_umem_get failed [%lx]\n", PTR_ERR(q->umem));
1674 		return PTR_ERR(q->umem);
1675 	}
1676 
1677 	page_cnt = ib_umem_page_count(q->umem);
1678 	rc = qlnxr_prepare_pbl_tbl(dev, &q->pbl_info, page_cnt,
1679 				  0 /* SQ and RQ don't support dual layer pbl.
1680 				     * CQ may, but this is yet uncoded.
1681 				     */);
1682 	if (rc) {
1683 		QL_DPRINT11(ha, "qlnxr_prepare_pbl_tbl failed [%d]\n", rc);
1684 		goto err;
1685 	}
1686 
1687 	if (alloc_and_init) {
1688 		q->pbl_tbl = qlnxr_alloc_pbl_tbl(dev, &q->pbl_info, GFP_KERNEL);
1689 
1690 		if (!q->pbl_tbl) {
1691 			QL_DPRINT11(ha, "qlnxr_alloc_pbl_tbl failed\n");
1692 			rc = -ENOMEM;
1693 			goto err;
1694 		}
1695 
1696 		qlnxr_populate_pbls(dev, q->umem, q->pbl_tbl, &q->pbl_info);
1697 	} else {
1698 		q->pbl_tbl = kzalloc(sizeof(*q->pbl_tbl), GFP_KERNEL);
1699 
1700 		if (!q->pbl_tbl) {
1701 			QL_DPRINT11(ha, "qlnxr_alloc_pbl_tbl failed\n");
1702 			rc = -ENOMEM;
1703 			goto err;
1704 		}
1705 	}
1706 
1707 	QL_DPRINT12(ha, "exit\n");
1708 	return 0;
1709 
1710 err:
1711 	ib_umem_release(q->umem);
1712 	q->umem = NULL;
1713 
1714 	QL_DPRINT12(ha, "exit [%d]\n", rc);
1715 	return rc;
1716 }
1717 
1718 int
1719 qlnxr_create_cq(struct ib_cq *ibcq,
1720 		const struct ib_cq_init_attr *attr,
1721 		struct ib_udata *udata)
1722 {
1723 	struct qlnxr_ucontext			*ctx;
1724 	struct ecore_rdma_destroy_cq_out_params destroy_oparams;
1725 	struct ecore_rdma_destroy_cq_in_params	destroy_iparams;
1726 	struct qlnxr_dev			*dev;
1727 	struct ecore_rdma_create_cq_in_params	params;
1728 	struct qlnxr_create_cq_ureq		ureq;
1729 
1730 	int					vector = attr->comp_vector;
1731 	int					entries = attr->cqe;
1732 	struct qlnxr_cq				*cq = get_qlnxr_cq(ibcq);
1733 	int					chain_entries, rc, page_cnt;
1734 	u64					pbl_ptr;
1735 	u16					icid;
1736 	qlnx_host_t				*ha;
1737 
1738 	dev = get_qlnxr_dev(ibcq->device);
1739 	ha = dev->ha;
1740 
1741 	QL_DPRINT12(ha, "called from %s. entries = %d, "
1742 		"vector = %d\n",
1743 		(udata ? "User Lib" : "Kernel"), entries, vector);
1744 
1745         memset(&params, 0, sizeof(struct ecore_rdma_create_cq_in_params));
1746         memset(&destroy_iparams, 0, sizeof(struct ecore_rdma_destroy_cq_in_params));
1747         memset(&destroy_oparams, 0, sizeof(struct ecore_rdma_destroy_cq_out_params));
1748 
1749 	if (entries > QLNXR_MAX_CQES) {
1750 		QL_DPRINT11(ha,
1751 			"the number of entries %d is too high. "
1752 			"Must be equal or below %d.\n",
1753 			entries, QLNXR_MAX_CQES);
1754 		return -EINVAL;
1755 	}
1756 	chain_entries = qlnxr_align_cq_entries(entries);
1757 	chain_entries = min_t(int, chain_entries, QLNXR_MAX_CQES);
1758 
1759 	if (udata) {
1760 		ctx = rdma_udata_to_drv_context(
1761 		    udata, struct qlnxr_ucontext, ibucontext);
1762 
1763 		memset(&ureq, 0, sizeof(ureq));
1764 
1765 		if (ib_copy_from_udata(&ureq, udata,
1766 			min(sizeof(ureq), udata->inlen))) {
1767 			QL_DPRINT11(ha, "ib_copy_from_udata failed\n");
1768 			goto err0;
1769 		}
1770 
1771 		if (!ureq.len) {
1772 			QL_DPRINT11(ha, "ureq.len == 0\n");
1773 			goto err0;
1774 		}
1775 
1776 		cq->cq_type = QLNXR_CQ_TYPE_USER;
1777 
1778 		qlnxr_init_user_queue(&ctx->ibucontext, dev, &cq->q, ureq.addr, ureq.len,
1779 				     IB_ACCESS_LOCAL_WRITE, 1, 1);
1780 
1781 		pbl_ptr = cq->q.pbl_tbl->pa;
1782 		page_cnt = cq->q.pbl_info.num_pbes;
1783 		cq->ibcq.cqe = chain_entries;
1784 	} else {
1785 		ctx = NULL;
1786 
1787 		cq->cq_type = QLNXR_CQ_TYPE_KERNEL;
1788 
1789                 rc = ecore_chain_alloc(&dev->ha->cdev,
1790                            ECORE_CHAIN_USE_TO_CONSUME,
1791                            ECORE_CHAIN_MODE_PBL,
1792                            ECORE_CHAIN_CNT_TYPE_U32,
1793                            chain_entries,
1794                            sizeof(union roce_cqe),
1795                            &cq->pbl, NULL);
1796 
1797 		if (rc)
1798 			goto err1;
1799 
1800 		page_cnt = ecore_chain_get_page_cnt(&cq->pbl);
1801 		pbl_ptr = ecore_chain_get_pbl_phys(&cq->pbl);
1802 		cq->ibcq.cqe = cq->pbl.capacity;
1803 	}
1804 
1805         params.cq_handle_hi = upper_32_bits((uintptr_t)cq);
1806         params.cq_handle_lo = lower_32_bits((uintptr_t)cq);
1807         params.cnq_id = vector;
1808         params.cq_size = chain_entries - 1;
1809         params.pbl_num_pages = page_cnt;
1810         params.pbl_ptr = pbl_ptr;
1811         params.pbl_two_level = 0;
1812 
1813 	if (udata) {
1814         	params.dpi = ctx->dpi;
1815 	} else {
1816         	params.dpi = dev->dpi;
1817 	}
1818 
1819 	rc = ecore_rdma_create_cq(dev->rdma_ctx, &params, &icid);
1820 	if (rc)
1821 		goto err2;
1822 
1823 	cq->icid = icid;
1824 	cq->sig = QLNXR_CQ_MAGIC_NUMBER;
1825 	spin_lock_init(&cq->cq_lock);
1826 
1827 	if (udata) {
1828 		rc = qlnxr_copy_cq_uresp(dev, cq, udata);
1829 		if (rc)
1830 			goto err3;
1831 	} else {
1832 		/* Generate doorbell address.
1833 		 * Configure bits 3-9 with DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT.
1834 		 * TODO: consider moving to device scope as it is a function of
1835 		 *       the device.
1836 		 * TODO: add ifdef if plan to support 16 bit.
1837 		 */
1838 		cq->db_addr = dev->db_addr +
1839 			DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT);
1840 		cq->db.data.icid = cq->icid;
1841 		cq->db.data.params = DB_AGG_CMD_SET <<
1842 				     RDMA_PWM_VAL32_DATA_AGG_CMD_SHIFT;
1843 
1844 		/* point to the very last element, passing it we will toggle */
1845 		cq->toggle_cqe = ecore_chain_get_last_elem(&cq->pbl);
1846 		cq->pbl_toggle = RDMA_RESIZE_CQ_RAMROD_DATA_TOGGLE_BIT_MASK;
1847 
1848 		/* must be different from pbl_toggle */
1849 		cq->latest_cqe = NULL;
1850 		consume_cqe(cq);
1851 		cq->cq_cons = ecore_chain_get_cons_idx_u32(&cq->pbl);
1852 	}
1853 
1854 	QL_DPRINT12(ha, "exit icid = 0x%0x, addr = %p,"
1855 		" number of entries = 0x%x\n",
1856 		cq->icid, cq, params.cq_size);
1857 	QL_DPRINT12(ha,"cq_addr = %p\n", cq);
1858 	return (0);
1859 
1860 err3:
1861 	destroy_iparams.icid = cq->icid;
1862 	ecore_rdma_destroy_cq(dev->rdma_ctx, &destroy_iparams, &destroy_oparams);
1863 err2:
1864 	if (udata)
1865 		qlnxr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
1866 	else
1867 		ecore_chain_free(&dev->ha->cdev, &cq->pbl);
1868 err1:
1869 	if (udata)
1870 		ib_umem_release(cq->q.umem);
1871 err0:
1872 	QL_DPRINT12(ha, "exit error\n");
1873 
1874 	return (-EINVAL);
1875 }
1876 
1877 int qlnxr_resize_cq(struct ib_cq *ibcq, int new_cnt, struct ib_udata *udata)
1878 {
1879 	int			status = 0;
1880 	struct qlnxr_dev	*dev = get_qlnxr_dev((ibcq->device));
1881 	qlnx_host_t		*ha;
1882 
1883 	ha = dev->ha;
1884 
1885 	QL_DPRINT12(ha, "enter/exit\n");
1886 
1887 	return status;
1888 }
1889 
1890 void
1891 qlnxr_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1892 {
1893 	struct qlnxr_dev			*dev = get_qlnxr_dev((ibcq->device));
1894 	struct ecore_rdma_destroy_cq_out_params oparams;
1895 	struct ecore_rdma_destroy_cq_in_params	iparams;
1896 	struct qlnxr_cq				*cq = get_qlnxr_cq(ibcq);
1897 	int					rc = 0;
1898 	qlnx_host_t				*ha;
1899 
1900 	ha = dev->ha;
1901 
1902 	QL_DPRINT12(ha, "enter cq_id = %d\n", cq->icid);
1903 
1904 	cq->destroyed = 1;
1905 
1906 	/* TODO: Syncronize irq of the CNQ the CQ belongs to for validation
1907 	 * that all completions with notification are dealt with. The rest
1908 	 * of the completions are not interesting
1909 	 */
1910 
1911 	/* GSIs CQs are handled by driver, so they don't exist in the FW */
1912 
1913 	if (cq->cq_type != QLNXR_CQ_TYPE_GSI) {
1914 		iparams.icid = cq->icid;
1915 
1916 		rc = ecore_rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams);
1917 
1918 		if (rc) {
1919 			QL_DPRINT12(ha, "ecore_rdma_destroy_cq failed cq_id = %d\n",
1920 				cq->icid);
1921 			return;
1922 		}
1923 
1924 		QL_DPRINT12(ha, "free cq->pbl cq_id = %d\n", cq->icid);
1925 		ecore_chain_free(&dev->ha->cdev, &cq->pbl);
1926 	}
1927 
1928 	if (udata) {
1929 		qlnxr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
1930 		ib_umem_release(cq->q.umem);
1931 	}
1932 
1933 	cq->sig = ~cq->sig;
1934 
1935 	QL_DPRINT12(ha, "exit cq_id = %d\n", cq->icid);
1936 }
1937 
1938 static int
1939 qlnxr_check_qp_attrs(struct ib_pd *ibpd,
1940 	struct qlnxr_dev *dev,
1941 	struct ib_qp_init_attr *attrs,
1942 	struct ib_udata *udata)
1943 {
1944 	struct ecore_rdma_device	*qattr;
1945 	qlnx_host_t			*ha;
1946 
1947 	qattr = ecore_rdma_query_device(dev->rdma_ctx);
1948 	ha = dev->ha;
1949 
1950 	QL_DPRINT12(ha, "enter\n");
1951 
1952 	QL_DPRINT12(ha, "attrs->sq_sig_type = %d\n", attrs->sq_sig_type);
1953 	QL_DPRINT12(ha, "attrs->qp_type = %d\n", attrs->qp_type);
1954 	QL_DPRINT12(ha, "attrs->create_flags = %d\n", attrs->create_flags);
1955 
1956 	QL_DPRINT12(ha, "attrs->port_num = %d\n", attrs->port_num);
1957 	QL_DPRINT12(ha, "attrs->cap.max_send_wr = 0x%x\n", attrs->cap.max_send_wr);
1958 	QL_DPRINT12(ha, "attrs->cap.max_recv_wr = 0x%x\n", attrs->cap.max_recv_wr);
1959 	QL_DPRINT12(ha, "attrs->cap.max_send_sge = 0x%x\n", attrs->cap.max_send_sge);
1960 	QL_DPRINT12(ha, "attrs->cap.max_recv_sge = 0x%x\n", attrs->cap.max_recv_sge);
1961 	QL_DPRINT12(ha, "attrs->cap.max_inline_data = 0x%x\n",
1962 		attrs->cap.max_inline_data);
1963 
1964 	QL_DPRINT12(ha, "\n\nqattr->vendor_id = 0x%x\n", qattr->vendor_id);
1965 	QL_DPRINT12(ha, "qattr->vendor_part_id = 0x%x\n", qattr->vendor_part_id);
1966 	QL_DPRINT12(ha, "qattr->hw_ver = 0x%x\n", qattr->hw_ver);
1967 	QL_DPRINT12(ha, "qattr->fw_ver = %p\n", (void *)qattr->fw_ver);
1968 	QL_DPRINT12(ha, "qattr->node_guid = %p\n", (void *)qattr->node_guid);
1969 	QL_DPRINT12(ha, "qattr->sys_image_guid = %p\n",
1970 		(void *)qattr->sys_image_guid);
1971 	QL_DPRINT12(ha, "qattr->max_cnq = 0x%x\n", qattr->max_cnq);
1972 	QL_DPRINT12(ha, "qattr->max_sge = 0x%x\n", qattr->max_sge);
1973 	QL_DPRINT12(ha, "qattr->max_srq_sge = 0x%x\n", qattr->max_srq_sge);
1974 	QL_DPRINT12(ha, "qattr->max_inline = 0x%x\n", qattr->max_inline);
1975 	QL_DPRINT12(ha, "qattr->max_wqe = 0x%x\n", qattr->max_wqe);
1976 	QL_DPRINT12(ha, "qattr->max_srq_wqe = 0x%x\n", qattr->max_srq_wqe);
1977 	QL_DPRINT12(ha, "qattr->max_qp_resp_rd_atomic_resc = 0x%x\n",
1978 		qattr->max_qp_resp_rd_atomic_resc);
1979 	QL_DPRINT12(ha, "qattr->max_qp_req_rd_atomic_resc = 0x%x\n",
1980 		qattr->max_qp_req_rd_atomic_resc);
1981 	QL_DPRINT12(ha, "qattr->max_dev_resp_rd_atomic_resc = 0x%x\n",
1982 		qattr->max_dev_resp_rd_atomic_resc);
1983 	QL_DPRINT12(ha, "qattr->max_cq = 0x%x\n", qattr->max_cq);
1984 	QL_DPRINT12(ha, "qattr->max_qp = 0x%x\n", qattr->max_qp);
1985 	QL_DPRINT12(ha, "qattr->max_srq = 0x%x\n", qattr->max_srq);
1986 	QL_DPRINT12(ha, "qattr->max_mr = 0x%x\n", qattr->max_mr);
1987 	QL_DPRINT12(ha, "qattr->max_mr_size = %p\n", (void *)qattr->max_mr_size);
1988 	QL_DPRINT12(ha, "qattr->max_cqe = 0x%x\n", qattr->max_cqe);
1989 	QL_DPRINT12(ha, "qattr->max_mw = 0x%x\n", qattr->max_mw);
1990 	QL_DPRINT12(ha, "qattr->max_fmr = 0x%x\n", qattr->max_fmr);
1991 	QL_DPRINT12(ha, "qattr->max_mr_mw_fmr_pbl = 0x%x\n",
1992 		qattr->max_mr_mw_fmr_pbl);
1993 	QL_DPRINT12(ha, "qattr->max_mr_mw_fmr_size = %p\n",
1994 		(void *)qattr->max_mr_mw_fmr_size);
1995 	QL_DPRINT12(ha, "qattr->max_pd = 0x%x\n", qattr->max_pd);
1996 	QL_DPRINT12(ha, "qattr->max_ah = 0x%x\n", qattr->max_ah);
1997 	QL_DPRINT12(ha, "qattr->max_pkey = 0x%x\n", qattr->max_pkey);
1998 	QL_DPRINT12(ha, "qattr->max_srq_wr = 0x%x\n", qattr->max_srq_wr);
1999 	QL_DPRINT12(ha, "qattr->max_stats_queues = 0x%x\n",
2000 		qattr->max_stats_queues);
2001 	//QL_DPRINT12(ha, "qattr->dev_caps = 0x%x\n", qattr->dev_caps);
2002 	QL_DPRINT12(ha, "qattr->page_size_caps = %p\n",
2003 		(void *)qattr->page_size_caps);
2004 	QL_DPRINT12(ha, "qattr->dev_ack_delay = 0x%x\n", qattr->dev_ack_delay);
2005 	QL_DPRINT12(ha, "qattr->reserved_lkey = 0x%x\n", qattr->reserved_lkey);
2006 	QL_DPRINT12(ha, "qattr->bad_pkey_counter = 0x%x\n",
2007 		qattr->bad_pkey_counter);
2008 
2009 	if ((attrs->qp_type == IB_QPT_GSI) && udata) {
2010 		QL_DPRINT12(ha, "unexpected udata when creating GSI QP\n");
2011 		return -EINVAL;
2012 	}
2013 
2014 	if (udata && !(ibpd->uobject && ibpd->uobject->context)) {
2015 		QL_DPRINT12(ha, "called from user without context\n");
2016 		return -EINVAL;
2017 	}
2018 
2019 	/* QP0... attrs->qp_type == IB_QPT_GSI */
2020 	if (attrs->qp_type != IB_QPT_RC && attrs->qp_type != IB_QPT_GSI) {
2021 		QL_DPRINT12(ha, "unsupported qp type=0x%x requested\n",
2022 			   attrs->qp_type);
2023 		return -EINVAL;
2024 	}
2025 	if (attrs->qp_type == IB_QPT_GSI && attrs->srq) {
2026 		QL_DPRINT12(ha, "cannot create GSI qp with SRQ\n");
2027 		return -EINVAL;
2028 	}
2029 	/* Skip the check for QP1 to support CM size of 128 */
2030 	if (attrs->cap.max_send_wr > qattr->max_wqe) {
2031 		QL_DPRINT12(ha, "cannot create a SQ with %d elements "
2032 			" (max_send_wr=0x%x)\n",
2033 			attrs->cap.max_send_wr, qattr->max_wqe);
2034 		return -EINVAL;
2035 	}
2036 	if (!attrs->srq && (attrs->cap.max_recv_wr > qattr->max_wqe)) {
2037 		QL_DPRINT12(ha, "cannot create a RQ with %d elements"
2038 			" (max_recv_wr=0x%x)\n",
2039 			attrs->cap.max_recv_wr, qattr->max_wqe);
2040 		return -EINVAL;
2041 	}
2042 	if (attrs->cap.max_inline_data > qattr->max_inline) {
2043 		QL_DPRINT12(ha,
2044 			"unsupported inline data size=0x%x "
2045 			"requested (max_inline=0x%x)\n",
2046 			attrs->cap.max_inline_data, qattr->max_inline);
2047 		return -EINVAL;
2048 	}
2049 	if (attrs->cap.max_send_sge > qattr->max_sge) {
2050 		QL_DPRINT12(ha,
2051 			"unsupported send_sge=0x%x "
2052 			"requested (max_send_sge=0x%x)\n",
2053 			attrs->cap.max_send_sge, qattr->max_sge);
2054 		return -EINVAL;
2055 	}
2056 	if (attrs->cap.max_recv_sge > qattr->max_sge) {
2057 		QL_DPRINT12(ha,
2058 			"unsupported recv_sge=0x%x requested "
2059 			" (max_recv_sge=0x%x)\n",
2060 			attrs->cap.max_recv_sge, qattr->max_sge);
2061 		return -EINVAL;
2062 	}
2063 	/* unprivileged user space cannot create special QP */
2064 	if (ibpd->uobject && attrs->qp_type == IB_QPT_GSI) {
2065 		QL_DPRINT12(ha,
2066 			"userspace can't create special QPs of type=0x%x\n",
2067 			attrs->qp_type);
2068 		return -EINVAL;
2069 	}
2070 	/* allow creating only one GSI type of QP */
2071 	if (attrs->qp_type == IB_QPT_GSI && dev->gsi_qp_created) {
2072 		QL_DPRINT12(ha,
2073 			"create qp: GSI special QPs already created.\n");
2074 		return -EINVAL;
2075 	}
2076 
2077 	/* verify consumer QPs are not trying to use GSI QP's CQ */
2078 	if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created)) {
2079 		struct qlnxr_cq *send_cq = get_qlnxr_cq(attrs->send_cq);
2080 		struct qlnxr_cq *recv_cq = get_qlnxr_cq(attrs->recv_cq);
2081 
2082 		if ((send_cq->cq_type == QLNXR_CQ_TYPE_GSI) ||
2083 		    (recv_cq->cq_type == QLNXR_CQ_TYPE_GSI)) {
2084 			QL_DPRINT11(ha, "consumer QP cannot use GSI CQs.\n");
2085 			return -EINVAL;
2086 		}
2087 	}
2088 	QL_DPRINT12(ha, "exit\n");
2089 	return 0;
2090 }
2091 
2092 static int
2093 qlnxr_copy_srq_uresp(struct qlnxr_dev *dev,
2094 	struct qlnxr_srq *srq,
2095 	struct ib_udata *udata)
2096 {
2097 	struct qlnxr_create_srq_uresp	uresp;
2098 	qlnx_host_t			*ha;
2099 	int				rc;
2100 
2101 	ha = dev->ha;
2102 
2103 	QL_DPRINT12(ha, "enter\n");
2104 
2105 	memset(&uresp, 0, sizeof(uresp));
2106 
2107 	uresp.srq_id = srq->srq_id;
2108 
2109 	rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
2110 
2111 	QL_DPRINT12(ha, "exit [%d]\n", rc);
2112 	return rc;
2113 }
2114 
2115 static void
2116 qlnxr_copy_rq_uresp(struct qlnxr_dev *dev,
2117 	struct qlnxr_create_qp_uresp *uresp,
2118 	struct qlnxr_qp *qp)
2119 {
2120 	qlnx_host_t	*ha;
2121 
2122 	ha = dev->ha;
2123 
2124 	/* Return if QP is associated with SRQ instead of RQ */
2125 	QL_DPRINT12(ha, "enter qp->srq = %p\n", qp->srq);
2126 
2127 	if (qp->srq)
2128 		return;
2129 
2130 	/* iWARP requires two doorbells per RQ. */
2131 	if (QLNX_IS_IWARP(dev)) {
2132 		uresp->rq_db_offset =
2133 			DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_IWARP_RQ_PROD);
2134 		uresp->rq_db2_offset =
2135 			DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_FLAGS);
2136 
2137 		QL_DPRINT12(ha, "uresp->rq_db_offset = 0x%x "
2138 			"uresp->rq_db2_offset = 0x%x\n",
2139 			uresp->rq_db_offset, uresp->rq_db2_offset);
2140 	} else {
2141 		uresp->rq_db_offset =
2142 			DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
2143 	}
2144 	uresp->rq_icid = qp->icid;
2145 
2146 	QL_DPRINT12(ha, "exit\n");
2147 	return;
2148 }
2149 
2150 static void
2151 qlnxr_copy_sq_uresp(struct qlnxr_dev *dev,
2152 	struct qlnxr_create_qp_uresp *uresp,
2153 	struct qlnxr_qp *qp)
2154 {
2155 	qlnx_host_t	*ha;
2156 
2157 	ha = dev->ha;
2158 
2159 	QL_DPRINT12(ha, "enter\n");
2160 
2161 	uresp->sq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
2162 
2163 	/* iWARP uses the same cid for rq and sq*/
2164 	if (QLNX_IS_IWARP(dev)) {
2165 		uresp->sq_icid = qp->icid;
2166 		QL_DPRINT12(ha, "uresp->sq_icid = 0x%x\n", uresp->sq_icid);
2167 	} else
2168 		uresp->sq_icid = qp->icid + 1;
2169 
2170 	QL_DPRINT12(ha, "exit\n");
2171 	return;
2172 }
2173 
2174 static int
2175 qlnxr_copy_qp_uresp(struct qlnxr_dev *dev,
2176 	struct qlnxr_qp *qp,
2177 	struct ib_udata *udata)
2178 {
2179 	int				rc;
2180 	struct qlnxr_create_qp_uresp	uresp;
2181 	qlnx_host_t			*ha;
2182 
2183 	ha = dev->ha;
2184 
2185 	QL_DPRINT12(ha, "enter qp->icid =0x%x\n", qp->icid);
2186 
2187 	memset(&uresp, 0, sizeof(uresp));
2188 	qlnxr_copy_sq_uresp(dev, &uresp, qp);
2189 	qlnxr_copy_rq_uresp(dev, &uresp, qp);
2190 
2191 	uresp.atomic_supported = dev->atomic_cap != IB_ATOMIC_NONE;
2192 	uresp.qp_id = qp->qp_id;
2193 
2194 	rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
2195 
2196 	QL_DPRINT12(ha, "exit [%d]\n", rc);
2197 	return rc;
2198 }
2199 
2200 static void
2201 qlnxr_set_common_qp_params(struct qlnxr_dev *dev,
2202 	struct qlnxr_qp *qp,
2203 	struct qlnxr_pd *pd,
2204 	struct ib_qp_init_attr *attrs)
2205 {
2206 	qlnx_host_t			*ha;
2207 
2208 	ha = dev->ha;
2209 
2210 	QL_DPRINT12(ha, "enter\n");
2211 
2212 	spin_lock_init(&qp->q_lock);
2213 
2214 	atomic_set(&qp->refcnt, 1);
2215 	qp->pd = pd;
2216 	qp->sig = QLNXR_QP_MAGIC_NUMBER;
2217 	qp->qp_type = attrs->qp_type;
2218 	qp->max_inline_data = ROCE_REQ_MAX_INLINE_DATA_SIZE;
2219 	qp->sq.max_sges = attrs->cap.max_send_sge;
2220 	qp->state = ECORE_ROCE_QP_STATE_RESET;
2221 	qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
2222 	qp->sq_cq = get_qlnxr_cq(attrs->send_cq);
2223 	qp->rq_cq = get_qlnxr_cq(attrs->recv_cq);
2224 	qp->dev = dev;
2225 
2226 	if (!attrs->srq) {
2227 		/* QP is associated with RQ instead of SRQ */
2228 		qp->rq.max_sges = attrs->cap.max_recv_sge;
2229 		QL_DPRINT12(ha, "RQ params:\trq_max_sges = %d, rq_cq_id = %d\n",
2230 			qp->rq.max_sges, qp->rq_cq->icid);
2231 	} else {
2232 		qp->srq = get_qlnxr_srq(attrs->srq);
2233 	}
2234 
2235 	QL_DPRINT12(ha,
2236 		"QP params:\tpd = %d, qp_type = %d, max_inline_data = %d,"
2237 		" state = %d, signaled = %d, use_srq=%d\n",
2238 		pd->pd_id, qp->qp_type, qp->max_inline_data,
2239 		qp->state, qp->signaled, ((attrs->srq) ? 1 : 0));
2240 	QL_DPRINT12(ha, "SQ params:\tsq_max_sges = %d, sq_cq_id = %d\n",
2241 		qp->sq.max_sges, qp->sq_cq->icid);
2242 	return;
2243 }
2244 
2245 static int
2246 qlnxr_check_srq_params(struct qlnxr_dev *dev,
2247 	struct ib_srq_init_attr *attrs)
2248 {
2249 	struct ecore_rdma_device *qattr;
2250 	qlnx_host_t		*ha;
2251 
2252 	ha = dev->ha;
2253 	qattr = ecore_rdma_query_device(dev->rdma_ctx);
2254 
2255 	QL_DPRINT12(ha, "enter\n");
2256 
2257 	if (attrs->attr.max_wr > qattr->max_srq_wqe) {
2258 		QL_DPRINT12(ha, "unsupported srq_wr=0x%x"
2259 			" requested (max_srq_wr=0x%x)\n",
2260 			attrs->attr.max_wr, qattr->max_srq_wr);
2261 		return -EINVAL;
2262 	}
2263 
2264 	if (attrs->attr.max_sge > qattr->max_sge) {
2265 		QL_DPRINT12(ha,
2266 			"unsupported sge=0x%x requested (max_srq_sge=0x%x)\n",
2267 			attrs->attr.max_sge, qattr->max_sge);
2268 		return -EINVAL;
2269 	}
2270 
2271 	if (attrs->attr.srq_limit > attrs->attr.max_wr) {
2272 		QL_DPRINT12(ha,
2273 		       "unsupported srq_limit=0x%x requested"
2274 			" (max_srq_limit=0x%x)\n",
2275 			attrs->attr.srq_limit, attrs->attr.srq_limit);
2276 		return -EINVAL;
2277 	}
2278 
2279 	QL_DPRINT12(ha, "exit\n");
2280 	return 0;
2281 }
2282 
2283 static void
2284 qlnxr_free_srq_user_params(struct qlnxr_srq *srq)
2285 {
2286 	struct qlnxr_dev	*dev = srq->dev;
2287 	qlnx_host_t		*ha;
2288 
2289 	ha = dev->ha;
2290 
2291 	QL_DPRINT12(ha, "enter\n");
2292 
2293 	qlnxr_free_pbl(srq->dev, &srq->usrq.pbl_info, srq->usrq.pbl_tbl);
2294 	ib_umem_release(srq->usrq.umem);
2295 	ib_umem_release(srq->prod_umem);
2296 
2297 	QL_DPRINT12(ha, "exit\n");
2298 	return;
2299 }
2300 
2301 static void
2302 qlnxr_free_srq_kernel_params(struct qlnxr_srq *srq)
2303 {
2304 	struct qlnxr_srq_hwq_info *hw_srq  = &srq->hw_srq;
2305 	struct qlnxr_dev	*dev = srq->dev;
2306 	qlnx_host_t		*ha;
2307 
2308 	ha = dev->ha;
2309 
2310 	QL_DPRINT12(ha, "enter\n");
2311 
2312 	ecore_chain_free(dev->cdev, &hw_srq->pbl);
2313 
2314 	qlnx_dma_free_coherent(&dev->cdev,
2315 		hw_srq->virt_prod_pair_addr,
2316 		hw_srq->phy_prod_pair_addr,
2317 		sizeof(struct rdma_srq_producers));
2318 
2319 	QL_DPRINT12(ha, "exit\n");
2320 
2321 	return;
2322 }
2323 
2324 static int
2325 qlnxr_init_srq_user_params(struct ib_ucontext *ib_ctx,
2326 	struct qlnxr_srq *srq,
2327 	struct qlnxr_create_srq_ureq *ureq,
2328 	int access, int dmasync)
2329 {
2330 	struct scatterlist	*sg;
2331 	int			rc;
2332 	struct qlnxr_dev	*dev = srq->dev;
2333 	qlnx_host_t		*ha;
2334 
2335 	ha = dev->ha;
2336 
2337 	QL_DPRINT12(ha, "enter\n");
2338 
2339 	rc = qlnxr_init_user_queue(ib_ctx, srq->dev, &srq->usrq, ureq->srq_addr,
2340 				  ureq->srq_len, access, dmasync, 1);
2341 	if (rc)
2342 		return rc;
2343 
2344 	srq->prod_umem = ib_umem_get(ib_ctx, ureq->prod_pair_addr,
2345 				     sizeof(struct rdma_srq_producers),
2346 				     access, dmasync);
2347 	if (IS_ERR(srq->prod_umem)) {
2348 		qlnxr_free_pbl(srq->dev, &srq->usrq.pbl_info, srq->usrq.pbl_tbl);
2349 		ib_umem_release(srq->usrq.umem);
2350 
2351 		QL_DPRINT12(ha, "ib_umem_get failed for producer [%p]\n",
2352 			PTR_ERR(srq->prod_umem));
2353 
2354 		return PTR_ERR(srq->prod_umem);
2355 	}
2356 
2357 	sg = srq->prod_umem->sg_head.sgl;
2358 	srq->hw_srq.phy_prod_pair_addr = sg_dma_address(sg);
2359 
2360 	QL_DPRINT12(ha, "exit\n");
2361 	return 0;
2362 }
2363 
2364 static int
2365 qlnxr_alloc_srq_kernel_params(struct qlnxr_srq *srq,
2366 	struct qlnxr_dev *dev,
2367 	struct ib_srq_init_attr *init_attr)
2368 {
2369 	struct qlnxr_srq_hwq_info	*hw_srq  = &srq->hw_srq;
2370 	dma_addr_t			phy_prod_pair_addr;
2371 	u32				num_elems, max_wr;
2372 	void				*va;
2373 	int				rc;
2374 	qlnx_host_t			*ha;
2375 
2376 	ha = dev->ha;
2377 
2378 	QL_DPRINT12(ha, "enter\n");
2379 
2380 	va = qlnx_dma_alloc_coherent(&dev->cdev,
2381 			&phy_prod_pair_addr,
2382 			sizeof(struct rdma_srq_producers));
2383 	if (!va) {
2384 		QL_DPRINT11(ha, "qlnx_dma_alloc_coherent failed for produceer\n");
2385 		return -ENOMEM;
2386 	}
2387 
2388 	hw_srq->phy_prod_pair_addr = phy_prod_pair_addr;
2389 	hw_srq->virt_prod_pair_addr = va;
2390 
2391 	max_wr = init_attr->attr.max_wr;
2392 
2393 	num_elems = max_wr * RDMA_MAX_SRQ_WQE_SIZE;
2394 
2395         rc = ecore_chain_alloc(dev->cdev,
2396                    ECORE_CHAIN_USE_TO_CONSUME_PRODUCE,
2397                    ECORE_CHAIN_MODE_PBL,
2398                    ECORE_CHAIN_CNT_TYPE_U32,
2399                    num_elems,
2400                    ECORE_RDMA_SRQ_WQE_ELEM_SIZE,
2401                    &hw_srq->pbl, NULL);
2402 
2403 	if (rc) {
2404 		QL_DPRINT11(ha, "ecore_chain_alloc failed [%d]\n", rc);
2405 		goto err0;
2406 	}
2407 
2408 	hw_srq->max_wr = max_wr;
2409 	hw_srq->num_elems = num_elems;
2410 	hw_srq->max_sges = RDMA_MAX_SGE_PER_SRQ;
2411 
2412 	QL_DPRINT12(ha, "exit\n");
2413 	return 0;
2414 
2415 err0:
2416 	qlnx_dma_free_coherent(&dev->cdev, va, phy_prod_pair_addr,
2417 		sizeof(struct rdma_srq_producers));
2418 
2419 	QL_DPRINT12(ha, "exit [%d]\n", rc);
2420 	return rc;
2421 }
2422 
2423 static inline void
2424 qlnxr_init_common_qp_in_params(struct qlnxr_dev *dev,
2425 	struct qlnxr_pd *pd,
2426 	struct qlnxr_qp *qp,
2427 	struct ib_qp_init_attr *attrs,
2428 	bool fmr_and_reserved_lkey,
2429 	struct ecore_rdma_create_qp_in_params *params)
2430 {
2431 	qlnx_host_t	*ha;
2432 
2433 	ha = dev->ha;
2434 
2435 	QL_DPRINT12(ha, "enter\n");
2436 
2437 	/* QP handle to be written in an async event */
2438 	params->qp_handle_async_lo = lower_32_bits((uintptr_t)qp);
2439 	params->qp_handle_async_hi = upper_32_bits((uintptr_t)qp);
2440 
2441 	params->signal_all = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR);
2442 	params->fmr_and_reserved_lkey = fmr_and_reserved_lkey;
2443 	params->pd = pd->pd_id;
2444 	params->dpi = pd->uctx ? pd->uctx->dpi : dev->dpi;
2445 	params->sq_cq_id = get_qlnxr_cq(attrs->send_cq)->icid;
2446 	params->stats_queue = 0;
2447 
2448 	params->rq_cq_id = get_qlnxr_cq(attrs->recv_cq)->icid;
2449 
2450 	if (qp->srq) {
2451 		/* QP is associated with SRQ instead of RQ */
2452 		params->srq_id = qp->srq->srq_id;
2453 		params->use_srq = true;
2454 		QL_DPRINT11(ha, "exit srq_id = 0x%x use_srq = 0x%x\n",
2455 			params->srq_id, params->use_srq);
2456 		return;
2457 	}
2458 
2459 	params->srq_id = 0;
2460 	params->use_srq = false;
2461 
2462 	QL_DPRINT12(ha, "exit\n");
2463 	return;
2464 }
2465 
2466 static inline void
2467 qlnxr_qp_user_print( struct qlnxr_dev *dev,
2468 	struct qlnxr_qp *qp)
2469 {
2470 	QL_DPRINT12((dev->ha), "qp=%p. sq_addr=0x%llx, sq_len=%zd, "
2471 		"rq_addr=0x%llx, rq_len=%zd\n",
2472 		qp, qp->usq.buf_addr, qp->usq.buf_len, qp->urq.buf_addr,
2473 		qp->urq.buf_len);
2474 	return;
2475 }
2476 
2477 static int
2478 qlnxr_idr_add(struct qlnxr_dev *dev, void *ptr, u32 id)
2479 {
2480 	u32		newid;
2481 	int		rc;
2482 	qlnx_host_t	*ha;
2483 
2484 	ha = dev->ha;
2485 
2486 	QL_DPRINT12(ha, "enter\n");
2487 
2488 	if (!QLNX_IS_IWARP(dev))
2489 		return 0;
2490 
2491 	do {
2492 		if (!idr_pre_get(&dev->qpidr, GFP_KERNEL)) {
2493 			QL_DPRINT11(ha, "idr_pre_get failed\n");
2494 			return -ENOMEM;
2495 		}
2496 
2497 		mtx_lock(&dev->idr_lock);
2498 
2499 		rc = idr_get_new_above(&dev->qpidr, ptr, id, &newid);
2500 
2501 		mtx_unlock(&dev->idr_lock);
2502 
2503 	} while (rc == -EAGAIN);
2504 
2505 	QL_DPRINT12(ha, "exit [%d]\n", rc);
2506 
2507 	return rc;
2508 }
2509 
2510 static void
2511 qlnxr_idr_remove(struct qlnxr_dev *dev, u32 id)
2512 {
2513 	qlnx_host_t	*ha;
2514 
2515 	ha = dev->ha;
2516 
2517 	QL_DPRINT12(ha, "enter\n");
2518 
2519 	if (!QLNX_IS_IWARP(dev))
2520 		return;
2521 
2522 	mtx_lock(&dev->idr_lock);
2523 	idr_remove(&dev->qpidr, id);
2524 	mtx_unlock(&dev->idr_lock);
2525 
2526 	QL_DPRINT12(ha, "exit \n");
2527 
2528 	return;
2529 }
2530 
2531 static inline void
2532 qlnxr_iwarp_populate_user_qp(struct qlnxr_dev *dev,
2533 	struct qlnxr_qp *qp,
2534 	struct ecore_rdma_create_qp_out_params *out_params)
2535 {
2536 	qlnx_host_t	*ha;
2537 
2538 	ha = dev->ha;
2539 
2540 	QL_DPRINT12(ha, "enter\n");
2541 
2542 	qp->usq.pbl_tbl->va = out_params->sq_pbl_virt;
2543 	qp->usq.pbl_tbl->pa = out_params->sq_pbl_phys;
2544 
2545 	qlnxr_populate_pbls(dev, qp->usq.umem, qp->usq.pbl_tbl,
2546 			   &qp->usq.pbl_info);
2547 
2548 	if (qp->srq) {
2549 		QL_DPRINT11(ha, "qp->srq = %p\n", qp->srq);
2550 		return;
2551 	}
2552 
2553 	qp->urq.pbl_tbl->va = out_params->rq_pbl_virt;
2554 	qp->urq.pbl_tbl->pa = out_params->rq_pbl_phys;
2555 
2556 	qlnxr_populate_pbls(dev, qp->urq.umem, qp->urq.pbl_tbl,
2557 			   &qp->urq.pbl_info);
2558 
2559 	QL_DPRINT12(ha, "exit\n");
2560 	return;
2561 }
2562 
2563 static int
2564 qlnxr_create_user_qp(struct qlnxr_dev *dev,
2565 	struct qlnxr_qp *qp,
2566 	struct ib_pd *ibpd,
2567 	struct ib_udata *udata,
2568 	struct ib_qp_init_attr *attrs)
2569 {
2570 	struct ecore_rdma_destroy_qp_out_params d_out_params;
2571 	struct ecore_rdma_create_qp_in_params in_params;
2572 	struct ecore_rdma_create_qp_out_params out_params;
2573 	struct qlnxr_pd *pd = get_qlnxr_pd(ibpd);
2574 	struct ib_ucontext *ib_ctx = NULL;
2575 	struct qlnxr_create_qp_ureq ureq;
2576 	int alloc_and_init = QLNX_IS_ROCE(dev);
2577 	int rc = -EINVAL;
2578 	qlnx_host_t	*ha;
2579 
2580 	ha = dev->ha;
2581 
2582 	QL_DPRINT12(ha, "enter\n");
2583 
2584 	ib_ctx = ibpd->uobject->context;
2585 
2586 	memset(&ureq, 0, sizeof(ureq));
2587 	rc = ib_copy_from_udata(&ureq, udata, sizeof(ureq));
2588 
2589 	if (rc) {
2590 		QL_DPRINT11(ha, "ib_copy_from_udata failed [%d]\n", rc);
2591 		return rc;
2592 	}
2593 
2594 	/* SQ - read access only (0), dma sync not required (0) */
2595 	rc = qlnxr_init_user_queue(ib_ctx, dev, &qp->usq, ureq.sq_addr,
2596 				  ureq.sq_len, 0, 0,
2597 				  alloc_and_init);
2598 	if (rc) {
2599 		QL_DPRINT11(ha, "qlnxr_init_user_queue failed [%d]\n", rc);
2600 		return rc;
2601 	}
2602 
2603 	if (!qp->srq) {
2604 		/* RQ - read access only (0), dma sync not required (0) */
2605 		rc = qlnxr_init_user_queue(ib_ctx, dev, &qp->urq, ureq.rq_addr,
2606 					  ureq.rq_len, 0, 0,
2607 					  alloc_and_init);
2608 
2609 		if (rc) {
2610 			QL_DPRINT11(ha, "qlnxr_init_user_queue failed [%d]\n", rc);
2611 			return rc;
2612 		}
2613 	}
2614 
2615 	memset(&in_params, 0, sizeof(in_params));
2616 	qlnxr_init_common_qp_in_params(dev, pd, qp, attrs, false, &in_params);
2617 	in_params.qp_handle_lo = ureq.qp_handle_lo;
2618 	in_params.qp_handle_hi = ureq.qp_handle_hi;
2619 	in_params.sq_num_pages = qp->usq.pbl_info.num_pbes;
2620 	in_params.sq_pbl_ptr = qp->usq.pbl_tbl->pa;
2621 
2622 	if (!qp->srq) {
2623 		in_params.rq_num_pages = qp->urq.pbl_info.num_pbes;
2624 		in_params.rq_pbl_ptr = qp->urq.pbl_tbl->pa;
2625 	}
2626 
2627 	qp->ecore_qp = ecore_rdma_create_qp(dev->rdma_ctx, &in_params, &out_params);
2628 
2629 	if (!qp->ecore_qp) {
2630 		rc = -ENOMEM;
2631 		QL_DPRINT11(ha, "ecore_rdma_create_qp failed\n");
2632 		goto err1;
2633 	}
2634 
2635 	if (QLNX_IS_IWARP(dev))
2636 		qlnxr_iwarp_populate_user_qp(dev, qp, &out_params);
2637 
2638 	qp->qp_id = out_params.qp_id;
2639 	qp->icid = out_params.icid;
2640 
2641 	rc = qlnxr_copy_qp_uresp(dev, qp, udata);
2642 
2643 	if (rc) {
2644 		QL_DPRINT11(ha, "qlnxr_copy_qp_uresp failed\n");
2645 		goto err;
2646 	}
2647 
2648 	qlnxr_qp_user_print(dev, qp);
2649 
2650 	QL_DPRINT12(ha, "exit\n");
2651 	return 0;
2652 err:
2653 	rc = ecore_rdma_destroy_qp(dev->rdma_ctx, qp->ecore_qp, &d_out_params);
2654 
2655 	if (rc)
2656 		QL_DPRINT12(ha, "fatal fault\n");
2657 
2658 err1:
2659 	qlnxr_cleanup_user(dev, qp);
2660 
2661 	QL_DPRINT12(ha, "exit[%d]\n", rc);
2662 	return rc;
2663 }
2664 
2665 static void
2666 qlnxr_set_roce_db_info(struct qlnxr_dev *dev,
2667 	struct qlnxr_qp *qp)
2668 {
2669 	qlnx_host_t	*ha;
2670 
2671 	ha = dev->ha;
2672 
2673 	QL_DPRINT12(ha, "enter qp = %p qp->srq %p\n", qp, qp->srq);
2674 
2675 	qp->sq.db = dev->db_addr +
2676 		DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
2677 	qp->sq.db_data.data.icid = qp->icid + 1;
2678 
2679 	if (!qp->srq) {
2680 		qp->rq.db = dev->db_addr +
2681 			DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
2682 		qp->rq.db_data.data.icid = qp->icid;
2683 	}
2684 
2685 	QL_DPRINT12(ha, "exit\n");
2686 	return;
2687 }
2688 
2689 static void
2690 qlnxr_set_iwarp_db_info(struct qlnxr_dev *dev,
2691 	struct qlnxr_qp *qp)
2692 
2693 {
2694 	qlnx_host_t	*ha;
2695 
2696 	ha = dev->ha;
2697 
2698 	QL_DPRINT12(ha, "enter qp = %p qp->srq %p\n", qp, qp->srq);
2699 
2700 	qp->sq.db = dev->db_addr +
2701 		DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
2702 	qp->sq.db_data.data.icid = qp->icid;
2703 
2704 	if (!qp->srq) {
2705 		qp->rq.db = dev->db_addr +
2706 			DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_IWARP_RQ_PROD);
2707 		qp->rq.db_data.data.icid = qp->icid;
2708 
2709 		qp->rq.iwarp_db2 = dev->db_addr +
2710 			DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_FLAGS);
2711 		qp->rq.iwarp_db2_data.data.icid = qp->icid;
2712 		qp->rq.iwarp_db2_data.data.value = DQ_TCM_IWARP_POST_RQ_CF_CMD;
2713 	}
2714 
2715 	QL_DPRINT12(ha,
2716 		"qp->sq.db = %p qp->sq.db_data.data.icid =0x%x\n"
2717 		"\t\t\tqp->rq.db = %p qp->rq.db_data.data.icid =0x%x\n"
2718 		"\t\t\tqp->rq.iwarp_db2 = %p qp->rq.iwarp_db2.data.icid =0x%x"
2719 		" qp->rq.iwarp_db2.data.prod_val =0x%x\n",
2720 		qp->sq.db, qp->sq.db_data.data.icid,
2721 		qp->rq.db, qp->rq.db_data.data.icid,
2722 		qp->rq.iwarp_db2, qp->rq.iwarp_db2_data.data.icid,
2723 		qp->rq.iwarp_db2_data.data.value);
2724 
2725 	QL_DPRINT12(ha, "exit\n");
2726 	return;
2727 }
2728 
2729 static int
2730 qlnxr_roce_create_kernel_qp(struct qlnxr_dev *dev,
2731 	struct qlnxr_qp *qp,
2732 	struct ecore_rdma_create_qp_in_params *in_params,
2733 	u32 n_sq_elems,
2734 	u32 n_rq_elems)
2735 {
2736 	struct ecore_rdma_create_qp_out_params out_params;
2737 	int		rc;
2738 	qlnx_host_t	*ha;
2739 
2740 	ha = dev->ha;
2741 
2742 	QL_DPRINT12(ha, "enter\n");
2743 
2744         rc = ecore_chain_alloc(
2745                 dev->cdev,
2746                 ECORE_CHAIN_USE_TO_PRODUCE,
2747                 ECORE_CHAIN_MODE_PBL,
2748                 ECORE_CHAIN_CNT_TYPE_U32,
2749                 n_sq_elems,
2750                 QLNXR_SQE_ELEMENT_SIZE,
2751                 &qp->sq.pbl,
2752                 NULL);
2753 
2754 	if (rc) {
2755 		QL_DPRINT11(ha, "ecore_chain_alloc qp->sq.pbl failed[%d]\n", rc);
2756 		return rc;
2757 	}
2758 
2759 	in_params->sq_num_pages = ecore_chain_get_page_cnt(&qp->sq.pbl);
2760 	in_params->sq_pbl_ptr = ecore_chain_get_pbl_phys(&qp->sq.pbl);
2761 
2762 	if (!qp->srq) {
2763                 rc = ecore_chain_alloc(
2764                         dev->cdev,
2765                         ECORE_CHAIN_USE_TO_CONSUME_PRODUCE,
2766                         ECORE_CHAIN_MODE_PBL,
2767                         ECORE_CHAIN_CNT_TYPE_U32,
2768                         n_rq_elems,
2769                         QLNXR_RQE_ELEMENT_SIZE,
2770                         &qp->rq.pbl,
2771                         NULL);
2772 
2773 		if (rc) {
2774 			QL_DPRINT11(ha,
2775 				"ecore_chain_alloc qp->rq.pbl failed[%d]\n", rc);
2776 			return rc;
2777 		}
2778 
2779 		in_params->rq_num_pages = ecore_chain_get_page_cnt(&qp->rq.pbl);
2780 		in_params->rq_pbl_ptr = ecore_chain_get_pbl_phys(&qp->rq.pbl);
2781 	}
2782 
2783 	qp->ecore_qp = ecore_rdma_create_qp(dev->rdma_ctx, in_params, &out_params);
2784 
2785 	if (!qp->ecore_qp) {
2786 		QL_DPRINT11(ha, "qp->ecore_qp == NULL\n");
2787 		return -EINVAL;
2788 	}
2789 
2790 	qp->qp_id = out_params.qp_id;
2791 	qp->icid = out_params.icid;
2792 
2793 	qlnxr_set_roce_db_info(dev, qp);
2794 
2795 	QL_DPRINT12(ha, "exit\n");
2796 	return 0;
2797 }
2798 
2799 static int
2800 qlnxr_iwarp_create_kernel_qp(struct qlnxr_dev *dev,
2801 	struct qlnxr_qp *qp,
2802 	struct ecore_rdma_create_qp_in_params *in_params,
2803 	u32 n_sq_elems,
2804 	u32 n_rq_elems)
2805 {
2806 	struct ecore_rdma_destroy_qp_out_params d_out_params;
2807 	struct ecore_rdma_create_qp_out_params out_params;
2808 	struct ecore_chain_ext_pbl ext_pbl;
2809 	int rc;
2810 	qlnx_host_t	*ha;
2811 
2812 	ha = dev->ha;
2813 
2814 	QL_DPRINT12(ha, "enter\n");
2815 
2816 	in_params->sq_num_pages = ECORE_CHAIN_PAGE_CNT(n_sq_elems,
2817 						     QLNXR_SQE_ELEMENT_SIZE,
2818 						     ECORE_CHAIN_MODE_PBL);
2819 	in_params->rq_num_pages = ECORE_CHAIN_PAGE_CNT(n_rq_elems,
2820 						     QLNXR_RQE_ELEMENT_SIZE,
2821 						     ECORE_CHAIN_MODE_PBL);
2822 
2823 	QL_DPRINT12(ha, "n_sq_elems = 0x%x"
2824 		" n_rq_elems = 0x%x in_params\n"
2825 		"\t\t\tqp_handle_lo\t\t= 0x%08x\n"
2826 		"\t\t\tqp_handle_hi\t\t= 0x%08x\n"
2827 		"\t\t\tqp_handle_async_lo\t\t= 0x%08x\n"
2828 		"\t\t\tqp_handle_async_hi\t\t= 0x%08x\n"
2829 		"\t\t\tuse_srq\t\t\t= 0x%x\n"
2830 		"\t\t\tsignal_all\t\t= 0x%x\n"
2831 		"\t\t\tfmr_and_reserved_lkey\t= 0x%x\n"
2832 		"\t\t\tpd\t\t\t= 0x%x\n"
2833 		"\t\t\tdpi\t\t\t= 0x%x\n"
2834 		"\t\t\tsq_cq_id\t\t\t= 0x%x\n"
2835 		"\t\t\tsq_num_pages\t\t= 0x%x\n"
2836 		"\t\t\tsq_pbl_ptr\t\t= %p\n"
2837 		"\t\t\tmax_sq_sges\t\t= 0x%x\n"
2838 		"\t\t\trq_cq_id\t\t\t= 0x%x\n"
2839 		"\t\t\trq_num_pages\t\t= 0x%x\n"
2840 		"\t\t\trq_pbl_ptr\t\t= %p\n"
2841 		"\t\t\tsrq_id\t\t\t= 0x%x\n"
2842 		"\t\t\tstats_queue\t\t= 0x%x\n",
2843 		n_sq_elems, n_rq_elems,
2844 		in_params->qp_handle_lo,
2845 		in_params->qp_handle_hi,
2846 		in_params->qp_handle_async_lo,
2847 		in_params->qp_handle_async_hi,
2848 		in_params->use_srq,
2849 		in_params->signal_all,
2850 		in_params->fmr_and_reserved_lkey,
2851 		in_params->pd,
2852 		in_params->dpi,
2853 		in_params->sq_cq_id,
2854 		in_params->sq_num_pages,
2855 		(void *)in_params->sq_pbl_ptr,
2856 		in_params->max_sq_sges,
2857 		in_params->rq_cq_id,
2858 		in_params->rq_num_pages,
2859 		(void *)in_params->rq_pbl_ptr,
2860 		in_params->srq_id,
2861 		in_params->stats_queue );
2862 
2863 	memset(&out_params, 0, sizeof (struct ecore_rdma_create_qp_out_params));
2864 	memset(&ext_pbl, 0, sizeof (struct ecore_chain_ext_pbl));
2865 
2866 	qp->ecore_qp = ecore_rdma_create_qp(dev->rdma_ctx, in_params, &out_params);
2867 
2868 	if (!qp->ecore_qp) {
2869 		QL_DPRINT11(ha, "ecore_rdma_create_qp failed\n");
2870 		return -EINVAL;
2871 	}
2872 
2873 	/* Now we allocate the chain */
2874 	ext_pbl.p_pbl_virt = out_params.sq_pbl_virt;
2875 	ext_pbl.p_pbl_phys = out_params.sq_pbl_phys;
2876 
2877 	QL_DPRINT12(ha, "ext_pbl.p_pbl_virt = %p "
2878 		"ext_pbl.p_pbl_phys = %p\n",
2879 		ext_pbl.p_pbl_virt, ext_pbl.p_pbl_phys);
2880 
2881         rc = ecore_chain_alloc(
2882                 dev->cdev,
2883                 ECORE_CHAIN_USE_TO_PRODUCE,
2884                 ECORE_CHAIN_MODE_PBL,
2885                 ECORE_CHAIN_CNT_TYPE_U32,
2886                 n_sq_elems,
2887                 QLNXR_SQE_ELEMENT_SIZE,
2888                 &qp->sq.pbl,
2889                 &ext_pbl);
2890 
2891 	if (rc) {
2892 		QL_DPRINT11(ha,
2893 			"ecore_chain_alloc qp->sq.pbl failed rc = %d\n", rc);
2894 		goto err;
2895 	}
2896 
2897 	ext_pbl.p_pbl_virt = out_params.rq_pbl_virt;
2898 	ext_pbl.p_pbl_phys = out_params.rq_pbl_phys;
2899 
2900 	QL_DPRINT12(ha, "ext_pbl.p_pbl_virt = %p "
2901 		"ext_pbl.p_pbl_phys = %p\n",
2902 		ext_pbl.p_pbl_virt, ext_pbl.p_pbl_phys);
2903 
2904 	if (!qp->srq) {
2905                 rc = ecore_chain_alloc(
2906                         dev->cdev,
2907                         ECORE_CHAIN_USE_TO_CONSUME_PRODUCE,
2908                         ECORE_CHAIN_MODE_PBL,
2909                         ECORE_CHAIN_CNT_TYPE_U32,
2910                         n_rq_elems,
2911                         QLNXR_RQE_ELEMENT_SIZE,
2912                         &qp->rq.pbl,
2913                         &ext_pbl);
2914 
2915 		if (rc) {
2916 			QL_DPRINT11(ha,, "ecore_chain_alloc qp->rq.pbl"
2917 				" failed rc = %d\n", rc);
2918 			goto err;
2919 		}
2920 	}
2921 
2922 	QL_DPRINT12(ha, "qp_id = 0x%x icid =0x%x\n",
2923 		out_params.qp_id, out_params.icid);
2924 
2925 	qp->qp_id = out_params.qp_id;
2926 	qp->icid = out_params.icid;
2927 
2928 	qlnxr_set_iwarp_db_info(dev, qp);
2929 
2930 	QL_DPRINT12(ha, "exit\n");
2931 	return 0;
2932 
2933 err:
2934 	ecore_rdma_destroy_qp(dev->rdma_ctx, qp->ecore_qp, &d_out_params);
2935 
2936 	QL_DPRINT12(ha, "exit rc = %d\n", rc);
2937 	return rc;
2938 }
2939 
2940 static int
2941 qlnxr_create_kernel_qp(struct qlnxr_dev *dev,
2942 	struct qlnxr_qp *qp,
2943 	struct ib_pd *ibpd,
2944 	struct ib_qp_init_attr *attrs)
2945 {
2946 	struct ecore_rdma_create_qp_in_params in_params;
2947 	struct qlnxr_pd *pd = get_qlnxr_pd(ibpd);
2948 	int rc = -EINVAL;
2949 	u32 n_rq_elems;
2950 	u32 n_sq_elems;
2951 	u32 n_sq_entries;
2952 	struct ecore_rdma_device *qattr = ecore_rdma_query_device(dev->rdma_ctx);
2953 	qlnx_host_t	*ha;
2954 
2955 	ha = dev->ha;
2956 
2957 	QL_DPRINT12(ha, "enter\n");
2958 
2959 	memset(&in_params, 0, sizeof(in_params));
2960 
2961 	/* A single work request may take up to MAX_SQ_WQE_SIZE elements in
2962 	 * the ring. The ring should allow at least a single WR, even if the
2963 	 * user requested none, due to allocation issues.
2964 	 * We should add an extra WR since the prod and cons indices of
2965 	 * wqe_wr_id are managed in such a way that the WQ is considered full
2966 	 * when (prod+1)%max_wr==cons. We currently don't do that because we
2967 	 * double the number of entries due an iSER issue that pushes far more
2968 	 * WRs than indicated. If we decline its ib_post_send() then we get
2969 	 * error prints in the dmesg we'd like to avoid.
2970 	 */
2971 	qp->sq.max_wr = min_t(u32, attrs->cap.max_send_wr * dev->wq_multiplier,
2972 			      qattr->max_wqe);
2973 
2974 	qp->wqe_wr_id = kzalloc(qp->sq.max_wr * sizeof(*qp->wqe_wr_id),
2975 			GFP_KERNEL);
2976 	if (!qp->wqe_wr_id) {
2977 		QL_DPRINT11(ha, "failed SQ shadow memory allocation\n");
2978 		return -ENOMEM;
2979 	}
2980 
2981 	/* QP handle to be written in CQE */
2982 	in_params.qp_handle_lo = lower_32_bits((uintptr_t)qp);
2983 	in_params.qp_handle_hi = upper_32_bits((uintptr_t)qp);
2984 
2985 	/* A single work request may take up to MAX_RQ_WQE_SIZE elements in
2986 	 * the ring. There ring should allow at least a single WR, even if the
2987 	 * user requested none, due to allocation issues.
2988 	 */
2989 	qp->rq.max_wr = (u16)max_t(u32, attrs->cap.max_recv_wr, 1);
2990 
2991 	/* Allocate driver internal RQ array */
2992 	if (!qp->srq) {
2993 		qp->rqe_wr_id = kzalloc(qp->rq.max_wr * sizeof(*qp->rqe_wr_id),
2994 					GFP_KERNEL);
2995 		if (!qp->rqe_wr_id) {
2996 			QL_DPRINT11(ha, "failed RQ shadow memory allocation\n");
2997 			kfree(qp->wqe_wr_id);
2998 			return -ENOMEM;
2999 		}
3000 	}
3001 
3002 	//qlnxr_init_common_qp_in_params(dev, pd, qp, attrs, true, &in_params);
3003 
3004         in_params.qp_handle_async_lo = lower_32_bits((uintptr_t)qp);
3005         in_params.qp_handle_async_hi = upper_32_bits((uintptr_t)qp);
3006 
3007         in_params.signal_all = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR);
3008         in_params.fmr_and_reserved_lkey = true;
3009         in_params.pd = pd->pd_id;
3010         in_params.dpi = pd->uctx ? pd->uctx->dpi : dev->dpi;
3011         in_params.sq_cq_id = get_qlnxr_cq(attrs->send_cq)->icid;
3012         in_params.stats_queue = 0;
3013 
3014         in_params.rq_cq_id = get_qlnxr_cq(attrs->recv_cq)->icid;
3015 
3016         if (qp->srq) {
3017                 /* QP is associated with SRQ instead of RQ */
3018                 in_params.srq_id = qp->srq->srq_id;
3019                 in_params.use_srq = true;
3020                 QL_DPRINT11(ha, "exit srq_id = 0x%x use_srq = 0x%x\n",
3021                         in_params.srq_id, in_params.use_srq);
3022         } else {
3023         	in_params.srq_id = 0;
3024 		in_params.use_srq = false;
3025 	}
3026 
3027 	n_sq_entries = attrs->cap.max_send_wr;
3028 	n_sq_entries = min_t(u32, n_sq_entries, qattr->max_wqe);
3029 	n_sq_entries = max_t(u32, n_sq_entries, 1);
3030 	n_sq_elems = n_sq_entries * QLNXR_MAX_SQE_ELEMENTS_PER_SQE;
3031 
3032 	n_rq_elems = qp->rq.max_wr * QLNXR_MAX_RQE_ELEMENTS_PER_RQE;
3033 
3034 	if (QLNX_IS_ROCE(dev)) {
3035 		rc = qlnxr_roce_create_kernel_qp(dev, qp, &in_params,
3036 						n_sq_elems, n_rq_elems);
3037 	} else {
3038 		rc = qlnxr_iwarp_create_kernel_qp(dev, qp, &in_params,
3039 						 n_sq_elems, n_rq_elems);
3040 	}
3041 
3042 	if (rc)
3043 		qlnxr_cleanup_kernel(dev, qp);
3044 
3045 	QL_DPRINT12(ha, "exit [%d]\n", rc);
3046 	return rc;
3047 }
3048 
3049 struct ib_qp *
3050 qlnxr_create_qp(struct ib_pd *ibpd,
3051 		struct ib_qp_init_attr *attrs,
3052 		struct ib_udata *udata)
3053 {
3054 	struct qlnxr_dev *dev = get_qlnxr_dev(ibpd->device);
3055 	struct qlnxr_pd *pd = get_qlnxr_pd(ibpd);
3056 	struct qlnxr_qp *qp;
3057 	int rc = 0;
3058 	qlnx_host_t	*ha;
3059 
3060 	ha = dev->ha;
3061 
3062 	QL_DPRINT12(ha, "enter\n");
3063 
3064 	rc = qlnxr_check_qp_attrs(ibpd, dev, attrs, udata);
3065 	if (rc) {
3066 		QL_DPRINT11(ha, "qlnxr_check_qp_attrs failed [%d]\n", rc);
3067 		return ERR_PTR(rc);
3068 	}
3069 
3070 	QL_DPRINT12(ha, "called from %s, event_handle=%p,"
3071 		" eepd=%p sq_cq=%p, sq_icid=%d, rq_cq=%p, rq_icid=%d\n",
3072 		(udata ? "user library" : "kernel"),
3073 		attrs->event_handler, pd,
3074 		get_qlnxr_cq(attrs->send_cq),
3075 		get_qlnxr_cq(attrs->send_cq)->icid,
3076 		get_qlnxr_cq(attrs->recv_cq),
3077 		get_qlnxr_cq(attrs->recv_cq)->icid);
3078 
3079 	qp = qlnx_zalloc(sizeof(struct qlnxr_qp));
3080 
3081 	if (!qp) {
3082 		QL_DPRINT11(ha, "kzalloc(qp) failed\n");
3083 		return ERR_PTR(-ENOMEM);
3084 	}
3085 
3086 	qlnxr_set_common_qp_params(dev, qp, pd, attrs);
3087 
3088 	if (attrs->qp_type == IB_QPT_GSI) {
3089 		QL_DPRINT11(ha, "calling qlnxr_create_gsi_qp\n");
3090 		return qlnxr_create_gsi_qp(dev, attrs, qp);
3091 	}
3092 
3093 	if (udata) {
3094 		rc = qlnxr_create_user_qp(dev, qp, ibpd, udata, attrs);
3095 
3096 		if (rc) {
3097 			QL_DPRINT11(ha, "qlnxr_create_user_qp failed\n");
3098 			goto err;
3099 		}
3100 	} else {
3101 		rc = qlnxr_create_kernel_qp(dev, qp, ibpd, attrs);
3102 
3103 		if (rc) {
3104 			QL_DPRINT11(ha, "qlnxr_create_kernel_qp failed\n");
3105 			goto err;
3106 		}
3107 	}
3108 
3109 	qp->ibqp.qp_num = qp->qp_id;
3110 
3111 	rc = qlnxr_idr_add(dev, qp, qp->qp_id);
3112 
3113 	if (rc) {
3114 		QL_DPRINT11(ha, "qlnxr_idr_add failed\n");
3115 		goto err;
3116 	}
3117 
3118 	QL_DPRINT12(ha, "exit [%p]\n", &qp->ibqp);
3119 
3120 	return &qp->ibqp;
3121 err:
3122 	QL_DPRINT12(ha, "failed exit\n");
3123 	return ERR_PTR(-EFAULT);
3124 }
3125 
3126 static enum ib_qp_state
3127 qlnxr_get_ibqp_state(enum ecore_roce_qp_state qp_state)
3128 {
3129 	enum ib_qp_state state = IB_QPS_ERR;
3130 
3131 	switch (qp_state) {
3132 	case ECORE_ROCE_QP_STATE_RESET:
3133 		state = IB_QPS_RESET;
3134 		break;
3135 
3136 	case ECORE_ROCE_QP_STATE_INIT:
3137 		state = IB_QPS_INIT;
3138 		break;
3139 
3140 	case ECORE_ROCE_QP_STATE_RTR:
3141 		state = IB_QPS_RTR;
3142 		break;
3143 
3144 	case ECORE_ROCE_QP_STATE_RTS:
3145 		state = IB_QPS_RTS;
3146 		break;
3147 
3148 	case ECORE_ROCE_QP_STATE_SQD:
3149 		state = IB_QPS_SQD;
3150 		break;
3151 
3152 	case ECORE_ROCE_QP_STATE_ERR:
3153 		state = IB_QPS_ERR;
3154 		break;
3155 
3156 	case ECORE_ROCE_QP_STATE_SQE:
3157 		state = IB_QPS_SQE;
3158 		break;
3159 	}
3160 	return state;
3161 }
3162 
3163 static enum ecore_roce_qp_state
3164 qlnxr_get_state_from_ibqp( enum ib_qp_state qp_state)
3165 {
3166 	enum ecore_roce_qp_state ecore_qp_state;
3167 
3168 	ecore_qp_state = ECORE_ROCE_QP_STATE_ERR;
3169 
3170 	switch (qp_state) {
3171 	case IB_QPS_RESET:
3172 		ecore_qp_state =  ECORE_ROCE_QP_STATE_RESET;
3173 		break;
3174 
3175 	case IB_QPS_INIT:
3176 		ecore_qp_state =  ECORE_ROCE_QP_STATE_INIT;
3177 		break;
3178 
3179 	case IB_QPS_RTR:
3180 		ecore_qp_state =  ECORE_ROCE_QP_STATE_RTR;
3181 		break;
3182 
3183 	case IB_QPS_RTS:
3184 		ecore_qp_state =  ECORE_ROCE_QP_STATE_RTS;
3185 		break;
3186 
3187 	case IB_QPS_SQD:
3188 		ecore_qp_state =  ECORE_ROCE_QP_STATE_SQD;
3189 		break;
3190 
3191 	case IB_QPS_ERR:
3192 		ecore_qp_state =  ECORE_ROCE_QP_STATE_ERR;
3193 		break;
3194 
3195 	default:
3196 		ecore_qp_state =  ECORE_ROCE_QP_STATE_ERR;
3197 		break;
3198 	}
3199 
3200 	return (ecore_qp_state);
3201 }
3202 
3203 static void
3204 qlnxr_reset_qp_hwq_info(struct qlnxr_qp_hwq_info *qph)
3205 {
3206 	ecore_chain_reset(&qph->pbl);
3207 	qph->prod = qph->cons = 0;
3208 	qph->wqe_cons = 0;
3209 	qph->db_data.data.value = cpu_to_le16(0);
3210 
3211 	return;
3212 }
3213 
3214 static int
3215 qlnxr_update_qp_state(struct qlnxr_dev *dev,
3216 	struct qlnxr_qp *qp,
3217 	enum ecore_roce_qp_state new_state)
3218 {
3219 	int		status = 0;
3220 	uint32_t	reg_addr;
3221 	struct ecore_dev *cdev;
3222 	qlnx_host_t	*ha;
3223 
3224 	ha = dev->ha;
3225 	cdev = &ha->cdev;
3226 
3227 	QL_DPRINT12(ha, "enter qp = %p new_state = 0x%x qp->state = 0x%x\n",
3228 		qp, new_state, qp->state);
3229 
3230 	if (new_state == qp->state) {
3231 		return 0;
3232 	}
3233 
3234 	switch (qp->state) {
3235 	case ECORE_ROCE_QP_STATE_RESET:
3236 		switch (new_state) {
3237 		case ECORE_ROCE_QP_STATE_INIT:
3238 			qp->prev_wqe_size = 0;
3239 			qlnxr_reset_qp_hwq_info(&qp->sq);
3240 			if (!(qp->srq))
3241 				qlnxr_reset_qp_hwq_info(&qp->rq);
3242 			break;
3243 		default:
3244 			status = -EINVAL;
3245 			break;
3246 		};
3247 		break;
3248 	case ECORE_ROCE_QP_STATE_INIT:
3249 		/* INIT->XXX */
3250 		switch (new_state) {
3251 		case ECORE_ROCE_QP_STATE_RTR:
3252 		/* Update doorbell (in case post_recv was done before move to RTR) */
3253 			if (qp->srq)
3254 				break;
3255 			wmb();
3256 			//writel(qp->rq.db_data.raw, qp->rq.db);
3257 			//if (QLNX_IS_IWARP(dev))
3258 			//	writel(qp->rq.iwarp_db2_data.raw,
3259 			//	       qp->rq.iwarp_db2);
3260 
3261 			reg_addr = (uint32_t)((uint8_t *)qp->rq.db -
3262 					(uint8_t *)cdev->doorbells);
3263 
3264 			bus_write_4(ha->pci_dbells, reg_addr, qp->rq.db_data.raw);
3265 			bus_barrier(ha->pci_dbells,  0, 0, BUS_SPACE_BARRIER_READ);
3266 
3267 			if (QLNX_IS_IWARP(dev)) {
3268 				reg_addr = (uint32_t)((uint8_t *)qp->rq.iwarp_db2 -
3269 					(uint8_t *)cdev->doorbells);
3270 				bus_write_4(ha->pci_dbells, reg_addr,\
3271 					qp->rq.iwarp_db2_data.raw);
3272 				bus_barrier(ha->pci_dbells,  0, 0,\
3273 					BUS_SPACE_BARRIER_READ);
3274 			}
3275 
3276 
3277 			mmiowb();
3278 			break;
3279 		case ECORE_ROCE_QP_STATE_ERR:
3280 			/* TBD:flush qps... */
3281 			break;
3282 		default:
3283 			/* invalid state change. */
3284 			status = -EINVAL;
3285 			break;
3286 		};
3287 		break;
3288 	case ECORE_ROCE_QP_STATE_RTR:
3289 		/* RTR->XXX */
3290 		switch (new_state) {
3291 		case ECORE_ROCE_QP_STATE_RTS:
3292 			break;
3293 		case ECORE_ROCE_QP_STATE_ERR:
3294 			break;
3295 		default:
3296 			/* invalid state change. */
3297 			status = -EINVAL;
3298 			break;
3299 		};
3300 		break;
3301 	case ECORE_ROCE_QP_STATE_RTS:
3302 		/* RTS->XXX */
3303 		switch (new_state) {
3304 		case ECORE_ROCE_QP_STATE_SQD:
3305 			break;
3306 		case ECORE_ROCE_QP_STATE_ERR:
3307 			break;
3308 		default:
3309 			/* invalid state change. */
3310 			status = -EINVAL;
3311 			break;
3312 		};
3313 		break;
3314 	case ECORE_ROCE_QP_STATE_SQD:
3315 		/* SQD->XXX */
3316 		switch (new_state) {
3317 		case ECORE_ROCE_QP_STATE_RTS:
3318 		case ECORE_ROCE_QP_STATE_ERR:
3319 			break;
3320 		default:
3321 			/* invalid state change. */
3322 			status = -EINVAL;
3323 			break;
3324 		};
3325 		break;
3326 	case ECORE_ROCE_QP_STATE_ERR:
3327 		/* ERR->XXX */
3328 		switch (new_state) {
3329 		case ECORE_ROCE_QP_STATE_RESET:
3330 			if ((qp->rq.prod != qp->rq.cons) ||
3331 			    (qp->sq.prod != qp->sq.cons)) {
3332 				QL_DPRINT11(ha,
3333 					"Error->Reset with rq/sq "
3334 					"not empty rq.prod=0x%x rq.cons=0x%x"
3335 					" sq.prod=0x%x sq.cons=0x%x\n",
3336 					qp->rq.prod, qp->rq.cons,
3337 					qp->sq.prod, qp->sq.cons);
3338 				status = -EINVAL;
3339 			}
3340 			break;
3341 		default:
3342 			status = -EINVAL;
3343 			break;
3344 		};
3345 		break;
3346 	default:
3347 		status = -EINVAL;
3348 		break;
3349 	};
3350 
3351 	QL_DPRINT12(ha, "exit\n");
3352 	return status;
3353 }
3354 
3355 int
3356 qlnxr_modify_qp(struct ib_qp	*ibqp,
3357 	struct ib_qp_attr	*attr,
3358 	int			attr_mask,
3359 	struct ib_udata		*udata)
3360 {
3361 	int rc = 0;
3362 	struct qlnxr_qp *qp = get_qlnxr_qp(ibqp);
3363 	struct qlnxr_dev *dev = get_qlnxr_dev(&qp->dev->ibdev);
3364 	struct ecore_rdma_modify_qp_in_params qp_params = { 0 };
3365 	enum ib_qp_state old_qp_state, new_qp_state;
3366 	struct ecore_rdma_device *qattr = ecore_rdma_query_device(dev->rdma_ctx);
3367 	qlnx_host_t	*ha;
3368 
3369 	ha = dev->ha;
3370 
3371 	QL_DPRINT12(ha,
3372 		"enter qp = %p attr_mask = 0x%x, state = %d udata = %p\n",
3373 		qp, attr_mask, attr->qp_state, udata);
3374 
3375 	old_qp_state = qlnxr_get_ibqp_state(qp->state);
3376 	if (attr_mask & IB_QP_STATE)
3377 		new_qp_state = attr->qp_state;
3378 	else
3379 		new_qp_state = old_qp_state;
3380 
3381 	if (QLNX_IS_ROCE(dev)) {
3382 		if (!ib_modify_qp_is_ok(old_qp_state,
3383 					new_qp_state,
3384 					ibqp->qp_type,
3385 					attr_mask )) {
3386 			QL_DPRINT12(ha,
3387 				"invalid attribute mask=0x%x"
3388 				" specified for qpn=0x%x of type=0x%x \n"
3389 				" old_qp_state=0x%x, new_qp_state=0x%x\n",
3390 				attr_mask, qp->qp_id, ibqp->qp_type,
3391 				old_qp_state, new_qp_state);
3392 			rc = -EINVAL;
3393 			goto err;
3394 		}
3395 	}
3396 	/* translate the masks... */
3397 	if (attr_mask & IB_QP_STATE) {
3398 		SET_FIELD(qp_params.modify_flags,
3399 			  ECORE_RDMA_MODIFY_QP_VALID_NEW_STATE, 1);
3400 		qp_params.new_state = qlnxr_get_state_from_ibqp(attr->qp_state);
3401 	}
3402 
3403 	// TBD consider changing ecore to be a flag as well...
3404 	if (attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
3405 		qp_params.sqd_async = true;
3406 
3407 	if (attr_mask & IB_QP_PKEY_INDEX) {
3408 		SET_FIELD(qp_params.modify_flags,
3409 			  ECORE_ROCE_MODIFY_QP_VALID_PKEY,
3410 			  1);
3411 		if (attr->pkey_index >= QLNXR_ROCE_PKEY_TABLE_LEN) {
3412 			rc = -EINVAL;
3413 			goto err;
3414 		}
3415 
3416 		qp_params.pkey = QLNXR_ROCE_PKEY_DEFAULT;
3417 	}
3418 
3419 	if (attr_mask & IB_QP_QKEY) {
3420 		qp->qkey = attr->qkey;
3421 	}
3422 
3423 	/* tbd consider splitting in ecore.. */
3424 	if (attr_mask & IB_QP_ACCESS_FLAGS) {
3425 		SET_FIELD(qp_params.modify_flags,
3426 			  ECORE_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN, 1);
3427 		qp_params.incoming_rdma_read_en =
3428 			attr->qp_access_flags & IB_ACCESS_REMOTE_READ;
3429 		qp_params.incoming_rdma_write_en =
3430 			attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE;
3431 		qp_params.incoming_atomic_en =
3432 			attr->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC;
3433 	}
3434 
3435 	if (attr_mask & (IB_QP_AV | IB_QP_PATH_MTU)) {
3436 		if (attr_mask & IB_QP_PATH_MTU) {
3437 			if (attr->path_mtu < IB_MTU_256 ||
3438 			    attr->path_mtu > IB_MTU_4096) {
3439 				QL_DPRINT12(ha,
3440 					"Only MTU sizes of 256, 512, 1024,"
3441 					" 2048 and 4096 are supported "
3442 					" attr->path_mtu = [%d]\n",
3443 					attr->path_mtu);
3444 
3445 				rc = -EINVAL;
3446 				goto err;
3447 			}
3448 			qp->mtu = min(ib_mtu_enum_to_int(attr->path_mtu),
3449 				      ib_mtu_enum_to_int(
3450 						iboe_get_mtu(if_getmtu(dev->ha->ifp))));
3451 		}
3452 
3453 		if (qp->mtu == 0) {
3454 			qp->mtu = ib_mtu_enum_to_int(
3455 					iboe_get_mtu(if_getmtu(dev->ha->ifp)));
3456 			QL_DPRINT12(ha, "fixing zetoed MTU to qp->mtu = %d\n",
3457 				qp->mtu);
3458 		}
3459 
3460 		SET_FIELD(qp_params.modify_flags,
3461 			  ECORE_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR,
3462 			  1);
3463 
3464 		qp_params.traffic_class_tos = attr->ah_attr.grh.traffic_class;
3465 		qp_params.flow_label = attr->ah_attr.grh.flow_label;
3466 		qp_params.hop_limit_ttl = attr->ah_attr.grh.hop_limit;
3467 
3468 		qp->sgid_idx = attr->ah_attr.grh.sgid_index;
3469 
3470 		get_gid_info(ibqp, attr, attr_mask, dev, qp, &qp_params);
3471 
3472 		rc = qlnxr_get_dmac(dev, &attr->ah_attr, qp_params.remote_mac_addr);
3473 		if (rc)
3474 			return rc;
3475 
3476 		qp_params.use_local_mac = true;
3477 		memcpy(qp_params.local_mac_addr, dev->ha->primary_mac, ETH_ALEN);
3478 
3479 		QL_DPRINT12(ha, "dgid=0x%x:0x%x:0x%x:0x%x\n",
3480 		       qp_params.dgid.dwords[0], qp_params.dgid.dwords[1],
3481 		       qp_params.dgid.dwords[2], qp_params.dgid.dwords[3]);
3482 		QL_DPRINT12(ha, "sgid=0x%x:0x%x:0x%x:0x%x\n",
3483 		       qp_params.sgid.dwords[0], qp_params.sgid.dwords[1],
3484 		       qp_params.sgid.dwords[2], qp_params.sgid.dwords[3]);
3485 		QL_DPRINT12(ha,
3486 			"remote_mac=[0x%x:0x%x:0x%x:0x%x:0x%x:0x%x]\n",
3487 			qp_params.remote_mac_addr[0],
3488 			qp_params.remote_mac_addr[1],
3489 			qp_params.remote_mac_addr[2],
3490 			qp_params.remote_mac_addr[3],
3491 			qp_params.remote_mac_addr[4],
3492 			qp_params.remote_mac_addr[5]);
3493 
3494 		qp_params.mtu = qp->mtu;
3495 	}
3496 
3497 	if (qp_params.mtu == 0) {
3498 		/* stay with current MTU */
3499 		if (qp->mtu) {
3500 			qp_params.mtu = qp->mtu;
3501 		} else {
3502 			qp_params.mtu = ib_mtu_enum_to_int(
3503 						iboe_get_mtu(if_getmtu(dev->ha->ifp)));
3504 		}
3505 	}
3506 
3507 	if (attr_mask & IB_QP_TIMEOUT) {
3508 		SET_FIELD(qp_params.modify_flags, \
3509 			ECORE_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1);
3510 
3511 		qp_params.ack_timeout = attr->timeout;
3512 		if (attr->timeout) {
3513 			u32 temp;
3514 
3515 			/* 12.7.34 LOCAL ACK TIMEOUT
3516 			 * Value representing the transport (ACK) timeout for
3517 			 * use by the remote, expressed as (4.096 μS*2Local ACK
3518 			 * Timeout)
3519 			 */
3520 			/* We use 1UL since the temporal value may be  overflow
3521 			 * 32 bits
3522 			 */
3523 			temp = 4096 * (1UL << attr->timeout) / 1000 / 1000;
3524 			qp_params.ack_timeout = temp; /* FW requires [msec] */
3525 		}
3526 		else
3527 			qp_params.ack_timeout = 0; /* infinite */
3528 	}
3529 	if (attr_mask & IB_QP_RETRY_CNT) {
3530 		SET_FIELD(qp_params.modify_flags,\
3531 			 ECORE_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1);
3532 		qp_params.retry_cnt = attr->retry_cnt;
3533 	}
3534 
3535 	if (attr_mask & IB_QP_RNR_RETRY) {
3536 		SET_FIELD(qp_params.modify_flags,
3537 			  ECORE_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT,
3538 			  1);
3539 		qp_params.rnr_retry_cnt = attr->rnr_retry;
3540 	}
3541 
3542 	if (attr_mask & IB_QP_RQ_PSN) {
3543 		SET_FIELD(qp_params.modify_flags,
3544 			  ECORE_ROCE_MODIFY_QP_VALID_RQ_PSN,
3545 			  1);
3546 		qp_params.rq_psn = attr->rq_psn;
3547 		qp->rq_psn = attr->rq_psn;
3548 	}
3549 
3550 	if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
3551 		if (attr->max_rd_atomic > qattr->max_qp_req_rd_atomic_resc) {
3552 			rc = -EINVAL;
3553 			QL_DPRINT12(ha,
3554 				"unsupported  max_rd_atomic=%d, supported=%d\n",
3555 				attr->max_rd_atomic,
3556 				qattr->max_qp_req_rd_atomic_resc);
3557 			goto err;
3558 		}
3559 
3560 		SET_FIELD(qp_params.modify_flags,
3561 			  ECORE_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ,
3562 			  1);
3563 		qp_params.max_rd_atomic_req = attr->max_rd_atomic;
3564 	}
3565 
3566 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
3567 		SET_FIELD(qp_params.modify_flags,
3568 			  ECORE_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER,
3569 			  1);
3570 		qp_params.min_rnr_nak_timer = attr->min_rnr_timer;
3571 	}
3572 
3573 	if (attr_mask & IB_QP_SQ_PSN) {
3574 		SET_FIELD(qp_params.modify_flags,
3575 			  ECORE_ROCE_MODIFY_QP_VALID_SQ_PSN,
3576 			  1);
3577 		qp_params.sq_psn = attr->sq_psn;
3578 		qp->sq_psn = attr->sq_psn;
3579 	}
3580 
3581 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
3582 		if (attr->max_dest_rd_atomic >
3583 		    qattr->max_qp_resp_rd_atomic_resc) {
3584 			QL_DPRINT12(ha,
3585 				"unsupported max_dest_rd_atomic=%d, "
3586 				"supported=%d\n",
3587 				attr->max_dest_rd_atomic,
3588 				qattr->max_qp_resp_rd_atomic_resc);
3589 
3590 			rc = -EINVAL;
3591 			goto err;
3592 		}
3593 
3594 		SET_FIELD(qp_params.modify_flags,
3595 			  ECORE_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP,
3596 			  1);
3597 		qp_params.max_rd_atomic_resp = attr->max_dest_rd_atomic;
3598 	}
3599 
3600  	if (attr_mask & IB_QP_DEST_QPN) {
3601 		SET_FIELD(qp_params.modify_flags,
3602 			  ECORE_ROCE_MODIFY_QP_VALID_DEST_QP,
3603 			  1);
3604 
3605 		qp_params.dest_qp = attr->dest_qp_num;
3606 		qp->dest_qp_num = attr->dest_qp_num;
3607 	}
3608 
3609 	/*
3610 	 * Update the QP state before the actual ramrod to prevent a race with
3611 	 * fast path. Modifying the QP state to error will cause the device to
3612 	 * flush the CQEs and while polling the flushed CQEs will considered as
3613 	 * a potential issue if the QP isn't in error state.
3614 	 */
3615 	if ((attr_mask & IB_QP_STATE) && (qp->qp_type != IB_QPT_GSI) &&
3616 		(!udata) && (qp_params.new_state == ECORE_ROCE_QP_STATE_ERR))
3617 		qp->state = ECORE_ROCE_QP_STATE_ERR;
3618 
3619 	if (qp->qp_type != IB_QPT_GSI)
3620 		rc = ecore_rdma_modify_qp(dev->rdma_ctx, qp->ecore_qp, &qp_params);
3621 
3622 	if (attr_mask & IB_QP_STATE) {
3623 		if ((qp->qp_type != IB_QPT_GSI) && (!udata))
3624 			rc = qlnxr_update_qp_state(dev, qp, qp_params.new_state);
3625 		qp->state = qp_params.new_state;
3626 	}
3627 
3628 err:
3629 	QL_DPRINT12(ha, "exit\n");
3630 	return rc;
3631 }
3632 
3633 static int
3634 qlnxr_to_ib_qp_acc_flags(struct ecore_rdma_query_qp_out_params *params)
3635 {
3636 	int ib_qp_acc_flags = 0;
3637 
3638 	if (params->incoming_rdma_write_en)
3639 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
3640 	if (params->incoming_rdma_read_en)
3641 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_READ;
3642 	if (params->incoming_atomic_en)
3643 		ib_qp_acc_flags |= IB_ACCESS_REMOTE_ATOMIC;
3644 	if (true) /* FIXME -> local write ?? */
3645 		ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
3646 
3647 	return ib_qp_acc_flags;
3648 }
3649 
3650 static enum ib_mtu
3651 qlnxr_mtu_int_to_enum(u16 mtu)
3652 {
3653 	enum ib_mtu ib_mtu_size;
3654 
3655 	switch (mtu) {
3656 	case 256:
3657 		ib_mtu_size = IB_MTU_256;
3658 		break;
3659 
3660 	case 512:
3661 		ib_mtu_size = IB_MTU_512;
3662 		break;
3663 
3664 	case 1024:
3665 		ib_mtu_size = IB_MTU_1024;
3666 		break;
3667 
3668 	case 2048:
3669 		ib_mtu_size = IB_MTU_2048;
3670 		break;
3671 
3672 	case 4096:
3673 		ib_mtu_size = IB_MTU_4096;
3674 		break;
3675 
3676 	default:
3677 		ib_mtu_size = IB_MTU_1024;
3678 		break;
3679 	}
3680 	return (ib_mtu_size);
3681 }
3682 
3683 int
3684 qlnxr_query_qp(struct ib_qp *ibqp,
3685 	struct ib_qp_attr *qp_attr,
3686 	int attr_mask,
3687 	struct ib_qp_init_attr *qp_init_attr)
3688 {
3689 	int rc = 0;
3690 	struct ecore_rdma_query_qp_out_params params;
3691 	struct qlnxr_qp *qp = get_qlnxr_qp(ibqp);
3692 	struct qlnxr_dev *dev = qp->dev;
3693 	qlnx_host_t	*ha;
3694 
3695 	ha = dev->ha;
3696 
3697 	QL_DPRINT12(ha, "enter\n");
3698 
3699 	memset(&params, 0, sizeof(params));
3700 
3701 	rc = ecore_rdma_query_qp(dev->rdma_ctx, qp->ecore_qp, &params);
3702 	if (rc)
3703 		goto err;
3704 
3705 	memset(qp_attr, 0, sizeof(*qp_attr));
3706 	memset(qp_init_attr, 0, sizeof(*qp_init_attr));
3707 
3708 	qp_attr->qp_state = qlnxr_get_ibqp_state(params.state);
3709 	qp_attr->cur_qp_state = qlnxr_get_ibqp_state(params.state);
3710 
3711 	/* In some cases in iWARP qelr will ask for the state only */
3712 	if (QLNX_IS_IWARP(dev) && (attr_mask == IB_QP_STATE)) {
3713 		QL_DPRINT11(ha, "only state requested\n");
3714 		return 0;
3715 	}
3716 
3717 	qp_attr->path_mtu = qlnxr_mtu_int_to_enum(params.mtu);
3718 	qp_attr->path_mig_state = IB_MIG_MIGRATED;
3719 	qp_attr->rq_psn = params.rq_psn;
3720 	qp_attr->sq_psn = params.sq_psn;
3721 	qp_attr->dest_qp_num = params.dest_qp;
3722 
3723 	qp_attr->qp_access_flags = qlnxr_to_ib_qp_acc_flags(&params);
3724 
3725 	QL_DPRINT12(ha, "qp_state = 0x%x cur_qp_state = 0x%x "
3726 		"path_mtu = %d qp_access_flags = 0x%x\n",
3727 		qp_attr->qp_state, qp_attr->cur_qp_state, qp_attr->path_mtu,
3728 		qp_attr->qp_access_flags);
3729 
3730 	qp_attr->cap.max_send_wr = qp->sq.max_wr;
3731 	qp_attr->cap.max_recv_wr = qp->rq.max_wr;
3732 	qp_attr->cap.max_send_sge = qp->sq.max_sges;
3733 	qp_attr->cap.max_recv_sge = qp->rq.max_sges;
3734 	qp_attr->cap.max_inline_data = qp->max_inline_data;
3735 	qp_init_attr->cap = qp_attr->cap;
3736 
3737 	memcpy(&qp_attr->ah_attr.grh.dgid.raw[0], &params.dgid.bytes[0],
3738 	       sizeof(qp_attr->ah_attr.grh.dgid.raw));
3739 
3740 	qp_attr->ah_attr.grh.flow_label = params.flow_label;
3741 	qp_attr->ah_attr.grh.sgid_index = qp->sgid_idx;
3742 	qp_attr->ah_attr.grh.hop_limit = params.hop_limit_ttl;
3743 	qp_attr->ah_attr.grh.traffic_class = params.traffic_class_tos;
3744 
3745 	qp_attr->ah_attr.ah_flags = IB_AH_GRH;
3746 	qp_attr->ah_attr.port_num = 1; /* FIXME -> check this */
3747 	qp_attr->ah_attr.sl = 0;/* FIXME -> check this */
3748 	qp_attr->timeout = params.timeout;
3749 	qp_attr->rnr_retry = params.rnr_retry;
3750 	qp_attr->retry_cnt = params.retry_cnt;
3751 	qp_attr->min_rnr_timer = params.min_rnr_nak_timer;
3752 	qp_attr->pkey_index = params.pkey_index;
3753 	qp_attr->port_num = 1; /* FIXME -> check this */
3754 	qp_attr->ah_attr.src_path_bits = 0;
3755 	qp_attr->ah_attr.static_rate = 0;
3756 	qp_attr->alt_pkey_index = 0;
3757 	qp_attr->alt_port_num = 0;
3758 	qp_attr->alt_timeout = 0;
3759 	memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
3760 
3761 	qp_attr->sq_draining = (params.state == ECORE_ROCE_QP_STATE_SQD) ? 1 : 0;
3762 	qp_attr->max_dest_rd_atomic = params.max_dest_rd_atomic;
3763 	qp_attr->max_rd_atomic = params.max_rd_atomic;
3764 	qp_attr->en_sqd_async_notify = (params.sqd_async)? 1 : 0;
3765 
3766 	QL_DPRINT12(ha, "max_inline_data=%d\n",
3767 		qp_attr->cap.max_inline_data);
3768 
3769 err:
3770 	QL_DPRINT12(ha, "exit\n");
3771 	return rc;
3772 }
3773 
3774 static void
3775 qlnxr_cleanup_user(struct qlnxr_dev *dev, struct qlnxr_qp *qp)
3776 {
3777 	qlnx_host_t	*ha;
3778 
3779 	ha = dev->ha;
3780 
3781 	QL_DPRINT12(ha, "enter\n");
3782 
3783 	if (qp->usq.umem)
3784 		ib_umem_release(qp->usq.umem);
3785 
3786 	qp->usq.umem = NULL;
3787 
3788 	if (qp->urq.umem)
3789 		ib_umem_release(qp->urq.umem);
3790 
3791 	qp->urq.umem = NULL;
3792 
3793 	QL_DPRINT12(ha, "exit\n");
3794 	return;
3795 }
3796 
3797 static void
3798 qlnxr_cleanup_kernel(struct qlnxr_dev *dev, struct qlnxr_qp *qp)
3799 {
3800 	qlnx_host_t	*ha;
3801 
3802 	ha = dev->ha;
3803 
3804 	QL_DPRINT12(ha, "enter\n");
3805 
3806 	if (qlnxr_qp_has_sq(qp)) {
3807 		QL_DPRINT12(ha, "freeing SQ\n");
3808 		ha->qlnxr_debug = 1;
3809 //		ecore_chain_free(dev->cdev, &qp->sq.pbl);
3810 		ha->qlnxr_debug = 0;
3811 		kfree(qp->wqe_wr_id);
3812 	}
3813 
3814 	if (qlnxr_qp_has_rq(qp)) {
3815 		QL_DPRINT12(ha, "freeing RQ\n");
3816 		ha->qlnxr_debug = 1;
3817 	//	ecore_chain_free(dev->cdev, &qp->rq.pbl);
3818 		ha->qlnxr_debug = 0;
3819 		kfree(qp->rqe_wr_id);
3820 	}
3821 
3822 	QL_DPRINT12(ha, "exit\n");
3823 	return;
3824 }
3825 
3826 static int
3827 qlnxr_free_qp_resources(struct qlnxr_dev *dev,
3828     struct qlnxr_qp *qp, struct ib_udata *udata)
3829 {
3830 	int		rc = 0;
3831 	qlnx_host_t	*ha;
3832 	struct ecore_rdma_destroy_qp_out_params d_out_params;
3833 
3834 	ha = dev->ha;
3835 
3836 	QL_DPRINT12(ha, "enter\n");
3837 
3838 #if 0
3839 	if (qp->qp_type != IB_QPT_GSI) {
3840 		rc = ecore_rdma_destroy_qp(dev->rdma_ctx, qp->ecore_qp,
3841 				&d_out_params);
3842 		if (rc)
3843 			return rc;
3844 	}
3845 
3846 	if (udata)
3847 		qlnxr_cleanup_user(dev, qp);
3848 	else
3849 		qlnxr_cleanup_kernel(dev, qp);
3850 #endif
3851 
3852 	if (udata)
3853 		qlnxr_cleanup_user(dev, qp);
3854 	else
3855 		qlnxr_cleanup_kernel(dev, qp);
3856 
3857 	if (qp->qp_type != IB_QPT_GSI) {
3858 		rc = ecore_rdma_destroy_qp(dev->rdma_ctx, qp->ecore_qp,
3859 				&d_out_params);
3860 		if (rc)
3861 			return rc;
3862 	}
3863 
3864 	QL_DPRINT12(ha, "exit\n");
3865 	return 0;
3866 }
3867 
3868 int
3869 qlnxr_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
3870 {
3871 	struct qlnxr_qp *qp = get_qlnxr_qp(ibqp);
3872 	struct qlnxr_dev *dev = qp->dev;
3873 	int rc = 0;
3874 	struct ib_qp_attr attr;
3875 	int attr_mask = 0;
3876 	qlnx_host_t	*ha;
3877 
3878 	ha = dev->ha;
3879 
3880 	QL_DPRINT12(ha, "enter qp = %p, qp_type=%d\n", qp, qp->qp_type);
3881 
3882 	qp->destroyed = 1;
3883 
3884 	if (QLNX_IS_ROCE(dev) && (qp->state != (ECORE_ROCE_QP_STATE_RESET |
3885 				  ECORE_ROCE_QP_STATE_ERR |
3886 				  ECORE_ROCE_QP_STATE_INIT))) {
3887 		attr.qp_state = IB_QPS_ERR;
3888 		attr_mask |= IB_QP_STATE;
3889 
3890 		/* change the QP state to ERROR */
3891 		qlnxr_modify_qp(ibqp, &attr, attr_mask, NULL);
3892 	}
3893 
3894 	if (qp->qp_type == IB_QPT_GSI)
3895 		qlnxr_destroy_gsi_qp(dev);
3896 
3897 	qp->sig = ~qp->sig;
3898 
3899 	qlnxr_free_qp_resources(dev, qp, udata);
3900 
3901 	if (atomic_dec_and_test(&qp->refcnt)) {
3902 		/* TODO: only for iWARP? */
3903 		qlnxr_idr_remove(dev, qp->qp_id);
3904 	}
3905 
3906 	QL_DPRINT12(ha, "exit\n");
3907 	return rc;
3908 }
3909 
3910 static inline int
3911 qlnxr_wq_is_full(struct qlnxr_qp_hwq_info *wq)
3912 {
3913 	return (((wq->prod + 1) % wq->max_wr) == wq->cons);
3914 }
3915 
3916 static int
3917 sge_data_len(struct ib_sge *sg_list, int num_sge)
3918 {
3919 	int i, len = 0;
3920 	for (i = 0; i < num_sge; i++)
3921 		len += sg_list[i].length;
3922 	return len;
3923 }
3924 
3925 static void
3926 swap_wqe_data64(u64 *p)
3927 {
3928 	int i;
3929 
3930 	for (i = 0; i < QLNXR_SQE_ELEMENT_SIZE / sizeof(u64); i++, p++)
3931 		*p = cpu_to_be64(cpu_to_le64(*p));
3932 }
3933 
3934 static u32
3935 qlnxr_prepare_sq_inline_data(struct qlnxr_dev *dev,
3936 	struct qlnxr_qp		*qp,
3937 	u8			*wqe_size,
3938 	const struct ib_send_wr	*wr,
3939 	const struct ib_send_wr	**bad_wr,
3940 	u8			*bits,
3941 	u8			bit)
3942 {
3943 	int i, seg_siz;
3944 	char *seg_prt, *wqe;
3945 	u32 data_size = sge_data_len(wr->sg_list, wr->num_sge);
3946 	qlnx_host_t	*ha;
3947 
3948 	ha = dev->ha;
3949 
3950 	QL_DPRINT12(ha, "enter[%d]\n", data_size);
3951 
3952 	if (data_size > ROCE_REQ_MAX_INLINE_DATA_SIZE) {
3953 		QL_DPRINT12(ha,
3954 			"Too much inline data in WR:[%d, %d]\n",
3955 			data_size, ROCE_REQ_MAX_INLINE_DATA_SIZE);
3956 		*bad_wr = wr;
3957 		return 0;
3958 	}
3959 
3960 	if (!data_size)
3961 		return data_size;
3962 
3963 	/* set the bit */
3964 	*bits |= bit;
3965 
3966 	seg_prt = wqe = NULL;
3967 	seg_siz = 0;
3968 
3969 	/* copy data inline */
3970 	for (i = 0; i < wr->num_sge; i++) {
3971 		u32 len = wr->sg_list[i].length;
3972 		void *src = (void *)(uintptr_t)wr->sg_list[i].addr;
3973 
3974 		while (len > 0) {
3975 			u32 cur;
3976 
3977 			/* new segment required */
3978 			if (!seg_siz) {
3979 				wqe = (char *)ecore_chain_produce(&qp->sq.pbl);
3980 				seg_prt = wqe;
3981 				seg_siz = sizeof(struct rdma_sq_common_wqe);
3982 				(*wqe_size)++;
3983 			}
3984 
3985 			/* calculate currently allowed length */
3986 			cur = MIN(len, seg_siz);
3987 
3988 			memcpy(seg_prt, src, cur);
3989 
3990 			/* update segment variables */
3991 			seg_prt += cur;
3992 			seg_siz -= cur;
3993 			/* update sge variables */
3994 			src += cur;
3995 			len -= cur;
3996 
3997 			/* swap fully-completed segments */
3998 			if (!seg_siz)
3999 				swap_wqe_data64((u64 *)wqe);
4000 		}
4001 	}
4002 
4003 	/* swap last not completed segment */
4004 	if (seg_siz)
4005 		swap_wqe_data64((u64 *)wqe);
4006 
4007 	QL_DPRINT12(ha, "exit\n");
4008 	return data_size;
4009 }
4010 
4011 static u32
4012 qlnxr_prepare_sq_sges(struct qlnxr_dev *dev, struct qlnxr_qp *qp,
4013 	u8 *wqe_size, const struct ib_send_wr *wr)
4014 {
4015 	int i;
4016 	u32 data_size = 0;
4017 	qlnx_host_t	*ha;
4018 
4019 	ha = dev->ha;
4020 
4021 	QL_DPRINT12(ha, "enter wr->num_sge = %d \n", wr->num_sge);
4022 
4023 	for (i = 0; i < wr->num_sge; i++) {
4024 		struct rdma_sq_sge *sge = ecore_chain_produce(&qp->sq.pbl);
4025 
4026 		TYPEPTR_ADDR_SET(sge, addr, wr->sg_list[i].addr);
4027 		sge->l_key = cpu_to_le32(wr->sg_list[i].lkey);
4028 		sge->length = cpu_to_le32(wr->sg_list[i].length);
4029 		data_size += wr->sg_list[i].length;
4030 	}
4031 
4032 	if (wqe_size)
4033 		*wqe_size += wr->num_sge;
4034 
4035 	QL_DPRINT12(ha, "exit data_size = %d\n", data_size);
4036 	return data_size;
4037 }
4038 
4039 static u32
4040 qlnxr_prepare_sq_rdma_data(struct qlnxr_dev *dev,
4041 	struct qlnxr_qp *qp,
4042 	struct rdma_sq_rdma_wqe_1st *rwqe,
4043 	struct rdma_sq_rdma_wqe_2nd *rwqe2,
4044 	const struct ib_send_wr *wr,
4045 	const struct ib_send_wr **bad_wr)
4046 {
4047 	qlnx_host_t	*ha;
4048 	u32             ret = 0;
4049 
4050 	ha = dev->ha;
4051 
4052 	QL_DPRINT12(ha, "enter\n");
4053 
4054 	rwqe2->r_key = cpu_to_le32(rdma_wr(wr)->rkey);
4055 	TYPEPTR_ADDR_SET(rwqe2, remote_va, rdma_wr(wr)->remote_addr);
4056 
4057 	if (wr->send_flags & IB_SEND_INLINE) {
4058 		u8 flags = 0;
4059 		SET_FIELD2(flags, RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG, 1);
4060 		return qlnxr_prepare_sq_inline_data(dev, qp, &rwqe->wqe_size,
4061 				wr, bad_wr, &rwqe->flags, flags);
4062 	}
4063 
4064 	ret = qlnxr_prepare_sq_sges(dev, qp, &rwqe->wqe_size, wr);
4065 
4066 	QL_DPRINT12(ha, "exit ret = 0x%x\n", ret);
4067 
4068 	return (ret);
4069 }
4070 
4071 static u32
4072 qlnxr_prepare_sq_send_data(struct qlnxr_dev *dev,
4073 	struct qlnxr_qp *qp,
4074 	struct rdma_sq_send_wqe *swqe,
4075 	struct rdma_sq_send_wqe *swqe2,
4076 	const struct ib_send_wr *wr,
4077 	const struct ib_send_wr **bad_wr)
4078 {
4079 	qlnx_host_t	*ha;
4080 	u32             ret = 0;
4081 
4082 	ha = dev->ha;
4083 
4084 	QL_DPRINT12(ha, "enter\n");
4085 
4086 	memset(swqe2, 0, sizeof(*swqe2));
4087 
4088 	if (wr->send_flags & IB_SEND_INLINE) {
4089 		u8 flags = 0;
4090 		SET_FIELD2(flags, RDMA_SQ_SEND_WQE_INLINE_FLG, 1);
4091 		return qlnxr_prepare_sq_inline_data(dev, qp, &swqe->wqe_size,
4092 				wr, bad_wr, &swqe->flags, flags);
4093 	}
4094 
4095 	ret = qlnxr_prepare_sq_sges(dev, qp, &swqe->wqe_size, wr);
4096 
4097 	QL_DPRINT12(ha, "exit ret = 0x%x\n", ret);
4098 
4099 	return (ret);
4100 }
4101 
4102 static void
4103 qlnx_handle_completed_mrs(struct qlnxr_dev *dev, struct mr_info *info)
4104 {
4105 	qlnx_host_t	*ha;
4106 
4107 	ha = dev->ha;
4108 
4109 	int work = info->completed - info->completed_handled - 1;
4110 
4111 	QL_DPRINT12(ha, "enter [%d]\n", work);
4112 
4113 	while (work-- > 0 && !list_empty(&info->inuse_pbl_list)) {
4114 		struct qlnxr_pbl *pbl;
4115 
4116 		/* Free all the page list that are possible to be freed
4117 		 * (all the ones that were invalidated), under the assumption
4118 		 * that if an FMR was completed successfully that means that
4119 		 * if there was an invalidate operation before it also ended
4120 		 */
4121 		pbl = list_first_entry(&info->inuse_pbl_list,
4122 				       struct qlnxr_pbl,
4123 				       list_entry);
4124 		list_del(&pbl->list_entry);
4125 		list_add_tail(&pbl->list_entry, &info->free_pbl_list);
4126 		info->completed_handled++;
4127 	}
4128 
4129 	QL_DPRINT12(ha, "exit\n");
4130 	return;
4131 }
4132 
4133 static int qlnxr_prepare_reg(struct qlnxr_qp *qp,
4134 		struct rdma_sq_fmr_wqe_1st *fwqe1,
4135 		const struct ib_reg_wr *wr)
4136 {
4137 	struct qlnxr_mr *mr = get_qlnxr_mr(wr->mr);
4138 	struct rdma_sq_fmr_wqe_2nd *fwqe2;
4139 
4140 	fwqe2 = (struct rdma_sq_fmr_wqe_2nd *)ecore_chain_produce(&qp->sq.pbl);
4141 	fwqe1->addr.hi = upper_32_bits(mr->ibmr.iova);
4142 	fwqe1->addr.lo = lower_32_bits(mr->ibmr.iova);
4143 	fwqe1->l_key = wr->key;
4144 
4145 	fwqe2->access_ctrl = 0;
4146 
4147 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_READ,
4148 		!!(wr->access & IB_ACCESS_REMOTE_READ));
4149 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE,
4150 		!!(wr->access & IB_ACCESS_REMOTE_WRITE));
4151 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC,
4152 		!!(wr->access & IB_ACCESS_REMOTE_ATOMIC));
4153 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_READ, 1);
4154 	SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE,
4155 		!!(wr->access & IB_ACCESS_LOCAL_WRITE));
4156 	fwqe2->fmr_ctrl = 0;
4157 
4158 	SET_FIELD2(fwqe2->fmr_ctrl, RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG,
4159 		ilog2(mr->ibmr.page_size) - 12);
4160 
4161 	fwqe2->length_hi = 0; /* TODO - figure out why length is only 32bit.. */
4162 	fwqe2->length_lo = mr->ibmr.length;
4163 	fwqe2->pbl_addr.hi = upper_32_bits(mr->info.pbl_table->pa);
4164 	fwqe2->pbl_addr.lo = lower_32_bits(mr->info.pbl_table->pa);
4165 
4166 	qp->wqe_wr_id[qp->sq.prod].mr = mr;
4167 
4168 	return 0;
4169 }
4170 
4171 static enum ib_wc_opcode
4172 qlnxr_ib_to_wc_opcode(enum ib_wr_opcode opcode)
4173 {
4174 	switch (opcode) {
4175 	case IB_WR_RDMA_WRITE:
4176 	case IB_WR_RDMA_WRITE_WITH_IMM:
4177 		return IB_WC_RDMA_WRITE;
4178 	case IB_WR_SEND_WITH_IMM:
4179 	case IB_WR_SEND:
4180 	case IB_WR_SEND_WITH_INV:
4181 		return IB_WC_SEND;
4182 	case IB_WR_RDMA_READ:
4183 		return IB_WC_RDMA_READ;
4184 	case IB_WR_ATOMIC_CMP_AND_SWP:
4185 		return IB_WC_COMP_SWAP;
4186 	case IB_WR_ATOMIC_FETCH_AND_ADD:
4187 		return IB_WC_FETCH_ADD;
4188 	case IB_WR_REG_MR:
4189 		return IB_WC_REG_MR;
4190 	case IB_WR_LOCAL_INV:
4191 		return IB_WC_LOCAL_INV;
4192 	default:
4193 		return IB_WC_SEND;
4194 	}
4195 }
4196 static inline bool
4197 qlnxr_can_post_send(struct qlnxr_qp *qp, const struct ib_send_wr *wr)
4198 {
4199 	int wq_is_full, err_wr, pbl_is_full;
4200 	struct qlnxr_dev *dev = qp->dev;
4201 	qlnx_host_t	*ha;
4202 
4203 	ha = dev->ha;
4204 
4205 	QL_DPRINT12(ha, "enter[qp, wr] = [%p,%p]\n", qp, wr);
4206 
4207 	/* prevent SQ overflow and/or processing of a bad WR */
4208 	err_wr = wr->num_sge > qp->sq.max_sges;
4209 	wq_is_full = qlnxr_wq_is_full(&qp->sq);
4210 	pbl_is_full = ecore_chain_get_elem_left_u32(&qp->sq.pbl) <
4211 		      QLNXR_MAX_SQE_ELEMENTS_PER_SQE;
4212 	if (wq_is_full || err_wr || pbl_is_full) {
4213 		if (wq_is_full &&
4214 		    !(qp->err_bitmap & QLNXR_QP_ERR_SQ_FULL)) {
4215 			qp->err_bitmap |= QLNXR_QP_ERR_SQ_FULL;
4216 
4217 			QL_DPRINT12(ha,
4218 				"error: WQ is full. Post send on QP failed"
4219 				" (this error appears only once) "
4220 				"[qp, wr, qp->err_bitmap]=[%p, %p, 0x%x]\n",
4221 				qp, wr, qp->err_bitmap);
4222 		}
4223 
4224 		if (err_wr &&
4225 		    !(qp->err_bitmap & QLNXR_QP_ERR_BAD_SR)) {
4226 			qp->err_bitmap |= QLNXR_QP_ERR_BAD_SR;
4227 
4228 			QL_DPRINT12(ha,
4229 				"error: WQ is bad. Post send on QP failed"
4230 				" (this error appears only once) "
4231 				"[qp, wr, qp->err_bitmap]=[%p, %p, 0x%x]\n",
4232 				qp, wr, qp->err_bitmap);
4233 		}
4234 
4235 		if (pbl_is_full &&
4236 		    !(qp->err_bitmap & QLNXR_QP_ERR_SQ_PBL_FULL)) {
4237 			qp->err_bitmap |= QLNXR_QP_ERR_SQ_PBL_FULL;
4238 
4239 			QL_DPRINT12(ha,
4240 				"error: WQ PBL is full. Post send on QP failed"
4241 				" (this error appears only once) "
4242 				"[qp, wr, qp->err_bitmap]=[%p, %p, 0x%x]\n",
4243 				qp, wr, qp->err_bitmap);
4244 		}
4245 		return false;
4246 	}
4247 	QL_DPRINT12(ha, "exit[qp, wr] = [%p,%p]\n", qp, wr);
4248 	return true;
4249 }
4250 
4251 int
4252 qlnxr_post_send(struct ib_qp *ibqp,
4253 	const struct ib_send_wr *wr,
4254 	const struct ib_send_wr **bad_wr)
4255 {
4256 	struct qlnxr_dev	*dev = get_qlnxr_dev(ibqp->device);
4257 	struct qlnxr_qp		*qp = get_qlnxr_qp(ibqp);
4258 	unsigned long 		flags;
4259 	int 			status = 0, rc = 0;
4260 	bool			comp;
4261 	qlnx_host_t		*ha;
4262 	uint32_t		reg_addr;
4263 
4264 	*bad_wr = NULL;
4265 	ha = dev->ha;
4266 
4267 	QL_DPRINT12(ha, "exit[ibqp, wr, bad_wr] = [%p, %p, %p]\n",
4268 		ibqp, wr, bad_wr);
4269 
4270 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
4271 		return -EINVAL;
4272 
4273 	if (qp->qp_type == IB_QPT_GSI)
4274 		return qlnxr_gsi_post_send(ibqp, wr, bad_wr);
4275 
4276 	spin_lock_irqsave(&qp->q_lock, flags);
4277 
4278 	if (QLNX_IS_ROCE(dev) && (qp->state != ECORE_ROCE_QP_STATE_RTS) &&
4279 	    (qp->state != ECORE_ROCE_QP_STATE_ERR) &&
4280 	    (qp->state != ECORE_ROCE_QP_STATE_SQD)) {
4281 		spin_unlock_irqrestore(&qp->q_lock, flags);
4282 		*bad_wr = wr;
4283 		QL_DPRINT11(ha, "QP in wrong state! QP icid=0x%x state %d\n",
4284 			qp->icid, qp->state);
4285 		return -EINVAL;
4286 	}
4287 
4288 	if (!wr) {
4289 		QL_DPRINT11(ha, "Got an empty post send???\n");
4290 	}
4291 
4292 	while (wr) {
4293 		struct rdma_sq_common_wqe	*wqe;
4294 		struct rdma_sq_send_wqe		*swqe;
4295 		struct rdma_sq_send_wqe		*swqe2;
4296 		struct rdma_sq_rdma_wqe_1st	*rwqe;
4297 		struct rdma_sq_rdma_wqe_2nd	*rwqe2;
4298 		struct rdma_sq_local_inv_wqe	*iwqe;
4299 		struct rdma_sq_atomic_wqe	*awqe1;
4300 		struct rdma_sq_atomic_wqe	*awqe2;
4301 		struct rdma_sq_atomic_wqe	*awqe3;
4302 		struct rdma_sq_fmr_wqe_1st	*fwqe1;
4303 
4304 		if (!qlnxr_can_post_send(qp, wr)) {
4305 			status = -ENOMEM;
4306 			*bad_wr = wr;
4307 			break;
4308 		}
4309 
4310 		wqe = ecore_chain_produce(&qp->sq.pbl);
4311 
4312 		qp->wqe_wr_id[qp->sq.prod].signaled =
4313 			!!(wr->send_flags & IB_SEND_SIGNALED) || qp->signaled;
4314 
4315 		/* common fields */
4316 		wqe->flags = 0;
4317 		wqe->flags |= (RDMA_SQ_SEND_WQE_COMP_FLG_MASK <<
4318 				RDMA_SQ_SEND_WQE_COMP_FLG_SHIFT);
4319 
4320 		SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_SE_FLG, \
4321 			!!(wr->send_flags & IB_SEND_SOLICITED));
4322 
4323 		comp = (!!(wr->send_flags & IB_SEND_SIGNALED)) ||
4324 				(qp->signaled);
4325 
4326 		SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_COMP_FLG, comp);
4327 		SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_RD_FENCE_FLG,  \
4328 			!!(wr->send_flags & IB_SEND_FENCE));
4329 
4330 		wqe->prev_wqe_size = qp->prev_wqe_size;
4331 
4332 		qp->wqe_wr_id[qp->sq.prod].opcode = qlnxr_ib_to_wc_opcode(wr->opcode);
4333 
4334 		switch (wr->opcode) {
4335 		case IB_WR_SEND_WITH_IMM:
4336 
4337 			wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_IMM;
4338 			swqe = (struct rdma_sq_send_wqe *)wqe;
4339 			swqe->wqe_size = 2;
4340 			swqe2 = (struct rdma_sq_send_wqe *)
4341 					ecore_chain_produce(&qp->sq.pbl);
4342 			swqe->inv_key_or_imm_data =
4343 				cpu_to_le32(wr->ex.imm_data);
4344 			swqe->length = cpu_to_le32(
4345 						qlnxr_prepare_sq_send_data(dev,
4346 							qp, swqe, swqe2, wr,
4347 							bad_wr));
4348 
4349 			qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
4350 			qp->prev_wqe_size = swqe->wqe_size;
4351 			qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
4352 
4353 			QL_DPRINT12(ha, "SEND w/ IMM length = %d imm data=%x\n",
4354 				swqe->length, wr->ex.imm_data);
4355 
4356 			break;
4357 
4358 		case IB_WR_SEND:
4359 
4360 			wqe->req_type = RDMA_SQ_REQ_TYPE_SEND;
4361 			swqe = (struct rdma_sq_send_wqe *)wqe;
4362 
4363 			swqe->wqe_size = 2;
4364 			swqe2 = (struct rdma_sq_send_wqe *)
4365 					ecore_chain_produce(&qp->sq.pbl);
4366 			swqe->length = cpu_to_le32(
4367 						qlnxr_prepare_sq_send_data(dev,
4368 							qp, swqe, swqe2, wr,
4369 							bad_wr));
4370 			qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
4371 			qp->prev_wqe_size = swqe->wqe_size;
4372 			qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
4373 
4374 			QL_DPRINT12(ha, "SEND w/o IMM length = %d\n",
4375 				swqe->length);
4376 
4377 			break;
4378 
4379 		case IB_WR_SEND_WITH_INV:
4380 
4381 			wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_INVALIDATE;
4382 			swqe = (struct rdma_sq_send_wqe *)wqe;
4383 			swqe2 = (struct rdma_sq_send_wqe *)
4384 					ecore_chain_produce(&qp->sq.pbl);
4385 			swqe->wqe_size = 2;
4386 			swqe->inv_key_or_imm_data =
4387 				cpu_to_le32(wr->ex.invalidate_rkey);
4388 			swqe->length = cpu_to_le32(qlnxr_prepare_sq_send_data(dev,
4389 						qp, swqe, swqe2, wr, bad_wr));
4390 			qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
4391 			qp->prev_wqe_size = swqe->wqe_size;
4392 			qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
4393 
4394 			QL_DPRINT12(ha, "SEND w INVALIDATE length = %d\n",
4395 				swqe->length);
4396 			break;
4397 
4398 		case IB_WR_RDMA_WRITE_WITH_IMM:
4399 
4400 			wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR_WITH_IMM;
4401 			rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
4402 
4403 			rwqe->wqe_size = 2;
4404 			rwqe->imm_data = htonl(cpu_to_le32(wr->ex.imm_data));
4405 			rwqe2 = (struct rdma_sq_rdma_wqe_2nd *)
4406 					ecore_chain_produce(&qp->sq.pbl);
4407 			rwqe->length = cpu_to_le32(qlnxr_prepare_sq_rdma_data(dev,
4408 						qp, rwqe, rwqe2, wr, bad_wr));
4409 			qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
4410 			qp->prev_wqe_size = rwqe->wqe_size;
4411 			qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
4412 
4413 			QL_DPRINT12(ha,
4414 				"RDMA WRITE w/ IMM length = %d imm data=%x\n",
4415 				rwqe->length, rwqe->imm_data);
4416 
4417 			break;
4418 
4419 		case IB_WR_RDMA_WRITE:
4420 
4421 			wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR;
4422 			rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
4423 
4424 			rwqe->wqe_size = 2;
4425 			rwqe2 = (struct rdma_sq_rdma_wqe_2nd *)
4426 					ecore_chain_produce(&qp->sq.pbl);
4427 			rwqe->length = cpu_to_le32(qlnxr_prepare_sq_rdma_data(dev,
4428 						qp, rwqe, rwqe2, wr, bad_wr));
4429 			qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
4430 			qp->prev_wqe_size = rwqe->wqe_size;
4431 			qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
4432 
4433 			QL_DPRINT12(ha,
4434 				"RDMA WRITE w/o IMM length = %d\n",
4435 				rwqe->length);
4436 
4437 			break;
4438 
4439 		case IB_WR_RDMA_READ_WITH_INV:
4440 
4441 			QL_DPRINT12(ha,
4442 				"RDMA READ WITH INVALIDATE not supported\n");
4443 
4444 			*bad_wr = wr;
4445 			rc = -EINVAL;
4446 
4447 			break;
4448 
4449 		case IB_WR_RDMA_READ:
4450 
4451 			wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_RD;
4452 			rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
4453 
4454 			rwqe->wqe_size = 2;
4455 			rwqe2 = (struct rdma_sq_rdma_wqe_2nd *)
4456 					ecore_chain_produce(&qp->sq.pbl);
4457 			rwqe->length = cpu_to_le32(qlnxr_prepare_sq_rdma_data(dev,
4458 						qp, rwqe, rwqe2, wr, bad_wr));
4459 
4460 			qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
4461 			qp->prev_wqe_size = rwqe->wqe_size;
4462 			qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
4463 
4464 			QL_DPRINT12(ha, "RDMA READ length = %d\n",
4465 				rwqe->length);
4466 
4467 			break;
4468 
4469 		case IB_WR_ATOMIC_CMP_AND_SWP:
4470 		case IB_WR_ATOMIC_FETCH_AND_ADD:
4471 
4472 			QL_DPRINT12(ha,
4473 				"ATOMIC operation = %s\n",
4474 				((wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) ?
4475 					"IB_WR_ATOMIC_CMP_AND_SWP" :
4476 					"IB_WR_ATOMIC_FETCH_AND_ADD"));
4477 
4478 			awqe1 = (struct rdma_sq_atomic_wqe *)wqe;
4479 			awqe1->prev_wqe_size = 4;
4480 
4481 			awqe2 = (struct rdma_sq_atomic_wqe *)
4482 					ecore_chain_produce(&qp->sq.pbl);
4483 
4484 			TYPEPTR_ADDR_SET(awqe2, remote_va, \
4485 				atomic_wr(wr)->remote_addr);
4486 
4487 			awqe2->r_key = cpu_to_le32(atomic_wr(wr)->rkey);
4488 
4489 			awqe3 = (struct rdma_sq_atomic_wqe *)
4490 					ecore_chain_produce(&qp->sq.pbl);
4491 
4492 			if (wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
4493 				wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_ADD;
4494 				TYPEPTR_ADDR_SET(awqe3, swap_data,
4495 						 atomic_wr(wr)->compare_add);
4496 			} else {
4497 				wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_CMP_AND_SWAP;
4498 				TYPEPTR_ADDR_SET(awqe3, swap_data,
4499 						 atomic_wr(wr)->swap);
4500 				TYPEPTR_ADDR_SET(awqe3, cmp_data,
4501 						 atomic_wr(wr)->compare_add);
4502 			}
4503 
4504 			qlnxr_prepare_sq_sges(dev, qp, NULL, wr);
4505 
4506 			qp->wqe_wr_id[qp->sq.prod].wqe_size = awqe1->prev_wqe_size;
4507 			qp->prev_wqe_size = awqe1->prev_wqe_size;
4508 
4509 			break;
4510 
4511 		case IB_WR_LOCAL_INV:
4512 
4513 			QL_DPRINT12(ha,
4514 				"INVALIDATE length (IB_WR_LOCAL_INV)\n");
4515 
4516 			iwqe = (struct rdma_sq_local_inv_wqe *)wqe;
4517 			iwqe->prev_wqe_size = 1;
4518 
4519 			iwqe->req_type = RDMA_SQ_REQ_TYPE_LOCAL_INVALIDATE;
4520 			iwqe->inv_l_key = wr->ex.invalidate_rkey;
4521 			qp->wqe_wr_id[qp->sq.prod].wqe_size = iwqe->prev_wqe_size;
4522 			qp->prev_wqe_size = iwqe->prev_wqe_size;
4523 
4524 			break;
4525 
4526 		case IB_WR_REG_MR:
4527 
4528 			QL_DPRINT12(ha, "IB_WR_REG_MR\n");
4529 
4530 			wqe->req_type = RDMA_SQ_REQ_TYPE_FAST_MR;
4531 			fwqe1 = (struct rdma_sq_fmr_wqe_1st *)wqe;
4532 			fwqe1->wqe_size = 2;
4533 
4534 			rc = qlnxr_prepare_reg(qp, fwqe1, reg_wr(wr));
4535 			if (rc) {
4536 				QL_DPRINT11(ha, "IB_WR_REG_MR failed rc=%d\n", rc);
4537 				*bad_wr = wr;
4538 				break;
4539 			}
4540 
4541 			qp->wqe_wr_id[qp->sq.prod].wqe_size = fwqe1->wqe_size;
4542 			qp->prev_wqe_size = fwqe1->wqe_size;
4543 
4544 			break;
4545 
4546 		default:
4547 
4548 			QL_DPRINT12(ha, "Invalid Opcode 0x%x!\n", wr->opcode);
4549 
4550 			rc = -EINVAL;
4551 			*bad_wr = wr;
4552 			break;
4553 		}
4554 
4555 		if (*bad_wr) {
4556 			/*
4557 			 * restore prod to its position before this WR was processed
4558 			 */
4559 			ecore_chain_set_prod(&qp->sq.pbl,
4560 			     le16_to_cpu(qp->sq.db_data.data.value),
4561 			     wqe);
4562 			/* restore prev_wqe_size */
4563 			qp->prev_wqe_size = wqe->prev_wqe_size;
4564 			status = rc;
4565 
4566 			QL_DPRINT12(ha, "failed *bad_wr = %p\n", *bad_wr);
4567 			break; /* out of the loop */
4568 		}
4569 
4570 		qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id;
4571 
4572 		qlnxr_inc_sw_prod(&qp->sq);
4573 
4574 		qp->sq.db_data.data.value++;
4575 
4576 		wr = wr->next;
4577 	}
4578 
4579 	/* Trigger doorbell
4580 	 * If there was a failure in the first WR then it will be triggered in
4581 	 * vane. However this is not harmful (as long as the producer value is
4582 	 * unchanged). For performance reasons we avoid checking for this
4583 	 * redundant doorbell.
4584 	 */
4585 	wmb();
4586 	//writel(qp->sq.db_data.raw, qp->sq.db);
4587 
4588 	reg_addr = (uint32_t)((uint8_t *)qp->sq.db - (uint8_t *)ha->cdev.doorbells);
4589         bus_write_4(ha->pci_dbells, reg_addr, qp->sq.db_data.raw);
4590         bus_barrier(ha->pci_dbells,  0, 0, BUS_SPACE_BARRIER_READ);
4591 
4592 	mmiowb();
4593 
4594 	spin_unlock_irqrestore(&qp->q_lock, flags);
4595 
4596 	QL_DPRINT12(ha, "exit[ibqp, wr, bad_wr] = [%p, %p, %p]\n",
4597 		ibqp, wr, bad_wr);
4598 
4599 	return status;
4600 }
4601 
4602 static u32
4603 qlnxr_srq_elem_left(struct qlnxr_srq_hwq_info *hw_srq)
4604 {
4605 	u32 used;
4606 
4607 	/* Calculate number of elements used based on producer
4608 	 * count and consumer count and subtract it from max
4609 	 * work request supported so that we get elements left.
4610 	 */
4611 	used = hw_srq->wr_prod_cnt - hw_srq->wr_cons_cnt;
4612 
4613 	return hw_srq->max_wr - used;
4614 }
4615 
4616 int
4617 qlnxr_post_recv(struct ib_qp *ibqp,
4618 	const struct ib_recv_wr *wr,
4619 	const struct ib_recv_wr **bad_wr)
4620 {
4621  	struct qlnxr_qp		*qp = get_qlnxr_qp(ibqp);
4622 	struct qlnxr_dev	*dev = qp->dev;
4623 	unsigned long		flags;
4624 	int			status = 0;
4625 	qlnx_host_t		*ha;
4626 	uint32_t		reg_addr;
4627 
4628 	ha = dev->ha;
4629 
4630 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
4631 		return -EINVAL;
4632 
4633 	QL_DPRINT12(ha, "enter\n");
4634 
4635 	if (qp->qp_type == IB_QPT_GSI) {
4636 		QL_DPRINT12(ha, "(qp->qp_type = IB_QPT_GSI)\n");
4637 		return qlnxr_gsi_post_recv(ibqp, wr, bad_wr);
4638 	}
4639 
4640 	if (qp->srq) {
4641 		QL_DPRINT11(ha, "qp->srq [%p]"
4642 			" QP is associated with SRQ, cannot post RQ buffers\n",
4643 			qp->srq);
4644 		return -EINVAL;
4645 	}
4646 
4647 	spin_lock_irqsave(&qp->q_lock, flags);
4648 
4649 	if (qp->state == ECORE_ROCE_QP_STATE_RESET) {
4650 		spin_unlock_irqrestore(&qp->q_lock, flags);
4651 		*bad_wr = wr;
4652 
4653 		QL_DPRINT11(ha, "qp->qp_type = ECORE_ROCE_QP_STATE_RESET\n");
4654 
4655 		return -EINVAL;
4656 	}
4657 
4658 	while (wr) {
4659 		int i;
4660 
4661 		if ((ecore_chain_get_elem_left_u32(&qp->rq.pbl) <
4662 			QLNXR_MAX_RQE_ELEMENTS_PER_RQE) ||
4663 			(wr->num_sge > qp->rq.max_sges)) {
4664 			status = -ENOMEM;
4665 			*bad_wr = wr;
4666 			break;
4667 		}
4668 		for (i = 0; i < wr->num_sge; i++) {
4669 			u32 flags = 0;
4670 			struct rdma_rq_sge *rqe = ecore_chain_produce(&qp->rq.pbl);
4671 
4672 			/* first one must include the number of SGE in the list */
4673 			if (!i)
4674 				SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, wr->num_sge);
4675 
4676 			SET_FIELD(flags, RDMA_RQ_SGE_L_KEY, wr->sg_list[i].lkey);
4677 
4678 			RQ_SGE_SET(rqe, wr->sg_list[i].addr, \
4679 				wr->sg_list[i].length, flags);
4680 		}
4681 		/* Special case of no sges. FW requires between 1-4 sges...
4682 		 * in this case we need to post 1 sge with length zero. this is
4683 		 * because rdma write with immediate consumes an RQ. */
4684 		if (!wr->num_sge) {
4685 			u32 flags = 0;
4686 			struct rdma_rq_sge *rqe = ecore_chain_produce(&qp->rq.pbl);
4687 
4688 			/* first one must include the number of SGE in the list */
4689 			SET_FIELD(flags, RDMA_RQ_SGE_L_KEY, 0);
4690 			SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, 1);
4691 
4692 			//RQ_SGE_SET(rqe, 0, 0, flags);
4693 			rqe->addr.hi = 0;
4694 			rqe->addr.lo = 0;
4695 
4696 			rqe->length = 0;
4697 			rqe->flags = cpu_to_le32(flags);
4698 
4699 			i = 1;
4700 		}
4701 
4702 		qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id;
4703 		qp->rqe_wr_id[qp->rq.prod].wqe_size = i;
4704 
4705 		qlnxr_inc_sw_prod(&qp->rq);
4706 
4707 		wmb();
4708 
4709 		qp->rq.db_data.data.value++;
4710 
4711 	//	writel(qp->rq.db_data.raw, qp->rq.db);
4712 		mmiowb();
4713 	//	if (QLNX_IS_IWARP(dev)) {
4714 	//		writel(qp->rq.iwarp_db2_data.raw, qp->rq.iwarp_db2);
4715 	//		mmiowb(); /* for second doorbell */
4716 	//	}
4717 
4718 		reg_addr = (uint32_t)((uint8_t *)qp->rq.db -
4719 				(uint8_t *)ha->cdev.doorbells);
4720 
4721 		bus_write_4(ha->pci_dbells, reg_addr, qp->rq.db_data.raw);
4722 		bus_barrier(ha->pci_dbells,  0, 0, BUS_SPACE_BARRIER_READ);
4723 
4724 		if (QLNX_IS_IWARP(dev)) {
4725 			reg_addr = (uint32_t)((uint8_t *)qp->rq.iwarp_db2 -
4726 						(uint8_t *)ha->cdev.doorbells);
4727 			bus_write_4(ha->pci_dbells, reg_addr, \
4728 				qp->rq.iwarp_db2_data.raw);
4729 			bus_barrier(ha->pci_dbells,  0, 0, \
4730 				BUS_SPACE_BARRIER_READ);
4731 		}
4732 
4733 		wr = wr->next;
4734 	}
4735 
4736 	spin_unlock_irqrestore(&qp->q_lock, flags);
4737 
4738 	QL_DPRINT12(ha, "exit status = 0x%x\n", status);
4739 
4740 	return status;
4741 }
4742 
4743 /* In fmr we need to increase the number of fmr completed counter for the fmr
4744  * algorithm determining whether we can free a pbl or not.
4745  * we need to perform this whether the work request was signaled or not. for
4746  * this purpose we call this function from the condition that checks if a wr
4747  * should be skipped, to make sure we don't miss it ( possibly this fmr
4748  * operation was not signalted)
4749  */
4750 static inline void
4751 qlnxr_chk_if_fmr(struct qlnxr_qp *qp)
4752 {
4753 
4754 	if (qp->wqe_wr_id[qp->sq.cons].opcode == IB_WC_REG_MR)
4755 		qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
4756 }
4757 
4758 static int
4759 process_req(struct qlnxr_dev *dev,
4760 	struct qlnxr_qp *qp,
4761 	struct qlnxr_cq *cq,
4762 	int num_entries,
4763 	struct ib_wc *wc,
4764 	u16 hw_cons,
4765 	enum ib_wc_status status,
4766 	int force)
4767 {
4768 	u16		cnt = 0;
4769 	qlnx_host_t	*ha = dev->ha;
4770 
4771 	QL_DPRINT12(ha, "enter\n");
4772 
4773 	while (num_entries && qp->sq.wqe_cons != hw_cons) {
4774 		if (!qp->wqe_wr_id[qp->sq.cons].signaled && !force) {
4775 			qlnxr_chk_if_fmr(qp);
4776 			/* skip WC */
4777 			goto next_cqe;
4778 		}
4779 
4780 		/* fill WC */
4781 		wc->status = status;
4782 		wc->vendor_err = 0;
4783 		wc->wc_flags = 0;
4784 		wc->src_qp = qp->id;
4785 		wc->qp = &qp->ibqp;
4786 
4787 		// common section
4788 		wc->wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id;
4789 		wc->opcode = qp->wqe_wr_id[qp->sq.cons].opcode;
4790 
4791 		switch (wc->opcode) {
4792 		case IB_WC_RDMA_WRITE:
4793 
4794 			wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len;
4795 
4796 			QL_DPRINT12(ha,
4797 				"opcode = IB_WC_RDMA_WRITE bytes = %d\n",
4798 				qp->wqe_wr_id[qp->sq.cons].bytes_len);
4799 			break;
4800 
4801 		case IB_WC_COMP_SWAP:
4802 		case IB_WC_FETCH_ADD:
4803 			wc->byte_len = 8;
4804 			break;
4805 
4806 		case IB_WC_REG_MR:
4807 			qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
4808 			break;
4809 
4810 		case IB_WC_RDMA_READ:
4811 		case IB_WC_SEND:
4812 
4813 			QL_DPRINT12(ha, "opcode = 0x%x \n", wc->opcode);
4814 			break;
4815 		default:
4816 			;//DP_ERR("TBD ERROR");
4817 		}
4818 
4819 		num_entries--;
4820 		wc++;
4821 		cnt++;
4822 next_cqe:
4823 		while (qp->wqe_wr_id[qp->sq.cons].wqe_size--)
4824 			ecore_chain_consume(&qp->sq.pbl);
4825 		qlnxr_inc_sw_cons(&qp->sq);
4826 	}
4827 
4828 	QL_DPRINT12(ha, "exit cnt = 0x%x\n", cnt);
4829 	return cnt;
4830 }
4831 
4832 static int
4833 qlnxr_poll_cq_req(struct qlnxr_dev *dev,
4834 	struct qlnxr_qp *qp,
4835 	struct qlnxr_cq *cq,
4836 	int num_entries,
4837 	struct ib_wc *wc,
4838 	struct rdma_cqe_requester *req)
4839 {
4840 	int		cnt = 0;
4841 	qlnx_host_t	*ha = dev->ha;
4842 
4843 	QL_DPRINT12(ha, "enter req->status = 0x%x\n", req->status);
4844 
4845 	switch (req->status) {
4846 	case RDMA_CQE_REQ_STS_OK:
4847 
4848 		cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
4849 			IB_WC_SUCCESS, 0);
4850 		break;
4851 
4852 	case RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR:
4853 
4854 		if (qp->state != ECORE_ROCE_QP_STATE_ERR)
4855 		cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
4856 				  IB_WC_WR_FLUSH_ERR, 1);
4857 		break;
4858 
4859 	default: /* other errors case */
4860 
4861 		/* process all WQE before the cosumer */
4862 		qp->state = ECORE_ROCE_QP_STATE_ERR;
4863 		cnt = process_req(dev, qp, cq, num_entries, wc,
4864 				req->sq_cons - 1, IB_WC_SUCCESS, 0);
4865 		wc += cnt;
4866 		/* if we have extra WC fill it with actual error info */
4867 
4868 		if (cnt < num_entries) {
4869 			enum ib_wc_status wc_status;
4870 
4871 			switch (req->status) {
4872 			case 	RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR:
4873 				wc_status = IB_WC_BAD_RESP_ERR;
4874 				break;
4875 			case 	RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR:
4876 				wc_status = IB_WC_LOC_LEN_ERR;
4877 				break;
4878 			case    RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR:
4879 				wc_status = IB_WC_LOC_QP_OP_ERR;
4880 				break;
4881 			case    RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR:
4882 				wc_status = IB_WC_LOC_PROT_ERR;
4883 				break;
4884 			case    RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR:
4885 				wc_status = IB_WC_MW_BIND_ERR;
4886 				break;
4887 			case    RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR:
4888 				wc_status = IB_WC_REM_INV_REQ_ERR;
4889 				break;
4890 			case    RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR:
4891 				wc_status = IB_WC_REM_ACCESS_ERR;
4892 				break;
4893 			case    RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR:
4894 				wc_status = IB_WC_REM_OP_ERR;
4895 				break;
4896 			case    RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR:
4897 				wc_status = IB_WC_RNR_RETRY_EXC_ERR;
4898 				break;
4899 			case    RDMA_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR:
4900 				wc_status = IB_WC_RETRY_EXC_ERR;
4901 				break;
4902 			default:
4903 				wc_status = IB_WC_GENERAL_ERR;
4904 			}
4905 
4906 			cnt += process_req(dev, qp, cq, 1, wc, req->sq_cons,
4907 					wc_status, 1 /* force use of WC */);
4908 		}
4909 	}
4910 
4911 	QL_DPRINT12(ha, "exit cnt = %d\n", cnt);
4912 	return cnt;
4913 }
4914 
4915 static void
4916 __process_resp_one(struct qlnxr_dev *dev,
4917 	struct qlnxr_qp *qp,
4918 	struct qlnxr_cq *cq,
4919 	struct ib_wc *wc,
4920 	struct rdma_cqe_responder *resp,
4921 	u64 wr_id)
4922 {
4923 	enum ib_wc_status	wc_status = IB_WC_SUCCESS;
4924 	qlnx_host_t		*ha = dev->ha;
4925 
4926 	QL_DPRINT12(ha, "enter qp = %p resp->status = 0x%x\n",
4927 		qp, resp->status);
4928 
4929 	wc->opcode = IB_WC_RECV;
4930 	wc->wc_flags = 0;
4931 
4932 	switch (resp->status) {
4933 	case RDMA_CQE_RESP_STS_LOCAL_ACCESS_ERR:
4934 		wc_status = IB_WC_LOC_ACCESS_ERR;
4935 		break;
4936 
4937 	case RDMA_CQE_RESP_STS_LOCAL_LENGTH_ERR:
4938 		wc_status = IB_WC_LOC_LEN_ERR;
4939 		break;
4940 
4941 	case RDMA_CQE_RESP_STS_LOCAL_QP_OPERATION_ERR:
4942 		wc_status = IB_WC_LOC_QP_OP_ERR;
4943 		break;
4944 
4945 	case RDMA_CQE_RESP_STS_LOCAL_PROTECTION_ERR:
4946 		wc_status = IB_WC_LOC_PROT_ERR;
4947 		break;
4948 
4949 	case RDMA_CQE_RESP_STS_MEMORY_MGT_OPERATION_ERR:
4950 		wc_status = IB_WC_MW_BIND_ERR;
4951 		break;
4952 
4953 	case RDMA_CQE_RESP_STS_REMOTE_INVALID_REQUEST_ERR:
4954 		wc_status = IB_WC_REM_INV_RD_REQ_ERR;
4955 		break;
4956 
4957 	case RDMA_CQE_RESP_STS_OK:
4958 		if (resp->flags & QLNXR_RESP_IMM) {
4959 			wc->ex.imm_data =
4960 				le32_to_cpu(resp->imm_data_or_inv_r_Key);
4961 			wc->wc_flags |= IB_WC_WITH_IMM;
4962 
4963 			if (resp->flags & QLNXR_RESP_RDMA)
4964 				wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
4965 
4966 			if (resp->flags & QLNXR_RESP_INV) {
4967 				QL_DPRINT11(ha,
4968 					"Invalid flags QLNXR_RESP_INV [0x%x]"
4969 					"qp = %p qp->id = 0x%x cq = %p"
4970 					" cq->icid = 0x%x\n",
4971 					resp->flags, qp, qp->id, cq, cq->icid );
4972 			}
4973 		} else if (resp->flags & QLNXR_RESP_INV) {
4974 			wc->ex.imm_data =
4975 				le32_to_cpu(resp->imm_data_or_inv_r_Key);
4976 			wc->wc_flags |= IB_WC_WITH_INVALIDATE;
4977 
4978 			if (resp->flags & QLNXR_RESP_RDMA) {
4979 				QL_DPRINT11(ha,
4980 					"Invalid flags QLNXR_RESP_RDMA [0x%x]"
4981 					"qp = %p qp->id = 0x%x cq = %p"
4982 					" cq->icid = 0x%x\n",
4983 					resp->flags, qp, qp->id, cq, cq->icid );
4984 			}
4985 		} else if (resp->flags & QLNXR_RESP_RDMA) {
4986 			QL_DPRINT11(ha, "Invalid flags QLNXR_RESP_RDMA [0x%x]"
4987 				"qp = %p qp->id = 0x%x cq = %p cq->icid = 0x%x\n",
4988 				resp->flags, qp, qp->id, cq, cq->icid );
4989 		}
4990 		break;
4991 
4992 	default:
4993 		wc_status = IB_WC_GENERAL_ERR;
4994 	}
4995 
4996 	/* fill WC */
4997 	wc->status = wc_status;
4998 	wc->vendor_err = 0;
4999 	wc->src_qp = qp->id;
5000 	wc->qp = &qp->ibqp;
5001 	wc->wr_id = wr_id;
5002 
5003 	QL_DPRINT12(ha, "exit status = 0x%x\n", wc_status);
5004 
5005 	return;
5006 }
5007 
5008 static int
5009 process_resp_one_srq(struct qlnxr_dev *dev,
5010 	struct qlnxr_qp *qp,
5011 	struct qlnxr_cq *cq,
5012 	struct ib_wc *wc,
5013 	struct rdma_cqe_responder *resp)
5014 {
5015 	struct qlnxr_srq	*srq = qp->srq;
5016 	u64			wr_id;
5017 	qlnx_host_t		*ha = dev->ha;
5018 
5019 	QL_DPRINT12(ha, "enter\n");
5020 
5021 	wr_id = HILO_U64(resp->srq_wr_id.hi, resp->srq_wr_id.lo);
5022 
5023 	if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
5024 		wc->status = IB_WC_WR_FLUSH_ERR;
5025 		wc->vendor_err = 0;
5026 		wc->wr_id = wr_id;
5027 		wc->byte_len = 0;
5028 		wc->src_qp = qp->id;
5029 		wc->qp = &qp->ibqp;
5030 		wc->wr_id = wr_id;
5031 	} else {
5032 		__process_resp_one(dev, qp, cq, wc, resp, wr_id);
5033 	}
5034 
5035 	/* PBL is maintained in case of WR granularity.
5036 	 * So increment WR consumer after consuming WR
5037 	 */
5038 	srq->hw_srq.wr_cons_cnt++;
5039 
5040 	QL_DPRINT12(ha, "exit\n");
5041 	return 1;
5042 }
5043 
5044 static int
5045 process_resp_one(struct qlnxr_dev *dev,
5046 	struct qlnxr_qp *qp,
5047 	struct qlnxr_cq *cq,
5048 	struct ib_wc *wc,
5049 	struct rdma_cqe_responder *resp)
5050 {
5051 	qlnx_host_t	*ha = dev->ha;
5052 	u64		wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
5053 
5054 	QL_DPRINT12(ha, "enter\n");
5055 
5056 	__process_resp_one(dev, qp, cq, wc, resp, wr_id);
5057 
5058 	while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
5059 		ecore_chain_consume(&qp->rq.pbl);
5060 	qlnxr_inc_sw_cons(&qp->rq);
5061 
5062 	QL_DPRINT12(ha, "exit\n");
5063 	return 1;
5064 }
5065 
5066 static int
5067 process_resp_flush(struct qlnxr_qp *qp,
5068 	int num_entries,
5069 	struct ib_wc *wc,
5070 	u16 hw_cons)
5071 {
5072 	u16		cnt = 0;
5073 	qlnx_host_t	*ha = qp->dev->ha;
5074 
5075 	QL_DPRINT12(ha, "enter\n");
5076 
5077 	while (num_entries && qp->rq.wqe_cons != hw_cons) {
5078 		/* fill WC */
5079 		wc->status = IB_WC_WR_FLUSH_ERR;
5080 		wc->vendor_err = 0;
5081 		wc->wc_flags = 0;
5082 		wc->src_qp = qp->id;
5083 		wc->byte_len = 0;
5084 		wc->wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
5085 		wc->qp = &qp->ibqp;
5086 		num_entries--;
5087 		wc++;
5088 		cnt++;
5089 		while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
5090 			ecore_chain_consume(&qp->rq.pbl);
5091 		qlnxr_inc_sw_cons(&qp->rq);
5092 	}
5093 
5094 	QL_DPRINT12(ha, "exit cnt = 0x%x\n", cnt);
5095 	return cnt;
5096 }
5097 
5098 static void
5099 try_consume_resp_cqe(struct qlnxr_cq *cq,
5100 	struct qlnxr_qp *qp,
5101 	struct rdma_cqe_responder *resp,
5102 	int *update)
5103 {
5104 	if (le16_to_cpu(resp->rq_cons) == qp->rq.wqe_cons) {
5105 		consume_cqe(cq);
5106 		*update |= 1;
5107 	}
5108 }
5109 
5110 static int
5111 qlnxr_poll_cq_resp_srq(struct qlnxr_dev *dev,
5112 	struct qlnxr_qp *qp,
5113 	struct qlnxr_cq *cq,
5114 	int num_entries,
5115 	struct ib_wc *wc,
5116 	struct rdma_cqe_responder *resp,
5117 	int *update)
5118 {
5119 	int		cnt;
5120 	qlnx_host_t	*ha = dev->ha;
5121 
5122 	QL_DPRINT12(ha, "enter\n");
5123 
5124 	cnt = process_resp_one_srq(dev, qp, cq, wc, resp);
5125 	consume_cqe(cq);
5126 	*update |= 1;
5127 
5128 	QL_DPRINT12(ha, "exit cnt = 0x%x\n", cnt);
5129 	return cnt;
5130 }
5131 
5132 static int
5133 qlnxr_poll_cq_resp(struct qlnxr_dev *dev,
5134 	struct qlnxr_qp *qp,
5135 	struct qlnxr_cq *cq,
5136 	int num_entries,
5137 	struct ib_wc *wc,
5138 	struct rdma_cqe_responder *resp,
5139 	int *update)
5140 {
5141 	int		cnt;
5142 	qlnx_host_t	*ha = dev->ha;
5143 
5144 	QL_DPRINT12(ha, "enter\n");
5145 
5146 	if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
5147 		cnt = process_resp_flush(qp, num_entries, wc,
5148 				resp->rq_cons);
5149 		try_consume_resp_cqe(cq, qp, resp, update);
5150 	} else {
5151 		cnt = process_resp_one(dev, qp, cq, wc, resp);
5152 		consume_cqe(cq);
5153 		*update |= 1;
5154 	}
5155 
5156 	QL_DPRINT12(ha, "exit cnt = 0x%x\n", cnt);
5157 	return cnt;
5158 }
5159 
5160 static void
5161 try_consume_req_cqe(struct qlnxr_cq *cq, struct qlnxr_qp *qp,
5162 	struct rdma_cqe_requester *req, int *update)
5163 {
5164 	if (le16_to_cpu(req->sq_cons) == qp->sq.wqe_cons) {
5165 		consume_cqe(cq);
5166 		*update |= 1;
5167 	}
5168 }
5169 
5170 static void
5171 doorbell_cq(struct qlnxr_dev *dev, struct qlnxr_cq *cq, u32 cons, u8 flags)
5172 {
5173 	uint64_t	reg_addr;
5174 	qlnx_host_t	*ha = dev->ha;
5175 
5176 	QL_DPRINT12(ha, "enter\n");
5177 
5178 	wmb();
5179 	cq->db.data.agg_flags = flags;
5180 	cq->db.data.value = cpu_to_le32(cons);
5181 
5182 	reg_addr = (uint64_t)((uint8_t *)cq->db_addr -
5183 				(uint8_t *)(ha->cdev.doorbells));
5184 
5185 	bus_write_8(ha->pci_dbells, reg_addr, cq->db.raw);
5186 	bus_barrier(ha->pci_dbells,  0, 0, BUS_SPACE_BARRIER_READ);
5187 
5188 	QL_DPRINT12(ha, "exit\n");
5189 	return;
5190 
5191 //#ifdef __LP64__
5192 //	writeq(cq->db.raw, cq->db_addr);
5193 //#else
5194 	/* Note that since the FW allows 64 bit write only, in 32bit systems
5195 	 * the value of db_addr must be low enough. This is currently not
5196 	 * enforced.
5197 	 */
5198 //	writel(cq->db.raw & 0xffffffff, cq->db_addr);
5199 //	mmiowb();
5200 //#endif
5201 }
5202 
5203 static int
5204 is_valid_cqe(struct qlnxr_cq *cq, union rdma_cqe *cqe)
5205 {
5206 	struct rdma_cqe_requester *resp_cqe = &cqe->req;
5207 	return (resp_cqe->flags & RDMA_RESIZE_CQ_RAMROD_DATA_TOGGLE_BIT_MASK) ==
5208 			cq->pbl_toggle;
5209 }
5210 
5211 int
5212 qlnxr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
5213 {
5214 	struct qlnxr_cq	*cq = get_qlnxr_cq(ibcq);
5215 	struct qlnxr_dev *dev = get_qlnxr_dev((ibcq->device));
5216 	int		done = 0;
5217 	union rdma_cqe	*cqe = cq->latest_cqe;
5218 	int 		update = 0;
5219 	u32		old_cons, new_cons;
5220 	unsigned long	flags;
5221 	qlnx_host_t	*ha = dev->ha;
5222 
5223 	QL_DPRINT12(ha, "enter\n");
5224 
5225 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
5226 		return -EINVAL;
5227 
5228 	if (cq->destroyed) {
5229 		QL_DPRINT11(ha, "called after destroy for cq %p (icid=%d)\n",
5230 			cq, cq->icid);
5231 		return 0;
5232 	}
5233 
5234 	if (cq->cq_type == QLNXR_CQ_TYPE_GSI)
5235 		return qlnxr_gsi_poll_cq(ibcq, num_entries, wc);
5236 
5237 	spin_lock_irqsave(&cq->cq_lock, flags);
5238 
5239 	old_cons = ecore_chain_get_cons_idx_u32(&cq->pbl);
5240 
5241 	while (num_entries && is_valid_cqe(cq, cqe)) {
5242 		int cnt = 0;
5243 		struct qlnxr_qp *qp;
5244 		struct rdma_cqe_requester *resp_cqe;
5245 		enum rdma_cqe_type cqe_type;
5246 
5247 		/* prevent speculative reads of any field of CQE */
5248 		rmb();
5249 
5250 		resp_cqe = &cqe->req;
5251 		qp = (struct qlnxr_qp *)(uintptr_t)HILO_U64(resp_cqe->qp_handle.hi,
5252 						resp_cqe->qp_handle.lo);
5253 
5254 		if (!qp) {
5255 			QL_DPRINT11(ha, "qp = NULL\n");
5256 			break;
5257 		}
5258 
5259 		wc->qp = &qp->ibqp;
5260 
5261 		cqe_type = GET_FIELD(resp_cqe->flags, RDMA_CQE_REQUESTER_TYPE);
5262 
5263 		switch (cqe_type) {
5264 		case RDMA_CQE_TYPE_REQUESTER:
5265 			cnt = qlnxr_poll_cq_req(dev, qp, cq, num_entries,
5266 					wc, &cqe->req);
5267 			try_consume_req_cqe(cq, qp, &cqe->req, &update);
5268 			break;
5269 		case RDMA_CQE_TYPE_RESPONDER_RQ:
5270 			cnt = qlnxr_poll_cq_resp(dev, qp, cq, num_entries,
5271 					wc, &cqe->resp, &update);
5272 			break;
5273 		case RDMA_CQE_TYPE_RESPONDER_SRQ:
5274 			cnt = qlnxr_poll_cq_resp_srq(dev, qp, cq, num_entries,
5275 					wc, &cqe->resp, &update);
5276 			break;
5277 		case RDMA_CQE_TYPE_INVALID:
5278 		default:
5279 			QL_DPRINT11(ha, "cqe type [0x%x] invalid\n", cqe_type);
5280 			break;
5281 		}
5282 		num_entries -= cnt;
5283 		wc += cnt;
5284 		done += cnt;
5285 
5286 		cqe = cq->latest_cqe;
5287 	}
5288 	new_cons = ecore_chain_get_cons_idx_u32(&cq->pbl);
5289 
5290 	cq->cq_cons += new_cons - old_cons;
5291 
5292 	if (update) {
5293 		/* doorbell notifies abount latest VALID entry,
5294 		 * but chain already point to the next INVALID one
5295 		 */
5296 		doorbell_cq(dev, cq, cq->cq_cons - 1, cq->arm_flags);
5297 		QL_DPRINT12(ha, "cq = %p cons = 0x%x "
5298 			"arm_flags = 0x%x db.icid = 0x%x\n", cq,
5299 			(cq->cq_cons - 1), cq->arm_flags, cq->db.data.icid);
5300 	}
5301 
5302 	spin_unlock_irqrestore(&cq->cq_lock, flags);
5303 
5304 	QL_DPRINT12(ha, "exit\n");
5305 
5306 	return done;
5307 }
5308 
5309 int
5310 qlnxr_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
5311 {
5312         struct qlnxr_cq *cq = get_qlnxr_cq(ibcq);
5313         unsigned long sflags;
5314         struct qlnxr_dev *dev;
5315 	qlnx_host_t	*ha;
5316 
5317 	dev = get_qlnxr_dev((ibcq->device));
5318 	ha = dev->ha;
5319 
5320 	QL_DPRINT12(ha, "enter ibcq = %p flags = 0x%x "
5321 		"cp = %p cons = 0x%x cq_type = 0x%x\n", ibcq,
5322 		flags, cq, cq->cq_cons, cq->cq_type);
5323 
5324 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
5325 		return -EINVAL;
5326 
5327 	if (cq->destroyed) {
5328 		QL_DPRINT11(ha, "cq was already destroyed cq = %p icid=%d\n",
5329 			cq, cq->icid);
5330 		return -EINVAL;
5331 	}
5332 
5333         if (cq->cq_type == QLNXR_CQ_TYPE_GSI) {
5334                 return 0;
5335         }
5336 
5337         spin_lock_irqsave(&cq->cq_lock, sflags);
5338 
5339         cq->arm_flags = 0;
5340 
5341         if (flags & IB_CQ_SOLICITED) {
5342                 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_SE_CF_CMD;
5343         }
5344         if (flags & IB_CQ_NEXT_COMP) {
5345                 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_CF_CMD;
5346         }
5347 
5348         doorbell_cq(dev, cq, (cq->cq_cons - 1), cq->arm_flags);
5349 
5350         spin_unlock_irqrestore(&cq->cq_lock, sflags);
5351 
5352 	QL_DPRINT12(ha, "exit ibcq = %p flags = 0x%x\n", ibcq, flags);
5353         return 0;
5354 }
5355 
5356 static struct qlnxr_mr *
5357 __qlnxr_alloc_mr(struct ib_pd *ibpd, int max_page_list_len)
5358 {
5359 	struct qlnxr_pd *pd = get_qlnxr_pd(ibpd);
5360 	struct qlnxr_dev *dev = get_qlnxr_dev((ibpd->device));
5361 	struct qlnxr_mr *mr;
5362 	int		rc = -ENOMEM;
5363 	qlnx_host_t	*ha;
5364 
5365 	ha = dev->ha;
5366 
5367 	QL_DPRINT12(ha, "enter ibpd = %p pd = %p "
5368 		" pd_id = %d max_page_list_len = %d\n",
5369 		ibpd, pd, pd->pd_id, max_page_list_len);
5370 
5371 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
5372 	if (!mr) {
5373 		QL_DPRINT11(ha, "kzalloc(mr) failed\n");
5374 		return ERR_PTR(rc);
5375 	}
5376 
5377 	mr->dev = dev;
5378 	mr->type = QLNXR_MR_FRMR;
5379 
5380 	rc = qlnxr_init_mr_info(dev, &mr->info, max_page_list_len,
5381 				  1 /* allow dual layer pbl */);
5382 	if (rc) {
5383 		QL_DPRINT11(ha, "qlnxr_init_mr_info failed\n");
5384 		goto err0;
5385 	}
5386 
5387 	rc = ecore_rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
5388 	if (rc) {
5389 		QL_DPRINT11(ha, "ecore_rdma_alloc_tid failed\n");
5390 		goto err0;
5391 	}
5392 
5393 	/* index only, 18 bit long, lkey = itid << 8 | key */
5394 	mr->hw_mr.tid_type = ECORE_RDMA_TID_FMR;
5395 	mr->hw_mr.key = 0;
5396 	mr->hw_mr.pd = pd->pd_id;
5397 	mr->hw_mr.local_read = 1;
5398 	mr->hw_mr.local_write = 0;
5399 	mr->hw_mr.remote_read = 0;
5400 	mr->hw_mr.remote_write = 0;
5401 	mr->hw_mr.remote_atomic = 0;
5402 	mr->hw_mr.mw_bind = false; /* TBD MW BIND */
5403 	mr->hw_mr.pbl_ptr = 0; /* Will be supplied during post */
5404 	mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
5405 	mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
5406 	mr->hw_mr.fbo = 0;
5407 	mr->hw_mr.length = 0;
5408 	mr->hw_mr.vaddr = 0;
5409 	mr->hw_mr.zbva = false; /* TBD figure when this should be true */
5410 	mr->hw_mr.phy_mr = true; /* Fast MR - True, Regular Register False */
5411 	mr->hw_mr.dma_mr = false;
5412 
5413 	rc = ecore_rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
5414 	if (rc) {
5415 		QL_DPRINT11(ha, "ecore_rdma_register_tid failed\n");
5416 		goto err1;
5417 	}
5418 
5419 	mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
5420 	mr->ibmr.rkey = mr->ibmr.lkey;
5421 
5422 	QL_DPRINT12(ha, "exit mr = %p mr->ibmr.lkey = 0x%x\n",
5423 		mr, mr->ibmr.lkey);
5424 
5425 	return mr;
5426 
5427 err1:
5428 	ecore_rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
5429 err0:
5430 	kfree(mr);
5431 
5432 	QL_DPRINT12(ha, "exit\n");
5433 
5434 	return ERR_PTR(rc);
5435 }
5436 
5437 struct ib_mr *
5438 qlnxr_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
5439     u32 max_num_sg, struct ib_udata *udata)
5440 {
5441 	struct qlnxr_dev *dev;
5442 	struct qlnxr_mr *mr;
5443 	qlnx_host_t     *ha;
5444 
5445 	dev = get_qlnxr_dev(ibpd->device);
5446 	ha = dev->ha;
5447 
5448 	QL_DPRINT12(ha, "enter\n");
5449 
5450 	if (mr_type != IB_MR_TYPE_MEM_REG)
5451 		return ERR_PTR(-EINVAL);
5452 
5453 	mr = __qlnxr_alloc_mr(ibpd, max_num_sg);
5454 
5455 	if (IS_ERR(mr))
5456 		return ERR_PTR(-EINVAL);
5457 
5458 	QL_DPRINT12(ha, "exit mr = %p &mr->ibmr = %p\n", mr, &mr->ibmr);
5459 
5460 	return &mr->ibmr;
5461 }
5462 
5463 static int
5464 qlnxr_set_page(struct ib_mr *ibmr, u64 addr)
5465 {
5466 	struct qlnxr_mr *mr = get_qlnxr_mr(ibmr);
5467 	struct qlnxr_pbl *pbl_table;
5468 	struct regpair *pbe;
5469 	struct qlnxr_dev *dev;
5470 	qlnx_host_t     *ha;
5471 	u32 pbes_in_page;
5472 
5473 	dev = mr->dev;
5474 	ha = dev->ha;
5475 
5476 	if (unlikely(mr->npages == mr->info.pbl_info.num_pbes)) {
5477 		QL_DPRINT12(ha, "fails mr->npages %d\n", mr->npages);
5478 		return -ENOMEM;
5479 	}
5480 
5481 	QL_DPRINT12(ha, "mr->npages %d addr = %p enter\n", mr->npages,
5482 		((void *)addr));
5483 
5484 	pbes_in_page = mr->info.pbl_info.pbl_size / sizeof(u64);
5485 	pbl_table = mr->info.pbl_table + (mr->npages / pbes_in_page);
5486 	pbe = (struct regpair *)pbl_table->va;
5487 	pbe +=  mr->npages % pbes_in_page;
5488 	pbe->lo = cpu_to_le32((u32)addr);
5489 	pbe->hi = cpu_to_le32((u32)upper_32_bits(addr));
5490 
5491 	mr->npages++;
5492 
5493 	QL_DPRINT12(ha, "mr->npages %d addr = %p exit \n", mr->npages,
5494 		((void *)addr));
5495 	return 0;
5496 }
5497 
5498 int
5499 qlnxr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
5500 	int sg_nents, unsigned int *sg_offset)
5501 {
5502 	int             ret;
5503 	struct qlnxr_mr *mr = get_qlnxr_mr(ibmr);
5504 	qlnx_host_t     *ha;
5505 
5506 	if (mr == NULL)
5507 		return (-1);
5508 
5509 	if (mr->dev == NULL)
5510 		return (-1);
5511 
5512 	ha = mr->dev->ha;
5513 
5514 	QL_DPRINT12(ha, "enter\n");
5515 
5516 	mr->npages = 0;
5517 	qlnx_handle_completed_mrs(mr->dev, &mr->info);
5518 
5519 	ret = ib_sg_to_pages(ibmr, sg, sg_nents, NULL, qlnxr_set_page);
5520 
5521 	QL_DPRINT12(ha, "exit ret = %d\n", ret);
5522 
5523 	return (ret);
5524 }
5525 
5526 int
5527 qlnxr_create_ah(struct ib_ah *ibah,
5528 	struct ib_ah_attr *attr, u32 flags,
5529 	struct ib_udata *udata)
5530 {
5531 	struct qlnxr_dev *dev;
5532 	qlnx_host_t	*ha;
5533 	struct qlnxr_ah *ah = get_qlnxr_ah(ibah);
5534 
5535 	dev = get_qlnxr_dev(ibah->device);
5536 	ha = dev->ha;
5537 
5538 	QL_DPRINT12(ha, "in create_ah\n");
5539 
5540 	ah->attr = *attr;
5541 
5542 	return (0);
5543 }
5544 
5545 void
5546 qlnxr_destroy_ah(struct ib_ah *ibah, u32 flags)
5547 {
5548 }
5549 
5550 int
5551 qlnxr_query_ah(struct ib_ah *ibah, struct ib_ah_attr *attr)
5552 {
5553 	struct qlnxr_dev *dev;
5554 	qlnx_host_t     *ha;
5555 
5556 	dev = get_qlnxr_dev((ibah->device));
5557 	ha = dev->ha;
5558 	QL_DPRINT12(ha, "Query AH not supported\n");
5559 	return -EINVAL;
5560 }
5561 
5562 int
5563 qlnxr_modify_ah(struct ib_ah *ibah, struct ib_ah_attr *attr)
5564 {
5565 	struct qlnxr_dev *dev;
5566 	qlnx_host_t     *ha;
5567 
5568 	dev = get_qlnxr_dev((ibah->device));
5569 	ha = dev->ha;
5570 	QL_DPRINT12(ha, "Modify AH not supported\n");
5571 	return -ENOSYS;
5572 }
5573 
5574 int
5575 qlnxr_process_mad(struct ib_device *ibdev,
5576 		int process_mad_flags,
5577 		u8 port_num,
5578 		const struct ib_wc *in_wc,
5579 		const struct ib_grh *in_grh,
5580 		const struct ib_mad_hdr *mad_hdr,
5581 		size_t in_mad_size,
5582 		struct ib_mad_hdr *out_mad,
5583 		size_t *out_mad_size,
5584 		u16 *out_mad_pkey_index)
5585 {
5586 	struct qlnxr_dev *dev;
5587 	qlnx_host_t	*ha;
5588 
5589 	dev = get_qlnxr_dev(ibdev);
5590 	ha = dev->ha;
5591 	QL_DPRINT12(ha, "process mad not supported\n");
5592 
5593 	return -ENOSYS;
5594 //	QL_DPRINT12(ha, "qlnxr_process_mad in_mad %x %x %x %x %x %x %x %x\n",
5595 //               in_mad->mad_hdr.attr_id, in_mad->mad_hdr.base_version,
5596 //               in_mad->mad_hdr.attr_mod, in_mad->mad_hdr.class_specific,
5597 //               in_mad->mad_hdr.class_version, in_mad->mad_hdr.method,
5598 //               in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.status);
5599 
5600 //	return IB_MAD_RESULT_SUCCESS;
5601 }
5602 
5603 int
5604 qlnxr_get_port_immutable(struct ib_device *ibdev, u8 port_num,
5605 	struct ib_port_immutable *immutable)
5606 {
5607 	struct qlnxr_dev        *dev;
5608 	qlnx_host_t             *ha;
5609 	struct ib_port_attr     attr;
5610 	int                     err;
5611 
5612 	dev = get_qlnxr_dev(ibdev);
5613 	ha = dev->ha;
5614 
5615 	QL_DPRINT12(ha, "enter\n");
5616 
5617 	err = qlnxr_query_port(ibdev, port_num, &attr);
5618 	if (err)
5619 		return err;
5620 
5621 	if (QLNX_IS_IWARP(dev)) {
5622 		immutable->pkey_tbl_len = 1;
5623 		immutable->gid_tbl_len = 1;
5624 		immutable->core_cap_flags = RDMA_CORE_PORT_IWARP;
5625 		immutable->max_mad_size = 0;
5626 	} else {
5627 		immutable->pkey_tbl_len = attr.pkey_tbl_len;
5628 		immutable->gid_tbl_len = attr.gid_tbl_len;
5629 		immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
5630 		immutable->max_mad_size = IB_MGMT_MAD_SIZE;
5631 	}
5632 
5633 	QL_DPRINT12(ha, "exit\n");
5634 	return 0;
5635 }
5636 
5637 /***** iWARP related functions *************/
5638 
5639 static void
5640 qlnxr_iw_mpa_request(void *context,
5641 	struct ecore_iwarp_cm_event_params *params)
5642 {
5643 	struct qlnxr_iw_listener *listener = (struct qlnxr_iw_listener *)context;
5644 	struct qlnxr_dev *dev = listener->dev;
5645 	struct qlnxr_iw_ep *ep;
5646 	struct iw_cm_event event;
5647 	struct sockaddr_in *laddr;
5648 	struct sockaddr_in *raddr;
5649 	qlnx_host_t	*ha;
5650 
5651 	ha = dev->ha;
5652 
5653 	QL_DPRINT12(ha, "enter\n");
5654 
5655 	if (params->cm_info->ip_version != ECORE_TCP_IPV4) {
5656 		QL_DPRINT11(ha, "only IPv4 supported [0x%x]\n",
5657 			params->cm_info->ip_version);
5658 		return;
5659 	}
5660 
5661 	ep = kzalloc(sizeof(*ep), GFP_ATOMIC);
5662 
5663 	if (!ep) {
5664 		QL_DPRINT11(ha, "kzalloc{ep) failed\n");
5665 		return;
5666 	}
5667 
5668 	ep->dev = dev;
5669 	ep->ecore_context = params->ep_context;
5670 
5671 	memset(&event, 0, sizeof(event));
5672 
5673 	event.event = IW_CM_EVENT_CONNECT_REQUEST;
5674 	event.status = params->status;
5675 
5676 	laddr = (struct sockaddr_in *)&event.local_addr;
5677 	raddr = (struct sockaddr_in *)&event.remote_addr;
5678 
5679 	laddr->sin_family = AF_INET;
5680 	raddr->sin_family = AF_INET;
5681 
5682 	laddr->sin_port = htons(params->cm_info->local_port);
5683 	raddr->sin_port = htons(params->cm_info->remote_port);
5684 
5685 	laddr->sin_addr.s_addr = htonl(params->cm_info->local_ip[0]);
5686 	raddr->sin_addr.s_addr = htonl(params->cm_info->remote_ip[0]);
5687 
5688 	event.provider_data = (void *)ep;
5689 	event.private_data = (void *)params->cm_info->private_data;
5690 	event.private_data_len = (u8)params->cm_info->private_data_len;
5691 
5692 	event.ord = params->cm_info->ord;
5693 	event.ird = params->cm_info->ird;
5694 
5695 	listener->cm_id->event_handler(listener->cm_id, &event);
5696 
5697 	QL_DPRINT12(ha, "exit\n");
5698 
5699 	return;
5700 }
5701 
5702 static void
5703 qlnxr_iw_issue_event(void *context,
5704 	 struct ecore_iwarp_cm_event_params *params,
5705 	 enum iw_cm_event_type event_type,
5706 	 char *str)
5707 {
5708 	struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)context;
5709 	struct qlnxr_dev *dev = ep->dev;
5710 	struct iw_cm_event event;
5711 	qlnx_host_t	*ha;
5712 
5713 	ha = dev->ha;
5714 
5715 	QL_DPRINT12(ha, "enter\n");
5716 
5717 	memset(&event, 0, sizeof(event));
5718 	event.status = params->status;
5719 	event.event = event_type;
5720 
5721 	if (params->cm_info != NULL) {
5722 		event.ird = params->cm_info->ird;
5723 		event.ord = params->cm_info->ord;
5724 		QL_DPRINT12(ha, "ord=[%d] \n", event.ord);
5725 		QL_DPRINT12(ha, "ird=[%d] \n", event.ird);
5726 
5727 		event.private_data_len = params->cm_info->private_data_len;
5728 		event.private_data = (void *)params->cm_info->private_data;
5729 		QL_DPRINT12(ha, "private_data_len=[%d] \n",
5730 			event.private_data_len);
5731 	}
5732 
5733 	QL_DPRINT12(ha, "event=[%d] %s\n", event.event, str);
5734 	QL_DPRINT12(ha, "status=[%d] \n", event.status);
5735 
5736 	if (ep) {
5737 		if (ep->cm_id)
5738 			ep->cm_id->event_handler(ep->cm_id, &event);
5739 		else
5740 			QL_DPRINT11(ha, "ep->cm_id == NULL \n");
5741 	} else {
5742 		QL_DPRINT11(ha, "ep == NULL \n");
5743 	}
5744 
5745 	QL_DPRINT12(ha, "exit\n");
5746 
5747 	return;
5748 }
5749 
5750 static void
5751 qlnxr_iw_close_event(void *context,
5752 	 struct ecore_iwarp_cm_event_params *params)
5753 {
5754 	struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)context;
5755 	struct qlnxr_dev *dev = ep->dev;
5756 	qlnx_host_t	*ha;
5757 
5758 	ha = dev->ha;
5759 
5760 	QL_DPRINT12(ha, "enter\n");
5761 
5762 	if (ep->cm_id) {
5763 		qlnxr_iw_issue_event(context,
5764 				    params,
5765 				    IW_CM_EVENT_CLOSE,
5766 				    "IW_CM_EVENT_EVENT_CLOSE");
5767 		ep->cm_id->rem_ref(ep->cm_id);
5768 		ep->cm_id = NULL;
5769 	}
5770 
5771 	QL_DPRINT12(ha, "exit\n");
5772 
5773 	return;
5774 }
5775 
5776 static void
5777 qlnxr_iw_passive_complete(void *context,
5778         struct ecore_iwarp_cm_event_params *params)
5779 {
5780         struct qlnxr_iw_ep      *ep = (struct qlnxr_iw_ep *)context;
5781         struct qlnxr_dev        *dev = ep->dev;
5782         qlnx_host_t             *ha;
5783 
5784         ha = dev->ha;
5785 
5786         /* We will only reach the following state if MPA_REJECT was called on
5787          * passive. In this case there will be no associated QP.
5788          */
5789         if ((params->status == -ECONNREFUSED) && (ep->qp == NULL)) {
5790                 QL_DPRINT11(ha, "PASSIVE connection refused releasing ep...\n");
5791                 kfree(ep);
5792                 return;
5793         }
5794 
5795         /* We always issue an established event, however, ofed does not look
5796          * at event code for established. So if there was a failure, we follow
5797          * with close...
5798          */
5799         qlnxr_iw_issue_event(context,
5800                 params,
5801                 IW_CM_EVENT_ESTABLISHED,
5802                 "IW_CM_EVENT_ESTABLISHED");
5803 
5804         if (params->status < 0) {
5805                 qlnxr_iw_close_event(context, params);
5806         }
5807 
5808         return;
5809 }
5810 
5811 struct qlnxr_discon_work {
5812         struct work_struct work;
5813         struct qlnxr_iw_ep *ep;
5814         enum ecore_iwarp_event_type event;
5815         int status;
5816 };
5817 
5818 static void
5819 qlnxr_iw_disconnect_worker(struct work_struct *work)
5820 {
5821         struct qlnxr_discon_work *dwork =
5822                 container_of(work, struct qlnxr_discon_work, work);
5823         struct ecore_rdma_modify_qp_in_params qp_params = { 0 };
5824         struct qlnxr_iw_ep *ep = dwork->ep;
5825         struct qlnxr_dev *dev = ep->dev;
5826         struct qlnxr_qp *qp = ep->qp;
5827         struct iw_cm_event event;
5828 
5829         if (qp->destroyed) {
5830                 kfree(dwork);
5831                 qlnxr_iw_qp_rem_ref(&qp->ibqp);
5832                 return;
5833         }
5834 
5835         memset(&event, 0, sizeof(event));
5836         event.status = dwork->status;
5837         event.event = IW_CM_EVENT_DISCONNECT;
5838 
5839         /* Success means graceful disconnect was requested. modifying
5840          * to SQD is translated to graceful disconnect. O/w reset is sent
5841          */
5842         if (dwork->status)
5843                 qp_params.new_state = ECORE_ROCE_QP_STATE_ERR;
5844         else
5845                 qp_params.new_state = ECORE_ROCE_QP_STATE_SQD;
5846 
5847         kfree(dwork);
5848 
5849         if (ep->cm_id)
5850                 ep->cm_id->event_handler(ep->cm_id, &event);
5851 
5852         SET_FIELD(qp_params.modify_flags,
5853                   ECORE_RDMA_MODIFY_QP_VALID_NEW_STATE, 1);
5854 
5855         ecore_rdma_modify_qp(dev->rdma_ctx, qp->ecore_qp, &qp_params);
5856 
5857         qlnxr_iw_qp_rem_ref(&qp->ibqp);
5858 
5859         return;
5860 }
5861 
5862 void
5863 qlnxr_iw_disconnect_event(void *context,
5864         struct ecore_iwarp_cm_event_params *params)
5865 {
5866         struct qlnxr_discon_work *work;
5867         struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)context;
5868         struct qlnxr_dev *dev = ep->dev;
5869         struct qlnxr_qp *qp = ep->qp;
5870 
5871         work = kzalloc(sizeof(*work), GFP_ATOMIC);
5872         if (!work)
5873                 return;
5874 
5875         qlnxr_iw_qp_add_ref(&qp->ibqp);
5876         work->ep = ep;
5877         work->event = params->event;
5878         work->status = params->status;
5879 
5880         INIT_WORK(&work->work, qlnxr_iw_disconnect_worker);
5881         queue_work(dev->iwarp_wq, &work->work);
5882 
5883         return;
5884 }
5885 
5886 static int
5887 qlnxr_iw_mpa_reply(void *context,
5888 	struct ecore_iwarp_cm_event_params *params)
5889 {
5890         struct qlnxr_iw_ep	*ep = (struct qlnxr_iw_ep *)context;
5891         struct qlnxr_dev	*dev = ep->dev;
5892         struct ecore_iwarp_send_rtr_in rtr_in;
5893         int			rc;
5894 	qlnx_host_t		*ha;
5895 
5896 	ha = dev->ha;
5897 
5898 	QL_DPRINT12(ha, "enter\n");
5899 
5900 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
5901 		return -EINVAL;
5902 
5903 	bzero(&rtr_in, sizeof(struct ecore_iwarp_send_rtr_in));
5904         rtr_in.ep_context = params->ep_context;
5905 
5906         rc = ecore_iwarp_send_rtr(dev->rdma_ctx, &rtr_in);
5907 
5908 	QL_DPRINT12(ha, "exit rc = %d\n", rc);
5909         return rc;
5910 }
5911 
5912 void
5913 qlnxr_iw_qp_event(void *context,
5914 	struct ecore_iwarp_cm_event_params *params,
5915 	enum ib_event_type ib_event,
5916 	char *str)
5917 {
5918         struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)context;
5919         struct qlnxr_dev *dev = ep->dev;
5920         struct ib_qp *ibqp = &(ep->qp->ibqp);
5921         struct ib_event event;
5922 	qlnx_host_t	*ha;
5923 
5924 	ha = dev->ha;
5925 
5926 	QL_DPRINT12(ha,
5927 		"[context, event, event_handler] = [%p, 0x%x, %s, %p] enter\n",
5928 		context, params->event, str, ibqp->event_handler);
5929 
5930         if (ibqp->event_handler) {
5931                 event.event = ib_event;
5932                 event.device = ibqp->device;
5933                 event.element.qp = ibqp;
5934                 ibqp->event_handler(&event, ibqp->qp_context);
5935         }
5936 
5937 	return;
5938 }
5939 
5940 int
5941 qlnxr_iw_event_handler(void *context,
5942 	struct ecore_iwarp_cm_event_params *params)
5943 {
5944 	struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)context;
5945 	struct qlnxr_dev *dev = ep->dev;
5946 	qlnx_host_t	*ha;
5947 
5948 	ha = dev->ha;
5949 
5950 	QL_DPRINT12(ha, "[context, event] = [%p, 0x%x] "
5951 		"enter\n", context, params->event);
5952 
5953 	switch (params->event) {
5954 	/* Passive side request received */
5955 	case ECORE_IWARP_EVENT_MPA_REQUEST:
5956 		qlnxr_iw_mpa_request(context, params);
5957 		break;
5958 
5959         case ECORE_IWARP_EVENT_ACTIVE_MPA_REPLY:
5960                 qlnxr_iw_mpa_reply(context, params);
5961                 break;
5962 
5963 	/* Passive side established ( ack on mpa response ) */
5964 	case ECORE_IWARP_EVENT_PASSIVE_COMPLETE:
5965 		ep->during_connect = 0;
5966 		qlnxr_iw_passive_complete(context, params);
5967 		break;
5968 
5969 	/* Active side reply received */
5970 	case ECORE_IWARP_EVENT_ACTIVE_COMPLETE:
5971 		ep->during_connect = 0;
5972 		qlnxr_iw_issue_event(context,
5973 				    params,
5974 				    IW_CM_EVENT_CONNECT_REPLY,
5975 				    "IW_CM_EVENT_CONNECT_REPLY");
5976 		if (params->status < 0) {
5977 			struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)context;
5978 
5979 			ep->cm_id->rem_ref(ep->cm_id);
5980 			ep->cm_id = NULL;
5981 		}
5982 		break;
5983 
5984 	case ECORE_IWARP_EVENT_DISCONNECT:
5985 		qlnxr_iw_disconnect_event(context, params);
5986 		break;
5987 
5988 	case ECORE_IWARP_EVENT_CLOSE:
5989 		ep->during_connect = 0;
5990 		qlnxr_iw_close_event(context, params);
5991 		break;
5992 
5993         case ECORE_IWARP_EVENT_RQ_EMPTY:
5994                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_FATAL,
5995                                  "IWARP_EVENT_RQ_EMPTY");
5996                 break;
5997 
5998         case ECORE_IWARP_EVENT_IRQ_FULL:
5999                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_FATAL,
6000                                  "IWARP_EVENT_IRQ_FULL");
6001                 break;
6002 
6003         case ECORE_IWARP_EVENT_LLP_TIMEOUT:
6004                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_FATAL,
6005                                  "IWARP_EVENT_LLP_TIMEOUT");
6006                 break;
6007 
6008         case ECORE_IWARP_EVENT_REMOTE_PROTECTION_ERROR:
6009                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_ACCESS_ERR,
6010                                  "IWARP_EVENT_REMOTE_PROTECTION_ERROR");
6011                 break;
6012 
6013         case ECORE_IWARP_EVENT_CQ_OVERFLOW:
6014                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_FATAL,
6015                                  "QED_IWARP_EVENT_CQ_OVERFLOW");
6016                 break;
6017 
6018         case ECORE_IWARP_EVENT_QP_CATASTROPHIC:
6019                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_FATAL,
6020                                  "QED_IWARP_EVENT_QP_CATASTROPHIC");
6021                 break;
6022 
6023         case ECORE_IWARP_EVENT_LOCAL_ACCESS_ERROR:
6024                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_ACCESS_ERR,
6025                                  "IWARP_EVENT_LOCAL_ACCESS_ERROR");
6026                 break;
6027 
6028         case ECORE_IWARP_EVENT_REMOTE_OPERATION_ERROR:
6029                 qlnxr_iw_qp_event(context, params, IB_EVENT_QP_FATAL,
6030                                  "IWARP_EVENT_REMOTE_OPERATION_ERROR");
6031                 break;
6032 
6033         case ECORE_IWARP_EVENT_TERMINATE_RECEIVED:
6034 		QL_DPRINT12(ha, "Got terminate message"
6035 			" ECORE_IWARP_EVENT_TERMINATE_RECEIVED\n");
6036                 break;
6037 
6038 	default:
6039 		QL_DPRINT12(ha,
6040 			"Unknown event [0x%x] received \n", params->event);
6041 		break;
6042 	};
6043 
6044 	QL_DPRINT12(ha, "[context, event] = [%p, 0x%x] "
6045 		"exit\n", context, params->event);
6046 	return 0;
6047 }
6048 
6049 static int
6050 qlnxr_addr4_resolve(struct qlnxr_dev *dev,
6051 			      struct sockaddr_in *src_in,
6052 			      struct sockaddr_in *dst_in,
6053 			      u8 *dst_mac)
6054 {
6055 	int rc;
6056 
6057 	rc = arpresolve(dev->ha->ifp, 0, NULL, (struct sockaddr *)dst_in,
6058 			dst_mac, NULL, NULL);
6059 
6060 	QL_DPRINT12(dev->ha, "rc = %d "
6061 		"sa_len = 0x%x sa_family = 0x%x IP Address = %d.%d.%d.%d "
6062 		"Dest MAC %02x:%02x:%02x:%02x:%02x:%02x\n", rc,
6063 		dst_in->sin_len, dst_in->sin_family,
6064 		NIPQUAD((dst_in->sin_addr.s_addr)),
6065 		dst_mac[0], dst_mac[1], dst_mac[2],
6066 		dst_mac[3], dst_mac[4], dst_mac[5]);
6067 
6068 	return rc;
6069 }
6070 
6071 int
6072 qlnxr_iw_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
6073 {
6074 	struct qlnxr_dev *dev;
6075 	struct ecore_iwarp_connect_out out_params;
6076 	struct ecore_iwarp_connect_in in_params;
6077 	struct qlnxr_iw_ep *ep;
6078 	struct qlnxr_qp *qp;
6079 	struct sockaddr_in *laddr;
6080 	struct sockaddr_in *raddr;
6081 	int rc = 0;
6082 	qlnx_host_t	*ha;
6083 
6084 	dev = get_qlnxr_dev((cm_id->device));
6085 	ha = dev->ha;
6086 
6087 	QL_DPRINT12(ha, "[cm_id, conn_param] = [%p, %p] "
6088 		"enter \n", cm_id, conn_param);
6089 
6090 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
6091 		return -EINVAL;
6092 
6093 	qp = idr_find(&dev->qpidr, conn_param->qpn);
6094 
6095 	laddr = (struct sockaddr_in *)&cm_id->local_addr;
6096 	raddr = (struct sockaddr_in *)&cm_id->remote_addr;
6097 
6098 	QL_DPRINT12(ha,
6099 		"local = [%d.%d.%d.%d, %d] remote = [%d.%d.%d.%d, %d]\n",
6100 		NIPQUAD((laddr->sin_addr.s_addr)), laddr->sin_port,
6101 		NIPQUAD((raddr->sin_addr.s_addr)), raddr->sin_port);
6102 
6103 	ep = kzalloc(sizeof(*ep), GFP_KERNEL);
6104 	if (!ep) {
6105 		QL_DPRINT11(ha, "struct qlnxr_iw_ep "
6106 			"alloc memory failed\n");
6107 		return -ENOMEM;
6108 	}
6109 
6110 	ep->dev = dev;
6111 	ep->qp = qp;
6112 	cm_id->add_ref(cm_id);
6113 	ep->cm_id = cm_id;
6114 
6115 	memset(&in_params, 0, sizeof (struct ecore_iwarp_connect_in));
6116 	memset(&out_params, 0, sizeof (struct ecore_iwarp_connect_out));
6117 
6118 	in_params.event_cb = qlnxr_iw_event_handler;
6119 	in_params.cb_context = ep;
6120 
6121 	in_params.cm_info.ip_version = ECORE_TCP_IPV4;
6122 
6123 	in_params.cm_info.remote_ip[0] = ntohl(raddr->sin_addr.s_addr);
6124 	in_params.cm_info.local_ip[0] = ntohl(laddr->sin_addr.s_addr);
6125 	in_params.cm_info.remote_port = ntohs(raddr->sin_port);
6126 	in_params.cm_info.local_port = ntohs(laddr->sin_port);
6127 	in_params.cm_info.vlan = 0;
6128 	in_params.mss = if_getmtu(dev->ha->ifp) - 40;
6129 
6130 	QL_DPRINT12(ha, "remote_ip = [%d.%d.%d.%d] "
6131 		"local_ip = [%d.%d.%d.%d] remote_port = %d local_port = %d "
6132 		"vlan = %d\n",
6133 		NIPQUAD((in_params.cm_info.remote_ip[0])),
6134 		NIPQUAD((in_params.cm_info.local_ip[0])),
6135 		in_params.cm_info.remote_port, in_params.cm_info.local_port,
6136 		in_params.cm_info.vlan);
6137 
6138 	rc = qlnxr_addr4_resolve(dev, laddr, raddr, (u8 *)in_params.remote_mac_addr);
6139 
6140 	if (rc) {
6141 		QL_DPRINT11(ha, "qlnxr_addr4_resolve failed\n");
6142 		goto err;
6143 	}
6144 
6145 	QL_DPRINT12(ha, "ord = %d ird=%d private_data=%p"
6146 		" private_data_len=%d rq_psn=%d\n",
6147 		conn_param->ord, conn_param->ird, conn_param->private_data,
6148 		conn_param->private_data_len, qp->rq_psn);
6149 
6150 	in_params.cm_info.ord = conn_param->ord;
6151 	in_params.cm_info.ird = conn_param->ird;
6152 	in_params.cm_info.private_data = conn_param->private_data;
6153 	in_params.cm_info.private_data_len = conn_param->private_data_len;
6154 	in_params.qp = qp->ecore_qp;
6155 
6156 	memcpy(in_params.local_mac_addr, dev->ha->primary_mac, ETH_ALEN);
6157 
6158 	rc = ecore_iwarp_connect(dev->rdma_ctx, &in_params, &out_params);
6159 
6160 	if (rc) {
6161 		QL_DPRINT12(ha, "ecore_iwarp_connect failed\n");
6162 		goto err;
6163 	}
6164 
6165 	QL_DPRINT12(ha, "exit\n");
6166 
6167 	return rc;
6168 
6169 err:
6170 	cm_id->rem_ref(cm_id);
6171 	kfree(ep);
6172 
6173 	QL_DPRINT12(ha, "exit [%d]\n", rc);
6174 	return rc;
6175 }
6176 
6177 int
6178 qlnxr_iw_create_listen(struct iw_cm_id *cm_id, int backlog)
6179 {
6180 	struct qlnxr_dev *dev;
6181 	struct qlnxr_iw_listener *listener;
6182 	struct ecore_iwarp_listen_in iparams;
6183 	struct ecore_iwarp_listen_out oparams;
6184 	struct sockaddr_in *laddr;
6185 	qlnx_host_t	*ha;
6186 	int rc;
6187 
6188 	dev = get_qlnxr_dev((cm_id->device));
6189 	ha = dev->ha;
6190 
6191 	QL_DPRINT12(ha, "enter\n");
6192 
6193 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
6194 		return -EINVAL;
6195 
6196 	laddr = (struct sockaddr_in *)&cm_id->local_addr;
6197 
6198 	listener = kzalloc(sizeof(*listener), GFP_KERNEL);
6199 
6200 	if (listener == NULL) {
6201 		QL_DPRINT11(ha, "listener memory alloc failed\n");
6202 		return -ENOMEM;
6203 	}
6204 
6205 	listener->dev = dev;
6206 	cm_id->add_ref(cm_id);
6207 	listener->cm_id = cm_id;
6208 	listener->backlog = backlog;
6209 
6210 	memset(&iparams, 0, sizeof (struct ecore_iwarp_listen_in));
6211 	memset(&oparams, 0, sizeof (struct ecore_iwarp_listen_out));
6212 
6213 	iparams.cb_context = listener;
6214 	iparams.event_cb = qlnxr_iw_event_handler;
6215 	iparams.max_backlog = backlog;
6216 
6217 	iparams.ip_version = ECORE_TCP_IPV4;
6218 
6219 	iparams.ip_addr[0] = ntohl(laddr->sin_addr.s_addr);
6220 	iparams.port = ntohs(laddr->sin_port);
6221 	iparams.vlan = 0;
6222 
6223 	QL_DPRINT12(ha, "[%d.%d.%d.%d, %d] iparamsport=%d\n",
6224 		NIPQUAD((laddr->sin_addr.s_addr)),
6225 		laddr->sin_port, iparams.port);
6226 
6227 	rc = ecore_iwarp_create_listen(dev->rdma_ctx, &iparams, &oparams);
6228 	if (rc) {
6229 		QL_DPRINT11(ha,
6230 			"ecore_iwarp_create_listen failed rc = %d\n", rc);
6231 		goto err;
6232 	}
6233 
6234 	listener->ecore_handle = oparams.handle;
6235 	cm_id->provider_data = listener;
6236 
6237 	QL_DPRINT12(ha, "exit\n");
6238 	return rc;
6239 
6240 err:
6241 	cm_id->rem_ref(cm_id);
6242 	kfree(listener);
6243 
6244 	QL_DPRINT12(ha, "exit [%d]\n", rc);
6245 	return rc;
6246 }
6247 
6248 void
6249 qlnxr_iw_destroy_listen(struct iw_cm_id *cm_id)
6250 {
6251 	struct qlnxr_iw_listener *listener = cm_id->provider_data;
6252 	struct qlnxr_dev *dev = get_qlnxr_dev((cm_id->device));
6253 	int rc = 0;
6254 	qlnx_host_t	*ha;
6255 
6256 	ha = dev->ha;
6257 
6258 	QL_DPRINT12(ha, "enter\n");
6259 
6260 	if (listener->ecore_handle)
6261 		rc = ecore_iwarp_destroy_listen(dev->rdma_ctx,
6262 				listener->ecore_handle);
6263 
6264 	cm_id->rem_ref(cm_id);
6265 
6266 	QL_DPRINT12(ha, "exit [%d]\n", rc);
6267 	return;
6268 }
6269 
6270 int
6271 qlnxr_iw_accept(struct iw_cm_id *cm_id,
6272 	struct iw_cm_conn_param *conn_param)
6273 {
6274 	struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)cm_id->provider_data;
6275 	struct qlnxr_dev *dev = ep->dev;
6276 	struct qlnxr_qp *qp;
6277 	struct ecore_iwarp_accept_in params;
6278 	int rc;
6279 	qlnx_host_t	*ha;
6280 
6281 	ha = dev->ha;
6282 
6283 	QL_DPRINT12(ha, "enter  qpid=%d\n", conn_param->qpn);
6284 
6285 	if (!(if_getdrvflags(ha->ifp) & IFF_DRV_RUNNING))
6286 		return -EINVAL;
6287 
6288 	qp = idr_find(&dev->qpidr, conn_param->qpn);
6289 	if (!qp) {
6290 		QL_DPRINT11(ha, "idr_find failed invalid qpn = %d\n",
6291 			conn_param->qpn);
6292 		return -EINVAL;
6293 	}
6294 	ep->qp = qp;
6295 	qp->ep = ep;
6296 	cm_id->add_ref(cm_id);
6297 	ep->cm_id = cm_id;
6298 
6299 	params.ep_context = ep->ecore_context;
6300 	params.cb_context = ep;
6301 	params.qp = ep->qp->ecore_qp;
6302 	params.private_data = conn_param->private_data;
6303 	params.private_data_len = conn_param->private_data_len;
6304 	params.ird = conn_param->ird;
6305 	params.ord = conn_param->ord;
6306 
6307 	rc = ecore_iwarp_accept(dev->rdma_ctx, &params);
6308 	if (rc) {
6309 		QL_DPRINT11(ha, "ecore_iwarp_accept failed %d\n", rc);
6310 		goto err;
6311 	}
6312 
6313 	QL_DPRINT12(ha, "exit\n");
6314 	return 0;
6315 err:
6316 	cm_id->rem_ref(cm_id);
6317 	QL_DPRINT12(ha, "exit rc = %d\n", rc);
6318 	return rc;
6319 }
6320 
6321 int
6322 qlnxr_iw_reject(struct iw_cm_id *cm_id, const void *pdata, u8 pdata_len)
6323 {
6324         struct qlnxr_iw_ep *ep = (struct qlnxr_iw_ep *)cm_id->provider_data;
6325         struct qlnxr_dev *dev = ep->dev;
6326         struct ecore_iwarp_reject_in params;
6327         int rc;
6328 
6329         params.ep_context = ep->ecore_context;
6330         params.cb_context = ep;
6331         params.private_data = pdata;
6332         params.private_data_len = pdata_len;
6333         ep->qp = NULL;
6334 
6335         rc = ecore_iwarp_reject(dev->rdma_ctx, &params);
6336 
6337         return rc;
6338 }
6339 
6340 void
6341 qlnxr_iw_qp_add_ref(struct ib_qp *ibqp)
6342 {
6343 	struct qlnxr_qp *qp = get_qlnxr_qp(ibqp);
6344 	qlnx_host_t	*ha;
6345 
6346 	ha = qp->dev->ha;
6347 
6348 	QL_DPRINT12(ha, "enter ibqp = %p\n", ibqp);
6349 
6350 	atomic_inc(&qp->refcnt);
6351 
6352 	QL_DPRINT12(ha, "exit \n");
6353 	return;
6354 }
6355 
6356 void
6357 qlnxr_iw_qp_rem_ref(struct ib_qp *ibqp)
6358 {
6359 	struct qlnxr_qp *qp = get_qlnxr_qp(ibqp);
6360 	qlnx_host_t	*ha;
6361 
6362 	ha = qp->dev->ha;
6363 
6364 	QL_DPRINT12(ha, "enter ibqp = %p qp = %p\n", ibqp, qp);
6365 
6366 	if (atomic_dec_and_test(&qp->refcnt)) {
6367 		qlnxr_idr_remove(qp->dev, qp->qp_id);
6368 	}
6369 
6370 	QL_DPRINT12(ha, "exit \n");
6371 	return;
6372 }
6373 
6374 struct ib_qp *
6375 qlnxr_iw_get_qp(struct ib_device *ibdev, int qpn)
6376 {
6377 	struct qlnxr_dev *dev = get_qlnxr_dev(ibdev);
6378 	struct ib_qp *qp;
6379 	qlnx_host_t	*ha;
6380 
6381 	ha = dev->ha;
6382 
6383 	QL_DPRINT12(ha, "enter dev = %p ibdev = %p qpn = %d\n", dev, ibdev, qpn);
6384 
6385 	qp = idr_find(&dev->qpidr, qpn);
6386 
6387 	QL_DPRINT12(ha, "exit qp = %p\n", qp);
6388 
6389 	return (qp);
6390 }
6391