xref: /linux/drivers/infiniband/hw/bnxt_re/ib_verbs.c (revision 0e2b2a76278153d1ac312b0691cb65dabb9aef3e)
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 = 0;
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 		ret = rdma_user_mmap_entry_insert(&uctx->ib_uctx,
569 						  &entry->rdma_entry, PAGE_SIZE);
570 		break;
571 	default:
572 		ret = -EINVAL;
573 		break;
574 	}
575 
576 	if (ret) {
577 		kfree(entry);
578 		return NULL;
579 	}
580 	if (offset)
581 		*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
582 
583 	return entry;
584 }
585 
586 /* Protection Domains */
587 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
588 {
589 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
590 	struct bnxt_re_dev *rdev = pd->rdev;
591 
592 	if (udata) {
593 		rdma_user_mmap_entry_remove(pd->pd_db_mmap);
594 		pd->pd_db_mmap = NULL;
595 	}
596 
597 	bnxt_re_destroy_fence_mr(pd);
598 
599 	if (pd->qplib_pd.id) {
600 		if (!bnxt_qplib_dealloc_pd(&rdev->qplib_res,
601 					   &rdev->qplib_res.pd_tbl,
602 					   &pd->qplib_pd))
603 			atomic_dec(&rdev->pd_count);
604 	}
605 	return 0;
606 }
607 
608 int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
609 {
610 	struct ib_device *ibdev = ibpd->device;
611 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
612 	struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context(
613 		udata, struct bnxt_re_ucontext, ib_uctx);
614 	struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd);
615 	struct bnxt_re_user_mmap_entry *entry = NULL;
616 	int rc = 0;
617 
618 	pd->rdev = rdev;
619 	if (bnxt_qplib_alloc_pd(&rdev->qplib_res.pd_tbl, &pd->qplib_pd)) {
620 		ibdev_err(&rdev->ibdev, "Failed to allocate HW PD");
621 		rc = -ENOMEM;
622 		goto fail;
623 	}
624 
625 	if (udata) {
626 		struct bnxt_re_pd_resp resp = {};
627 
628 		if (!ucntx->dpi.dbr) {
629 			/* Allocate DPI in alloc_pd to avoid failing of
630 			 * ibv_devinfo and family of application when DPIs
631 			 * are depleted.
632 			 */
633 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res,
634 						 &ucntx->dpi, ucntx, BNXT_QPLIB_DPI_TYPE_UC)) {
635 				rc = -ENOMEM;
636 				goto dbfail;
637 			}
638 		}
639 
640 		resp.pdid = pd->qplib_pd.id;
641 		/* Still allow mapping this DBR to the new user PD. */
642 		resp.dpi = ucntx->dpi.dpi;
643 
644 		entry = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->dpi.umdbr,
645 						  BNXT_RE_MMAP_UC_DB, &resp.dbr);
646 
647 		if (!entry) {
648 			rc = -ENOMEM;
649 			goto dbfail;
650 		}
651 
652 		pd->pd_db_mmap = &entry->rdma_entry;
653 
654 		rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
655 		if (rc) {
656 			rdma_user_mmap_entry_remove(pd->pd_db_mmap);
657 			rc = -EFAULT;
658 			goto dbfail;
659 		}
660 	}
661 
662 	if (!udata)
663 		if (bnxt_re_create_fence_mr(pd))
664 			ibdev_warn(&rdev->ibdev,
665 				   "Failed to create Fence-MR\n");
666 	atomic_inc(&rdev->pd_count);
667 
668 	return 0;
669 dbfail:
670 	bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
671 			      &pd->qplib_pd);
672 fail:
673 	return rc;
674 }
675 
676 /* Address Handles */
677 int bnxt_re_destroy_ah(struct ib_ah *ib_ah, u32 flags)
678 {
679 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
680 	struct bnxt_re_dev *rdev = ah->rdev;
681 	bool block = true;
682 	int rc = 0;
683 
684 	block = !(flags & RDMA_DESTROY_AH_SLEEPABLE);
685 	rc = bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah, block);
686 	if (BNXT_RE_CHECK_RC(rc)) {
687 		if (rc == -ETIMEDOUT)
688 			rc = 0;
689 		else
690 			goto fail;
691 	}
692 	atomic_dec(&rdev->ah_count);
693 fail:
694 	return rc;
695 }
696 
697 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)
698 {
699 	u8 nw_type;
700 
701 	switch (ntype) {
702 	case RDMA_NETWORK_IPV4:
703 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4;
704 		break;
705 	case RDMA_NETWORK_IPV6:
706 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6;
707 		break;
708 	default:
709 		nw_type = CMDQ_CREATE_AH_TYPE_V1;
710 		break;
711 	}
712 	return nw_type;
713 }
714 
715 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr,
716 		      struct ib_udata *udata)
717 {
718 	struct ib_pd *ib_pd = ib_ah->pd;
719 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
720 	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
721 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
722 	struct bnxt_re_dev *rdev = pd->rdev;
723 	const struct ib_gid_attr *sgid_attr;
724 	struct bnxt_re_gid_ctx *ctx;
725 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
726 	u8 nw_type;
727 	int rc;
728 
729 	if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) {
730 		ibdev_err(&rdev->ibdev, "Failed to alloc AH: GRH not set");
731 		return -EINVAL;
732 	}
733 
734 	ah->rdev = rdev;
735 	ah->qplib_ah.pd = &pd->qplib_pd;
736 
737 	/* Supply the configuration for the HW */
738 	memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw,
739 	       sizeof(union ib_gid));
740 	sgid_attr = grh->sgid_attr;
741 	/* Get the HW context of the GID. The reference
742 	 * of GID table entry is already taken by the caller.
743 	 */
744 	ctx = rdma_read_gid_hw_context(sgid_attr);
745 	ah->qplib_ah.sgid_index = ctx->idx;
746 	ah->qplib_ah.host_sgid_index = grh->sgid_index;
747 	ah->qplib_ah.traffic_class = grh->traffic_class;
748 	ah->qplib_ah.flow_label = grh->flow_label;
749 	ah->qplib_ah.hop_limit = grh->hop_limit;
750 	ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr);
751 
752 	/* Get network header type for this GID */
753 	nw_type = rdma_gid_attr_network_type(sgid_attr);
754 	ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type);
755 
756 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
757 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah,
758 				  !(init_attr->flags &
759 				    RDMA_CREATE_AH_SLEEPABLE));
760 	if (rc) {
761 		ibdev_err(&rdev->ibdev, "Failed to allocate HW AH");
762 		return rc;
763 	}
764 
765 	/* Write AVID to shared page. */
766 	if (udata) {
767 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
768 			udata, struct bnxt_re_ucontext, ib_uctx);
769 		unsigned long flag;
770 		u32 *wrptr;
771 
772 		spin_lock_irqsave(&uctx->sh_lock, flag);
773 		wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT);
774 		*wrptr = ah->qplib_ah.id;
775 		wmb(); /* make sure cache is updated. */
776 		spin_unlock_irqrestore(&uctx->sh_lock, flag);
777 	}
778 	atomic_inc(&rdev->ah_count);
779 
780 	return 0;
781 }
782 
783 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
784 {
785 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
786 
787 	ah_attr->type = ib_ah->type;
788 	rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl);
789 	memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN);
790 	rdma_ah_set_grh(ah_attr, NULL, 0,
791 			ah->qplib_ah.host_sgid_index,
792 			0, ah->qplib_ah.traffic_class);
793 	rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data);
794 	rdma_ah_set_port_num(ah_attr, 1);
795 	rdma_ah_set_static_rate(ah_attr, 0);
796 	return 0;
797 }
798 
799 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp)
800 	__acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock)
801 {
802 	unsigned long flags;
803 
804 	spin_lock_irqsave(&qp->scq->cq_lock, flags);
805 	if (qp->rcq != qp->scq)
806 		spin_lock(&qp->rcq->cq_lock);
807 	else
808 		__acquire(&qp->rcq->cq_lock);
809 
810 	return flags;
811 }
812 
813 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp,
814 			unsigned long flags)
815 	__releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock)
816 {
817 	if (qp->rcq != qp->scq)
818 		spin_unlock(&qp->rcq->cq_lock);
819 	else
820 		__release(&qp->rcq->cq_lock);
821 	spin_unlock_irqrestore(&qp->scq->cq_lock, flags);
822 }
823 
824 static int bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp *qp)
825 {
826 	struct bnxt_re_qp *gsi_sqp;
827 	struct bnxt_re_ah *gsi_sah;
828 	struct bnxt_re_dev *rdev;
829 	int rc = 0;
830 
831 	rdev = qp->rdev;
832 	gsi_sqp = rdev->gsi_ctx.gsi_sqp;
833 	gsi_sah = rdev->gsi_ctx.gsi_sah;
834 
835 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow AH\n");
836 	bnxt_qplib_destroy_ah(&rdev->qplib_res,
837 			      &gsi_sah->qplib_ah,
838 			      true);
839 	atomic_dec(&rdev->ah_count);
840 	bnxt_qplib_clean_qp(&qp->qplib_qp);
841 
842 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow QP\n");
843 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &gsi_sqp->qplib_qp);
844 	if (rc) {
845 		ibdev_err(&rdev->ibdev, "Destroy Shadow QP failed");
846 		goto fail;
847 	}
848 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &gsi_sqp->qplib_qp);
849 
850 	/* remove from active qp list */
851 	mutex_lock(&rdev->qp_lock);
852 	list_del(&gsi_sqp->list);
853 	mutex_unlock(&rdev->qp_lock);
854 	atomic_dec(&rdev->qp_count);
855 
856 	kfree(rdev->gsi_ctx.sqp_tbl);
857 	kfree(gsi_sah);
858 	kfree(gsi_sqp);
859 	rdev->gsi_ctx.gsi_sqp = NULL;
860 	rdev->gsi_ctx.gsi_sah = NULL;
861 	rdev->gsi_ctx.sqp_tbl = NULL;
862 
863 	return 0;
864 fail:
865 	return rc;
866 }
867 
868 /* Queue Pairs */
869 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
870 {
871 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
872 	struct bnxt_re_dev *rdev = qp->rdev;
873 	unsigned int flags;
874 	int rc;
875 
876 	bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
877 
878 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
879 	if (rc) {
880 		ibdev_err(&rdev->ibdev, "Failed to destroy HW QP");
881 		return rc;
882 	}
883 
884 	if (rdma_is_kernel_res(&qp->ib_qp.res)) {
885 		flags = bnxt_re_lock_cqs(qp);
886 		bnxt_qplib_clean_qp(&qp->qplib_qp);
887 		bnxt_re_unlock_cqs(qp, flags);
888 	}
889 
890 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
891 
892 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) {
893 		rc = bnxt_re_destroy_gsi_sqp(qp);
894 		if (rc)
895 			return rc;
896 	}
897 
898 	mutex_lock(&rdev->qp_lock);
899 	list_del(&qp->list);
900 	mutex_unlock(&rdev->qp_lock);
901 	atomic_dec(&rdev->qp_count);
902 
903 	ib_umem_release(qp->rumem);
904 	ib_umem_release(qp->sumem);
905 
906 	return 0;
907 }
908 
909 static u8 __from_ib_qp_type(enum ib_qp_type type)
910 {
911 	switch (type) {
912 	case IB_QPT_GSI:
913 		return CMDQ_CREATE_QP1_TYPE_GSI;
914 	case IB_QPT_RC:
915 		return CMDQ_CREATE_QP_TYPE_RC;
916 	case IB_QPT_UD:
917 		return CMDQ_CREATE_QP_TYPE_UD;
918 	default:
919 		return IB_QPT_MAX;
920 	}
921 }
922 
923 static u16 bnxt_re_setup_rwqe_size(struct bnxt_qplib_qp *qplqp,
924 				   int rsge, int max)
925 {
926 	if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
927 		rsge = max;
928 	return bnxt_re_get_rwqe_size(rsge);
929 }
930 
931 static u16 bnxt_re_get_wqe_size(int ilsize, int nsge)
932 {
933 	u16 wqe_size, calc_ils;
934 
935 	wqe_size = bnxt_re_get_swqe_size(nsge);
936 	if (ilsize) {
937 		calc_ils = sizeof(struct sq_send_hdr) + ilsize;
938 		wqe_size = max_t(u16, calc_ils, wqe_size);
939 		wqe_size = ALIGN(wqe_size, sizeof(struct sq_send_hdr));
940 	}
941 	return wqe_size;
942 }
943 
944 static int bnxt_re_setup_swqe_size(struct bnxt_re_qp *qp,
945 				   struct ib_qp_init_attr *init_attr)
946 {
947 	struct bnxt_qplib_dev_attr *dev_attr;
948 	struct bnxt_qplib_qp *qplqp;
949 	struct bnxt_re_dev *rdev;
950 	struct bnxt_qplib_q *sq;
951 	int align, ilsize;
952 
953 	rdev = qp->rdev;
954 	qplqp = &qp->qplib_qp;
955 	sq = &qplqp->sq;
956 	dev_attr = &rdev->dev_attr;
957 
958 	align = sizeof(struct sq_send_hdr);
959 	ilsize = ALIGN(init_attr->cap.max_inline_data, align);
960 
961 	sq->wqe_size = bnxt_re_get_wqe_size(ilsize, sq->max_sge);
962 	if (sq->wqe_size > bnxt_re_get_swqe_size(dev_attr->max_qp_sges))
963 		return -EINVAL;
964 	/* For gen p4 and gen p5 backward compatibility mode
965 	 * wqe size is fixed to 128 bytes
966 	 */
967 	if (sq->wqe_size < bnxt_re_get_swqe_size(dev_attr->max_qp_sges) &&
968 			qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
969 		sq->wqe_size = bnxt_re_get_swqe_size(dev_attr->max_qp_sges);
970 
971 	if (init_attr->cap.max_inline_data) {
972 		qplqp->max_inline_data = sq->wqe_size -
973 			sizeof(struct sq_send_hdr);
974 		init_attr->cap.max_inline_data = qplqp->max_inline_data;
975 		if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
976 			sq->max_sge = qplqp->max_inline_data /
977 				sizeof(struct sq_sge);
978 	}
979 
980 	return 0;
981 }
982 
983 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
984 				struct bnxt_re_qp *qp, struct ib_udata *udata)
985 {
986 	struct bnxt_qplib_qp *qplib_qp;
987 	struct bnxt_re_ucontext *cntx;
988 	struct bnxt_re_qp_req ureq;
989 	int bytes = 0, psn_sz;
990 	struct ib_umem *umem;
991 	int psn_nume;
992 
993 	qplib_qp = &qp->qplib_qp;
994 	cntx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext,
995 					 ib_uctx);
996 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
997 		return -EFAULT;
998 
999 	bytes = (qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size);
1000 	/* Consider mapping PSN search memory only for RC QPs. */
1001 	if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) {
1002 		psn_sz = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
1003 						   sizeof(struct sq_psn_search_ext) :
1004 						   sizeof(struct sq_psn_search);
1005 		psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ?
1006 			    qplib_qp->sq.max_wqe :
1007 			    ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) /
1008 			      sizeof(struct bnxt_qplib_sge));
1009 		bytes += (psn_nume * psn_sz);
1010 	}
1011 
1012 	bytes = PAGE_ALIGN(bytes);
1013 	umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes,
1014 			   IB_ACCESS_LOCAL_WRITE);
1015 	if (IS_ERR(umem))
1016 		return PTR_ERR(umem);
1017 
1018 	qp->sumem = umem;
1019 	qplib_qp->sq.sg_info.umem = umem;
1020 	qplib_qp->sq.sg_info.pgsize = PAGE_SIZE;
1021 	qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT;
1022 	qplib_qp->qp_handle = ureq.qp_handle;
1023 
1024 	if (!qp->qplib_qp.srq) {
1025 		bytes = (qplib_qp->rq.max_wqe * qplib_qp->rq.wqe_size);
1026 		bytes = PAGE_ALIGN(bytes);
1027 		umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes,
1028 				   IB_ACCESS_LOCAL_WRITE);
1029 		if (IS_ERR(umem))
1030 			goto rqfail;
1031 		qp->rumem = umem;
1032 		qplib_qp->rq.sg_info.umem = umem;
1033 		qplib_qp->rq.sg_info.pgsize = PAGE_SIZE;
1034 		qplib_qp->rq.sg_info.pgshft = PAGE_SHIFT;
1035 	}
1036 
1037 	qplib_qp->dpi = &cntx->dpi;
1038 	return 0;
1039 rqfail:
1040 	ib_umem_release(qp->sumem);
1041 	qp->sumem = NULL;
1042 	memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info));
1043 
1044 	return PTR_ERR(umem);
1045 }
1046 
1047 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
1048 				(struct bnxt_re_pd *pd,
1049 				 struct bnxt_qplib_res *qp1_res,
1050 				 struct bnxt_qplib_qp *qp1_qp)
1051 {
1052 	struct bnxt_re_dev *rdev = pd->rdev;
1053 	struct bnxt_re_ah *ah;
1054 	union ib_gid sgid;
1055 	int rc;
1056 
1057 	ah = kzalloc(sizeof(*ah), GFP_KERNEL);
1058 	if (!ah)
1059 		return NULL;
1060 
1061 	ah->rdev = rdev;
1062 	ah->qplib_ah.pd = &pd->qplib_pd;
1063 
1064 	rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid);
1065 	if (rc)
1066 		goto fail;
1067 
1068 	/* supply the dgid data same as sgid */
1069 	memcpy(ah->qplib_ah.dgid.data, &sgid.raw,
1070 	       sizeof(union ib_gid));
1071 	ah->qplib_ah.sgid_index = 0;
1072 
1073 	ah->qplib_ah.traffic_class = 0;
1074 	ah->qplib_ah.flow_label = 0;
1075 	ah->qplib_ah.hop_limit = 1;
1076 	ah->qplib_ah.sl = 0;
1077 	/* Have DMAC same as SMAC */
1078 	ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr);
1079 
1080 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false);
1081 	if (rc) {
1082 		ibdev_err(&rdev->ibdev,
1083 			  "Failed to allocate HW AH for Shadow QP");
1084 		goto fail;
1085 	}
1086 	atomic_inc(&rdev->ah_count);
1087 
1088 	return ah;
1089 
1090 fail:
1091 	kfree(ah);
1092 	return NULL;
1093 }
1094 
1095 static struct bnxt_re_qp *bnxt_re_create_shadow_qp
1096 				(struct bnxt_re_pd *pd,
1097 				 struct bnxt_qplib_res *qp1_res,
1098 				 struct bnxt_qplib_qp *qp1_qp)
1099 {
1100 	struct bnxt_re_dev *rdev = pd->rdev;
1101 	struct bnxt_re_qp *qp;
1102 	int rc;
1103 
1104 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1105 	if (!qp)
1106 		return NULL;
1107 
1108 	qp->rdev = rdev;
1109 
1110 	/* Initialize the shadow QP structure from the QP1 values */
1111 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
1112 
1113 	qp->qplib_qp.pd = &pd->qplib_pd;
1114 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
1115 	qp->qplib_qp.type = IB_QPT_UD;
1116 
1117 	qp->qplib_qp.max_inline_data = 0;
1118 	qp->qplib_qp.sig_type = true;
1119 
1120 	/* Shadow QP SQ depth should be same as QP1 RQ depth */
1121 	qp->qplib_qp.sq.wqe_size = bnxt_re_get_wqe_size(0, 6);
1122 	qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
1123 	qp->qplib_qp.sq.max_sge = 2;
1124 	/* Q full delta can be 1 since it is internal QP */
1125 	qp->qplib_qp.sq.q_full_delta = 1;
1126 	qp->qplib_qp.sq.sg_info.pgsize = PAGE_SIZE;
1127 	qp->qplib_qp.sq.sg_info.pgshft = PAGE_SHIFT;
1128 
1129 	qp->qplib_qp.scq = qp1_qp->scq;
1130 	qp->qplib_qp.rcq = qp1_qp->rcq;
1131 
1132 	qp->qplib_qp.rq.wqe_size = bnxt_re_get_rwqe_size(6);
1133 	qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
1134 	qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
1135 	/* Q full delta can be 1 since it is internal QP */
1136 	qp->qplib_qp.rq.q_full_delta = 1;
1137 	qp->qplib_qp.rq.sg_info.pgsize = PAGE_SIZE;
1138 	qp->qplib_qp.rq.sg_info.pgshft = PAGE_SHIFT;
1139 
1140 	qp->qplib_qp.mtu = qp1_qp->mtu;
1141 
1142 	qp->qplib_qp.sq_hdr_buf_size = 0;
1143 	qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
1144 	qp->qplib_qp.dpi = &rdev->dpi_privileged;
1145 
1146 	rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
1147 	if (rc)
1148 		goto fail;
1149 
1150 	spin_lock_init(&qp->sq_lock);
1151 	INIT_LIST_HEAD(&qp->list);
1152 	mutex_lock(&rdev->qp_lock);
1153 	list_add_tail(&qp->list, &rdev->qp_list);
1154 	atomic_inc(&rdev->qp_count);
1155 	mutex_unlock(&rdev->qp_lock);
1156 	return qp;
1157 fail:
1158 	kfree(qp);
1159 	return NULL;
1160 }
1161 
1162 static int bnxt_re_init_rq_attr(struct bnxt_re_qp *qp,
1163 				struct ib_qp_init_attr *init_attr)
1164 {
1165 	struct bnxt_qplib_dev_attr *dev_attr;
1166 	struct bnxt_qplib_qp *qplqp;
1167 	struct bnxt_re_dev *rdev;
1168 	struct bnxt_qplib_q *rq;
1169 	int entries;
1170 
1171 	rdev = qp->rdev;
1172 	qplqp = &qp->qplib_qp;
1173 	rq = &qplqp->rq;
1174 	dev_attr = &rdev->dev_attr;
1175 
1176 	if (init_attr->srq) {
1177 		struct bnxt_re_srq *srq;
1178 
1179 		srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq);
1180 		qplqp->srq = &srq->qplib_srq;
1181 		rq->max_wqe = 0;
1182 	} else {
1183 		rq->max_sge = init_attr->cap.max_recv_sge;
1184 		if (rq->max_sge > dev_attr->max_qp_sges)
1185 			rq->max_sge = dev_attr->max_qp_sges;
1186 		init_attr->cap.max_recv_sge = rq->max_sge;
1187 		rq->wqe_size = bnxt_re_setup_rwqe_size(qplqp, rq->max_sge,
1188 						       dev_attr->max_qp_sges);
1189 		/* Allocate 1 more than what's provided so posting max doesn't
1190 		 * mean empty.
1191 		 */
1192 		entries = roundup_pow_of_two(init_attr->cap.max_recv_wr + 1);
1193 		rq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1194 		rq->q_full_delta = 0;
1195 		rq->sg_info.pgsize = PAGE_SIZE;
1196 		rq->sg_info.pgshft = PAGE_SHIFT;
1197 	}
1198 
1199 	return 0;
1200 }
1201 
1202 static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp)
1203 {
1204 	struct bnxt_qplib_dev_attr *dev_attr;
1205 	struct bnxt_qplib_qp *qplqp;
1206 	struct bnxt_re_dev *rdev;
1207 
1208 	rdev = qp->rdev;
1209 	qplqp = &qp->qplib_qp;
1210 	dev_attr = &rdev->dev_attr;
1211 
1212 	if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) {
1213 		qplqp->rq.max_sge = dev_attr->max_qp_sges;
1214 		if (qplqp->rq.max_sge > dev_attr->max_qp_sges)
1215 			qplqp->rq.max_sge = dev_attr->max_qp_sges;
1216 		qplqp->rq.max_sge = 6;
1217 	}
1218 }
1219 
1220 static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp,
1221 				struct ib_qp_init_attr *init_attr,
1222 				struct ib_udata *udata)
1223 {
1224 	struct bnxt_qplib_dev_attr *dev_attr;
1225 	struct bnxt_qplib_qp *qplqp;
1226 	struct bnxt_re_dev *rdev;
1227 	struct bnxt_qplib_q *sq;
1228 	int entries;
1229 	int diff;
1230 	int rc;
1231 
1232 	rdev = qp->rdev;
1233 	qplqp = &qp->qplib_qp;
1234 	sq = &qplqp->sq;
1235 	dev_attr = &rdev->dev_attr;
1236 
1237 	sq->max_sge = init_attr->cap.max_send_sge;
1238 	if (sq->max_sge > dev_attr->max_qp_sges) {
1239 		sq->max_sge = dev_attr->max_qp_sges;
1240 		init_attr->cap.max_send_sge = sq->max_sge;
1241 	}
1242 
1243 	rc = bnxt_re_setup_swqe_size(qp, init_attr);
1244 	if (rc)
1245 		return rc;
1246 
1247 	entries = init_attr->cap.max_send_wr;
1248 	/* Allocate 128 + 1 more than what's provided */
1249 	diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ?
1250 		0 : BNXT_QPLIB_RESERVED_QP_WRS;
1251 	entries = roundup_pow_of_two(entries + diff + 1);
1252 	sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1);
1253 	sq->q_full_delta = diff + 1;
1254 	/*
1255 	 * Reserving one slot for Phantom WQE. Application can
1256 	 * post one extra entry in this case. But allowing this to avoid
1257 	 * unexpected Queue full condition
1258 	 */
1259 	qplqp->sq.q_full_delta -= 1;
1260 	qplqp->sq.sg_info.pgsize = PAGE_SIZE;
1261 	qplqp->sq.sg_info.pgshft = PAGE_SHIFT;
1262 
1263 	return 0;
1264 }
1265 
1266 static void bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp *qp,
1267 				       struct ib_qp_init_attr *init_attr)
1268 {
1269 	struct bnxt_qplib_dev_attr *dev_attr;
1270 	struct bnxt_qplib_qp *qplqp;
1271 	struct bnxt_re_dev *rdev;
1272 	int entries;
1273 
1274 	rdev = qp->rdev;
1275 	qplqp = &qp->qplib_qp;
1276 	dev_attr = &rdev->dev_attr;
1277 
1278 	if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) {
1279 		entries = roundup_pow_of_two(init_attr->cap.max_send_wr + 1);
1280 		qplqp->sq.max_wqe = min_t(u32, entries,
1281 					  dev_attr->max_qp_wqes + 1);
1282 		qplqp->sq.q_full_delta = qplqp->sq.max_wqe -
1283 			init_attr->cap.max_send_wr;
1284 		qplqp->sq.max_sge++; /* Need one extra sge to put UD header */
1285 		if (qplqp->sq.max_sge > dev_attr->max_qp_sges)
1286 			qplqp->sq.max_sge = dev_attr->max_qp_sges;
1287 	}
1288 }
1289 
1290 static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev,
1291 				struct ib_qp_init_attr *init_attr)
1292 {
1293 	struct bnxt_qplib_chip_ctx *chip_ctx;
1294 	int qptype;
1295 
1296 	chip_ctx = rdev->chip_ctx;
1297 
1298 	qptype = __from_ib_qp_type(init_attr->qp_type);
1299 	if (qptype == IB_QPT_MAX) {
1300 		ibdev_err(&rdev->ibdev, "QP type 0x%x not supported", qptype);
1301 		qptype = -EOPNOTSUPP;
1302 		goto out;
1303 	}
1304 
1305 	if (bnxt_qplib_is_chip_gen_p5(chip_ctx) &&
1306 	    init_attr->qp_type == IB_QPT_GSI)
1307 		qptype = CMDQ_CREATE_QP_TYPE_GSI;
1308 out:
1309 	return qptype;
1310 }
1311 
1312 static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1313 				struct ib_qp_init_attr *init_attr,
1314 				struct ib_udata *udata)
1315 {
1316 	struct bnxt_qplib_dev_attr *dev_attr;
1317 	struct bnxt_qplib_qp *qplqp;
1318 	struct bnxt_re_dev *rdev;
1319 	struct bnxt_re_cq *cq;
1320 	int rc = 0, qptype;
1321 
1322 	rdev = qp->rdev;
1323 	qplqp = &qp->qplib_qp;
1324 	dev_attr = &rdev->dev_attr;
1325 
1326 	/* Setup misc params */
1327 	ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr);
1328 	qplqp->pd = &pd->qplib_pd;
1329 	qplqp->qp_handle = (u64)qplqp;
1330 	qplqp->max_inline_data = init_attr->cap.max_inline_data;
1331 	qplqp->sig_type = ((init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ?
1332 			    true : false);
1333 	qptype = bnxt_re_init_qp_type(rdev, init_attr);
1334 	if (qptype < 0) {
1335 		rc = qptype;
1336 		goto out;
1337 	}
1338 	qplqp->type = (u8)qptype;
1339 	qplqp->wqe_mode = rdev->chip_ctx->modes.wqe_mode;
1340 
1341 	if (init_attr->qp_type == IB_QPT_RC) {
1342 		qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom;
1343 		qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
1344 	}
1345 	qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1346 	qplqp->dpi = &rdev->dpi_privileged; /* Doorbell page */
1347 	if (init_attr->create_flags) {
1348 		ibdev_dbg(&rdev->ibdev,
1349 			  "QP create flags 0x%x not supported",
1350 			  init_attr->create_flags);
1351 		return -EOPNOTSUPP;
1352 	}
1353 
1354 	/* Setup CQs */
1355 	if (init_attr->send_cq) {
1356 		cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq);
1357 		qplqp->scq = &cq->qplib_cq;
1358 		qp->scq = cq;
1359 	}
1360 
1361 	if (init_attr->recv_cq) {
1362 		cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq);
1363 		qplqp->rcq = &cq->qplib_cq;
1364 		qp->rcq = cq;
1365 	}
1366 
1367 	/* Setup RQ/SRQ */
1368 	rc = bnxt_re_init_rq_attr(qp, init_attr);
1369 	if (rc)
1370 		goto out;
1371 	if (init_attr->qp_type == IB_QPT_GSI)
1372 		bnxt_re_adjust_gsi_rq_attr(qp);
1373 
1374 	/* Setup SQ */
1375 	rc = bnxt_re_init_sq_attr(qp, init_attr, udata);
1376 	if (rc)
1377 		goto out;
1378 	if (init_attr->qp_type == IB_QPT_GSI)
1379 		bnxt_re_adjust_gsi_sq_attr(qp, init_attr);
1380 
1381 	if (udata) /* This will update DPI and qp_handle */
1382 		rc = bnxt_re_init_user_qp(rdev, pd, qp, udata);
1383 out:
1384 	return rc;
1385 }
1386 
1387 static int bnxt_re_create_shadow_gsi(struct bnxt_re_qp *qp,
1388 				     struct bnxt_re_pd *pd)
1389 {
1390 	struct bnxt_re_sqp_entries *sqp_tbl;
1391 	struct bnxt_re_dev *rdev;
1392 	struct bnxt_re_qp *sqp;
1393 	struct bnxt_re_ah *sah;
1394 	int rc = 0;
1395 
1396 	rdev = qp->rdev;
1397 	/* Create a shadow QP to handle the QP1 traffic */
1398 	sqp_tbl = kcalloc(BNXT_RE_MAX_GSI_SQP_ENTRIES, sizeof(*sqp_tbl),
1399 			  GFP_KERNEL);
1400 	if (!sqp_tbl)
1401 		return -ENOMEM;
1402 	rdev->gsi_ctx.sqp_tbl = sqp_tbl;
1403 
1404 	sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res, &qp->qplib_qp);
1405 	if (!sqp) {
1406 		rc = -ENODEV;
1407 		ibdev_err(&rdev->ibdev, "Failed to create Shadow QP for QP1");
1408 		goto out;
1409 	}
1410 	rdev->gsi_ctx.gsi_sqp = sqp;
1411 
1412 	sqp->rcq = qp->rcq;
1413 	sqp->scq = qp->scq;
1414 	sah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res,
1415 					  &qp->qplib_qp);
1416 	if (!sah) {
1417 		bnxt_qplib_destroy_qp(&rdev->qplib_res,
1418 				      &sqp->qplib_qp);
1419 		rc = -ENODEV;
1420 		ibdev_err(&rdev->ibdev,
1421 			  "Failed to create AH entry for ShadowQP");
1422 		goto out;
1423 	}
1424 	rdev->gsi_ctx.gsi_sah = sah;
1425 
1426 	return 0;
1427 out:
1428 	kfree(sqp_tbl);
1429 	return rc;
1430 }
1431 
1432 static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1433 				 struct ib_qp_init_attr *init_attr)
1434 {
1435 	struct bnxt_re_dev *rdev;
1436 	struct bnxt_qplib_qp *qplqp;
1437 	int rc = 0;
1438 
1439 	rdev = qp->rdev;
1440 	qplqp = &qp->qplib_qp;
1441 
1442 	qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
1443 	qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
1444 
1445 	rc = bnxt_qplib_create_qp1(&rdev->qplib_res, qplqp);
1446 	if (rc) {
1447 		ibdev_err(&rdev->ibdev, "create HW QP1 failed!");
1448 		goto out;
1449 	}
1450 
1451 	rc = bnxt_re_create_shadow_gsi(qp, pd);
1452 out:
1453 	return rc;
1454 }
1455 
1456 static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev,
1457 				   struct ib_qp_init_attr *init_attr,
1458 				   struct bnxt_qplib_dev_attr *dev_attr)
1459 {
1460 	bool rc = true;
1461 
1462 	if (init_attr->cap.max_send_wr > dev_attr->max_qp_wqes ||
1463 	    init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes ||
1464 	    init_attr->cap.max_send_sge > dev_attr->max_qp_sges ||
1465 	    init_attr->cap.max_recv_sge > dev_attr->max_qp_sges ||
1466 	    init_attr->cap.max_inline_data > dev_attr->max_inline_data) {
1467 		ibdev_err(&rdev->ibdev,
1468 			  "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",
1469 			  init_attr->cap.max_send_wr, dev_attr->max_qp_wqes,
1470 			  init_attr->cap.max_recv_wr, dev_attr->max_qp_wqes,
1471 			  init_attr->cap.max_send_sge, dev_attr->max_qp_sges,
1472 			  init_attr->cap.max_recv_sge, dev_attr->max_qp_sges,
1473 			  init_attr->cap.max_inline_data,
1474 			  dev_attr->max_inline_data);
1475 		rc = false;
1476 	}
1477 	return rc;
1478 }
1479 
1480 int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr,
1481 		      struct ib_udata *udata)
1482 {
1483 	struct ib_pd *ib_pd = ib_qp->pd;
1484 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1485 	struct bnxt_re_dev *rdev = pd->rdev;
1486 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1487 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1488 	int rc;
1489 
1490 	rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr);
1491 	if (!rc) {
1492 		rc = -EINVAL;
1493 		goto fail;
1494 	}
1495 
1496 	qp->rdev = rdev;
1497 	rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, udata);
1498 	if (rc)
1499 		goto fail;
1500 
1501 	if (qp_init_attr->qp_type == IB_QPT_GSI &&
1502 	    !(bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx))) {
1503 		rc = bnxt_re_create_gsi_qp(qp, pd, qp_init_attr);
1504 		if (rc == -ENODEV)
1505 			goto qp_destroy;
1506 		if (rc)
1507 			goto fail;
1508 	} else {
1509 		rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
1510 		if (rc) {
1511 			ibdev_err(&rdev->ibdev, "Failed to create HW QP");
1512 			goto free_umem;
1513 		}
1514 		if (udata) {
1515 			struct bnxt_re_qp_resp resp;
1516 
1517 			resp.qpid = qp->qplib_qp.id;
1518 			resp.rsvd = 0;
1519 			rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1520 			if (rc) {
1521 				ibdev_err(&rdev->ibdev, "Failed to copy QP udata");
1522 				goto qp_destroy;
1523 			}
1524 		}
1525 	}
1526 
1527 	qp->ib_qp.qp_num = qp->qplib_qp.id;
1528 	if (qp_init_attr->qp_type == IB_QPT_GSI)
1529 		rdev->gsi_ctx.gsi_qp = qp;
1530 	spin_lock_init(&qp->sq_lock);
1531 	spin_lock_init(&qp->rq_lock);
1532 	INIT_LIST_HEAD(&qp->list);
1533 	mutex_lock(&rdev->qp_lock);
1534 	list_add_tail(&qp->list, &rdev->qp_list);
1535 	mutex_unlock(&rdev->qp_lock);
1536 	atomic_inc(&rdev->qp_count);
1537 
1538 	return 0;
1539 qp_destroy:
1540 	bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
1541 free_umem:
1542 	ib_umem_release(qp->rumem);
1543 	ib_umem_release(qp->sumem);
1544 fail:
1545 	return rc;
1546 }
1547 
1548 static u8 __from_ib_qp_state(enum ib_qp_state state)
1549 {
1550 	switch (state) {
1551 	case IB_QPS_RESET:
1552 		return CMDQ_MODIFY_QP_NEW_STATE_RESET;
1553 	case IB_QPS_INIT:
1554 		return CMDQ_MODIFY_QP_NEW_STATE_INIT;
1555 	case IB_QPS_RTR:
1556 		return CMDQ_MODIFY_QP_NEW_STATE_RTR;
1557 	case IB_QPS_RTS:
1558 		return CMDQ_MODIFY_QP_NEW_STATE_RTS;
1559 	case IB_QPS_SQD:
1560 		return CMDQ_MODIFY_QP_NEW_STATE_SQD;
1561 	case IB_QPS_SQE:
1562 		return CMDQ_MODIFY_QP_NEW_STATE_SQE;
1563 	case IB_QPS_ERR:
1564 	default:
1565 		return CMDQ_MODIFY_QP_NEW_STATE_ERR;
1566 	}
1567 }
1568 
1569 static enum ib_qp_state __to_ib_qp_state(u8 state)
1570 {
1571 	switch (state) {
1572 	case CMDQ_MODIFY_QP_NEW_STATE_RESET:
1573 		return IB_QPS_RESET;
1574 	case CMDQ_MODIFY_QP_NEW_STATE_INIT:
1575 		return IB_QPS_INIT;
1576 	case CMDQ_MODIFY_QP_NEW_STATE_RTR:
1577 		return IB_QPS_RTR;
1578 	case CMDQ_MODIFY_QP_NEW_STATE_RTS:
1579 		return IB_QPS_RTS;
1580 	case CMDQ_MODIFY_QP_NEW_STATE_SQD:
1581 		return IB_QPS_SQD;
1582 	case CMDQ_MODIFY_QP_NEW_STATE_SQE:
1583 		return IB_QPS_SQE;
1584 	case CMDQ_MODIFY_QP_NEW_STATE_ERR:
1585 	default:
1586 		return IB_QPS_ERR;
1587 	}
1588 }
1589 
1590 static u32 __from_ib_mtu(enum ib_mtu mtu)
1591 {
1592 	switch (mtu) {
1593 	case IB_MTU_256:
1594 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_256;
1595 	case IB_MTU_512:
1596 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_512;
1597 	case IB_MTU_1024:
1598 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024;
1599 	case IB_MTU_2048:
1600 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1601 	case IB_MTU_4096:
1602 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096;
1603 	default:
1604 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1605 	}
1606 }
1607 
1608 static enum ib_mtu __to_ib_mtu(u32 mtu)
1609 {
1610 	switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) {
1611 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_256:
1612 		return IB_MTU_256;
1613 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_512:
1614 		return IB_MTU_512;
1615 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024:
1616 		return IB_MTU_1024;
1617 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048:
1618 		return IB_MTU_2048;
1619 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096:
1620 		return IB_MTU_4096;
1621 	default:
1622 		return IB_MTU_2048;
1623 	}
1624 }
1625 
1626 /* Shared Receive Queues */
1627 int bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata)
1628 {
1629 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1630 					       ib_srq);
1631 	struct bnxt_re_dev *rdev = srq->rdev;
1632 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1633 	struct bnxt_qplib_nq *nq = NULL;
1634 
1635 	if (qplib_srq->cq)
1636 		nq = qplib_srq->cq->nq;
1637 	bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
1638 	ib_umem_release(srq->umem);
1639 	atomic_dec(&rdev->srq_count);
1640 	if (nq)
1641 		nq->budget--;
1642 	return 0;
1643 }
1644 
1645 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
1646 				 struct bnxt_re_pd *pd,
1647 				 struct bnxt_re_srq *srq,
1648 				 struct ib_udata *udata)
1649 {
1650 	struct bnxt_re_srq_req ureq;
1651 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1652 	struct ib_umem *umem;
1653 	int bytes = 0;
1654 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
1655 		udata, struct bnxt_re_ucontext, ib_uctx);
1656 
1657 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1658 		return -EFAULT;
1659 
1660 	bytes = (qplib_srq->max_wqe * qplib_srq->wqe_size);
1661 	bytes = PAGE_ALIGN(bytes);
1662 	umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes,
1663 			   IB_ACCESS_LOCAL_WRITE);
1664 	if (IS_ERR(umem))
1665 		return PTR_ERR(umem);
1666 
1667 	srq->umem = umem;
1668 	qplib_srq->sg_info.umem = umem;
1669 	qplib_srq->sg_info.pgsize = PAGE_SIZE;
1670 	qplib_srq->sg_info.pgshft = PAGE_SHIFT;
1671 	qplib_srq->srq_handle = ureq.srq_handle;
1672 	qplib_srq->dpi = &cntx->dpi;
1673 
1674 	return 0;
1675 }
1676 
1677 int bnxt_re_create_srq(struct ib_srq *ib_srq,
1678 		       struct ib_srq_init_attr *srq_init_attr,
1679 		       struct ib_udata *udata)
1680 {
1681 	struct bnxt_qplib_dev_attr *dev_attr;
1682 	struct bnxt_qplib_nq *nq = NULL;
1683 	struct bnxt_re_dev *rdev;
1684 	struct bnxt_re_srq *srq;
1685 	struct bnxt_re_pd *pd;
1686 	struct ib_pd *ib_pd;
1687 	int rc, entries;
1688 
1689 	ib_pd = ib_srq->pd;
1690 	pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1691 	rdev = pd->rdev;
1692 	dev_attr = &rdev->dev_attr;
1693 	srq = container_of(ib_srq, struct bnxt_re_srq, ib_srq);
1694 
1695 	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
1696 		ibdev_err(&rdev->ibdev, "Create CQ failed - max exceeded");
1697 		rc = -EINVAL;
1698 		goto exit;
1699 	}
1700 
1701 	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
1702 		rc = -EOPNOTSUPP;
1703 		goto exit;
1704 	}
1705 
1706 	srq->rdev = rdev;
1707 	srq->qplib_srq.pd = &pd->qplib_pd;
1708 	srq->qplib_srq.dpi = &rdev->dpi_privileged;
1709 	/* Allocate 1 more than what's provided so posting max doesn't
1710 	 * mean empty
1711 	 */
1712 	entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
1713 	if (entries > dev_attr->max_srq_wqes + 1)
1714 		entries = dev_attr->max_srq_wqes + 1;
1715 	srq->qplib_srq.max_wqe = entries;
1716 
1717 	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
1718 	 /* 128 byte wqe size for SRQ . So use max sges */
1719 	srq->qplib_srq.wqe_size = bnxt_re_get_rwqe_size(dev_attr->max_srq_sges);
1720 	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
1721 	srq->srq_limit = srq_init_attr->attr.srq_limit;
1722 	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
1723 	nq = &rdev->nq[0];
1724 
1725 	if (udata) {
1726 		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
1727 		if (rc)
1728 			goto fail;
1729 	}
1730 
1731 	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
1732 	if (rc) {
1733 		ibdev_err(&rdev->ibdev, "Create HW SRQ failed!");
1734 		goto fail;
1735 	}
1736 
1737 	if (udata) {
1738 		struct bnxt_re_srq_resp resp;
1739 
1740 		resp.srqid = srq->qplib_srq.id;
1741 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1742 		if (rc) {
1743 			ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!");
1744 			bnxt_qplib_destroy_srq(&rdev->qplib_res,
1745 					       &srq->qplib_srq);
1746 			goto fail;
1747 		}
1748 	}
1749 	if (nq)
1750 		nq->budget++;
1751 	atomic_inc(&rdev->srq_count);
1752 	spin_lock_init(&srq->lock);
1753 
1754 	return 0;
1755 
1756 fail:
1757 	ib_umem_release(srq->umem);
1758 exit:
1759 	return rc;
1760 }
1761 
1762 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
1763 		       enum ib_srq_attr_mask srq_attr_mask,
1764 		       struct ib_udata *udata)
1765 {
1766 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1767 					       ib_srq);
1768 	struct bnxt_re_dev *rdev = srq->rdev;
1769 	int rc;
1770 
1771 	switch (srq_attr_mask) {
1772 	case IB_SRQ_MAX_WR:
1773 		/* SRQ resize is not supported */
1774 		break;
1775 	case IB_SRQ_LIMIT:
1776 		/* Change the SRQ threshold */
1777 		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
1778 			return -EINVAL;
1779 
1780 		srq->qplib_srq.threshold = srq_attr->srq_limit;
1781 		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
1782 		if (rc) {
1783 			ibdev_err(&rdev->ibdev, "Modify HW SRQ failed!");
1784 			return rc;
1785 		}
1786 		/* On success, update the shadow */
1787 		srq->srq_limit = srq_attr->srq_limit;
1788 		/* No need to Build and send response back to udata */
1789 		break;
1790 	default:
1791 		ibdev_err(&rdev->ibdev,
1792 			  "Unsupported srq_attr_mask 0x%x", srq_attr_mask);
1793 		return -EINVAL;
1794 	}
1795 	return 0;
1796 }
1797 
1798 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
1799 {
1800 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1801 					       ib_srq);
1802 	struct bnxt_re_srq tsrq;
1803 	struct bnxt_re_dev *rdev = srq->rdev;
1804 	int rc;
1805 
1806 	/* Get live SRQ attr */
1807 	tsrq.qplib_srq.id = srq->qplib_srq.id;
1808 	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
1809 	if (rc) {
1810 		ibdev_err(&rdev->ibdev, "Query HW SRQ failed!");
1811 		return rc;
1812 	}
1813 	srq_attr->max_wr = srq->qplib_srq.max_wqe;
1814 	srq_attr->max_sge = srq->qplib_srq.max_sge;
1815 	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
1816 
1817 	return 0;
1818 }
1819 
1820 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr,
1821 			  const struct ib_recv_wr **bad_wr)
1822 {
1823 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1824 					       ib_srq);
1825 	struct bnxt_qplib_swqe wqe;
1826 	unsigned long flags;
1827 	int rc = 0;
1828 
1829 	spin_lock_irqsave(&srq->lock, flags);
1830 	while (wr) {
1831 		/* Transcribe each ib_recv_wr to qplib_swqe */
1832 		wqe.num_sge = wr->num_sge;
1833 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
1834 		wqe.wr_id = wr->wr_id;
1835 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
1836 
1837 		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
1838 		if (rc) {
1839 			*bad_wr = wr;
1840 			break;
1841 		}
1842 		wr = wr->next;
1843 	}
1844 	spin_unlock_irqrestore(&srq->lock, flags);
1845 
1846 	return rc;
1847 }
1848 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
1849 				    struct bnxt_re_qp *qp1_qp,
1850 				    int qp_attr_mask)
1851 {
1852 	struct bnxt_re_qp *qp = rdev->gsi_ctx.gsi_sqp;
1853 	int rc = 0;
1854 
1855 	if (qp_attr_mask & IB_QP_STATE) {
1856 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1857 		qp->qplib_qp.state = qp1_qp->qplib_qp.state;
1858 	}
1859 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1860 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1861 		qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index;
1862 	}
1863 
1864 	if (qp_attr_mask & IB_QP_QKEY) {
1865 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1866 		/* Using a Random  QKEY */
1867 		qp->qplib_qp.qkey = 0x81818181;
1868 	}
1869 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1870 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1871 		qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn;
1872 	}
1873 
1874 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1875 	if (rc)
1876 		ibdev_err(&rdev->ibdev, "Failed to modify Shadow QP for QP1");
1877 	return rc;
1878 }
1879 
1880 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1881 		      int qp_attr_mask, struct ib_udata *udata)
1882 {
1883 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1884 	struct bnxt_re_dev *rdev = qp->rdev;
1885 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1886 	enum ib_qp_state curr_qp_state, new_qp_state;
1887 	int rc, entries;
1888 	unsigned int flags;
1889 	u8 nw_type;
1890 
1891 	if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1892 		return -EOPNOTSUPP;
1893 
1894 	qp->qplib_qp.modify_flags = 0;
1895 	if (qp_attr_mask & IB_QP_STATE) {
1896 		curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state);
1897 		new_qp_state = qp_attr->qp_state;
1898 		if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state,
1899 					ib_qp->qp_type, qp_attr_mask)) {
1900 			ibdev_err(&rdev->ibdev,
1901 				  "Invalid attribute mask: %#x specified ",
1902 				  qp_attr_mask);
1903 			ibdev_err(&rdev->ibdev,
1904 				  "for qpn: %#x type: %#x",
1905 				  ib_qp->qp_num, ib_qp->qp_type);
1906 			ibdev_err(&rdev->ibdev,
1907 				  "curr_qp_state=0x%x, new_qp_state=0x%x\n",
1908 				  curr_qp_state, new_qp_state);
1909 			return -EINVAL;
1910 		}
1911 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1912 		qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state);
1913 
1914 		if (!qp->sumem &&
1915 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1916 			ibdev_dbg(&rdev->ibdev,
1917 				  "Move QP = %p to flush list\n", qp);
1918 			flags = bnxt_re_lock_cqs(qp);
1919 			bnxt_qplib_add_flush_qp(&qp->qplib_qp);
1920 			bnxt_re_unlock_cqs(qp, flags);
1921 		}
1922 		if (!qp->sumem &&
1923 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) {
1924 			ibdev_dbg(&rdev->ibdev,
1925 				  "Move QP = %p out of flush list\n", qp);
1926 			flags = bnxt_re_lock_cqs(qp);
1927 			bnxt_qplib_clean_qp(&qp->qplib_qp);
1928 			bnxt_re_unlock_cqs(qp, flags);
1929 		}
1930 	}
1931 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1932 		qp->qplib_qp.modify_flags |=
1933 				CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY;
1934 		qp->qplib_qp.en_sqd_async_notify = true;
1935 	}
1936 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
1937 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS;
1938 		qp->qplib_qp.access =
1939 			__from_ib_access_flags(qp_attr->qp_access_flags);
1940 		/* LOCAL_WRITE access must be set to allow RC receive */
1941 		qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
1942 		/* Temp: Set all params on QP as of now */
1943 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE;
1944 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ;
1945 	}
1946 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1947 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1948 		qp->qplib_qp.pkey_index = qp_attr->pkey_index;
1949 	}
1950 	if (qp_attr_mask & IB_QP_QKEY) {
1951 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1952 		qp->qplib_qp.qkey = qp_attr->qkey;
1953 	}
1954 	if (qp_attr_mask & IB_QP_AV) {
1955 		const struct ib_global_route *grh =
1956 			rdma_ah_read_grh(&qp_attr->ah_attr);
1957 		const struct ib_gid_attr *sgid_attr;
1958 		struct bnxt_re_gid_ctx *ctx;
1959 
1960 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID |
1961 				     CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL |
1962 				     CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX |
1963 				     CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT |
1964 				     CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS |
1965 				     CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC |
1966 				     CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID;
1967 		memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw,
1968 		       sizeof(qp->qplib_qp.ah.dgid.data));
1969 		qp->qplib_qp.ah.flow_label = grh->flow_label;
1970 		sgid_attr = grh->sgid_attr;
1971 		/* Get the HW context of the GID. The reference
1972 		 * of GID table entry is already taken by the caller.
1973 		 */
1974 		ctx = rdma_read_gid_hw_context(sgid_attr);
1975 		qp->qplib_qp.ah.sgid_index = ctx->idx;
1976 		qp->qplib_qp.ah.host_sgid_index = grh->sgid_index;
1977 		qp->qplib_qp.ah.hop_limit = grh->hop_limit;
1978 		qp->qplib_qp.ah.traffic_class = grh->traffic_class;
1979 		qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr);
1980 		ether_addr_copy(qp->qplib_qp.ah.dmac,
1981 				qp_attr->ah_attr.roce.dmac);
1982 
1983 		rc = rdma_read_gid_l2_fields(sgid_attr, NULL,
1984 					     &qp->qplib_qp.smac[0]);
1985 		if (rc)
1986 			return rc;
1987 
1988 		nw_type = rdma_gid_attr_network_type(sgid_attr);
1989 		switch (nw_type) {
1990 		case RDMA_NETWORK_IPV4:
1991 			qp->qplib_qp.nw_type =
1992 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4;
1993 			break;
1994 		case RDMA_NETWORK_IPV6:
1995 			qp->qplib_qp.nw_type =
1996 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6;
1997 			break;
1998 		default:
1999 			qp->qplib_qp.nw_type =
2000 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1;
2001 			break;
2002 		}
2003 	}
2004 
2005 	if (qp_attr_mask & IB_QP_PATH_MTU) {
2006 		qp->qplib_qp.modify_flags |=
2007 				CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
2008 		qp->qplib_qp.path_mtu = __from_ib_mtu(qp_attr->path_mtu);
2009 		qp->qplib_qp.mtu = ib_mtu_enum_to_int(qp_attr->path_mtu);
2010 	} else if (qp_attr->qp_state == IB_QPS_RTR) {
2011 		qp->qplib_qp.modify_flags |=
2012 			CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
2013 		qp->qplib_qp.path_mtu =
2014 			__from_ib_mtu(iboe_get_mtu(rdev->netdev->mtu));
2015 		qp->qplib_qp.mtu =
2016 			ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
2017 	}
2018 
2019 	if (qp_attr_mask & IB_QP_TIMEOUT) {
2020 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT;
2021 		qp->qplib_qp.timeout = qp_attr->timeout;
2022 	}
2023 	if (qp_attr_mask & IB_QP_RETRY_CNT) {
2024 		qp->qplib_qp.modify_flags |=
2025 				CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT;
2026 		qp->qplib_qp.retry_cnt = qp_attr->retry_cnt;
2027 	}
2028 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
2029 		qp->qplib_qp.modify_flags |=
2030 				CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY;
2031 		qp->qplib_qp.rnr_retry = qp_attr->rnr_retry;
2032 	}
2033 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) {
2034 		qp->qplib_qp.modify_flags |=
2035 				CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER;
2036 		qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer;
2037 	}
2038 	if (qp_attr_mask & IB_QP_RQ_PSN) {
2039 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN;
2040 		qp->qplib_qp.rq.psn = qp_attr->rq_psn;
2041 	}
2042 	if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2043 		qp->qplib_qp.modify_flags |=
2044 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC;
2045 		/* Cap the max_rd_atomic to device max */
2046 		qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic,
2047 						   dev_attr->max_qp_rd_atom);
2048 	}
2049 	if (qp_attr_mask & IB_QP_SQ_PSN) {
2050 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
2051 		qp->qplib_qp.sq.psn = qp_attr->sq_psn;
2052 	}
2053 	if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2054 		if (qp_attr->max_dest_rd_atomic >
2055 		    dev_attr->max_qp_init_rd_atom) {
2056 			ibdev_err(&rdev->ibdev,
2057 				  "max_dest_rd_atomic requested%d is > dev_max%d",
2058 				  qp_attr->max_dest_rd_atomic,
2059 				  dev_attr->max_qp_init_rd_atom);
2060 			return -EINVAL;
2061 		}
2062 
2063 		qp->qplib_qp.modify_flags |=
2064 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC;
2065 		qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic;
2066 	}
2067 	if (qp_attr_mask & IB_QP_CAP) {
2068 		qp->qplib_qp.modify_flags |=
2069 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE |
2070 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE |
2071 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE |
2072 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE |
2073 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA;
2074 		if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) ||
2075 		    (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) ||
2076 		    (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) ||
2077 		    (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) ||
2078 		    (qp_attr->cap.max_inline_data >=
2079 						dev_attr->max_inline_data)) {
2080 			ibdev_err(&rdev->ibdev,
2081 				  "Create QP failed - max exceeded");
2082 			return -EINVAL;
2083 		}
2084 		entries = roundup_pow_of_two(qp_attr->cap.max_send_wr);
2085 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
2086 						dev_attr->max_qp_wqes + 1);
2087 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
2088 						qp_attr->cap.max_send_wr;
2089 		/*
2090 		 * Reserving one slot for Phantom WQE. Some application can
2091 		 * post one extra entry in this case. Allowing this to avoid
2092 		 * unexpected Queue full condition
2093 		 */
2094 		qp->qplib_qp.sq.q_full_delta -= 1;
2095 		qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
2096 		if (qp->qplib_qp.rq.max_wqe) {
2097 			entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr);
2098 			qp->qplib_qp.rq.max_wqe =
2099 				min_t(u32, entries, dev_attr->max_qp_wqes + 1);
2100 			qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
2101 						       qp_attr->cap.max_recv_wr;
2102 			qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
2103 		} else {
2104 			/* SRQ was used prior, just ignore the RQ caps */
2105 		}
2106 	}
2107 	if (qp_attr_mask & IB_QP_DEST_QPN) {
2108 		qp->qplib_qp.modify_flags |=
2109 				CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID;
2110 		qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num;
2111 	}
2112 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
2113 	if (rc) {
2114 		ibdev_err(&rdev->ibdev, "Failed to modify HW QP");
2115 		return rc;
2116 	}
2117 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp)
2118 		rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask);
2119 	return rc;
2120 }
2121 
2122 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
2123 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
2124 {
2125 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2126 	struct bnxt_re_dev *rdev = qp->rdev;
2127 	struct bnxt_qplib_qp *qplib_qp;
2128 	int rc;
2129 
2130 	qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL);
2131 	if (!qplib_qp)
2132 		return -ENOMEM;
2133 
2134 	qplib_qp->id = qp->qplib_qp.id;
2135 	qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index;
2136 
2137 	rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp);
2138 	if (rc) {
2139 		ibdev_err(&rdev->ibdev, "Failed to query HW QP");
2140 		goto out;
2141 	}
2142 	qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state);
2143 	qp_attr->cur_qp_state = __to_ib_qp_state(qplib_qp->cur_qp_state);
2144 	qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0;
2145 	qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access);
2146 	qp_attr->pkey_index = qplib_qp->pkey_index;
2147 	qp_attr->qkey = qplib_qp->qkey;
2148 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2149 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label,
2150 			qplib_qp->ah.host_sgid_index,
2151 			qplib_qp->ah.hop_limit,
2152 			qplib_qp->ah.traffic_class);
2153 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data);
2154 	rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl);
2155 	ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac);
2156 	qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu);
2157 	qp_attr->timeout = qplib_qp->timeout;
2158 	qp_attr->retry_cnt = qplib_qp->retry_cnt;
2159 	qp_attr->rnr_retry = qplib_qp->rnr_retry;
2160 	qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer;
2161 	qp_attr->rq_psn = qplib_qp->rq.psn;
2162 	qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic;
2163 	qp_attr->sq_psn = qplib_qp->sq.psn;
2164 	qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic;
2165 	qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR :
2166 							 IB_SIGNAL_REQ_WR;
2167 	qp_attr->dest_qp_num = qplib_qp->dest_qpn;
2168 
2169 	qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe;
2170 	qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge;
2171 	qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe;
2172 	qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge;
2173 	qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data;
2174 	qp_init_attr->cap = qp_attr->cap;
2175 
2176 out:
2177 	kfree(qplib_qp);
2178 	return rc;
2179 }
2180 
2181 /* Routine for sending QP1 packets for RoCE V1 an V2
2182  */
2183 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp,
2184 				     const struct ib_send_wr *wr,
2185 				     struct bnxt_qplib_swqe *wqe,
2186 				     int payload_size)
2187 {
2188 	struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah,
2189 					     ib_ah);
2190 	struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah;
2191 	const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr;
2192 	struct bnxt_qplib_sge sge;
2193 	u8 nw_type;
2194 	u16 ether_type;
2195 	union ib_gid dgid;
2196 	bool is_eth = false;
2197 	bool is_vlan = false;
2198 	bool is_grh = false;
2199 	bool is_udp = false;
2200 	u8 ip_version = 0;
2201 	u16 vlan_id = 0xFFFF;
2202 	void *buf;
2203 	int i, rc = 0;
2204 
2205 	memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr));
2206 
2207 	rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL);
2208 	if (rc)
2209 		return rc;
2210 
2211 	/* Get network header type for this GID */
2212 	nw_type = rdma_gid_attr_network_type(sgid_attr);
2213 	switch (nw_type) {
2214 	case RDMA_NETWORK_IPV4:
2215 		nw_type = BNXT_RE_ROCEV2_IPV4_PACKET;
2216 		break;
2217 	case RDMA_NETWORK_IPV6:
2218 		nw_type = BNXT_RE_ROCEV2_IPV6_PACKET;
2219 		break;
2220 	default:
2221 		nw_type = BNXT_RE_ROCE_V1_PACKET;
2222 		break;
2223 	}
2224 	memcpy(&dgid.raw, &qplib_ah->dgid, 16);
2225 	is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2226 	if (is_udp) {
2227 		if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) {
2228 			ip_version = 4;
2229 			ether_type = ETH_P_IP;
2230 		} else {
2231 			ip_version = 6;
2232 			ether_type = ETH_P_IPV6;
2233 		}
2234 		is_grh = false;
2235 	} else {
2236 		ether_type = ETH_P_IBOE;
2237 		is_grh = true;
2238 	}
2239 
2240 	is_eth = true;
2241 	is_vlan = (vlan_id && (vlan_id < 0x1000)) ? true : false;
2242 
2243 	ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh,
2244 			  ip_version, is_udp, 0, &qp->qp1_hdr);
2245 
2246 	/* ETH */
2247 	ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac);
2248 	ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac);
2249 
2250 	/* For vlan, check the sgid for vlan existence */
2251 
2252 	if (!is_vlan) {
2253 		qp->qp1_hdr.eth.type = cpu_to_be16(ether_type);
2254 	} else {
2255 		qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type);
2256 		qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id);
2257 	}
2258 
2259 	if (is_grh || (ip_version == 6)) {
2260 		memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw,
2261 		       sizeof(sgid_attr->gid));
2262 		memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data,
2263 		       sizeof(sgid_attr->gid));
2264 		qp->qp1_hdr.grh.hop_limit     = qplib_ah->hop_limit;
2265 	}
2266 
2267 	if (ip_version == 4) {
2268 		qp->qp1_hdr.ip4.tos = 0;
2269 		qp->qp1_hdr.ip4.id = 0;
2270 		qp->qp1_hdr.ip4.frag_off = htons(IP_DF);
2271 		qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit;
2272 
2273 		memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4);
2274 		memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4);
2275 		qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr);
2276 	}
2277 
2278 	if (is_udp) {
2279 		qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT);
2280 		qp->qp1_hdr.udp.sport = htons(0x8CD1);
2281 		qp->qp1_hdr.udp.csum = 0;
2282 	}
2283 
2284 	/* BTH */
2285 	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
2286 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2287 		qp->qp1_hdr.immediate_present = 1;
2288 	} else {
2289 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2290 	}
2291 	if (wr->send_flags & IB_SEND_SOLICITED)
2292 		qp->qp1_hdr.bth.solicited_event = 1;
2293 	/* pad_count */
2294 	qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3;
2295 
2296 	/* P_key for QP1 is for all members */
2297 	qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF);
2298 	qp->qp1_hdr.bth.destination_qpn = IB_QP1;
2299 	qp->qp1_hdr.bth.ack_req = 0;
2300 	qp->send_psn++;
2301 	qp->send_psn &= BTH_PSN_MASK;
2302 	qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn);
2303 	/* DETH */
2304 	/* Use the priviledged Q_Key for QP1 */
2305 	qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY);
2306 	qp->qp1_hdr.deth.source_qpn = IB_QP1;
2307 
2308 	/* Pack the QP1 to the transmit buffer */
2309 	buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge);
2310 	if (buf) {
2311 		ib_ud_header_pack(&qp->qp1_hdr, buf);
2312 		for (i = wqe->num_sge; i; i--) {
2313 			wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr;
2314 			wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey;
2315 			wqe->sg_list[i].size = wqe->sg_list[i - 1].size;
2316 		}
2317 
2318 		/*
2319 		 * Max Header buf size for IPV6 RoCE V2 is 86,
2320 		 * which is same as the QP1 SQ header buffer.
2321 		 * Header buf size for IPV4 RoCE V2 can be 66.
2322 		 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
2323 		 * Subtract 20 bytes from QP1 SQ header buf size
2324 		 */
2325 		if (is_udp && ip_version == 4)
2326 			sge.size -= 20;
2327 		/*
2328 		 * Max Header buf size for RoCE V1 is 78.
2329 		 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
2330 		 * Subtract 8 bytes from QP1 SQ header buf size
2331 		 */
2332 		if (!is_udp)
2333 			sge.size -= 8;
2334 
2335 		/* Subtract 4 bytes for non vlan packets */
2336 		if (!is_vlan)
2337 			sge.size -= 4;
2338 
2339 		wqe->sg_list[0].addr = sge.addr;
2340 		wqe->sg_list[0].lkey = sge.lkey;
2341 		wqe->sg_list[0].size = sge.size;
2342 		wqe->num_sge++;
2343 
2344 	} else {
2345 		ibdev_err(&qp->rdev->ibdev, "QP1 buffer is empty!");
2346 		rc = -ENOMEM;
2347 	}
2348 	return rc;
2349 }
2350 
2351 /* For the MAD layer, it only provides the recv SGE the size of
2352  * ib_grh + MAD datagram.  No Ethernet headers, Ethertype, BTH, DETH,
2353  * nor RoCE iCRC.  The Cu+ solution must provide buffer for the entire
2354  * receive packet (334 bytes) with no VLAN and then copy the GRH
2355  * and the MAD datagram out to the provided SGE.
2356  */
2357 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp,
2358 					    const struct ib_recv_wr *wr,
2359 					    struct bnxt_qplib_swqe *wqe,
2360 					    int payload_size)
2361 {
2362 	struct bnxt_re_sqp_entries *sqp_entry;
2363 	struct bnxt_qplib_sge ref, sge;
2364 	struct bnxt_re_dev *rdev;
2365 	u32 rq_prod_index;
2366 
2367 	rdev = qp->rdev;
2368 
2369 	rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp);
2370 
2371 	if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge))
2372 		return -ENOMEM;
2373 
2374 	/* Create 1 SGE to receive the entire
2375 	 * ethernet packet
2376 	 */
2377 	/* Save the reference from ULP */
2378 	ref.addr = wqe->sg_list[0].addr;
2379 	ref.lkey = wqe->sg_list[0].lkey;
2380 	ref.size = wqe->sg_list[0].size;
2381 
2382 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[rq_prod_index];
2383 
2384 	/* SGE 1 */
2385 	wqe->sg_list[0].addr = sge.addr;
2386 	wqe->sg_list[0].lkey = sge.lkey;
2387 	wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
2388 	sge.size -= wqe->sg_list[0].size;
2389 
2390 	sqp_entry->sge.addr = ref.addr;
2391 	sqp_entry->sge.lkey = ref.lkey;
2392 	sqp_entry->sge.size = ref.size;
2393 	/* Store the wrid for reporting completion */
2394 	sqp_entry->wrid = wqe->wr_id;
2395 	/* change the wqe->wrid to table index */
2396 	wqe->wr_id = rq_prod_index;
2397 	return 0;
2398 }
2399 
2400 static int is_ud_qp(struct bnxt_re_qp *qp)
2401 {
2402 	return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD ||
2403 		qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI);
2404 }
2405 
2406 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp,
2407 				  const struct ib_send_wr *wr,
2408 				  struct bnxt_qplib_swqe *wqe)
2409 {
2410 	struct bnxt_re_ah *ah = NULL;
2411 
2412 	if (is_ud_qp(qp)) {
2413 		ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah);
2414 		wqe->send.q_key = ud_wr(wr)->remote_qkey;
2415 		wqe->send.dst_qp = ud_wr(wr)->remote_qpn;
2416 		wqe->send.avid = ah->qplib_ah.id;
2417 	}
2418 	switch (wr->opcode) {
2419 	case IB_WR_SEND:
2420 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND;
2421 		break;
2422 	case IB_WR_SEND_WITH_IMM:
2423 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM;
2424 		wqe->send.imm_data = wr->ex.imm_data;
2425 		break;
2426 	case IB_WR_SEND_WITH_INV:
2427 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV;
2428 		wqe->send.inv_key = wr->ex.invalidate_rkey;
2429 		break;
2430 	default:
2431 		return -EINVAL;
2432 	}
2433 	if (wr->send_flags & IB_SEND_SIGNALED)
2434 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2435 	if (wr->send_flags & IB_SEND_FENCE)
2436 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2437 	if (wr->send_flags & IB_SEND_SOLICITED)
2438 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2439 	if (wr->send_flags & IB_SEND_INLINE)
2440 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2441 
2442 	return 0;
2443 }
2444 
2445 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr,
2446 				  struct bnxt_qplib_swqe *wqe)
2447 {
2448 	switch (wr->opcode) {
2449 	case IB_WR_RDMA_WRITE:
2450 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE;
2451 		break;
2452 	case IB_WR_RDMA_WRITE_WITH_IMM:
2453 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM;
2454 		wqe->rdma.imm_data = wr->ex.imm_data;
2455 		break;
2456 	case IB_WR_RDMA_READ:
2457 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ;
2458 		wqe->rdma.inv_key = wr->ex.invalidate_rkey;
2459 		break;
2460 	default:
2461 		return -EINVAL;
2462 	}
2463 	wqe->rdma.remote_va = rdma_wr(wr)->remote_addr;
2464 	wqe->rdma.r_key = rdma_wr(wr)->rkey;
2465 	if (wr->send_flags & IB_SEND_SIGNALED)
2466 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2467 	if (wr->send_flags & IB_SEND_FENCE)
2468 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2469 	if (wr->send_flags & IB_SEND_SOLICITED)
2470 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2471 	if (wr->send_flags & IB_SEND_INLINE)
2472 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2473 
2474 	return 0;
2475 }
2476 
2477 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr,
2478 				    struct bnxt_qplib_swqe *wqe)
2479 {
2480 	switch (wr->opcode) {
2481 	case IB_WR_ATOMIC_CMP_AND_SWP:
2482 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP;
2483 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2484 		wqe->atomic.swap_data = atomic_wr(wr)->swap;
2485 		break;
2486 	case IB_WR_ATOMIC_FETCH_AND_ADD:
2487 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD;
2488 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2489 		break;
2490 	default:
2491 		return -EINVAL;
2492 	}
2493 	wqe->atomic.remote_va = atomic_wr(wr)->remote_addr;
2494 	wqe->atomic.r_key = atomic_wr(wr)->rkey;
2495 	if (wr->send_flags & IB_SEND_SIGNALED)
2496 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2497 	if (wr->send_flags & IB_SEND_FENCE)
2498 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2499 	if (wr->send_flags & IB_SEND_SOLICITED)
2500 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2501 	return 0;
2502 }
2503 
2504 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr,
2505 				 struct bnxt_qplib_swqe *wqe)
2506 {
2507 	wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV;
2508 	wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey;
2509 
2510 	/* Need unconditional fence for local invalidate
2511 	 * opcode to work as expected.
2512 	 */
2513 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2514 
2515 	if (wr->send_flags & IB_SEND_SIGNALED)
2516 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2517 	if (wr->send_flags & IB_SEND_SOLICITED)
2518 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2519 
2520 	return 0;
2521 }
2522 
2523 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr,
2524 				 struct bnxt_qplib_swqe *wqe)
2525 {
2526 	struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr);
2527 	struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl;
2528 	int access = wr->access;
2529 
2530 	wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0];
2531 	wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0];
2532 	wqe->frmr.page_list = mr->pages;
2533 	wqe->frmr.page_list_len = mr->npages;
2534 	wqe->frmr.levels = qplib_frpl->hwq.level;
2535 	wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR;
2536 
2537 	/* Need unconditional fence for reg_mr
2538 	 * opcode to function as expected.
2539 	 */
2540 
2541 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2542 
2543 	if (wr->wr.send_flags & IB_SEND_SIGNALED)
2544 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2545 
2546 	if (access & IB_ACCESS_LOCAL_WRITE)
2547 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE;
2548 	if (access & IB_ACCESS_REMOTE_READ)
2549 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ;
2550 	if (access & IB_ACCESS_REMOTE_WRITE)
2551 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE;
2552 	if (access & IB_ACCESS_REMOTE_ATOMIC)
2553 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC;
2554 	if (access & IB_ACCESS_MW_BIND)
2555 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND;
2556 
2557 	wqe->frmr.l_key = wr->key;
2558 	wqe->frmr.length = wr->mr->length;
2559 	wqe->frmr.pbl_pg_sz_log = ilog2(PAGE_SIZE >> PAGE_SHIFT_4K);
2560 	wqe->frmr.pg_sz_log = ilog2(wr->mr->page_size >> PAGE_SHIFT_4K);
2561 	wqe->frmr.va = wr->mr->iova;
2562 	return 0;
2563 }
2564 
2565 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev,
2566 				    const struct ib_send_wr *wr,
2567 				    struct bnxt_qplib_swqe *wqe)
2568 {
2569 	/*  Copy the inline data to the data  field */
2570 	u8 *in_data;
2571 	u32 i, sge_len;
2572 	void *sge_addr;
2573 
2574 	in_data = wqe->inline_data;
2575 	for (i = 0; i < wr->num_sge; i++) {
2576 		sge_addr = (void *)(unsigned long)
2577 				wr->sg_list[i].addr;
2578 		sge_len = wr->sg_list[i].length;
2579 
2580 		if ((sge_len + wqe->inline_len) >
2581 		    BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) {
2582 			ibdev_err(&rdev->ibdev,
2583 				  "Inline data size requested > supported value");
2584 			return -EINVAL;
2585 		}
2586 		sge_len = wr->sg_list[i].length;
2587 
2588 		memcpy(in_data, sge_addr, sge_len);
2589 		in_data += wr->sg_list[i].length;
2590 		wqe->inline_len += wr->sg_list[i].length;
2591 	}
2592 	return wqe->inline_len;
2593 }
2594 
2595 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
2596 				   const struct ib_send_wr *wr,
2597 				   struct bnxt_qplib_swqe *wqe)
2598 {
2599 	int payload_sz = 0;
2600 
2601 	if (wr->send_flags & IB_SEND_INLINE)
2602 		payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe);
2603 	else
2604 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list,
2605 					       wqe->num_sge);
2606 
2607 	return payload_sz;
2608 }
2609 
2610 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
2611 {
2612 	if ((qp->ib_qp.qp_type == IB_QPT_UD ||
2613 	     qp->ib_qp.qp_type == IB_QPT_GSI ||
2614 	     qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
2615 	     qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
2616 		int qp_attr_mask;
2617 		struct ib_qp_attr qp_attr;
2618 
2619 		qp_attr_mask = IB_QP_STATE;
2620 		qp_attr.qp_state = IB_QPS_RTS;
2621 		bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
2622 		qp->qplib_qp.wqe_cnt = 0;
2623 	}
2624 }
2625 
2626 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
2627 				       struct bnxt_re_qp *qp,
2628 				       const struct ib_send_wr *wr)
2629 {
2630 	int rc = 0, payload_sz = 0;
2631 	unsigned long flags;
2632 
2633 	spin_lock_irqsave(&qp->sq_lock, flags);
2634 	while (wr) {
2635 		struct bnxt_qplib_swqe wqe = {};
2636 
2637 		/* Common */
2638 		wqe.num_sge = wr->num_sge;
2639 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2640 			ibdev_err(&rdev->ibdev,
2641 				  "Limit exceeded for Send SGEs");
2642 			rc = -EINVAL;
2643 			goto bad;
2644 		}
2645 
2646 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2647 		if (payload_sz < 0) {
2648 			rc = -EINVAL;
2649 			goto bad;
2650 		}
2651 		wqe.wr_id = wr->wr_id;
2652 
2653 		wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND;
2654 
2655 		rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2656 		if (!rc)
2657 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2658 bad:
2659 		if (rc) {
2660 			ibdev_err(&rdev->ibdev,
2661 				  "Post send failed opcode = %#x rc = %d",
2662 				  wr->opcode, rc);
2663 			break;
2664 		}
2665 		wr = wr->next;
2666 	}
2667 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2668 	bnxt_ud_qp_hw_stall_workaround(qp);
2669 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2670 	return rc;
2671 }
2672 
2673 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr,
2674 		      const struct ib_send_wr **bad_wr)
2675 {
2676 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2677 	struct bnxt_qplib_swqe wqe;
2678 	int rc = 0, payload_sz = 0;
2679 	unsigned long flags;
2680 
2681 	spin_lock_irqsave(&qp->sq_lock, flags);
2682 	while (wr) {
2683 		/* House keeping */
2684 		memset(&wqe, 0, sizeof(wqe));
2685 
2686 		/* Common */
2687 		wqe.num_sge = wr->num_sge;
2688 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2689 			ibdev_err(&qp->rdev->ibdev,
2690 				  "Limit exceeded for Send SGEs");
2691 			rc = -EINVAL;
2692 			goto bad;
2693 		}
2694 
2695 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2696 		if (payload_sz < 0) {
2697 			rc = -EINVAL;
2698 			goto bad;
2699 		}
2700 		wqe.wr_id = wr->wr_id;
2701 
2702 		switch (wr->opcode) {
2703 		case IB_WR_SEND:
2704 		case IB_WR_SEND_WITH_IMM:
2705 			if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) {
2706 				rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe,
2707 							       payload_sz);
2708 				if (rc)
2709 					goto bad;
2710 				wqe.rawqp1.lflags |=
2711 					SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC;
2712 			}
2713 			switch (wr->send_flags) {
2714 			case IB_SEND_IP_CSUM:
2715 				wqe.rawqp1.lflags |=
2716 					SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM;
2717 				break;
2718 			default:
2719 				break;
2720 			}
2721 			fallthrough;
2722 		case IB_WR_SEND_WITH_INV:
2723 			rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2724 			break;
2725 		case IB_WR_RDMA_WRITE:
2726 		case IB_WR_RDMA_WRITE_WITH_IMM:
2727 		case IB_WR_RDMA_READ:
2728 			rc = bnxt_re_build_rdma_wqe(wr, &wqe);
2729 			break;
2730 		case IB_WR_ATOMIC_CMP_AND_SWP:
2731 		case IB_WR_ATOMIC_FETCH_AND_ADD:
2732 			rc = bnxt_re_build_atomic_wqe(wr, &wqe);
2733 			break;
2734 		case IB_WR_RDMA_READ_WITH_INV:
2735 			ibdev_err(&qp->rdev->ibdev,
2736 				  "RDMA Read with Invalidate is not supported");
2737 			rc = -EINVAL;
2738 			goto bad;
2739 		case IB_WR_LOCAL_INV:
2740 			rc = bnxt_re_build_inv_wqe(wr, &wqe);
2741 			break;
2742 		case IB_WR_REG_MR:
2743 			rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe);
2744 			break;
2745 		default:
2746 			/* Unsupported WRs */
2747 			ibdev_err(&qp->rdev->ibdev,
2748 				  "WR (%#x) is not supported", wr->opcode);
2749 			rc = -EINVAL;
2750 			goto bad;
2751 		}
2752 		if (!rc)
2753 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2754 bad:
2755 		if (rc) {
2756 			ibdev_err(&qp->rdev->ibdev,
2757 				  "post_send failed op:%#x qps = %#x rc = %d\n",
2758 				  wr->opcode, qp->qplib_qp.state, rc);
2759 			*bad_wr = wr;
2760 			break;
2761 		}
2762 		wr = wr->next;
2763 	}
2764 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2765 	bnxt_ud_qp_hw_stall_workaround(qp);
2766 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2767 
2768 	return rc;
2769 }
2770 
2771 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev,
2772 				       struct bnxt_re_qp *qp,
2773 				       const struct ib_recv_wr *wr)
2774 {
2775 	struct bnxt_qplib_swqe wqe;
2776 	int rc = 0;
2777 
2778 	memset(&wqe, 0, sizeof(wqe));
2779 	while (wr) {
2780 		/* House keeping */
2781 		memset(&wqe, 0, sizeof(wqe));
2782 
2783 		/* Common */
2784 		wqe.num_sge = wr->num_sge;
2785 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2786 			ibdev_err(&rdev->ibdev,
2787 				  "Limit exceeded for Receive SGEs");
2788 			rc = -EINVAL;
2789 			break;
2790 		}
2791 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
2792 		wqe.wr_id = wr->wr_id;
2793 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2794 
2795 		rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2796 		if (rc)
2797 			break;
2798 
2799 		wr = wr->next;
2800 	}
2801 	if (!rc)
2802 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2803 	return rc;
2804 }
2805 
2806 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
2807 		      const struct ib_recv_wr **bad_wr)
2808 {
2809 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2810 	struct bnxt_qplib_swqe wqe;
2811 	int rc = 0, payload_sz = 0;
2812 	unsigned long flags;
2813 	u32 count = 0;
2814 
2815 	spin_lock_irqsave(&qp->rq_lock, flags);
2816 	while (wr) {
2817 		/* House keeping */
2818 		memset(&wqe, 0, sizeof(wqe));
2819 
2820 		/* Common */
2821 		wqe.num_sge = wr->num_sge;
2822 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2823 			ibdev_err(&qp->rdev->ibdev,
2824 				  "Limit exceeded for Receive SGEs");
2825 			rc = -EINVAL;
2826 			*bad_wr = wr;
2827 			break;
2828 		}
2829 
2830 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
2831 					       wr->num_sge);
2832 		wqe.wr_id = wr->wr_id;
2833 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2834 
2835 		if (ib_qp->qp_type == IB_QPT_GSI &&
2836 		    qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI)
2837 			rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe,
2838 							      payload_sz);
2839 		if (!rc)
2840 			rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2841 		if (rc) {
2842 			*bad_wr = wr;
2843 			break;
2844 		}
2845 
2846 		/* Ring DB if the RQEs posted reaches a threshold value */
2847 		if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
2848 			bnxt_qplib_post_recv_db(&qp->qplib_qp);
2849 			count = 0;
2850 		}
2851 
2852 		wr = wr->next;
2853 	}
2854 
2855 	if (count)
2856 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2857 
2858 	spin_unlock_irqrestore(&qp->rq_lock, flags);
2859 
2860 	return rc;
2861 }
2862 
2863 /* Completion Queues */
2864 int bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
2865 {
2866 	struct bnxt_re_cq *cq;
2867 	struct bnxt_qplib_nq *nq;
2868 	struct bnxt_re_dev *rdev;
2869 
2870 	cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
2871 	rdev = cq->rdev;
2872 	nq = cq->qplib_cq.nq;
2873 
2874 	bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2875 	ib_umem_release(cq->umem);
2876 
2877 	atomic_dec(&rdev->cq_count);
2878 	nq->budget--;
2879 	kfree(cq->cql);
2880 	return 0;
2881 }
2882 
2883 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
2884 		      struct ib_udata *udata)
2885 {
2886 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev);
2887 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
2888 	struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq);
2889 	int rc, entries;
2890 	int cqe = attr->cqe;
2891 	struct bnxt_qplib_nq *nq = NULL;
2892 	unsigned int nq_alloc_cnt;
2893 
2894 	if (attr->flags)
2895 		return -EOPNOTSUPP;
2896 
2897 	/* Validate CQ fields */
2898 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
2899 		ibdev_err(&rdev->ibdev, "Failed to create CQ -max exceeded");
2900 		return -EINVAL;
2901 	}
2902 
2903 	cq->rdev = rdev;
2904 	cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
2905 
2906 	entries = roundup_pow_of_two(cqe + 1);
2907 	if (entries > dev_attr->max_cq_wqes + 1)
2908 		entries = dev_attr->max_cq_wqes + 1;
2909 
2910 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
2911 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
2912 	if (udata) {
2913 		struct bnxt_re_cq_req req;
2914 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
2915 			udata, struct bnxt_re_ucontext, ib_uctx);
2916 		if (ib_copy_from_udata(&req, udata, sizeof(req))) {
2917 			rc = -EFAULT;
2918 			goto fail;
2919 		}
2920 
2921 		cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va,
2922 				       entries * sizeof(struct cq_base),
2923 				       IB_ACCESS_LOCAL_WRITE);
2924 		if (IS_ERR(cq->umem)) {
2925 			rc = PTR_ERR(cq->umem);
2926 			goto fail;
2927 		}
2928 		cq->qplib_cq.sg_info.umem = cq->umem;
2929 		cq->qplib_cq.dpi = &uctx->dpi;
2930 	} else {
2931 		cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL);
2932 		cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe),
2933 				  GFP_KERNEL);
2934 		if (!cq->cql) {
2935 			rc = -ENOMEM;
2936 			goto fail;
2937 		}
2938 
2939 		cq->qplib_cq.dpi = &rdev->dpi_privileged;
2940 	}
2941 	/*
2942 	 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
2943 	 * used for getting the NQ index.
2944 	 */
2945 	nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
2946 	nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
2947 	cq->qplib_cq.max_wqe = entries;
2948 	cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
2949 	cq->qplib_cq.nq	= nq;
2950 
2951 	rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
2952 	if (rc) {
2953 		ibdev_err(&rdev->ibdev, "Failed to create HW CQ");
2954 		goto fail;
2955 	}
2956 
2957 	cq->ib_cq.cqe = entries;
2958 	cq->cq_period = cq->qplib_cq.period;
2959 	nq->budget++;
2960 
2961 	atomic_inc(&rdev->cq_count);
2962 	spin_lock_init(&cq->cq_lock);
2963 
2964 	if (udata) {
2965 		struct bnxt_re_cq_resp resp;
2966 
2967 		resp.cqid = cq->qplib_cq.id;
2968 		resp.tail = cq->qplib_cq.hwq.cons;
2969 		resp.phase = cq->qplib_cq.period;
2970 		resp.rsvd = 0;
2971 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
2972 		if (rc) {
2973 			ibdev_err(&rdev->ibdev, "Failed to copy CQ udata");
2974 			bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2975 			goto c2fail;
2976 		}
2977 	}
2978 
2979 	return 0;
2980 
2981 c2fail:
2982 	ib_umem_release(cq->umem);
2983 fail:
2984 	kfree(cq->cql);
2985 	return rc;
2986 }
2987 
2988 static void bnxt_re_resize_cq_complete(struct bnxt_re_cq *cq)
2989 {
2990 	struct bnxt_re_dev *rdev = cq->rdev;
2991 
2992 	bnxt_qplib_resize_cq_complete(&rdev->qplib_res, &cq->qplib_cq);
2993 
2994 	cq->qplib_cq.max_wqe = cq->resize_cqe;
2995 	if (cq->resize_umem) {
2996 		ib_umem_release(cq->umem);
2997 		cq->umem = cq->resize_umem;
2998 		cq->resize_umem = NULL;
2999 		cq->resize_cqe = 0;
3000 	}
3001 }
3002 
3003 int bnxt_re_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
3004 {
3005 	struct bnxt_qplib_sg_info sg_info = {};
3006 	struct bnxt_qplib_dpi *orig_dpi = NULL;
3007 	struct bnxt_qplib_dev_attr *dev_attr;
3008 	struct bnxt_re_ucontext *uctx = NULL;
3009 	struct bnxt_re_resize_cq_req req;
3010 	struct bnxt_re_dev *rdev;
3011 	struct bnxt_re_cq *cq;
3012 	int rc, entries;
3013 
3014 	cq =  container_of(ibcq, struct bnxt_re_cq, ib_cq);
3015 	rdev = cq->rdev;
3016 	dev_attr = &rdev->dev_attr;
3017 	if (!ibcq->uobject) {
3018 		ibdev_err(&rdev->ibdev, "Kernel CQ Resize not supported");
3019 		return -EOPNOTSUPP;
3020 	}
3021 
3022 	if (cq->resize_umem) {
3023 		ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - Busy",
3024 			  cq->qplib_cq.id);
3025 		return -EBUSY;
3026 	}
3027 
3028 	/* Check the requested cq depth out of supported depth */
3029 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
3030 		ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - out of range cqe %d",
3031 			  cq->qplib_cq.id, cqe);
3032 		return -EINVAL;
3033 	}
3034 
3035 	entries = roundup_pow_of_two(cqe + 1);
3036 	if (entries > dev_attr->max_cq_wqes + 1)
3037 		entries = dev_attr->max_cq_wqes + 1;
3038 
3039 	uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext,
3040 					 ib_uctx);
3041 	/* uverbs consumer */
3042 	if (ib_copy_from_udata(&req, udata, sizeof(req))) {
3043 		rc = -EFAULT;
3044 		goto fail;
3045 	}
3046 
3047 	cq->resize_umem = ib_umem_get(&rdev->ibdev, req.cq_va,
3048 				      entries * sizeof(struct cq_base),
3049 				      IB_ACCESS_LOCAL_WRITE);
3050 	if (IS_ERR(cq->resize_umem)) {
3051 		rc = PTR_ERR(cq->resize_umem);
3052 		cq->resize_umem = NULL;
3053 		ibdev_err(&rdev->ibdev, "%s: ib_umem_get failed! rc = %d\n",
3054 			  __func__, rc);
3055 		goto fail;
3056 	}
3057 	cq->resize_cqe = entries;
3058 	memcpy(&sg_info, &cq->qplib_cq.sg_info, sizeof(sg_info));
3059 	orig_dpi = cq->qplib_cq.dpi;
3060 
3061 	cq->qplib_cq.sg_info.umem = cq->resize_umem;
3062 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
3063 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
3064 	cq->qplib_cq.dpi = &uctx->dpi;
3065 
3066 	rc = bnxt_qplib_resize_cq(&rdev->qplib_res, &cq->qplib_cq, entries);
3067 	if (rc) {
3068 		ibdev_err(&rdev->ibdev, "Resize HW CQ %#x failed!",
3069 			  cq->qplib_cq.id);
3070 		goto fail;
3071 	}
3072 
3073 	cq->ib_cq.cqe = cq->resize_cqe;
3074 
3075 	return 0;
3076 
3077 fail:
3078 	if (cq->resize_umem) {
3079 		ib_umem_release(cq->resize_umem);
3080 		cq->resize_umem = NULL;
3081 		cq->resize_cqe = 0;
3082 		memcpy(&cq->qplib_cq.sg_info, &sg_info, sizeof(sg_info));
3083 		cq->qplib_cq.dpi = orig_dpi;
3084 	}
3085 	return rc;
3086 }
3087 
3088 static u8 __req_to_ib_wc_status(u8 qstatus)
3089 {
3090 	switch (qstatus) {
3091 	case CQ_REQ_STATUS_OK:
3092 		return IB_WC_SUCCESS;
3093 	case CQ_REQ_STATUS_BAD_RESPONSE_ERR:
3094 		return IB_WC_BAD_RESP_ERR;
3095 	case CQ_REQ_STATUS_LOCAL_LENGTH_ERR:
3096 		return IB_WC_LOC_LEN_ERR;
3097 	case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR:
3098 		return IB_WC_LOC_QP_OP_ERR;
3099 	case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR:
3100 		return IB_WC_LOC_PROT_ERR;
3101 	case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR:
3102 		return IB_WC_GENERAL_ERR;
3103 	case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR:
3104 		return IB_WC_REM_INV_REQ_ERR;
3105 	case CQ_REQ_STATUS_REMOTE_ACCESS_ERR:
3106 		return IB_WC_REM_ACCESS_ERR;
3107 	case CQ_REQ_STATUS_REMOTE_OPERATION_ERR:
3108 		return IB_WC_REM_OP_ERR;
3109 	case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR:
3110 		return IB_WC_RNR_RETRY_EXC_ERR;
3111 	case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR:
3112 		return IB_WC_RETRY_EXC_ERR;
3113 	case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR:
3114 		return IB_WC_WR_FLUSH_ERR;
3115 	default:
3116 		return IB_WC_GENERAL_ERR;
3117 	}
3118 	return 0;
3119 }
3120 
3121 static u8 __rawqp1_to_ib_wc_status(u8 qstatus)
3122 {
3123 	switch (qstatus) {
3124 	case CQ_RES_RAWETH_QP1_STATUS_OK:
3125 		return IB_WC_SUCCESS;
3126 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR:
3127 		return IB_WC_LOC_ACCESS_ERR;
3128 	case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR:
3129 		return IB_WC_LOC_LEN_ERR;
3130 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR:
3131 		return IB_WC_LOC_PROT_ERR;
3132 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR:
3133 		return IB_WC_LOC_QP_OP_ERR;
3134 	case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR:
3135 		return IB_WC_GENERAL_ERR;
3136 	case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR:
3137 		return IB_WC_WR_FLUSH_ERR;
3138 	case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR:
3139 		return IB_WC_WR_FLUSH_ERR;
3140 	default:
3141 		return IB_WC_GENERAL_ERR;
3142 	}
3143 }
3144 
3145 static u8 __rc_to_ib_wc_status(u8 qstatus)
3146 {
3147 	switch (qstatus) {
3148 	case CQ_RES_RC_STATUS_OK:
3149 		return IB_WC_SUCCESS;
3150 	case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR:
3151 		return IB_WC_LOC_ACCESS_ERR;
3152 	case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR:
3153 		return IB_WC_LOC_LEN_ERR;
3154 	case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR:
3155 		return IB_WC_LOC_PROT_ERR;
3156 	case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR:
3157 		return IB_WC_LOC_QP_OP_ERR;
3158 	case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR:
3159 		return IB_WC_GENERAL_ERR;
3160 	case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR:
3161 		return IB_WC_REM_INV_REQ_ERR;
3162 	case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR:
3163 		return IB_WC_WR_FLUSH_ERR;
3164 	case CQ_RES_RC_STATUS_HW_FLUSH_ERR:
3165 		return IB_WC_WR_FLUSH_ERR;
3166 	default:
3167 		return IB_WC_GENERAL_ERR;
3168 	}
3169 }
3170 
3171 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe)
3172 {
3173 	switch (cqe->type) {
3174 	case BNXT_QPLIB_SWQE_TYPE_SEND:
3175 		wc->opcode = IB_WC_SEND;
3176 		break;
3177 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM:
3178 		wc->opcode = IB_WC_SEND;
3179 		wc->wc_flags |= IB_WC_WITH_IMM;
3180 		break;
3181 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV:
3182 		wc->opcode = IB_WC_SEND;
3183 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3184 		break;
3185 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE:
3186 		wc->opcode = IB_WC_RDMA_WRITE;
3187 		break;
3188 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM:
3189 		wc->opcode = IB_WC_RDMA_WRITE;
3190 		wc->wc_flags |= IB_WC_WITH_IMM;
3191 		break;
3192 	case BNXT_QPLIB_SWQE_TYPE_RDMA_READ:
3193 		wc->opcode = IB_WC_RDMA_READ;
3194 		break;
3195 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP:
3196 		wc->opcode = IB_WC_COMP_SWAP;
3197 		break;
3198 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD:
3199 		wc->opcode = IB_WC_FETCH_ADD;
3200 		break;
3201 	case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV:
3202 		wc->opcode = IB_WC_LOCAL_INV;
3203 		break;
3204 	case BNXT_QPLIB_SWQE_TYPE_REG_MR:
3205 		wc->opcode = IB_WC_REG_MR;
3206 		break;
3207 	default:
3208 		wc->opcode = IB_WC_SEND;
3209 		break;
3210 	}
3211 
3212 	wc->status = __req_to_ib_wc_status(cqe->status);
3213 }
3214 
3215 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags,
3216 				     u16 raweth_qp1_flags2)
3217 {
3218 	bool is_ipv6 = false, is_ipv4 = false;
3219 
3220 	/* raweth_qp1_flags Bit 9-6 indicates itype */
3221 	if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3222 	    != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3223 		return -1;
3224 
3225 	if (raweth_qp1_flags2 &
3226 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC &&
3227 	    raweth_qp1_flags2 &
3228 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) {
3229 		/* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
3230 		(raweth_qp1_flags2 &
3231 		 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ?
3232 			(is_ipv6 = true) : (is_ipv4 = true);
3233 		return ((is_ipv6) ?
3234 			 BNXT_RE_ROCEV2_IPV6_PACKET :
3235 			 BNXT_RE_ROCEV2_IPV4_PACKET);
3236 	} else {
3237 		return BNXT_RE_ROCE_V1_PACKET;
3238 	}
3239 }
3240 
3241 static int bnxt_re_to_ib_nw_type(int nw_type)
3242 {
3243 	u8 nw_hdr_type = 0xFF;
3244 
3245 	switch (nw_type) {
3246 	case BNXT_RE_ROCE_V1_PACKET:
3247 		nw_hdr_type = RDMA_NETWORK_ROCE_V1;
3248 		break;
3249 	case BNXT_RE_ROCEV2_IPV4_PACKET:
3250 		nw_hdr_type = RDMA_NETWORK_IPV4;
3251 		break;
3252 	case BNXT_RE_ROCEV2_IPV6_PACKET:
3253 		nw_hdr_type = RDMA_NETWORK_IPV6;
3254 		break;
3255 	}
3256 	return nw_hdr_type;
3257 }
3258 
3259 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev,
3260 				       void *rq_hdr_buf)
3261 {
3262 	u8 *tmp_buf = NULL;
3263 	struct ethhdr *eth_hdr;
3264 	u16 eth_type;
3265 	bool rc = false;
3266 
3267 	tmp_buf = (u8 *)rq_hdr_buf;
3268 	/*
3269 	 * If dest mac is not same as I/F mac, this could be a
3270 	 * loopback address or multicast address, check whether
3271 	 * it is a loopback packet
3272 	 */
3273 	if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) {
3274 		tmp_buf += 4;
3275 		/* Check the  ether type */
3276 		eth_hdr = (struct ethhdr *)tmp_buf;
3277 		eth_type = ntohs(eth_hdr->h_proto);
3278 		switch (eth_type) {
3279 		case ETH_P_IBOE:
3280 			rc = true;
3281 			break;
3282 		case ETH_P_IP:
3283 		case ETH_P_IPV6: {
3284 			u32 len;
3285 			struct udphdr *udp_hdr;
3286 
3287 			len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) :
3288 						      sizeof(struct ipv6hdr));
3289 			tmp_buf += sizeof(struct ethhdr) + len;
3290 			udp_hdr = (struct udphdr *)tmp_buf;
3291 			if (ntohs(udp_hdr->dest) ==
3292 				    ROCE_V2_UDP_DPORT)
3293 				rc = true;
3294 			break;
3295 			}
3296 		default:
3297 			break;
3298 		}
3299 	}
3300 
3301 	return rc;
3302 }
3303 
3304 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *gsi_qp,
3305 					 struct bnxt_qplib_cqe *cqe)
3306 {
3307 	struct bnxt_re_dev *rdev = gsi_qp->rdev;
3308 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3309 	struct bnxt_re_qp *gsi_sqp = rdev->gsi_ctx.gsi_sqp;
3310 	struct bnxt_re_ah *gsi_sah;
3311 	struct ib_send_wr *swr;
3312 	struct ib_ud_wr udwr;
3313 	struct ib_recv_wr rwr;
3314 	int pkt_type = 0;
3315 	u32 tbl_idx;
3316 	void *rq_hdr_buf;
3317 	dma_addr_t rq_hdr_buf_map;
3318 	dma_addr_t shrq_hdr_buf_map;
3319 	u32 offset = 0;
3320 	u32 skip_bytes = 0;
3321 	struct ib_sge s_sge[2];
3322 	struct ib_sge r_sge[2];
3323 	int rc;
3324 
3325 	memset(&udwr, 0, sizeof(udwr));
3326 	memset(&rwr, 0, sizeof(rwr));
3327 	memset(&s_sge, 0, sizeof(s_sge));
3328 	memset(&r_sge, 0, sizeof(r_sge));
3329 
3330 	swr = &udwr.wr;
3331 	tbl_idx = cqe->wr_id;
3332 
3333 	rq_hdr_buf = gsi_qp->qplib_qp.rq_hdr_buf +
3334 			(tbl_idx * gsi_qp->qplib_qp.rq_hdr_buf_size);
3335 	rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3336 							  tbl_idx);
3337 
3338 	/* Shadow QP header buffer */
3339 	shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3340 							    tbl_idx);
3341 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3342 
3343 	/* Store this cqe */
3344 	memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe));
3345 	sqp_entry->qp1_qp = gsi_qp;
3346 
3347 	/* Find packet type from the cqe */
3348 
3349 	pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags,
3350 					     cqe->raweth_qp1_flags2);
3351 	if (pkt_type < 0) {
3352 		ibdev_err(&rdev->ibdev, "Invalid packet\n");
3353 		return -EINVAL;
3354 	}
3355 
3356 	/* Adjust the offset for the user buffer and post in the rq */
3357 
3358 	if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET)
3359 		offset = 20;
3360 
3361 	/*
3362 	 * QP1 loopback packet has 4 bytes of internal header before
3363 	 * ether header. Skip these four bytes.
3364 	 */
3365 	if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf))
3366 		skip_bytes = 4;
3367 
3368 	/* First send SGE . Skip the ether header*/
3369 	s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
3370 			+ skip_bytes;
3371 	s_sge[0].lkey = 0xFFFFFFFF;
3372 	s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 :
3373 				BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
3374 
3375 	/* Second Send SGE */
3376 	s_sge[1].addr = s_sge[0].addr + s_sge[0].length +
3377 			BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE;
3378 	if (pkt_type != BNXT_RE_ROCE_V1_PACKET)
3379 		s_sge[1].addr += 8;
3380 	s_sge[1].lkey = 0xFFFFFFFF;
3381 	s_sge[1].length = 256;
3382 
3383 	/* First recv SGE */
3384 
3385 	r_sge[0].addr = shrq_hdr_buf_map;
3386 	r_sge[0].lkey = 0xFFFFFFFF;
3387 	r_sge[0].length = 40;
3388 
3389 	r_sge[1].addr = sqp_entry->sge.addr + offset;
3390 	r_sge[1].lkey = sqp_entry->sge.lkey;
3391 	r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset;
3392 
3393 	/* Create receive work request */
3394 	rwr.num_sge = 2;
3395 	rwr.sg_list = r_sge;
3396 	rwr.wr_id = tbl_idx;
3397 	rwr.next = NULL;
3398 
3399 	rc = bnxt_re_post_recv_shadow_qp(rdev, gsi_sqp, &rwr);
3400 	if (rc) {
3401 		ibdev_err(&rdev->ibdev,
3402 			  "Failed to post Rx buffers to shadow QP");
3403 		return -ENOMEM;
3404 	}
3405 
3406 	swr->num_sge = 2;
3407 	swr->sg_list = s_sge;
3408 	swr->wr_id = tbl_idx;
3409 	swr->opcode = IB_WR_SEND;
3410 	swr->next = NULL;
3411 	gsi_sah = rdev->gsi_ctx.gsi_sah;
3412 	udwr.ah = &gsi_sah->ib_ah;
3413 	udwr.remote_qpn = gsi_sqp->qplib_qp.id;
3414 	udwr.remote_qkey = gsi_sqp->qplib_qp.qkey;
3415 
3416 	/* post data received  in the send queue */
3417 	return bnxt_re_post_send_shadow_qp(rdev, gsi_sqp, swr);
3418 }
3419 
3420 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc,
3421 					  struct bnxt_qplib_cqe *cqe)
3422 {
3423 	wc->opcode = IB_WC_RECV;
3424 	wc->status = __rawqp1_to_ib_wc_status(cqe->status);
3425 	wc->wc_flags |= IB_WC_GRH;
3426 }
3427 
3428 static bool bnxt_re_check_if_vlan_valid(struct bnxt_re_dev *rdev,
3429 					u16 vlan_id)
3430 {
3431 	/*
3432 	 * Check if the vlan is configured in the host.  If not configured, it
3433 	 * can be a transparent VLAN. So dont report the vlan id.
3434 	 */
3435 	if (!__vlan_find_dev_deep_rcu(rdev->netdev,
3436 				      htons(ETH_P_8021Q), vlan_id))
3437 		return false;
3438 	return true;
3439 }
3440 
3441 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe,
3442 				u16 *vid, u8 *sl)
3443 {
3444 	bool ret = false;
3445 	u32 metadata;
3446 	u16 tpid;
3447 
3448 	metadata = orig_cqe->raweth_qp1_metadata;
3449 	if (orig_cqe->raweth_qp1_flags2 &
3450 		CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) {
3451 		tpid = ((metadata &
3452 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >>
3453 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT);
3454 		if (tpid == ETH_P_8021Q) {
3455 			*vid = metadata &
3456 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK;
3457 			*sl = (metadata &
3458 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >>
3459 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT;
3460 			ret = true;
3461 		}
3462 	}
3463 
3464 	return ret;
3465 }
3466 
3467 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc,
3468 				      struct bnxt_qplib_cqe *cqe)
3469 {
3470 	wc->opcode = IB_WC_RECV;
3471 	wc->status = __rc_to_ib_wc_status(cqe->status);
3472 
3473 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3474 		wc->wc_flags |= IB_WC_WITH_IMM;
3475 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3476 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3477 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3478 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3479 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3480 }
3481 
3482 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *gsi_sqp,
3483 					     struct ib_wc *wc,
3484 					     struct bnxt_qplib_cqe *cqe)
3485 {
3486 	struct bnxt_re_dev *rdev = gsi_sqp->rdev;
3487 	struct bnxt_re_qp *gsi_qp = NULL;
3488 	struct bnxt_qplib_cqe *orig_cqe = NULL;
3489 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3490 	int nw_type;
3491 	u32 tbl_idx;
3492 	u16 vlan_id;
3493 	u8 sl;
3494 
3495 	tbl_idx = cqe->wr_id;
3496 
3497 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3498 	gsi_qp = sqp_entry->qp1_qp;
3499 	orig_cqe = &sqp_entry->cqe;
3500 
3501 	wc->wr_id = sqp_entry->wrid;
3502 	wc->byte_len = orig_cqe->length;
3503 	wc->qp = &gsi_qp->ib_qp;
3504 
3505 	wc->ex.imm_data = orig_cqe->immdata;
3506 	wc->src_qp = orig_cqe->src_qp;
3507 	memcpy(wc->smac, orig_cqe->smac, ETH_ALEN);
3508 	if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) {
3509 		if (bnxt_re_check_if_vlan_valid(rdev, vlan_id)) {
3510 			wc->vlan_id = vlan_id;
3511 			wc->sl = sl;
3512 			wc->wc_flags |= IB_WC_WITH_VLAN;
3513 		}
3514 	}
3515 	wc->port_num = 1;
3516 	wc->vendor_err = orig_cqe->status;
3517 
3518 	wc->opcode = IB_WC_RECV;
3519 	wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status);
3520 	wc->wc_flags |= IB_WC_GRH;
3521 
3522 	nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags,
3523 					    orig_cqe->raweth_qp1_flags2);
3524 	if (nw_type >= 0) {
3525 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3526 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3527 	}
3528 }
3529 
3530 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp,
3531 				      struct ib_wc *wc,
3532 				      struct bnxt_qplib_cqe *cqe)
3533 {
3534 	struct bnxt_re_dev *rdev;
3535 	u16 vlan_id = 0;
3536 	u8 nw_type;
3537 
3538 	rdev = qp->rdev;
3539 	wc->opcode = IB_WC_RECV;
3540 	wc->status = __rc_to_ib_wc_status(cqe->status);
3541 
3542 	if (cqe->flags & CQ_RES_UD_FLAGS_IMM)
3543 		wc->wc_flags |= IB_WC_WITH_IMM;
3544 	/* report only on GSI QP for Thor */
3545 	if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) {
3546 		wc->wc_flags |= IB_WC_GRH;
3547 		memcpy(wc->smac, cqe->smac, ETH_ALEN);
3548 		wc->wc_flags |= IB_WC_WITH_SMAC;
3549 		if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) {
3550 			vlan_id = (cqe->cfa_meta & 0xFFF);
3551 		}
3552 		/* Mark only if vlan_id is non zero */
3553 		if (vlan_id && bnxt_re_check_if_vlan_valid(rdev, vlan_id)) {
3554 			wc->vlan_id = vlan_id;
3555 			wc->wc_flags |= IB_WC_WITH_VLAN;
3556 		}
3557 		nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >>
3558 			   CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT;
3559 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3560 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3561 	}
3562 
3563 }
3564 
3565 static int send_phantom_wqe(struct bnxt_re_qp *qp)
3566 {
3567 	struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
3568 	unsigned long flags;
3569 	int rc = 0;
3570 
3571 	spin_lock_irqsave(&qp->sq_lock, flags);
3572 
3573 	rc = bnxt_re_bind_fence_mw(lib_qp);
3574 	if (!rc) {
3575 		lib_qp->sq.phantom_wqe_cnt++;
3576 		ibdev_dbg(&qp->rdev->ibdev,
3577 			  "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
3578 			  lib_qp->id, lib_qp->sq.hwq.prod,
3579 			  HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
3580 			  lib_qp->sq.phantom_wqe_cnt);
3581 	}
3582 
3583 	spin_unlock_irqrestore(&qp->sq_lock, flags);
3584 	return rc;
3585 }
3586 
3587 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
3588 {
3589 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3590 	struct bnxt_re_qp *qp, *sh_qp;
3591 	struct bnxt_qplib_cqe *cqe;
3592 	int i, ncqe, budget;
3593 	struct bnxt_qplib_q *sq;
3594 	struct bnxt_qplib_qp *lib_qp;
3595 	u32 tbl_idx;
3596 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3597 	unsigned long flags;
3598 
3599 	/* User CQ; the only processing we do is to
3600 	 * complete any pending CQ resize operation.
3601 	 */
3602 	if (cq->umem) {
3603 		if (cq->resize_umem)
3604 			bnxt_re_resize_cq_complete(cq);
3605 		return 0;
3606 	}
3607 
3608 	spin_lock_irqsave(&cq->cq_lock, flags);
3609 	budget = min_t(u32, num_entries, cq->max_cql);
3610 	num_entries = budget;
3611 	if (!cq->cql) {
3612 		ibdev_err(&cq->rdev->ibdev, "POLL CQ : no CQL to use");
3613 		goto exit;
3614 	}
3615 	cqe = &cq->cql[0];
3616 	while (budget) {
3617 		lib_qp = NULL;
3618 		ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
3619 		if (lib_qp) {
3620 			sq = &lib_qp->sq;
3621 			if (sq->send_phantom) {
3622 				qp = container_of(lib_qp,
3623 						  struct bnxt_re_qp, qplib_qp);
3624 				if (send_phantom_wqe(qp) == -ENOMEM)
3625 					ibdev_err(&cq->rdev->ibdev,
3626 						  "Phantom failed! Scheduled to send again\n");
3627 				else
3628 					sq->send_phantom = false;
3629 			}
3630 		}
3631 		if (ncqe < budget)
3632 			ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq,
3633 							      cqe + ncqe,
3634 							      budget - ncqe);
3635 
3636 		if (!ncqe)
3637 			break;
3638 
3639 		for (i = 0; i < ncqe; i++, cqe++) {
3640 			/* Transcribe each qplib_wqe back to ib_wc */
3641 			memset(wc, 0, sizeof(*wc));
3642 
3643 			wc->wr_id = cqe->wr_id;
3644 			wc->byte_len = cqe->length;
3645 			qp = container_of
3646 				((struct bnxt_qplib_qp *)
3647 				 (unsigned long)(cqe->qp_handle),
3648 				 struct bnxt_re_qp, qplib_qp);
3649 			wc->qp = &qp->ib_qp;
3650 			wc->ex.imm_data = cqe->immdata;
3651 			wc->src_qp = cqe->src_qp;
3652 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
3653 			wc->port_num = 1;
3654 			wc->vendor_err = cqe->status;
3655 
3656 			switch (cqe->opcode) {
3657 			case CQ_BASE_CQE_TYPE_REQ:
3658 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3659 				if (sh_qp &&
3660 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3661 					/* Handle this completion with
3662 					 * the stored completion
3663 					 */
3664 					memset(wc, 0, sizeof(*wc));
3665 					continue;
3666 				}
3667 				bnxt_re_process_req_wc(wc, cqe);
3668 				break;
3669 			case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1:
3670 				if (!cqe->status) {
3671 					int rc = 0;
3672 
3673 					rc = bnxt_re_process_raw_qp_pkt_rx
3674 								(qp, cqe);
3675 					if (!rc) {
3676 						memset(wc, 0, sizeof(*wc));
3677 						continue;
3678 					}
3679 					cqe->status = -1;
3680 				}
3681 				/* Errors need not be looped back.
3682 				 * But change the wr_id to the one
3683 				 * stored in the table
3684 				 */
3685 				tbl_idx = cqe->wr_id;
3686 				sqp_entry = &cq->rdev->gsi_ctx.sqp_tbl[tbl_idx];
3687 				wc->wr_id = sqp_entry->wrid;
3688 				bnxt_re_process_res_rawqp1_wc(wc, cqe);
3689 				break;
3690 			case CQ_BASE_CQE_TYPE_RES_RC:
3691 				bnxt_re_process_res_rc_wc(wc, cqe);
3692 				break;
3693 			case CQ_BASE_CQE_TYPE_RES_UD:
3694 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3695 				if (sh_qp &&
3696 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3697 					/* Handle this completion with
3698 					 * the stored completion
3699 					 */
3700 					if (cqe->status) {
3701 						continue;
3702 					} else {
3703 						bnxt_re_process_res_shadow_qp_wc
3704 								(qp, wc, cqe);
3705 						break;
3706 					}
3707 				}
3708 				bnxt_re_process_res_ud_wc(qp, wc, cqe);
3709 				break;
3710 			default:
3711 				ibdev_err(&cq->rdev->ibdev,
3712 					  "POLL CQ : type 0x%x not handled",
3713 					  cqe->opcode);
3714 				continue;
3715 			}
3716 			wc++;
3717 			budget--;
3718 		}
3719 	}
3720 exit:
3721 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3722 	return num_entries - budget;
3723 }
3724 
3725 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq,
3726 			  enum ib_cq_notify_flags ib_cqn_flags)
3727 {
3728 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3729 	int type = 0, rc = 0;
3730 	unsigned long flags;
3731 
3732 	spin_lock_irqsave(&cq->cq_lock, flags);
3733 	/* Trigger on the very next completion */
3734 	if (ib_cqn_flags & IB_CQ_NEXT_COMP)
3735 		type = DBC_DBC_TYPE_CQ_ARMALL;
3736 	/* Trigger on the next solicited completion */
3737 	else if (ib_cqn_flags & IB_CQ_SOLICITED)
3738 		type = DBC_DBC_TYPE_CQ_ARMSE;
3739 
3740 	/* Poll to see if there are missed events */
3741 	if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3742 	    !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) {
3743 		rc = 1;
3744 		goto exit;
3745 	}
3746 	bnxt_qplib_req_notify_cq(&cq->qplib_cq, type);
3747 
3748 exit:
3749 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3750 	return rc;
3751 }
3752 
3753 /* Memory Regions */
3754 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags)
3755 {
3756 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3757 	struct bnxt_re_dev *rdev = pd->rdev;
3758 	struct bnxt_re_mr *mr;
3759 	int rc;
3760 
3761 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3762 	if (!mr)
3763 		return ERR_PTR(-ENOMEM);
3764 
3765 	mr->rdev = rdev;
3766 	mr->qplib_mr.pd = &pd->qplib_pd;
3767 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3768 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3769 
3770 	/* Allocate and register 0 as the address */
3771 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3772 	if (rc)
3773 		goto fail;
3774 
3775 	mr->qplib_mr.hwq.level = PBL_LVL_MAX;
3776 	mr->qplib_mr.total_size = -1; /* Infinte length */
3777 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL, 0,
3778 			       PAGE_SIZE);
3779 	if (rc)
3780 		goto fail_mr;
3781 
3782 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3783 	if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ |
3784 			       IB_ACCESS_REMOTE_ATOMIC))
3785 		mr->ib_mr.rkey = mr->ib_mr.lkey;
3786 	atomic_inc(&rdev->mr_count);
3787 
3788 	return &mr->ib_mr;
3789 
3790 fail_mr:
3791 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3792 fail:
3793 	kfree(mr);
3794 	return ERR_PTR(rc);
3795 }
3796 
3797 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3798 {
3799 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3800 	struct bnxt_re_dev *rdev = mr->rdev;
3801 	int rc;
3802 
3803 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3804 	if (rc) {
3805 		ibdev_err(&rdev->ibdev, "Dereg MR failed: %#x\n", rc);
3806 		return rc;
3807 	}
3808 
3809 	if (mr->pages) {
3810 		rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
3811 							&mr->qplib_frpl);
3812 		kfree(mr->pages);
3813 		mr->npages = 0;
3814 		mr->pages = NULL;
3815 	}
3816 	ib_umem_release(mr->ib_umem);
3817 
3818 	kfree(mr);
3819 	atomic_dec(&rdev->mr_count);
3820 	return rc;
3821 }
3822 
3823 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr)
3824 {
3825 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3826 
3827 	if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs))
3828 		return -ENOMEM;
3829 
3830 	mr->pages[mr->npages++] = addr;
3831 	return 0;
3832 }
3833 
3834 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents,
3835 		      unsigned int *sg_offset)
3836 {
3837 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3838 
3839 	mr->npages = 0;
3840 	return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page);
3841 }
3842 
3843 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type,
3844 			       u32 max_num_sg)
3845 {
3846 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3847 	struct bnxt_re_dev *rdev = pd->rdev;
3848 	struct bnxt_re_mr *mr = NULL;
3849 	int rc;
3850 
3851 	if (type != IB_MR_TYPE_MEM_REG) {
3852 		ibdev_dbg(&rdev->ibdev, "MR type 0x%x not supported", type);
3853 		return ERR_PTR(-EINVAL);
3854 	}
3855 	if (max_num_sg > MAX_PBL_LVL_1_PGS)
3856 		return ERR_PTR(-EINVAL);
3857 
3858 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3859 	if (!mr)
3860 		return ERR_PTR(-ENOMEM);
3861 
3862 	mr->rdev = rdev;
3863 	mr->qplib_mr.pd = &pd->qplib_pd;
3864 	mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR;
3865 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3866 
3867 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3868 	if (rc)
3869 		goto bail;
3870 
3871 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3872 	mr->ib_mr.rkey = mr->ib_mr.lkey;
3873 
3874 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3875 	if (!mr->pages) {
3876 		rc = -ENOMEM;
3877 		goto fail;
3878 	}
3879 	rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res,
3880 						 &mr->qplib_frpl, max_num_sg);
3881 	if (rc) {
3882 		ibdev_err(&rdev->ibdev,
3883 			  "Failed to allocate HW FR page list");
3884 		goto fail_mr;
3885 	}
3886 
3887 	atomic_inc(&rdev->mr_count);
3888 	return &mr->ib_mr;
3889 
3890 fail_mr:
3891 	kfree(mr->pages);
3892 fail:
3893 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3894 bail:
3895 	kfree(mr);
3896 	return ERR_PTR(rc);
3897 }
3898 
3899 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
3900 			       struct ib_udata *udata)
3901 {
3902 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3903 	struct bnxt_re_dev *rdev = pd->rdev;
3904 	struct bnxt_re_mw *mw;
3905 	int rc;
3906 
3907 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
3908 	if (!mw)
3909 		return ERR_PTR(-ENOMEM);
3910 	mw->rdev = rdev;
3911 	mw->qplib_mw.pd = &pd->qplib_pd;
3912 
3913 	mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
3914 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
3915 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
3916 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
3917 	if (rc) {
3918 		ibdev_err(&rdev->ibdev, "Allocate MW failed!");
3919 		goto fail;
3920 	}
3921 	mw->ib_mw.rkey = mw->qplib_mw.rkey;
3922 
3923 	atomic_inc(&rdev->mw_count);
3924 	return &mw->ib_mw;
3925 
3926 fail:
3927 	kfree(mw);
3928 	return ERR_PTR(rc);
3929 }
3930 
3931 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
3932 {
3933 	struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
3934 	struct bnxt_re_dev *rdev = mw->rdev;
3935 	int rc;
3936 
3937 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
3938 	if (rc) {
3939 		ibdev_err(&rdev->ibdev, "Free MW failed: %#x\n", rc);
3940 		return rc;
3941 	}
3942 
3943 	kfree(mw);
3944 	atomic_dec(&rdev->mw_count);
3945 	return rc;
3946 }
3947 
3948 /* uverbs */
3949 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
3950 				  u64 virt_addr, int mr_access_flags,
3951 				  struct ib_udata *udata)
3952 {
3953 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3954 	struct bnxt_re_dev *rdev = pd->rdev;
3955 	struct bnxt_re_mr *mr;
3956 	struct ib_umem *umem;
3957 	unsigned long page_size;
3958 	int umem_pgs, rc;
3959 
3960 	if (length > BNXT_RE_MAX_MR_SIZE) {
3961 		ibdev_err(&rdev->ibdev, "MR Size: %lld > Max supported:%lld\n",
3962 			  length, BNXT_RE_MAX_MR_SIZE);
3963 		return ERR_PTR(-ENOMEM);
3964 	}
3965 
3966 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3967 	if (!mr)
3968 		return ERR_PTR(-ENOMEM);
3969 
3970 	mr->rdev = rdev;
3971 	mr->qplib_mr.pd = &pd->qplib_pd;
3972 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3973 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
3974 
3975 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3976 	if (rc) {
3977 		ibdev_err(&rdev->ibdev, "Failed to allocate MR");
3978 		goto free_mr;
3979 	}
3980 	/* The fixed portion of the rkey is the same as the lkey */
3981 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
3982 
3983 	umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags);
3984 	if (IS_ERR(umem)) {
3985 		ibdev_err(&rdev->ibdev, "Failed to get umem");
3986 		rc = -EFAULT;
3987 		goto free_mrw;
3988 	}
3989 	mr->ib_umem = umem;
3990 
3991 	mr->qplib_mr.va = virt_addr;
3992 	page_size = ib_umem_find_best_pgsz(
3993 		umem, BNXT_RE_PAGE_SIZE_SUPPORTED, virt_addr);
3994 	if (!page_size) {
3995 		ibdev_err(&rdev->ibdev, "umem page size unsupported!");
3996 		rc = -EFAULT;
3997 		goto free_umem;
3998 	}
3999 	mr->qplib_mr.total_size = length;
4000 
4001 	umem_pgs = ib_umem_num_dma_blocks(umem, page_size);
4002 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, umem,
4003 			       umem_pgs, page_size);
4004 	if (rc) {
4005 		ibdev_err(&rdev->ibdev, "Failed to register user MR");
4006 		goto free_umem;
4007 	}
4008 
4009 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
4010 	mr->ib_mr.rkey = mr->qplib_mr.lkey;
4011 	atomic_inc(&rdev->mr_count);
4012 
4013 	return &mr->ib_mr;
4014 free_umem:
4015 	ib_umem_release(umem);
4016 free_mrw:
4017 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
4018 free_mr:
4019 	kfree(mr);
4020 	return ERR_PTR(rc);
4021 }
4022 
4023 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
4024 {
4025 	struct ib_device *ibdev = ctx->device;
4026 	struct bnxt_re_ucontext *uctx =
4027 		container_of(ctx, struct bnxt_re_ucontext, ib_uctx);
4028 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
4029 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
4030 	struct bnxt_re_user_mmap_entry *entry;
4031 	struct bnxt_re_uctx_resp resp = {};
4032 	u32 chip_met_rev_num = 0;
4033 	int rc;
4034 
4035 	ibdev_dbg(ibdev, "ABI version requested %u", ibdev->ops.uverbs_abi_ver);
4036 
4037 	if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) {
4038 		ibdev_dbg(ibdev, " is different from the device %d ",
4039 			  BNXT_RE_ABI_VERSION);
4040 		return -EPERM;
4041 	}
4042 
4043 	uctx->rdev = rdev;
4044 
4045 	uctx->shpg = (void *)__get_free_page(GFP_KERNEL);
4046 	if (!uctx->shpg) {
4047 		rc = -ENOMEM;
4048 		goto fail;
4049 	}
4050 	spin_lock_init(&uctx->sh_lock);
4051 
4052 	resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX;
4053 	chip_met_rev_num = rdev->chip_ctx->chip_num;
4054 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_rev & 0xFF) <<
4055 			     BNXT_RE_CHIP_ID0_CHIP_REV_SFT;
4056 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_metal & 0xFF) <<
4057 			     BNXT_RE_CHIP_ID0_CHIP_MET_SFT;
4058 	resp.chip_id0 = chip_met_rev_num;
4059 	/*Temp, Use xa_alloc instead */
4060 	resp.dev_id = rdev->en_dev->pdev->devfn;
4061 	resp.max_qp = rdev->qplib_ctx.qpc_count;
4062 	resp.pg_size = PAGE_SIZE;
4063 	resp.cqe_sz = sizeof(struct cq_base);
4064 	resp.max_cqd = dev_attr->max_cq_wqes;
4065 
4066 	resp.comp_mask |= BNXT_RE_UCNTX_CMASK_HAVE_MODE;
4067 	resp.mode = rdev->chip_ctx->modes.wqe_mode;
4068 
4069 	if (rdev->chip_ctx->modes.db_push)
4070 		resp.comp_mask |= BNXT_RE_UCNTX_CMASK_WC_DPI_ENABLED;
4071 
4072 	entry = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_SH_PAGE, NULL);
4073 	if (!entry) {
4074 		rc = -ENOMEM;
4075 		goto cfail;
4076 	}
4077 	uctx->shpage_mmap = &entry->rdma_entry;
4078 
4079 	rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
4080 	if (rc) {
4081 		ibdev_err(ibdev, "Failed to copy user context");
4082 		rc = -EFAULT;
4083 		goto cfail;
4084 	}
4085 
4086 	return 0;
4087 cfail:
4088 	free_page((unsigned long)uctx->shpg);
4089 	uctx->shpg = NULL;
4090 fail:
4091 	return rc;
4092 }
4093 
4094 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
4095 {
4096 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
4097 						   struct bnxt_re_ucontext,
4098 						   ib_uctx);
4099 
4100 	struct bnxt_re_dev *rdev = uctx->rdev;
4101 
4102 	rdma_user_mmap_entry_remove(uctx->shpage_mmap);
4103 	uctx->shpage_mmap = NULL;
4104 	if (uctx->shpg)
4105 		free_page((unsigned long)uctx->shpg);
4106 
4107 	if (uctx->dpi.dbr) {
4108 		/* Free DPI only if this is the first PD allocated by the
4109 		 * application and mark the context dpi as NULL
4110 		 */
4111 		bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->dpi);
4112 		uctx->dpi.dbr = NULL;
4113 	}
4114 }
4115 
4116 /* Helper function to mmap the virtual memory from user app */
4117 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
4118 {
4119 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
4120 						   struct bnxt_re_ucontext,
4121 						   ib_uctx);
4122 	struct bnxt_re_user_mmap_entry *bnxt_entry;
4123 	struct rdma_user_mmap_entry *rdma_entry;
4124 	int ret = 0;
4125 	u64 pfn;
4126 
4127 	rdma_entry = rdma_user_mmap_entry_get(&uctx->ib_uctx, vma);
4128 	if (!rdma_entry)
4129 		return -EINVAL;
4130 
4131 	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
4132 				  rdma_entry);
4133 
4134 	switch (bnxt_entry->mmap_flag) {
4135 	case BNXT_RE_MMAP_WC_DB:
4136 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4137 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4138 					pgprot_writecombine(vma->vm_page_prot),
4139 					rdma_entry);
4140 		break;
4141 	case BNXT_RE_MMAP_UC_DB:
4142 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4143 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4144 					pgprot_noncached(vma->vm_page_prot),
4145 				rdma_entry);
4146 		break;
4147 	case BNXT_RE_MMAP_SH_PAGE:
4148 		ret = vm_insert_page(vma, vma->vm_start, virt_to_page(uctx->shpg));
4149 		break;
4150 	default:
4151 		ret = -EINVAL;
4152 		break;
4153 	}
4154 
4155 	rdma_user_mmap_entry_put(rdma_entry);
4156 	return ret;
4157 }
4158 
4159 void bnxt_re_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
4160 {
4161 	struct bnxt_re_user_mmap_entry *bnxt_entry;
4162 
4163 	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
4164 				  rdma_entry);
4165 
4166 	kfree(bnxt_entry);
4167 }
4168 
4169 static int UVERBS_HANDLER(BNXT_RE_METHOD_ALLOC_PAGE)(struct uverbs_attr_bundle *attrs)
4170 {
4171 	struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
4172 	enum bnxt_re_alloc_page_type alloc_type;
4173 	struct bnxt_re_user_mmap_entry *entry;
4174 	enum bnxt_re_mmap_flag mmap_flag;
4175 	struct bnxt_qplib_chip_ctx *cctx;
4176 	struct bnxt_re_ucontext *uctx;
4177 	struct bnxt_re_dev *rdev;
4178 	u64 mmap_offset;
4179 	u32 length;
4180 	u32 dpi;
4181 	u64 dbr;
4182 	int err;
4183 
4184 	uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx);
4185 	if (IS_ERR(uctx))
4186 		return PTR_ERR(uctx);
4187 
4188 	err = uverbs_get_const(&alloc_type, attrs, BNXT_RE_ALLOC_PAGE_TYPE);
4189 	if (err)
4190 		return err;
4191 
4192 	rdev = uctx->rdev;
4193 	cctx = rdev->chip_ctx;
4194 
4195 	switch (alloc_type) {
4196 	case BNXT_RE_ALLOC_WC_PAGE:
4197 		if (cctx->modes.db_push)  {
4198 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res, &uctx->wcdpi,
4199 						 uctx, BNXT_QPLIB_DPI_TYPE_WC))
4200 				return -ENOMEM;
4201 			length = PAGE_SIZE;
4202 			dpi = uctx->wcdpi.dpi;
4203 			dbr = (u64)uctx->wcdpi.umdbr;
4204 			mmap_flag = BNXT_RE_MMAP_WC_DB;
4205 		} else {
4206 			return -EINVAL;
4207 		}
4208 
4209 		break;
4210 
4211 	default:
4212 		return -EOPNOTSUPP;
4213 	}
4214 
4215 	entry = bnxt_re_mmap_entry_insert(uctx, dbr, mmap_flag, &mmap_offset);
4216 	if (!entry)
4217 		return -ENOMEM;
4218 
4219 	uobj->object = entry;
4220 	uverbs_finalize_uobj_create(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
4221 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
4222 			     &mmap_offset, sizeof(mmap_offset));
4223 	if (err)
4224 		return err;
4225 
4226 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
4227 			     &length, sizeof(length));
4228 	if (err)
4229 		return err;
4230 
4231 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_DPI,
4232 			     &dpi, sizeof(length));
4233 	if (err)
4234 		return err;
4235 
4236 	return 0;
4237 }
4238 
4239 static int alloc_page_obj_cleanup(struct ib_uobject *uobject,
4240 				  enum rdma_remove_reason why,
4241 			    struct uverbs_attr_bundle *attrs)
4242 {
4243 	struct  bnxt_re_user_mmap_entry *entry = uobject->object;
4244 	struct bnxt_re_ucontext *uctx = entry->uctx;
4245 
4246 	switch (entry->mmap_flag) {
4247 	case BNXT_RE_MMAP_WC_DB:
4248 		if (uctx && uctx->wcdpi.dbr) {
4249 			struct bnxt_re_dev *rdev = uctx->rdev;
4250 
4251 			bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->wcdpi);
4252 			uctx->wcdpi.dbr = NULL;
4253 		}
4254 		break;
4255 	default:
4256 		goto exit;
4257 	}
4258 	rdma_user_mmap_entry_remove(&entry->rdma_entry);
4259 exit:
4260 	return 0;
4261 }
4262 
4263 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_ALLOC_PAGE,
4264 			    UVERBS_ATTR_IDR(BNXT_RE_ALLOC_PAGE_HANDLE,
4265 					    BNXT_RE_OBJECT_ALLOC_PAGE,
4266 					    UVERBS_ACCESS_NEW,
4267 					    UA_MANDATORY),
4268 			    UVERBS_ATTR_CONST_IN(BNXT_RE_ALLOC_PAGE_TYPE,
4269 						 enum bnxt_re_alloc_page_type,
4270 						 UA_MANDATORY),
4271 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
4272 						UVERBS_ATTR_TYPE(u64),
4273 						UA_MANDATORY),
4274 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
4275 						UVERBS_ATTR_TYPE(u32),
4276 						UA_MANDATORY),
4277 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_DPI,
4278 						UVERBS_ATTR_TYPE(u32),
4279 						UA_MANDATORY));
4280 
4281 DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_DESTROY_PAGE,
4282 				    UVERBS_ATTR_IDR(BNXT_RE_DESTROY_PAGE_HANDLE,
4283 						    BNXT_RE_OBJECT_ALLOC_PAGE,
4284 						    UVERBS_ACCESS_DESTROY,
4285 						    UA_MANDATORY));
4286 
4287 DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_ALLOC_PAGE,
4288 			    UVERBS_TYPE_ALLOC_IDR(alloc_page_obj_cleanup),
4289 			    &UVERBS_METHOD(BNXT_RE_METHOD_ALLOC_PAGE),
4290 			    &UVERBS_METHOD(BNXT_RE_METHOD_DESTROY_PAGE));
4291 
4292 const struct uapi_definition bnxt_re_uapi_defs[] = {
4293 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE),
4294 	{}
4295 };
4296