1 /*
2  * Broadcom NetXtreme-E RoCE driver.
3  *
4  * Copyright (c) 2016 - 2017, Broadcom. All rights reserved.  The term
5  * Broadcom refers to Broadcom Limited and/or its subsidiaries.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * BSD license below:
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  *
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in
21  *    the documentation and/or other materials provided with the
22  *    distribution.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS''
25  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
26  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
31  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
32  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
33  * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
34  * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35  *
36  * Description: IB Verbs interpreter
37  */
38 
39 #include <linux/interrupt.h>
40 #include <linux/types.h>
41 #include <linux/pci.h>
42 #include <linux/netdevice.h>
43 #include <linux/if_ether.h>
44 
45 #include <rdma/ib_verbs.h>
46 #include <rdma/ib_user_verbs.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_mad.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/uverbs_ioctl.h>
52 
53 #include "bnxt_ulp.h"
54 
55 #include "roce_hsi.h"
56 #include "qplib_res.h"
57 #include "qplib_sp.h"
58 #include "qplib_fp.h"
59 #include "qplib_rcfw.h"
60 
61 #include "bnxt_re.h"
62 #include "ib_verbs.h"
63 #include <rdma/bnxt_re-abi.h>
64 
65 static int __from_ib_access_flags(int iflags)
66 {
67 	int qflags = 0;
68 
69 	if (iflags & IB_ACCESS_LOCAL_WRITE)
70 		qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
71 	if (iflags & IB_ACCESS_REMOTE_READ)
72 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ;
73 	if (iflags & IB_ACCESS_REMOTE_WRITE)
74 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE;
75 	if (iflags & IB_ACCESS_REMOTE_ATOMIC)
76 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC;
77 	if (iflags & IB_ACCESS_MW_BIND)
78 		qflags |= BNXT_QPLIB_ACCESS_MW_BIND;
79 	if (iflags & IB_ZERO_BASED)
80 		qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED;
81 	if (iflags & IB_ACCESS_ON_DEMAND)
82 		qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND;
83 	return qflags;
84 };
85 
86 static enum ib_access_flags __to_ib_access_flags(int qflags)
87 {
88 	enum ib_access_flags iflags = 0;
89 
90 	if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE)
91 		iflags |= IB_ACCESS_LOCAL_WRITE;
92 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE)
93 		iflags |= IB_ACCESS_REMOTE_WRITE;
94 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ)
95 		iflags |= IB_ACCESS_REMOTE_READ;
96 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC)
97 		iflags |= IB_ACCESS_REMOTE_ATOMIC;
98 	if (qflags & BNXT_QPLIB_ACCESS_MW_BIND)
99 		iflags |= IB_ACCESS_MW_BIND;
100 	if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED)
101 		iflags |= IB_ZERO_BASED;
102 	if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND)
103 		iflags |= IB_ACCESS_ON_DEMAND;
104 	return iflags;
105 };
106 
107 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list,
108 			     struct bnxt_qplib_sge *sg_list, int num)
109 {
110 	int i, total = 0;
111 
112 	for (i = 0; i < num; i++) {
113 		sg_list[i].addr = ib_sg_list[i].addr;
114 		sg_list[i].lkey = ib_sg_list[i].lkey;
115 		sg_list[i].size = ib_sg_list[i].length;
116 		total += sg_list[i].size;
117 	}
118 	return total;
119 }
120 
121 /* Device */
122 int bnxt_re_query_device(struct ib_device *ibdev,
123 			 struct ib_device_attr *ib_attr,
124 			 struct ib_udata *udata)
125 {
126 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
127 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
128 
129 	memset(ib_attr, 0, sizeof(*ib_attr));
130 	memcpy(&ib_attr->fw_ver, dev_attr->fw_ver,
131 	       min(sizeof(dev_attr->fw_ver),
132 		   sizeof(ib_attr->fw_ver)));
133 	bnxt_qplib_get_guid(rdev->netdev->dev_addr,
134 			    (u8 *)&ib_attr->sys_image_guid);
135 	ib_attr->max_mr_size = BNXT_RE_MAX_MR_SIZE;
136 	ib_attr->page_size_cap = BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M;
137 
138 	ib_attr->vendor_id = rdev->en_dev->pdev->vendor;
139 	ib_attr->vendor_part_id = rdev->en_dev->pdev->device;
140 	ib_attr->hw_ver = rdev->en_dev->pdev->subsystem_device;
141 	ib_attr->max_qp = dev_attr->max_qp;
142 	ib_attr->max_qp_wr = dev_attr->max_qp_wqes;
143 	ib_attr->device_cap_flags =
144 				    IB_DEVICE_CURR_QP_STATE_MOD
145 				    | IB_DEVICE_RC_RNR_NAK_GEN
146 				    | IB_DEVICE_SHUTDOWN_PORT
147 				    | IB_DEVICE_SYS_IMAGE_GUID
148 				    | IB_DEVICE_LOCAL_DMA_LKEY
149 				    | IB_DEVICE_RESIZE_MAX_WR
150 				    | IB_DEVICE_PORT_ACTIVE_EVENT
151 				    | IB_DEVICE_N_NOTIFY_CQ
152 				    | IB_DEVICE_MEM_WINDOW
153 				    | IB_DEVICE_MEM_WINDOW_TYPE_2B
154 				    | IB_DEVICE_MEM_MGT_EXTENSIONS;
155 	ib_attr->max_send_sge = dev_attr->max_qp_sges;
156 	ib_attr->max_recv_sge = dev_attr->max_qp_sges;
157 	ib_attr->max_sge_rd = dev_attr->max_qp_sges;
158 	ib_attr->max_cq = dev_attr->max_cq;
159 	ib_attr->max_cqe = dev_attr->max_cq_wqes;
160 	ib_attr->max_mr = dev_attr->max_mr;
161 	ib_attr->max_pd = dev_attr->max_pd;
162 	ib_attr->max_qp_rd_atom = dev_attr->max_qp_rd_atom;
163 	ib_attr->max_qp_init_rd_atom = dev_attr->max_qp_init_rd_atom;
164 	ib_attr->atomic_cap = IB_ATOMIC_NONE;
165 	ib_attr->masked_atomic_cap = IB_ATOMIC_NONE;
166 
167 	ib_attr->max_ee_rd_atom = 0;
168 	ib_attr->max_res_rd_atom = 0;
169 	ib_attr->max_ee_init_rd_atom = 0;
170 	ib_attr->max_ee = 0;
171 	ib_attr->max_rdd = 0;
172 	ib_attr->max_mw = dev_attr->max_mw;
173 	ib_attr->max_raw_ipv6_qp = 0;
174 	ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp;
175 	ib_attr->max_mcast_grp = 0;
176 	ib_attr->max_mcast_qp_attach = 0;
177 	ib_attr->max_total_mcast_qp_attach = 0;
178 	ib_attr->max_ah = dev_attr->max_ah;
179 
180 	ib_attr->max_fmr = 0;
181 	ib_attr->max_map_per_fmr = 0;
182 
183 	ib_attr->max_srq = dev_attr->max_srq;
184 	ib_attr->max_srq_wr = dev_attr->max_srq_wqes;
185 	ib_attr->max_srq_sge = dev_attr->max_srq_sges;
186 
187 	ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS;
188 
189 	ib_attr->max_pkeys = 1;
190 	ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY;
191 	return 0;
192 }
193 
194 int bnxt_re_modify_device(struct ib_device *ibdev,
195 			  int device_modify_mask,
196 			  struct ib_device_modify *device_modify)
197 {
198 	switch (device_modify_mask) {
199 	case IB_DEVICE_MODIFY_SYS_IMAGE_GUID:
200 		/* Modify the GUID requires the modification of the GID table */
201 		/* GUID should be made as READ-ONLY */
202 		break;
203 	case IB_DEVICE_MODIFY_NODE_DESC:
204 		/* Node Desc should be made as READ-ONLY */
205 		break;
206 	default:
207 		break;
208 	}
209 	return 0;
210 }
211 
212 /* Port */
213 int bnxt_re_query_port(struct ib_device *ibdev, u8 port_num,
214 		       struct ib_port_attr *port_attr)
215 {
216 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
217 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
218 
219 	memset(port_attr, 0, sizeof(*port_attr));
220 
221 	if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) {
222 		port_attr->state = IB_PORT_ACTIVE;
223 		port_attr->phys_state = 5;
224 	} else {
225 		port_attr->state = IB_PORT_DOWN;
226 		port_attr->phys_state = 3;
227 	}
228 	port_attr->max_mtu = IB_MTU_4096;
229 	port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu);
230 	port_attr->gid_tbl_len = dev_attr->max_sgid;
231 	port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
232 				    IB_PORT_DEVICE_MGMT_SUP |
233 				    IB_PORT_VENDOR_CLASS_SUP;
234 	port_attr->ip_gids = true;
235 
236 	port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW;
237 	port_attr->bad_pkey_cntr = 0;
238 	port_attr->qkey_viol_cntr = 0;
239 	port_attr->pkey_tbl_len = dev_attr->max_pkey;
240 	port_attr->lid = 0;
241 	port_attr->sm_lid = 0;
242 	port_attr->lmc = 0;
243 	port_attr->max_vl_num = 4;
244 	port_attr->sm_sl = 0;
245 	port_attr->subnet_timeout = 0;
246 	port_attr->init_type_reply = 0;
247 	port_attr->active_speed = rdev->active_speed;
248 	port_attr->active_width = rdev->active_width;
249 
250 	return 0;
251 }
252 
253 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u8 port_num,
254 			       struct ib_port_immutable *immutable)
255 {
256 	struct ib_port_attr port_attr;
257 
258 	if (bnxt_re_query_port(ibdev, port_num, &port_attr))
259 		return -EINVAL;
260 
261 	immutable->pkey_tbl_len = port_attr.pkey_tbl_len;
262 	immutable->gid_tbl_len = port_attr.gid_tbl_len;
263 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
264 	immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP;
265 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
266 	return 0;
267 }
268 
269 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str)
270 {
271 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
272 
273 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d",
274 		 rdev->dev_attr.fw_ver[0], rdev->dev_attr.fw_ver[1],
275 		 rdev->dev_attr.fw_ver[2], rdev->dev_attr.fw_ver[3]);
276 }
277 
278 int bnxt_re_query_pkey(struct ib_device *ibdev, u8 port_num,
279 		       u16 index, u16 *pkey)
280 {
281 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
282 
283 	/* Ignore port_num */
284 
285 	memset(pkey, 0, sizeof(*pkey));
286 	return bnxt_qplib_get_pkey(&rdev->qplib_res,
287 				   &rdev->qplib_res.pkey_tbl, index, pkey);
288 }
289 
290 int bnxt_re_query_gid(struct ib_device *ibdev, u8 port_num,
291 		      int index, union ib_gid *gid)
292 {
293 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
294 	int rc = 0;
295 
296 	/* Ignore port_num */
297 	memset(gid, 0, sizeof(*gid));
298 	rc = bnxt_qplib_get_sgid(&rdev->qplib_res,
299 				 &rdev->qplib_res.sgid_tbl, index,
300 				 (struct bnxt_qplib_gid *)gid);
301 	return rc;
302 }
303 
304 int bnxt_re_del_gid(const struct ib_gid_attr *attr, void **context)
305 {
306 	int rc = 0;
307 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
308 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
309 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
310 	struct bnxt_qplib_gid *gid_to_del;
311 
312 	/* Delete the entry from the hardware */
313 	ctx = *context;
314 	if (!ctx)
315 		return -EINVAL;
316 
317 	if (sgid_tbl && sgid_tbl->active) {
318 		if (ctx->idx >= sgid_tbl->max)
319 			return -EINVAL;
320 		gid_to_del = &sgid_tbl->tbl[ctx->idx];
321 		/* DEL_GID is called in WQ context(netdevice_event_work_handler)
322 		 * or via the ib_unregister_device path. In the former case QP1
323 		 * may not be destroyed yet, in which case just return as FW
324 		 * needs that entry to be present and will fail it's deletion.
325 		 * We could get invoked again after QP1 is destroyed OR get an
326 		 * ADD_GID call with a different GID value for the same index
327 		 * where we issue MODIFY_GID cmd to update the GID entry -- TBD
328 		 */
329 		if (ctx->idx == 0 &&
330 		    rdma_link_local_addr((struct in6_addr *)gid_to_del) &&
331 		    ctx->refcnt == 1 && rdev->qp1_sqp) {
332 			dev_dbg(rdev_to_dev(rdev),
333 				"Trying to delete GID0 while QP1 is alive\n");
334 			return -EFAULT;
335 		}
336 		ctx->refcnt--;
337 		if (!ctx->refcnt) {
338 			rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del, true);
339 			if (rc) {
340 				dev_err(rdev_to_dev(rdev),
341 					"Failed to remove GID: %#x", rc);
342 			} else {
343 				ctx_tbl = sgid_tbl->ctx;
344 				ctx_tbl[ctx->idx] = NULL;
345 				kfree(ctx);
346 			}
347 		}
348 	} else {
349 		return -EINVAL;
350 	}
351 	return rc;
352 }
353 
354 int bnxt_re_add_gid(const struct ib_gid_attr *attr, void **context)
355 {
356 	int rc;
357 	u32 tbl_idx = 0;
358 	u16 vlan_id = 0xFFFF;
359 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
360 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
361 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
362 
363 	rc = rdma_read_gid_l2_fields(attr, &vlan_id, NULL);
364 	if (rc)
365 		return rc;
366 
367 	rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)&attr->gid,
368 				 rdev->qplib_res.netdev->dev_addr,
369 				 vlan_id, true, &tbl_idx);
370 	if (rc == -EALREADY) {
371 		ctx_tbl = sgid_tbl->ctx;
372 		ctx_tbl[tbl_idx]->refcnt++;
373 		*context = ctx_tbl[tbl_idx];
374 		return 0;
375 	}
376 
377 	if (rc < 0) {
378 		dev_err(rdev_to_dev(rdev), "Failed to add GID: %#x", rc);
379 		return rc;
380 	}
381 
382 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
383 	if (!ctx)
384 		return -ENOMEM;
385 	ctx_tbl = sgid_tbl->ctx;
386 	ctx->idx = tbl_idx;
387 	ctx->refcnt = 1;
388 	ctx_tbl[tbl_idx] = ctx;
389 	*context = ctx;
390 
391 	return rc;
392 }
393 
394 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
395 					    u8 port_num)
396 {
397 	return IB_LINK_LAYER_ETHERNET;
398 }
399 
400 #define	BNXT_RE_FENCE_PBL_SIZE	DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
401 
402 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd)
403 {
404 	struct bnxt_re_fence_data *fence = &pd->fence;
405 	struct ib_mr *ib_mr = &fence->mr->ib_mr;
406 	struct bnxt_qplib_swqe *wqe = &fence->bind_wqe;
407 
408 	memset(wqe, 0, sizeof(*wqe));
409 	wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW;
410 	wqe->wr_id = BNXT_QPLIB_FENCE_WRID;
411 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
412 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
413 	wqe->bind.zero_based = false;
414 	wqe->bind.parent_l_key = ib_mr->lkey;
415 	wqe->bind.va = (u64)(unsigned long)fence->va;
416 	wqe->bind.length = fence->size;
417 	wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ);
418 	wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1;
419 
420 	/* Save the initial rkey in fence structure for now;
421 	 * wqe->bind.r_key will be set at (re)bind time.
422 	 */
423 	fence->bind_rkey = ib_inc_rkey(fence->mw->rkey);
424 }
425 
426 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp)
427 {
428 	struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp,
429 					     qplib_qp);
430 	struct ib_pd *ib_pd = qp->ib_qp.pd;
431 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
432 	struct bnxt_re_fence_data *fence = &pd->fence;
433 	struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe;
434 	struct bnxt_qplib_swqe wqe;
435 	int rc;
436 
437 	memcpy(&wqe, fence_wqe, sizeof(wqe));
438 	wqe.bind.r_key = fence->bind_rkey;
439 	fence->bind_rkey = ib_inc_rkey(fence->bind_rkey);
440 
441 	dev_dbg(rdev_to_dev(qp->rdev),
442 		"Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
443 		wqe.bind.r_key, qp->qplib_qp.id, pd);
444 	rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
445 	if (rc) {
446 		dev_err(rdev_to_dev(qp->rdev), "Failed to bind fence-WQE\n");
447 		return rc;
448 	}
449 	bnxt_qplib_post_send_db(&qp->qplib_qp);
450 
451 	return rc;
452 }
453 
454 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd)
455 {
456 	struct bnxt_re_fence_data *fence = &pd->fence;
457 	struct bnxt_re_dev *rdev = pd->rdev;
458 	struct device *dev = &rdev->en_dev->pdev->dev;
459 	struct bnxt_re_mr *mr = fence->mr;
460 
461 	if (fence->mw) {
462 		bnxt_re_dealloc_mw(fence->mw);
463 		fence->mw = NULL;
464 	}
465 	if (mr) {
466 		if (mr->ib_mr.rkey)
467 			bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr,
468 					     true);
469 		if (mr->ib_mr.lkey)
470 			bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
471 		kfree(mr);
472 		fence->mr = NULL;
473 	}
474 	if (fence->dma_addr) {
475 		dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES,
476 				 DMA_BIDIRECTIONAL);
477 		fence->dma_addr = 0;
478 	}
479 }
480 
481 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
482 {
483 	int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND;
484 	struct bnxt_re_fence_data *fence = &pd->fence;
485 	struct bnxt_re_dev *rdev = pd->rdev;
486 	struct device *dev = &rdev->en_dev->pdev->dev;
487 	struct bnxt_re_mr *mr = NULL;
488 	dma_addr_t dma_addr = 0;
489 	struct ib_mw *mw;
490 	u64 pbl_tbl;
491 	int rc;
492 
493 	dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES,
494 				  DMA_BIDIRECTIONAL);
495 	rc = dma_mapping_error(dev, dma_addr);
496 	if (rc) {
497 		dev_err(rdev_to_dev(rdev), "Failed to dma-map fence-MR-mem\n");
498 		rc = -EIO;
499 		fence->dma_addr = 0;
500 		goto fail;
501 	}
502 	fence->dma_addr = dma_addr;
503 
504 	/* Allocate a MR */
505 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
506 	if (!mr) {
507 		rc = -ENOMEM;
508 		goto fail;
509 	}
510 	fence->mr = mr;
511 	mr->rdev = rdev;
512 	mr->qplib_mr.pd = &pd->qplib_pd;
513 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
514 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
515 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
516 	if (rc) {
517 		dev_err(rdev_to_dev(rdev), "Failed to alloc fence-HW-MR\n");
518 		goto fail;
519 	}
520 
521 	/* Register MR */
522 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
523 	mr->qplib_mr.va = (u64)(unsigned long)fence->va;
524 	mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES;
525 	pbl_tbl = dma_addr;
526 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl_tbl,
527 			       BNXT_RE_FENCE_PBL_SIZE, false, PAGE_SIZE);
528 	if (rc) {
529 		dev_err(rdev_to_dev(rdev), "Failed to register fence-MR\n");
530 		goto fail;
531 	}
532 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
533 
534 	/* Create a fence MW only for kernel consumers */
535 	mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL);
536 	if (IS_ERR(mw)) {
537 		dev_err(rdev_to_dev(rdev),
538 			"Failed to create fence-MW for PD: %p\n", pd);
539 		rc = PTR_ERR(mw);
540 		goto fail;
541 	}
542 	fence->mw = mw;
543 
544 	bnxt_re_create_fence_wqe(pd);
545 	return 0;
546 
547 fail:
548 	bnxt_re_destroy_fence_mr(pd);
549 	return rc;
550 }
551 
552 /* Protection Domains */
553 void bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
554 {
555 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
556 	struct bnxt_re_dev *rdev = pd->rdev;
557 
558 	bnxt_re_destroy_fence_mr(pd);
559 
560 	if (pd->qplib_pd.id)
561 		bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
562 				      &pd->qplib_pd);
563 }
564 
565 int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
566 {
567 	struct ib_device *ibdev = ibpd->device;
568 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
569 	struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context(
570 		udata, struct bnxt_re_ucontext, ib_uctx);
571 	struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd);
572 	int rc;
573 
574 	pd->rdev = rdev;
575 	if (bnxt_qplib_alloc_pd(&rdev->qplib_res.pd_tbl, &pd->qplib_pd)) {
576 		dev_err(rdev_to_dev(rdev), "Failed to allocate HW PD");
577 		rc = -ENOMEM;
578 		goto fail;
579 	}
580 
581 	if (udata) {
582 		struct bnxt_re_pd_resp resp;
583 
584 		if (!ucntx->dpi.dbr) {
585 			/* Allocate DPI in alloc_pd to avoid failing of
586 			 * ibv_devinfo and family of application when DPIs
587 			 * are depleted.
588 			 */
589 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res.dpi_tbl,
590 						 &ucntx->dpi, ucntx)) {
591 				rc = -ENOMEM;
592 				goto dbfail;
593 			}
594 		}
595 
596 		resp.pdid = pd->qplib_pd.id;
597 		/* Still allow mapping this DBR to the new user PD. */
598 		resp.dpi = ucntx->dpi.dpi;
599 		resp.dbr = (u64)ucntx->dpi.umdbr;
600 
601 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
602 		if (rc) {
603 			dev_err(rdev_to_dev(rdev),
604 				"Failed to copy user response\n");
605 			goto dbfail;
606 		}
607 	}
608 
609 	if (!udata)
610 		if (bnxt_re_create_fence_mr(pd))
611 			dev_warn(rdev_to_dev(rdev),
612 				 "Failed to create Fence-MR\n");
613 	return 0;
614 dbfail:
615 	bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
616 			      &pd->qplib_pd);
617 fail:
618 	return rc;
619 }
620 
621 /* Address Handles */
622 void bnxt_re_destroy_ah(struct ib_ah *ib_ah, u32 flags)
623 {
624 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
625 	struct bnxt_re_dev *rdev = ah->rdev;
626 
627 	bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah,
628 			      !(flags & RDMA_DESTROY_AH_SLEEPABLE));
629 }
630 
631 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)
632 {
633 	u8 nw_type;
634 
635 	switch (ntype) {
636 	case RDMA_NETWORK_IPV4:
637 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4;
638 		break;
639 	case RDMA_NETWORK_IPV6:
640 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6;
641 		break;
642 	default:
643 		nw_type = CMDQ_CREATE_AH_TYPE_V1;
644 		break;
645 	}
646 	return nw_type;
647 }
648 
649 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr,
650 		      u32 flags, struct ib_udata *udata)
651 {
652 	struct ib_pd *ib_pd = ib_ah->pd;
653 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
654 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
655 	struct bnxt_re_dev *rdev = pd->rdev;
656 	const struct ib_gid_attr *sgid_attr;
657 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
658 	u8 nw_type;
659 	int rc;
660 
661 	if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) {
662 		dev_err(rdev_to_dev(rdev), "Failed to alloc AH: GRH not set");
663 		return -EINVAL;
664 	}
665 
666 	ah->rdev = rdev;
667 	ah->qplib_ah.pd = &pd->qplib_pd;
668 
669 	/* Supply the configuration for the HW */
670 	memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw,
671 	       sizeof(union ib_gid));
672 	/*
673 	 * If RoCE V2 is enabled, stack will have two entries for
674 	 * each GID entry. Avoiding this duplicte entry in HW. Dividing
675 	 * the GID index by 2 for RoCE V2
676 	 */
677 	ah->qplib_ah.sgid_index = grh->sgid_index / 2;
678 	ah->qplib_ah.host_sgid_index = grh->sgid_index;
679 	ah->qplib_ah.traffic_class = grh->traffic_class;
680 	ah->qplib_ah.flow_label = grh->flow_label;
681 	ah->qplib_ah.hop_limit = grh->hop_limit;
682 	ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr);
683 
684 	sgid_attr = grh->sgid_attr;
685 	/* Get network header type for this GID */
686 	nw_type = rdma_gid_attr_network_type(sgid_attr);
687 	ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type);
688 
689 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
690 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah,
691 				  !(flags & RDMA_CREATE_AH_SLEEPABLE));
692 	if (rc) {
693 		dev_err(rdev_to_dev(rdev), "Failed to allocate HW AH");
694 		return rc;
695 	}
696 
697 	/* Write AVID to shared page. */
698 	if (udata) {
699 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
700 			udata, struct bnxt_re_ucontext, ib_uctx);
701 		unsigned long flag;
702 		u32 *wrptr;
703 
704 		spin_lock_irqsave(&uctx->sh_lock, flag);
705 		wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT);
706 		*wrptr = ah->qplib_ah.id;
707 		wmb(); /* make sure cache is updated. */
708 		spin_unlock_irqrestore(&uctx->sh_lock, flag);
709 	}
710 
711 	return 0;
712 }
713 
714 int bnxt_re_modify_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
715 {
716 	return 0;
717 }
718 
719 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
720 {
721 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
722 
723 	ah_attr->type = ib_ah->type;
724 	rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl);
725 	memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN);
726 	rdma_ah_set_grh(ah_attr, NULL, 0,
727 			ah->qplib_ah.host_sgid_index,
728 			0, ah->qplib_ah.traffic_class);
729 	rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data);
730 	rdma_ah_set_port_num(ah_attr, 1);
731 	rdma_ah_set_static_rate(ah_attr, 0);
732 	return 0;
733 }
734 
735 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp)
736 	__acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock)
737 {
738 	unsigned long flags;
739 
740 	spin_lock_irqsave(&qp->scq->cq_lock, flags);
741 	if (qp->rcq != qp->scq)
742 		spin_lock(&qp->rcq->cq_lock);
743 	else
744 		__acquire(&qp->rcq->cq_lock);
745 
746 	return flags;
747 }
748 
749 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp,
750 			unsigned long flags)
751 	__releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock)
752 {
753 	if (qp->rcq != qp->scq)
754 		spin_unlock(&qp->rcq->cq_lock);
755 	else
756 		__release(&qp->rcq->cq_lock);
757 	spin_unlock_irqrestore(&qp->scq->cq_lock, flags);
758 }
759 
760 /* Queue Pairs */
761 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
762 {
763 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
764 	struct bnxt_re_dev *rdev = qp->rdev;
765 	unsigned int flags;
766 	int rc;
767 
768 	bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
769 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
770 	if (rc) {
771 		dev_err(rdev_to_dev(rdev), "Failed to destroy HW QP");
772 		return rc;
773 	}
774 
775 	if (rdma_is_kernel_res(&qp->ib_qp.res)) {
776 		flags = bnxt_re_lock_cqs(qp);
777 		bnxt_qplib_clean_qp(&qp->qplib_qp);
778 		bnxt_re_unlock_cqs(qp, flags);
779 	}
780 
781 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
782 
783 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->qp1_sqp) {
784 		bnxt_qplib_destroy_ah(&rdev->qplib_res, &rdev->sqp_ah->qplib_ah,
785 				      false);
786 
787 		bnxt_qplib_clean_qp(&qp->qplib_qp);
788 		rc = bnxt_qplib_destroy_qp(&rdev->qplib_res,
789 					   &rdev->qp1_sqp->qplib_qp);
790 		if (rc) {
791 			dev_err(rdev_to_dev(rdev),
792 				"Failed to destroy Shadow QP");
793 			return rc;
794 		}
795 		bnxt_qplib_free_qp_res(&rdev->qplib_res,
796 				       &rdev->qp1_sqp->qplib_qp);
797 		mutex_lock(&rdev->qp_lock);
798 		list_del(&rdev->qp1_sqp->list);
799 		atomic_dec(&rdev->qp_count);
800 		mutex_unlock(&rdev->qp_lock);
801 
802 		kfree(rdev->sqp_ah);
803 		kfree(rdev->qp1_sqp);
804 		rdev->qp1_sqp = NULL;
805 		rdev->sqp_ah = NULL;
806 	}
807 
808 	ib_umem_release(qp->rumem);
809 	ib_umem_release(qp->sumem);
810 
811 	mutex_lock(&rdev->qp_lock);
812 	list_del(&qp->list);
813 	atomic_dec(&rdev->qp_count);
814 	mutex_unlock(&rdev->qp_lock);
815 	kfree(qp);
816 	return 0;
817 }
818 
819 static u8 __from_ib_qp_type(enum ib_qp_type type)
820 {
821 	switch (type) {
822 	case IB_QPT_GSI:
823 		return CMDQ_CREATE_QP1_TYPE_GSI;
824 	case IB_QPT_RC:
825 		return CMDQ_CREATE_QP_TYPE_RC;
826 	case IB_QPT_UD:
827 		return CMDQ_CREATE_QP_TYPE_UD;
828 	default:
829 		return IB_QPT_MAX;
830 	}
831 }
832 
833 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
834 				struct bnxt_re_qp *qp, struct ib_udata *udata)
835 {
836 	struct bnxt_re_qp_req ureq;
837 	struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
838 	struct ib_umem *umem;
839 	int bytes = 0, psn_sz;
840 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
841 		udata, struct bnxt_re_ucontext, ib_uctx);
842 
843 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
844 		return -EFAULT;
845 
846 	bytes = (qplib_qp->sq.max_wqe * BNXT_QPLIB_MAX_SQE_ENTRY_SIZE);
847 	/* Consider mapping PSN search memory only for RC QPs. */
848 	if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) {
849 		psn_sz = bnxt_qplib_is_chip_gen_p5(&rdev->chip_ctx) ?
850 					sizeof(struct sq_psn_search_ext) :
851 					sizeof(struct sq_psn_search);
852 		bytes += (qplib_qp->sq.max_wqe * psn_sz);
853 	}
854 	bytes = PAGE_ALIGN(bytes);
855 	umem = ib_umem_get(udata, ureq.qpsva, bytes, IB_ACCESS_LOCAL_WRITE, 1);
856 	if (IS_ERR(umem))
857 		return PTR_ERR(umem);
858 
859 	qp->sumem = umem;
860 	qplib_qp->sq.sg_info.sglist = umem->sg_head.sgl;
861 	qplib_qp->sq.sg_info.npages = ib_umem_num_pages(umem);
862 	qplib_qp->sq.sg_info.nmap = umem->nmap;
863 	qplib_qp->qp_handle = ureq.qp_handle;
864 
865 	if (!qp->qplib_qp.srq) {
866 		bytes = (qplib_qp->rq.max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
867 		bytes = PAGE_ALIGN(bytes);
868 		umem = ib_umem_get(udata, ureq.qprva, bytes,
869 				   IB_ACCESS_LOCAL_WRITE, 1);
870 		if (IS_ERR(umem))
871 			goto rqfail;
872 		qp->rumem = umem;
873 		qplib_qp->rq.sg_info.sglist = umem->sg_head.sgl;
874 		qplib_qp->rq.sg_info.npages = ib_umem_num_pages(umem);
875 		qplib_qp->rq.sg_info.nmap = umem->nmap;
876 	}
877 
878 	qplib_qp->dpi = &cntx->dpi;
879 	return 0;
880 rqfail:
881 	ib_umem_release(qp->sumem);
882 	qp->sumem = NULL;
883 	memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info));
884 
885 	return PTR_ERR(umem);
886 }
887 
888 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
889 				(struct bnxt_re_pd *pd,
890 				 struct bnxt_qplib_res *qp1_res,
891 				 struct bnxt_qplib_qp *qp1_qp)
892 {
893 	struct bnxt_re_dev *rdev = pd->rdev;
894 	struct bnxt_re_ah *ah;
895 	union ib_gid sgid;
896 	int rc;
897 
898 	ah = kzalloc(sizeof(*ah), GFP_KERNEL);
899 	if (!ah)
900 		return NULL;
901 
902 	ah->rdev = rdev;
903 	ah->qplib_ah.pd = &pd->qplib_pd;
904 
905 	rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid);
906 	if (rc)
907 		goto fail;
908 
909 	/* supply the dgid data same as sgid */
910 	memcpy(ah->qplib_ah.dgid.data, &sgid.raw,
911 	       sizeof(union ib_gid));
912 	ah->qplib_ah.sgid_index = 0;
913 
914 	ah->qplib_ah.traffic_class = 0;
915 	ah->qplib_ah.flow_label = 0;
916 	ah->qplib_ah.hop_limit = 1;
917 	ah->qplib_ah.sl = 0;
918 	/* Have DMAC same as SMAC */
919 	ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr);
920 
921 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false);
922 	if (rc) {
923 		dev_err(rdev_to_dev(rdev),
924 			"Failed to allocate HW AH for Shadow QP");
925 		goto fail;
926 	}
927 
928 	return ah;
929 
930 fail:
931 	kfree(ah);
932 	return NULL;
933 }
934 
935 static struct bnxt_re_qp *bnxt_re_create_shadow_qp
936 				(struct bnxt_re_pd *pd,
937 				 struct bnxt_qplib_res *qp1_res,
938 				 struct bnxt_qplib_qp *qp1_qp)
939 {
940 	struct bnxt_re_dev *rdev = pd->rdev;
941 	struct bnxt_re_qp *qp;
942 	int rc;
943 
944 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
945 	if (!qp)
946 		return NULL;
947 
948 	qp->rdev = rdev;
949 
950 	/* Initialize the shadow QP structure from the QP1 values */
951 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
952 
953 	qp->qplib_qp.pd = &pd->qplib_pd;
954 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
955 	qp->qplib_qp.type = IB_QPT_UD;
956 
957 	qp->qplib_qp.max_inline_data = 0;
958 	qp->qplib_qp.sig_type = true;
959 
960 	/* Shadow QP SQ depth should be same as QP1 RQ depth */
961 	qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
962 	qp->qplib_qp.sq.max_sge = 2;
963 	/* Q full delta can be 1 since it is internal QP */
964 	qp->qplib_qp.sq.q_full_delta = 1;
965 
966 	qp->qplib_qp.scq = qp1_qp->scq;
967 	qp->qplib_qp.rcq = qp1_qp->rcq;
968 
969 	qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
970 	qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
971 	/* Q full delta can be 1 since it is internal QP */
972 	qp->qplib_qp.rq.q_full_delta = 1;
973 
974 	qp->qplib_qp.mtu = qp1_qp->mtu;
975 
976 	qp->qplib_qp.sq_hdr_buf_size = 0;
977 	qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
978 	qp->qplib_qp.dpi = &rdev->dpi_privileged;
979 
980 	rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
981 	if (rc)
982 		goto fail;
983 
984 	rdev->sqp_id = qp->qplib_qp.id;
985 
986 	spin_lock_init(&qp->sq_lock);
987 	INIT_LIST_HEAD(&qp->list);
988 	mutex_lock(&rdev->qp_lock);
989 	list_add_tail(&qp->list, &rdev->qp_list);
990 	atomic_inc(&rdev->qp_count);
991 	mutex_unlock(&rdev->qp_lock);
992 	return qp;
993 fail:
994 	kfree(qp);
995 	return NULL;
996 }
997 
998 struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
999 				struct ib_qp_init_attr *qp_init_attr,
1000 				struct ib_udata *udata)
1001 {
1002 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1003 	struct bnxt_re_dev *rdev = pd->rdev;
1004 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1005 	struct bnxt_re_qp *qp;
1006 	struct bnxt_re_cq *cq;
1007 	struct bnxt_re_srq *srq;
1008 	int rc, entries;
1009 
1010 	if ((qp_init_attr->cap.max_send_wr > dev_attr->max_qp_wqes) ||
1011 	    (qp_init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes) ||
1012 	    (qp_init_attr->cap.max_send_sge > dev_attr->max_qp_sges) ||
1013 	    (qp_init_attr->cap.max_recv_sge > dev_attr->max_qp_sges) ||
1014 	    (qp_init_attr->cap.max_inline_data > dev_attr->max_inline_data))
1015 		return ERR_PTR(-EINVAL);
1016 
1017 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1018 	if (!qp)
1019 		return ERR_PTR(-ENOMEM);
1020 
1021 	qp->rdev = rdev;
1022 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
1023 	qp->qplib_qp.pd = &pd->qplib_pd;
1024 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
1025 	qp->qplib_qp.type = __from_ib_qp_type(qp_init_attr->qp_type);
1026 
1027 	if (qp_init_attr->qp_type == IB_QPT_GSI &&
1028 	    bnxt_qplib_is_chip_gen_p5(&rdev->chip_ctx))
1029 		qp->qplib_qp.type = CMDQ_CREATE_QP_TYPE_GSI;
1030 	if (qp->qplib_qp.type == IB_QPT_MAX) {
1031 		dev_err(rdev_to_dev(rdev), "QP type 0x%x not supported",
1032 			qp->qplib_qp.type);
1033 		rc = -EINVAL;
1034 		goto fail;
1035 	}
1036 
1037 	qp->qplib_qp.max_inline_data = qp_init_attr->cap.max_inline_data;
1038 	qp->qplib_qp.sig_type = ((qp_init_attr->sq_sig_type ==
1039 				  IB_SIGNAL_ALL_WR) ? true : false);
1040 
1041 	qp->qplib_qp.sq.max_sge = qp_init_attr->cap.max_send_sge;
1042 	if (qp->qplib_qp.sq.max_sge > dev_attr->max_qp_sges)
1043 		qp->qplib_qp.sq.max_sge = dev_attr->max_qp_sges;
1044 
1045 	if (qp_init_attr->send_cq) {
1046 		cq = container_of(qp_init_attr->send_cq, struct bnxt_re_cq,
1047 				  ib_cq);
1048 		if (!cq) {
1049 			dev_err(rdev_to_dev(rdev), "Send CQ not found");
1050 			rc = -EINVAL;
1051 			goto fail;
1052 		}
1053 		qp->qplib_qp.scq = &cq->qplib_cq;
1054 		qp->scq = cq;
1055 	}
1056 
1057 	if (qp_init_attr->recv_cq) {
1058 		cq = container_of(qp_init_attr->recv_cq, struct bnxt_re_cq,
1059 				  ib_cq);
1060 		if (!cq) {
1061 			dev_err(rdev_to_dev(rdev), "Receive CQ not found");
1062 			rc = -EINVAL;
1063 			goto fail;
1064 		}
1065 		qp->qplib_qp.rcq = &cq->qplib_cq;
1066 		qp->rcq = cq;
1067 	}
1068 
1069 	if (qp_init_attr->srq) {
1070 		srq = container_of(qp_init_attr->srq, struct bnxt_re_srq,
1071 				   ib_srq);
1072 		if (!srq) {
1073 			dev_err(rdev_to_dev(rdev), "SRQ not found");
1074 			rc = -EINVAL;
1075 			goto fail;
1076 		}
1077 		qp->qplib_qp.srq = &srq->qplib_srq;
1078 		qp->qplib_qp.rq.max_wqe = 0;
1079 	} else {
1080 		/* Allocate 1 more than what's provided so posting max doesn't
1081 		 * mean empty
1082 		 */
1083 		entries = roundup_pow_of_two(qp_init_attr->cap.max_recv_wr + 1);
1084 		qp->qplib_qp.rq.max_wqe = min_t(u32, entries,
1085 						dev_attr->max_qp_wqes + 1);
1086 
1087 		qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
1088 						qp_init_attr->cap.max_recv_wr;
1089 
1090 		qp->qplib_qp.rq.max_sge = qp_init_attr->cap.max_recv_sge;
1091 		if (qp->qplib_qp.rq.max_sge > dev_attr->max_qp_sges)
1092 			qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
1093 	}
1094 
1095 	qp->qplib_qp.mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1096 
1097 	if (qp_init_attr->qp_type == IB_QPT_GSI &&
1098 	    !(bnxt_qplib_is_chip_gen_p5(&rdev->chip_ctx))) {
1099 		/* Allocate 1 more than what's provided */
1100 		entries = roundup_pow_of_two(qp_init_attr->cap.max_send_wr + 1);
1101 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
1102 						dev_attr->max_qp_wqes + 1);
1103 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
1104 						qp_init_attr->cap.max_send_wr;
1105 		qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
1106 		if (qp->qplib_qp.rq.max_sge > dev_attr->max_qp_sges)
1107 			qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
1108 		qp->qplib_qp.sq.max_sge++;
1109 		if (qp->qplib_qp.sq.max_sge > dev_attr->max_qp_sges)
1110 			qp->qplib_qp.sq.max_sge = dev_attr->max_qp_sges;
1111 
1112 		qp->qplib_qp.rq_hdr_buf_size =
1113 					BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
1114 
1115 		qp->qplib_qp.sq_hdr_buf_size =
1116 					BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
1117 		qp->qplib_qp.dpi = &rdev->dpi_privileged;
1118 		rc = bnxt_qplib_create_qp1(&rdev->qplib_res, &qp->qplib_qp);
1119 		if (rc) {
1120 			dev_err(rdev_to_dev(rdev), "Failed to create HW QP1");
1121 			goto fail;
1122 		}
1123 		/* Create a shadow QP to handle the QP1 traffic */
1124 		rdev->qp1_sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res,
1125 							 &qp->qplib_qp);
1126 		if (!rdev->qp1_sqp) {
1127 			rc = -EINVAL;
1128 			dev_err(rdev_to_dev(rdev),
1129 				"Failed to create Shadow QP for QP1");
1130 			goto qp_destroy;
1131 		}
1132 		rdev->sqp_ah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res,
1133 							   &qp->qplib_qp);
1134 		if (!rdev->sqp_ah) {
1135 			bnxt_qplib_destroy_qp(&rdev->qplib_res,
1136 					      &rdev->qp1_sqp->qplib_qp);
1137 			rc = -EINVAL;
1138 			dev_err(rdev_to_dev(rdev),
1139 				"Failed to create AH entry for ShadowQP");
1140 			goto qp_destroy;
1141 		}
1142 
1143 	} else {
1144 		/* Allocate 128 + 1 more than what's provided */
1145 		entries = roundup_pow_of_two(qp_init_attr->cap.max_send_wr +
1146 					     BNXT_QPLIB_RESERVED_QP_WRS + 1);
1147 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
1148 						dev_attr->max_qp_wqes +
1149 						BNXT_QPLIB_RESERVED_QP_WRS + 1);
1150 		qp->qplib_qp.sq.q_full_delta = BNXT_QPLIB_RESERVED_QP_WRS + 1;
1151 
1152 		/*
1153 		 * Reserving one slot for Phantom WQE. Application can
1154 		 * post one extra entry in this case. But allowing this to avoid
1155 		 * unexpected Queue full condition
1156 		 */
1157 
1158 		qp->qplib_qp.sq.q_full_delta -= 1;
1159 
1160 		qp->qplib_qp.max_rd_atomic = dev_attr->max_qp_rd_atom;
1161 		qp->qplib_qp.max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
1162 		if (udata) {
1163 			rc = bnxt_re_init_user_qp(rdev, pd, qp, udata);
1164 			if (rc)
1165 				goto fail;
1166 		} else {
1167 			qp->qplib_qp.dpi = &rdev->dpi_privileged;
1168 		}
1169 
1170 		rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
1171 		if (rc) {
1172 			dev_err(rdev_to_dev(rdev), "Failed to create HW QP");
1173 			goto free_umem;
1174 		}
1175 	}
1176 
1177 	qp->ib_qp.qp_num = qp->qplib_qp.id;
1178 	spin_lock_init(&qp->sq_lock);
1179 	spin_lock_init(&qp->rq_lock);
1180 
1181 	if (udata) {
1182 		struct bnxt_re_qp_resp resp;
1183 
1184 		resp.qpid = qp->ib_qp.qp_num;
1185 		resp.rsvd = 0;
1186 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1187 		if (rc) {
1188 			dev_err(rdev_to_dev(rdev), "Failed to copy QP udata");
1189 			goto qp_destroy;
1190 		}
1191 	}
1192 	INIT_LIST_HEAD(&qp->list);
1193 	mutex_lock(&rdev->qp_lock);
1194 	list_add_tail(&qp->list, &rdev->qp_list);
1195 	atomic_inc(&rdev->qp_count);
1196 	mutex_unlock(&rdev->qp_lock);
1197 
1198 	return &qp->ib_qp;
1199 qp_destroy:
1200 	bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
1201 free_umem:
1202 	ib_umem_release(qp->rumem);
1203 	ib_umem_release(qp->sumem);
1204 fail:
1205 	kfree(qp);
1206 	return ERR_PTR(rc);
1207 }
1208 
1209 static u8 __from_ib_qp_state(enum ib_qp_state state)
1210 {
1211 	switch (state) {
1212 	case IB_QPS_RESET:
1213 		return CMDQ_MODIFY_QP_NEW_STATE_RESET;
1214 	case IB_QPS_INIT:
1215 		return CMDQ_MODIFY_QP_NEW_STATE_INIT;
1216 	case IB_QPS_RTR:
1217 		return CMDQ_MODIFY_QP_NEW_STATE_RTR;
1218 	case IB_QPS_RTS:
1219 		return CMDQ_MODIFY_QP_NEW_STATE_RTS;
1220 	case IB_QPS_SQD:
1221 		return CMDQ_MODIFY_QP_NEW_STATE_SQD;
1222 	case IB_QPS_SQE:
1223 		return CMDQ_MODIFY_QP_NEW_STATE_SQE;
1224 	case IB_QPS_ERR:
1225 	default:
1226 		return CMDQ_MODIFY_QP_NEW_STATE_ERR;
1227 	}
1228 }
1229 
1230 static enum ib_qp_state __to_ib_qp_state(u8 state)
1231 {
1232 	switch (state) {
1233 	case CMDQ_MODIFY_QP_NEW_STATE_RESET:
1234 		return IB_QPS_RESET;
1235 	case CMDQ_MODIFY_QP_NEW_STATE_INIT:
1236 		return IB_QPS_INIT;
1237 	case CMDQ_MODIFY_QP_NEW_STATE_RTR:
1238 		return IB_QPS_RTR;
1239 	case CMDQ_MODIFY_QP_NEW_STATE_RTS:
1240 		return IB_QPS_RTS;
1241 	case CMDQ_MODIFY_QP_NEW_STATE_SQD:
1242 		return IB_QPS_SQD;
1243 	case CMDQ_MODIFY_QP_NEW_STATE_SQE:
1244 		return IB_QPS_SQE;
1245 	case CMDQ_MODIFY_QP_NEW_STATE_ERR:
1246 	default:
1247 		return IB_QPS_ERR;
1248 	}
1249 }
1250 
1251 static u32 __from_ib_mtu(enum ib_mtu mtu)
1252 {
1253 	switch (mtu) {
1254 	case IB_MTU_256:
1255 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_256;
1256 	case IB_MTU_512:
1257 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_512;
1258 	case IB_MTU_1024:
1259 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024;
1260 	case IB_MTU_2048:
1261 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1262 	case IB_MTU_4096:
1263 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096;
1264 	default:
1265 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1266 	}
1267 }
1268 
1269 static enum ib_mtu __to_ib_mtu(u32 mtu)
1270 {
1271 	switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) {
1272 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_256:
1273 		return IB_MTU_256;
1274 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_512:
1275 		return IB_MTU_512;
1276 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024:
1277 		return IB_MTU_1024;
1278 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048:
1279 		return IB_MTU_2048;
1280 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096:
1281 		return IB_MTU_4096;
1282 	default:
1283 		return IB_MTU_2048;
1284 	}
1285 }
1286 
1287 /* Shared Receive Queues */
1288 void bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata)
1289 {
1290 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1291 					       ib_srq);
1292 	struct bnxt_re_dev *rdev = srq->rdev;
1293 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1294 	struct bnxt_qplib_nq *nq = NULL;
1295 
1296 	if (qplib_srq->cq)
1297 		nq = qplib_srq->cq->nq;
1298 	bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
1299 	ib_umem_release(srq->umem);
1300 	atomic_dec(&rdev->srq_count);
1301 	if (nq)
1302 		nq->budget--;
1303 }
1304 
1305 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
1306 				 struct bnxt_re_pd *pd,
1307 				 struct bnxt_re_srq *srq,
1308 				 struct ib_udata *udata)
1309 {
1310 	struct bnxt_re_srq_req ureq;
1311 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1312 	struct ib_umem *umem;
1313 	int bytes = 0;
1314 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
1315 		udata, struct bnxt_re_ucontext, ib_uctx);
1316 
1317 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1318 		return -EFAULT;
1319 
1320 	bytes = (qplib_srq->max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
1321 	bytes = PAGE_ALIGN(bytes);
1322 	umem = ib_umem_get(udata, ureq.srqva, bytes, IB_ACCESS_LOCAL_WRITE, 1);
1323 	if (IS_ERR(umem))
1324 		return PTR_ERR(umem);
1325 
1326 	srq->umem = umem;
1327 	qplib_srq->sg_info.sglist = umem->sg_head.sgl;
1328 	qplib_srq->sg_info.npages = ib_umem_num_pages(umem);
1329 	qplib_srq->sg_info.nmap = umem->nmap;
1330 	qplib_srq->srq_handle = ureq.srq_handle;
1331 	qplib_srq->dpi = &cntx->dpi;
1332 
1333 	return 0;
1334 }
1335 
1336 int bnxt_re_create_srq(struct ib_srq *ib_srq,
1337 		       struct ib_srq_init_attr *srq_init_attr,
1338 		       struct ib_udata *udata)
1339 {
1340 	struct ib_pd *ib_pd = ib_srq->pd;
1341 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1342 	struct bnxt_re_dev *rdev = pd->rdev;
1343 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1344 	struct bnxt_re_srq *srq =
1345 		container_of(ib_srq, struct bnxt_re_srq, ib_srq);
1346 	struct bnxt_qplib_nq *nq = NULL;
1347 	int rc, entries;
1348 
1349 	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
1350 		dev_err(rdev_to_dev(rdev), "Create CQ failed - max exceeded");
1351 		rc = -EINVAL;
1352 		goto exit;
1353 	}
1354 
1355 	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
1356 		rc = -EOPNOTSUPP;
1357 		goto exit;
1358 	}
1359 
1360 	srq->rdev = rdev;
1361 	srq->qplib_srq.pd = &pd->qplib_pd;
1362 	srq->qplib_srq.dpi = &rdev->dpi_privileged;
1363 	/* Allocate 1 more than what's provided so posting max doesn't
1364 	 * mean empty
1365 	 */
1366 	entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
1367 	if (entries > dev_attr->max_srq_wqes + 1)
1368 		entries = dev_attr->max_srq_wqes + 1;
1369 
1370 	srq->qplib_srq.max_wqe = entries;
1371 	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
1372 	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
1373 	srq->srq_limit = srq_init_attr->attr.srq_limit;
1374 	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
1375 	nq = &rdev->nq[0];
1376 
1377 	if (udata) {
1378 		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
1379 		if (rc)
1380 			goto fail;
1381 	}
1382 
1383 	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
1384 	if (rc) {
1385 		dev_err(rdev_to_dev(rdev), "Create HW SRQ failed!");
1386 		goto fail;
1387 	}
1388 
1389 	if (udata) {
1390 		struct bnxt_re_srq_resp resp;
1391 
1392 		resp.srqid = srq->qplib_srq.id;
1393 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1394 		if (rc) {
1395 			dev_err(rdev_to_dev(rdev), "SRQ copy to udata failed!");
1396 			bnxt_qplib_destroy_srq(&rdev->qplib_res,
1397 					       &srq->qplib_srq);
1398 			goto exit;
1399 		}
1400 	}
1401 	if (nq)
1402 		nq->budget++;
1403 	atomic_inc(&rdev->srq_count);
1404 
1405 	return 0;
1406 
1407 fail:
1408 	ib_umem_release(srq->umem);
1409 exit:
1410 	return rc;
1411 }
1412 
1413 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
1414 		       enum ib_srq_attr_mask srq_attr_mask,
1415 		       struct ib_udata *udata)
1416 {
1417 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1418 					       ib_srq);
1419 	struct bnxt_re_dev *rdev = srq->rdev;
1420 	int rc;
1421 
1422 	switch (srq_attr_mask) {
1423 	case IB_SRQ_MAX_WR:
1424 		/* SRQ resize is not supported */
1425 		break;
1426 	case IB_SRQ_LIMIT:
1427 		/* Change the SRQ threshold */
1428 		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
1429 			return -EINVAL;
1430 
1431 		srq->qplib_srq.threshold = srq_attr->srq_limit;
1432 		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
1433 		if (rc) {
1434 			dev_err(rdev_to_dev(rdev), "Modify HW SRQ failed!");
1435 			return rc;
1436 		}
1437 		/* On success, update the shadow */
1438 		srq->srq_limit = srq_attr->srq_limit;
1439 		/* No need to Build and send response back to udata */
1440 		break;
1441 	default:
1442 		dev_err(rdev_to_dev(rdev),
1443 			"Unsupported srq_attr_mask 0x%x", srq_attr_mask);
1444 		return -EINVAL;
1445 	}
1446 	return 0;
1447 }
1448 
1449 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
1450 {
1451 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1452 					       ib_srq);
1453 	struct bnxt_re_srq tsrq;
1454 	struct bnxt_re_dev *rdev = srq->rdev;
1455 	int rc;
1456 
1457 	/* Get live SRQ attr */
1458 	tsrq.qplib_srq.id = srq->qplib_srq.id;
1459 	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
1460 	if (rc) {
1461 		dev_err(rdev_to_dev(rdev), "Query HW SRQ failed!");
1462 		return rc;
1463 	}
1464 	srq_attr->max_wr = srq->qplib_srq.max_wqe;
1465 	srq_attr->max_sge = srq->qplib_srq.max_sge;
1466 	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
1467 
1468 	return 0;
1469 }
1470 
1471 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr,
1472 			  const struct ib_recv_wr **bad_wr)
1473 {
1474 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1475 					       ib_srq);
1476 	struct bnxt_qplib_swqe wqe;
1477 	unsigned long flags;
1478 	int rc = 0;
1479 
1480 	spin_lock_irqsave(&srq->lock, flags);
1481 	while (wr) {
1482 		/* Transcribe each ib_recv_wr to qplib_swqe */
1483 		wqe.num_sge = wr->num_sge;
1484 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
1485 		wqe.wr_id = wr->wr_id;
1486 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
1487 
1488 		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
1489 		if (rc) {
1490 			*bad_wr = wr;
1491 			break;
1492 		}
1493 		wr = wr->next;
1494 	}
1495 	spin_unlock_irqrestore(&srq->lock, flags);
1496 
1497 	return rc;
1498 }
1499 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
1500 				    struct bnxt_re_qp *qp1_qp,
1501 				    int qp_attr_mask)
1502 {
1503 	struct bnxt_re_qp *qp = rdev->qp1_sqp;
1504 	int rc = 0;
1505 
1506 	if (qp_attr_mask & IB_QP_STATE) {
1507 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1508 		qp->qplib_qp.state = qp1_qp->qplib_qp.state;
1509 	}
1510 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1511 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1512 		qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index;
1513 	}
1514 
1515 	if (qp_attr_mask & IB_QP_QKEY) {
1516 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1517 		/* Using a Random  QKEY */
1518 		qp->qplib_qp.qkey = 0x81818181;
1519 	}
1520 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1521 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1522 		qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn;
1523 	}
1524 
1525 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1526 	if (rc)
1527 		dev_err(rdev_to_dev(rdev),
1528 			"Failed to modify Shadow QP for QP1");
1529 	return rc;
1530 }
1531 
1532 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1533 		      int qp_attr_mask, struct ib_udata *udata)
1534 {
1535 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1536 	struct bnxt_re_dev *rdev = qp->rdev;
1537 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1538 	enum ib_qp_state curr_qp_state, new_qp_state;
1539 	int rc, entries;
1540 	unsigned int flags;
1541 	u8 nw_type;
1542 
1543 	qp->qplib_qp.modify_flags = 0;
1544 	if (qp_attr_mask & IB_QP_STATE) {
1545 		curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state);
1546 		new_qp_state = qp_attr->qp_state;
1547 		if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state,
1548 					ib_qp->qp_type, qp_attr_mask)) {
1549 			dev_err(rdev_to_dev(rdev),
1550 				"Invalid attribute mask: %#x specified ",
1551 				qp_attr_mask);
1552 			dev_err(rdev_to_dev(rdev),
1553 				"for qpn: %#x type: %#x",
1554 				ib_qp->qp_num, ib_qp->qp_type);
1555 			dev_err(rdev_to_dev(rdev),
1556 				"curr_qp_state=0x%x, new_qp_state=0x%x\n",
1557 				curr_qp_state, new_qp_state);
1558 			return -EINVAL;
1559 		}
1560 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1561 		qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state);
1562 
1563 		if (!qp->sumem &&
1564 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1565 			dev_dbg(rdev_to_dev(rdev),
1566 				"Move QP = %p to flush list\n",
1567 				qp);
1568 			flags = bnxt_re_lock_cqs(qp);
1569 			bnxt_qplib_add_flush_qp(&qp->qplib_qp);
1570 			bnxt_re_unlock_cqs(qp, flags);
1571 		}
1572 		if (!qp->sumem &&
1573 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) {
1574 			dev_dbg(rdev_to_dev(rdev),
1575 				"Move QP = %p out of flush list\n",
1576 				qp);
1577 			flags = bnxt_re_lock_cqs(qp);
1578 			bnxt_qplib_clean_qp(&qp->qplib_qp);
1579 			bnxt_re_unlock_cqs(qp, flags);
1580 		}
1581 	}
1582 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1583 		qp->qplib_qp.modify_flags |=
1584 				CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY;
1585 		qp->qplib_qp.en_sqd_async_notify = true;
1586 	}
1587 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
1588 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS;
1589 		qp->qplib_qp.access =
1590 			__from_ib_access_flags(qp_attr->qp_access_flags);
1591 		/* LOCAL_WRITE access must be set to allow RC receive */
1592 		qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
1593 		/* Temp: Set all params on QP as of now */
1594 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE;
1595 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ;
1596 	}
1597 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1598 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1599 		qp->qplib_qp.pkey_index = qp_attr->pkey_index;
1600 	}
1601 	if (qp_attr_mask & IB_QP_QKEY) {
1602 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1603 		qp->qplib_qp.qkey = qp_attr->qkey;
1604 	}
1605 	if (qp_attr_mask & IB_QP_AV) {
1606 		const struct ib_global_route *grh =
1607 			rdma_ah_read_grh(&qp_attr->ah_attr);
1608 		const struct ib_gid_attr *sgid_attr;
1609 
1610 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID |
1611 				     CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL |
1612 				     CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX |
1613 				     CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT |
1614 				     CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS |
1615 				     CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC |
1616 				     CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID;
1617 		memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw,
1618 		       sizeof(qp->qplib_qp.ah.dgid.data));
1619 		qp->qplib_qp.ah.flow_label = grh->flow_label;
1620 		/* If RoCE V2 is enabled, stack will have two entries for
1621 		 * each GID entry. Avoiding this duplicte entry in HW. Dividing
1622 		 * the GID index by 2 for RoCE V2
1623 		 */
1624 		qp->qplib_qp.ah.sgid_index = grh->sgid_index / 2;
1625 		qp->qplib_qp.ah.host_sgid_index = grh->sgid_index;
1626 		qp->qplib_qp.ah.hop_limit = grh->hop_limit;
1627 		qp->qplib_qp.ah.traffic_class = grh->traffic_class;
1628 		qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr);
1629 		ether_addr_copy(qp->qplib_qp.ah.dmac,
1630 				qp_attr->ah_attr.roce.dmac);
1631 
1632 		sgid_attr = qp_attr->ah_attr.grh.sgid_attr;
1633 		rc = rdma_read_gid_l2_fields(sgid_attr, NULL,
1634 					     &qp->qplib_qp.smac[0]);
1635 		if (rc)
1636 			return rc;
1637 
1638 		nw_type = rdma_gid_attr_network_type(sgid_attr);
1639 		switch (nw_type) {
1640 		case RDMA_NETWORK_IPV4:
1641 			qp->qplib_qp.nw_type =
1642 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4;
1643 			break;
1644 		case RDMA_NETWORK_IPV6:
1645 			qp->qplib_qp.nw_type =
1646 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6;
1647 			break;
1648 		default:
1649 			qp->qplib_qp.nw_type =
1650 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1;
1651 			break;
1652 		}
1653 	}
1654 
1655 	if (qp_attr_mask & IB_QP_PATH_MTU) {
1656 		qp->qplib_qp.modify_flags |=
1657 				CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1658 		qp->qplib_qp.path_mtu = __from_ib_mtu(qp_attr->path_mtu);
1659 		qp->qplib_qp.mtu = ib_mtu_enum_to_int(qp_attr->path_mtu);
1660 	} else if (qp_attr->qp_state == IB_QPS_RTR) {
1661 		qp->qplib_qp.modify_flags |=
1662 			CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1663 		qp->qplib_qp.path_mtu =
1664 			__from_ib_mtu(iboe_get_mtu(rdev->netdev->mtu));
1665 		qp->qplib_qp.mtu =
1666 			ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1667 	}
1668 
1669 	if (qp_attr_mask & IB_QP_TIMEOUT) {
1670 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT;
1671 		qp->qplib_qp.timeout = qp_attr->timeout;
1672 	}
1673 	if (qp_attr_mask & IB_QP_RETRY_CNT) {
1674 		qp->qplib_qp.modify_flags |=
1675 				CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT;
1676 		qp->qplib_qp.retry_cnt = qp_attr->retry_cnt;
1677 	}
1678 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
1679 		qp->qplib_qp.modify_flags |=
1680 				CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY;
1681 		qp->qplib_qp.rnr_retry = qp_attr->rnr_retry;
1682 	}
1683 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) {
1684 		qp->qplib_qp.modify_flags |=
1685 				CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER;
1686 		qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer;
1687 	}
1688 	if (qp_attr_mask & IB_QP_RQ_PSN) {
1689 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN;
1690 		qp->qplib_qp.rq.psn = qp_attr->rq_psn;
1691 	}
1692 	if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1693 		qp->qplib_qp.modify_flags |=
1694 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC;
1695 		/* Cap the max_rd_atomic to device max */
1696 		qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic,
1697 						   dev_attr->max_qp_rd_atom);
1698 	}
1699 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1700 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1701 		qp->qplib_qp.sq.psn = qp_attr->sq_psn;
1702 	}
1703 	if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1704 		if (qp_attr->max_dest_rd_atomic >
1705 		    dev_attr->max_qp_init_rd_atom) {
1706 			dev_err(rdev_to_dev(rdev),
1707 				"max_dest_rd_atomic requested%d is > dev_max%d",
1708 				qp_attr->max_dest_rd_atomic,
1709 				dev_attr->max_qp_init_rd_atom);
1710 			return -EINVAL;
1711 		}
1712 
1713 		qp->qplib_qp.modify_flags |=
1714 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC;
1715 		qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic;
1716 	}
1717 	if (qp_attr_mask & IB_QP_CAP) {
1718 		qp->qplib_qp.modify_flags |=
1719 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE |
1720 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE |
1721 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE |
1722 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE |
1723 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA;
1724 		if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) ||
1725 		    (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) ||
1726 		    (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) ||
1727 		    (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) ||
1728 		    (qp_attr->cap.max_inline_data >=
1729 						dev_attr->max_inline_data)) {
1730 			dev_err(rdev_to_dev(rdev),
1731 				"Create QP failed - max exceeded");
1732 			return -EINVAL;
1733 		}
1734 		entries = roundup_pow_of_two(qp_attr->cap.max_send_wr);
1735 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
1736 						dev_attr->max_qp_wqes + 1);
1737 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
1738 						qp_attr->cap.max_send_wr;
1739 		/*
1740 		 * Reserving one slot for Phantom WQE. Some application can
1741 		 * post one extra entry in this case. Allowing this to avoid
1742 		 * unexpected Queue full condition
1743 		 */
1744 		qp->qplib_qp.sq.q_full_delta -= 1;
1745 		qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
1746 		if (qp->qplib_qp.rq.max_wqe) {
1747 			entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr);
1748 			qp->qplib_qp.rq.max_wqe =
1749 				min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1750 			qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
1751 						       qp_attr->cap.max_recv_wr;
1752 			qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
1753 		} else {
1754 			/* SRQ was used prior, just ignore the RQ caps */
1755 		}
1756 	}
1757 	if (qp_attr_mask & IB_QP_DEST_QPN) {
1758 		qp->qplib_qp.modify_flags |=
1759 				CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID;
1760 		qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num;
1761 	}
1762 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1763 	if (rc) {
1764 		dev_err(rdev_to_dev(rdev), "Failed to modify HW QP");
1765 		return rc;
1766 	}
1767 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->qp1_sqp)
1768 		rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask);
1769 	return rc;
1770 }
1771 
1772 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1773 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
1774 {
1775 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1776 	struct bnxt_re_dev *rdev = qp->rdev;
1777 	struct bnxt_qplib_qp *qplib_qp;
1778 	int rc;
1779 
1780 	qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL);
1781 	if (!qplib_qp)
1782 		return -ENOMEM;
1783 
1784 	qplib_qp->id = qp->qplib_qp.id;
1785 	qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index;
1786 
1787 	rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp);
1788 	if (rc) {
1789 		dev_err(rdev_to_dev(rdev), "Failed to query HW QP");
1790 		goto out;
1791 	}
1792 	qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state);
1793 	qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0;
1794 	qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access);
1795 	qp_attr->pkey_index = qplib_qp->pkey_index;
1796 	qp_attr->qkey = qplib_qp->qkey;
1797 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1798 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label,
1799 			qplib_qp->ah.host_sgid_index,
1800 			qplib_qp->ah.hop_limit,
1801 			qplib_qp->ah.traffic_class);
1802 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data);
1803 	rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl);
1804 	ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac);
1805 	qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu);
1806 	qp_attr->timeout = qplib_qp->timeout;
1807 	qp_attr->retry_cnt = qplib_qp->retry_cnt;
1808 	qp_attr->rnr_retry = qplib_qp->rnr_retry;
1809 	qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer;
1810 	qp_attr->rq_psn = qplib_qp->rq.psn;
1811 	qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic;
1812 	qp_attr->sq_psn = qplib_qp->sq.psn;
1813 	qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic;
1814 	qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR :
1815 							 IB_SIGNAL_REQ_WR;
1816 	qp_attr->dest_qp_num = qplib_qp->dest_qpn;
1817 
1818 	qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe;
1819 	qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge;
1820 	qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe;
1821 	qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge;
1822 	qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data;
1823 	qp_init_attr->cap = qp_attr->cap;
1824 
1825 out:
1826 	kfree(qplib_qp);
1827 	return rc;
1828 }
1829 
1830 /* Routine for sending QP1 packets for RoCE V1 an V2
1831  */
1832 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp,
1833 				     const struct ib_send_wr *wr,
1834 				     struct bnxt_qplib_swqe *wqe,
1835 				     int payload_size)
1836 {
1837 	struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah,
1838 					     ib_ah);
1839 	struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah;
1840 	const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr;
1841 	struct bnxt_qplib_sge sge;
1842 	u8 nw_type;
1843 	u16 ether_type;
1844 	union ib_gid dgid;
1845 	bool is_eth = false;
1846 	bool is_vlan = false;
1847 	bool is_grh = false;
1848 	bool is_udp = false;
1849 	u8 ip_version = 0;
1850 	u16 vlan_id = 0xFFFF;
1851 	void *buf;
1852 	int i, rc = 0;
1853 
1854 	memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr));
1855 
1856 	rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL);
1857 	if (rc)
1858 		return rc;
1859 
1860 	/* Get network header type for this GID */
1861 	nw_type = rdma_gid_attr_network_type(sgid_attr);
1862 	switch (nw_type) {
1863 	case RDMA_NETWORK_IPV4:
1864 		nw_type = BNXT_RE_ROCEV2_IPV4_PACKET;
1865 		break;
1866 	case RDMA_NETWORK_IPV6:
1867 		nw_type = BNXT_RE_ROCEV2_IPV6_PACKET;
1868 		break;
1869 	default:
1870 		nw_type = BNXT_RE_ROCE_V1_PACKET;
1871 		break;
1872 	}
1873 	memcpy(&dgid.raw, &qplib_ah->dgid, 16);
1874 	is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
1875 	if (is_udp) {
1876 		if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) {
1877 			ip_version = 4;
1878 			ether_type = ETH_P_IP;
1879 		} else {
1880 			ip_version = 6;
1881 			ether_type = ETH_P_IPV6;
1882 		}
1883 		is_grh = false;
1884 	} else {
1885 		ether_type = ETH_P_IBOE;
1886 		is_grh = true;
1887 	}
1888 
1889 	is_eth = true;
1890 	is_vlan = (vlan_id && (vlan_id < 0x1000)) ? true : false;
1891 
1892 	ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh,
1893 			  ip_version, is_udp, 0, &qp->qp1_hdr);
1894 
1895 	/* ETH */
1896 	ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac);
1897 	ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac);
1898 
1899 	/* For vlan, check the sgid for vlan existence */
1900 
1901 	if (!is_vlan) {
1902 		qp->qp1_hdr.eth.type = cpu_to_be16(ether_type);
1903 	} else {
1904 		qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type);
1905 		qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id);
1906 	}
1907 
1908 	if (is_grh || (ip_version == 6)) {
1909 		memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw,
1910 		       sizeof(sgid_attr->gid));
1911 		memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data,
1912 		       sizeof(sgid_attr->gid));
1913 		qp->qp1_hdr.grh.hop_limit     = qplib_ah->hop_limit;
1914 	}
1915 
1916 	if (ip_version == 4) {
1917 		qp->qp1_hdr.ip4.tos = 0;
1918 		qp->qp1_hdr.ip4.id = 0;
1919 		qp->qp1_hdr.ip4.frag_off = htons(IP_DF);
1920 		qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit;
1921 
1922 		memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4);
1923 		memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4);
1924 		qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr);
1925 	}
1926 
1927 	if (is_udp) {
1928 		qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT);
1929 		qp->qp1_hdr.udp.sport = htons(0x8CD1);
1930 		qp->qp1_hdr.udp.csum = 0;
1931 	}
1932 
1933 	/* BTH */
1934 	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
1935 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1936 		qp->qp1_hdr.immediate_present = 1;
1937 	} else {
1938 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1939 	}
1940 	if (wr->send_flags & IB_SEND_SOLICITED)
1941 		qp->qp1_hdr.bth.solicited_event = 1;
1942 	/* pad_count */
1943 	qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3;
1944 
1945 	/* P_key for QP1 is for all members */
1946 	qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF);
1947 	qp->qp1_hdr.bth.destination_qpn = IB_QP1;
1948 	qp->qp1_hdr.bth.ack_req = 0;
1949 	qp->send_psn++;
1950 	qp->send_psn &= BTH_PSN_MASK;
1951 	qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn);
1952 	/* DETH */
1953 	/* Use the priviledged Q_Key for QP1 */
1954 	qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY);
1955 	qp->qp1_hdr.deth.source_qpn = IB_QP1;
1956 
1957 	/* Pack the QP1 to the transmit buffer */
1958 	buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge);
1959 	if (buf) {
1960 		ib_ud_header_pack(&qp->qp1_hdr, buf);
1961 		for (i = wqe->num_sge; i; i--) {
1962 			wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr;
1963 			wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey;
1964 			wqe->sg_list[i].size = wqe->sg_list[i - 1].size;
1965 		}
1966 
1967 		/*
1968 		 * Max Header buf size for IPV6 RoCE V2 is 86,
1969 		 * which is same as the QP1 SQ header buffer.
1970 		 * Header buf size for IPV4 RoCE V2 can be 66.
1971 		 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
1972 		 * Subtract 20 bytes from QP1 SQ header buf size
1973 		 */
1974 		if (is_udp && ip_version == 4)
1975 			sge.size -= 20;
1976 		/*
1977 		 * Max Header buf size for RoCE V1 is 78.
1978 		 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
1979 		 * Subtract 8 bytes from QP1 SQ header buf size
1980 		 */
1981 		if (!is_udp)
1982 			sge.size -= 8;
1983 
1984 		/* Subtract 4 bytes for non vlan packets */
1985 		if (!is_vlan)
1986 			sge.size -= 4;
1987 
1988 		wqe->sg_list[0].addr = sge.addr;
1989 		wqe->sg_list[0].lkey = sge.lkey;
1990 		wqe->sg_list[0].size = sge.size;
1991 		wqe->num_sge++;
1992 
1993 	} else {
1994 		dev_err(rdev_to_dev(qp->rdev), "QP1 buffer is empty!");
1995 		rc = -ENOMEM;
1996 	}
1997 	return rc;
1998 }
1999 
2000 /* For the MAD layer, it only provides the recv SGE the size of
2001  * ib_grh + MAD datagram.  No Ethernet headers, Ethertype, BTH, DETH,
2002  * nor RoCE iCRC.  The Cu+ solution must provide buffer for the entire
2003  * receive packet (334 bytes) with no VLAN and then copy the GRH
2004  * and the MAD datagram out to the provided SGE.
2005  */
2006 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp,
2007 					    const struct ib_recv_wr *wr,
2008 					    struct bnxt_qplib_swqe *wqe,
2009 					    int payload_size)
2010 {
2011 	struct bnxt_qplib_sge ref, sge;
2012 	u32 rq_prod_index;
2013 	struct bnxt_re_sqp_entries *sqp_entry;
2014 
2015 	rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp);
2016 
2017 	if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge))
2018 		return -ENOMEM;
2019 
2020 	/* Create 1 SGE to receive the entire
2021 	 * ethernet packet
2022 	 */
2023 	/* Save the reference from ULP */
2024 	ref.addr = wqe->sg_list[0].addr;
2025 	ref.lkey = wqe->sg_list[0].lkey;
2026 	ref.size = wqe->sg_list[0].size;
2027 
2028 	sqp_entry = &qp->rdev->sqp_tbl[rq_prod_index];
2029 
2030 	/* SGE 1 */
2031 	wqe->sg_list[0].addr = sge.addr;
2032 	wqe->sg_list[0].lkey = sge.lkey;
2033 	wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
2034 	sge.size -= wqe->sg_list[0].size;
2035 
2036 	sqp_entry->sge.addr = ref.addr;
2037 	sqp_entry->sge.lkey = ref.lkey;
2038 	sqp_entry->sge.size = ref.size;
2039 	/* Store the wrid for reporting completion */
2040 	sqp_entry->wrid = wqe->wr_id;
2041 	/* change the wqe->wrid to table index */
2042 	wqe->wr_id = rq_prod_index;
2043 	return 0;
2044 }
2045 
2046 static int is_ud_qp(struct bnxt_re_qp *qp)
2047 {
2048 	return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD ||
2049 		qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI);
2050 }
2051 
2052 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp,
2053 				  const struct ib_send_wr *wr,
2054 				  struct bnxt_qplib_swqe *wqe)
2055 {
2056 	struct bnxt_re_ah *ah = NULL;
2057 
2058 	if (is_ud_qp(qp)) {
2059 		ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah);
2060 		wqe->send.q_key = ud_wr(wr)->remote_qkey;
2061 		wqe->send.dst_qp = ud_wr(wr)->remote_qpn;
2062 		wqe->send.avid = ah->qplib_ah.id;
2063 	}
2064 	switch (wr->opcode) {
2065 	case IB_WR_SEND:
2066 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND;
2067 		break;
2068 	case IB_WR_SEND_WITH_IMM:
2069 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM;
2070 		wqe->send.imm_data = wr->ex.imm_data;
2071 		break;
2072 	case IB_WR_SEND_WITH_INV:
2073 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV;
2074 		wqe->send.inv_key = wr->ex.invalidate_rkey;
2075 		break;
2076 	default:
2077 		return -EINVAL;
2078 	}
2079 	if (wr->send_flags & IB_SEND_SIGNALED)
2080 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2081 	if (wr->send_flags & IB_SEND_FENCE)
2082 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2083 	if (wr->send_flags & IB_SEND_SOLICITED)
2084 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2085 	if (wr->send_flags & IB_SEND_INLINE)
2086 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2087 
2088 	return 0;
2089 }
2090 
2091 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr,
2092 				  struct bnxt_qplib_swqe *wqe)
2093 {
2094 	switch (wr->opcode) {
2095 	case IB_WR_RDMA_WRITE:
2096 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE;
2097 		break;
2098 	case IB_WR_RDMA_WRITE_WITH_IMM:
2099 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM;
2100 		wqe->rdma.imm_data = wr->ex.imm_data;
2101 		break;
2102 	case IB_WR_RDMA_READ:
2103 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ;
2104 		wqe->rdma.inv_key = wr->ex.invalidate_rkey;
2105 		break;
2106 	default:
2107 		return -EINVAL;
2108 	}
2109 	wqe->rdma.remote_va = rdma_wr(wr)->remote_addr;
2110 	wqe->rdma.r_key = rdma_wr(wr)->rkey;
2111 	if (wr->send_flags & IB_SEND_SIGNALED)
2112 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2113 	if (wr->send_flags & IB_SEND_FENCE)
2114 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2115 	if (wr->send_flags & IB_SEND_SOLICITED)
2116 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2117 	if (wr->send_flags & IB_SEND_INLINE)
2118 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2119 
2120 	return 0;
2121 }
2122 
2123 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr,
2124 				    struct bnxt_qplib_swqe *wqe)
2125 {
2126 	switch (wr->opcode) {
2127 	case IB_WR_ATOMIC_CMP_AND_SWP:
2128 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP;
2129 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2130 		wqe->atomic.swap_data = atomic_wr(wr)->swap;
2131 		break;
2132 	case IB_WR_ATOMIC_FETCH_AND_ADD:
2133 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD;
2134 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2135 		break;
2136 	default:
2137 		return -EINVAL;
2138 	}
2139 	wqe->atomic.remote_va = atomic_wr(wr)->remote_addr;
2140 	wqe->atomic.r_key = atomic_wr(wr)->rkey;
2141 	if (wr->send_flags & IB_SEND_SIGNALED)
2142 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2143 	if (wr->send_flags & IB_SEND_FENCE)
2144 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2145 	if (wr->send_flags & IB_SEND_SOLICITED)
2146 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2147 	return 0;
2148 }
2149 
2150 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr,
2151 				 struct bnxt_qplib_swqe *wqe)
2152 {
2153 	wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV;
2154 	wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey;
2155 
2156 	/* Need unconditional fence for local invalidate
2157 	 * opcode to work as expected.
2158 	 */
2159 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2160 
2161 	if (wr->send_flags & IB_SEND_SIGNALED)
2162 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2163 	if (wr->send_flags & IB_SEND_SOLICITED)
2164 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2165 
2166 	return 0;
2167 }
2168 
2169 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr,
2170 				 struct bnxt_qplib_swqe *wqe)
2171 {
2172 	struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr);
2173 	struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl;
2174 	int access = wr->access;
2175 
2176 	wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0];
2177 	wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0];
2178 	wqe->frmr.page_list = mr->pages;
2179 	wqe->frmr.page_list_len = mr->npages;
2180 	wqe->frmr.levels = qplib_frpl->hwq.level + 1;
2181 	wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR;
2182 
2183 	/* Need unconditional fence for reg_mr
2184 	 * opcode to function as expected.
2185 	 */
2186 
2187 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2188 
2189 	if (wr->wr.send_flags & IB_SEND_SIGNALED)
2190 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2191 
2192 	if (access & IB_ACCESS_LOCAL_WRITE)
2193 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE;
2194 	if (access & IB_ACCESS_REMOTE_READ)
2195 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ;
2196 	if (access & IB_ACCESS_REMOTE_WRITE)
2197 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE;
2198 	if (access & IB_ACCESS_REMOTE_ATOMIC)
2199 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC;
2200 	if (access & IB_ACCESS_MW_BIND)
2201 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND;
2202 
2203 	wqe->frmr.l_key = wr->key;
2204 	wqe->frmr.length = wr->mr->length;
2205 	wqe->frmr.pbl_pg_sz_log = (wr->mr->page_size >> PAGE_SHIFT_4K) - 1;
2206 	wqe->frmr.va = wr->mr->iova;
2207 	return 0;
2208 }
2209 
2210 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev,
2211 				    const struct ib_send_wr *wr,
2212 				    struct bnxt_qplib_swqe *wqe)
2213 {
2214 	/*  Copy the inline data to the data  field */
2215 	u8 *in_data;
2216 	u32 i, sge_len;
2217 	void *sge_addr;
2218 
2219 	in_data = wqe->inline_data;
2220 	for (i = 0; i < wr->num_sge; i++) {
2221 		sge_addr = (void *)(unsigned long)
2222 				wr->sg_list[i].addr;
2223 		sge_len = wr->sg_list[i].length;
2224 
2225 		if ((sge_len + wqe->inline_len) >
2226 		    BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) {
2227 			dev_err(rdev_to_dev(rdev),
2228 				"Inline data size requested > supported value");
2229 			return -EINVAL;
2230 		}
2231 		sge_len = wr->sg_list[i].length;
2232 
2233 		memcpy(in_data, sge_addr, sge_len);
2234 		in_data += wr->sg_list[i].length;
2235 		wqe->inline_len += wr->sg_list[i].length;
2236 	}
2237 	return wqe->inline_len;
2238 }
2239 
2240 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
2241 				   const struct ib_send_wr *wr,
2242 				   struct bnxt_qplib_swqe *wqe)
2243 {
2244 	int payload_sz = 0;
2245 
2246 	if (wr->send_flags & IB_SEND_INLINE)
2247 		payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe);
2248 	else
2249 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list,
2250 					       wqe->num_sge);
2251 
2252 	return payload_sz;
2253 }
2254 
2255 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
2256 {
2257 	if ((qp->ib_qp.qp_type == IB_QPT_UD ||
2258 	     qp->ib_qp.qp_type == IB_QPT_GSI ||
2259 	     qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
2260 	     qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
2261 		int qp_attr_mask;
2262 		struct ib_qp_attr qp_attr;
2263 
2264 		qp_attr_mask = IB_QP_STATE;
2265 		qp_attr.qp_state = IB_QPS_RTS;
2266 		bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
2267 		qp->qplib_qp.wqe_cnt = 0;
2268 	}
2269 }
2270 
2271 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
2272 				       struct bnxt_re_qp *qp,
2273 				       const struct ib_send_wr *wr)
2274 {
2275 	struct bnxt_qplib_swqe wqe;
2276 	int rc = 0, payload_sz = 0;
2277 	unsigned long flags;
2278 
2279 	spin_lock_irqsave(&qp->sq_lock, flags);
2280 	memset(&wqe, 0, sizeof(wqe));
2281 	while (wr) {
2282 		/* House keeping */
2283 		memset(&wqe, 0, sizeof(wqe));
2284 
2285 		/* Common */
2286 		wqe.num_sge = wr->num_sge;
2287 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2288 			dev_err(rdev_to_dev(rdev),
2289 				"Limit exceeded for Send SGEs");
2290 			rc = -EINVAL;
2291 			goto bad;
2292 		}
2293 
2294 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2295 		if (payload_sz < 0) {
2296 			rc = -EINVAL;
2297 			goto bad;
2298 		}
2299 		wqe.wr_id = wr->wr_id;
2300 
2301 		wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND;
2302 
2303 		rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2304 		if (!rc)
2305 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2306 bad:
2307 		if (rc) {
2308 			dev_err(rdev_to_dev(rdev),
2309 				"Post send failed opcode = %#x rc = %d",
2310 				wr->opcode, rc);
2311 			break;
2312 		}
2313 		wr = wr->next;
2314 	}
2315 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2316 	bnxt_ud_qp_hw_stall_workaround(qp);
2317 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2318 	return rc;
2319 }
2320 
2321 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr,
2322 		      const struct ib_send_wr **bad_wr)
2323 {
2324 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2325 	struct bnxt_qplib_swqe wqe;
2326 	int rc = 0, payload_sz = 0;
2327 	unsigned long flags;
2328 
2329 	spin_lock_irqsave(&qp->sq_lock, flags);
2330 	while (wr) {
2331 		/* House keeping */
2332 		memset(&wqe, 0, sizeof(wqe));
2333 
2334 		/* Common */
2335 		wqe.num_sge = wr->num_sge;
2336 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2337 			dev_err(rdev_to_dev(qp->rdev),
2338 				"Limit exceeded for Send SGEs");
2339 			rc = -EINVAL;
2340 			goto bad;
2341 		}
2342 
2343 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2344 		if (payload_sz < 0) {
2345 			rc = -EINVAL;
2346 			goto bad;
2347 		}
2348 		wqe.wr_id = wr->wr_id;
2349 
2350 		switch (wr->opcode) {
2351 		case IB_WR_SEND:
2352 		case IB_WR_SEND_WITH_IMM:
2353 			if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) {
2354 				rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe,
2355 							       payload_sz);
2356 				if (rc)
2357 					goto bad;
2358 				wqe.rawqp1.lflags |=
2359 					SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC;
2360 			}
2361 			switch (wr->send_flags) {
2362 			case IB_SEND_IP_CSUM:
2363 				wqe.rawqp1.lflags |=
2364 					SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM;
2365 				break;
2366 			default:
2367 				break;
2368 			}
2369 			/* fall through */
2370 		case IB_WR_SEND_WITH_INV:
2371 			rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2372 			break;
2373 		case IB_WR_RDMA_WRITE:
2374 		case IB_WR_RDMA_WRITE_WITH_IMM:
2375 		case IB_WR_RDMA_READ:
2376 			rc = bnxt_re_build_rdma_wqe(wr, &wqe);
2377 			break;
2378 		case IB_WR_ATOMIC_CMP_AND_SWP:
2379 		case IB_WR_ATOMIC_FETCH_AND_ADD:
2380 			rc = bnxt_re_build_atomic_wqe(wr, &wqe);
2381 			break;
2382 		case IB_WR_RDMA_READ_WITH_INV:
2383 			dev_err(rdev_to_dev(qp->rdev),
2384 				"RDMA Read with Invalidate is not supported");
2385 			rc = -EINVAL;
2386 			goto bad;
2387 		case IB_WR_LOCAL_INV:
2388 			rc = bnxt_re_build_inv_wqe(wr, &wqe);
2389 			break;
2390 		case IB_WR_REG_MR:
2391 			rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe);
2392 			break;
2393 		default:
2394 			/* Unsupported WRs */
2395 			dev_err(rdev_to_dev(qp->rdev),
2396 				"WR (%#x) is not supported", wr->opcode);
2397 			rc = -EINVAL;
2398 			goto bad;
2399 		}
2400 		if (!rc)
2401 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2402 bad:
2403 		if (rc) {
2404 			dev_err(rdev_to_dev(qp->rdev),
2405 				"post_send failed op:%#x qps = %#x rc = %d\n",
2406 				wr->opcode, qp->qplib_qp.state, rc);
2407 			*bad_wr = wr;
2408 			break;
2409 		}
2410 		wr = wr->next;
2411 	}
2412 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2413 	bnxt_ud_qp_hw_stall_workaround(qp);
2414 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2415 
2416 	return rc;
2417 }
2418 
2419 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev,
2420 				       struct bnxt_re_qp *qp,
2421 				       const struct ib_recv_wr *wr)
2422 {
2423 	struct bnxt_qplib_swqe wqe;
2424 	int rc = 0;
2425 
2426 	memset(&wqe, 0, sizeof(wqe));
2427 	while (wr) {
2428 		/* House keeping */
2429 		memset(&wqe, 0, sizeof(wqe));
2430 
2431 		/* Common */
2432 		wqe.num_sge = wr->num_sge;
2433 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2434 			dev_err(rdev_to_dev(rdev),
2435 				"Limit exceeded for Receive SGEs");
2436 			rc = -EINVAL;
2437 			break;
2438 		}
2439 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
2440 		wqe.wr_id = wr->wr_id;
2441 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2442 
2443 		rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2444 		if (rc)
2445 			break;
2446 
2447 		wr = wr->next;
2448 	}
2449 	if (!rc)
2450 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2451 	return rc;
2452 }
2453 
2454 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
2455 		      const struct ib_recv_wr **bad_wr)
2456 {
2457 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2458 	struct bnxt_qplib_swqe wqe;
2459 	int rc = 0, payload_sz = 0;
2460 	unsigned long flags;
2461 	u32 count = 0;
2462 
2463 	spin_lock_irqsave(&qp->rq_lock, flags);
2464 	while (wr) {
2465 		/* House keeping */
2466 		memset(&wqe, 0, sizeof(wqe));
2467 
2468 		/* Common */
2469 		wqe.num_sge = wr->num_sge;
2470 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2471 			dev_err(rdev_to_dev(qp->rdev),
2472 				"Limit exceeded for Receive SGEs");
2473 			rc = -EINVAL;
2474 			*bad_wr = wr;
2475 			break;
2476 		}
2477 
2478 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
2479 					       wr->num_sge);
2480 		wqe.wr_id = wr->wr_id;
2481 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2482 
2483 		if (ib_qp->qp_type == IB_QPT_GSI &&
2484 		    qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI)
2485 			rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe,
2486 							      payload_sz);
2487 		if (!rc)
2488 			rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2489 		if (rc) {
2490 			*bad_wr = wr;
2491 			break;
2492 		}
2493 
2494 		/* Ring DB if the RQEs posted reaches a threshold value */
2495 		if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
2496 			bnxt_qplib_post_recv_db(&qp->qplib_qp);
2497 			count = 0;
2498 		}
2499 
2500 		wr = wr->next;
2501 	}
2502 
2503 	if (count)
2504 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2505 
2506 	spin_unlock_irqrestore(&qp->rq_lock, flags);
2507 
2508 	return rc;
2509 }
2510 
2511 /* Completion Queues */
2512 void bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
2513 {
2514 	struct bnxt_re_cq *cq;
2515 	struct bnxt_qplib_nq *nq;
2516 	struct bnxt_re_dev *rdev;
2517 
2518 	cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
2519 	rdev = cq->rdev;
2520 	nq = cq->qplib_cq.nq;
2521 
2522 	bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2523 	ib_umem_release(cq->umem);
2524 
2525 	atomic_dec(&rdev->cq_count);
2526 	nq->budget--;
2527 	kfree(cq->cql);
2528 }
2529 
2530 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
2531 		      struct ib_udata *udata)
2532 {
2533 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev);
2534 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
2535 	struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq);
2536 	int rc, entries;
2537 	int cqe = attr->cqe;
2538 	struct bnxt_qplib_nq *nq = NULL;
2539 	unsigned int nq_alloc_cnt;
2540 
2541 	/* Validate CQ fields */
2542 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
2543 		dev_err(rdev_to_dev(rdev), "Failed to create CQ -max exceeded");
2544 		return -EINVAL;
2545 	}
2546 
2547 	cq->rdev = rdev;
2548 	cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
2549 
2550 	entries = roundup_pow_of_two(cqe + 1);
2551 	if (entries > dev_attr->max_cq_wqes + 1)
2552 		entries = dev_attr->max_cq_wqes + 1;
2553 
2554 	if (udata) {
2555 		struct bnxt_re_cq_req req;
2556 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
2557 			udata, struct bnxt_re_ucontext, ib_uctx);
2558 		if (ib_copy_from_udata(&req, udata, sizeof(req))) {
2559 			rc = -EFAULT;
2560 			goto fail;
2561 		}
2562 
2563 		cq->umem = ib_umem_get(udata, req.cq_va,
2564 				       entries * sizeof(struct cq_base),
2565 				       IB_ACCESS_LOCAL_WRITE, 1);
2566 		if (IS_ERR(cq->umem)) {
2567 			rc = PTR_ERR(cq->umem);
2568 			goto fail;
2569 		}
2570 		cq->qplib_cq.sg_info.sglist = cq->umem->sg_head.sgl;
2571 		cq->qplib_cq.sg_info.npages = ib_umem_num_pages(cq->umem);
2572 		cq->qplib_cq.sg_info.nmap = cq->umem->nmap;
2573 		cq->qplib_cq.dpi = &uctx->dpi;
2574 	} else {
2575 		cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL);
2576 		cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe),
2577 				  GFP_KERNEL);
2578 		if (!cq->cql) {
2579 			rc = -ENOMEM;
2580 			goto fail;
2581 		}
2582 
2583 		cq->qplib_cq.dpi = &rdev->dpi_privileged;
2584 	}
2585 	/*
2586 	 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
2587 	 * used for getting the NQ index.
2588 	 */
2589 	nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
2590 	nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
2591 	cq->qplib_cq.max_wqe = entries;
2592 	cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
2593 	cq->qplib_cq.nq	= nq;
2594 
2595 	rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
2596 	if (rc) {
2597 		dev_err(rdev_to_dev(rdev), "Failed to create HW CQ");
2598 		goto fail;
2599 	}
2600 
2601 	cq->ib_cq.cqe = entries;
2602 	cq->cq_period = cq->qplib_cq.period;
2603 	nq->budget++;
2604 
2605 	atomic_inc(&rdev->cq_count);
2606 	spin_lock_init(&cq->cq_lock);
2607 
2608 	if (udata) {
2609 		struct bnxt_re_cq_resp resp;
2610 
2611 		resp.cqid = cq->qplib_cq.id;
2612 		resp.tail = cq->qplib_cq.hwq.cons;
2613 		resp.phase = cq->qplib_cq.period;
2614 		resp.rsvd = 0;
2615 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
2616 		if (rc) {
2617 			dev_err(rdev_to_dev(rdev), "Failed to copy CQ udata");
2618 			bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2619 			goto c2fail;
2620 		}
2621 	}
2622 
2623 	return 0;
2624 
2625 c2fail:
2626 	ib_umem_release(cq->umem);
2627 fail:
2628 	kfree(cq->cql);
2629 	return rc;
2630 }
2631 
2632 static u8 __req_to_ib_wc_status(u8 qstatus)
2633 {
2634 	switch (qstatus) {
2635 	case CQ_REQ_STATUS_OK:
2636 		return IB_WC_SUCCESS;
2637 	case CQ_REQ_STATUS_BAD_RESPONSE_ERR:
2638 		return IB_WC_BAD_RESP_ERR;
2639 	case CQ_REQ_STATUS_LOCAL_LENGTH_ERR:
2640 		return IB_WC_LOC_LEN_ERR;
2641 	case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR:
2642 		return IB_WC_LOC_QP_OP_ERR;
2643 	case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR:
2644 		return IB_WC_LOC_PROT_ERR;
2645 	case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR:
2646 		return IB_WC_GENERAL_ERR;
2647 	case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR:
2648 		return IB_WC_REM_INV_REQ_ERR;
2649 	case CQ_REQ_STATUS_REMOTE_ACCESS_ERR:
2650 		return IB_WC_REM_ACCESS_ERR;
2651 	case CQ_REQ_STATUS_REMOTE_OPERATION_ERR:
2652 		return IB_WC_REM_OP_ERR;
2653 	case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR:
2654 		return IB_WC_RNR_RETRY_EXC_ERR;
2655 	case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR:
2656 		return IB_WC_RETRY_EXC_ERR;
2657 	case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR:
2658 		return IB_WC_WR_FLUSH_ERR;
2659 	default:
2660 		return IB_WC_GENERAL_ERR;
2661 	}
2662 	return 0;
2663 }
2664 
2665 static u8 __rawqp1_to_ib_wc_status(u8 qstatus)
2666 {
2667 	switch (qstatus) {
2668 	case CQ_RES_RAWETH_QP1_STATUS_OK:
2669 		return IB_WC_SUCCESS;
2670 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR:
2671 		return IB_WC_LOC_ACCESS_ERR;
2672 	case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR:
2673 		return IB_WC_LOC_LEN_ERR;
2674 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR:
2675 		return IB_WC_LOC_PROT_ERR;
2676 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR:
2677 		return IB_WC_LOC_QP_OP_ERR;
2678 	case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR:
2679 		return IB_WC_GENERAL_ERR;
2680 	case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR:
2681 		return IB_WC_WR_FLUSH_ERR;
2682 	case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR:
2683 		return IB_WC_WR_FLUSH_ERR;
2684 	default:
2685 		return IB_WC_GENERAL_ERR;
2686 	}
2687 }
2688 
2689 static u8 __rc_to_ib_wc_status(u8 qstatus)
2690 {
2691 	switch (qstatus) {
2692 	case CQ_RES_RC_STATUS_OK:
2693 		return IB_WC_SUCCESS;
2694 	case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR:
2695 		return IB_WC_LOC_ACCESS_ERR;
2696 	case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR:
2697 		return IB_WC_LOC_LEN_ERR;
2698 	case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR:
2699 		return IB_WC_LOC_PROT_ERR;
2700 	case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR:
2701 		return IB_WC_LOC_QP_OP_ERR;
2702 	case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR:
2703 		return IB_WC_GENERAL_ERR;
2704 	case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR:
2705 		return IB_WC_REM_INV_REQ_ERR;
2706 	case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR:
2707 		return IB_WC_WR_FLUSH_ERR;
2708 	case CQ_RES_RC_STATUS_HW_FLUSH_ERR:
2709 		return IB_WC_WR_FLUSH_ERR;
2710 	default:
2711 		return IB_WC_GENERAL_ERR;
2712 	}
2713 }
2714 
2715 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe)
2716 {
2717 	switch (cqe->type) {
2718 	case BNXT_QPLIB_SWQE_TYPE_SEND:
2719 		wc->opcode = IB_WC_SEND;
2720 		break;
2721 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM:
2722 		wc->opcode = IB_WC_SEND;
2723 		wc->wc_flags |= IB_WC_WITH_IMM;
2724 		break;
2725 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV:
2726 		wc->opcode = IB_WC_SEND;
2727 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
2728 		break;
2729 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE:
2730 		wc->opcode = IB_WC_RDMA_WRITE;
2731 		break;
2732 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM:
2733 		wc->opcode = IB_WC_RDMA_WRITE;
2734 		wc->wc_flags |= IB_WC_WITH_IMM;
2735 		break;
2736 	case BNXT_QPLIB_SWQE_TYPE_RDMA_READ:
2737 		wc->opcode = IB_WC_RDMA_READ;
2738 		break;
2739 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP:
2740 		wc->opcode = IB_WC_COMP_SWAP;
2741 		break;
2742 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD:
2743 		wc->opcode = IB_WC_FETCH_ADD;
2744 		break;
2745 	case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV:
2746 		wc->opcode = IB_WC_LOCAL_INV;
2747 		break;
2748 	case BNXT_QPLIB_SWQE_TYPE_REG_MR:
2749 		wc->opcode = IB_WC_REG_MR;
2750 		break;
2751 	default:
2752 		wc->opcode = IB_WC_SEND;
2753 		break;
2754 	}
2755 
2756 	wc->status = __req_to_ib_wc_status(cqe->status);
2757 }
2758 
2759 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags,
2760 				     u16 raweth_qp1_flags2)
2761 {
2762 	bool is_ipv6 = false, is_ipv4 = false;
2763 
2764 	/* raweth_qp1_flags Bit 9-6 indicates itype */
2765 	if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
2766 	    != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
2767 		return -1;
2768 
2769 	if (raweth_qp1_flags2 &
2770 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC &&
2771 	    raweth_qp1_flags2 &
2772 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) {
2773 		/* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
2774 		(raweth_qp1_flags2 &
2775 		 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ?
2776 			(is_ipv6 = true) : (is_ipv4 = true);
2777 		return ((is_ipv6) ?
2778 			 BNXT_RE_ROCEV2_IPV6_PACKET :
2779 			 BNXT_RE_ROCEV2_IPV4_PACKET);
2780 	} else {
2781 		return BNXT_RE_ROCE_V1_PACKET;
2782 	}
2783 }
2784 
2785 static int bnxt_re_to_ib_nw_type(int nw_type)
2786 {
2787 	u8 nw_hdr_type = 0xFF;
2788 
2789 	switch (nw_type) {
2790 	case BNXT_RE_ROCE_V1_PACKET:
2791 		nw_hdr_type = RDMA_NETWORK_ROCE_V1;
2792 		break;
2793 	case BNXT_RE_ROCEV2_IPV4_PACKET:
2794 		nw_hdr_type = RDMA_NETWORK_IPV4;
2795 		break;
2796 	case BNXT_RE_ROCEV2_IPV6_PACKET:
2797 		nw_hdr_type = RDMA_NETWORK_IPV6;
2798 		break;
2799 	}
2800 	return nw_hdr_type;
2801 }
2802 
2803 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev,
2804 				       void *rq_hdr_buf)
2805 {
2806 	u8 *tmp_buf = NULL;
2807 	struct ethhdr *eth_hdr;
2808 	u16 eth_type;
2809 	bool rc = false;
2810 
2811 	tmp_buf = (u8 *)rq_hdr_buf;
2812 	/*
2813 	 * If dest mac is not same as I/F mac, this could be a
2814 	 * loopback address or multicast address, check whether
2815 	 * it is a loopback packet
2816 	 */
2817 	if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) {
2818 		tmp_buf += 4;
2819 		/* Check the  ether type */
2820 		eth_hdr = (struct ethhdr *)tmp_buf;
2821 		eth_type = ntohs(eth_hdr->h_proto);
2822 		switch (eth_type) {
2823 		case ETH_P_IBOE:
2824 			rc = true;
2825 			break;
2826 		case ETH_P_IP:
2827 		case ETH_P_IPV6: {
2828 			u32 len;
2829 			struct udphdr *udp_hdr;
2830 
2831 			len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) :
2832 						      sizeof(struct ipv6hdr));
2833 			tmp_buf += sizeof(struct ethhdr) + len;
2834 			udp_hdr = (struct udphdr *)tmp_buf;
2835 			if (ntohs(udp_hdr->dest) ==
2836 				    ROCE_V2_UDP_DPORT)
2837 				rc = true;
2838 			break;
2839 			}
2840 		default:
2841 			break;
2842 		}
2843 	}
2844 
2845 	return rc;
2846 }
2847 
2848 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *qp1_qp,
2849 					 struct bnxt_qplib_cqe *cqe)
2850 {
2851 	struct bnxt_re_dev *rdev = qp1_qp->rdev;
2852 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
2853 	struct bnxt_re_qp *qp = rdev->qp1_sqp;
2854 	struct ib_send_wr *swr;
2855 	struct ib_ud_wr udwr;
2856 	struct ib_recv_wr rwr;
2857 	int pkt_type = 0;
2858 	u32 tbl_idx;
2859 	void *rq_hdr_buf;
2860 	dma_addr_t rq_hdr_buf_map;
2861 	dma_addr_t shrq_hdr_buf_map;
2862 	u32 offset = 0;
2863 	u32 skip_bytes = 0;
2864 	struct ib_sge s_sge[2];
2865 	struct ib_sge r_sge[2];
2866 	int rc;
2867 
2868 	memset(&udwr, 0, sizeof(udwr));
2869 	memset(&rwr, 0, sizeof(rwr));
2870 	memset(&s_sge, 0, sizeof(s_sge));
2871 	memset(&r_sge, 0, sizeof(r_sge));
2872 
2873 	swr = &udwr.wr;
2874 	tbl_idx = cqe->wr_id;
2875 
2876 	rq_hdr_buf = qp1_qp->qplib_qp.rq_hdr_buf +
2877 			(tbl_idx * qp1_qp->qplib_qp.rq_hdr_buf_size);
2878 	rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&qp1_qp->qplib_qp,
2879 							  tbl_idx);
2880 
2881 	/* Shadow QP header buffer */
2882 	shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&qp->qplib_qp,
2883 							    tbl_idx);
2884 	sqp_entry = &rdev->sqp_tbl[tbl_idx];
2885 
2886 	/* Store this cqe */
2887 	memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe));
2888 	sqp_entry->qp1_qp = qp1_qp;
2889 
2890 	/* Find packet type from the cqe */
2891 
2892 	pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags,
2893 					     cqe->raweth_qp1_flags2);
2894 	if (pkt_type < 0) {
2895 		dev_err(rdev_to_dev(rdev), "Invalid packet\n");
2896 		return -EINVAL;
2897 	}
2898 
2899 	/* Adjust the offset for the user buffer and post in the rq */
2900 
2901 	if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET)
2902 		offset = 20;
2903 
2904 	/*
2905 	 * QP1 loopback packet has 4 bytes of internal header before
2906 	 * ether header. Skip these four bytes.
2907 	 */
2908 	if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf))
2909 		skip_bytes = 4;
2910 
2911 	/* First send SGE . Skip the ether header*/
2912 	s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
2913 			+ skip_bytes;
2914 	s_sge[0].lkey = 0xFFFFFFFF;
2915 	s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 :
2916 				BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
2917 
2918 	/* Second Send SGE */
2919 	s_sge[1].addr = s_sge[0].addr + s_sge[0].length +
2920 			BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE;
2921 	if (pkt_type != BNXT_RE_ROCE_V1_PACKET)
2922 		s_sge[1].addr += 8;
2923 	s_sge[1].lkey = 0xFFFFFFFF;
2924 	s_sge[1].length = 256;
2925 
2926 	/* First recv SGE */
2927 
2928 	r_sge[0].addr = shrq_hdr_buf_map;
2929 	r_sge[0].lkey = 0xFFFFFFFF;
2930 	r_sge[0].length = 40;
2931 
2932 	r_sge[1].addr = sqp_entry->sge.addr + offset;
2933 	r_sge[1].lkey = sqp_entry->sge.lkey;
2934 	r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset;
2935 
2936 	/* Create receive work request */
2937 	rwr.num_sge = 2;
2938 	rwr.sg_list = r_sge;
2939 	rwr.wr_id = tbl_idx;
2940 	rwr.next = NULL;
2941 
2942 	rc = bnxt_re_post_recv_shadow_qp(rdev, qp, &rwr);
2943 	if (rc) {
2944 		dev_err(rdev_to_dev(rdev),
2945 			"Failed to post Rx buffers to shadow QP");
2946 		return -ENOMEM;
2947 	}
2948 
2949 	swr->num_sge = 2;
2950 	swr->sg_list = s_sge;
2951 	swr->wr_id = tbl_idx;
2952 	swr->opcode = IB_WR_SEND;
2953 	swr->next = NULL;
2954 
2955 	udwr.ah = &rdev->sqp_ah->ib_ah;
2956 	udwr.remote_qpn = rdev->qp1_sqp->qplib_qp.id;
2957 	udwr.remote_qkey = rdev->qp1_sqp->qplib_qp.qkey;
2958 
2959 	/* post data received  in the send queue */
2960 	rc = bnxt_re_post_send_shadow_qp(rdev, qp, swr);
2961 
2962 	return 0;
2963 }
2964 
2965 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc,
2966 					  struct bnxt_qplib_cqe *cqe)
2967 {
2968 	wc->opcode = IB_WC_RECV;
2969 	wc->status = __rawqp1_to_ib_wc_status(cqe->status);
2970 	wc->wc_flags |= IB_WC_GRH;
2971 }
2972 
2973 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe,
2974 				u16 *vid, u8 *sl)
2975 {
2976 	bool ret = false;
2977 	u32 metadata;
2978 	u16 tpid;
2979 
2980 	metadata = orig_cqe->raweth_qp1_metadata;
2981 	if (orig_cqe->raweth_qp1_flags2 &
2982 		CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) {
2983 		tpid = ((metadata &
2984 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >>
2985 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT);
2986 		if (tpid == ETH_P_8021Q) {
2987 			*vid = metadata &
2988 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK;
2989 			*sl = (metadata &
2990 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >>
2991 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT;
2992 			ret = true;
2993 		}
2994 	}
2995 
2996 	return ret;
2997 }
2998 
2999 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc,
3000 				      struct bnxt_qplib_cqe *cqe)
3001 {
3002 	wc->opcode = IB_WC_RECV;
3003 	wc->status = __rc_to_ib_wc_status(cqe->status);
3004 
3005 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3006 		wc->wc_flags |= IB_WC_WITH_IMM;
3007 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3008 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3009 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3010 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3011 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3012 }
3013 
3014 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *qp,
3015 					     struct ib_wc *wc,
3016 					     struct bnxt_qplib_cqe *cqe)
3017 {
3018 	struct bnxt_re_dev *rdev = qp->rdev;
3019 	struct bnxt_re_qp *qp1_qp = NULL;
3020 	struct bnxt_qplib_cqe *orig_cqe = NULL;
3021 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3022 	int nw_type;
3023 	u32 tbl_idx;
3024 	u16 vlan_id;
3025 	u8 sl;
3026 
3027 	tbl_idx = cqe->wr_id;
3028 
3029 	sqp_entry = &rdev->sqp_tbl[tbl_idx];
3030 	qp1_qp = sqp_entry->qp1_qp;
3031 	orig_cqe = &sqp_entry->cqe;
3032 
3033 	wc->wr_id = sqp_entry->wrid;
3034 	wc->byte_len = orig_cqe->length;
3035 	wc->qp = &qp1_qp->ib_qp;
3036 
3037 	wc->ex.imm_data = orig_cqe->immdata;
3038 	wc->src_qp = orig_cqe->src_qp;
3039 	memcpy(wc->smac, orig_cqe->smac, ETH_ALEN);
3040 	if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) {
3041 		wc->vlan_id = vlan_id;
3042 		wc->sl = sl;
3043 		wc->wc_flags |= IB_WC_WITH_VLAN;
3044 	}
3045 	wc->port_num = 1;
3046 	wc->vendor_err = orig_cqe->status;
3047 
3048 	wc->opcode = IB_WC_RECV;
3049 	wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status);
3050 	wc->wc_flags |= IB_WC_GRH;
3051 
3052 	nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags,
3053 					    orig_cqe->raweth_qp1_flags2);
3054 	if (nw_type >= 0) {
3055 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3056 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3057 	}
3058 }
3059 
3060 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp,
3061 				      struct ib_wc *wc,
3062 				      struct bnxt_qplib_cqe *cqe)
3063 {
3064 	u8 nw_type;
3065 
3066 	wc->opcode = IB_WC_RECV;
3067 	wc->status = __rc_to_ib_wc_status(cqe->status);
3068 
3069 	if (cqe->flags & CQ_RES_UD_FLAGS_IMM)
3070 		wc->wc_flags |= IB_WC_WITH_IMM;
3071 	/* report only on GSI QP for Thor */
3072 	if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) {
3073 		wc->wc_flags |= IB_WC_GRH;
3074 		memcpy(wc->smac, cqe->smac, ETH_ALEN);
3075 		wc->wc_flags |= IB_WC_WITH_SMAC;
3076 		if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) {
3077 			wc->vlan_id = (cqe->cfa_meta & 0xFFF);
3078 			if (wc->vlan_id < 0x1000)
3079 				wc->wc_flags |= IB_WC_WITH_VLAN;
3080 		}
3081 		nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >>
3082 			   CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT;
3083 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3084 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3085 	}
3086 
3087 }
3088 
3089 static int send_phantom_wqe(struct bnxt_re_qp *qp)
3090 {
3091 	struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
3092 	unsigned long flags;
3093 	int rc = 0;
3094 
3095 	spin_lock_irqsave(&qp->sq_lock, flags);
3096 
3097 	rc = bnxt_re_bind_fence_mw(lib_qp);
3098 	if (!rc) {
3099 		lib_qp->sq.phantom_wqe_cnt++;
3100 		dev_dbg(&lib_qp->sq.hwq.pdev->dev,
3101 			"qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
3102 			lib_qp->id, lib_qp->sq.hwq.prod,
3103 			HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
3104 			lib_qp->sq.phantom_wqe_cnt);
3105 	}
3106 
3107 	spin_unlock_irqrestore(&qp->sq_lock, flags);
3108 	return rc;
3109 }
3110 
3111 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
3112 {
3113 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3114 	struct bnxt_re_qp *qp;
3115 	struct bnxt_qplib_cqe *cqe;
3116 	int i, ncqe, budget;
3117 	struct bnxt_qplib_q *sq;
3118 	struct bnxt_qplib_qp *lib_qp;
3119 	u32 tbl_idx;
3120 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3121 	unsigned long flags;
3122 
3123 	spin_lock_irqsave(&cq->cq_lock, flags);
3124 	budget = min_t(u32, num_entries, cq->max_cql);
3125 	num_entries = budget;
3126 	if (!cq->cql) {
3127 		dev_err(rdev_to_dev(cq->rdev), "POLL CQ : no CQL to use");
3128 		goto exit;
3129 	}
3130 	cqe = &cq->cql[0];
3131 	while (budget) {
3132 		lib_qp = NULL;
3133 		ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
3134 		if (lib_qp) {
3135 			sq = &lib_qp->sq;
3136 			if (sq->send_phantom) {
3137 				qp = container_of(lib_qp,
3138 						  struct bnxt_re_qp, qplib_qp);
3139 				if (send_phantom_wqe(qp) == -ENOMEM)
3140 					dev_err(rdev_to_dev(cq->rdev),
3141 						"Phantom failed! Scheduled to send again\n");
3142 				else
3143 					sq->send_phantom = false;
3144 			}
3145 		}
3146 		if (ncqe < budget)
3147 			ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq,
3148 							      cqe + ncqe,
3149 							      budget - ncqe);
3150 
3151 		if (!ncqe)
3152 			break;
3153 
3154 		for (i = 0; i < ncqe; i++, cqe++) {
3155 			/* Transcribe each qplib_wqe back to ib_wc */
3156 			memset(wc, 0, sizeof(*wc));
3157 
3158 			wc->wr_id = cqe->wr_id;
3159 			wc->byte_len = cqe->length;
3160 			qp = container_of
3161 				((struct bnxt_qplib_qp *)
3162 				 (unsigned long)(cqe->qp_handle),
3163 				 struct bnxt_re_qp, qplib_qp);
3164 			if (!qp) {
3165 				dev_err(rdev_to_dev(cq->rdev),
3166 					"POLL CQ : bad QP handle");
3167 				continue;
3168 			}
3169 			wc->qp = &qp->ib_qp;
3170 			wc->ex.imm_data = cqe->immdata;
3171 			wc->src_qp = cqe->src_qp;
3172 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
3173 			wc->port_num = 1;
3174 			wc->vendor_err = cqe->status;
3175 
3176 			switch (cqe->opcode) {
3177 			case CQ_BASE_CQE_TYPE_REQ:
3178 				if (qp->rdev->qp1_sqp && qp->qplib_qp.id ==
3179 				    qp->rdev->qp1_sqp->qplib_qp.id) {
3180 					/* Handle this completion with
3181 					 * the stored completion
3182 					 */
3183 					memset(wc, 0, sizeof(*wc));
3184 					continue;
3185 				}
3186 				bnxt_re_process_req_wc(wc, cqe);
3187 				break;
3188 			case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1:
3189 				if (!cqe->status) {
3190 					int rc = 0;
3191 
3192 					rc = bnxt_re_process_raw_qp_pkt_rx
3193 								(qp, cqe);
3194 					if (!rc) {
3195 						memset(wc, 0, sizeof(*wc));
3196 						continue;
3197 					}
3198 					cqe->status = -1;
3199 				}
3200 				/* Errors need not be looped back.
3201 				 * But change the wr_id to the one
3202 				 * stored in the table
3203 				 */
3204 				tbl_idx = cqe->wr_id;
3205 				sqp_entry = &cq->rdev->sqp_tbl[tbl_idx];
3206 				wc->wr_id = sqp_entry->wrid;
3207 				bnxt_re_process_res_rawqp1_wc(wc, cqe);
3208 				break;
3209 			case CQ_BASE_CQE_TYPE_RES_RC:
3210 				bnxt_re_process_res_rc_wc(wc, cqe);
3211 				break;
3212 			case CQ_BASE_CQE_TYPE_RES_UD:
3213 				if (qp->rdev->qp1_sqp && qp->qplib_qp.id ==
3214 				    qp->rdev->qp1_sqp->qplib_qp.id) {
3215 					/* Handle this completion with
3216 					 * the stored completion
3217 					 */
3218 					if (cqe->status) {
3219 						continue;
3220 					} else {
3221 						bnxt_re_process_res_shadow_qp_wc
3222 								(qp, wc, cqe);
3223 						break;
3224 					}
3225 				}
3226 				bnxt_re_process_res_ud_wc(qp, wc, cqe);
3227 				break;
3228 			default:
3229 				dev_err(rdev_to_dev(cq->rdev),
3230 					"POLL CQ : type 0x%x not handled",
3231 					cqe->opcode);
3232 				continue;
3233 			}
3234 			wc++;
3235 			budget--;
3236 		}
3237 	}
3238 exit:
3239 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3240 	return num_entries - budget;
3241 }
3242 
3243 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq,
3244 			  enum ib_cq_notify_flags ib_cqn_flags)
3245 {
3246 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3247 	int type = 0, rc = 0;
3248 	unsigned long flags;
3249 
3250 	spin_lock_irqsave(&cq->cq_lock, flags);
3251 	/* Trigger on the very next completion */
3252 	if (ib_cqn_flags & IB_CQ_NEXT_COMP)
3253 		type = DBC_DBC_TYPE_CQ_ARMALL;
3254 	/* Trigger on the next solicited completion */
3255 	else if (ib_cqn_flags & IB_CQ_SOLICITED)
3256 		type = DBC_DBC_TYPE_CQ_ARMSE;
3257 
3258 	/* Poll to see if there are missed events */
3259 	if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3260 	    !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) {
3261 		rc = 1;
3262 		goto exit;
3263 	}
3264 	bnxt_qplib_req_notify_cq(&cq->qplib_cq, type);
3265 
3266 exit:
3267 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3268 	return rc;
3269 }
3270 
3271 /* Memory Regions */
3272 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags)
3273 {
3274 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3275 	struct bnxt_re_dev *rdev = pd->rdev;
3276 	struct bnxt_re_mr *mr;
3277 	u64 pbl = 0;
3278 	int rc;
3279 
3280 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3281 	if (!mr)
3282 		return ERR_PTR(-ENOMEM);
3283 
3284 	mr->rdev = rdev;
3285 	mr->qplib_mr.pd = &pd->qplib_pd;
3286 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3287 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3288 
3289 	/* Allocate and register 0 as the address */
3290 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3291 	if (rc)
3292 		goto fail;
3293 
3294 	mr->qplib_mr.hwq.level = PBL_LVL_MAX;
3295 	mr->qplib_mr.total_size = -1; /* Infinte length */
3296 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl, 0, false,
3297 			       PAGE_SIZE);
3298 	if (rc)
3299 		goto fail_mr;
3300 
3301 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3302 	if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ |
3303 			       IB_ACCESS_REMOTE_ATOMIC))
3304 		mr->ib_mr.rkey = mr->ib_mr.lkey;
3305 	atomic_inc(&rdev->mr_count);
3306 
3307 	return &mr->ib_mr;
3308 
3309 fail_mr:
3310 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3311 fail:
3312 	kfree(mr);
3313 	return ERR_PTR(rc);
3314 }
3315 
3316 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3317 {
3318 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3319 	struct bnxt_re_dev *rdev = mr->rdev;
3320 	int rc;
3321 
3322 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3323 	if (rc)
3324 		dev_err(rdev_to_dev(rdev), "Dereg MR failed: %#x\n", rc);
3325 
3326 	if (mr->pages) {
3327 		rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
3328 							&mr->qplib_frpl);
3329 		kfree(mr->pages);
3330 		mr->npages = 0;
3331 		mr->pages = NULL;
3332 	}
3333 	ib_umem_release(mr->ib_umem);
3334 
3335 	kfree(mr);
3336 	atomic_dec(&rdev->mr_count);
3337 	return rc;
3338 }
3339 
3340 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr)
3341 {
3342 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3343 
3344 	if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs))
3345 		return -ENOMEM;
3346 
3347 	mr->pages[mr->npages++] = addr;
3348 	return 0;
3349 }
3350 
3351 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents,
3352 		      unsigned int *sg_offset)
3353 {
3354 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3355 
3356 	mr->npages = 0;
3357 	return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page);
3358 }
3359 
3360 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type,
3361 			       u32 max_num_sg, struct ib_udata *udata)
3362 {
3363 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3364 	struct bnxt_re_dev *rdev = pd->rdev;
3365 	struct bnxt_re_mr *mr = NULL;
3366 	int rc;
3367 
3368 	if (type != IB_MR_TYPE_MEM_REG) {
3369 		dev_dbg(rdev_to_dev(rdev), "MR type 0x%x not supported", type);
3370 		return ERR_PTR(-EINVAL);
3371 	}
3372 	if (max_num_sg > MAX_PBL_LVL_1_PGS)
3373 		return ERR_PTR(-EINVAL);
3374 
3375 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3376 	if (!mr)
3377 		return ERR_PTR(-ENOMEM);
3378 
3379 	mr->rdev = rdev;
3380 	mr->qplib_mr.pd = &pd->qplib_pd;
3381 	mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR;
3382 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3383 
3384 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3385 	if (rc)
3386 		goto bail;
3387 
3388 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3389 	mr->ib_mr.rkey = mr->ib_mr.lkey;
3390 
3391 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3392 	if (!mr->pages) {
3393 		rc = -ENOMEM;
3394 		goto fail;
3395 	}
3396 	rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res,
3397 						 &mr->qplib_frpl, max_num_sg);
3398 	if (rc) {
3399 		dev_err(rdev_to_dev(rdev),
3400 			"Failed to allocate HW FR page list");
3401 		goto fail_mr;
3402 	}
3403 
3404 	atomic_inc(&rdev->mr_count);
3405 	return &mr->ib_mr;
3406 
3407 fail_mr:
3408 	kfree(mr->pages);
3409 fail:
3410 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3411 bail:
3412 	kfree(mr);
3413 	return ERR_PTR(rc);
3414 }
3415 
3416 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
3417 			       struct ib_udata *udata)
3418 {
3419 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3420 	struct bnxt_re_dev *rdev = pd->rdev;
3421 	struct bnxt_re_mw *mw;
3422 	int rc;
3423 
3424 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
3425 	if (!mw)
3426 		return ERR_PTR(-ENOMEM);
3427 	mw->rdev = rdev;
3428 	mw->qplib_mw.pd = &pd->qplib_pd;
3429 
3430 	mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
3431 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
3432 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
3433 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
3434 	if (rc) {
3435 		dev_err(rdev_to_dev(rdev), "Allocate MW failed!");
3436 		goto fail;
3437 	}
3438 	mw->ib_mw.rkey = mw->qplib_mw.rkey;
3439 
3440 	atomic_inc(&rdev->mw_count);
3441 	return &mw->ib_mw;
3442 
3443 fail:
3444 	kfree(mw);
3445 	return ERR_PTR(rc);
3446 }
3447 
3448 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
3449 {
3450 	struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
3451 	struct bnxt_re_dev *rdev = mw->rdev;
3452 	int rc;
3453 
3454 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
3455 	if (rc) {
3456 		dev_err(rdev_to_dev(rdev), "Free MW failed: %#x\n", rc);
3457 		return rc;
3458 	}
3459 
3460 	kfree(mw);
3461 	atomic_dec(&rdev->mw_count);
3462 	return rc;
3463 }
3464 
3465 static int bnxt_re_page_size_ok(int page_shift)
3466 {
3467 	switch (page_shift) {
3468 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4K:
3469 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_8K:
3470 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_64K:
3471 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_2M:
3472 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_256K:
3473 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1M:
3474 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4M:
3475 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1G:
3476 		return 1;
3477 	default:
3478 		return 0;
3479 	}
3480 }
3481 
3482 static int fill_umem_pbl_tbl(struct ib_umem *umem, u64 *pbl_tbl_orig,
3483 			     int page_shift)
3484 {
3485 	u64 *pbl_tbl = pbl_tbl_orig;
3486 	u64 page_size =  BIT_ULL(page_shift);
3487 	struct ib_block_iter biter;
3488 
3489 	rdma_for_each_block(umem->sg_head.sgl, &biter, umem->nmap, page_size)
3490 		*pbl_tbl++ = rdma_block_iter_dma_address(&biter);
3491 
3492 	return pbl_tbl - pbl_tbl_orig;
3493 }
3494 
3495 /* uverbs */
3496 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
3497 				  u64 virt_addr, int mr_access_flags,
3498 				  struct ib_udata *udata)
3499 {
3500 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3501 	struct bnxt_re_dev *rdev = pd->rdev;
3502 	struct bnxt_re_mr *mr;
3503 	struct ib_umem *umem;
3504 	u64 *pbl_tbl = NULL;
3505 	int umem_pgs, page_shift, rc;
3506 
3507 	if (length > BNXT_RE_MAX_MR_SIZE) {
3508 		dev_err(rdev_to_dev(rdev), "MR Size: %lld > Max supported:%lld\n",
3509 			length, BNXT_RE_MAX_MR_SIZE);
3510 		return ERR_PTR(-ENOMEM);
3511 	}
3512 
3513 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3514 	if (!mr)
3515 		return ERR_PTR(-ENOMEM);
3516 
3517 	mr->rdev = rdev;
3518 	mr->qplib_mr.pd = &pd->qplib_pd;
3519 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3520 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
3521 
3522 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3523 	if (rc) {
3524 		dev_err(rdev_to_dev(rdev), "Failed to allocate MR");
3525 		goto free_mr;
3526 	}
3527 	/* The fixed portion of the rkey is the same as the lkey */
3528 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
3529 
3530 	umem = ib_umem_get(udata, start, length, mr_access_flags, 0);
3531 	if (IS_ERR(umem)) {
3532 		dev_err(rdev_to_dev(rdev), "Failed to get umem");
3533 		rc = -EFAULT;
3534 		goto free_mrw;
3535 	}
3536 	mr->ib_umem = umem;
3537 
3538 	mr->qplib_mr.va = virt_addr;
3539 	umem_pgs = ib_umem_page_count(umem);
3540 	if (!umem_pgs) {
3541 		dev_err(rdev_to_dev(rdev), "umem is invalid!");
3542 		rc = -EINVAL;
3543 		goto free_umem;
3544 	}
3545 	mr->qplib_mr.total_size = length;
3546 
3547 	pbl_tbl = kcalloc(umem_pgs, sizeof(u64 *), GFP_KERNEL);
3548 	if (!pbl_tbl) {
3549 		rc = -ENOMEM;
3550 		goto free_umem;
3551 	}
3552 
3553 	page_shift = __ffs(ib_umem_find_best_pgsz(umem,
3554 				BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M,
3555 				virt_addr));
3556 
3557 	if (!bnxt_re_page_size_ok(page_shift)) {
3558 		dev_err(rdev_to_dev(rdev), "umem page size unsupported!");
3559 		rc = -EFAULT;
3560 		goto fail;
3561 	}
3562 
3563 	if (page_shift == BNXT_RE_PAGE_SHIFT_4K &&
3564 	    length > BNXT_RE_MAX_MR_SIZE_LOW) {
3565 		dev_err(rdev_to_dev(rdev), "Requested MR Sz:%llu Max sup:%llu",
3566 			length,	(u64)BNXT_RE_MAX_MR_SIZE_LOW);
3567 		rc = -EINVAL;
3568 		goto fail;
3569 	}
3570 
3571 	/* Map umem buf ptrs to the PBL */
3572 	umem_pgs = fill_umem_pbl_tbl(umem, pbl_tbl, page_shift);
3573 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, pbl_tbl,
3574 			       umem_pgs, false, 1 << page_shift);
3575 	if (rc) {
3576 		dev_err(rdev_to_dev(rdev), "Failed to register user MR");
3577 		goto fail;
3578 	}
3579 
3580 	kfree(pbl_tbl);
3581 
3582 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3583 	mr->ib_mr.rkey = mr->qplib_mr.lkey;
3584 	atomic_inc(&rdev->mr_count);
3585 
3586 	return &mr->ib_mr;
3587 fail:
3588 	kfree(pbl_tbl);
3589 free_umem:
3590 	ib_umem_release(umem);
3591 free_mrw:
3592 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3593 free_mr:
3594 	kfree(mr);
3595 	return ERR_PTR(rc);
3596 }
3597 
3598 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
3599 {
3600 	struct ib_device *ibdev = ctx->device;
3601 	struct bnxt_re_ucontext *uctx =
3602 		container_of(ctx, struct bnxt_re_ucontext, ib_uctx);
3603 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
3604 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
3605 	struct bnxt_re_uctx_resp resp;
3606 	u32 chip_met_rev_num = 0;
3607 	int rc;
3608 
3609 	dev_dbg(rdev_to_dev(rdev), "ABI version requested %u",
3610 		ibdev->ops.uverbs_abi_ver);
3611 
3612 	if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) {
3613 		dev_dbg(rdev_to_dev(rdev), " is different from the device %d ",
3614 			BNXT_RE_ABI_VERSION);
3615 		return -EPERM;
3616 	}
3617 
3618 	uctx->rdev = rdev;
3619 
3620 	uctx->shpg = (void *)__get_free_page(GFP_KERNEL);
3621 	if (!uctx->shpg) {
3622 		rc = -ENOMEM;
3623 		goto fail;
3624 	}
3625 	spin_lock_init(&uctx->sh_lock);
3626 
3627 	resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX;
3628 	chip_met_rev_num = rdev->chip_ctx.chip_num;
3629 	chip_met_rev_num |= ((u32)rdev->chip_ctx.chip_rev & 0xFF) <<
3630 			     BNXT_RE_CHIP_ID0_CHIP_REV_SFT;
3631 	chip_met_rev_num |= ((u32)rdev->chip_ctx.chip_metal & 0xFF) <<
3632 			     BNXT_RE_CHIP_ID0_CHIP_MET_SFT;
3633 	resp.chip_id0 = chip_met_rev_num;
3634 	/* Future extension of chip info */
3635 	resp.chip_id1 = 0;
3636 	/*Temp, Use xa_alloc instead */
3637 	resp.dev_id = rdev->en_dev->pdev->devfn;
3638 	resp.max_qp = rdev->qplib_ctx.qpc_count;
3639 	resp.pg_size = PAGE_SIZE;
3640 	resp.cqe_sz = sizeof(struct cq_base);
3641 	resp.max_cqd = dev_attr->max_cq_wqes;
3642 	resp.rsvd    = 0;
3643 
3644 	rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
3645 	if (rc) {
3646 		dev_err(rdev_to_dev(rdev), "Failed to copy user context");
3647 		rc = -EFAULT;
3648 		goto cfail;
3649 	}
3650 
3651 	return 0;
3652 cfail:
3653 	free_page((unsigned long)uctx->shpg);
3654 	uctx->shpg = NULL;
3655 fail:
3656 	return rc;
3657 }
3658 
3659 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
3660 {
3661 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3662 						   struct bnxt_re_ucontext,
3663 						   ib_uctx);
3664 
3665 	struct bnxt_re_dev *rdev = uctx->rdev;
3666 
3667 	if (uctx->shpg)
3668 		free_page((unsigned long)uctx->shpg);
3669 
3670 	if (uctx->dpi.dbr) {
3671 		/* Free DPI only if this is the first PD allocated by the
3672 		 * application and mark the context dpi as NULL
3673 		 */
3674 		bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
3675 				       &rdev->qplib_res.dpi_tbl, &uctx->dpi);
3676 		uctx->dpi.dbr = NULL;
3677 	}
3678 }
3679 
3680 /* Helper function to mmap the virtual memory from user app */
3681 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
3682 {
3683 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3684 						   struct bnxt_re_ucontext,
3685 						   ib_uctx);
3686 	struct bnxt_re_dev *rdev = uctx->rdev;
3687 	u64 pfn;
3688 
3689 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3690 		return -EINVAL;
3691 
3692 	if (vma->vm_pgoff) {
3693 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3694 		if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
3695 				       PAGE_SIZE, vma->vm_page_prot)) {
3696 			dev_err(rdev_to_dev(rdev), "Failed to map DPI");
3697 			return -EAGAIN;
3698 		}
3699 	} else {
3700 		pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT;
3701 		if (remap_pfn_range(vma, vma->vm_start,
3702 				    pfn, PAGE_SIZE, vma->vm_page_prot)) {
3703 			dev_err(rdev_to_dev(rdev),
3704 				"Failed to map shared page");
3705 			return -EAGAIN;
3706 		}
3707 	}
3708 
3709 	return 0;
3710 }
3711