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 #include <net/addrconf.h>
45 
46 #include <rdma/ib_verbs.h>
47 #include <rdma/ib_user_verbs.h>
48 #include <rdma/ib_umem.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_mad.h>
51 #include <rdma/ib_cache.h>
52 #include <rdma/uverbs_ioctl.h>
53 
54 #include "bnxt_ulp.h"
55 
56 #include "roce_hsi.h"
57 #include "qplib_res.h"
58 #include "qplib_sp.h"
59 #include "qplib_fp.h"
60 #include "qplib_rcfw.h"
61 
62 #include "bnxt_re.h"
63 #include "ib_verbs.h"
64 
65 #include <rdma/uverbs_types.h>
66 #include <rdma/uverbs_std_types.h>
67 
68 #include <rdma/ib_user_ioctl_cmds.h>
69 
70 #define UVERBS_MODULE_NAME bnxt_re
71 #include <rdma/uverbs_named_ioctl.h>
72 
73 #include <rdma/bnxt_re-abi.h>
74 
75 static int __from_ib_access_flags(int iflags)
76 {
77 	int qflags = 0;
78 
79 	if (iflags & IB_ACCESS_LOCAL_WRITE)
80 		qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
81 	if (iflags & IB_ACCESS_REMOTE_READ)
82 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ;
83 	if (iflags & IB_ACCESS_REMOTE_WRITE)
84 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE;
85 	if (iflags & IB_ACCESS_REMOTE_ATOMIC)
86 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC;
87 	if (iflags & IB_ACCESS_MW_BIND)
88 		qflags |= BNXT_QPLIB_ACCESS_MW_BIND;
89 	if (iflags & IB_ZERO_BASED)
90 		qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED;
91 	if (iflags & IB_ACCESS_ON_DEMAND)
92 		qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND;
93 	return qflags;
94 };
95 
96 static enum ib_access_flags __to_ib_access_flags(int qflags)
97 {
98 	enum ib_access_flags iflags = 0;
99 
100 	if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE)
101 		iflags |= IB_ACCESS_LOCAL_WRITE;
102 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE)
103 		iflags |= IB_ACCESS_REMOTE_WRITE;
104 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ)
105 		iflags |= IB_ACCESS_REMOTE_READ;
106 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC)
107 		iflags |= IB_ACCESS_REMOTE_ATOMIC;
108 	if (qflags & BNXT_QPLIB_ACCESS_MW_BIND)
109 		iflags |= IB_ACCESS_MW_BIND;
110 	if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED)
111 		iflags |= IB_ZERO_BASED;
112 	if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND)
113 		iflags |= IB_ACCESS_ON_DEMAND;
114 	return iflags;
115 };
116 
117 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list,
118 			     struct bnxt_qplib_sge *sg_list, int num)
119 {
120 	int i, total = 0;
121 
122 	for (i = 0; i < num; i++) {
123 		sg_list[i].addr = ib_sg_list[i].addr;
124 		sg_list[i].lkey = ib_sg_list[i].lkey;
125 		sg_list[i].size = ib_sg_list[i].length;
126 		total += sg_list[i].size;
127 	}
128 	return total;
129 }
130 
131 /* Device */
132 int bnxt_re_query_device(struct ib_device *ibdev,
133 			 struct ib_device_attr *ib_attr,
134 			 struct ib_udata *udata)
135 {
136 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
137 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
138 
139 	memset(ib_attr, 0, sizeof(*ib_attr));
140 	memcpy(&ib_attr->fw_ver, dev_attr->fw_ver,
141 	       min(sizeof(dev_attr->fw_ver),
142 		   sizeof(ib_attr->fw_ver)));
143 	addrconf_addr_eui48((u8 *)&ib_attr->sys_image_guid,
144 			    rdev->netdev->dev_addr);
145 	ib_attr->max_mr_size = BNXT_RE_MAX_MR_SIZE;
146 	ib_attr->page_size_cap = BNXT_RE_PAGE_SIZE_SUPPORTED;
147 
148 	ib_attr->vendor_id = rdev->en_dev->pdev->vendor;
149 	ib_attr->vendor_part_id = rdev->en_dev->pdev->device;
150 	ib_attr->hw_ver = rdev->en_dev->pdev->subsystem_device;
151 	ib_attr->max_qp = dev_attr->max_qp;
152 	ib_attr->max_qp_wr = dev_attr->max_qp_wqes;
153 	ib_attr->device_cap_flags =
154 				    IB_DEVICE_CURR_QP_STATE_MOD
155 				    | IB_DEVICE_RC_RNR_NAK_GEN
156 				    | IB_DEVICE_SHUTDOWN_PORT
157 				    | IB_DEVICE_SYS_IMAGE_GUID
158 				    | IB_DEVICE_RESIZE_MAX_WR
159 				    | IB_DEVICE_PORT_ACTIVE_EVENT
160 				    | IB_DEVICE_N_NOTIFY_CQ
161 				    | IB_DEVICE_MEM_WINDOW
162 				    | IB_DEVICE_MEM_WINDOW_TYPE_2B
163 				    | IB_DEVICE_MEM_MGT_EXTENSIONS;
164 	ib_attr->kernel_cap_flags = IBK_LOCAL_DMA_LKEY;
165 	ib_attr->max_send_sge = dev_attr->max_qp_sges;
166 	ib_attr->max_recv_sge = dev_attr->max_qp_sges;
167 	ib_attr->max_sge_rd = dev_attr->max_qp_sges;
168 	ib_attr->max_cq = dev_attr->max_cq;
169 	ib_attr->max_cqe = dev_attr->max_cq_wqes;
170 	ib_attr->max_mr = dev_attr->max_mr;
171 	ib_attr->max_pd = dev_attr->max_pd;
172 	ib_attr->max_qp_rd_atom = dev_attr->max_qp_rd_atom;
173 	ib_attr->max_qp_init_rd_atom = dev_attr->max_qp_init_rd_atom;
174 	ib_attr->atomic_cap = IB_ATOMIC_NONE;
175 	ib_attr->masked_atomic_cap = IB_ATOMIC_NONE;
176 	if (dev_attr->is_atomic) {
177 		ib_attr->atomic_cap = IB_ATOMIC_GLOB;
178 		ib_attr->masked_atomic_cap = IB_ATOMIC_GLOB;
179 	}
180 
181 	ib_attr->max_ee_rd_atom = 0;
182 	ib_attr->max_res_rd_atom = 0;
183 	ib_attr->max_ee_init_rd_atom = 0;
184 	ib_attr->max_ee = 0;
185 	ib_attr->max_rdd = 0;
186 	ib_attr->max_mw = dev_attr->max_mw;
187 	ib_attr->max_raw_ipv6_qp = 0;
188 	ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp;
189 	ib_attr->max_mcast_grp = 0;
190 	ib_attr->max_mcast_qp_attach = 0;
191 	ib_attr->max_total_mcast_qp_attach = 0;
192 	ib_attr->max_ah = dev_attr->max_ah;
193 
194 	ib_attr->max_srq = dev_attr->max_srq;
195 	ib_attr->max_srq_wr = dev_attr->max_srq_wqes;
196 	ib_attr->max_srq_sge = dev_attr->max_srq_sges;
197 
198 	ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS;
199 
200 	ib_attr->max_pkeys = 1;
201 	ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY;
202 	return 0;
203 }
204 
205 /* Port */
206 int bnxt_re_query_port(struct ib_device *ibdev, u32 port_num,
207 		       struct ib_port_attr *port_attr)
208 {
209 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
210 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
211 	int rc;
212 
213 	memset(port_attr, 0, sizeof(*port_attr));
214 
215 	if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) {
216 		port_attr->state = IB_PORT_ACTIVE;
217 		port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
218 	} else {
219 		port_attr->state = IB_PORT_DOWN;
220 		port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
221 	}
222 	port_attr->max_mtu = IB_MTU_4096;
223 	port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu);
224 	port_attr->gid_tbl_len = dev_attr->max_sgid;
225 	port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
226 				    IB_PORT_DEVICE_MGMT_SUP |
227 				    IB_PORT_VENDOR_CLASS_SUP;
228 	port_attr->ip_gids = true;
229 
230 	port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW;
231 	port_attr->bad_pkey_cntr = 0;
232 	port_attr->qkey_viol_cntr = 0;
233 	port_attr->pkey_tbl_len = dev_attr->max_pkey;
234 	port_attr->lid = 0;
235 	port_attr->sm_lid = 0;
236 	port_attr->lmc = 0;
237 	port_attr->max_vl_num = 4;
238 	port_attr->sm_sl = 0;
239 	port_attr->subnet_timeout = 0;
240 	port_attr->init_type_reply = 0;
241 	rc = ib_get_eth_speed(&rdev->ibdev, port_num, &port_attr->active_speed,
242 			      &port_attr->active_width);
243 
244 	return rc;
245 }
246 
247 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u32 port_num,
248 			       struct ib_port_immutable *immutable)
249 {
250 	struct ib_port_attr port_attr;
251 
252 	if (bnxt_re_query_port(ibdev, port_num, &port_attr))
253 		return -EINVAL;
254 
255 	immutable->pkey_tbl_len = port_attr.pkey_tbl_len;
256 	immutable->gid_tbl_len = port_attr.gid_tbl_len;
257 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
258 	immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP;
259 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
260 	return 0;
261 }
262 
263 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str)
264 {
265 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
266 
267 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d",
268 		 rdev->dev_attr.fw_ver[0], rdev->dev_attr.fw_ver[1],
269 		 rdev->dev_attr.fw_ver[2], rdev->dev_attr.fw_ver[3]);
270 }
271 
272 int bnxt_re_query_pkey(struct ib_device *ibdev, u32 port_num,
273 		       u16 index, u16 *pkey)
274 {
275 	if (index > 0)
276 		return -EINVAL;
277 
278 	*pkey = IB_DEFAULT_PKEY_FULL;
279 
280 	return 0;
281 }
282 
283 int bnxt_re_query_gid(struct ib_device *ibdev, u32 port_num,
284 		      int index, union ib_gid *gid)
285 {
286 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
287 	int rc;
288 
289 	/* Ignore port_num */
290 	memset(gid, 0, sizeof(*gid));
291 	rc = bnxt_qplib_get_sgid(&rdev->qplib_res,
292 				 &rdev->qplib_res.sgid_tbl, index,
293 				 (struct bnxt_qplib_gid *)gid);
294 	return rc;
295 }
296 
297 int bnxt_re_del_gid(const struct ib_gid_attr *attr, void **context)
298 {
299 	int rc = 0;
300 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
301 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
302 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
303 	struct bnxt_qplib_gid *gid_to_del;
304 	u16 vlan_id = 0xFFFF;
305 
306 	/* Delete the entry from the hardware */
307 	ctx = *context;
308 	if (!ctx)
309 		return -EINVAL;
310 
311 	if (sgid_tbl && sgid_tbl->active) {
312 		if (ctx->idx >= sgid_tbl->max)
313 			return -EINVAL;
314 		gid_to_del = &sgid_tbl->tbl[ctx->idx].gid;
315 		vlan_id = sgid_tbl->tbl[ctx->idx].vlan_id;
316 		/* DEL_GID is called in WQ context(netdevice_event_work_handler)
317 		 * or via the ib_unregister_device path. In the former case QP1
318 		 * may not be destroyed yet, in which case just return as FW
319 		 * needs that entry to be present and will fail it's deletion.
320 		 * We could get invoked again after QP1 is destroyed OR get an
321 		 * ADD_GID call with a different GID value for the same index
322 		 * where we issue MODIFY_GID cmd to update the GID entry -- TBD
323 		 */
324 		if (ctx->idx == 0 &&
325 		    rdma_link_local_addr((struct in6_addr *)gid_to_del) &&
326 		    ctx->refcnt == 1 && rdev->gsi_ctx.gsi_sqp) {
327 			ibdev_dbg(&rdev->ibdev,
328 				  "Trying to delete GID0 while QP1 is alive\n");
329 			return -EFAULT;
330 		}
331 		ctx->refcnt--;
332 		if (!ctx->refcnt) {
333 			rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del,
334 						 vlan_id,  true);
335 			if (rc) {
336 				ibdev_err(&rdev->ibdev,
337 					  "Failed to remove GID: %#x", rc);
338 			} else {
339 				ctx_tbl = sgid_tbl->ctx;
340 				ctx_tbl[ctx->idx] = NULL;
341 				kfree(ctx);
342 			}
343 		}
344 	} else {
345 		return -EINVAL;
346 	}
347 	return rc;
348 }
349 
350 int bnxt_re_add_gid(const struct ib_gid_attr *attr, void **context)
351 {
352 	int rc;
353 	u32 tbl_idx = 0;
354 	u16 vlan_id = 0xFFFF;
355 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
356 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
357 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
358 
359 	rc = rdma_read_gid_l2_fields(attr, &vlan_id, NULL);
360 	if (rc)
361 		return rc;
362 
363 	rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)&attr->gid,
364 				 rdev->qplib_res.netdev->dev_addr,
365 				 vlan_id, true, &tbl_idx);
366 	if (rc == -EALREADY) {
367 		ctx_tbl = sgid_tbl->ctx;
368 		ctx_tbl[tbl_idx]->refcnt++;
369 		*context = ctx_tbl[tbl_idx];
370 		return 0;
371 	}
372 
373 	if (rc < 0) {
374 		ibdev_err(&rdev->ibdev, "Failed to add GID: %#x", rc);
375 		return rc;
376 	}
377 
378 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
379 	if (!ctx)
380 		return -ENOMEM;
381 	ctx_tbl = sgid_tbl->ctx;
382 	ctx->idx = tbl_idx;
383 	ctx->refcnt = 1;
384 	ctx_tbl[tbl_idx] = ctx;
385 	*context = ctx;
386 
387 	return rc;
388 }
389 
390 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
391 					    u32 port_num)
392 {
393 	return IB_LINK_LAYER_ETHERNET;
394 }
395 
396 #define	BNXT_RE_FENCE_PBL_SIZE	DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
397 
398 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd)
399 {
400 	struct bnxt_re_fence_data *fence = &pd->fence;
401 	struct ib_mr *ib_mr = &fence->mr->ib_mr;
402 	struct bnxt_qplib_swqe *wqe = &fence->bind_wqe;
403 
404 	memset(wqe, 0, sizeof(*wqe));
405 	wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW;
406 	wqe->wr_id = BNXT_QPLIB_FENCE_WRID;
407 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
408 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
409 	wqe->bind.zero_based = false;
410 	wqe->bind.parent_l_key = ib_mr->lkey;
411 	wqe->bind.va = (u64)(unsigned long)fence->va;
412 	wqe->bind.length = fence->size;
413 	wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ);
414 	wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1;
415 
416 	/* Save the initial rkey in fence structure for now;
417 	 * wqe->bind.r_key will be set at (re)bind time.
418 	 */
419 	fence->bind_rkey = ib_inc_rkey(fence->mw->rkey);
420 }
421 
422 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp)
423 {
424 	struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp,
425 					     qplib_qp);
426 	struct ib_pd *ib_pd = qp->ib_qp.pd;
427 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
428 	struct bnxt_re_fence_data *fence = &pd->fence;
429 	struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe;
430 	struct bnxt_qplib_swqe wqe;
431 	int rc;
432 
433 	memcpy(&wqe, fence_wqe, sizeof(wqe));
434 	wqe.bind.r_key = fence->bind_rkey;
435 	fence->bind_rkey = ib_inc_rkey(fence->bind_rkey);
436 
437 	ibdev_dbg(&qp->rdev->ibdev,
438 		  "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
439 		wqe.bind.r_key, qp->qplib_qp.id, pd);
440 	rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
441 	if (rc) {
442 		ibdev_err(&qp->rdev->ibdev, "Failed to bind fence-WQE\n");
443 		return rc;
444 	}
445 	bnxt_qplib_post_send_db(&qp->qplib_qp);
446 
447 	return rc;
448 }
449 
450 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd)
451 {
452 	struct bnxt_re_fence_data *fence = &pd->fence;
453 	struct bnxt_re_dev *rdev = pd->rdev;
454 	struct device *dev = &rdev->en_dev->pdev->dev;
455 	struct bnxt_re_mr *mr = fence->mr;
456 
457 	if (fence->mw) {
458 		bnxt_re_dealloc_mw(fence->mw);
459 		fence->mw = NULL;
460 	}
461 	if (mr) {
462 		if (mr->ib_mr.rkey)
463 			bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr,
464 					     true);
465 		if (mr->ib_mr.lkey)
466 			bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
467 		kfree(mr);
468 		fence->mr = NULL;
469 	}
470 	if (fence->dma_addr) {
471 		dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES,
472 				 DMA_BIDIRECTIONAL);
473 		fence->dma_addr = 0;
474 	}
475 }
476 
477 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
478 {
479 	int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND;
480 	struct bnxt_re_fence_data *fence = &pd->fence;
481 	struct bnxt_re_dev *rdev = pd->rdev;
482 	struct device *dev = &rdev->en_dev->pdev->dev;
483 	struct bnxt_re_mr *mr = NULL;
484 	dma_addr_t dma_addr = 0;
485 	struct ib_mw *mw;
486 	int rc;
487 
488 	dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES,
489 				  DMA_BIDIRECTIONAL);
490 	rc = dma_mapping_error(dev, dma_addr);
491 	if (rc) {
492 		ibdev_err(&rdev->ibdev, "Failed to dma-map fence-MR-mem\n");
493 		rc = -EIO;
494 		fence->dma_addr = 0;
495 		goto fail;
496 	}
497 	fence->dma_addr = dma_addr;
498 
499 	/* Allocate a MR */
500 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
501 	if (!mr) {
502 		rc = -ENOMEM;
503 		goto fail;
504 	}
505 	fence->mr = mr;
506 	mr->rdev = rdev;
507 	mr->qplib_mr.pd = &pd->qplib_pd;
508 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
509 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
510 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
511 	if (rc) {
512 		ibdev_err(&rdev->ibdev, "Failed to alloc fence-HW-MR\n");
513 		goto fail;
514 	}
515 
516 	/* Register MR */
517 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
518 	mr->qplib_mr.va = (u64)(unsigned long)fence->va;
519 	mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES;
520 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL,
521 			       BNXT_RE_FENCE_PBL_SIZE, PAGE_SIZE);
522 	if (rc) {
523 		ibdev_err(&rdev->ibdev, "Failed to register fence-MR\n");
524 		goto fail;
525 	}
526 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
527 
528 	/* Create a fence MW only for kernel consumers */
529 	mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL);
530 	if (IS_ERR(mw)) {
531 		ibdev_err(&rdev->ibdev,
532 			  "Failed to create fence-MW for PD: %p\n", pd);
533 		rc = PTR_ERR(mw);
534 		goto fail;
535 	}
536 	fence->mw = mw;
537 
538 	bnxt_re_create_fence_wqe(pd);
539 	return 0;
540 
541 fail:
542 	bnxt_re_destroy_fence_mr(pd);
543 	return rc;
544 }
545 
546 static struct bnxt_re_user_mmap_entry*
547 bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset,
548 			  enum bnxt_re_mmap_flag mmap_flag, u64 *offset)
549 {
550 	struct bnxt_re_user_mmap_entry *entry;
551 	int ret;
552 
553 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
554 	if (!entry)
555 		return NULL;
556 
557 	entry->mem_offset = mem_offset;
558 	entry->mmap_flag = mmap_flag;
559 	entry->uctx = uctx;
560 
561 	switch (mmap_flag) {
562 	case BNXT_RE_MMAP_SH_PAGE:
563 		ret = rdma_user_mmap_entry_insert_exact(&uctx->ib_uctx,
564 							&entry->rdma_entry, PAGE_SIZE, 0);
565 		break;
566 	case BNXT_RE_MMAP_UC_DB:
567 	case BNXT_RE_MMAP_WC_DB:
568 	case BNXT_RE_MMAP_DBR_BAR:
569 	case BNXT_RE_MMAP_DBR_PAGE:
570 		ret = rdma_user_mmap_entry_insert(&uctx->ib_uctx,
571 						  &entry->rdma_entry, PAGE_SIZE);
572 		break;
573 	default:
574 		ret = -EINVAL;
575 		break;
576 	}
577 
578 	if (ret) {
579 		kfree(entry);
580 		return NULL;
581 	}
582 	if (offset)
583 		*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
584 
585 	return entry;
586 }
587 
588 /* Protection Domains */
589 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
590 {
591 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
592 	struct bnxt_re_dev *rdev = pd->rdev;
593 
594 	if (udata) {
595 		rdma_user_mmap_entry_remove(pd->pd_db_mmap);
596 		pd->pd_db_mmap = NULL;
597 	}
598 
599 	bnxt_re_destroy_fence_mr(pd);
600 
601 	if (pd->qplib_pd.id) {
602 		if (!bnxt_qplib_dealloc_pd(&rdev->qplib_res,
603 					   &rdev->qplib_res.pd_tbl,
604 					   &pd->qplib_pd))
605 			atomic_dec(&rdev->stats.res.pd_count);
606 	}
607 	return 0;
608 }
609 
610 int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
611 {
612 	struct ib_device *ibdev = ibpd->device;
613 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
614 	struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context(
615 		udata, struct bnxt_re_ucontext, ib_uctx);
616 	struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd);
617 	struct bnxt_re_user_mmap_entry *entry = NULL;
618 	u32 active_pds;
619 	int rc = 0;
620 
621 	pd->rdev = rdev;
622 	if (bnxt_qplib_alloc_pd(&rdev->qplib_res, &pd->qplib_pd)) {
623 		ibdev_err(&rdev->ibdev, "Failed to allocate HW PD");
624 		rc = -ENOMEM;
625 		goto fail;
626 	}
627 
628 	if (udata) {
629 		struct bnxt_re_pd_resp resp = {};
630 
631 		if (!ucntx->dpi.dbr) {
632 			/* Allocate DPI in alloc_pd to avoid failing of
633 			 * ibv_devinfo and family of application when DPIs
634 			 * are depleted.
635 			 */
636 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res,
637 						 &ucntx->dpi, ucntx, BNXT_QPLIB_DPI_TYPE_UC)) {
638 				rc = -ENOMEM;
639 				goto dbfail;
640 			}
641 		}
642 
643 		resp.pdid = pd->qplib_pd.id;
644 		/* Still allow mapping this DBR to the new user PD. */
645 		resp.dpi = ucntx->dpi.dpi;
646 
647 		entry = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->dpi.umdbr,
648 						  BNXT_RE_MMAP_UC_DB, &resp.dbr);
649 
650 		if (!entry) {
651 			rc = -ENOMEM;
652 			goto dbfail;
653 		}
654 
655 		pd->pd_db_mmap = &entry->rdma_entry;
656 
657 		rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
658 		if (rc) {
659 			rdma_user_mmap_entry_remove(pd->pd_db_mmap);
660 			rc = -EFAULT;
661 			goto dbfail;
662 		}
663 	}
664 
665 	if (!udata)
666 		if (bnxt_re_create_fence_mr(pd))
667 			ibdev_warn(&rdev->ibdev,
668 				   "Failed to create Fence-MR\n");
669 	active_pds = atomic_inc_return(&rdev->stats.res.pd_count);
670 	if (active_pds > rdev->stats.res.pd_watermark)
671 		rdev->stats.res.pd_watermark = active_pds;
672 
673 	return 0;
674 dbfail:
675 	bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
676 			      &pd->qplib_pd);
677 fail:
678 	return rc;
679 }
680 
681 /* Address Handles */
682 int bnxt_re_destroy_ah(struct ib_ah *ib_ah, u32 flags)
683 {
684 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
685 	struct bnxt_re_dev *rdev = ah->rdev;
686 	bool block = true;
687 	int rc;
688 
689 	block = !(flags & RDMA_DESTROY_AH_SLEEPABLE);
690 	rc = bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah, block);
691 	if (BNXT_RE_CHECK_RC(rc)) {
692 		if (rc == -ETIMEDOUT)
693 			rc = 0;
694 		else
695 			goto fail;
696 	}
697 	atomic_dec(&rdev->stats.res.ah_count);
698 fail:
699 	return rc;
700 }
701 
702 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)
703 {
704 	u8 nw_type;
705 
706 	switch (ntype) {
707 	case RDMA_NETWORK_IPV4:
708 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4;
709 		break;
710 	case RDMA_NETWORK_IPV6:
711 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6;
712 		break;
713 	default:
714 		nw_type = CMDQ_CREATE_AH_TYPE_V1;
715 		break;
716 	}
717 	return nw_type;
718 }
719 
720 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr,
721 		      struct ib_udata *udata)
722 {
723 	struct ib_pd *ib_pd = ib_ah->pd;
724 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
725 	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
726 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
727 	struct bnxt_re_dev *rdev = pd->rdev;
728 	const struct ib_gid_attr *sgid_attr;
729 	struct bnxt_re_gid_ctx *ctx;
730 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
731 	u32 active_ahs;
732 	u8 nw_type;
733 	int rc;
734 
735 	if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) {
736 		ibdev_err(&rdev->ibdev, "Failed to alloc AH: GRH not set");
737 		return -EINVAL;
738 	}
739 
740 	ah->rdev = rdev;
741 	ah->qplib_ah.pd = &pd->qplib_pd;
742 
743 	/* Supply the configuration for the HW */
744 	memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw,
745 	       sizeof(union ib_gid));
746 	sgid_attr = grh->sgid_attr;
747 	/* Get the HW context of the GID. The reference
748 	 * of GID table entry is already taken by the caller.
749 	 */
750 	ctx = rdma_read_gid_hw_context(sgid_attr);
751 	ah->qplib_ah.sgid_index = ctx->idx;
752 	ah->qplib_ah.host_sgid_index = grh->sgid_index;
753 	ah->qplib_ah.traffic_class = grh->traffic_class;
754 	ah->qplib_ah.flow_label = grh->flow_label;
755 	ah->qplib_ah.hop_limit = grh->hop_limit;
756 	ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr);
757 
758 	/* Get network header type for this GID */
759 	nw_type = rdma_gid_attr_network_type(sgid_attr);
760 	ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type);
761 
762 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
763 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah,
764 				  !(init_attr->flags &
765 				    RDMA_CREATE_AH_SLEEPABLE));
766 	if (rc) {
767 		ibdev_err(&rdev->ibdev, "Failed to allocate HW AH");
768 		return rc;
769 	}
770 
771 	/* Write AVID to shared page. */
772 	if (udata) {
773 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
774 			udata, struct bnxt_re_ucontext, ib_uctx);
775 		unsigned long flag;
776 		u32 *wrptr;
777 
778 		spin_lock_irqsave(&uctx->sh_lock, flag);
779 		wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT);
780 		*wrptr = ah->qplib_ah.id;
781 		wmb(); /* make sure cache is updated. */
782 		spin_unlock_irqrestore(&uctx->sh_lock, flag);
783 	}
784 	active_ahs = atomic_inc_return(&rdev->stats.res.ah_count);
785 	if (active_ahs > rdev->stats.res.ah_watermark)
786 		rdev->stats.res.ah_watermark = active_ahs;
787 
788 	return 0;
789 }
790 
791 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
792 {
793 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
794 
795 	ah_attr->type = ib_ah->type;
796 	rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl);
797 	memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN);
798 	rdma_ah_set_grh(ah_attr, NULL, 0,
799 			ah->qplib_ah.host_sgid_index,
800 			0, ah->qplib_ah.traffic_class);
801 	rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data);
802 	rdma_ah_set_port_num(ah_attr, 1);
803 	rdma_ah_set_static_rate(ah_attr, 0);
804 	return 0;
805 }
806 
807 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp)
808 	__acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock)
809 {
810 	unsigned long flags;
811 
812 	spin_lock_irqsave(&qp->scq->cq_lock, flags);
813 	if (qp->rcq != qp->scq)
814 		spin_lock(&qp->rcq->cq_lock);
815 	else
816 		__acquire(&qp->rcq->cq_lock);
817 
818 	return flags;
819 }
820 
821 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp,
822 			unsigned long flags)
823 	__releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock)
824 {
825 	if (qp->rcq != qp->scq)
826 		spin_unlock(&qp->rcq->cq_lock);
827 	else
828 		__release(&qp->rcq->cq_lock);
829 	spin_unlock_irqrestore(&qp->scq->cq_lock, flags);
830 }
831 
832 static int bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp *qp)
833 {
834 	struct bnxt_re_qp *gsi_sqp;
835 	struct bnxt_re_ah *gsi_sah;
836 	struct bnxt_re_dev *rdev;
837 	int rc;
838 
839 	rdev = qp->rdev;
840 	gsi_sqp = rdev->gsi_ctx.gsi_sqp;
841 	gsi_sah = rdev->gsi_ctx.gsi_sah;
842 
843 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow AH\n");
844 	bnxt_qplib_destroy_ah(&rdev->qplib_res,
845 			      &gsi_sah->qplib_ah,
846 			      true);
847 	atomic_dec(&rdev->stats.res.ah_count);
848 	bnxt_qplib_clean_qp(&qp->qplib_qp);
849 
850 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow QP\n");
851 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &gsi_sqp->qplib_qp);
852 	if (rc) {
853 		ibdev_err(&rdev->ibdev, "Destroy Shadow QP failed");
854 		goto fail;
855 	}
856 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &gsi_sqp->qplib_qp);
857 
858 	/* remove from active qp list */
859 	mutex_lock(&rdev->qp_lock);
860 	list_del(&gsi_sqp->list);
861 	mutex_unlock(&rdev->qp_lock);
862 	atomic_dec(&rdev->stats.res.qp_count);
863 
864 	kfree(rdev->gsi_ctx.sqp_tbl);
865 	kfree(gsi_sah);
866 	kfree(gsi_sqp);
867 	rdev->gsi_ctx.gsi_sqp = NULL;
868 	rdev->gsi_ctx.gsi_sah = NULL;
869 	rdev->gsi_ctx.sqp_tbl = NULL;
870 
871 	return 0;
872 fail:
873 	return rc;
874 }
875 
876 /* Queue Pairs */
877 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
878 {
879 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
880 	struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
881 	struct bnxt_re_dev *rdev = qp->rdev;
882 	struct bnxt_qplib_nq *scq_nq = NULL;
883 	struct bnxt_qplib_nq *rcq_nq = NULL;
884 	unsigned int flags;
885 	int rc;
886 
887 	bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
888 
889 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
890 	if (rc) {
891 		ibdev_err(&rdev->ibdev, "Failed to destroy HW QP");
892 		return rc;
893 	}
894 
895 	if (rdma_is_kernel_res(&qp->ib_qp.res)) {
896 		flags = bnxt_re_lock_cqs(qp);
897 		bnxt_qplib_clean_qp(&qp->qplib_qp);
898 		bnxt_re_unlock_cqs(qp, flags);
899 	}
900 
901 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
902 
903 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) {
904 		rc = bnxt_re_destroy_gsi_sqp(qp);
905 		if (rc)
906 			return rc;
907 	}
908 
909 	mutex_lock(&rdev->qp_lock);
910 	list_del(&qp->list);
911 	mutex_unlock(&rdev->qp_lock);
912 	atomic_dec(&rdev->stats.res.qp_count);
913 
914 	ib_umem_release(qp->rumem);
915 	ib_umem_release(qp->sumem);
916 
917 	/* Flush all the entries of notification queue associated with
918 	 * given qp.
919 	 */
920 	scq_nq = qplib_qp->scq->nq;
921 	rcq_nq = qplib_qp->rcq->nq;
922 	bnxt_re_synchronize_nq(scq_nq);
923 	if (scq_nq != rcq_nq)
924 		bnxt_re_synchronize_nq(rcq_nq);
925 
926 	return 0;
927 }
928 
929 static u8 __from_ib_qp_type(enum ib_qp_type type)
930 {
931 	switch (type) {
932 	case IB_QPT_GSI:
933 		return CMDQ_CREATE_QP1_TYPE_GSI;
934 	case IB_QPT_RC:
935 		return CMDQ_CREATE_QP_TYPE_RC;
936 	case IB_QPT_UD:
937 		return CMDQ_CREATE_QP_TYPE_UD;
938 	default:
939 		return IB_QPT_MAX;
940 	}
941 }
942 
943 static u16 bnxt_re_setup_rwqe_size(struct bnxt_qplib_qp *qplqp,
944 				   int rsge, int max)
945 {
946 	if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
947 		rsge = max;
948 	return bnxt_re_get_rwqe_size(rsge);
949 }
950 
951 static u16 bnxt_re_get_wqe_size(int ilsize, int nsge)
952 {
953 	u16 wqe_size, calc_ils;
954 
955 	wqe_size = bnxt_re_get_swqe_size(nsge);
956 	if (ilsize) {
957 		calc_ils = sizeof(struct sq_send_hdr) + ilsize;
958 		wqe_size = max_t(u16, calc_ils, wqe_size);
959 		wqe_size = ALIGN(wqe_size, sizeof(struct sq_send_hdr));
960 	}
961 	return wqe_size;
962 }
963 
964 static int bnxt_re_setup_swqe_size(struct bnxt_re_qp *qp,
965 				   struct ib_qp_init_attr *init_attr)
966 {
967 	struct bnxt_qplib_dev_attr *dev_attr;
968 	struct bnxt_qplib_qp *qplqp;
969 	struct bnxt_re_dev *rdev;
970 	struct bnxt_qplib_q *sq;
971 	int align, ilsize;
972 
973 	rdev = qp->rdev;
974 	qplqp = &qp->qplib_qp;
975 	sq = &qplqp->sq;
976 	dev_attr = &rdev->dev_attr;
977 
978 	align = sizeof(struct sq_send_hdr);
979 	ilsize = ALIGN(init_attr->cap.max_inline_data, align);
980 
981 	sq->wqe_size = bnxt_re_get_wqe_size(ilsize, sq->max_sge);
982 	if (sq->wqe_size > bnxt_re_get_swqe_size(dev_attr->max_qp_sges))
983 		return -EINVAL;
984 	/* For gen p4 and gen p5 backward compatibility mode
985 	 * wqe size is fixed to 128 bytes
986 	 */
987 	if (sq->wqe_size < bnxt_re_get_swqe_size(dev_attr->max_qp_sges) &&
988 			qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
989 		sq->wqe_size = bnxt_re_get_swqe_size(dev_attr->max_qp_sges);
990 
991 	if (init_attr->cap.max_inline_data) {
992 		qplqp->max_inline_data = sq->wqe_size -
993 			sizeof(struct sq_send_hdr);
994 		init_attr->cap.max_inline_data = qplqp->max_inline_data;
995 		if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
996 			sq->max_sge = qplqp->max_inline_data /
997 				sizeof(struct sq_sge);
998 	}
999 
1000 	return 0;
1001 }
1002 
1003 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
1004 				struct bnxt_re_qp *qp, struct ib_udata *udata)
1005 {
1006 	struct bnxt_qplib_qp *qplib_qp;
1007 	struct bnxt_re_ucontext *cntx;
1008 	struct bnxt_re_qp_req ureq;
1009 	int bytes = 0, psn_sz;
1010 	struct ib_umem *umem;
1011 	int psn_nume;
1012 
1013 	qplib_qp = &qp->qplib_qp;
1014 	cntx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext,
1015 					 ib_uctx);
1016 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1017 		return -EFAULT;
1018 
1019 	bytes = (qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size);
1020 	/* Consider mapping PSN search memory only for RC QPs. */
1021 	if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) {
1022 		psn_sz = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
1023 						   sizeof(struct sq_psn_search_ext) :
1024 						   sizeof(struct sq_psn_search);
1025 		psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ?
1026 			    qplib_qp->sq.max_wqe :
1027 			    ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) /
1028 			      sizeof(struct bnxt_qplib_sge));
1029 		bytes += (psn_nume * psn_sz);
1030 	}
1031 
1032 	bytes = PAGE_ALIGN(bytes);
1033 	umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes,
1034 			   IB_ACCESS_LOCAL_WRITE);
1035 	if (IS_ERR(umem))
1036 		return PTR_ERR(umem);
1037 
1038 	qp->sumem = umem;
1039 	qplib_qp->sq.sg_info.umem = umem;
1040 	qplib_qp->sq.sg_info.pgsize = PAGE_SIZE;
1041 	qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT;
1042 	qplib_qp->qp_handle = ureq.qp_handle;
1043 
1044 	if (!qp->qplib_qp.srq) {
1045 		bytes = (qplib_qp->rq.max_wqe * qplib_qp->rq.wqe_size);
1046 		bytes = PAGE_ALIGN(bytes);
1047 		umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes,
1048 				   IB_ACCESS_LOCAL_WRITE);
1049 		if (IS_ERR(umem))
1050 			goto rqfail;
1051 		qp->rumem = umem;
1052 		qplib_qp->rq.sg_info.umem = umem;
1053 		qplib_qp->rq.sg_info.pgsize = PAGE_SIZE;
1054 		qplib_qp->rq.sg_info.pgshft = PAGE_SHIFT;
1055 	}
1056 
1057 	qplib_qp->dpi = &cntx->dpi;
1058 	return 0;
1059 rqfail:
1060 	ib_umem_release(qp->sumem);
1061 	qp->sumem = NULL;
1062 	memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info));
1063 
1064 	return PTR_ERR(umem);
1065 }
1066 
1067 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
1068 				(struct bnxt_re_pd *pd,
1069 				 struct bnxt_qplib_res *qp1_res,
1070 				 struct bnxt_qplib_qp *qp1_qp)
1071 {
1072 	struct bnxt_re_dev *rdev = pd->rdev;
1073 	struct bnxt_re_ah *ah;
1074 	union ib_gid sgid;
1075 	int rc;
1076 
1077 	ah = kzalloc(sizeof(*ah), GFP_KERNEL);
1078 	if (!ah)
1079 		return NULL;
1080 
1081 	ah->rdev = rdev;
1082 	ah->qplib_ah.pd = &pd->qplib_pd;
1083 
1084 	rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid);
1085 	if (rc)
1086 		goto fail;
1087 
1088 	/* supply the dgid data same as sgid */
1089 	memcpy(ah->qplib_ah.dgid.data, &sgid.raw,
1090 	       sizeof(union ib_gid));
1091 	ah->qplib_ah.sgid_index = 0;
1092 
1093 	ah->qplib_ah.traffic_class = 0;
1094 	ah->qplib_ah.flow_label = 0;
1095 	ah->qplib_ah.hop_limit = 1;
1096 	ah->qplib_ah.sl = 0;
1097 	/* Have DMAC same as SMAC */
1098 	ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr);
1099 
1100 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false);
1101 	if (rc) {
1102 		ibdev_err(&rdev->ibdev,
1103 			  "Failed to allocate HW AH for Shadow QP");
1104 		goto fail;
1105 	}
1106 	atomic_inc(&rdev->stats.res.ah_count);
1107 
1108 	return ah;
1109 
1110 fail:
1111 	kfree(ah);
1112 	return NULL;
1113 }
1114 
1115 static struct bnxt_re_qp *bnxt_re_create_shadow_qp
1116 				(struct bnxt_re_pd *pd,
1117 				 struct bnxt_qplib_res *qp1_res,
1118 				 struct bnxt_qplib_qp *qp1_qp)
1119 {
1120 	struct bnxt_re_dev *rdev = pd->rdev;
1121 	struct bnxt_re_qp *qp;
1122 	int rc;
1123 
1124 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1125 	if (!qp)
1126 		return NULL;
1127 
1128 	qp->rdev = rdev;
1129 
1130 	/* Initialize the shadow QP structure from the QP1 values */
1131 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
1132 
1133 	qp->qplib_qp.pd = &pd->qplib_pd;
1134 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
1135 	qp->qplib_qp.type = IB_QPT_UD;
1136 
1137 	qp->qplib_qp.max_inline_data = 0;
1138 	qp->qplib_qp.sig_type = true;
1139 
1140 	/* Shadow QP SQ depth should be same as QP1 RQ depth */
1141 	qp->qplib_qp.sq.wqe_size = bnxt_re_get_wqe_size(0, 6);
1142 	qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
1143 	qp->qplib_qp.sq.max_sge = 2;
1144 	/* Q full delta can be 1 since it is internal QP */
1145 	qp->qplib_qp.sq.q_full_delta = 1;
1146 	qp->qplib_qp.sq.sg_info.pgsize = PAGE_SIZE;
1147 	qp->qplib_qp.sq.sg_info.pgshft = PAGE_SHIFT;
1148 
1149 	qp->qplib_qp.scq = qp1_qp->scq;
1150 	qp->qplib_qp.rcq = qp1_qp->rcq;
1151 
1152 	qp->qplib_qp.rq.wqe_size = bnxt_re_get_rwqe_size(6);
1153 	qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
1154 	qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
1155 	/* Q full delta can be 1 since it is internal QP */
1156 	qp->qplib_qp.rq.q_full_delta = 1;
1157 	qp->qplib_qp.rq.sg_info.pgsize = PAGE_SIZE;
1158 	qp->qplib_qp.rq.sg_info.pgshft = PAGE_SHIFT;
1159 
1160 	qp->qplib_qp.mtu = qp1_qp->mtu;
1161 
1162 	qp->qplib_qp.sq_hdr_buf_size = 0;
1163 	qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
1164 	qp->qplib_qp.dpi = &rdev->dpi_privileged;
1165 
1166 	rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
1167 	if (rc)
1168 		goto fail;
1169 
1170 	spin_lock_init(&qp->sq_lock);
1171 	INIT_LIST_HEAD(&qp->list);
1172 	mutex_lock(&rdev->qp_lock);
1173 	list_add_tail(&qp->list, &rdev->qp_list);
1174 	atomic_inc(&rdev->stats.res.qp_count);
1175 	mutex_unlock(&rdev->qp_lock);
1176 	return qp;
1177 fail:
1178 	kfree(qp);
1179 	return NULL;
1180 }
1181 
1182 static int bnxt_re_init_rq_attr(struct bnxt_re_qp *qp,
1183 				struct ib_qp_init_attr *init_attr)
1184 {
1185 	struct bnxt_qplib_dev_attr *dev_attr;
1186 	struct bnxt_qplib_qp *qplqp;
1187 	struct bnxt_re_dev *rdev;
1188 	struct bnxt_qplib_q *rq;
1189 	int entries;
1190 
1191 	rdev = qp->rdev;
1192 	qplqp = &qp->qplib_qp;
1193 	rq = &qplqp->rq;
1194 	dev_attr = &rdev->dev_attr;
1195 
1196 	if (init_attr->srq) {
1197 		struct bnxt_re_srq *srq;
1198 
1199 		srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq);
1200 		qplqp->srq = &srq->qplib_srq;
1201 		rq->max_wqe = 0;
1202 	} else {
1203 		rq->max_sge = init_attr->cap.max_recv_sge;
1204 		if (rq->max_sge > dev_attr->max_qp_sges)
1205 			rq->max_sge = dev_attr->max_qp_sges;
1206 		init_attr->cap.max_recv_sge = rq->max_sge;
1207 		rq->wqe_size = bnxt_re_setup_rwqe_size(qplqp, rq->max_sge,
1208 						       dev_attr->max_qp_sges);
1209 		/* Allocate 1 more than what's provided so posting max doesn't
1210 		 * mean empty.
1211 		 */
1212 		entries = roundup_pow_of_two(init_attr->cap.max_recv_wr + 1);
1213 		rq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1214 		rq->q_full_delta = 0;
1215 		rq->sg_info.pgsize = PAGE_SIZE;
1216 		rq->sg_info.pgshft = PAGE_SHIFT;
1217 	}
1218 
1219 	return 0;
1220 }
1221 
1222 static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp)
1223 {
1224 	struct bnxt_qplib_dev_attr *dev_attr;
1225 	struct bnxt_qplib_qp *qplqp;
1226 	struct bnxt_re_dev *rdev;
1227 
1228 	rdev = qp->rdev;
1229 	qplqp = &qp->qplib_qp;
1230 	dev_attr = &rdev->dev_attr;
1231 
1232 	if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) {
1233 		qplqp->rq.max_sge = dev_attr->max_qp_sges;
1234 		if (qplqp->rq.max_sge > dev_attr->max_qp_sges)
1235 			qplqp->rq.max_sge = dev_attr->max_qp_sges;
1236 		qplqp->rq.max_sge = 6;
1237 	}
1238 }
1239 
1240 static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp,
1241 				struct ib_qp_init_attr *init_attr,
1242 				struct ib_udata *udata)
1243 {
1244 	struct bnxt_qplib_dev_attr *dev_attr;
1245 	struct bnxt_qplib_qp *qplqp;
1246 	struct bnxt_re_dev *rdev;
1247 	struct bnxt_qplib_q *sq;
1248 	int entries;
1249 	int diff;
1250 	int rc;
1251 
1252 	rdev = qp->rdev;
1253 	qplqp = &qp->qplib_qp;
1254 	sq = &qplqp->sq;
1255 	dev_attr = &rdev->dev_attr;
1256 
1257 	sq->max_sge = init_attr->cap.max_send_sge;
1258 	if (sq->max_sge > dev_attr->max_qp_sges) {
1259 		sq->max_sge = dev_attr->max_qp_sges;
1260 		init_attr->cap.max_send_sge = sq->max_sge;
1261 	}
1262 
1263 	rc = bnxt_re_setup_swqe_size(qp, init_attr);
1264 	if (rc)
1265 		return rc;
1266 
1267 	entries = init_attr->cap.max_send_wr;
1268 	/* Allocate 128 + 1 more than what's provided */
1269 	diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ?
1270 		0 : BNXT_QPLIB_RESERVED_QP_WRS;
1271 	entries = roundup_pow_of_two(entries + diff + 1);
1272 	sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1);
1273 	sq->q_full_delta = diff + 1;
1274 	/*
1275 	 * Reserving one slot for Phantom WQE. Application can
1276 	 * post one extra entry in this case. But allowing this to avoid
1277 	 * unexpected Queue full condition
1278 	 */
1279 	qplqp->sq.q_full_delta -= 1;
1280 	qplqp->sq.sg_info.pgsize = PAGE_SIZE;
1281 	qplqp->sq.sg_info.pgshft = PAGE_SHIFT;
1282 
1283 	return 0;
1284 }
1285 
1286 static void bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp *qp,
1287 				       struct ib_qp_init_attr *init_attr)
1288 {
1289 	struct bnxt_qplib_dev_attr *dev_attr;
1290 	struct bnxt_qplib_qp *qplqp;
1291 	struct bnxt_re_dev *rdev;
1292 	int entries;
1293 
1294 	rdev = qp->rdev;
1295 	qplqp = &qp->qplib_qp;
1296 	dev_attr = &rdev->dev_attr;
1297 
1298 	if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) {
1299 		entries = roundup_pow_of_two(init_attr->cap.max_send_wr + 1);
1300 		qplqp->sq.max_wqe = min_t(u32, entries,
1301 					  dev_attr->max_qp_wqes + 1);
1302 		qplqp->sq.q_full_delta = qplqp->sq.max_wqe -
1303 			init_attr->cap.max_send_wr;
1304 		qplqp->sq.max_sge++; /* Need one extra sge to put UD header */
1305 		if (qplqp->sq.max_sge > dev_attr->max_qp_sges)
1306 			qplqp->sq.max_sge = dev_attr->max_qp_sges;
1307 	}
1308 }
1309 
1310 static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev,
1311 				struct ib_qp_init_attr *init_attr)
1312 {
1313 	struct bnxt_qplib_chip_ctx *chip_ctx;
1314 	int qptype;
1315 
1316 	chip_ctx = rdev->chip_ctx;
1317 
1318 	qptype = __from_ib_qp_type(init_attr->qp_type);
1319 	if (qptype == IB_QPT_MAX) {
1320 		ibdev_err(&rdev->ibdev, "QP type 0x%x not supported", qptype);
1321 		qptype = -EOPNOTSUPP;
1322 		goto out;
1323 	}
1324 
1325 	if (bnxt_qplib_is_chip_gen_p5(chip_ctx) &&
1326 	    init_attr->qp_type == IB_QPT_GSI)
1327 		qptype = CMDQ_CREATE_QP_TYPE_GSI;
1328 out:
1329 	return qptype;
1330 }
1331 
1332 static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1333 				struct ib_qp_init_attr *init_attr,
1334 				struct ib_udata *udata)
1335 {
1336 	struct bnxt_qplib_dev_attr *dev_attr;
1337 	struct bnxt_qplib_qp *qplqp;
1338 	struct bnxt_re_dev *rdev;
1339 	struct bnxt_re_cq *cq;
1340 	int rc = 0, qptype;
1341 
1342 	rdev = qp->rdev;
1343 	qplqp = &qp->qplib_qp;
1344 	dev_attr = &rdev->dev_attr;
1345 
1346 	/* Setup misc params */
1347 	ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr);
1348 	qplqp->pd = &pd->qplib_pd;
1349 	qplqp->qp_handle = (u64)qplqp;
1350 	qplqp->max_inline_data = init_attr->cap.max_inline_data;
1351 	qplqp->sig_type = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1352 	qptype = bnxt_re_init_qp_type(rdev, init_attr);
1353 	if (qptype < 0) {
1354 		rc = qptype;
1355 		goto out;
1356 	}
1357 	qplqp->type = (u8)qptype;
1358 	qplqp->wqe_mode = rdev->chip_ctx->modes.wqe_mode;
1359 
1360 	if (init_attr->qp_type == IB_QPT_RC) {
1361 		qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom;
1362 		qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
1363 	}
1364 	qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1365 	qplqp->dpi = &rdev->dpi_privileged; /* Doorbell page */
1366 	if (init_attr->create_flags) {
1367 		ibdev_dbg(&rdev->ibdev,
1368 			  "QP create flags 0x%x not supported",
1369 			  init_attr->create_flags);
1370 		return -EOPNOTSUPP;
1371 	}
1372 
1373 	/* Setup CQs */
1374 	if (init_attr->send_cq) {
1375 		cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq);
1376 		qplqp->scq = &cq->qplib_cq;
1377 		qp->scq = cq;
1378 	}
1379 
1380 	if (init_attr->recv_cq) {
1381 		cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq);
1382 		qplqp->rcq = &cq->qplib_cq;
1383 		qp->rcq = cq;
1384 	}
1385 
1386 	/* Setup RQ/SRQ */
1387 	rc = bnxt_re_init_rq_attr(qp, init_attr);
1388 	if (rc)
1389 		goto out;
1390 	if (init_attr->qp_type == IB_QPT_GSI)
1391 		bnxt_re_adjust_gsi_rq_attr(qp);
1392 
1393 	/* Setup SQ */
1394 	rc = bnxt_re_init_sq_attr(qp, init_attr, udata);
1395 	if (rc)
1396 		goto out;
1397 	if (init_attr->qp_type == IB_QPT_GSI)
1398 		bnxt_re_adjust_gsi_sq_attr(qp, init_attr);
1399 
1400 	if (udata) /* This will update DPI and qp_handle */
1401 		rc = bnxt_re_init_user_qp(rdev, pd, qp, udata);
1402 out:
1403 	return rc;
1404 }
1405 
1406 static int bnxt_re_create_shadow_gsi(struct bnxt_re_qp *qp,
1407 				     struct bnxt_re_pd *pd)
1408 {
1409 	struct bnxt_re_sqp_entries *sqp_tbl;
1410 	struct bnxt_re_dev *rdev;
1411 	struct bnxt_re_qp *sqp;
1412 	struct bnxt_re_ah *sah;
1413 	int rc = 0;
1414 
1415 	rdev = qp->rdev;
1416 	/* Create a shadow QP to handle the QP1 traffic */
1417 	sqp_tbl = kcalloc(BNXT_RE_MAX_GSI_SQP_ENTRIES, sizeof(*sqp_tbl),
1418 			  GFP_KERNEL);
1419 	if (!sqp_tbl)
1420 		return -ENOMEM;
1421 	rdev->gsi_ctx.sqp_tbl = sqp_tbl;
1422 
1423 	sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res, &qp->qplib_qp);
1424 	if (!sqp) {
1425 		rc = -ENODEV;
1426 		ibdev_err(&rdev->ibdev, "Failed to create Shadow QP for QP1");
1427 		goto out;
1428 	}
1429 	rdev->gsi_ctx.gsi_sqp = sqp;
1430 
1431 	sqp->rcq = qp->rcq;
1432 	sqp->scq = qp->scq;
1433 	sah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res,
1434 					  &qp->qplib_qp);
1435 	if (!sah) {
1436 		bnxt_qplib_destroy_qp(&rdev->qplib_res,
1437 				      &sqp->qplib_qp);
1438 		rc = -ENODEV;
1439 		ibdev_err(&rdev->ibdev,
1440 			  "Failed to create AH entry for ShadowQP");
1441 		goto out;
1442 	}
1443 	rdev->gsi_ctx.gsi_sah = sah;
1444 
1445 	return 0;
1446 out:
1447 	kfree(sqp_tbl);
1448 	return rc;
1449 }
1450 
1451 static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1452 				 struct ib_qp_init_attr *init_attr)
1453 {
1454 	struct bnxt_re_dev *rdev;
1455 	struct bnxt_qplib_qp *qplqp;
1456 	int rc;
1457 
1458 	rdev = qp->rdev;
1459 	qplqp = &qp->qplib_qp;
1460 
1461 	qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
1462 	qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
1463 
1464 	rc = bnxt_qplib_create_qp1(&rdev->qplib_res, qplqp);
1465 	if (rc) {
1466 		ibdev_err(&rdev->ibdev, "create HW QP1 failed!");
1467 		goto out;
1468 	}
1469 
1470 	rc = bnxt_re_create_shadow_gsi(qp, pd);
1471 out:
1472 	return rc;
1473 }
1474 
1475 static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev,
1476 				   struct ib_qp_init_attr *init_attr,
1477 				   struct bnxt_qplib_dev_attr *dev_attr)
1478 {
1479 	bool rc = true;
1480 
1481 	if (init_attr->cap.max_send_wr > dev_attr->max_qp_wqes ||
1482 	    init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes ||
1483 	    init_attr->cap.max_send_sge > dev_attr->max_qp_sges ||
1484 	    init_attr->cap.max_recv_sge > dev_attr->max_qp_sges ||
1485 	    init_attr->cap.max_inline_data > dev_attr->max_inline_data) {
1486 		ibdev_err(&rdev->ibdev,
1487 			  "Create QP failed - max exceeded! 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x",
1488 			  init_attr->cap.max_send_wr, dev_attr->max_qp_wqes,
1489 			  init_attr->cap.max_recv_wr, dev_attr->max_qp_wqes,
1490 			  init_attr->cap.max_send_sge, dev_attr->max_qp_sges,
1491 			  init_attr->cap.max_recv_sge, dev_attr->max_qp_sges,
1492 			  init_attr->cap.max_inline_data,
1493 			  dev_attr->max_inline_data);
1494 		rc = false;
1495 	}
1496 	return rc;
1497 }
1498 
1499 int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr,
1500 		      struct ib_udata *udata)
1501 {
1502 	struct ib_pd *ib_pd = ib_qp->pd;
1503 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1504 	struct bnxt_re_dev *rdev = pd->rdev;
1505 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1506 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1507 	u32 active_qps;
1508 	int rc;
1509 
1510 	rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr);
1511 	if (!rc) {
1512 		rc = -EINVAL;
1513 		goto fail;
1514 	}
1515 
1516 	qp->rdev = rdev;
1517 	rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, udata);
1518 	if (rc)
1519 		goto fail;
1520 
1521 	if (qp_init_attr->qp_type == IB_QPT_GSI &&
1522 	    !(bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx))) {
1523 		rc = bnxt_re_create_gsi_qp(qp, pd, qp_init_attr);
1524 		if (rc == -ENODEV)
1525 			goto qp_destroy;
1526 		if (rc)
1527 			goto fail;
1528 	} else {
1529 		rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
1530 		if (rc) {
1531 			ibdev_err(&rdev->ibdev, "Failed to create HW QP");
1532 			goto free_umem;
1533 		}
1534 		if (udata) {
1535 			struct bnxt_re_qp_resp resp;
1536 
1537 			resp.qpid = qp->qplib_qp.id;
1538 			resp.rsvd = 0;
1539 			rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1540 			if (rc) {
1541 				ibdev_err(&rdev->ibdev, "Failed to copy QP udata");
1542 				goto qp_destroy;
1543 			}
1544 		}
1545 	}
1546 
1547 	qp->ib_qp.qp_num = qp->qplib_qp.id;
1548 	if (qp_init_attr->qp_type == IB_QPT_GSI)
1549 		rdev->gsi_ctx.gsi_qp = qp;
1550 	spin_lock_init(&qp->sq_lock);
1551 	spin_lock_init(&qp->rq_lock);
1552 	INIT_LIST_HEAD(&qp->list);
1553 	mutex_lock(&rdev->qp_lock);
1554 	list_add_tail(&qp->list, &rdev->qp_list);
1555 	mutex_unlock(&rdev->qp_lock);
1556 	active_qps = atomic_inc_return(&rdev->stats.res.qp_count);
1557 	if (active_qps > rdev->stats.res.qp_watermark)
1558 		rdev->stats.res.qp_watermark = active_qps;
1559 	if (qp_init_attr->qp_type == IB_QPT_RC) {
1560 		active_qps = atomic_inc_return(&rdev->stats.res.rc_qp_count);
1561 		if (active_qps > rdev->stats.res.rc_qp_watermark)
1562 			rdev->stats.res.rc_qp_watermark = active_qps;
1563 	} else if (qp_init_attr->qp_type == IB_QPT_UD) {
1564 		active_qps = atomic_inc_return(&rdev->stats.res.ud_qp_count);
1565 		if (active_qps > rdev->stats.res.ud_qp_watermark)
1566 			rdev->stats.res.ud_qp_watermark = active_qps;
1567 	}
1568 
1569 	return 0;
1570 qp_destroy:
1571 	bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
1572 free_umem:
1573 	ib_umem_release(qp->rumem);
1574 	ib_umem_release(qp->sumem);
1575 fail:
1576 	return rc;
1577 }
1578 
1579 static u8 __from_ib_qp_state(enum ib_qp_state state)
1580 {
1581 	switch (state) {
1582 	case IB_QPS_RESET:
1583 		return CMDQ_MODIFY_QP_NEW_STATE_RESET;
1584 	case IB_QPS_INIT:
1585 		return CMDQ_MODIFY_QP_NEW_STATE_INIT;
1586 	case IB_QPS_RTR:
1587 		return CMDQ_MODIFY_QP_NEW_STATE_RTR;
1588 	case IB_QPS_RTS:
1589 		return CMDQ_MODIFY_QP_NEW_STATE_RTS;
1590 	case IB_QPS_SQD:
1591 		return CMDQ_MODIFY_QP_NEW_STATE_SQD;
1592 	case IB_QPS_SQE:
1593 		return CMDQ_MODIFY_QP_NEW_STATE_SQE;
1594 	case IB_QPS_ERR:
1595 	default:
1596 		return CMDQ_MODIFY_QP_NEW_STATE_ERR;
1597 	}
1598 }
1599 
1600 static enum ib_qp_state __to_ib_qp_state(u8 state)
1601 {
1602 	switch (state) {
1603 	case CMDQ_MODIFY_QP_NEW_STATE_RESET:
1604 		return IB_QPS_RESET;
1605 	case CMDQ_MODIFY_QP_NEW_STATE_INIT:
1606 		return IB_QPS_INIT;
1607 	case CMDQ_MODIFY_QP_NEW_STATE_RTR:
1608 		return IB_QPS_RTR;
1609 	case CMDQ_MODIFY_QP_NEW_STATE_RTS:
1610 		return IB_QPS_RTS;
1611 	case CMDQ_MODIFY_QP_NEW_STATE_SQD:
1612 		return IB_QPS_SQD;
1613 	case CMDQ_MODIFY_QP_NEW_STATE_SQE:
1614 		return IB_QPS_SQE;
1615 	case CMDQ_MODIFY_QP_NEW_STATE_ERR:
1616 	default:
1617 		return IB_QPS_ERR;
1618 	}
1619 }
1620 
1621 static u32 __from_ib_mtu(enum ib_mtu mtu)
1622 {
1623 	switch (mtu) {
1624 	case IB_MTU_256:
1625 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_256;
1626 	case IB_MTU_512:
1627 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_512;
1628 	case IB_MTU_1024:
1629 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024;
1630 	case IB_MTU_2048:
1631 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1632 	case IB_MTU_4096:
1633 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096;
1634 	default:
1635 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1636 	}
1637 }
1638 
1639 static enum ib_mtu __to_ib_mtu(u32 mtu)
1640 {
1641 	switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) {
1642 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_256:
1643 		return IB_MTU_256;
1644 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_512:
1645 		return IB_MTU_512;
1646 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024:
1647 		return IB_MTU_1024;
1648 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048:
1649 		return IB_MTU_2048;
1650 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096:
1651 		return IB_MTU_4096;
1652 	default:
1653 		return IB_MTU_2048;
1654 	}
1655 }
1656 
1657 /* Shared Receive Queues */
1658 int bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata)
1659 {
1660 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1661 					       ib_srq);
1662 	struct bnxt_re_dev *rdev = srq->rdev;
1663 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1664 	struct bnxt_qplib_nq *nq = NULL;
1665 
1666 	if (qplib_srq->cq)
1667 		nq = qplib_srq->cq->nq;
1668 	bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
1669 	ib_umem_release(srq->umem);
1670 	atomic_dec(&rdev->stats.res.srq_count);
1671 	if (nq)
1672 		nq->budget--;
1673 	return 0;
1674 }
1675 
1676 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
1677 				 struct bnxt_re_pd *pd,
1678 				 struct bnxt_re_srq *srq,
1679 				 struct ib_udata *udata)
1680 {
1681 	struct bnxt_re_srq_req ureq;
1682 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1683 	struct ib_umem *umem;
1684 	int bytes = 0;
1685 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
1686 		udata, struct bnxt_re_ucontext, ib_uctx);
1687 
1688 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1689 		return -EFAULT;
1690 
1691 	bytes = (qplib_srq->max_wqe * qplib_srq->wqe_size);
1692 	bytes = PAGE_ALIGN(bytes);
1693 	umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes,
1694 			   IB_ACCESS_LOCAL_WRITE);
1695 	if (IS_ERR(umem))
1696 		return PTR_ERR(umem);
1697 
1698 	srq->umem = umem;
1699 	qplib_srq->sg_info.umem = umem;
1700 	qplib_srq->sg_info.pgsize = PAGE_SIZE;
1701 	qplib_srq->sg_info.pgshft = PAGE_SHIFT;
1702 	qplib_srq->srq_handle = ureq.srq_handle;
1703 	qplib_srq->dpi = &cntx->dpi;
1704 
1705 	return 0;
1706 }
1707 
1708 int bnxt_re_create_srq(struct ib_srq *ib_srq,
1709 		       struct ib_srq_init_attr *srq_init_attr,
1710 		       struct ib_udata *udata)
1711 {
1712 	struct bnxt_qplib_dev_attr *dev_attr;
1713 	struct bnxt_qplib_nq *nq = NULL;
1714 	struct bnxt_re_dev *rdev;
1715 	struct bnxt_re_srq *srq;
1716 	struct bnxt_re_pd *pd;
1717 	struct ib_pd *ib_pd;
1718 	u32 active_srqs;
1719 	int rc, entries;
1720 
1721 	ib_pd = ib_srq->pd;
1722 	pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1723 	rdev = pd->rdev;
1724 	dev_attr = &rdev->dev_attr;
1725 	srq = container_of(ib_srq, struct bnxt_re_srq, ib_srq);
1726 
1727 	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
1728 		ibdev_err(&rdev->ibdev, "Create CQ failed - max exceeded");
1729 		rc = -EINVAL;
1730 		goto exit;
1731 	}
1732 
1733 	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
1734 		rc = -EOPNOTSUPP;
1735 		goto exit;
1736 	}
1737 
1738 	srq->rdev = rdev;
1739 	srq->qplib_srq.pd = &pd->qplib_pd;
1740 	srq->qplib_srq.dpi = &rdev->dpi_privileged;
1741 	/* Allocate 1 more than what's provided so posting max doesn't
1742 	 * mean empty
1743 	 */
1744 	entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
1745 	if (entries > dev_attr->max_srq_wqes + 1)
1746 		entries = dev_attr->max_srq_wqes + 1;
1747 	srq->qplib_srq.max_wqe = entries;
1748 
1749 	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
1750 	 /* 128 byte wqe size for SRQ . So use max sges */
1751 	srq->qplib_srq.wqe_size = bnxt_re_get_rwqe_size(dev_attr->max_srq_sges);
1752 	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
1753 	srq->srq_limit = srq_init_attr->attr.srq_limit;
1754 	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
1755 	nq = &rdev->nq[0];
1756 
1757 	if (udata) {
1758 		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
1759 		if (rc)
1760 			goto fail;
1761 	}
1762 
1763 	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
1764 	if (rc) {
1765 		ibdev_err(&rdev->ibdev, "Create HW SRQ failed!");
1766 		goto fail;
1767 	}
1768 
1769 	if (udata) {
1770 		struct bnxt_re_srq_resp resp;
1771 
1772 		resp.srqid = srq->qplib_srq.id;
1773 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1774 		if (rc) {
1775 			ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!");
1776 			bnxt_qplib_destroy_srq(&rdev->qplib_res,
1777 					       &srq->qplib_srq);
1778 			goto fail;
1779 		}
1780 	}
1781 	if (nq)
1782 		nq->budget++;
1783 	active_srqs = atomic_inc_return(&rdev->stats.res.srq_count);
1784 	if (active_srqs > rdev->stats.res.srq_watermark)
1785 		rdev->stats.res.srq_watermark = active_srqs;
1786 	spin_lock_init(&srq->lock);
1787 
1788 	return 0;
1789 
1790 fail:
1791 	ib_umem_release(srq->umem);
1792 exit:
1793 	return rc;
1794 }
1795 
1796 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
1797 		       enum ib_srq_attr_mask srq_attr_mask,
1798 		       struct ib_udata *udata)
1799 {
1800 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1801 					       ib_srq);
1802 	struct bnxt_re_dev *rdev = srq->rdev;
1803 	int rc;
1804 
1805 	switch (srq_attr_mask) {
1806 	case IB_SRQ_MAX_WR:
1807 		/* SRQ resize is not supported */
1808 		break;
1809 	case IB_SRQ_LIMIT:
1810 		/* Change the SRQ threshold */
1811 		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
1812 			return -EINVAL;
1813 
1814 		srq->qplib_srq.threshold = srq_attr->srq_limit;
1815 		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
1816 		if (rc) {
1817 			ibdev_err(&rdev->ibdev, "Modify HW SRQ failed!");
1818 			return rc;
1819 		}
1820 		/* On success, update the shadow */
1821 		srq->srq_limit = srq_attr->srq_limit;
1822 		/* No need to Build and send response back to udata */
1823 		break;
1824 	default:
1825 		ibdev_err(&rdev->ibdev,
1826 			  "Unsupported srq_attr_mask 0x%x", srq_attr_mask);
1827 		return -EINVAL;
1828 	}
1829 	return 0;
1830 }
1831 
1832 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
1833 {
1834 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1835 					       ib_srq);
1836 	struct bnxt_re_srq tsrq;
1837 	struct bnxt_re_dev *rdev = srq->rdev;
1838 	int rc;
1839 
1840 	/* Get live SRQ attr */
1841 	tsrq.qplib_srq.id = srq->qplib_srq.id;
1842 	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
1843 	if (rc) {
1844 		ibdev_err(&rdev->ibdev, "Query HW SRQ failed!");
1845 		return rc;
1846 	}
1847 	srq_attr->max_wr = srq->qplib_srq.max_wqe;
1848 	srq_attr->max_sge = srq->qplib_srq.max_sge;
1849 	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
1850 
1851 	return 0;
1852 }
1853 
1854 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr,
1855 			  const struct ib_recv_wr **bad_wr)
1856 {
1857 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1858 					       ib_srq);
1859 	struct bnxt_qplib_swqe wqe;
1860 	unsigned long flags;
1861 	int rc = 0;
1862 
1863 	spin_lock_irqsave(&srq->lock, flags);
1864 	while (wr) {
1865 		/* Transcribe each ib_recv_wr to qplib_swqe */
1866 		wqe.num_sge = wr->num_sge;
1867 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
1868 		wqe.wr_id = wr->wr_id;
1869 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
1870 
1871 		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
1872 		if (rc) {
1873 			*bad_wr = wr;
1874 			break;
1875 		}
1876 		wr = wr->next;
1877 	}
1878 	spin_unlock_irqrestore(&srq->lock, flags);
1879 
1880 	return rc;
1881 }
1882 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
1883 				    struct bnxt_re_qp *qp1_qp,
1884 				    int qp_attr_mask)
1885 {
1886 	struct bnxt_re_qp *qp = rdev->gsi_ctx.gsi_sqp;
1887 	int rc;
1888 
1889 	if (qp_attr_mask & IB_QP_STATE) {
1890 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1891 		qp->qplib_qp.state = qp1_qp->qplib_qp.state;
1892 	}
1893 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1894 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1895 		qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index;
1896 	}
1897 
1898 	if (qp_attr_mask & IB_QP_QKEY) {
1899 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1900 		/* Using a Random  QKEY */
1901 		qp->qplib_qp.qkey = 0x81818181;
1902 	}
1903 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1904 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1905 		qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn;
1906 	}
1907 
1908 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1909 	if (rc)
1910 		ibdev_err(&rdev->ibdev, "Failed to modify Shadow QP for QP1");
1911 	return rc;
1912 }
1913 
1914 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1915 		      int qp_attr_mask, struct ib_udata *udata)
1916 {
1917 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1918 	struct bnxt_re_dev *rdev = qp->rdev;
1919 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1920 	enum ib_qp_state curr_qp_state, new_qp_state;
1921 	int rc, entries;
1922 	unsigned int flags;
1923 	u8 nw_type;
1924 
1925 	if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1926 		return -EOPNOTSUPP;
1927 
1928 	qp->qplib_qp.modify_flags = 0;
1929 	if (qp_attr_mask & IB_QP_STATE) {
1930 		curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state);
1931 		new_qp_state = qp_attr->qp_state;
1932 		if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state,
1933 					ib_qp->qp_type, qp_attr_mask)) {
1934 			ibdev_err(&rdev->ibdev,
1935 				  "Invalid attribute mask: %#x specified ",
1936 				  qp_attr_mask);
1937 			ibdev_err(&rdev->ibdev,
1938 				  "for qpn: %#x type: %#x",
1939 				  ib_qp->qp_num, ib_qp->qp_type);
1940 			ibdev_err(&rdev->ibdev,
1941 				  "curr_qp_state=0x%x, new_qp_state=0x%x\n",
1942 				  curr_qp_state, new_qp_state);
1943 			return -EINVAL;
1944 		}
1945 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1946 		qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state);
1947 
1948 		if (!qp->sumem &&
1949 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1950 			ibdev_dbg(&rdev->ibdev,
1951 				  "Move QP = %p to flush list\n", qp);
1952 			flags = bnxt_re_lock_cqs(qp);
1953 			bnxt_qplib_add_flush_qp(&qp->qplib_qp);
1954 			bnxt_re_unlock_cqs(qp, flags);
1955 		}
1956 		if (!qp->sumem &&
1957 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) {
1958 			ibdev_dbg(&rdev->ibdev,
1959 				  "Move QP = %p out of flush list\n", qp);
1960 			flags = bnxt_re_lock_cqs(qp);
1961 			bnxt_qplib_clean_qp(&qp->qplib_qp);
1962 			bnxt_re_unlock_cqs(qp, flags);
1963 		}
1964 	}
1965 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1966 		qp->qplib_qp.modify_flags |=
1967 				CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY;
1968 		qp->qplib_qp.en_sqd_async_notify = true;
1969 	}
1970 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
1971 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS;
1972 		qp->qplib_qp.access =
1973 			__from_ib_access_flags(qp_attr->qp_access_flags);
1974 		/* LOCAL_WRITE access must be set to allow RC receive */
1975 		qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
1976 		/* Temp: Set all params on QP as of now */
1977 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE;
1978 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ;
1979 	}
1980 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1981 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1982 		qp->qplib_qp.pkey_index = qp_attr->pkey_index;
1983 	}
1984 	if (qp_attr_mask & IB_QP_QKEY) {
1985 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1986 		qp->qplib_qp.qkey = qp_attr->qkey;
1987 	}
1988 	if (qp_attr_mask & IB_QP_AV) {
1989 		const struct ib_global_route *grh =
1990 			rdma_ah_read_grh(&qp_attr->ah_attr);
1991 		const struct ib_gid_attr *sgid_attr;
1992 		struct bnxt_re_gid_ctx *ctx;
1993 
1994 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID |
1995 				     CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL |
1996 				     CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX |
1997 				     CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT |
1998 				     CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS |
1999 				     CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC |
2000 				     CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID;
2001 		memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw,
2002 		       sizeof(qp->qplib_qp.ah.dgid.data));
2003 		qp->qplib_qp.ah.flow_label = grh->flow_label;
2004 		sgid_attr = grh->sgid_attr;
2005 		/* Get the HW context of the GID. The reference
2006 		 * of GID table entry is already taken by the caller.
2007 		 */
2008 		ctx = rdma_read_gid_hw_context(sgid_attr);
2009 		qp->qplib_qp.ah.sgid_index = ctx->idx;
2010 		qp->qplib_qp.ah.host_sgid_index = grh->sgid_index;
2011 		qp->qplib_qp.ah.hop_limit = grh->hop_limit;
2012 		qp->qplib_qp.ah.traffic_class = grh->traffic_class;
2013 		qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr);
2014 		ether_addr_copy(qp->qplib_qp.ah.dmac,
2015 				qp_attr->ah_attr.roce.dmac);
2016 
2017 		rc = rdma_read_gid_l2_fields(sgid_attr, NULL,
2018 					     &qp->qplib_qp.smac[0]);
2019 		if (rc)
2020 			return rc;
2021 
2022 		nw_type = rdma_gid_attr_network_type(sgid_attr);
2023 		switch (nw_type) {
2024 		case RDMA_NETWORK_IPV4:
2025 			qp->qplib_qp.nw_type =
2026 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4;
2027 			break;
2028 		case RDMA_NETWORK_IPV6:
2029 			qp->qplib_qp.nw_type =
2030 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6;
2031 			break;
2032 		default:
2033 			qp->qplib_qp.nw_type =
2034 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1;
2035 			break;
2036 		}
2037 	}
2038 
2039 	if (qp_attr_mask & IB_QP_PATH_MTU) {
2040 		qp->qplib_qp.modify_flags |=
2041 				CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
2042 		qp->qplib_qp.path_mtu = __from_ib_mtu(qp_attr->path_mtu);
2043 		qp->qplib_qp.mtu = ib_mtu_enum_to_int(qp_attr->path_mtu);
2044 	} else if (qp_attr->qp_state == IB_QPS_RTR) {
2045 		qp->qplib_qp.modify_flags |=
2046 			CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
2047 		qp->qplib_qp.path_mtu =
2048 			__from_ib_mtu(iboe_get_mtu(rdev->netdev->mtu));
2049 		qp->qplib_qp.mtu =
2050 			ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
2051 	}
2052 
2053 	if (qp_attr_mask & IB_QP_TIMEOUT) {
2054 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT;
2055 		qp->qplib_qp.timeout = qp_attr->timeout;
2056 	}
2057 	if (qp_attr_mask & IB_QP_RETRY_CNT) {
2058 		qp->qplib_qp.modify_flags |=
2059 				CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT;
2060 		qp->qplib_qp.retry_cnt = qp_attr->retry_cnt;
2061 	}
2062 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
2063 		qp->qplib_qp.modify_flags |=
2064 				CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY;
2065 		qp->qplib_qp.rnr_retry = qp_attr->rnr_retry;
2066 	}
2067 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) {
2068 		qp->qplib_qp.modify_flags |=
2069 				CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER;
2070 		qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer;
2071 	}
2072 	if (qp_attr_mask & IB_QP_RQ_PSN) {
2073 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN;
2074 		qp->qplib_qp.rq.psn = qp_attr->rq_psn;
2075 	}
2076 	if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2077 		qp->qplib_qp.modify_flags |=
2078 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC;
2079 		/* Cap the max_rd_atomic to device max */
2080 		qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic,
2081 						   dev_attr->max_qp_rd_atom);
2082 	}
2083 	if (qp_attr_mask & IB_QP_SQ_PSN) {
2084 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
2085 		qp->qplib_qp.sq.psn = qp_attr->sq_psn;
2086 	}
2087 	if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2088 		if (qp_attr->max_dest_rd_atomic >
2089 		    dev_attr->max_qp_init_rd_atom) {
2090 			ibdev_err(&rdev->ibdev,
2091 				  "max_dest_rd_atomic requested%d is > dev_max%d",
2092 				  qp_attr->max_dest_rd_atomic,
2093 				  dev_attr->max_qp_init_rd_atom);
2094 			return -EINVAL;
2095 		}
2096 
2097 		qp->qplib_qp.modify_flags |=
2098 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC;
2099 		qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic;
2100 	}
2101 	if (qp_attr_mask & IB_QP_CAP) {
2102 		qp->qplib_qp.modify_flags |=
2103 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE |
2104 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE |
2105 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE |
2106 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE |
2107 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA;
2108 		if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) ||
2109 		    (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) ||
2110 		    (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) ||
2111 		    (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) ||
2112 		    (qp_attr->cap.max_inline_data >=
2113 						dev_attr->max_inline_data)) {
2114 			ibdev_err(&rdev->ibdev,
2115 				  "Create QP failed - max exceeded");
2116 			return -EINVAL;
2117 		}
2118 		entries = roundup_pow_of_two(qp_attr->cap.max_send_wr);
2119 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
2120 						dev_attr->max_qp_wqes + 1);
2121 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
2122 						qp_attr->cap.max_send_wr;
2123 		/*
2124 		 * Reserving one slot for Phantom WQE. Some application can
2125 		 * post one extra entry in this case. Allowing this to avoid
2126 		 * unexpected Queue full condition
2127 		 */
2128 		qp->qplib_qp.sq.q_full_delta -= 1;
2129 		qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
2130 		if (qp->qplib_qp.rq.max_wqe) {
2131 			entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr);
2132 			qp->qplib_qp.rq.max_wqe =
2133 				min_t(u32, entries, dev_attr->max_qp_wqes + 1);
2134 			qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
2135 						       qp_attr->cap.max_recv_wr;
2136 			qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
2137 		} else {
2138 			/* SRQ was used prior, just ignore the RQ caps */
2139 		}
2140 	}
2141 	if (qp_attr_mask & IB_QP_DEST_QPN) {
2142 		qp->qplib_qp.modify_flags |=
2143 				CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID;
2144 		qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num;
2145 	}
2146 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
2147 	if (rc) {
2148 		ibdev_err(&rdev->ibdev, "Failed to modify HW QP");
2149 		return rc;
2150 	}
2151 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp)
2152 		rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask);
2153 	return rc;
2154 }
2155 
2156 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
2157 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
2158 {
2159 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2160 	struct bnxt_re_dev *rdev = qp->rdev;
2161 	struct bnxt_qplib_qp *qplib_qp;
2162 	int rc;
2163 
2164 	qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL);
2165 	if (!qplib_qp)
2166 		return -ENOMEM;
2167 
2168 	qplib_qp->id = qp->qplib_qp.id;
2169 	qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index;
2170 
2171 	rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp);
2172 	if (rc) {
2173 		ibdev_err(&rdev->ibdev, "Failed to query HW QP");
2174 		goto out;
2175 	}
2176 	qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state);
2177 	qp_attr->cur_qp_state = __to_ib_qp_state(qplib_qp->cur_qp_state);
2178 	qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0;
2179 	qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access);
2180 	qp_attr->pkey_index = qplib_qp->pkey_index;
2181 	qp_attr->qkey = qplib_qp->qkey;
2182 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2183 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label,
2184 			qplib_qp->ah.host_sgid_index,
2185 			qplib_qp->ah.hop_limit,
2186 			qplib_qp->ah.traffic_class);
2187 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data);
2188 	rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl);
2189 	ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac);
2190 	qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu);
2191 	qp_attr->timeout = qplib_qp->timeout;
2192 	qp_attr->retry_cnt = qplib_qp->retry_cnt;
2193 	qp_attr->rnr_retry = qplib_qp->rnr_retry;
2194 	qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer;
2195 	qp_attr->rq_psn = qplib_qp->rq.psn;
2196 	qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic;
2197 	qp_attr->sq_psn = qplib_qp->sq.psn;
2198 	qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic;
2199 	qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR :
2200 							 IB_SIGNAL_REQ_WR;
2201 	qp_attr->dest_qp_num = qplib_qp->dest_qpn;
2202 
2203 	qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe;
2204 	qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge;
2205 	qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe;
2206 	qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge;
2207 	qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data;
2208 	qp_init_attr->cap = qp_attr->cap;
2209 
2210 out:
2211 	kfree(qplib_qp);
2212 	return rc;
2213 }
2214 
2215 /* Routine for sending QP1 packets for RoCE V1 an V2
2216  */
2217 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp,
2218 				     const struct ib_send_wr *wr,
2219 				     struct bnxt_qplib_swqe *wqe,
2220 				     int payload_size)
2221 {
2222 	struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah,
2223 					     ib_ah);
2224 	struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah;
2225 	const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr;
2226 	struct bnxt_qplib_sge sge;
2227 	u8 nw_type;
2228 	u16 ether_type;
2229 	union ib_gid dgid;
2230 	bool is_eth = false;
2231 	bool is_vlan = false;
2232 	bool is_grh = false;
2233 	bool is_udp = false;
2234 	u8 ip_version = 0;
2235 	u16 vlan_id = 0xFFFF;
2236 	void *buf;
2237 	int i, rc;
2238 
2239 	memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr));
2240 
2241 	rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL);
2242 	if (rc)
2243 		return rc;
2244 
2245 	/* Get network header type for this GID */
2246 	nw_type = rdma_gid_attr_network_type(sgid_attr);
2247 	switch (nw_type) {
2248 	case RDMA_NETWORK_IPV4:
2249 		nw_type = BNXT_RE_ROCEV2_IPV4_PACKET;
2250 		break;
2251 	case RDMA_NETWORK_IPV6:
2252 		nw_type = BNXT_RE_ROCEV2_IPV6_PACKET;
2253 		break;
2254 	default:
2255 		nw_type = BNXT_RE_ROCE_V1_PACKET;
2256 		break;
2257 	}
2258 	memcpy(&dgid.raw, &qplib_ah->dgid, 16);
2259 	is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2260 	if (is_udp) {
2261 		if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) {
2262 			ip_version = 4;
2263 			ether_type = ETH_P_IP;
2264 		} else {
2265 			ip_version = 6;
2266 			ether_type = ETH_P_IPV6;
2267 		}
2268 		is_grh = false;
2269 	} else {
2270 		ether_type = ETH_P_IBOE;
2271 		is_grh = true;
2272 	}
2273 
2274 	is_eth = true;
2275 	is_vlan = vlan_id && (vlan_id < 0x1000);
2276 
2277 	ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh,
2278 			  ip_version, is_udp, 0, &qp->qp1_hdr);
2279 
2280 	/* ETH */
2281 	ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac);
2282 	ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac);
2283 
2284 	/* For vlan, check the sgid for vlan existence */
2285 
2286 	if (!is_vlan) {
2287 		qp->qp1_hdr.eth.type = cpu_to_be16(ether_type);
2288 	} else {
2289 		qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type);
2290 		qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id);
2291 	}
2292 
2293 	if (is_grh || (ip_version == 6)) {
2294 		memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw,
2295 		       sizeof(sgid_attr->gid));
2296 		memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data,
2297 		       sizeof(sgid_attr->gid));
2298 		qp->qp1_hdr.grh.hop_limit     = qplib_ah->hop_limit;
2299 	}
2300 
2301 	if (ip_version == 4) {
2302 		qp->qp1_hdr.ip4.tos = 0;
2303 		qp->qp1_hdr.ip4.id = 0;
2304 		qp->qp1_hdr.ip4.frag_off = htons(IP_DF);
2305 		qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit;
2306 
2307 		memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4);
2308 		memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4);
2309 		qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr);
2310 	}
2311 
2312 	if (is_udp) {
2313 		qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT);
2314 		qp->qp1_hdr.udp.sport = htons(0x8CD1);
2315 		qp->qp1_hdr.udp.csum = 0;
2316 	}
2317 
2318 	/* BTH */
2319 	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
2320 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2321 		qp->qp1_hdr.immediate_present = 1;
2322 	} else {
2323 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2324 	}
2325 	if (wr->send_flags & IB_SEND_SOLICITED)
2326 		qp->qp1_hdr.bth.solicited_event = 1;
2327 	/* pad_count */
2328 	qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3;
2329 
2330 	/* P_key for QP1 is for all members */
2331 	qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF);
2332 	qp->qp1_hdr.bth.destination_qpn = IB_QP1;
2333 	qp->qp1_hdr.bth.ack_req = 0;
2334 	qp->send_psn++;
2335 	qp->send_psn &= BTH_PSN_MASK;
2336 	qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn);
2337 	/* DETH */
2338 	/* Use the priviledged Q_Key for QP1 */
2339 	qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY);
2340 	qp->qp1_hdr.deth.source_qpn = IB_QP1;
2341 
2342 	/* Pack the QP1 to the transmit buffer */
2343 	buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge);
2344 	if (buf) {
2345 		ib_ud_header_pack(&qp->qp1_hdr, buf);
2346 		for (i = wqe->num_sge; i; i--) {
2347 			wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr;
2348 			wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey;
2349 			wqe->sg_list[i].size = wqe->sg_list[i - 1].size;
2350 		}
2351 
2352 		/*
2353 		 * Max Header buf size for IPV6 RoCE V2 is 86,
2354 		 * which is same as the QP1 SQ header buffer.
2355 		 * Header buf size for IPV4 RoCE V2 can be 66.
2356 		 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
2357 		 * Subtract 20 bytes from QP1 SQ header buf size
2358 		 */
2359 		if (is_udp && ip_version == 4)
2360 			sge.size -= 20;
2361 		/*
2362 		 * Max Header buf size for RoCE V1 is 78.
2363 		 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
2364 		 * Subtract 8 bytes from QP1 SQ header buf size
2365 		 */
2366 		if (!is_udp)
2367 			sge.size -= 8;
2368 
2369 		/* Subtract 4 bytes for non vlan packets */
2370 		if (!is_vlan)
2371 			sge.size -= 4;
2372 
2373 		wqe->sg_list[0].addr = sge.addr;
2374 		wqe->sg_list[0].lkey = sge.lkey;
2375 		wqe->sg_list[0].size = sge.size;
2376 		wqe->num_sge++;
2377 
2378 	} else {
2379 		ibdev_err(&qp->rdev->ibdev, "QP1 buffer is empty!");
2380 		rc = -ENOMEM;
2381 	}
2382 	return rc;
2383 }
2384 
2385 /* For the MAD layer, it only provides the recv SGE the size of
2386  * ib_grh + MAD datagram.  No Ethernet headers, Ethertype, BTH, DETH,
2387  * nor RoCE iCRC.  The Cu+ solution must provide buffer for the entire
2388  * receive packet (334 bytes) with no VLAN and then copy the GRH
2389  * and the MAD datagram out to the provided SGE.
2390  */
2391 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp,
2392 					    const struct ib_recv_wr *wr,
2393 					    struct bnxt_qplib_swqe *wqe,
2394 					    int payload_size)
2395 {
2396 	struct bnxt_re_sqp_entries *sqp_entry;
2397 	struct bnxt_qplib_sge ref, sge;
2398 	struct bnxt_re_dev *rdev;
2399 	u32 rq_prod_index;
2400 
2401 	rdev = qp->rdev;
2402 
2403 	rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp);
2404 
2405 	if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge))
2406 		return -ENOMEM;
2407 
2408 	/* Create 1 SGE to receive the entire
2409 	 * ethernet packet
2410 	 */
2411 	/* Save the reference from ULP */
2412 	ref.addr = wqe->sg_list[0].addr;
2413 	ref.lkey = wqe->sg_list[0].lkey;
2414 	ref.size = wqe->sg_list[0].size;
2415 
2416 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[rq_prod_index];
2417 
2418 	/* SGE 1 */
2419 	wqe->sg_list[0].addr = sge.addr;
2420 	wqe->sg_list[0].lkey = sge.lkey;
2421 	wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
2422 	sge.size -= wqe->sg_list[0].size;
2423 
2424 	sqp_entry->sge.addr = ref.addr;
2425 	sqp_entry->sge.lkey = ref.lkey;
2426 	sqp_entry->sge.size = ref.size;
2427 	/* Store the wrid for reporting completion */
2428 	sqp_entry->wrid = wqe->wr_id;
2429 	/* change the wqe->wrid to table index */
2430 	wqe->wr_id = rq_prod_index;
2431 	return 0;
2432 }
2433 
2434 static int is_ud_qp(struct bnxt_re_qp *qp)
2435 {
2436 	return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD ||
2437 		qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI);
2438 }
2439 
2440 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp,
2441 				  const struct ib_send_wr *wr,
2442 				  struct bnxt_qplib_swqe *wqe)
2443 {
2444 	struct bnxt_re_ah *ah = NULL;
2445 
2446 	if (is_ud_qp(qp)) {
2447 		ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah);
2448 		wqe->send.q_key = ud_wr(wr)->remote_qkey;
2449 		wqe->send.dst_qp = ud_wr(wr)->remote_qpn;
2450 		wqe->send.avid = ah->qplib_ah.id;
2451 	}
2452 	switch (wr->opcode) {
2453 	case IB_WR_SEND:
2454 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND;
2455 		break;
2456 	case IB_WR_SEND_WITH_IMM:
2457 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM;
2458 		wqe->send.imm_data = wr->ex.imm_data;
2459 		break;
2460 	case IB_WR_SEND_WITH_INV:
2461 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV;
2462 		wqe->send.inv_key = wr->ex.invalidate_rkey;
2463 		break;
2464 	default:
2465 		return -EINVAL;
2466 	}
2467 	if (wr->send_flags & IB_SEND_SIGNALED)
2468 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2469 	if (wr->send_flags & IB_SEND_FENCE)
2470 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2471 	if (wr->send_flags & IB_SEND_SOLICITED)
2472 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2473 	if (wr->send_flags & IB_SEND_INLINE)
2474 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2475 
2476 	return 0;
2477 }
2478 
2479 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr,
2480 				  struct bnxt_qplib_swqe *wqe)
2481 {
2482 	switch (wr->opcode) {
2483 	case IB_WR_RDMA_WRITE:
2484 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE;
2485 		break;
2486 	case IB_WR_RDMA_WRITE_WITH_IMM:
2487 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM;
2488 		wqe->rdma.imm_data = wr->ex.imm_data;
2489 		break;
2490 	case IB_WR_RDMA_READ:
2491 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ;
2492 		wqe->rdma.inv_key = wr->ex.invalidate_rkey;
2493 		break;
2494 	default:
2495 		return -EINVAL;
2496 	}
2497 	wqe->rdma.remote_va = rdma_wr(wr)->remote_addr;
2498 	wqe->rdma.r_key = rdma_wr(wr)->rkey;
2499 	if (wr->send_flags & IB_SEND_SIGNALED)
2500 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2501 	if (wr->send_flags & IB_SEND_FENCE)
2502 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2503 	if (wr->send_flags & IB_SEND_SOLICITED)
2504 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2505 	if (wr->send_flags & IB_SEND_INLINE)
2506 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2507 
2508 	return 0;
2509 }
2510 
2511 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr,
2512 				    struct bnxt_qplib_swqe *wqe)
2513 {
2514 	switch (wr->opcode) {
2515 	case IB_WR_ATOMIC_CMP_AND_SWP:
2516 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP;
2517 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2518 		wqe->atomic.swap_data = atomic_wr(wr)->swap;
2519 		break;
2520 	case IB_WR_ATOMIC_FETCH_AND_ADD:
2521 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD;
2522 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2523 		break;
2524 	default:
2525 		return -EINVAL;
2526 	}
2527 	wqe->atomic.remote_va = atomic_wr(wr)->remote_addr;
2528 	wqe->atomic.r_key = atomic_wr(wr)->rkey;
2529 	if (wr->send_flags & IB_SEND_SIGNALED)
2530 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2531 	if (wr->send_flags & IB_SEND_FENCE)
2532 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2533 	if (wr->send_flags & IB_SEND_SOLICITED)
2534 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2535 	return 0;
2536 }
2537 
2538 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr,
2539 				 struct bnxt_qplib_swqe *wqe)
2540 {
2541 	wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV;
2542 	wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey;
2543 
2544 	/* Need unconditional fence for local invalidate
2545 	 * opcode to work as expected.
2546 	 */
2547 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2548 
2549 	if (wr->send_flags & IB_SEND_SIGNALED)
2550 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2551 	if (wr->send_flags & IB_SEND_SOLICITED)
2552 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2553 
2554 	return 0;
2555 }
2556 
2557 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr,
2558 				 struct bnxt_qplib_swqe *wqe)
2559 {
2560 	struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr);
2561 	struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl;
2562 	int access = wr->access;
2563 
2564 	wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0];
2565 	wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0];
2566 	wqe->frmr.page_list = mr->pages;
2567 	wqe->frmr.page_list_len = mr->npages;
2568 	wqe->frmr.levels = qplib_frpl->hwq.level;
2569 	wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR;
2570 
2571 	/* Need unconditional fence for reg_mr
2572 	 * opcode to function as expected.
2573 	 */
2574 
2575 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2576 
2577 	if (wr->wr.send_flags & IB_SEND_SIGNALED)
2578 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2579 
2580 	if (access & IB_ACCESS_LOCAL_WRITE)
2581 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE;
2582 	if (access & IB_ACCESS_REMOTE_READ)
2583 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ;
2584 	if (access & IB_ACCESS_REMOTE_WRITE)
2585 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE;
2586 	if (access & IB_ACCESS_REMOTE_ATOMIC)
2587 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC;
2588 	if (access & IB_ACCESS_MW_BIND)
2589 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND;
2590 
2591 	wqe->frmr.l_key = wr->key;
2592 	wqe->frmr.length = wr->mr->length;
2593 	wqe->frmr.pbl_pg_sz_log = ilog2(PAGE_SIZE >> PAGE_SHIFT_4K);
2594 	wqe->frmr.pg_sz_log = ilog2(wr->mr->page_size >> PAGE_SHIFT_4K);
2595 	wqe->frmr.va = wr->mr->iova;
2596 	return 0;
2597 }
2598 
2599 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev,
2600 				    const struct ib_send_wr *wr,
2601 				    struct bnxt_qplib_swqe *wqe)
2602 {
2603 	/*  Copy the inline data to the data  field */
2604 	u8 *in_data;
2605 	u32 i, sge_len;
2606 	void *sge_addr;
2607 
2608 	in_data = wqe->inline_data;
2609 	for (i = 0; i < wr->num_sge; i++) {
2610 		sge_addr = (void *)(unsigned long)
2611 				wr->sg_list[i].addr;
2612 		sge_len = wr->sg_list[i].length;
2613 
2614 		if ((sge_len + wqe->inline_len) >
2615 		    BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) {
2616 			ibdev_err(&rdev->ibdev,
2617 				  "Inline data size requested > supported value");
2618 			return -EINVAL;
2619 		}
2620 		sge_len = wr->sg_list[i].length;
2621 
2622 		memcpy(in_data, sge_addr, sge_len);
2623 		in_data += wr->sg_list[i].length;
2624 		wqe->inline_len += wr->sg_list[i].length;
2625 	}
2626 	return wqe->inline_len;
2627 }
2628 
2629 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
2630 				   const struct ib_send_wr *wr,
2631 				   struct bnxt_qplib_swqe *wqe)
2632 {
2633 	int payload_sz = 0;
2634 
2635 	if (wr->send_flags & IB_SEND_INLINE)
2636 		payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe);
2637 	else
2638 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list,
2639 					       wqe->num_sge);
2640 
2641 	return payload_sz;
2642 }
2643 
2644 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
2645 {
2646 	if ((qp->ib_qp.qp_type == IB_QPT_UD ||
2647 	     qp->ib_qp.qp_type == IB_QPT_GSI ||
2648 	     qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
2649 	     qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
2650 		int qp_attr_mask;
2651 		struct ib_qp_attr qp_attr;
2652 
2653 		qp_attr_mask = IB_QP_STATE;
2654 		qp_attr.qp_state = IB_QPS_RTS;
2655 		bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
2656 		qp->qplib_qp.wqe_cnt = 0;
2657 	}
2658 }
2659 
2660 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
2661 				       struct bnxt_re_qp *qp,
2662 				       const struct ib_send_wr *wr)
2663 {
2664 	int rc = 0, payload_sz = 0;
2665 	unsigned long flags;
2666 
2667 	spin_lock_irqsave(&qp->sq_lock, flags);
2668 	while (wr) {
2669 		struct bnxt_qplib_swqe wqe = {};
2670 
2671 		/* Common */
2672 		wqe.num_sge = wr->num_sge;
2673 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2674 			ibdev_err(&rdev->ibdev,
2675 				  "Limit exceeded for Send SGEs");
2676 			rc = -EINVAL;
2677 			goto bad;
2678 		}
2679 
2680 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2681 		if (payload_sz < 0) {
2682 			rc = -EINVAL;
2683 			goto bad;
2684 		}
2685 		wqe.wr_id = wr->wr_id;
2686 
2687 		wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND;
2688 
2689 		rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2690 		if (!rc)
2691 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2692 bad:
2693 		if (rc) {
2694 			ibdev_err(&rdev->ibdev,
2695 				  "Post send failed opcode = %#x rc = %d",
2696 				  wr->opcode, rc);
2697 			break;
2698 		}
2699 		wr = wr->next;
2700 	}
2701 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2702 	bnxt_ud_qp_hw_stall_workaround(qp);
2703 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2704 	return rc;
2705 }
2706 
2707 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr,
2708 		      const struct ib_send_wr **bad_wr)
2709 {
2710 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2711 	struct bnxt_qplib_swqe wqe;
2712 	int rc = 0, payload_sz = 0;
2713 	unsigned long flags;
2714 
2715 	spin_lock_irqsave(&qp->sq_lock, flags);
2716 	while (wr) {
2717 		/* House keeping */
2718 		memset(&wqe, 0, sizeof(wqe));
2719 
2720 		/* Common */
2721 		wqe.num_sge = wr->num_sge;
2722 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2723 			ibdev_err(&qp->rdev->ibdev,
2724 				  "Limit exceeded for Send SGEs");
2725 			rc = -EINVAL;
2726 			goto bad;
2727 		}
2728 
2729 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2730 		if (payload_sz < 0) {
2731 			rc = -EINVAL;
2732 			goto bad;
2733 		}
2734 		wqe.wr_id = wr->wr_id;
2735 
2736 		switch (wr->opcode) {
2737 		case IB_WR_SEND:
2738 		case IB_WR_SEND_WITH_IMM:
2739 			if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) {
2740 				rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe,
2741 							       payload_sz);
2742 				if (rc)
2743 					goto bad;
2744 				wqe.rawqp1.lflags |=
2745 					SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC;
2746 			}
2747 			switch (wr->send_flags) {
2748 			case IB_SEND_IP_CSUM:
2749 				wqe.rawqp1.lflags |=
2750 					SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM;
2751 				break;
2752 			default:
2753 				break;
2754 			}
2755 			fallthrough;
2756 		case IB_WR_SEND_WITH_INV:
2757 			rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2758 			break;
2759 		case IB_WR_RDMA_WRITE:
2760 		case IB_WR_RDMA_WRITE_WITH_IMM:
2761 		case IB_WR_RDMA_READ:
2762 			rc = bnxt_re_build_rdma_wqe(wr, &wqe);
2763 			break;
2764 		case IB_WR_ATOMIC_CMP_AND_SWP:
2765 		case IB_WR_ATOMIC_FETCH_AND_ADD:
2766 			rc = bnxt_re_build_atomic_wqe(wr, &wqe);
2767 			break;
2768 		case IB_WR_RDMA_READ_WITH_INV:
2769 			ibdev_err(&qp->rdev->ibdev,
2770 				  "RDMA Read with Invalidate is not supported");
2771 			rc = -EINVAL;
2772 			goto bad;
2773 		case IB_WR_LOCAL_INV:
2774 			rc = bnxt_re_build_inv_wqe(wr, &wqe);
2775 			break;
2776 		case IB_WR_REG_MR:
2777 			rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe);
2778 			break;
2779 		default:
2780 			/* Unsupported WRs */
2781 			ibdev_err(&qp->rdev->ibdev,
2782 				  "WR (%#x) is not supported", wr->opcode);
2783 			rc = -EINVAL;
2784 			goto bad;
2785 		}
2786 		if (!rc)
2787 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2788 bad:
2789 		if (rc) {
2790 			ibdev_err(&qp->rdev->ibdev,
2791 				  "post_send failed op:%#x qps = %#x rc = %d\n",
2792 				  wr->opcode, qp->qplib_qp.state, rc);
2793 			*bad_wr = wr;
2794 			break;
2795 		}
2796 		wr = wr->next;
2797 	}
2798 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2799 	bnxt_ud_qp_hw_stall_workaround(qp);
2800 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2801 
2802 	return rc;
2803 }
2804 
2805 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev,
2806 				       struct bnxt_re_qp *qp,
2807 				       const struct ib_recv_wr *wr)
2808 {
2809 	struct bnxt_qplib_swqe wqe;
2810 	int rc = 0;
2811 
2812 	while (wr) {
2813 		/* House keeping */
2814 		memset(&wqe, 0, sizeof(wqe));
2815 
2816 		/* Common */
2817 		wqe.num_sge = wr->num_sge;
2818 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2819 			ibdev_err(&rdev->ibdev,
2820 				  "Limit exceeded for Receive SGEs");
2821 			rc = -EINVAL;
2822 			break;
2823 		}
2824 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
2825 		wqe.wr_id = wr->wr_id;
2826 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2827 
2828 		rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2829 		if (rc)
2830 			break;
2831 
2832 		wr = wr->next;
2833 	}
2834 	if (!rc)
2835 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2836 	return rc;
2837 }
2838 
2839 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
2840 		      const struct ib_recv_wr **bad_wr)
2841 {
2842 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2843 	struct bnxt_qplib_swqe wqe;
2844 	int rc = 0, payload_sz = 0;
2845 	unsigned long flags;
2846 	u32 count = 0;
2847 
2848 	spin_lock_irqsave(&qp->rq_lock, flags);
2849 	while (wr) {
2850 		/* House keeping */
2851 		memset(&wqe, 0, sizeof(wqe));
2852 
2853 		/* Common */
2854 		wqe.num_sge = wr->num_sge;
2855 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2856 			ibdev_err(&qp->rdev->ibdev,
2857 				  "Limit exceeded for Receive SGEs");
2858 			rc = -EINVAL;
2859 			*bad_wr = wr;
2860 			break;
2861 		}
2862 
2863 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
2864 					       wr->num_sge);
2865 		wqe.wr_id = wr->wr_id;
2866 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2867 
2868 		if (ib_qp->qp_type == IB_QPT_GSI &&
2869 		    qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI)
2870 			rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe,
2871 							      payload_sz);
2872 		if (!rc)
2873 			rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2874 		if (rc) {
2875 			*bad_wr = wr;
2876 			break;
2877 		}
2878 
2879 		/* Ring DB if the RQEs posted reaches a threshold value */
2880 		if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
2881 			bnxt_qplib_post_recv_db(&qp->qplib_qp);
2882 			count = 0;
2883 		}
2884 
2885 		wr = wr->next;
2886 	}
2887 
2888 	if (count)
2889 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2890 
2891 	spin_unlock_irqrestore(&qp->rq_lock, flags);
2892 
2893 	return rc;
2894 }
2895 
2896 /* Completion Queues */
2897 int bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
2898 {
2899 	struct bnxt_re_cq *cq;
2900 	struct bnxt_qplib_nq *nq;
2901 	struct bnxt_re_dev *rdev;
2902 
2903 	cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
2904 	rdev = cq->rdev;
2905 	nq = cq->qplib_cq.nq;
2906 
2907 	bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2908 	ib_umem_release(cq->umem);
2909 
2910 	atomic_dec(&rdev->stats.res.cq_count);
2911 	nq->budget--;
2912 	kfree(cq->cql);
2913 	return 0;
2914 }
2915 
2916 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
2917 		      struct ib_udata *udata)
2918 {
2919 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev);
2920 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
2921 	struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq);
2922 	int rc, entries;
2923 	int cqe = attr->cqe;
2924 	struct bnxt_qplib_nq *nq = NULL;
2925 	unsigned int nq_alloc_cnt;
2926 	u32 active_cqs;
2927 
2928 	if (attr->flags)
2929 		return -EOPNOTSUPP;
2930 
2931 	/* Validate CQ fields */
2932 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
2933 		ibdev_err(&rdev->ibdev, "Failed to create CQ -max exceeded");
2934 		return -EINVAL;
2935 	}
2936 
2937 	cq->rdev = rdev;
2938 	cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
2939 
2940 	entries = roundup_pow_of_two(cqe + 1);
2941 	if (entries > dev_attr->max_cq_wqes + 1)
2942 		entries = dev_attr->max_cq_wqes + 1;
2943 
2944 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
2945 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
2946 	if (udata) {
2947 		struct bnxt_re_cq_req req;
2948 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
2949 			udata, struct bnxt_re_ucontext, ib_uctx);
2950 		if (ib_copy_from_udata(&req, udata, sizeof(req))) {
2951 			rc = -EFAULT;
2952 			goto fail;
2953 		}
2954 
2955 		cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va,
2956 				       entries * sizeof(struct cq_base),
2957 				       IB_ACCESS_LOCAL_WRITE);
2958 		if (IS_ERR(cq->umem)) {
2959 			rc = PTR_ERR(cq->umem);
2960 			goto fail;
2961 		}
2962 		cq->qplib_cq.sg_info.umem = cq->umem;
2963 		cq->qplib_cq.dpi = &uctx->dpi;
2964 	} else {
2965 		cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL);
2966 		cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe),
2967 				  GFP_KERNEL);
2968 		if (!cq->cql) {
2969 			rc = -ENOMEM;
2970 			goto fail;
2971 		}
2972 
2973 		cq->qplib_cq.dpi = &rdev->dpi_privileged;
2974 	}
2975 	/*
2976 	 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
2977 	 * used for getting the NQ index.
2978 	 */
2979 	nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
2980 	nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
2981 	cq->qplib_cq.max_wqe = entries;
2982 	cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
2983 	cq->qplib_cq.nq	= nq;
2984 
2985 	rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
2986 	if (rc) {
2987 		ibdev_err(&rdev->ibdev, "Failed to create HW CQ");
2988 		goto fail;
2989 	}
2990 
2991 	cq->ib_cq.cqe = entries;
2992 	cq->cq_period = cq->qplib_cq.period;
2993 	nq->budget++;
2994 
2995 	active_cqs = atomic_inc_return(&rdev->stats.res.cq_count);
2996 	if (active_cqs > rdev->stats.res.cq_watermark)
2997 		rdev->stats.res.cq_watermark = active_cqs;
2998 	spin_lock_init(&cq->cq_lock);
2999 
3000 	if (udata) {
3001 		struct bnxt_re_cq_resp resp;
3002 
3003 		resp.cqid = cq->qplib_cq.id;
3004 		resp.tail = cq->qplib_cq.hwq.cons;
3005 		resp.phase = cq->qplib_cq.period;
3006 		resp.rsvd = 0;
3007 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
3008 		if (rc) {
3009 			ibdev_err(&rdev->ibdev, "Failed to copy CQ udata");
3010 			bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
3011 			goto c2fail;
3012 		}
3013 	}
3014 
3015 	return 0;
3016 
3017 c2fail:
3018 	ib_umem_release(cq->umem);
3019 fail:
3020 	kfree(cq->cql);
3021 	return rc;
3022 }
3023 
3024 static void bnxt_re_resize_cq_complete(struct bnxt_re_cq *cq)
3025 {
3026 	struct bnxt_re_dev *rdev = cq->rdev;
3027 
3028 	bnxt_qplib_resize_cq_complete(&rdev->qplib_res, &cq->qplib_cq);
3029 
3030 	cq->qplib_cq.max_wqe = cq->resize_cqe;
3031 	if (cq->resize_umem) {
3032 		ib_umem_release(cq->umem);
3033 		cq->umem = cq->resize_umem;
3034 		cq->resize_umem = NULL;
3035 		cq->resize_cqe = 0;
3036 	}
3037 }
3038 
3039 int bnxt_re_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
3040 {
3041 	struct bnxt_qplib_sg_info sg_info = {};
3042 	struct bnxt_qplib_dpi *orig_dpi = NULL;
3043 	struct bnxt_qplib_dev_attr *dev_attr;
3044 	struct bnxt_re_ucontext *uctx = NULL;
3045 	struct bnxt_re_resize_cq_req req;
3046 	struct bnxt_re_dev *rdev;
3047 	struct bnxt_re_cq *cq;
3048 	int rc, entries;
3049 
3050 	cq =  container_of(ibcq, struct bnxt_re_cq, ib_cq);
3051 	rdev = cq->rdev;
3052 	dev_attr = &rdev->dev_attr;
3053 	if (!ibcq->uobject) {
3054 		ibdev_err(&rdev->ibdev, "Kernel CQ Resize not supported");
3055 		return -EOPNOTSUPP;
3056 	}
3057 
3058 	if (cq->resize_umem) {
3059 		ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - Busy",
3060 			  cq->qplib_cq.id);
3061 		return -EBUSY;
3062 	}
3063 
3064 	/* Check the requested cq depth out of supported depth */
3065 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
3066 		ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - out of range cqe %d",
3067 			  cq->qplib_cq.id, cqe);
3068 		return -EINVAL;
3069 	}
3070 
3071 	entries = roundup_pow_of_two(cqe + 1);
3072 	if (entries > dev_attr->max_cq_wqes + 1)
3073 		entries = dev_attr->max_cq_wqes + 1;
3074 
3075 	uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext,
3076 					 ib_uctx);
3077 	/* uverbs consumer */
3078 	if (ib_copy_from_udata(&req, udata, sizeof(req))) {
3079 		rc = -EFAULT;
3080 		goto fail;
3081 	}
3082 
3083 	cq->resize_umem = ib_umem_get(&rdev->ibdev, req.cq_va,
3084 				      entries * sizeof(struct cq_base),
3085 				      IB_ACCESS_LOCAL_WRITE);
3086 	if (IS_ERR(cq->resize_umem)) {
3087 		rc = PTR_ERR(cq->resize_umem);
3088 		cq->resize_umem = NULL;
3089 		ibdev_err(&rdev->ibdev, "%s: ib_umem_get failed! rc = %d\n",
3090 			  __func__, rc);
3091 		goto fail;
3092 	}
3093 	cq->resize_cqe = entries;
3094 	memcpy(&sg_info, &cq->qplib_cq.sg_info, sizeof(sg_info));
3095 	orig_dpi = cq->qplib_cq.dpi;
3096 
3097 	cq->qplib_cq.sg_info.umem = cq->resize_umem;
3098 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
3099 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
3100 	cq->qplib_cq.dpi = &uctx->dpi;
3101 
3102 	rc = bnxt_qplib_resize_cq(&rdev->qplib_res, &cq->qplib_cq, entries);
3103 	if (rc) {
3104 		ibdev_err(&rdev->ibdev, "Resize HW CQ %#x failed!",
3105 			  cq->qplib_cq.id);
3106 		goto fail;
3107 	}
3108 
3109 	cq->ib_cq.cqe = cq->resize_cqe;
3110 	atomic_inc(&rdev->stats.res.resize_count);
3111 
3112 	return 0;
3113 
3114 fail:
3115 	if (cq->resize_umem) {
3116 		ib_umem_release(cq->resize_umem);
3117 		cq->resize_umem = NULL;
3118 		cq->resize_cqe = 0;
3119 		memcpy(&cq->qplib_cq.sg_info, &sg_info, sizeof(sg_info));
3120 		cq->qplib_cq.dpi = orig_dpi;
3121 	}
3122 	return rc;
3123 }
3124 
3125 static u8 __req_to_ib_wc_status(u8 qstatus)
3126 {
3127 	switch (qstatus) {
3128 	case CQ_REQ_STATUS_OK:
3129 		return IB_WC_SUCCESS;
3130 	case CQ_REQ_STATUS_BAD_RESPONSE_ERR:
3131 		return IB_WC_BAD_RESP_ERR;
3132 	case CQ_REQ_STATUS_LOCAL_LENGTH_ERR:
3133 		return IB_WC_LOC_LEN_ERR;
3134 	case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR:
3135 		return IB_WC_LOC_QP_OP_ERR;
3136 	case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR:
3137 		return IB_WC_LOC_PROT_ERR;
3138 	case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR:
3139 		return IB_WC_GENERAL_ERR;
3140 	case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR:
3141 		return IB_WC_REM_INV_REQ_ERR;
3142 	case CQ_REQ_STATUS_REMOTE_ACCESS_ERR:
3143 		return IB_WC_REM_ACCESS_ERR;
3144 	case CQ_REQ_STATUS_REMOTE_OPERATION_ERR:
3145 		return IB_WC_REM_OP_ERR;
3146 	case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR:
3147 		return IB_WC_RNR_RETRY_EXC_ERR;
3148 	case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR:
3149 		return IB_WC_RETRY_EXC_ERR;
3150 	case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR:
3151 		return IB_WC_WR_FLUSH_ERR;
3152 	default:
3153 		return IB_WC_GENERAL_ERR;
3154 	}
3155 	return 0;
3156 }
3157 
3158 static u8 __rawqp1_to_ib_wc_status(u8 qstatus)
3159 {
3160 	switch (qstatus) {
3161 	case CQ_RES_RAWETH_QP1_STATUS_OK:
3162 		return IB_WC_SUCCESS;
3163 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR:
3164 		return IB_WC_LOC_ACCESS_ERR;
3165 	case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR:
3166 		return IB_WC_LOC_LEN_ERR;
3167 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR:
3168 		return IB_WC_LOC_PROT_ERR;
3169 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR:
3170 		return IB_WC_LOC_QP_OP_ERR;
3171 	case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR:
3172 		return IB_WC_GENERAL_ERR;
3173 	case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR:
3174 		return IB_WC_WR_FLUSH_ERR;
3175 	case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR:
3176 		return IB_WC_WR_FLUSH_ERR;
3177 	default:
3178 		return IB_WC_GENERAL_ERR;
3179 	}
3180 }
3181 
3182 static u8 __rc_to_ib_wc_status(u8 qstatus)
3183 {
3184 	switch (qstatus) {
3185 	case CQ_RES_RC_STATUS_OK:
3186 		return IB_WC_SUCCESS;
3187 	case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR:
3188 		return IB_WC_LOC_ACCESS_ERR;
3189 	case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR:
3190 		return IB_WC_LOC_LEN_ERR;
3191 	case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR:
3192 		return IB_WC_LOC_PROT_ERR;
3193 	case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR:
3194 		return IB_WC_LOC_QP_OP_ERR;
3195 	case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR:
3196 		return IB_WC_GENERAL_ERR;
3197 	case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR:
3198 		return IB_WC_REM_INV_REQ_ERR;
3199 	case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR:
3200 		return IB_WC_WR_FLUSH_ERR;
3201 	case CQ_RES_RC_STATUS_HW_FLUSH_ERR:
3202 		return IB_WC_WR_FLUSH_ERR;
3203 	default:
3204 		return IB_WC_GENERAL_ERR;
3205 	}
3206 }
3207 
3208 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe)
3209 {
3210 	switch (cqe->type) {
3211 	case BNXT_QPLIB_SWQE_TYPE_SEND:
3212 		wc->opcode = IB_WC_SEND;
3213 		break;
3214 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM:
3215 		wc->opcode = IB_WC_SEND;
3216 		wc->wc_flags |= IB_WC_WITH_IMM;
3217 		break;
3218 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV:
3219 		wc->opcode = IB_WC_SEND;
3220 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3221 		break;
3222 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE:
3223 		wc->opcode = IB_WC_RDMA_WRITE;
3224 		break;
3225 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM:
3226 		wc->opcode = IB_WC_RDMA_WRITE;
3227 		wc->wc_flags |= IB_WC_WITH_IMM;
3228 		break;
3229 	case BNXT_QPLIB_SWQE_TYPE_RDMA_READ:
3230 		wc->opcode = IB_WC_RDMA_READ;
3231 		break;
3232 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP:
3233 		wc->opcode = IB_WC_COMP_SWAP;
3234 		break;
3235 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD:
3236 		wc->opcode = IB_WC_FETCH_ADD;
3237 		break;
3238 	case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV:
3239 		wc->opcode = IB_WC_LOCAL_INV;
3240 		break;
3241 	case BNXT_QPLIB_SWQE_TYPE_REG_MR:
3242 		wc->opcode = IB_WC_REG_MR;
3243 		break;
3244 	default:
3245 		wc->opcode = IB_WC_SEND;
3246 		break;
3247 	}
3248 
3249 	wc->status = __req_to_ib_wc_status(cqe->status);
3250 }
3251 
3252 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags,
3253 				     u16 raweth_qp1_flags2)
3254 {
3255 	bool is_ipv6 = false, is_ipv4 = false;
3256 
3257 	/* raweth_qp1_flags Bit 9-6 indicates itype */
3258 	if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3259 	    != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3260 		return -1;
3261 
3262 	if (raweth_qp1_flags2 &
3263 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC &&
3264 	    raweth_qp1_flags2 &
3265 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) {
3266 		/* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
3267 		(raweth_qp1_flags2 &
3268 		 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ?
3269 			(is_ipv6 = true) : (is_ipv4 = true);
3270 		return ((is_ipv6) ?
3271 			 BNXT_RE_ROCEV2_IPV6_PACKET :
3272 			 BNXT_RE_ROCEV2_IPV4_PACKET);
3273 	} else {
3274 		return BNXT_RE_ROCE_V1_PACKET;
3275 	}
3276 }
3277 
3278 static int bnxt_re_to_ib_nw_type(int nw_type)
3279 {
3280 	u8 nw_hdr_type = 0xFF;
3281 
3282 	switch (nw_type) {
3283 	case BNXT_RE_ROCE_V1_PACKET:
3284 		nw_hdr_type = RDMA_NETWORK_ROCE_V1;
3285 		break;
3286 	case BNXT_RE_ROCEV2_IPV4_PACKET:
3287 		nw_hdr_type = RDMA_NETWORK_IPV4;
3288 		break;
3289 	case BNXT_RE_ROCEV2_IPV6_PACKET:
3290 		nw_hdr_type = RDMA_NETWORK_IPV6;
3291 		break;
3292 	}
3293 	return nw_hdr_type;
3294 }
3295 
3296 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev,
3297 				       void *rq_hdr_buf)
3298 {
3299 	u8 *tmp_buf = NULL;
3300 	struct ethhdr *eth_hdr;
3301 	u16 eth_type;
3302 	bool rc = false;
3303 
3304 	tmp_buf = (u8 *)rq_hdr_buf;
3305 	/*
3306 	 * If dest mac is not same as I/F mac, this could be a
3307 	 * loopback address or multicast address, check whether
3308 	 * it is a loopback packet
3309 	 */
3310 	if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) {
3311 		tmp_buf += 4;
3312 		/* Check the  ether type */
3313 		eth_hdr = (struct ethhdr *)tmp_buf;
3314 		eth_type = ntohs(eth_hdr->h_proto);
3315 		switch (eth_type) {
3316 		case ETH_P_IBOE:
3317 			rc = true;
3318 			break;
3319 		case ETH_P_IP:
3320 		case ETH_P_IPV6: {
3321 			u32 len;
3322 			struct udphdr *udp_hdr;
3323 
3324 			len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) :
3325 						      sizeof(struct ipv6hdr));
3326 			tmp_buf += sizeof(struct ethhdr) + len;
3327 			udp_hdr = (struct udphdr *)tmp_buf;
3328 			if (ntohs(udp_hdr->dest) ==
3329 				    ROCE_V2_UDP_DPORT)
3330 				rc = true;
3331 			break;
3332 			}
3333 		default:
3334 			break;
3335 		}
3336 	}
3337 
3338 	return rc;
3339 }
3340 
3341 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *gsi_qp,
3342 					 struct bnxt_qplib_cqe *cqe)
3343 {
3344 	struct bnxt_re_dev *rdev = gsi_qp->rdev;
3345 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3346 	struct bnxt_re_qp *gsi_sqp = rdev->gsi_ctx.gsi_sqp;
3347 	dma_addr_t shrq_hdr_buf_map;
3348 	struct ib_sge s_sge[2] = {};
3349 	struct ib_sge r_sge[2] = {};
3350 	struct bnxt_re_ah *gsi_sah;
3351 	struct ib_recv_wr rwr = {};
3352 	dma_addr_t rq_hdr_buf_map;
3353 	struct ib_ud_wr udwr = {};
3354 	struct ib_send_wr *swr;
3355 	u32 skip_bytes = 0;
3356 	int pkt_type = 0;
3357 	void *rq_hdr_buf;
3358 	u32 offset = 0;
3359 	u32 tbl_idx;
3360 	int rc;
3361 
3362 	swr = &udwr.wr;
3363 	tbl_idx = cqe->wr_id;
3364 
3365 	rq_hdr_buf = gsi_qp->qplib_qp.rq_hdr_buf +
3366 			(tbl_idx * gsi_qp->qplib_qp.rq_hdr_buf_size);
3367 	rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3368 							  tbl_idx);
3369 
3370 	/* Shadow QP header buffer */
3371 	shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3372 							    tbl_idx);
3373 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3374 
3375 	/* Store this cqe */
3376 	memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe));
3377 	sqp_entry->qp1_qp = gsi_qp;
3378 
3379 	/* Find packet type from the cqe */
3380 
3381 	pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags,
3382 					     cqe->raweth_qp1_flags2);
3383 	if (pkt_type < 0) {
3384 		ibdev_err(&rdev->ibdev, "Invalid packet\n");
3385 		return -EINVAL;
3386 	}
3387 
3388 	/* Adjust the offset for the user buffer and post in the rq */
3389 
3390 	if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET)
3391 		offset = 20;
3392 
3393 	/*
3394 	 * QP1 loopback packet has 4 bytes of internal header before
3395 	 * ether header. Skip these four bytes.
3396 	 */
3397 	if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf))
3398 		skip_bytes = 4;
3399 
3400 	/* First send SGE . Skip the ether header*/
3401 	s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
3402 			+ skip_bytes;
3403 	s_sge[0].lkey = 0xFFFFFFFF;
3404 	s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 :
3405 				BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
3406 
3407 	/* Second Send SGE */
3408 	s_sge[1].addr = s_sge[0].addr + s_sge[0].length +
3409 			BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE;
3410 	if (pkt_type != BNXT_RE_ROCE_V1_PACKET)
3411 		s_sge[1].addr += 8;
3412 	s_sge[1].lkey = 0xFFFFFFFF;
3413 	s_sge[1].length = 256;
3414 
3415 	/* First recv SGE */
3416 
3417 	r_sge[0].addr = shrq_hdr_buf_map;
3418 	r_sge[0].lkey = 0xFFFFFFFF;
3419 	r_sge[0].length = 40;
3420 
3421 	r_sge[1].addr = sqp_entry->sge.addr + offset;
3422 	r_sge[1].lkey = sqp_entry->sge.lkey;
3423 	r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset;
3424 
3425 	/* Create receive work request */
3426 	rwr.num_sge = 2;
3427 	rwr.sg_list = r_sge;
3428 	rwr.wr_id = tbl_idx;
3429 	rwr.next = NULL;
3430 
3431 	rc = bnxt_re_post_recv_shadow_qp(rdev, gsi_sqp, &rwr);
3432 	if (rc) {
3433 		ibdev_err(&rdev->ibdev,
3434 			  "Failed to post Rx buffers to shadow QP");
3435 		return -ENOMEM;
3436 	}
3437 
3438 	swr->num_sge = 2;
3439 	swr->sg_list = s_sge;
3440 	swr->wr_id = tbl_idx;
3441 	swr->opcode = IB_WR_SEND;
3442 	swr->next = NULL;
3443 	gsi_sah = rdev->gsi_ctx.gsi_sah;
3444 	udwr.ah = &gsi_sah->ib_ah;
3445 	udwr.remote_qpn = gsi_sqp->qplib_qp.id;
3446 	udwr.remote_qkey = gsi_sqp->qplib_qp.qkey;
3447 
3448 	/* post data received  in the send queue */
3449 	return bnxt_re_post_send_shadow_qp(rdev, gsi_sqp, swr);
3450 }
3451 
3452 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc,
3453 					  struct bnxt_qplib_cqe *cqe)
3454 {
3455 	wc->opcode = IB_WC_RECV;
3456 	wc->status = __rawqp1_to_ib_wc_status(cqe->status);
3457 	wc->wc_flags |= IB_WC_GRH;
3458 }
3459 
3460 static bool bnxt_re_check_if_vlan_valid(struct bnxt_re_dev *rdev,
3461 					u16 vlan_id)
3462 {
3463 	/*
3464 	 * Check if the vlan is configured in the host.  If not configured, it
3465 	 * can be a transparent VLAN. So dont report the vlan id.
3466 	 */
3467 	if (!__vlan_find_dev_deep_rcu(rdev->netdev,
3468 				      htons(ETH_P_8021Q), vlan_id))
3469 		return false;
3470 	return true;
3471 }
3472 
3473 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe,
3474 				u16 *vid, u8 *sl)
3475 {
3476 	bool ret = false;
3477 	u32 metadata;
3478 	u16 tpid;
3479 
3480 	metadata = orig_cqe->raweth_qp1_metadata;
3481 	if (orig_cqe->raweth_qp1_flags2 &
3482 		CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) {
3483 		tpid = ((metadata &
3484 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >>
3485 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT);
3486 		if (tpid == ETH_P_8021Q) {
3487 			*vid = metadata &
3488 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK;
3489 			*sl = (metadata &
3490 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >>
3491 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT;
3492 			ret = true;
3493 		}
3494 	}
3495 
3496 	return ret;
3497 }
3498 
3499 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc,
3500 				      struct bnxt_qplib_cqe *cqe)
3501 {
3502 	wc->opcode = IB_WC_RECV;
3503 	wc->status = __rc_to_ib_wc_status(cqe->status);
3504 
3505 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3506 		wc->wc_flags |= IB_WC_WITH_IMM;
3507 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3508 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3509 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3510 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3511 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3512 }
3513 
3514 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *gsi_sqp,
3515 					     struct ib_wc *wc,
3516 					     struct bnxt_qplib_cqe *cqe)
3517 {
3518 	struct bnxt_re_dev *rdev = gsi_sqp->rdev;
3519 	struct bnxt_re_qp *gsi_qp = NULL;
3520 	struct bnxt_qplib_cqe *orig_cqe = NULL;
3521 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3522 	int nw_type;
3523 	u32 tbl_idx;
3524 	u16 vlan_id;
3525 	u8 sl;
3526 
3527 	tbl_idx = cqe->wr_id;
3528 
3529 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3530 	gsi_qp = sqp_entry->qp1_qp;
3531 	orig_cqe = &sqp_entry->cqe;
3532 
3533 	wc->wr_id = sqp_entry->wrid;
3534 	wc->byte_len = orig_cqe->length;
3535 	wc->qp = &gsi_qp->ib_qp;
3536 
3537 	wc->ex.imm_data = orig_cqe->immdata;
3538 	wc->src_qp = orig_cqe->src_qp;
3539 	memcpy(wc->smac, orig_cqe->smac, ETH_ALEN);
3540 	if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) {
3541 		if (bnxt_re_check_if_vlan_valid(rdev, vlan_id)) {
3542 			wc->vlan_id = vlan_id;
3543 			wc->sl = sl;
3544 			wc->wc_flags |= IB_WC_WITH_VLAN;
3545 		}
3546 	}
3547 	wc->port_num = 1;
3548 	wc->vendor_err = orig_cqe->status;
3549 
3550 	wc->opcode = IB_WC_RECV;
3551 	wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status);
3552 	wc->wc_flags |= IB_WC_GRH;
3553 
3554 	nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags,
3555 					    orig_cqe->raweth_qp1_flags2);
3556 	if (nw_type >= 0) {
3557 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3558 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3559 	}
3560 }
3561 
3562 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp,
3563 				      struct ib_wc *wc,
3564 				      struct bnxt_qplib_cqe *cqe)
3565 {
3566 	struct bnxt_re_dev *rdev;
3567 	u16 vlan_id = 0;
3568 	u8 nw_type;
3569 
3570 	rdev = qp->rdev;
3571 	wc->opcode = IB_WC_RECV;
3572 	wc->status = __rc_to_ib_wc_status(cqe->status);
3573 
3574 	if (cqe->flags & CQ_RES_UD_FLAGS_IMM)
3575 		wc->wc_flags |= IB_WC_WITH_IMM;
3576 	/* report only on GSI QP for Thor */
3577 	if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) {
3578 		wc->wc_flags |= IB_WC_GRH;
3579 		memcpy(wc->smac, cqe->smac, ETH_ALEN);
3580 		wc->wc_flags |= IB_WC_WITH_SMAC;
3581 		if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) {
3582 			vlan_id = (cqe->cfa_meta & 0xFFF);
3583 		}
3584 		/* Mark only if vlan_id is non zero */
3585 		if (vlan_id && bnxt_re_check_if_vlan_valid(rdev, vlan_id)) {
3586 			wc->vlan_id = vlan_id;
3587 			wc->wc_flags |= IB_WC_WITH_VLAN;
3588 		}
3589 		nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >>
3590 			   CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT;
3591 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3592 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3593 	}
3594 
3595 }
3596 
3597 static int send_phantom_wqe(struct bnxt_re_qp *qp)
3598 {
3599 	struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
3600 	unsigned long flags;
3601 	int rc;
3602 
3603 	spin_lock_irqsave(&qp->sq_lock, flags);
3604 
3605 	rc = bnxt_re_bind_fence_mw(lib_qp);
3606 	if (!rc) {
3607 		lib_qp->sq.phantom_wqe_cnt++;
3608 		ibdev_dbg(&qp->rdev->ibdev,
3609 			  "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
3610 			  lib_qp->id, lib_qp->sq.hwq.prod,
3611 			  HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
3612 			  lib_qp->sq.phantom_wqe_cnt);
3613 	}
3614 
3615 	spin_unlock_irqrestore(&qp->sq_lock, flags);
3616 	return rc;
3617 }
3618 
3619 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
3620 {
3621 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3622 	struct bnxt_re_qp *qp, *sh_qp;
3623 	struct bnxt_qplib_cqe *cqe;
3624 	int i, ncqe, budget;
3625 	struct bnxt_qplib_q *sq;
3626 	struct bnxt_qplib_qp *lib_qp;
3627 	u32 tbl_idx;
3628 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3629 	unsigned long flags;
3630 
3631 	/* User CQ; the only processing we do is to
3632 	 * complete any pending CQ resize operation.
3633 	 */
3634 	if (cq->umem) {
3635 		if (cq->resize_umem)
3636 			bnxt_re_resize_cq_complete(cq);
3637 		return 0;
3638 	}
3639 
3640 	spin_lock_irqsave(&cq->cq_lock, flags);
3641 	budget = min_t(u32, num_entries, cq->max_cql);
3642 	num_entries = budget;
3643 	if (!cq->cql) {
3644 		ibdev_err(&cq->rdev->ibdev, "POLL CQ : no CQL to use");
3645 		goto exit;
3646 	}
3647 	cqe = &cq->cql[0];
3648 	while (budget) {
3649 		lib_qp = NULL;
3650 		ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
3651 		if (lib_qp) {
3652 			sq = &lib_qp->sq;
3653 			if (sq->send_phantom) {
3654 				qp = container_of(lib_qp,
3655 						  struct bnxt_re_qp, qplib_qp);
3656 				if (send_phantom_wqe(qp) == -ENOMEM)
3657 					ibdev_err(&cq->rdev->ibdev,
3658 						  "Phantom failed! Scheduled to send again\n");
3659 				else
3660 					sq->send_phantom = false;
3661 			}
3662 		}
3663 		if (ncqe < budget)
3664 			ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq,
3665 							      cqe + ncqe,
3666 							      budget - ncqe);
3667 
3668 		if (!ncqe)
3669 			break;
3670 
3671 		for (i = 0; i < ncqe; i++, cqe++) {
3672 			/* Transcribe each qplib_wqe back to ib_wc */
3673 			memset(wc, 0, sizeof(*wc));
3674 
3675 			wc->wr_id = cqe->wr_id;
3676 			wc->byte_len = cqe->length;
3677 			qp = container_of
3678 				((struct bnxt_qplib_qp *)
3679 				 (unsigned long)(cqe->qp_handle),
3680 				 struct bnxt_re_qp, qplib_qp);
3681 			wc->qp = &qp->ib_qp;
3682 			wc->ex.imm_data = cqe->immdata;
3683 			wc->src_qp = cqe->src_qp;
3684 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
3685 			wc->port_num = 1;
3686 			wc->vendor_err = cqe->status;
3687 
3688 			switch (cqe->opcode) {
3689 			case CQ_BASE_CQE_TYPE_REQ:
3690 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3691 				if (sh_qp &&
3692 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3693 					/* Handle this completion with
3694 					 * the stored completion
3695 					 */
3696 					memset(wc, 0, sizeof(*wc));
3697 					continue;
3698 				}
3699 				bnxt_re_process_req_wc(wc, cqe);
3700 				break;
3701 			case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1:
3702 				if (!cqe->status) {
3703 					int rc = 0;
3704 
3705 					rc = bnxt_re_process_raw_qp_pkt_rx
3706 								(qp, cqe);
3707 					if (!rc) {
3708 						memset(wc, 0, sizeof(*wc));
3709 						continue;
3710 					}
3711 					cqe->status = -1;
3712 				}
3713 				/* Errors need not be looped back.
3714 				 * But change the wr_id to the one
3715 				 * stored in the table
3716 				 */
3717 				tbl_idx = cqe->wr_id;
3718 				sqp_entry = &cq->rdev->gsi_ctx.sqp_tbl[tbl_idx];
3719 				wc->wr_id = sqp_entry->wrid;
3720 				bnxt_re_process_res_rawqp1_wc(wc, cqe);
3721 				break;
3722 			case CQ_BASE_CQE_TYPE_RES_RC:
3723 				bnxt_re_process_res_rc_wc(wc, cqe);
3724 				break;
3725 			case CQ_BASE_CQE_TYPE_RES_UD:
3726 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3727 				if (sh_qp &&
3728 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3729 					/* Handle this completion with
3730 					 * the stored completion
3731 					 */
3732 					if (cqe->status) {
3733 						continue;
3734 					} else {
3735 						bnxt_re_process_res_shadow_qp_wc
3736 								(qp, wc, cqe);
3737 						break;
3738 					}
3739 				}
3740 				bnxt_re_process_res_ud_wc(qp, wc, cqe);
3741 				break;
3742 			default:
3743 				ibdev_err(&cq->rdev->ibdev,
3744 					  "POLL CQ : type 0x%x not handled",
3745 					  cqe->opcode);
3746 				continue;
3747 			}
3748 			wc++;
3749 			budget--;
3750 		}
3751 	}
3752 exit:
3753 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3754 	return num_entries - budget;
3755 }
3756 
3757 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq,
3758 			  enum ib_cq_notify_flags ib_cqn_flags)
3759 {
3760 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3761 	int type = 0, rc = 0;
3762 	unsigned long flags;
3763 
3764 	spin_lock_irqsave(&cq->cq_lock, flags);
3765 	/* Trigger on the very next completion */
3766 	if (ib_cqn_flags & IB_CQ_NEXT_COMP)
3767 		type = DBC_DBC_TYPE_CQ_ARMALL;
3768 	/* Trigger on the next solicited completion */
3769 	else if (ib_cqn_flags & IB_CQ_SOLICITED)
3770 		type = DBC_DBC_TYPE_CQ_ARMSE;
3771 
3772 	/* Poll to see if there are missed events */
3773 	if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3774 	    !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) {
3775 		rc = 1;
3776 		goto exit;
3777 	}
3778 	bnxt_qplib_req_notify_cq(&cq->qplib_cq, type);
3779 
3780 exit:
3781 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3782 	return rc;
3783 }
3784 
3785 /* Memory Regions */
3786 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags)
3787 {
3788 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3789 	struct bnxt_re_dev *rdev = pd->rdev;
3790 	struct bnxt_re_mr *mr;
3791 	u32 active_mrs;
3792 	int rc;
3793 
3794 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3795 	if (!mr)
3796 		return ERR_PTR(-ENOMEM);
3797 
3798 	mr->rdev = rdev;
3799 	mr->qplib_mr.pd = &pd->qplib_pd;
3800 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3801 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3802 
3803 	/* Allocate and register 0 as the address */
3804 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3805 	if (rc)
3806 		goto fail;
3807 
3808 	mr->qplib_mr.hwq.level = PBL_LVL_MAX;
3809 	mr->qplib_mr.total_size = -1; /* Infinte length */
3810 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL, 0,
3811 			       PAGE_SIZE);
3812 	if (rc)
3813 		goto fail_mr;
3814 
3815 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3816 	if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ |
3817 			       IB_ACCESS_REMOTE_ATOMIC))
3818 		mr->ib_mr.rkey = mr->ib_mr.lkey;
3819 	active_mrs = atomic_inc_return(&rdev->stats.res.mr_count);
3820 	if (active_mrs > rdev->stats.res.mr_watermark)
3821 		rdev->stats.res.mr_watermark = active_mrs;
3822 
3823 	return &mr->ib_mr;
3824 
3825 fail_mr:
3826 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3827 fail:
3828 	kfree(mr);
3829 	return ERR_PTR(rc);
3830 }
3831 
3832 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3833 {
3834 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3835 	struct bnxt_re_dev *rdev = mr->rdev;
3836 	int rc;
3837 
3838 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3839 	if (rc) {
3840 		ibdev_err(&rdev->ibdev, "Dereg MR failed: %#x\n", rc);
3841 		return rc;
3842 	}
3843 
3844 	if (mr->pages) {
3845 		rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
3846 							&mr->qplib_frpl);
3847 		kfree(mr->pages);
3848 		mr->npages = 0;
3849 		mr->pages = NULL;
3850 	}
3851 	ib_umem_release(mr->ib_umem);
3852 
3853 	kfree(mr);
3854 	atomic_dec(&rdev->stats.res.mr_count);
3855 	return rc;
3856 }
3857 
3858 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr)
3859 {
3860 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3861 
3862 	if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs))
3863 		return -ENOMEM;
3864 
3865 	mr->pages[mr->npages++] = addr;
3866 	return 0;
3867 }
3868 
3869 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents,
3870 		      unsigned int *sg_offset)
3871 {
3872 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3873 
3874 	mr->npages = 0;
3875 	return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page);
3876 }
3877 
3878 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type,
3879 			       u32 max_num_sg)
3880 {
3881 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3882 	struct bnxt_re_dev *rdev = pd->rdev;
3883 	struct bnxt_re_mr *mr = NULL;
3884 	u32 active_mrs;
3885 	int rc;
3886 
3887 	if (type != IB_MR_TYPE_MEM_REG) {
3888 		ibdev_dbg(&rdev->ibdev, "MR type 0x%x not supported", type);
3889 		return ERR_PTR(-EINVAL);
3890 	}
3891 	if (max_num_sg > MAX_PBL_LVL_1_PGS)
3892 		return ERR_PTR(-EINVAL);
3893 
3894 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3895 	if (!mr)
3896 		return ERR_PTR(-ENOMEM);
3897 
3898 	mr->rdev = rdev;
3899 	mr->qplib_mr.pd = &pd->qplib_pd;
3900 	mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR;
3901 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3902 
3903 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3904 	if (rc)
3905 		goto bail;
3906 
3907 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3908 	mr->ib_mr.rkey = mr->ib_mr.lkey;
3909 
3910 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3911 	if (!mr->pages) {
3912 		rc = -ENOMEM;
3913 		goto fail;
3914 	}
3915 	rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res,
3916 						 &mr->qplib_frpl, max_num_sg);
3917 	if (rc) {
3918 		ibdev_err(&rdev->ibdev,
3919 			  "Failed to allocate HW FR page list");
3920 		goto fail_mr;
3921 	}
3922 
3923 	active_mrs = atomic_inc_return(&rdev->stats.res.mr_count);
3924 	if (active_mrs > rdev->stats.res.mr_watermark)
3925 		rdev->stats.res.mr_watermark = active_mrs;
3926 	return &mr->ib_mr;
3927 
3928 fail_mr:
3929 	kfree(mr->pages);
3930 fail:
3931 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3932 bail:
3933 	kfree(mr);
3934 	return ERR_PTR(rc);
3935 }
3936 
3937 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
3938 			       struct ib_udata *udata)
3939 {
3940 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3941 	struct bnxt_re_dev *rdev = pd->rdev;
3942 	struct bnxt_re_mw *mw;
3943 	u32 active_mws;
3944 	int rc;
3945 
3946 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
3947 	if (!mw)
3948 		return ERR_PTR(-ENOMEM);
3949 	mw->rdev = rdev;
3950 	mw->qplib_mw.pd = &pd->qplib_pd;
3951 
3952 	mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
3953 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
3954 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
3955 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
3956 	if (rc) {
3957 		ibdev_err(&rdev->ibdev, "Allocate MW failed!");
3958 		goto fail;
3959 	}
3960 	mw->ib_mw.rkey = mw->qplib_mw.rkey;
3961 
3962 	active_mws = atomic_inc_return(&rdev->stats.res.mw_count);
3963 	if (active_mws > rdev->stats.res.mw_watermark)
3964 		rdev->stats.res.mw_watermark = active_mws;
3965 	return &mw->ib_mw;
3966 
3967 fail:
3968 	kfree(mw);
3969 	return ERR_PTR(rc);
3970 }
3971 
3972 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
3973 {
3974 	struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
3975 	struct bnxt_re_dev *rdev = mw->rdev;
3976 	int rc;
3977 
3978 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
3979 	if (rc) {
3980 		ibdev_err(&rdev->ibdev, "Free MW failed: %#x\n", rc);
3981 		return rc;
3982 	}
3983 
3984 	kfree(mw);
3985 	atomic_dec(&rdev->stats.res.mw_count);
3986 	return rc;
3987 }
3988 
3989 static struct ib_mr *__bnxt_re_user_reg_mr(struct ib_pd *ib_pd, u64 length, u64 virt_addr,
3990 					   int mr_access_flags, struct ib_umem *umem)
3991 {
3992 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3993 	struct bnxt_re_dev *rdev = pd->rdev;
3994 	unsigned long page_size;
3995 	struct bnxt_re_mr *mr;
3996 	int umem_pgs, rc;
3997 	u32 active_mrs;
3998 
3999 	if (length > BNXT_RE_MAX_MR_SIZE) {
4000 		ibdev_err(&rdev->ibdev, "MR Size: %lld > Max supported:%lld\n",
4001 			  length, BNXT_RE_MAX_MR_SIZE);
4002 		return ERR_PTR(-ENOMEM);
4003 	}
4004 
4005 	page_size = ib_umem_find_best_pgsz(umem, BNXT_RE_PAGE_SIZE_SUPPORTED, virt_addr);
4006 	if (!page_size) {
4007 		ibdev_err(&rdev->ibdev, "umem page size unsupported!");
4008 		return ERR_PTR(-EINVAL);
4009 	}
4010 
4011 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
4012 	if (!mr)
4013 		return ERR_PTR(-ENOMEM);
4014 
4015 	mr->rdev = rdev;
4016 	mr->qplib_mr.pd = &pd->qplib_pd;
4017 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
4018 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
4019 
4020 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
4021 	if (rc) {
4022 		ibdev_err(&rdev->ibdev, "Failed to allocate MR rc = %d", rc);
4023 		rc = -EIO;
4024 		goto free_mr;
4025 	}
4026 	/* The fixed portion of the rkey is the same as the lkey */
4027 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
4028 	mr->ib_umem = umem;
4029 	mr->qplib_mr.va = virt_addr;
4030 	mr->qplib_mr.total_size = length;
4031 
4032 	umem_pgs = ib_umem_num_dma_blocks(umem, page_size);
4033 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, umem,
4034 			       umem_pgs, page_size);
4035 	if (rc) {
4036 		ibdev_err(&rdev->ibdev, "Failed to register user MR - rc = %d\n", rc);
4037 		rc = -EIO;
4038 		goto free_mrw;
4039 	}
4040 
4041 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
4042 	mr->ib_mr.rkey = mr->qplib_mr.lkey;
4043 	active_mrs = atomic_inc_return(&rdev->stats.res.mr_count);
4044 	if (active_mrs > rdev->stats.res.mr_watermark)
4045 		rdev->stats.res.mr_watermark = active_mrs;
4046 
4047 	return &mr->ib_mr;
4048 
4049 free_mrw:
4050 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
4051 free_mr:
4052 	kfree(mr);
4053 	return ERR_PTR(rc);
4054 }
4055 
4056 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
4057 				  u64 virt_addr, int mr_access_flags,
4058 				  struct ib_udata *udata)
4059 {
4060 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
4061 	struct bnxt_re_dev *rdev = pd->rdev;
4062 	struct ib_umem *umem;
4063 	struct ib_mr *ib_mr;
4064 
4065 	umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags);
4066 	if (IS_ERR(umem))
4067 		return ERR_CAST(umem);
4068 
4069 	ib_mr = __bnxt_re_user_reg_mr(ib_pd, length, virt_addr, mr_access_flags, umem);
4070 	if (IS_ERR(ib_mr))
4071 		ib_umem_release(umem);
4072 	return ib_mr;
4073 }
4074 
4075 struct ib_mr *bnxt_re_reg_user_mr_dmabuf(struct ib_pd *ib_pd, u64 start,
4076 					 u64 length, u64 virt_addr, int fd,
4077 					 int mr_access_flags, struct ib_udata *udata)
4078 {
4079 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
4080 	struct bnxt_re_dev *rdev = pd->rdev;
4081 	struct ib_umem_dmabuf *umem_dmabuf;
4082 	struct ib_umem *umem;
4083 	struct ib_mr *ib_mr;
4084 
4085 	umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, start, length,
4086 						fd, mr_access_flags);
4087 	if (IS_ERR(umem_dmabuf))
4088 		return ERR_CAST(umem_dmabuf);
4089 
4090 	umem = &umem_dmabuf->umem;
4091 
4092 	ib_mr = __bnxt_re_user_reg_mr(ib_pd, length, virt_addr, mr_access_flags, umem);
4093 	if (IS_ERR(ib_mr))
4094 		ib_umem_release(umem);
4095 	return ib_mr;
4096 }
4097 
4098 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
4099 {
4100 	struct ib_device *ibdev = ctx->device;
4101 	struct bnxt_re_ucontext *uctx =
4102 		container_of(ctx, struct bnxt_re_ucontext, ib_uctx);
4103 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
4104 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
4105 	struct bnxt_re_user_mmap_entry *entry;
4106 	struct bnxt_re_uctx_resp resp = {};
4107 	u32 chip_met_rev_num = 0;
4108 	int rc;
4109 
4110 	ibdev_dbg(ibdev, "ABI version requested %u", ibdev->ops.uverbs_abi_ver);
4111 
4112 	if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) {
4113 		ibdev_dbg(ibdev, " is different from the device %d ",
4114 			  BNXT_RE_ABI_VERSION);
4115 		return -EPERM;
4116 	}
4117 
4118 	uctx->rdev = rdev;
4119 
4120 	uctx->shpg = (void *)__get_free_page(GFP_KERNEL);
4121 	if (!uctx->shpg) {
4122 		rc = -ENOMEM;
4123 		goto fail;
4124 	}
4125 	spin_lock_init(&uctx->sh_lock);
4126 
4127 	resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX;
4128 	chip_met_rev_num = rdev->chip_ctx->chip_num;
4129 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_rev & 0xFF) <<
4130 			     BNXT_RE_CHIP_ID0_CHIP_REV_SFT;
4131 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_metal & 0xFF) <<
4132 			     BNXT_RE_CHIP_ID0_CHIP_MET_SFT;
4133 	resp.chip_id0 = chip_met_rev_num;
4134 	/*Temp, Use xa_alloc instead */
4135 	resp.dev_id = rdev->en_dev->pdev->devfn;
4136 	resp.max_qp = rdev->qplib_ctx.qpc_count;
4137 	resp.pg_size = PAGE_SIZE;
4138 	resp.cqe_sz = sizeof(struct cq_base);
4139 	resp.max_cqd = dev_attr->max_cq_wqes;
4140 
4141 	resp.comp_mask |= BNXT_RE_UCNTX_CMASK_HAVE_MODE;
4142 	resp.mode = rdev->chip_ctx->modes.wqe_mode;
4143 
4144 	if (rdev->chip_ctx->modes.db_push)
4145 		resp.comp_mask |= BNXT_RE_UCNTX_CMASK_WC_DPI_ENABLED;
4146 
4147 	entry = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_SH_PAGE, NULL);
4148 	if (!entry) {
4149 		rc = -ENOMEM;
4150 		goto cfail;
4151 	}
4152 	uctx->shpage_mmap = &entry->rdma_entry;
4153 	if (rdev->pacing.dbr_pacing)
4154 		resp.comp_mask |= BNXT_RE_UCNTX_CMASK_DBR_PACING_ENABLED;
4155 
4156 	rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
4157 	if (rc) {
4158 		ibdev_err(ibdev, "Failed to copy user context");
4159 		rc = -EFAULT;
4160 		goto cfail;
4161 	}
4162 
4163 	return 0;
4164 cfail:
4165 	free_page((unsigned long)uctx->shpg);
4166 	uctx->shpg = NULL;
4167 fail:
4168 	return rc;
4169 }
4170 
4171 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
4172 {
4173 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
4174 						   struct bnxt_re_ucontext,
4175 						   ib_uctx);
4176 
4177 	struct bnxt_re_dev *rdev = uctx->rdev;
4178 
4179 	rdma_user_mmap_entry_remove(uctx->shpage_mmap);
4180 	uctx->shpage_mmap = NULL;
4181 	if (uctx->shpg)
4182 		free_page((unsigned long)uctx->shpg);
4183 
4184 	if (uctx->dpi.dbr) {
4185 		/* Free DPI only if this is the first PD allocated by the
4186 		 * application and mark the context dpi as NULL
4187 		 */
4188 		bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->dpi);
4189 		uctx->dpi.dbr = NULL;
4190 	}
4191 }
4192 
4193 /* Helper function to mmap the virtual memory from user app */
4194 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
4195 {
4196 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
4197 						   struct bnxt_re_ucontext,
4198 						   ib_uctx);
4199 	struct bnxt_re_user_mmap_entry *bnxt_entry;
4200 	struct rdma_user_mmap_entry *rdma_entry;
4201 	int ret = 0;
4202 	u64 pfn;
4203 
4204 	rdma_entry = rdma_user_mmap_entry_get(&uctx->ib_uctx, vma);
4205 	if (!rdma_entry)
4206 		return -EINVAL;
4207 
4208 	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
4209 				  rdma_entry);
4210 
4211 	switch (bnxt_entry->mmap_flag) {
4212 	case BNXT_RE_MMAP_WC_DB:
4213 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4214 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4215 					pgprot_writecombine(vma->vm_page_prot),
4216 					rdma_entry);
4217 		break;
4218 	case BNXT_RE_MMAP_UC_DB:
4219 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4220 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4221 					pgprot_noncached(vma->vm_page_prot),
4222 				rdma_entry);
4223 		break;
4224 	case BNXT_RE_MMAP_SH_PAGE:
4225 		ret = vm_insert_page(vma, vma->vm_start, virt_to_page(uctx->shpg));
4226 		break;
4227 	case BNXT_RE_MMAP_DBR_BAR:
4228 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4229 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4230 					pgprot_noncached(vma->vm_page_prot),
4231 					rdma_entry);
4232 		break;
4233 	case BNXT_RE_MMAP_DBR_PAGE:
4234 		/* Driver doesn't expect write access for user space */
4235 		if (vma->vm_flags & VM_WRITE)
4236 			return -EFAULT;
4237 		ret = vm_insert_page(vma, vma->vm_start,
4238 				     virt_to_page((void *)bnxt_entry->mem_offset));
4239 		break;
4240 	default:
4241 		ret = -EINVAL;
4242 		break;
4243 	}
4244 
4245 	rdma_user_mmap_entry_put(rdma_entry);
4246 	return ret;
4247 }
4248 
4249 void bnxt_re_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
4250 {
4251 	struct bnxt_re_user_mmap_entry *bnxt_entry;
4252 
4253 	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
4254 				  rdma_entry);
4255 
4256 	kfree(bnxt_entry);
4257 }
4258 
4259 static int UVERBS_HANDLER(BNXT_RE_METHOD_NOTIFY_DRV)(struct uverbs_attr_bundle *attrs)
4260 {
4261 	struct bnxt_re_ucontext *uctx;
4262 
4263 	uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx);
4264 	bnxt_re_pacing_alert(uctx->rdev);
4265 	return 0;
4266 }
4267 
4268 static int UVERBS_HANDLER(BNXT_RE_METHOD_ALLOC_PAGE)(struct uverbs_attr_bundle *attrs)
4269 {
4270 	struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
4271 	enum bnxt_re_alloc_page_type alloc_type;
4272 	struct bnxt_re_user_mmap_entry *entry;
4273 	enum bnxt_re_mmap_flag mmap_flag;
4274 	struct bnxt_qplib_chip_ctx *cctx;
4275 	struct bnxt_re_ucontext *uctx;
4276 	struct bnxt_re_dev *rdev;
4277 	u64 mmap_offset;
4278 	u32 length;
4279 	u32 dpi;
4280 	u64 addr;
4281 	int err;
4282 
4283 	uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx);
4284 	if (IS_ERR(uctx))
4285 		return PTR_ERR(uctx);
4286 
4287 	err = uverbs_get_const(&alloc_type, attrs, BNXT_RE_ALLOC_PAGE_TYPE);
4288 	if (err)
4289 		return err;
4290 
4291 	rdev = uctx->rdev;
4292 	cctx = rdev->chip_ctx;
4293 
4294 	switch (alloc_type) {
4295 	case BNXT_RE_ALLOC_WC_PAGE:
4296 		if (cctx->modes.db_push)  {
4297 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res, &uctx->wcdpi,
4298 						 uctx, BNXT_QPLIB_DPI_TYPE_WC))
4299 				return -ENOMEM;
4300 			length = PAGE_SIZE;
4301 			dpi = uctx->wcdpi.dpi;
4302 			addr = (u64)uctx->wcdpi.umdbr;
4303 			mmap_flag = BNXT_RE_MMAP_WC_DB;
4304 		} else {
4305 			return -EINVAL;
4306 		}
4307 
4308 		break;
4309 	case BNXT_RE_ALLOC_DBR_BAR_PAGE:
4310 		length = PAGE_SIZE;
4311 		addr = (u64)rdev->pacing.dbr_bar_addr;
4312 		mmap_flag = BNXT_RE_MMAP_DBR_BAR;
4313 		break;
4314 
4315 	case BNXT_RE_ALLOC_DBR_PAGE:
4316 		length = PAGE_SIZE;
4317 		addr = (u64)rdev->pacing.dbr_page;
4318 		mmap_flag = BNXT_RE_MMAP_DBR_PAGE;
4319 		break;
4320 
4321 	default:
4322 		return -EOPNOTSUPP;
4323 	}
4324 
4325 	entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mmap_offset);
4326 	if (!entry)
4327 		return -ENOMEM;
4328 
4329 	uobj->object = entry;
4330 	uverbs_finalize_uobj_create(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
4331 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
4332 			     &mmap_offset, sizeof(mmap_offset));
4333 	if (err)
4334 		return err;
4335 
4336 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
4337 			     &length, sizeof(length));
4338 	if (err)
4339 		return err;
4340 
4341 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_DPI,
4342 			     &dpi, sizeof(length));
4343 	if (err)
4344 		return err;
4345 
4346 	return 0;
4347 }
4348 
4349 static int alloc_page_obj_cleanup(struct ib_uobject *uobject,
4350 				  enum rdma_remove_reason why,
4351 			    struct uverbs_attr_bundle *attrs)
4352 {
4353 	struct  bnxt_re_user_mmap_entry *entry = uobject->object;
4354 	struct bnxt_re_ucontext *uctx = entry->uctx;
4355 
4356 	switch (entry->mmap_flag) {
4357 	case BNXT_RE_MMAP_WC_DB:
4358 		if (uctx && uctx->wcdpi.dbr) {
4359 			struct bnxt_re_dev *rdev = uctx->rdev;
4360 
4361 			bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->wcdpi);
4362 			uctx->wcdpi.dbr = NULL;
4363 		}
4364 		break;
4365 	case BNXT_RE_MMAP_DBR_BAR:
4366 	case BNXT_RE_MMAP_DBR_PAGE:
4367 		break;
4368 	default:
4369 		goto exit;
4370 	}
4371 	rdma_user_mmap_entry_remove(&entry->rdma_entry);
4372 exit:
4373 	return 0;
4374 }
4375 
4376 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_ALLOC_PAGE,
4377 			    UVERBS_ATTR_IDR(BNXT_RE_ALLOC_PAGE_HANDLE,
4378 					    BNXT_RE_OBJECT_ALLOC_PAGE,
4379 					    UVERBS_ACCESS_NEW,
4380 					    UA_MANDATORY),
4381 			    UVERBS_ATTR_CONST_IN(BNXT_RE_ALLOC_PAGE_TYPE,
4382 						 enum bnxt_re_alloc_page_type,
4383 						 UA_MANDATORY),
4384 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
4385 						UVERBS_ATTR_TYPE(u64),
4386 						UA_MANDATORY),
4387 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
4388 						UVERBS_ATTR_TYPE(u32),
4389 						UA_MANDATORY),
4390 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_DPI,
4391 						UVERBS_ATTR_TYPE(u32),
4392 						UA_MANDATORY));
4393 
4394 DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_DESTROY_PAGE,
4395 				    UVERBS_ATTR_IDR(BNXT_RE_DESTROY_PAGE_HANDLE,
4396 						    BNXT_RE_OBJECT_ALLOC_PAGE,
4397 						    UVERBS_ACCESS_DESTROY,
4398 						    UA_MANDATORY));
4399 
4400 DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_ALLOC_PAGE,
4401 			    UVERBS_TYPE_ALLOC_IDR(alloc_page_obj_cleanup),
4402 			    &UVERBS_METHOD(BNXT_RE_METHOD_ALLOC_PAGE),
4403 			    &UVERBS_METHOD(BNXT_RE_METHOD_DESTROY_PAGE));
4404 
4405 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_NOTIFY_DRV);
4406 
4407 DECLARE_UVERBS_GLOBAL_METHODS(BNXT_RE_OBJECT_NOTIFY_DRV,
4408 			      &UVERBS_METHOD(BNXT_RE_METHOD_NOTIFY_DRV));
4409 
4410 const struct uapi_definition bnxt_re_uapi_defs[] = {
4411 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE),
4412 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_NOTIFY_DRV),
4413 	{}
4414 };
4415