1 /*
2  * Broadcom NetXtreme-E RoCE driver.
3  *
4  * Copyright (c) 2016 - 2017, Broadcom. All rights reserved.  The term
5  * Broadcom refers to Broadcom Limited and/or its subsidiaries.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * BSD license below:
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  *
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in
21  *    the documentation and/or other materials provided with the
22  *    distribution.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS''
25  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
26  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
31  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
32  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
33  * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
34  * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35  *
36  * Description: IB Verbs interpreter
37  */
38 
39 #include <linux/interrupt.h>
40 #include <linux/types.h>
41 #include <linux/pci.h>
42 #include <linux/netdevice.h>
43 #include <linux/if_ether.h>
44 
45 #include <rdma/ib_verbs.h>
46 #include <rdma/ib_user_verbs.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_mad.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/uverbs_ioctl.h>
52 
53 #include "bnxt_ulp.h"
54 
55 #include "roce_hsi.h"
56 #include "qplib_res.h"
57 #include "qplib_sp.h"
58 #include "qplib_fp.h"
59 #include "qplib_rcfw.h"
60 
61 #include "bnxt_re.h"
62 #include "ib_verbs.h"
63 #include <rdma/bnxt_re-abi.h>
64 
65 static int __from_ib_access_flags(int iflags)
66 {
67 	int qflags = 0;
68 
69 	if (iflags & IB_ACCESS_LOCAL_WRITE)
70 		qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
71 	if (iflags & IB_ACCESS_REMOTE_READ)
72 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ;
73 	if (iflags & IB_ACCESS_REMOTE_WRITE)
74 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE;
75 	if (iflags & IB_ACCESS_REMOTE_ATOMIC)
76 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC;
77 	if (iflags & IB_ACCESS_MW_BIND)
78 		qflags |= BNXT_QPLIB_ACCESS_MW_BIND;
79 	if (iflags & IB_ZERO_BASED)
80 		qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED;
81 	if (iflags & IB_ACCESS_ON_DEMAND)
82 		qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND;
83 	return qflags;
84 };
85 
86 static enum ib_access_flags __to_ib_access_flags(int qflags)
87 {
88 	enum ib_access_flags iflags = 0;
89 
90 	if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE)
91 		iflags |= IB_ACCESS_LOCAL_WRITE;
92 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE)
93 		iflags |= IB_ACCESS_REMOTE_WRITE;
94 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ)
95 		iflags |= IB_ACCESS_REMOTE_READ;
96 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC)
97 		iflags |= IB_ACCESS_REMOTE_ATOMIC;
98 	if (qflags & BNXT_QPLIB_ACCESS_MW_BIND)
99 		iflags |= IB_ACCESS_MW_BIND;
100 	if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED)
101 		iflags |= IB_ZERO_BASED;
102 	if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND)
103 		iflags |= IB_ACCESS_ON_DEMAND;
104 	return iflags;
105 };
106 
107 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list,
108 			     struct bnxt_qplib_sge *sg_list, int num)
109 {
110 	int i, total = 0;
111 
112 	for (i = 0; i < num; i++) {
113 		sg_list[i].addr = ib_sg_list[i].addr;
114 		sg_list[i].lkey = ib_sg_list[i].lkey;
115 		sg_list[i].size = ib_sg_list[i].length;
116 		total += sg_list[i].size;
117 	}
118 	return total;
119 }
120 
121 /* Device */
122 int bnxt_re_query_device(struct ib_device *ibdev,
123 			 struct ib_device_attr *ib_attr,
124 			 struct ib_udata *udata)
125 {
126 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
127 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
128 
129 	memset(ib_attr, 0, sizeof(*ib_attr));
130 	memcpy(&ib_attr->fw_ver, dev_attr->fw_ver,
131 	       min(sizeof(dev_attr->fw_ver),
132 		   sizeof(ib_attr->fw_ver)));
133 	bnxt_qplib_get_guid(rdev->netdev->dev_addr,
134 			    (u8 *)&ib_attr->sys_image_guid);
135 	ib_attr->max_mr_size = BNXT_RE_MAX_MR_SIZE;
136 	ib_attr->page_size_cap = BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M;
137 
138 	ib_attr->vendor_id = rdev->en_dev->pdev->vendor;
139 	ib_attr->vendor_part_id = rdev->en_dev->pdev->device;
140 	ib_attr->hw_ver = rdev->en_dev->pdev->subsystem_device;
141 	ib_attr->max_qp = dev_attr->max_qp;
142 	ib_attr->max_qp_wr = dev_attr->max_qp_wqes;
143 	ib_attr->device_cap_flags =
144 				    IB_DEVICE_CURR_QP_STATE_MOD
145 				    | IB_DEVICE_RC_RNR_NAK_GEN
146 				    | IB_DEVICE_SHUTDOWN_PORT
147 				    | IB_DEVICE_SYS_IMAGE_GUID
148 				    | IB_DEVICE_LOCAL_DMA_LKEY
149 				    | IB_DEVICE_RESIZE_MAX_WR
150 				    | IB_DEVICE_PORT_ACTIVE_EVENT
151 				    | IB_DEVICE_N_NOTIFY_CQ
152 				    | IB_DEVICE_MEM_WINDOW
153 				    | IB_DEVICE_MEM_WINDOW_TYPE_2B
154 				    | IB_DEVICE_MEM_MGT_EXTENSIONS;
155 	ib_attr->max_send_sge = dev_attr->max_qp_sges;
156 	ib_attr->max_recv_sge = dev_attr->max_qp_sges;
157 	ib_attr->max_sge_rd = dev_attr->max_qp_sges;
158 	ib_attr->max_cq = dev_attr->max_cq;
159 	ib_attr->max_cqe = dev_attr->max_cq_wqes;
160 	ib_attr->max_mr = dev_attr->max_mr;
161 	ib_attr->max_pd = dev_attr->max_pd;
162 	ib_attr->max_qp_rd_atom = dev_attr->max_qp_rd_atom;
163 	ib_attr->max_qp_init_rd_atom = dev_attr->max_qp_init_rd_atom;
164 	ib_attr->atomic_cap = IB_ATOMIC_NONE;
165 	ib_attr->masked_atomic_cap = IB_ATOMIC_NONE;
166 
167 	ib_attr->max_ee_rd_atom = 0;
168 	ib_attr->max_res_rd_atom = 0;
169 	ib_attr->max_ee_init_rd_atom = 0;
170 	ib_attr->max_ee = 0;
171 	ib_attr->max_rdd = 0;
172 	ib_attr->max_mw = dev_attr->max_mw;
173 	ib_attr->max_raw_ipv6_qp = 0;
174 	ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp;
175 	ib_attr->max_mcast_grp = 0;
176 	ib_attr->max_mcast_qp_attach = 0;
177 	ib_attr->max_total_mcast_qp_attach = 0;
178 	ib_attr->max_ah = dev_attr->max_ah;
179 
180 	ib_attr->max_fmr = 0;
181 	ib_attr->max_map_per_fmr = 0;
182 
183 	ib_attr->max_srq = dev_attr->max_srq;
184 	ib_attr->max_srq_wr = dev_attr->max_srq_wqes;
185 	ib_attr->max_srq_sge = dev_attr->max_srq_sges;
186 
187 	ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS;
188 
189 	ib_attr->max_pkeys = 1;
190 	ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY;
191 	return 0;
192 }
193 
194 /* Port */
195 int bnxt_re_query_port(struct ib_device *ibdev, u8 port_num,
196 		       struct ib_port_attr *port_attr)
197 {
198 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
199 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
200 
201 	memset(port_attr, 0, sizeof(*port_attr));
202 
203 	if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) {
204 		port_attr->state = IB_PORT_ACTIVE;
205 		port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
206 	} else {
207 		port_attr->state = IB_PORT_DOWN;
208 		port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
209 	}
210 	port_attr->max_mtu = IB_MTU_4096;
211 	port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu);
212 	port_attr->gid_tbl_len = dev_attr->max_sgid;
213 	port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
214 				    IB_PORT_DEVICE_MGMT_SUP |
215 				    IB_PORT_VENDOR_CLASS_SUP;
216 	port_attr->ip_gids = true;
217 
218 	port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW;
219 	port_attr->bad_pkey_cntr = 0;
220 	port_attr->qkey_viol_cntr = 0;
221 	port_attr->pkey_tbl_len = dev_attr->max_pkey;
222 	port_attr->lid = 0;
223 	port_attr->sm_lid = 0;
224 	port_attr->lmc = 0;
225 	port_attr->max_vl_num = 4;
226 	port_attr->sm_sl = 0;
227 	port_attr->subnet_timeout = 0;
228 	port_attr->init_type_reply = 0;
229 	port_attr->active_speed = rdev->active_speed;
230 	port_attr->active_width = rdev->active_width;
231 
232 	return 0;
233 }
234 
235 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u8 port_num,
236 			       struct ib_port_immutable *immutable)
237 {
238 	struct ib_port_attr port_attr;
239 
240 	if (bnxt_re_query_port(ibdev, port_num, &port_attr))
241 		return -EINVAL;
242 
243 	immutable->pkey_tbl_len = port_attr.pkey_tbl_len;
244 	immutable->gid_tbl_len = port_attr.gid_tbl_len;
245 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
246 	immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP;
247 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
248 	return 0;
249 }
250 
251 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str)
252 {
253 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
254 
255 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d",
256 		 rdev->dev_attr.fw_ver[0], rdev->dev_attr.fw_ver[1],
257 		 rdev->dev_attr.fw_ver[2], rdev->dev_attr.fw_ver[3]);
258 }
259 
260 int bnxt_re_query_pkey(struct ib_device *ibdev, u8 port_num,
261 		       u16 index, u16 *pkey)
262 {
263 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
264 
265 	/* Ignore port_num */
266 
267 	memset(pkey, 0, sizeof(*pkey));
268 	return bnxt_qplib_get_pkey(&rdev->qplib_res,
269 				   &rdev->qplib_res.pkey_tbl, index, pkey);
270 }
271 
272 int bnxt_re_query_gid(struct ib_device *ibdev, u8 port_num,
273 		      int index, union ib_gid *gid)
274 {
275 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
276 	int rc = 0;
277 
278 	/* Ignore port_num */
279 	memset(gid, 0, sizeof(*gid));
280 	rc = bnxt_qplib_get_sgid(&rdev->qplib_res,
281 				 &rdev->qplib_res.sgid_tbl, index,
282 				 (struct bnxt_qplib_gid *)gid);
283 	return rc;
284 }
285 
286 int bnxt_re_del_gid(const struct ib_gid_attr *attr, void **context)
287 {
288 	int rc = 0;
289 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
290 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
291 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
292 	struct bnxt_qplib_gid *gid_to_del;
293 	u16 vlan_id = 0xFFFF;
294 
295 	/* Delete the entry from the hardware */
296 	ctx = *context;
297 	if (!ctx)
298 		return -EINVAL;
299 
300 	if (sgid_tbl && sgid_tbl->active) {
301 		if (ctx->idx >= sgid_tbl->max)
302 			return -EINVAL;
303 		gid_to_del = &sgid_tbl->tbl[ctx->idx].gid;
304 		vlan_id = sgid_tbl->tbl[ctx->idx].vlan_id;
305 		/* DEL_GID is called in WQ context(netdevice_event_work_handler)
306 		 * or via the ib_unregister_device path. In the former case QP1
307 		 * may not be destroyed yet, in which case just return as FW
308 		 * needs that entry to be present and will fail it's deletion.
309 		 * We could get invoked again after QP1 is destroyed OR get an
310 		 * ADD_GID call with a different GID value for the same index
311 		 * where we issue MODIFY_GID cmd to update the GID entry -- TBD
312 		 */
313 		if (ctx->idx == 0 &&
314 		    rdma_link_local_addr((struct in6_addr *)gid_to_del) &&
315 		    ctx->refcnt == 1 && rdev->gsi_ctx.gsi_sqp) {
316 			ibdev_dbg(&rdev->ibdev,
317 				  "Trying to delete GID0 while QP1 is alive\n");
318 			return -EFAULT;
319 		}
320 		ctx->refcnt--;
321 		if (!ctx->refcnt) {
322 			rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del,
323 						 vlan_id,  true);
324 			if (rc) {
325 				ibdev_err(&rdev->ibdev,
326 					  "Failed to remove GID: %#x", rc);
327 			} else {
328 				ctx_tbl = sgid_tbl->ctx;
329 				ctx_tbl[ctx->idx] = NULL;
330 				kfree(ctx);
331 			}
332 		}
333 	} else {
334 		return -EINVAL;
335 	}
336 	return rc;
337 }
338 
339 int bnxt_re_add_gid(const struct ib_gid_attr *attr, void **context)
340 {
341 	int rc;
342 	u32 tbl_idx = 0;
343 	u16 vlan_id = 0xFFFF;
344 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
345 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
346 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
347 
348 	rc = rdma_read_gid_l2_fields(attr, &vlan_id, NULL);
349 	if (rc)
350 		return rc;
351 
352 	rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)&attr->gid,
353 				 rdev->qplib_res.netdev->dev_addr,
354 				 vlan_id, true, &tbl_idx);
355 	if (rc == -EALREADY) {
356 		ctx_tbl = sgid_tbl->ctx;
357 		ctx_tbl[tbl_idx]->refcnt++;
358 		*context = ctx_tbl[tbl_idx];
359 		return 0;
360 	}
361 
362 	if (rc < 0) {
363 		ibdev_err(&rdev->ibdev, "Failed to add GID: %#x", rc);
364 		return rc;
365 	}
366 
367 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
368 	if (!ctx)
369 		return -ENOMEM;
370 	ctx_tbl = sgid_tbl->ctx;
371 	ctx->idx = tbl_idx;
372 	ctx->refcnt = 1;
373 	ctx_tbl[tbl_idx] = ctx;
374 	*context = ctx;
375 
376 	return rc;
377 }
378 
379 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
380 					    u8 port_num)
381 {
382 	return IB_LINK_LAYER_ETHERNET;
383 }
384 
385 #define	BNXT_RE_FENCE_PBL_SIZE	DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
386 
387 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd)
388 {
389 	struct bnxt_re_fence_data *fence = &pd->fence;
390 	struct ib_mr *ib_mr = &fence->mr->ib_mr;
391 	struct bnxt_qplib_swqe *wqe = &fence->bind_wqe;
392 
393 	memset(wqe, 0, sizeof(*wqe));
394 	wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW;
395 	wqe->wr_id = BNXT_QPLIB_FENCE_WRID;
396 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
397 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
398 	wqe->bind.zero_based = false;
399 	wqe->bind.parent_l_key = ib_mr->lkey;
400 	wqe->bind.va = (u64)(unsigned long)fence->va;
401 	wqe->bind.length = fence->size;
402 	wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ);
403 	wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1;
404 
405 	/* Save the initial rkey in fence structure for now;
406 	 * wqe->bind.r_key will be set at (re)bind time.
407 	 */
408 	fence->bind_rkey = ib_inc_rkey(fence->mw->rkey);
409 }
410 
411 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp)
412 {
413 	struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp,
414 					     qplib_qp);
415 	struct ib_pd *ib_pd = qp->ib_qp.pd;
416 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
417 	struct bnxt_re_fence_data *fence = &pd->fence;
418 	struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe;
419 	struct bnxt_qplib_swqe wqe;
420 	int rc;
421 
422 	memcpy(&wqe, fence_wqe, sizeof(wqe));
423 	wqe.bind.r_key = fence->bind_rkey;
424 	fence->bind_rkey = ib_inc_rkey(fence->bind_rkey);
425 
426 	ibdev_dbg(&qp->rdev->ibdev,
427 		  "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
428 		wqe.bind.r_key, qp->qplib_qp.id, pd);
429 	rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
430 	if (rc) {
431 		ibdev_err(&qp->rdev->ibdev, "Failed to bind fence-WQE\n");
432 		return rc;
433 	}
434 	bnxt_qplib_post_send_db(&qp->qplib_qp);
435 
436 	return rc;
437 }
438 
439 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd)
440 {
441 	struct bnxt_re_fence_data *fence = &pd->fence;
442 	struct bnxt_re_dev *rdev = pd->rdev;
443 	struct device *dev = &rdev->en_dev->pdev->dev;
444 	struct bnxt_re_mr *mr = fence->mr;
445 
446 	if (fence->mw) {
447 		bnxt_re_dealloc_mw(fence->mw);
448 		fence->mw = NULL;
449 	}
450 	if (mr) {
451 		if (mr->ib_mr.rkey)
452 			bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr,
453 					     true);
454 		if (mr->ib_mr.lkey)
455 			bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
456 		kfree(mr);
457 		fence->mr = NULL;
458 	}
459 	if (fence->dma_addr) {
460 		dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES,
461 				 DMA_BIDIRECTIONAL);
462 		fence->dma_addr = 0;
463 	}
464 }
465 
466 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
467 {
468 	int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND;
469 	struct bnxt_re_fence_data *fence = &pd->fence;
470 	struct bnxt_re_dev *rdev = pd->rdev;
471 	struct device *dev = &rdev->en_dev->pdev->dev;
472 	struct bnxt_re_mr *mr = NULL;
473 	dma_addr_t dma_addr = 0;
474 	struct ib_mw *mw;
475 	u64 pbl_tbl;
476 	int rc;
477 
478 	dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES,
479 				  DMA_BIDIRECTIONAL);
480 	rc = dma_mapping_error(dev, dma_addr);
481 	if (rc) {
482 		ibdev_err(&rdev->ibdev, "Failed to dma-map fence-MR-mem\n");
483 		rc = -EIO;
484 		fence->dma_addr = 0;
485 		goto fail;
486 	}
487 	fence->dma_addr = dma_addr;
488 
489 	/* Allocate a MR */
490 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
491 	if (!mr) {
492 		rc = -ENOMEM;
493 		goto fail;
494 	}
495 	fence->mr = mr;
496 	mr->rdev = rdev;
497 	mr->qplib_mr.pd = &pd->qplib_pd;
498 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
499 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
500 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
501 	if (rc) {
502 		ibdev_err(&rdev->ibdev, "Failed to alloc fence-HW-MR\n");
503 		goto fail;
504 	}
505 
506 	/* Register MR */
507 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
508 	mr->qplib_mr.va = (u64)(unsigned long)fence->va;
509 	mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES;
510 	pbl_tbl = dma_addr;
511 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl_tbl,
512 			       BNXT_RE_FENCE_PBL_SIZE, false, PAGE_SIZE);
513 	if (rc) {
514 		ibdev_err(&rdev->ibdev, "Failed to register fence-MR\n");
515 		goto fail;
516 	}
517 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
518 
519 	/* Create a fence MW only for kernel consumers */
520 	mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL);
521 	if (IS_ERR(mw)) {
522 		ibdev_err(&rdev->ibdev,
523 			  "Failed to create fence-MW for PD: %p\n", pd);
524 		rc = PTR_ERR(mw);
525 		goto fail;
526 	}
527 	fence->mw = mw;
528 
529 	bnxt_re_create_fence_wqe(pd);
530 	return 0;
531 
532 fail:
533 	bnxt_re_destroy_fence_mr(pd);
534 	return rc;
535 }
536 
537 /* Protection Domains */
538 void bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
539 {
540 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
541 	struct bnxt_re_dev *rdev = pd->rdev;
542 
543 	bnxt_re_destroy_fence_mr(pd);
544 
545 	if (pd->qplib_pd.id)
546 		bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
547 				      &pd->qplib_pd);
548 }
549 
550 int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
551 {
552 	struct ib_device *ibdev = ibpd->device;
553 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
554 	struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context(
555 		udata, struct bnxt_re_ucontext, ib_uctx);
556 	struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd);
557 	int rc;
558 
559 	pd->rdev = rdev;
560 	if (bnxt_qplib_alloc_pd(&rdev->qplib_res.pd_tbl, &pd->qplib_pd)) {
561 		ibdev_err(&rdev->ibdev, "Failed to allocate HW PD");
562 		rc = -ENOMEM;
563 		goto fail;
564 	}
565 
566 	if (udata) {
567 		struct bnxt_re_pd_resp resp;
568 
569 		if (!ucntx->dpi.dbr) {
570 			/* Allocate DPI in alloc_pd to avoid failing of
571 			 * ibv_devinfo and family of application when DPIs
572 			 * are depleted.
573 			 */
574 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res.dpi_tbl,
575 						 &ucntx->dpi, ucntx)) {
576 				rc = -ENOMEM;
577 				goto dbfail;
578 			}
579 		}
580 
581 		resp.pdid = pd->qplib_pd.id;
582 		/* Still allow mapping this DBR to the new user PD. */
583 		resp.dpi = ucntx->dpi.dpi;
584 		resp.dbr = (u64)ucntx->dpi.umdbr;
585 
586 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
587 		if (rc) {
588 			ibdev_err(&rdev->ibdev,
589 				  "Failed to copy user response\n");
590 			goto dbfail;
591 		}
592 	}
593 
594 	if (!udata)
595 		if (bnxt_re_create_fence_mr(pd))
596 			ibdev_warn(&rdev->ibdev,
597 				   "Failed to create Fence-MR\n");
598 	return 0;
599 dbfail:
600 	bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
601 			      &pd->qplib_pd);
602 fail:
603 	return rc;
604 }
605 
606 /* Address Handles */
607 void bnxt_re_destroy_ah(struct ib_ah *ib_ah, u32 flags)
608 {
609 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
610 	struct bnxt_re_dev *rdev = ah->rdev;
611 
612 	bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah,
613 			      !(flags & RDMA_DESTROY_AH_SLEEPABLE));
614 }
615 
616 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)
617 {
618 	u8 nw_type;
619 
620 	switch (ntype) {
621 	case RDMA_NETWORK_IPV4:
622 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4;
623 		break;
624 	case RDMA_NETWORK_IPV6:
625 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6;
626 		break;
627 	default:
628 		nw_type = CMDQ_CREATE_AH_TYPE_V1;
629 		break;
630 	}
631 	return nw_type;
632 }
633 
634 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr,
635 		      u32 flags, struct ib_udata *udata)
636 {
637 	struct ib_pd *ib_pd = ib_ah->pd;
638 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
639 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
640 	struct bnxt_re_dev *rdev = pd->rdev;
641 	const struct ib_gid_attr *sgid_attr;
642 	struct bnxt_re_gid_ctx *ctx;
643 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
644 	u8 nw_type;
645 	int rc;
646 
647 	if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) {
648 		ibdev_err(&rdev->ibdev, "Failed to alloc AH: GRH not set");
649 		return -EINVAL;
650 	}
651 
652 	ah->rdev = rdev;
653 	ah->qplib_ah.pd = &pd->qplib_pd;
654 
655 	/* Supply the configuration for the HW */
656 	memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw,
657 	       sizeof(union ib_gid));
658 	sgid_attr = grh->sgid_attr;
659 	/* Get the HW context of the GID. The reference
660 	 * of GID table entry is already taken by the caller.
661 	 */
662 	ctx = rdma_read_gid_hw_context(sgid_attr);
663 	ah->qplib_ah.sgid_index = ctx->idx;
664 	ah->qplib_ah.host_sgid_index = grh->sgid_index;
665 	ah->qplib_ah.traffic_class = grh->traffic_class;
666 	ah->qplib_ah.flow_label = grh->flow_label;
667 	ah->qplib_ah.hop_limit = grh->hop_limit;
668 	ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr);
669 
670 	/* Get network header type for this GID */
671 	nw_type = rdma_gid_attr_network_type(sgid_attr);
672 	ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type);
673 
674 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
675 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah,
676 				  !(flags & RDMA_CREATE_AH_SLEEPABLE));
677 	if (rc) {
678 		ibdev_err(&rdev->ibdev, "Failed to allocate HW AH");
679 		return rc;
680 	}
681 
682 	/* Write AVID to shared page. */
683 	if (udata) {
684 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
685 			udata, struct bnxt_re_ucontext, ib_uctx);
686 		unsigned long flag;
687 		u32 *wrptr;
688 
689 		spin_lock_irqsave(&uctx->sh_lock, flag);
690 		wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT);
691 		*wrptr = ah->qplib_ah.id;
692 		wmb(); /* make sure cache is updated. */
693 		spin_unlock_irqrestore(&uctx->sh_lock, flag);
694 	}
695 
696 	return 0;
697 }
698 
699 int bnxt_re_modify_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
700 {
701 	return 0;
702 }
703 
704 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
705 {
706 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
707 
708 	ah_attr->type = ib_ah->type;
709 	rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl);
710 	memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN);
711 	rdma_ah_set_grh(ah_attr, NULL, 0,
712 			ah->qplib_ah.host_sgid_index,
713 			0, ah->qplib_ah.traffic_class);
714 	rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data);
715 	rdma_ah_set_port_num(ah_attr, 1);
716 	rdma_ah_set_static_rate(ah_attr, 0);
717 	return 0;
718 }
719 
720 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp)
721 	__acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock)
722 {
723 	unsigned long flags;
724 
725 	spin_lock_irqsave(&qp->scq->cq_lock, flags);
726 	if (qp->rcq != qp->scq)
727 		spin_lock(&qp->rcq->cq_lock);
728 	else
729 		__acquire(&qp->rcq->cq_lock);
730 
731 	return flags;
732 }
733 
734 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp,
735 			unsigned long flags)
736 	__releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock)
737 {
738 	if (qp->rcq != qp->scq)
739 		spin_unlock(&qp->rcq->cq_lock);
740 	else
741 		__release(&qp->rcq->cq_lock);
742 	spin_unlock_irqrestore(&qp->scq->cq_lock, flags);
743 }
744 
745 static int bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp *qp)
746 {
747 	struct bnxt_re_qp *gsi_sqp;
748 	struct bnxt_re_ah *gsi_sah;
749 	struct bnxt_re_dev *rdev;
750 	int rc = 0;
751 
752 	rdev = qp->rdev;
753 	gsi_sqp = rdev->gsi_ctx.gsi_sqp;
754 	gsi_sah = rdev->gsi_ctx.gsi_sah;
755 
756 	/* remove from active qp list */
757 	mutex_lock(&rdev->qp_lock);
758 	list_del(&gsi_sqp->list);
759 	mutex_unlock(&rdev->qp_lock);
760 	atomic_dec(&rdev->qp_count);
761 
762 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow AH\n");
763 	bnxt_qplib_destroy_ah(&rdev->qplib_res,
764 			      &gsi_sah->qplib_ah,
765 			      true);
766 	bnxt_qplib_clean_qp(&qp->qplib_qp);
767 
768 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow QP\n");
769 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &gsi_sqp->qplib_qp);
770 	if (rc) {
771 		ibdev_err(&rdev->ibdev, "Destroy Shadow QP failed");
772 		goto fail;
773 	}
774 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &gsi_sqp->qplib_qp);
775 
776 	kfree(rdev->gsi_ctx.sqp_tbl);
777 	kfree(gsi_sah);
778 	kfree(gsi_sqp);
779 	rdev->gsi_ctx.gsi_sqp = NULL;
780 	rdev->gsi_ctx.gsi_sah = NULL;
781 	rdev->gsi_ctx.sqp_tbl = NULL;
782 
783 	return 0;
784 fail:
785 	return rc;
786 }
787 
788 /* Queue Pairs */
789 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
790 {
791 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
792 	struct bnxt_re_dev *rdev = qp->rdev;
793 	unsigned int flags;
794 	int rc;
795 
796 	mutex_lock(&rdev->qp_lock);
797 	list_del(&qp->list);
798 	mutex_unlock(&rdev->qp_lock);
799 	atomic_dec(&rdev->qp_count);
800 
801 	bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
802 
803 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
804 	if (rc) {
805 		ibdev_err(&rdev->ibdev, "Failed to destroy HW QP");
806 		return rc;
807 	}
808 
809 	if (rdma_is_kernel_res(&qp->ib_qp.res)) {
810 		flags = bnxt_re_lock_cqs(qp);
811 		bnxt_qplib_clean_qp(&qp->qplib_qp);
812 		bnxt_re_unlock_cqs(qp, flags);
813 	}
814 
815 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
816 
817 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) {
818 		rc = bnxt_re_destroy_gsi_sqp(qp);
819 		if (rc)
820 			goto sh_fail;
821 	}
822 
823 	ib_umem_release(qp->rumem);
824 	ib_umem_release(qp->sumem);
825 
826 	kfree(qp);
827 	return 0;
828 sh_fail:
829 	return rc;
830 }
831 
832 static u8 __from_ib_qp_type(enum ib_qp_type type)
833 {
834 	switch (type) {
835 	case IB_QPT_GSI:
836 		return CMDQ_CREATE_QP1_TYPE_GSI;
837 	case IB_QPT_RC:
838 		return CMDQ_CREATE_QP_TYPE_RC;
839 	case IB_QPT_UD:
840 		return CMDQ_CREATE_QP_TYPE_UD;
841 	default:
842 		return IB_QPT_MAX;
843 	}
844 }
845 
846 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
847 				struct bnxt_re_qp *qp, struct ib_udata *udata)
848 {
849 	struct bnxt_re_qp_req ureq;
850 	struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
851 	struct ib_umem *umem;
852 	int bytes = 0, psn_sz;
853 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
854 		udata, struct bnxt_re_ucontext, ib_uctx);
855 
856 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
857 		return -EFAULT;
858 
859 	bytes = (qplib_qp->sq.max_wqe * BNXT_QPLIB_MAX_SQE_ENTRY_SIZE);
860 	/* Consider mapping PSN search memory only for RC QPs. */
861 	if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) {
862 		psn_sz = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
863 					sizeof(struct sq_psn_search_ext) :
864 					sizeof(struct sq_psn_search);
865 		bytes += (qplib_qp->sq.max_wqe * psn_sz);
866 	}
867 	bytes = PAGE_ALIGN(bytes);
868 	umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes,
869 			   IB_ACCESS_LOCAL_WRITE);
870 	if (IS_ERR(umem))
871 		return PTR_ERR(umem);
872 
873 	qp->sumem = umem;
874 	qplib_qp->sq.sg_info.sghead = umem->sg_head.sgl;
875 	qplib_qp->sq.sg_info.npages = ib_umem_num_pages(umem);
876 	qplib_qp->sq.sg_info.nmap = umem->nmap;
877 	qplib_qp->sq.sg_info.pgsize = PAGE_SIZE;
878 	qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT;
879 	qplib_qp->qp_handle = ureq.qp_handle;
880 
881 	if (!qp->qplib_qp.srq) {
882 		bytes = (qplib_qp->rq.max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
883 		bytes = PAGE_ALIGN(bytes);
884 		umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes,
885 				   IB_ACCESS_LOCAL_WRITE);
886 		if (IS_ERR(umem))
887 			goto rqfail;
888 		qp->rumem = umem;
889 		qplib_qp->rq.sg_info.sghead = umem->sg_head.sgl;
890 		qplib_qp->rq.sg_info.npages = ib_umem_num_pages(umem);
891 		qplib_qp->rq.sg_info.nmap = umem->nmap;
892 		qplib_qp->rq.sg_info.pgsize = PAGE_SIZE;
893 		qplib_qp->rq.sg_info.pgshft = PAGE_SHIFT;
894 	}
895 
896 	qplib_qp->dpi = &cntx->dpi;
897 	return 0;
898 rqfail:
899 	ib_umem_release(qp->sumem);
900 	qp->sumem = NULL;
901 	memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info));
902 
903 	return PTR_ERR(umem);
904 }
905 
906 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
907 				(struct bnxt_re_pd *pd,
908 				 struct bnxt_qplib_res *qp1_res,
909 				 struct bnxt_qplib_qp *qp1_qp)
910 {
911 	struct bnxt_re_dev *rdev = pd->rdev;
912 	struct bnxt_re_ah *ah;
913 	union ib_gid sgid;
914 	int rc;
915 
916 	ah = kzalloc(sizeof(*ah), GFP_KERNEL);
917 	if (!ah)
918 		return NULL;
919 
920 	ah->rdev = rdev;
921 	ah->qplib_ah.pd = &pd->qplib_pd;
922 
923 	rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid);
924 	if (rc)
925 		goto fail;
926 
927 	/* supply the dgid data same as sgid */
928 	memcpy(ah->qplib_ah.dgid.data, &sgid.raw,
929 	       sizeof(union ib_gid));
930 	ah->qplib_ah.sgid_index = 0;
931 
932 	ah->qplib_ah.traffic_class = 0;
933 	ah->qplib_ah.flow_label = 0;
934 	ah->qplib_ah.hop_limit = 1;
935 	ah->qplib_ah.sl = 0;
936 	/* Have DMAC same as SMAC */
937 	ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr);
938 
939 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false);
940 	if (rc) {
941 		ibdev_err(&rdev->ibdev,
942 			  "Failed to allocate HW AH for Shadow QP");
943 		goto fail;
944 	}
945 
946 	return ah;
947 
948 fail:
949 	kfree(ah);
950 	return NULL;
951 }
952 
953 static struct bnxt_re_qp *bnxt_re_create_shadow_qp
954 				(struct bnxt_re_pd *pd,
955 				 struct bnxt_qplib_res *qp1_res,
956 				 struct bnxt_qplib_qp *qp1_qp)
957 {
958 	struct bnxt_re_dev *rdev = pd->rdev;
959 	struct bnxt_re_qp *qp;
960 	int rc;
961 
962 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
963 	if (!qp)
964 		return NULL;
965 
966 	qp->rdev = rdev;
967 
968 	/* Initialize the shadow QP structure from the QP1 values */
969 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
970 
971 	qp->qplib_qp.pd = &pd->qplib_pd;
972 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
973 	qp->qplib_qp.type = IB_QPT_UD;
974 
975 	qp->qplib_qp.max_inline_data = 0;
976 	qp->qplib_qp.sig_type = true;
977 
978 	/* Shadow QP SQ depth should be same as QP1 RQ depth */
979 	qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
980 	qp->qplib_qp.sq.max_sge = 2;
981 	/* Q full delta can be 1 since it is internal QP */
982 	qp->qplib_qp.sq.q_full_delta = 1;
983 	qp->qplib_qp.sq.sg_info.pgsize = PAGE_SIZE;
984 	qp->qplib_qp.sq.sg_info.pgshft = PAGE_SHIFT;
985 
986 	qp->qplib_qp.scq = qp1_qp->scq;
987 	qp->qplib_qp.rcq = qp1_qp->rcq;
988 
989 	qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
990 	qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
991 	/* Q full delta can be 1 since it is internal QP */
992 	qp->qplib_qp.rq.q_full_delta = 1;
993 	qp->qplib_qp.rq.sg_info.pgsize = PAGE_SIZE;
994 	qp->qplib_qp.rq.sg_info.pgshft = PAGE_SHIFT;
995 
996 	qp->qplib_qp.mtu = qp1_qp->mtu;
997 
998 	qp->qplib_qp.sq_hdr_buf_size = 0;
999 	qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
1000 	qp->qplib_qp.dpi = &rdev->dpi_privileged;
1001 
1002 	rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
1003 	if (rc)
1004 		goto fail;
1005 
1006 	spin_lock_init(&qp->sq_lock);
1007 	INIT_LIST_HEAD(&qp->list);
1008 	mutex_lock(&rdev->qp_lock);
1009 	list_add_tail(&qp->list, &rdev->qp_list);
1010 	atomic_inc(&rdev->qp_count);
1011 	mutex_unlock(&rdev->qp_lock);
1012 	return qp;
1013 fail:
1014 	kfree(qp);
1015 	return NULL;
1016 }
1017 
1018 static int bnxt_re_init_rq_attr(struct bnxt_re_qp *qp,
1019 				struct ib_qp_init_attr *init_attr)
1020 {
1021 	struct bnxt_qplib_dev_attr *dev_attr;
1022 	struct bnxt_qplib_qp *qplqp;
1023 	struct bnxt_re_dev *rdev;
1024 	int entries;
1025 
1026 	rdev = qp->rdev;
1027 	qplqp = &qp->qplib_qp;
1028 	dev_attr = &rdev->dev_attr;
1029 
1030 	if (init_attr->srq) {
1031 		struct bnxt_re_srq *srq;
1032 
1033 		srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq);
1034 		if (!srq) {
1035 			ibdev_err(&rdev->ibdev, "SRQ not found");
1036 			return -EINVAL;
1037 		}
1038 		qplqp->srq = &srq->qplib_srq;
1039 		qplqp->rq.max_wqe = 0;
1040 	} else {
1041 		/* Allocate 1 more than what's provided so posting max doesn't
1042 		 * mean empty.
1043 		 */
1044 		entries = roundup_pow_of_two(init_attr->cap.max_recv_wr + 1);
1045 		qplqp->rq.max_wqe = min_t(u32, entries,
1046 					  dev_attr->max_qp_wqes + 1);
1047 
1048 		qplqp->rq.q_full_delta = qplqp->rq.max_wqe -
1049 					 init_attr->cap.max_recv_wr;
1050 		qplqp->rq.max_sge = init_attr->cap.max_recv_sge;
1051 		if (qplqp->rq.max_sge > dev_attr->max_qp_sges)
1052 			qplqp->rq.max_sge = dev_attr->max_qp_sges;
1053 	}
1054 	qplqp->rq.sg_info.pgsize = PAGE_SIZE;
1055 	qplqp->rq.sg_info.pgshft = PAGE_SHIFT;
1056 
1057 	return 0;
1058 }
1059 
1060 static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp)
1061 {
1062 	struct bnxt_qplib_dev_attr *dev_attr;
1063 	struct bnxt_qplib_qp *qplqp;
1064 	struct bnxt_re_dev *rdev;
1065 
1066 	rdev = qp->rdev;
1067 	qplqp = &qp->qplib_qp;
1068 	dev_attr = &rdev->dev_attr;
1069 
1070 	qplqp->rq.max_sge = dev_attr->max_qp_sges;
1071 	if (qplqp->rq.max_sge > dev_attr->max_qp_sges)
1072 		qplqp->rq.max_sge = dev_attr->max_qp_sges;
1073 	qplqp->rq.max_sge = 6;
1074 }
1075 
1076 static void bnxt_re_init_sq_attr(struct bnxt_re_qp *qp,
1077 				 struct ib_qp_init_attr *init_attr,
1078 				 struct ib_udata *udata)
1079 {
1080 	struct bnxt_qplib_dev_attr *dev_attr;
1081 	struct bnxt_qplib_qp *qplqp;
1082 	struct bnxt_re_dev *rdev;
1083 	int entries;
1084 
1085 	rdev = qp->rdev;
1086 	qplqp = &qp->qplib_qp;
1087 	dev_attr = &rdev->dev_attr;
1088 
1089 	qplqp->sq.max_sge = init_attr->cap.max_send_sge;
1090 	if (qplqp->sq.max_sge > dev_attr->max_qp_sges)
1091 		qplqp->sq.max_sge = dev_attr->max_qp_sges;
1092 	/*
1093 	 * Change the SQ depth if user has requested minimum using
1094 	 * configfs. Only supported for kernel consumers
1095 	 */
1096 	entries = init_attr->cap.max_send_wr;
1097 	/* Allocate 128 + 1 more than what's provided */
1098 	entries = roundup_pow_of_two(entries + BNXT_QPLIB_RESERVED_QP_WRS + 1);
1099 	qplqp->sq.max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes +
1100 			BNXT_QPLIB_RESERVED_QP_WRS + 1);
1101 	qplqp->sq.q_full_delta = BNXT_QPLIB_RESERVED_QP_WRS + 1;
1102 	/*
1103 	 * Reserving one slot for Phantom WQE. Application can
1104 	 * post one extra entry in this case. But allowing this to avoid
1105 	 * unexpected Queue full condition
1106 	 */
1107 	qplqp->sq.q_full_delta -= 1;
1108 	qplqp->sq.sg_info.pgsize = PAGE_SIZE;
1109 	qplqp->sq.sg_info.pgshft = PAGE_SHIFT;
1110 }
1111 
1112 static void bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp *qp,
1113 				       struct ib_qp_init_attr *init_attr)
1114 {
1115 	struct bnxt_qplib_dev_attr *dev_attr;
1116 	struct bnxt_qplib_qp *qplqp;
1117 	struct bnxt_re_dev *rdev;
1118 	int entries;
1119 
1120 	rdev = qp->rdev;
1121 	qplqp = &qp->qplib_qp;
1122 	dev_attr = &rdev->dev_attr;
1123 
1124 	entries = roundup_pow_of_two(init_attr->cap.max_send_wr + 1);
1125 	qplqp->sq.max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1126 	qplqp->sq.q_full_delta = qplqp->sq.max_wqe -
1127 				 init_attr->cap.max_send_wr;
1128 	qplqp->sq.max_sge++; /* Need one extra sge to put UD header */
1129 	if (qplqp->sq.max_sge > dev_attr->max_qp_sges)
1130 		qplqp->sq.max_sge = dev_attr->max_qp_sges;
1131 }
1132 
1133 static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev,
1134 				struct ib_qp_init_attr *init_attr)
1135 {
1136 	struct bnxt_qplib_chip_ctx *chip_ctx;
1137 	int qptype;
1138 
1139 	chip_ctx = rdev->chip_ctx;
1140 
1141 	qptype = __from_ib_qp_type(init_attr->qp_type);
1142 	if (qptype == IB_QPT_MAX) {
1143 		ibdev_err(&rdev->ibdev, "QP type 0x%x not supported", qptype);
1144 		qptype = -EOPNOTSUPP;
1145 		goto out;
1146 	}
1147 
1148 	if (bnxt_qplib_is_chip_gen_p5(chip_ctx) &&
1149 	    init_attr->qp_type == IB_QPT_GSI)
1150 		qptype = CMDQ_CREATE_QP_TYPE_GSI;
1151 out:
1152 	return qptype;
1153 }
1154 
1155 static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1156 				struct ib_qp_init_attr *init_attr,
1157 				struct ib_udata *udata)
1158 {
1159 	struct bnxt_qplib_dev_attr *dev_attr;
1160 	struct bnxt_qplib_qp *qplqp;
1161 	struct bnxt_re_dev *rdev;
1162 	struct bnxt_re_cq *cq;
1163 	int rc = 0, qptype;
1164 
1165 	rdev = qp->rdev;
1166 	qplqp = &qp->qplib_qp;
1167 	dev_attr = &rdev->dev_attr;
1168 
1169 	/* Setup misc params */
1170 	ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr);
1171 	qplqp->pd = &pd->qplib_pd;
1172 	qplqp->qp_handle = (u64)qplqp;
1173 	qplqp->max_inline_data = init_attr->cap.max_inline_data;
1174 	qplqp->sig_type = ((init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ?
1175 			    true : false);
1176 	qptype = bnxt_re_init_qp_type(rdev, init_attr);
1177 	if (qptype < 0) {
1178 		rc = qptype;
1179 		goto out;
1180 	}
1181 	qplqp->type = (u8)qptype;
1182 
1183 	if (init_attr->qp_type == IB_QPT_RC) {
1184 		qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom;
1185 		qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
1186 	}
1187 	qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1188 	qplqp->dpi = &rdev->dpi_privileged; /* Doorbell page */
1189 	if (init_attr->create_flags)
1190 		ibdev_dbg(&rdev->ibdev,
1191 			  "QP create flags 0x%x not supported",
1192 			  init_attr->create_flags);
1193 
1194 	/* Setup CQs */
1195 	if (init_attr->send_cq) {
1196 		cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq);
1197 		if (!cq) {
1198 			ibdev_err(&rdev->ibdev, "Send CQ not found");
1199 			rc = -EINVAL;
1200 			goto out;
1201 		}
1202 		qplqp->scq = &cq->qplib_cq;
1203 		qp->scq = cq;
1204 	}
1205 
1206 	if (init_attr->recv_cq) {
1207 		cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq);
1208 		if (!cq) {
1209 			ibdev_err(&rdev->ibdev, "Receive CQ not found");
1210 			rc = -EINVAL;
1211 			goto out;
1212 		}
1213 		qplqp->rcq = &cq->qplib_cq;
1214 		qp->rcq = cq;
1215 	}
1216 
1217 	/* Setup RQ/SRQ */
1218 	rc = bnxt_re_init_rq_attr(qp, init_attr);
1219 	if (rc)
1220 		goto out;
1221 	if (init_attr->qp_type == IB_QPT_GSI)
1222 		bnxt_re_adjust_gsi_rq_attr(qp);
1223 
1224 	/* Setup SQ */
1225 	bnxt_re_init_sq_attr(qp, init_attr, udata);
1226 	if (init_attr->qp_type == IB_QPT_GSI)
1227 		bnxt_re_adjust_gsi_sq_attr(qp, init_attr);
1228 
1229 	if (udata) /* This will update DPI and qp_handle */
1230 		rc = bnxt_re_init_user_qp(rdev, pd, qp, udata);
1231 out:
1232 	return rc;
1233 }
1234 
1235 static int bnxt_re_create_shadow_gsi(struct bnxt_re_qp *qp,
1236 				     struct bnxt_re_pd *pd)
1237 {
1238 	struct bnxt_re_sqp_entries *sqp_tbl = NULL;
1239 	struct bnxt_re_dev *rdev;
1240 	struct bnxt_re_qp *sqp;
1241 	struct bnxt_re_ah *sah;
1242 	int rc = 0;
1243 
1244 	rdev = qp->rdev;
1245 	/* Create a shadow QP to handle the QP1 traffic */
1246 	sqp_tbl = kzalloc(sizeof(*sqp_tbl) * BNXT_RE_MAX_GSI_SQP_ENTRIES,
1247 			  GFP_KERNEL);
1248 	if (!sqp_tbl)
1249 		return -ENOMEM;
1250 	rdev->gsi_ctx.sqp_tbl = sqp_tbl;
1251 
1252 	sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res, &qp->qplib_qp);
1253 	if (!sqp) {
1254 		rc = -ENODEV;
1255 		ibdev_err(&rdev->ibdev, "Failed to create Shadow QP for QP1");
1256 		goto out;
1257 	}
1258 	rdev->gsi_ctx.gsi_sqp = sqp;
1259 
1260 	sqp->rcq = qp->rcq;
1261 	sqp->scq = qp->scq;
1262 	sah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res,
1263 					  &qp->qplib_qp);
1264 	if (!sah) {
1265 		bnxt_qplib_destroy_qp(&rdev->qplib_res,
1266 				      &sqp->qplib_qp);
1267 		rc = -ENODEV;
1268 		ibdev_err(&rdev->ibdev,
1269 			  "Failed to create AH entry for ShadowQP");
1270 		goto out;
1271 	}
1272 	rdev->gsi_ctx.gsi_sah = sah;
1273 
1274 	return 0;
1275 out:
1276 	kfree(sqp_tbl);
1277 	return rc;
1278 }
1279 
1280 static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1281 				 struct ib_qp_init_attr *init_attr)
1282 {
1283 	struct bnxt_re_dev *rdev;
1284 	struct bnxt_qplib_qp *qplqp;
1285 	int rc = 0;
1286 
1287 	rdev = qp->rdev;
1288 	qplqp = &qp->qplib_qp;
1289 
1290 	qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
1291 	qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
1292 
1293 	rc = bnxt_qplib_create_qp1(&rdev->qplib_res, qplqp);
1294 	if (rc) {
1295 		ibdev_err(&rdev->ibdev, "create HW QP1 failed!");
1296 		goto out;
1297 	}
1298 
1299 	rc = bnxt_re_create_shadow_gsi(qp, pd);
1300 out:
1301 	return rc;
1302 }
1303 
1304 static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev,
1305 				   struct ib_qp_init_attr *init_attr,
1306 				   struct bnxt_qplib_dev_attr *dev_attr)
1307 {
1308 	bool rc = true;
1309 
1310 	if (init_attr->cap.max_send_wr > dev_attr->max_qp_wqes ||
1311 	    init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes ||
1312 	    init_attr->cap.max_send_sge > dev_attr->max_qp_sges ||
1313 	    init_attr->cap.max_recv_sge > dev_attr->max_qp_sges ||
1314 	    init_attr->cap.max_inline_data > dev_attr->max_inline_data) {
1315 		ibdev_err(&rdev->ibdev,
1316 			  "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",
1317 			  init_attr->cap.max_send_wr, dev_attr->max_qp_wqes,
1318 			  init_attr->cap.max_recv_wr, dev_attr->max_qp_wqes,
1319 			  init_attr->cap.max_send_sge, dev_attr->max_qp_sges,
1320 			  init_attr->cap.max_recv_sge, dev_attr->max_qp_sges,
1321 			  init_attr->cap.max_inline_data,
1322 			  dev_attr->max_inline_data);
1323 		rc = false;
1324 	}
1325 	return rc;
1326 }
1327 
1328 struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
1329 				struct ib_qp_init_attr *qp_init_attr,
1330 				struct ib_udata *udata)
1331 {
1332 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1333 	struct bnxt_re_dev *rdev = pd->rdev;
1334 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1335 	struct bnxt_re_qp *qp;
1336 	int rc;
1337 
1338 	rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr);
1339 	if (!rc) {
1340 		rc = -EINVAL;
1341 		goto exit;
1342 	}
1343 
1344 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1345 	if (!qp) {
1346 		rc = -ENOMEM;
1347 		goto exit;
1348 	}
1349 	qp->rdev = rdev;
1350 	rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, udata);
1351 	if (rc)
1352 		goto fail;
1353 
1354 	if (qp_init_attr->qp_type == IB_QPT_GSI &&
1355 	    !(bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx))) {
1356 		rc = bnxt_re_create_gsi_qp(qp, pd, qp_init_attr);
1357 		if (rc == -ENODEV)
1358 			goto qp_destroy;
1359 		if (rc)
1360 			goto fail;
1361 	} else {
1362 		rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
1363 		if (rc) {
1364 			ibdev_err(&rdev->ibdev, "Failed to create HW QP");
1365 			goto free_umem;
1366 		}
1367 		if (udata) {
1368 			struct bnxt_re_qp_resp resp;
1369 
1370 			resp.qpid = qp->qplib_qp.id;
1371 			resp.rsvd = 0;
1372 			rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1373 			if (rc) {
1374 				ibdev_err(&rdev->ibdev, "Failed to copy QP udata");
1375 				goto qp_destroy;
1376 			}
1377 		}
1378 	}
1379 
1380 	qp->ib_qp.qp_num = qp->qplib_qp.id;
1381 	if (qp_init_attr->qp_type == IB_QPT_GSI)
1382 		rdev->gsi_ctx.gsi_qp = qp;
1383 	spin_lock_init(&qp->sq_lock);
1384 	spin_lock_init(&qp->rq_lock);
1385 	INIT_LIST_HEAD(&qp->list);
1386 	mutex_lock(&rdev->qp_lock);
1387 	list_add_tail(&qp->list, &rdev->qp_list);
1388 	mutex_unlock(&rdev->qp_lock);
1389 	atomic_inc(&rdev->qp_count);
1390 
1391 	return &qp->ib_qp;
1392 qp_destroy:
1393 	bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
1394 free_umem:
1395 	ib_umem_release(qp->rumem);
1396 	ib_umem_release(qp->sumem);
1397 fail:
1398 	kfree(qp);
1399 exit:
1400 	return ERR_PTR(rc);
1401 }
1402 
1403 static u8 __from_ib_qp_state(enum ib_qp_state state)
1404 {
1405 	switch (state) {
1406 	case IB_QPS_RESET:
1407 		return CMDQ_MODIFY_QP_NEW_STATE_RESET;
1408 	case IB_QPS_INIT:
1409 		return CMDQ_MODIFY_QP_NEW_STATE_INIT;
1410 	case IB_QPS_RTR:
1411 		return CMDQ_MODIFY_QP_NEW_STATE_RTR;
1412 	case IB_QPS_RTS:
1413 		return CMDQ_MODIFY_QP_NEW_STATE_RTS;
1414 	case IB_QPS_SQD:
1415 		return CMDQ_MODIFY_QP_NEW_STATE_SQD;
1416 	case IB_QPS_SQE:
1417 		return CMDQ_MODIFY_QP_NEW_STATE_SQE;
1418 	case IB_QPS_ERR:
1419 	default:
1420 		return CMDQ_MODIFY_QP_NEW_STATE_ERR;
1421 	}
1422 }
1423 
1424 static enum ib_qp_state __to_ib_qp_state(u8 state)
1425 {
1426 	switch (state) {
1427 	case CMDQ_MODIFY_QP_NEW_STATE_RESET:
1428 		return IB_QPS_RESET;
1429 	case CMDQ_MODIFY_QP_NEW_STATE_INIT:
1430 		return IB_QPS_INIT;
1431 	case CMDQ_MODIFY_QP_NEW_STATE_RTR:
1432 		return IB_QPS_RTR;
1433 	case CMDQ_MODIFY_QP_NEW_STATE_RTS:
1434 		return IB_QPS_RTS;
1435 	case CMDQ_MODIFY_QP_NEW_STATE_SQD:
1436 		return IB_QPS_SQD;
1437 	case CMDQ_MODIFY_QP_NEW_STATE_SQE:
1438 		return IB_QPS_SQE;
1439 	case CMDQ_MODIFY_QP_NEW_STATE_ERR:
1440 	default:
1441 		return IB_QPS_ERR;
1442 	}
1443 }
1444 
1445 static u32 __from_ib_mtu(enum ib_mtu mtu)
1446 {
1447 	switch (mtu) {
1448 	case IB_MTU_256:
1449 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_256;
1450 	case IB_MTU_512:
1451 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_512;
1452 	case IB_MTU_1024:
1453 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024;
1454 	case IB_MTU_2048:
1455 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1456 	case IB_MTU_4096:
1457 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096;
1458 	default:
1459 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1460 	}
1461 }
1462 
1463 static enum ib_mtu __to_ib_mtu(u32 mtu)
1464 {
1465 	switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) {
1466 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_256:
1467 		return IB_MTU_256;
1468 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_512:
1469 		return IB_MTU_512;
1470 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024:
1471 		return IB_MTU_1024;
1472 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048:
1473 		return IB_MTU_2048;
1474 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096:
1475 		return IB_MTU_4096;
1476 	default:
1477 		return IB_MTU_2048;
1478 	}
1479 }
1480 
1481 /* Shared Receive Queues */
1482 void bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata)
1483 {
1484 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1485 					       ib_srq);
1486 	struct bnxt_re_dev *rdev = srq->rdev;
1487 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1488 	struct bnxt_qplib_nq *nq = NULL;
1489 
1490 	if (qplib_srq->cq)
1491 		nq = qplib_srq->cq->nq;
1492 	bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
1493 	ib_umem_release(srq->umem);
1494 	atomic_dec(&rdev->srq_count);
1495 	if (nq)
1496 		nq->budget--;
1497 }
1498 
1499 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
1500 				 struct bnxt_re_pd *pd,
1501 				 struct bnxt_re_srq *srq,
1502 				 struct ib_udata *udata)
1503 {
1504 	struct bnxt_re_srq_req ureq;
1505 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1506 	struct ib_umem *umem;
1507 	int bytes = 0;
1508 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
1509 		udata, struct bnxt_re_ucontext, ib_uctx);
1510 
1511 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1512 		return -EFAULT;
1513 
1514 	bytes = (qplib_srq->max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
1515 	bytes = PAGE_ALIGN(bytes);
1516 	umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes,
1517 			   IB_ACCESS_LOCAL_WRITE);
1518 	if (IS_ERR(umem))
1519 		return PTR_ERR(umem);
1520 
1521 	srq->umem = umem;
1522 	qplib_srq->sg_info.sghead = umem->sg_head.sgl;
1523 	qplib_srq->sg_info.npages = ib_umem_num_pages(umem);
1524 	qplib_srq->sg_info.nmap = umem->nmap;
1525 	qplib_srq->sg_info.pgsize = PAGE_SIZE;
1526 	qplib_srq->sg_info.pgshft = PAGE_SHIFT;
1527 	qplib_srq->srq_handle = ureq.srq_handle;
1528 	qplib_srq->dpi = &cntx->dpi;
1529 
1530 	return 0;
1531 }
1532 
1533 int bnxt_re_create_srq(struct ib_srq *ib_srq,
1534 		       struct ib_srq_init_attr *srq_init_attr,
1535 		       struct ib_udata *udata)
1536 {
1537 	struct ib_pd *ib_pd = ib_srq->pd;
1538 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1539 	struct bnxt_re_dev *rdev = pd->rdev;
1540 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1541 	struct bnxt_re_srq *srq =
1542 		container_of(ib_srq, struct bnxt_re_srq, ib_srq);
1543 	struct bnxt_qplib_nq *nq = NULL;
1544 	int rc, entries;
1545 
1546 	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
1547 		ibdev_err(&rdev->ibdev, "Create CQ failed - max exceeded");
1548 		rc = -EINVAL;
1549 		goto exit;
1550 	}
1551 
1552 	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
1553 		rc = -EOPNOTSUPP;
1554 		goto exit;
1555 	}
1556 
1557 	srq->rdev = rdev;
1558 	srq->qplib_srq.pd = &pd->qplib_pd;
1559 	srq->qplib_srq.dpi = &rdev->dpi_privileged;
1560 	/* Allocate 1 more than what's provided so posting max doesn't
1561 	 * mean empty
1562 	 */
1563 	entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
1564 	if (entries > dev_attr->max_srq_wqes + 1)
1565 		entries = dev_attr->max_srq_wqes + 1;
1566 
1567 	srq->qplib_srq.max_wqe = entries;
1568 	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
1569 	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
1570 	srq->srq_limit = srq_init_attr->attr.srq_limit;
1571 	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
1572 	nq = &rdev->nq[0];
1573 
1574 	if (udata) {
1575 		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
1576 		if (rc)
1577 			goto fail;
1578 	}
1579 
1580 	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
1581 	if (rc) {
1582 		ibdev_err(&rdev->ibdev, "Create HW SRQ failed!");
1583 		goto fail;
1584 	}
1585 
1586 	if (udata) {
1587 		struct bnxt_re_srq_resp resp;
1588 
1589 		resp.srqid = srq->qplib_srq.id;
1590 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1591 		if (rc) {
1592 			ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!");
1593 			bnxt_qplib_destroy_srq(&rdev->qplib_res,
1594 					       &srq->qplib_srq);
1595 			goto fail;
1596 		}
1597 	}
1598 	if (nq)
1599 		nq->budget++;
1600 	atomic_inc(&rdev->srq_count);
1601 
1602 	return 0;
1603 
1604 fail:
1605 	ib_umem_release(srq->umem);
1606 exit:
1607 	return rc;
1608 }
1609 
1610 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
1611 		       enum ib_srq_attr_mask srq_attr_mask,
1612 		       struct ib_udata *udata)
1613 {
1614 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1615 					       ib_srq);
1616 	struct bnxt_re_dev *rdev = srq->rdev;
1617 	int rc;
1618 
1619 	switch (srq_attr_mask) {
1620 	case IB_SRQ_MAX_WR:
1621 		/* SRQ resize is not supported */
1622 		break;
1623 	case IB_SRQ_LIMIT:
1624 		/* Change the SRQ threshold */
1625 		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
1626 			return -EINVAL;
1627 
1628 		srq->qplib_srq.threshold = srq_attr->srq_limit;
1629 		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
1630 		if (rc) {
1631 			ibdev_err(&rdev->ibdev, "Modify HW SRQ failed!");
1632 			return rc;
1633 		}
1634 		/* On success, update the shadow */
1635 		srq->srq_limit = srq_attr->srq_limit;
1636 		/* No need to Build and send response back to udata */
1637 		break;
1638 	default:
1639 		ibdev_err(&rdev->ibdev,
1640 			  "Unsupported srq_attr_mask 0x%x", srq_attr_mask);
1641 		return -EINVAL;
1642 	}
1643 	return 0;
1644 }
1645 
1646 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
1647 {
1648 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1649 					       ib_srq);
1650 	struct bnxt_re_srq tsrq;
1651 	struct bnxt_re_dev *rdev = srq->rdev;
1652 	int rc;
1653 
1654 	/* Get live SRQ attr */
1655 	tsrq.qplib_srq.id = srq->qplib_srq.id;
1656 	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
1657 	if (rc) {
1658 		ibdev_err(&rdev->ibdev, "Query HW SRQ failed!");
1659 		return rc;
1660 	}
1661 	srq_attr->max_wr = srq->qplib_srq.max_wqe;
1662 	srq_attr->max_sge = srq->qplib_srq.max_sge;
1663 	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
1664 
1665 	return 0;
1666 }
1667 
1668 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr,
1669 			  const struct ib_recv_wr **bad_wr)
1670 {
1671 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1672 					       ib_srq);
1673 	struct bnxt_qplib_swqe wqe;
1674 	unsigned long flags;
1675 	int rc = 0;
1676 
1677 	spin_lock_irqsave(&srq->lock, flags);
1678 	while (wr) {
1679 		/* Transcribe each ib_recv_wr to qplib_swqe */
1680 		wqe.num_sge = wr->num_sge;
1681 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
1682 		wqe.wr_id = wr->wr_id;
1683 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
1684 
1685 		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
1686 		if (rc) {
1687 			*bad_wr = wr;
1688 			break;
1689 		}
1690 		wr = wr->next;
1691 	}
1692 	spin_unlock_irqrestore(&srq->lock, flags);
1693 
1694 	return rc;
1695 }
1696 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
1697 				    struct bnxt_re_qp *qp1_qp,
1698 				    int qp_attr_mask)
1699 {
1700 	struct bnxt_re_qp *qp = rdev->gsi_ctx.gsi_sqp;
1701 	int rc = 0;
1702 
1703 	if (qp_attr_mask & IB_QP_STATE) {
1704 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1705 		qp->qplib_qp.state = qp1_qp->qplib_qp.state;
1706 	}
1707 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1708 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1709 		qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index;
1710 	}
1711 
1712 	if (qp_attr_mask & IB_QP_QKEY) {
1713 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1714 		/* Using a Random  QKEY */
1715 		qp->qplib_qp.qkey = 0x81818181;
1716 	}
1717 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1718 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1719 		qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn;
1720 	}
1721 
1722 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1723 	if (rc)
1724 		ibdev_err(&rdev->ibdev, "Failed to modify Shadow QP for QP1");
1725 	return rc;
1726 }
1727 
1728 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1729 		      int qp_attr_mask, struct ib_udata *udata)
1730 {
1731 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1732 	struct bnxt_re_dev *rdev = qp->rdev;
1733 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1734 	enum ib_qp_state curr_qp_state, new_qp_state;
1735 	int rc, entries;
1736 	unsigned int flags;
1737 	u8 nw_type;
1738 
1739 	qp->qplib_qp.modify_flags = 0;
1740 	if (qp_attr_mask & IB_QP_STATE) {
1741 		curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state);
1742 		new_qp_state = qp_attr->qp_state;
1743 		if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state,
1744 					ib_qp->qp_type, qp_attr_mask)) {
1745 			ibdev_err(&rdev->ibdev,
1746 				  "Invalid attribute mask: %#x specified ",
1747 				  qp_attr_mask);
1748 			ibdev_err(&rdev->ibdev,
1749 				  "for qpn: %#x type: %#x",
1750 				  ib_qp->qp_num, ib_qp->qp_type);
1751 			ibdev_err(&rdev->ibdev,
1752 				  "curr_qp_state=0x%x, new_qp_state=0x%x\n",
1753 				  curr_qp_state, new_qp_state);
1754 			return -EINVAL;
1755 		}
1756 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1757 		qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state);
1758 
1759 		if (!qp->sumem &&
1760 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1761 			ibdev_dbg(&rdev->ibdev,
1762 				  "Move QP = %p to flush list\n", qp);
1763 			flags = bnxt_re_lock_cqs(qp);
1764 			bnxt_qplib_add_flush_qp(&qp->qplib_qp);
1765 			bnxt_re_unlock_cqs(qp, flags);
1766 		}
1767 		if (!qp->sumem &&
1768 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) {
1769 			ibdev_dbg(&rdev->ibdev,
1770 				  "Move QP = %p out of flush list\n", qp);
1771 			flags = bnxt_re_lock_cqs(qp);
1772 			bnxt_qplib_clean_qp(&qp->qplib_qp);
1773 			bnxt_re_unlock_cqs(qp, flags);
1774 		}
1775 	}
1776 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1777 		qp->qplib_qp.modify_flags |=
1778 				CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY;
1779 		qp->qplib_qp.en_sqd_async_notify = true;
1780 	}
1781 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
1782 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS;
1783 		qp->qplib_qp.access =
1784 			__from_ib_access_flags(qp_attr->qp_access_flags);
1785 		/* LOCAL_WRITE access must be set to allow RC receive */
1786 		qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
1787 		/* Temp: Set all params on QP as of now */
1788 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE;
1789 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ;
1790 	}
1791 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1792 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1793 		qp->qplib_qp.pkey_index = qp_attr->pkey_index;
1794 	}
1795 	if (qp_attr_mask & IB_QP_QKEY) {
1796 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1797 		qp->qplib_qp.qkey = qp_attr->qkey;
1798 	}
1799 	if (qp_attr_mask & IB_QP_AV) {
1800 		const struct ib_global_route *grh =
1801 			rdma_ah_read_grh(&qp_attr->ah_attr);
1802 		const struct ib_gid_attr *sgid_attr;
1803 		struct bnxt_re_gid_ctx *ctx;
1804 
1805 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID |
1806 				     CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL |
1807 				     CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX |
1808 				     CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT |
1809 				     CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS |
1810 				     CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC |
1811 				     CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID;
1812 		memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw,
1813 		       sizeof(qp->qplib_qp.ah.dgid.data));
1814 		qp->qplib_qp.ah.flow_label = grh->flow_label;
1815 		sgid_attr = grh->sgid_attr;
1816 		/* Get the HW context of the GID. The reference
1817 		 * of GID table entry is already taken by the caller.
1818 		 */
1819 		ctx = rdma_read_gid_hw_context(sgid_attr);
1820 		qp->qplib_qp.ah.sgid_index = ctx->idx;
1821 		qp->qplib_qp.ah.host_sgid_index = grh->sgid_index;
1822 		qp->qplib_qp.ah.hop_limit = grh->hop_limit;
1823 		qp->qplib_qp.ah.traffic_class = grh->traffic_class;
1824 		qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr);
1825 		ether_addr_copy(qp->qplib_qp.ah.dmac,
1826 				qp_attr->ah_attr.roce.dmac);
1827 
1828 		rc = rdma_read_gid_l2_fields(sgid_attr, NULL,
1829 					     &qp->qplib_qp.smac[0]);
1830 		if (rc)
1831 			return rc;
1832 
1833 		nw_type = rdma_gid_attr_network_type(sgid_attr);
1834 		switch (nw_type) {
1835 		case RDMA_NETWORK_IPV4:
1836 			qp->qplib_qp.nw_type =
1837 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4;
1838 			break;
1839 		case RDMA_NETWORK_IPV6:
1840 			qp->qplib_qp.nw_type =
1841 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6;
1842 			break;
1843 		default:
1844 			qp->qplib_qp.nw_type =
1845 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1;
1846 			break;
1847 		}
1848 	}
1849 
1850 	if (qp_attr_mask & IB_QP_PATH_MTU) {
1851 		qp->qplib_qp.modify_flags |=
1852 				CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1853 		qp->qplib_qp.path_mtu = __from_ib_mtu(qp_attr->path_mtu);
1854 		qp->qplib_qp.mtu = ib_mtu_enum_to_int(qp_attr->path_mtu);
1855 	} else if (qp_attr->qp_state == IB_QPS_RTR) {
1856 		qp->qplib_qp.modify_flags |=
1857 			CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1858 		qp->qplib_qp.path_mtu =
1859 			__from_ib_mtu(iboe_get_mtu(rdev->netdev->mtu));
1860 		qp->qplib_qp.mtu =
1861 			ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1862 	}
1863 
1864 	if (qp_attr_mask & IB_QP_TIMEOUT) {
1865 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT;
1866 		qp->qplib_qp.timeout = qp_attr->timeout;
1867 	}
1868 	if (qp_attr_mask & IB_QP_RETRY_CNT) {
1869 		qp->qplib_qp.modify_flags |=
1870 				CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT;
1871 		qp->qplib_qp.retry_cnt = qp_attr->retry_cnt;
1872 	}
1873 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
1874 		qp->qplib_qp.modify_flags |=
1875 				CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY;
1876 		qp->qplib_qp.rnr_retry = qp_attr->rnr_retry;
1877 	}
1878 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) {
1879 		qp->qplib_qp.modify_flags |=
1880 				CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER;
1881 		qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer;
1882 	}
1883 	if (qp_attr_mask & IB_QP_RQ_PSN) {
1884 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN;
1885 		qp->qplib_qp.rq.psn = qp_attr->rq_psn;
1886 	}
1887 	if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1888 		qp->qplib_qp.modify_flags |=
1889 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC;
1890 		/* Cap the max_rd_atomic to device max */
1891 		qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic,
1892 						   dev_attr->max_qp_rd_atom);
1893 	}
1894 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1895 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1896 		qp->qplib_qp.sq.psn = qp_attr->sq_psn;
1897 	}
1898 	if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1899 		if (qp_attr->max_dest_rd_atomic >
1900 		    dev_attr->max_qp_init_rd_atom) {
1901 			ibdev_err(&rdev->ibdev,
1902 				  "max_dest_rd_atomic requested%d is > dev_max%d",
1903 				  qp_attr->max_dest_rd_atomic,
1904 				  dev_attr->max_qp_init_rd_atom);
1905 			return -EINVAL;
1906 		}
1907 
1908 		qp->qplib_qp.modify_flags |=
1909 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC;
1910 		qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic;
1911 	}
1912 	if (qp_attr_mask & IB_QP_CAP) {
1913 		qp->qplib_qp.modify_flags |=
1914 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE |
1915 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE |
1916 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE |
1917 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE |
1918 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA;
1919 		if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) ||
1920 		    (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) ||
1921 		    (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) ||
1922 		    (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) ||
1923 		    (qp_attr->cap.max_inline_data >=
1924 						dev_attr->max_inline_data)) {
1925 			ibdev_err(&rdev->ibdev,
1926 				  "Create QP failed - max exceeded");
1927 			return -EINVAL;
1928 		}
1929 		entries = roundup_pow_of_two(qp_attr->cap.max_send_wr);
1930 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
1931 						dev_attr->max_qp_wqes + 1);
1932 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
1933 						qp_attr->cap.max_send_wr;
1934 		/*
1935 		 * Reserving one slot for Phantom WQE. Some application can
1936 		 * post one extra entry in this case. Allowing this to avoid
1937 		 * unexpected Queue full condition
1938 		 */
1939 		qp->qplib_qp.sq.q_full_delta -= 1;
1940 		qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
1941 		if (qp->qplib_qp.rq.max_wqe) {
1942 			entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr);
1943 			qp->qplib_qp.rq.max_wqe =
1944 				min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1945 			qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
1946 						       qp_attr->cap.max_recv_wr;
1947 			qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
1948 		} else {
1949 			/* SRQ was used prior, just ignore the RQ caps */
1950 		}
1951 	}
1952 	if (qp_attr_mask & IB_QP_DEST_QPN) {
1953 		qp->qplib_qp.modify_flags |=
1954 				CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID;
1955 		qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num;
1956 	}
1957 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1958 	if (rc) {
1959 		ibdev_err(&rdev->ibdev, "Failed to modify HW QP");
1960 		return rc;
1961 	}
1962 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp)
1963 		rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask);
1964 	return rc;
1965 }
1966 
1967 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1968 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
1969 {
1970 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1971 	struct bnxt_re_dev *rdev = qp->rdev;
1972 	struct bnxt_qplib_qp *qplib_qp;
1973 	int rc;
1974 
1975 	qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL);
1976 	if (!qplib_qp)
1977 		return -ENOMEM;
1978 
1979 	qplib_qp->id = qp->qplib_qp.id;
1980 	qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index;
1981 
1982 	rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp);
1983 	if (rc) {
1984 		ibdev_err(&rdev->ibdev, "Failed to query HW QP");
1985 		goto out;
1986 	}
1987 	qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state);
1988 	qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0;
1989 	qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access);
1990 	qp_attr->pkey_index = qplib_qp->pkey_index;
1991 	qp_attr->qkey = qplib_qp->qkey;
1992 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1993 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label,
1994 			qplib_qp->ah.host_sgid_index,
1995 			qplib_qp->ah.hop_limit,
1996 			qplib_qp->ah.traffic_class);
1997 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data);
1998 	rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl);
1999 	ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac);
2000 	qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu);
2001 	qp_attr->timeout = qplib_qp->timeout;
2002 	qp_attr->retry_cnt = qplib_qp->retry_cnt;
2003 	qp_attr->rnr_retry = qplib_qp->rnr_retry;
2004 	qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer;
2005 	qp_attr->rq_psn = qplib_qp->rq.psn;
2006 	qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic;
2007 	qp_attr->sq_psn = qplib_qp->sq.psn;
2008 	qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic;
2009 	qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR :
2010 							 IB_SIGNAL_REQ_WR;
2011 	qp_attr->dest_qp_num = qplib_qp->dest_qpn;
2012 
2013 	qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe;
2014 	qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge;
2015 	qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe;
2016 	qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge;
2017 	qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data;
2018 	qp_init_attr->cap = qp_attr->cap;
2019 
2020 out:
2021 	kfree(qplib_qp);
2022 	return rc;
2023 }
2024 
2025 /* Routine for sending QP1 packets for RoCE V1 an V2
2026  */
2027 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp,
2028 				     const struct ib_send_wr *wr,
2029 				     struct bnxt_qplib_swqe *wqe,
2030 				     int payload_size)
2031 {
2032 	struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah,
2033 					     ib_ah);
2034 	struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah;
2035 	const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr;
2036 	struct bnxt_qplib_sge sge;
2037 	u8 nw_type;
2038 	u16 ether_type;
2039 	union ib_gid dgid;
2040 	bool is_eth = false;
2041 	bool is_vlan = false;
2042 	bool is_grh = false;
2043 	bool is_udp = false;
2044 	u8 ip_version = 0;
2045 	u16 vlan_id = 0xFFFF;
2046 	void *buf;
2047 	int i, rc = 0;
2048 
2049 	memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr));
2050 
2051 	rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL);
2052 	if (rc)
2053 		return rc;
2054 
2055 	/* Get network header type for this GID */
2056 	nw_type = rdma_gid_attr_network_type(sgid_attr);
2057 	switch (nw_type) {
2058 	case RDMA_NETWORK_IPV4:
2059 		nw_type = BNXT_RE_ROCEV2_IPV4_PACKET;
2060 		break;
2061 	case RDMA_NETWORK_IPV6:
2062 		nw_type = BNXT_RE_ROCEV2_IPV6_PACKET;
2063 		break;
2064 	default:
2065 		nw_type = BNXT_RE_ROCE_V1_PACKET;
2066 		break;
2067 	}
2068 	memcpy(&dgid.raw, &qplib_ah->dgid, 16);
2069 	is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2070 	if (is_udp) {
2071 		if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) {
2072 			ip_version = 4;
2073 			ether_type = ETH_P_IP;
2074 		} else {
2075 			ip_version = 6;
2076 			ether_type = ETH_P_IPV6;
2077 		}
2078 		is_grh = false;
2079 	} else {
2080 		ether_type = ETH_P_IBOE;
2081 		is_grh = true;
2082 	}
2083 
2084 	is_eth = true;
2085 	is_vlan = (vlan_id && (vlan_id < 0x1000)) ? true : false;
2086 
2087 	ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh,
2088 			  ip_version, is_udp, 0, &qp->qp1_hdr);
2089 
2090 	/* ETH */
2091 	ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac);
2092 	ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac);
2093 
2094 	/* For vlan, check the sgid for vlan existence */
2095 
2096 	if (!is_vlan) {
2097 		qp->qp1_hdr.eth.type = cpu_to_be16(ether_type);
2098 	} else {
2099 		qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type);
2100 		qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id);
2101 	}
2102 
2103 	if (is_grh || (ip_version == 6)) {
2104 		memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw,
2105 		       sizeof(sgid_attr->gid));
2106 		memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data,
2107 		       sizeof(sgid_attr->gid));
2108 		qp->qp1_hdr.grh.hop_limit     = qplib_ah->hop_limit;
2109 	}
2110 
2111 	if (ip_version == 4) {
2112 		qp->qp1_hdr.ip4.tos = 0;
2113 		qp->qp1_hdr.ip4.id = 0;
2114 		qp->qp1_hdr.ip4.frag_off = htons(IP_DF);
2115 		qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit;
2116 
2117 		memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4);
2118 		memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4);
2119 		qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr);
2120 	}
2121 
2122 	if (is_udp) {
2123 		qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT);
2124 		qp->qp1_hdr.udp.sport = htons(0x8CD1);
2125 		qp->qp1_hdr.udp.csum = 0;
2126 	}
2127 
2128 	/* BTH */
2129 	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
2130 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2131 		qp->qp1_hdr.immediate_present = 1;
2132 	} else {
2133 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2134 	}
2135 	if (wr->send_flags & IB_SEND_SOLICITED)
2136 		qp->qp1_hdr.bth.solicited_event = 1;
2137 	/* pad_count */
2138 	qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3;
2139 
2140 	/* P_key for QP1 is for all members */
2141 	qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF);
2142 	qp->qp1_hdr.bth.destination_qpn = IB_QP1;
2143 	qp->qp1_hdr.bth.ack_req = 0;
2144 	qp->send_psn++;
2145 	qp->send_psn &= BTH_PSN_MASK;
2146 	qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn);
2147 	/* DETH */
2148 	/* Use the priviledged Q_Key for QP1 */
2149 	qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY);
2150 	qp->qp1_hdr.deth.source_qpn = IB_QP1;
2151 
2152 	/* Pack the QP1 to the transmit buffer */
2153 	buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge);
2154 	if (buf) {
2155 		ib_ud_header_pack(&qp->qp1_hdr, buf);
2156 		for (i = wqe->num_sge; i; i--) {
2157 			wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr;
2158 			wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey;
2159 			wqe->sg_list[i].size = wqe->sg_list[i - 1].size;
2160 		}
2161 
2162 		/*
2163 		 * Max Header buf size for IPV6 RoCE V2 is 86,
2164 		 * which is same as the QP1 SQ header buffer.
2165 		 * Header buf size for IPV4 RoCE V2 can be 66.
2166 		 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
2167 		 * Subtract 20 bytes from QP1 SQ header buf size
2168 		 */
2169 		if (is_udp && ip_version == 4)
2170 			sge.size -= 20;
2171 		/*
2172 		 * Max Header buf size for RoCE V1 is 78.
2173 		 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
2174 		 * Subtract 8 bytes from QP1 SQ header buf size
2175 		 */
2176 		if (!is_udp)
2177 			sge.size -= 8;
2178 
2179 		/* Subtract 4 bytes for non vlan packets */
2180 		if (!is_vlan)
2181 			sge.size -= 4;
2182 
2183 		wqe->sg_list[0].addr = sge.addr;
2184 		wqe->sg_list[0].lkey = sge.lkey;
2185 		wqe->sg_list[0].size = sge.size;
2186 		wqe->num_sge++;
2187 
2188 	} else {
2189 		ibdev_err(&qp->rdev->ibdev, "QP1 buffer is empty!");
2190 		rc = -ENOMEM;
2191 	}
2192 	return rc;
2193 }
2194 
2195 /* For the MAD layer, it only provides the recv SGE the size of
2196  * ib_grh + MAD datagram.  No Ethernet headers, Ethertype, BTH, DETH,
2197  * nor RoCE iCRC.  The Cu+ solution must provide buffer for the entire
2198  * receive packet (334 bytes) with no VLAN and then copy the GRH
2199  * and the MAD datagram out to the provided SGE.
2200  */
2201 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp,
2202 					    const struct ib_recv_wr *wr,
2203 					    struct bnxt_qplib_swqe *wqe,
2204 					    int payload_size)
2205 {
2206 	struct bnxt_re_sqp_entries *sqp_entry;
2207 	struct bnxt_qplib_sge ref, sge;
2208 	struct bnxt_re_dev *rdev;
2209 	u32 rq_prod_index;
2210 
2211 	rdev = qp->rdev;
2212 
2213 	rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp);
2214 
2215 	if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge))
2216 		return -ENOMEM;
2217 
2218 	/* Create 1 SGE to receive the entire
2219 	 * ethernet packet
2220 	 */
2221 	/* Save the reference from ULP */
2222 	ref.addr = wqe->sg_list[0].addr;
2223 	ref.lkey = wqe->sg_list[0].lkey;
2224 	ref.size = wqe->sg_list[0].size;
2225 
2226 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[rq_prod_index];
2227 
2228 	/* SGE 1 */
2229 	wqe->sg_list[0].addr = sge.addr;
2230 	wqe->sg_list[0].lkey = sge.lkey;
2231 	wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
2232 	sge.size -= wqe->sg_list[0].size;
2233 
2234 	sqp_entry->sge.addr = ref.addr;
2235 	sqp_entry->sge.lkey = ref.lkey;
2236 	sqp_entry->sge.size = ref.size;
2237 	/* Store the wrid for reporting completion */
2238 	sqp_entry->wrid = wqe->wr_id;
2239 	/* change the wqe->wrid to table index */
2240 	wqe->wr_id = rq_prod_index;
2241 	return 0;
2242 }
2243 
2244 static int is_ud_qp(struct bnxt_re_qp *qp)
2245 {
2246 	return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD ||
2247 		qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI);
2248 }
2249 
2250 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp,
2251 				  const struct ib_send_wr *wr,
2252 				  struct bnxt_qplib_swqe *wqe)
2253 {
2254 	struct bnxt_re_ah *ah = NULL;
2255 
2256 	if (is_ud_qp(qp)) {
2257 		ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah);
2258 		wqe->send.q_key = ud_wr(wr)->remote_qkey;
2259 		wqe->send.dst_qp = ud_wr(wr)->remote_qpn;
2260 		wqe->send.avid = ah->qplib_ah.id;
2261 	}
2262 	switch (wr->opcode) {
2263 	case IB_WR_SEND:
2264 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND;
2265 		break;
2266 	case IB_WR_SEND_WITH_IMM:
2267 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM;
2268 		wqe->send.imm_data = wr->ex.imm_data;
2269 		break;
2270 	case IB_WR_SEND_WITH_INV:
2271 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV;
2272 		wqe->send.inv_key = wr->ex.invalidate_rkey;
2273 		break;
2274 	default:
2275 		return -EINVAL;
2276 	}
2277 	if (wr->send_flags & IB_SEND_SIGNALED)
2278 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2279 	if (wr->send_flags & IB_SEND_FENCE)
2280 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2281 	if (wr->send_flags & IB_SEND_SOLICITED)
2282 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2283 	if (wr->send_flags & IB_SEND_INLINE)
2284 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2285 
2286 	return 0;
2287 }
2288 
2289 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr,
2290 				  struct bnxt_qplib_swqe *wqe)
2291 {
2292 	switch (wr->opcode) {
2293 	case IB_WR_RDMA_WRITE:
2294 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE;
2295 		break;
2296 	case IB_WR_RDMA_WRITE_WITH_IMM:
2297 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM;
2298 		wqe->rdma.imm_data = wr->ex.imm_data;
2299 		break;
2300 	case IB_WR_RDMA_READ:
2301 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ;
2302 		wqe->rdma.inv_key = wr->ex.invalidate_rkey;
2303 		break;
2304 	default:
2305 		return -EINVAL;
2306 	}
2307 	wqe->rdma.remote_va = rdma_wr(wr)->remote_addr;
2308 	wqe->rdma.r_key = rdma_wr(wr)->rkey;
2309 	if (wr->send_flags & IB_SEND_SIGNALED)
2310 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2311 	if (wr->send_flags & IB_SEND_FENCE)
2312 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2313 	if (wr->send_flags & IB_SEND_SOLICITED)
2314 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2315 	if (wr->send_flags & IB_SEND_INLINE)
2316 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2317 
2318 	return 0;
2319 }
2320 
2321 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr,
2322 				    struct bnxt_qplib_swqe *wqe)
2323 {
2324 	switch (wr->opcode) {
2325 	case IB_WR_ATOMIC_CMP_AND_SWP:
2326 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP;
2327 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2328 		wqe->atomic.swap_data = atomic_wr(wr)->swap;
2329 		break;
2330 	case IB_WR_ATOMIC_FETCH_AND_ADD:
2331 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD;
2332 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2333 		break;
2334 	default:
2335 		return -EINVAL;
2336 	}
2337 	wqe->atomic.remote_va = atomic_wr(wr)->remote_addr;
2338 	wqe->atomic.r_key = atomic_wr(wr)->rkey;
2339 	if (wr->send_flags & IB_SEND_SIGNALED)
2340 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2341 	if (wr->send_flags & IB_SEND_FENCE)
2342 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2343 	if (wr->send_flags & IB_SEND_SOLICITED)
2344 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2345 	return 0;
2346 }
2347 
2348 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr,
2349 				 struct bnxt_qplib_swqe *wqe)
2350 {
2351 	wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV;
2352 	wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey;
2353 
2354 	/* Need unconditional fence for local invalidate
2355 	 * opcode to work as expected.
2356 	 */
2357 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2358 
2359 	if (wr->send_flags & IB_SEND_SIGNALED)
2360 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2361 	if (wr->send_flags & IB_SEND_SOLICITED)
2362 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2363 
2364 	return 0;
2365 }
2366 
2367 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr,
2368 				 struct bnxt_qplib_swqe *wqe)
2369 {
2370 	struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr);
2371 	struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl;
2372 	int access = wr->access;
2373 
2374 	wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0];
2375 	wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0];
2376 	wqe->frmr.page_list = mr->pages;
2377 	wqe->frmr.page_list_len = mr->npages;
2378 	wqe->frmr.levels = qplib_frpl->hwq.level;
2379 	wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR;
2380 
2381 	/* Need unconditional fence for reg_mr
2382 	 * opcode to function as expected.
2383 	 */
2384 
2385 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2386 
2387 	if (wr->wr.send_flags & IB_SEND_SIGNALED)
2388 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2389 
2390 	if (access & IB_ACCESS_LOCAL_WRITE)
2391 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE;
2392 	if (access & IB_ACCESS_REMOTE_READ)
2393 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ;
2394 	if (access & IB_ACCESS_REMOTE_WRITE)
2395 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE;
2396 	if (access & IB_ACCESS_REMOTE_ATOMIC)
2397 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC;
2398 	if (access & IB_ACCESS_MW_BIND)
2399 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND;
2400 
2401 	wqe->frmr.l_key = wr->key;
2402 	wqe->frmr.length = wr->mr->length;
2403 	wqe->frmr.pbl_pg_sz_log = (wr->mr->page_size >> PAGE_SHIFT_4K) - 1;
2404 	wqe->frmr.va = wr->mr->iova;
2405 	return 0;
2406 }
2407 
2408 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev,
2409 				    const struct ib_send_wr *wr,
2410 				    struct bnxt_qplib_swqe *wqe)
2411 {
2412 	/*  Copy the inline data to the data  field */
2413 	u8 *in_data;
2414 	u32 i, sge_len;
2415 	void *sge_addr;
2416 
2417 	in_data = wqe->inline_data;
2418 	for (i = 0; i < wr->num_sge; i++) {
2419 		sge_addr = (void *)(unsigned long)
2420 				wr->sg_list[i].addr;
2421 		sge_len = wr->sg_list[i].length;
2422 
2423 		if ((sge_len + wqe->inline_len) >
2424 		    BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) {
2425 			ibdev_err(&rdev->ibdev,
2426 				  "Inline data size requested > supported value");
2427 			return -EINVAL;
2428 		}
2429 		sge_len = wr->sg_list[i].length;
2430 
2431 		memcpy(in_data, sge_addr, sge_len);
2432 		in_data += wr->sg_list[i].length;
2433 		wqe->inline_len += wr->sg_list[i].length;
2434 	}
2435 	return wqe->inline_len;
2436 }
2437 
2438 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
2439 				   const struct ib_send_wr *wr,
2440 				   struct bnxt_qplib_swqe *wqe)
2441 {
2442 	int payload_sz = 0;
2443 
2444 	if (wr->send_flags & IB_SEND_INLINE)
2445 		payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe);
2446 	else
2447 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list,
2448 					       wqe->num_sge);
2449 
2450 	return payload_sz;
2451 }
2452 
2453 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
2454 {
2455 	if ((qp->ib_qp.qp_type == IB_QPT_UD ||
2456 	     qp->ib_qp.qp_type == IB_QPT_GSI ||
2457 	     qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
2458 	     qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
2459 		int qp_attr_mask;
2460 		struct ib_qp_attr qp_attr;
2461 
2462 		qp_attr_mask = IB_QP_STATE;
2463 		qp_attr.qp_state = IB_QPS_RTS;
2464 		bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
2465 		qp->qplib_qp.wqe_cnt = 0;
2466 	}
2467 }
2468 
2469 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
2470 				       struct bnxt_re_qp *qp,
2471 				       const struct ib_send_wr *wr)
2472 {
2473 	int rc = 0, payload_sz = 0;
2474 	unsigned long flags;
2475 
2476 	spin_lock_irqsave(&qp->sq_lock, flags);
2477 	while (wr) {
2478 		struct bnxt_qplib_swqe wqe = {};
2479 
2480 		/* Common */
2481 		wqe.num_sge = wr->num_sge;
2482 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2483 			ibdev_err(&rdev->ibdev,
2484 				  "Limit exceeded for Send SGEs");
2485 			rc = -EINVAL;
2486 			goto bad;
2487 		}
2488 
2489 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2490 		if (payload_sz < 0) {
2491 			rc = -EINVAL;
2492 			goto bad;
2493 		}
2494 		wqe.wr_id = wr->wr_id;
2495 
2496 		wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND;
2497 
2498 		rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2499 		if (!rc)
2500 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2501 bad:
2502 		if (rc) {
2503 			ibdev_err(&rdev->ibdev,
2504 				  "Post send failed opcode = %#x rc = %d",
2505 				  wr->opcode, rc);
2506 			break;
2507 		}
2508 		wr = wr->next;
2509 	}
2510 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2511 	bnxt_ud_qp_hw_stall_workaround(qp);
2512 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2513 	return rc;
2514 }
2515 
2516 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr,
2517 		      const struct ib_send_wr **bad_wr)
2518 {
2519 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2520 	struct bnxt_qplib_swqe wqe;
2521 	int rc = 0, payload_sz = 0;
2522 	unsigned long flags;
2523 
2524 	spin_lock_irqsave(&qp->sq_lock, flags);
2525 	while (wr) {
2526 		/* House keeping */
2527 		memset(&wqe, 0, sizeof(wqe));
2528 
2529 		/* Common */
2530 		wqe.num_sge = wr->num_sge;
2531 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2532 			ibdev_err(&qp->rdev->ibdev,
2533 				  "Limit exceeded for Send SGEs");
2534 			rc = -EINVAL;
2535 			goto bad;
2536 		}
2537 
2538 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2539 		if (payload_sz < 0) {
2540 			rc = -EINVAL;
2541 			goto bad;
2542 		}
2543 		wqe.wr_id = wr->wr_id;
2544 
2545 		switch (wr->opcode) {
2546 		case IB_WR_SEND:
2547 		case IB_WR_SEND_WITH_IMM:
2548 			if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) {
2549 				rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe,
2550 							       payload_sz);
2551 				if (rc)
2552 					goto bad;
2553 				wqe.rawqp1.lflags |=
2554 					SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC;
2555 			}
2556 			switch (wr->send_flags) {
2557 			case IB_SEND_IP_CSUM:
2558 				wqe.rawqp1.lflags |=
2559 					SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM;
2560 				break;
2561 			default:
2562 				break;
2563 			}
2564 			/* fall through */
2565 		case IB_WR_SEND_WITH_INV:
2566 			rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2567 			break;
2568 		case IB_WR_RDMA_WRITE:
2569 		case IB_WR_RDMA_WRITE_WITH_IMM:
2570 		case IB_WR_RDMA_READ:
2571 			rc = bnxt_re_build_rdma_wqe(wr, &wqe);
2572 			break;
2573 		case IB_WR_ATOMIC_CMP_AND_SWP:
2574 		case IB_WR_ATOMIC_FETCH_AND_ADD:
2575 			rc = bnxt_re_build_atomic_wqe(wr, &wqe);
2576 			break;
2577 		case IB_WR_RDMA_READ_WITH_INV:
2578 			ibdev_err(&qp->rdev->ibdev,
2579 				  "RDMA Read with Invalidate is not supported");
2580 			rc = -EINVAL;
2581 			goto bad;
2582 		case IB_WR_LOCAL_INV:
2583 			rc = bnxt_re_build_inv_wqe(wr, &wqe);
2584 			break;
2585 		case IB_WR_REG_MR:
2586 			rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe);
2587 			break;
2588 		default:
2589 			/* Unsupported WRs */
2590 			ibdev_err(&qp->rdev->ibdev,
2591 				  "WR (%#x) is not supported", wr->opcode);
2592 			rc = -EINVAL;
2593 			goto bad;
2594 		}
2595 		if (!rc)
2596 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2597 bad:
2598 		if (rc) {
2599 			ibdev_err(&qp->rdev->ibdev,
2600 				  "post_send failed op:%#x qps = %#x rc = %d\n",
2601 				  wr->opcode, qp->qplib_qp.state, rc);
2602 			*bad_wr = wr;
2603 			break;
2604 		}
2605 		wr = wr->next;
2606 	}
2607 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2608 	bnxt_ud_qp_hw_stall_workaround(qp);
2609 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2610 
2611 	return rc;
2612 }
2613 
2614 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev,
2615 				       struct bnxt_re_qp *qp,
2616 				       const struct ib_recv_wr *wr)
2617 {
2618 	struct bnxt_qplib_swqe wqe;
2619 	int rc = 0;
2620 
2621 	memset(&wqe, 0, sizeof(wqe));
2622 	while (wr) {
2623 		/* House keeping */
2624 		memset(&wqe, 0, sizeof(wqe));
2625 
2626 		/* Common */
2627 		wqe.num_sge = wr->num_sge;
2628 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2629 			ibdev_err(&rdev->ibdev,
2630 				  "Limit exceeded for Receive SGEs");
2631 			rc = -EINVAL;
2632 			break;
2633 		}
2634 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
2635 		wqe.wr_id = wr->wr_id;
2636 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2637 
2638 		rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2639 		if (rc)
2640 			break;
2641 
2642 		wr = wr->next;
2643 	}
2644 	if (!rc)
2645 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2646 	return rc;
2647 }
2648 
2649 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
2650 		      const struct ib_recv_wr **bad_wr)
2651 {
2652 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2653 	struct bnxt_qplib_swqe wqe;
2654 	int rc = 0, payload_sz = 0;
2655 	unsigned long flags;
2656 	u32 count = 0;
2657 
2658 	spin_lock_irqsave(&qp->rq_lock, flags);
2659 	while (wr) {
2660 		/* House keeping */
2661 		memset(&wqe, 0, sizeof(wqe));
2662 
2663 		/* Common */
2664 		wqe.num_sge = wr->num_sge;
2665 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2666 			ibdev_err(&qp->rdev->ibdev,
2667 				  "Limit exceeded for Receive SGEs");
2668 			rc = -EINVAL;
2669 			*bad_wr = wr;
2670 			break;
2671 		}
2672 
2673 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
2674 					       wr->num_sge);
2675 		wqe.wr_id = wr->wr_id;
2676 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2677 
2678 		if (ib_qp->qp_type == IB_QPT_GSI &&
2679 		    qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI)
2680 			rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe,
2681 							      payload_sz);
2682 		if (!rc)
2683 			rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2684 		if (rc) {
2685 			*bad_wr = wr;
2686 			break;
2687 		}
2688 
2689 		/* Ring DB if the RQEs posted reaches a threshold value */
2690 		if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
2691 			bnxt_qplib_post_recv_db(&qp->qplib_qp);
2692 			count = 0;
2693 		}
2694 
2695 		wr = wr->next;
2696 	}
2697 
2698 	if (count)
2699 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2700 
2701 	spin_unlock_irqrestore(&qp->rq_lock, flags);
2702 
2703 	return rc;
2704 }
2705 
2706 /* Completion Queues */
2707 void bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
2708 {
2709 	struct bnxt_re_cq *cq;
2710 	struct bnxt_qplib_nq *nq;
2711 	struct bnxt_re_dev *rdev;
2712 
2713 	cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
2714 	rdev = cq->rdev;
2715 	nq = cq->qplib_cq.nq;
2716 
2717 	bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2718 	ib_umem_release(cq->umem);
2719 
2720 	atomic_dec(&rdev->cq_count);
2721 	nq->budget--;
2722 	kfree(cq->cql);
2723 }
2724 
2725 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
2726 		      struct ib_udata *udata)
2727 {
2728 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev);
2729 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
2730 	struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq);
2731 	int rc, entries;
2732 	int cqe = attr->cqe;
2733 	struct bnxt_qplib_nq *nq = NULL;
2734 	unsigned int nq_alloc_cnt;
2735 
2736 	/* Validate CQ fields */
2737 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
2738 		ibdev_err(&rdev->ibdev, "Failed to create CQ -max exceeded");
2739 		return -EINVAL;
2740 	}
2741 
2742 	cq->rdev = rdev;
2743 	cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
2744 
2745 	entries = roundup_pow_of_two(cqe + 1);
2746 	if (entries > dev_attr->max_cq_wqes + 1)
2747 		entries = dev_attr->max_cq_wqes + 1;
2748 
2749 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
2750 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
2751 	if (udata) {
2752 		struct bnxt_re_cq_req req;
2753 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
2754 			udata, struct bnxt_re_ucontext, ib_uctx);
2755 		if (ib_copy_from_udata(&req, udata, sizeof(req))) {
2756 			rc = -EFAULT;
2757 			goto fail;
2758 		}
2759 
2760 		cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va,
2761 				       entries * sizeof(struct cq_base),
2762 				       IB_ACCESS_LOCAL_WRITE);
2763 		if (IS_ERR(cq->umem)) {
2764 			rc = PTR_ERR(cq->umem);
2765 			goto fail;
2766 		}
2767 		cq->qplib_cq.sg_info.sghead = cq->umem->sg_head.sgl;
2768 		cq->qplib_cq.sg_info.npages = ib_umem_num_pages(cq->umem);
2769 		cq->qplib_cq.sg_info.nmap = cq->umem->nmap;
2770 		cq->qplib_cq.dpi = &uctx->dpi;
2771 	} else {
2772 		cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL);
2773 		cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe),
2774 				  GFP_KERNEL);
2775 		if (!cq->cql) {
2776 			rc = -ENOMEM;
2777 			goto fail;
2778 		}
2779 
2780 		cq->qplib_cq.dpi = &rdev->dpi_privileged;
2781 	}
2782 	/*
2783 	 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
2784 	 * used for getting the NQ index.
2785 	 */
2786 	nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
2787 	nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
2788 	cq->qplib_cq.max_wqe = entries;
2789 	cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
2790 	cq->qplib_cq.nq	= nq;
2791 
2792 	rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
2793 	if (rc) {
2794 		ibdev_err(&rdev->ibdev, "Failed to create HW CQ");
2795 		goto fail;
2796 	}
2797 
2798 	cq->ib_cq.cqe = entries;
2799 	cq->cq_period = cq->qplib_cq.period;
2800 	nq->budget++;
2801 
2802 	atomic_inc(&rdev->cq_count);
2803 	spin_lock_init(&cq->cq_lock);
2804 
2805 	if (udata) {
2806 		struct bnxt_re_cq_resp resp;
2807 
2808 		resp.cqid = cq->qplib_cq.id;
2809 		resp.tail = cq->qplib_cq.hwq.cons;
2810 		resp.phase = cq->qplib_cq.period;
2811 		resp.rsvd = 0;
2812 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
2813 		if (rc) {
2814 			ibdev_err(&rdev->ibdev, "Failed to copy CQ udata");
2815 			bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2816 			goto c2fail;
2817 		}
2818 	}
2819 
2820 	return 0;
2821 
2822 c2fail:
2823 	ib_umem_release(cq->umem);
2824 fail:
2825 	kfree(cq->cql);
2826 	return rc;
2827 }
2828 
2829 static u8 __req_to_ib_wc_status(u8 qstatus)
2830 {
2831 	switch (qstatus) {
2832 	case CQ_REQ_STATUS_OK:
2833 		return IB_WC_SUCCESS;
2834 	case CQ_REQ_STATUS_BAD_RESPONSE_ERR:
2835 		return IB_WC_BAD_RESP_ERR;
2836 	case CQ_REQ_STATUS_LOCAL_LENGTH_ERR:
2837 		return IB_WC_LOC_LEN_ERR;
2838 	case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR:
2839 		return IB_WC_LOC_QP_OP_ERR;
2840 	case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR:
2841 		return IB_WC_LOC_PROT_ERR;
2842 	case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR:
2843 		return IB_WC_GENERAL_ERR;
2844 	case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR:
2845 		return IB_WC_REM_INV_REQ_ERR;
2846 	case CQ_REQ_STATUS_REMOTE_ACCESS_ERR:
2847 		return IB_WC_REM_ACCESS_ERR;
2848 	case CQ_REQ_STATUS_REMOTE_OPERATION_ERR:
2849 		return IB_WC_REM_OP_ERR;
2850 	case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR:
2851 		return IB_WC_RNR_RETRY_EXC_ERR;
2852 	case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR:
2853 		return IB_WC_RETRY_EXC_ERR;
2854 	case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR:
2855 		return IB_WC_WR_FLUSH_ERR;
2856 	default:
2857 		return IB_WC_GENERAL_ERR;
2858 	}
2859 	return 0;
2860 }
2861 
2862 static u8 __rawqp1_to_ib_wc_status(u8 qstatus)
2863 {
2864 	switch (qstatus) {
2865 	case CQ_RES_RAWETH_QP1_STATUS_OK:
2866 		return IB_WC_SUCCESS;
2867 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR:
2868 		return IB_WC_LOC_ACCESS_ERR;
2869 	case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR:
2870 		return IB_WC_LOC_LEN_ERR;
2871 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR:
2872 		return IB_WC_LOC_PROT_ERR;
2873 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR:
2874 		return IB_WC_LOC_QP_OP_ERR;
2875 	case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR:
2876 		return IB_WC_GENERAL_ERR;
2877 	case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR:
2878 		return IB_WC_WR_FLUSH_ERR;
2879 	case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR:
2880 		return IB_WC_WR_FLUSH_ERR;
2881 	default:
2882 		return IB_WC_GENERAL_ERR;
2883 	}
2884 }
2885 
2886 static u8 __rc_to_ib_wc_status(u8 qstatus)
2887 {
2888 	switch (qstatus) {
2889 	case CQ_RES_RC_STATUS_OK:
2890 		return IB_WC_SUCCESS;
2891 	case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR:
2892 		return IB_WC_LOC_ACCESS_ERR;
2893 	case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR:
2894 		return IB_WC_LOC_LEN_ERR;
2895 	case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR:
2896 		return IB_WC_LOC_PROT_ERR;
2897 	case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR:
2898 		return IB_WC_LOC_QP_OP_ERR;
2899 	case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR:
2900 		return IB_WC_GENERAL_ERR;
2901 	case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR:
2902 		return IB_WC_REM_INV_REQ_ERR;
2903 	case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR:
2904 		return IB_WC_WR_FLUSH_ERR;
2905 	case CQ_RES_RC_STATUS_HW_FLUSH_ERR:
2906 		return IB_WC_WR_FLUSH_ERR;
2907 	default:
2908 		return IB_WC_GENERAL_ERR;
2909 	}
2910 }
2911 
2912 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe)
2913 {
2914 	switch (cqe->type) {
2915 	case BNXT_QPLIB_SWQE_TYPE_SEND:
2916 		wc->opcode = IB_WC_SEND;
2917 		break;
2918 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM:
2919 		wc->opcode = IB_WC_SEND;
2920 		wc->wc_flags |= IB_WC_WITH_IMM;
2921 		break;
2922 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV:
2923 		wc->opcode = IB_WC_SEND;
2924 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
2925 		break;
2926 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE:
2927 		wc->opcode = IB_WC_RDMA_WRITE;
2928 		break;
2929 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM:
2930 		wc->opcode = IB_WC_RDMA_WRITE;
2931 		wc->wc_flags |= IB_WC_WITH_IMM;
2932 		break;
2933 	case BNXT_QPLIB_SWQE_TYPE_RDMA_READ:
2934 		wc->opcode = IB_WC_RDMA_READ;
2935 		break;
2936 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP:
2937 		wc->opcode = IB_WC_COMP_SWAP;
2938 		break;
2939 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD:
2940 		wc->opcode = IB_WC_FETCH_ADD;
2941 		break;
2942 	case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV:
2943 		wc->opcode = IB_WC_LOCAL_INV;
2944 		break;
2945 	case BNXT_QPLIB_SWQE_TYPE_REG_MR:
2946 		wc->opcode = IB_WC_REG_MR;
2947 		break;
2948 	default:
2949 		wc->opcode = IB_WC_SEND;
2950 		break;
2951 	}
2952 
2953 	wc->status = __req_to_ib_wc_status(cqe->status);
2954 }
2955 
2956 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags,
2957 				     u16 raweth_qp1_flags2)
2958 {
2959 	bool is_ipv6 = false, is_ipv4 = false;
2960 
2961 	/* raweth_qp1_flags Bit 9-6 indicates itype */
2962 	if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
2963 	    != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
2964 		return -1;
2965 
2966 	if (raweth_qp1_flags2 &
2967 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC &&
2968 	    raweth_qp1_flags2 &
2969 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) {
2970 		/* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
2971 		(raweth_qp1_flags2 &
2972 		 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ?
2973 			(is_ipv6 = true) : (is_ipv4 = true);
2974 		return ((is_ipv6) ?
2975 			 BNXT_RE_ROCEV2_IPV6_PACKET :
2976 			 BNXT_RE_ROCEV2_IPV4_PACKET);
2977 	} else {
2978 		return BNXT_RE_ROCE_V1_PACKET;
2979 	}
2980 }
2981 
2982 static int bnxt_re_to_ib_nw_type(int nw_type)
2983 {
2984 	u8 nw_hdr_type = 0xFF;
2985 
2986 	switch (nw_type) {
2987 	case BNXT_RE_ROCE_V1_PACKET:
2988 		nw_hdr_type = RDMA_NETWORK_ROCE_V1;
2989 		break;
2990 	case BNXT_RE_ROCEV2_IPV4_PACKET:
2991 		nw_hdr_type = RDMA_NETWORK_IPV4;
2992 		break;
2993 	case BNXT_RE_ROCEV2_IPV6_PACKET:
2994 		nw_hdr_type = RDMA_NETWORK_IPV6;
2995 		break;
2996 	}
2997 	return nw_hdr_type;
2998 }
2999 
3000 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev,
3001 				       void *rq_hdr_buf)
3002 {
3003 	u8 *tmp_buf = NULL;
3004 	struct ethhdr *eth_hdr;
3005 	u16 eth_type;
3006 	bool rc = false;
3007 
3008 	tmp_buf = (u8 *)rq_hdr_buf;
3009 	/*
3010 	 * If dest mac is not same as I/F mac, this could be a
3011 	 * loopback address or multicast address, check whether
3012 	 * it is a loopback packet
3013 	 */
3014 	if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) {
3015 		tmp_buf += 4;
3016 		/* Check the  ether type */
3017 		eth_hdr = (struct ethhdr *)tmp_buf;
3018 		eth_type = ntohs(eth_hdr->h_proto);
3019 		switch (eth_type) {
3020 		case ETH_P_IBOE:
3021 			rc = true;
3022 			break;
3023 		case ETH_P_IP:
3024 		case ETH_P_IPV6: {
3025 			u32 len;
3026 			struct udphdr *udp_hdr;
3027 
3028 			len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) :
3029 						      sizeof(struct ipv6hdr));
3030 			tmp_buf += sizeof(struct ethhdr) + len;
3031 			udp_hdr = (struct udphdr *)tmp_buf;
3032 			if (ntohs(udp_hdr->dest) ==
3033 				    ROCE_V2_UDP_DPORT)
3034 				rc = true;
3035 			break;
3036 			}
3037 		default:
3038 			break;
3039 		}
3040 	}
3041 
3042 	return rc;
3043 }
3044 
3045 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *gsi_qp,
3046 					 struct bnxt_qplib_cqe *cqe)
3047 {
3048 	struct bnxt_re_dev *rdev = gsi_qp->rdev;
3049 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3050 	struct bnxt_re_qp *gsi_sqp = rdev->gsi_ctx.gsi_sqp;
3051 	struct bnxt_re_ah *gsi_sah;
3052 	struct ib_send_wr *swr;
3053 	struct ib_ud_wr udwr;
3054 	struct ib_recv_wr rwr;
3055 	int pkt_type = 0;
3056 	u32 tbl_idx;
3057 	void *rq_hdr_buf;
3058 	dma_addr_t rq_hdr_buf_map;
3059 	dma_addr_t shrq_hdr_buf_map;
3060 	u32 offset = 0;
3061 	u32 skip_bytes = 0;
3062 	struct ib_sge s_sge[2];
3063 	struct ib_sge r_sge[2];
3064 	int rc;
3065 
3066 	memset(&udwr, 0, sizeof(udwr));
3067 	memset(&rwr, 0, sizeof(rwr));
3068 	memset(&s_sge, 0, sizeof(s_sge));
3069 	memset(&r_sge, 0, sizeof(r_sge));
3070 
3071 	swr = &udwr.wr;
3072 	tbl_idx = cqe->wr_id;
3073 
3074 	rq_hdr_buf = gsi_qp->qplib_qp.rq_hdr_buf +
3075 			(tbl_idx * gsi_qp->qplib_qp.rq_hdr_buf_size);
3076 	rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3077 							  tbl_idx);
3078 
3079 	/* Shadow QP header buffer */
3080 	shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3081 							    tbl_idx);
3082 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3083 
3084 	/* Store this cqe */
3085 	memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe));
3086 	sqp_entry->qp1_qp = gsi_qp;
3087 
3088 	/* Find packet type from the cqe */
3089 
3090 	pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags,
3091 					     cqe->raweth_qp1_flags2);
3092 	if (pkt_type < 0) {
3093 		ibdev_err(&rdev->ibdev, "Invalid packet\n");
3094 		return -EINVAL;
3095 	}
3096 
3097 	/* Adjust the offset for the user buffer and post in the rq */
3098 
3099 	if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET)
3100 		offset = 20;
3101 
3102 	/*
3103 	 * QP1 loopback packet has 4 bytes of internal header before
3104 	 * ether header. Skip these four bytes.
3105 	 */
3106 	if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf))
3107 		skip_bytes = 4;
3108 
3109 	/* First send SGE . Skip the ether header*/
3110 	s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
3111 			+ skip_bytes;
3112 	s_sge[0].lkey = 0xFFFFFFFF;
3113 	s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 :
3114 				BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
3115 
3116 	/* Second Send SGE */
3117 	s_sge[1].addr = s_sge[0].addr + s_sge[0].length +
3118 			BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE;
3119 	if (pkt_type != BNXT_RE_ROCE_V1_PACKET)
3120 		s_sge[1].addr += 8;
3121 	s_sge[1].lkey = 0xFFFFFFFF;
3122 	s_sge[1].length = 256;
3123 
3124 	/* First recv SGE */
3125 
3126 	r_sge[0].addr = shrq_hdr_buf_map;
3127 	r_sge[0].lkey = 0xFFFFFFFF;
3128 	r_sge[0].length = 40;
3129 
3130 	r_sge[1].addr = sqp_entry->sge.addr + offset;
3131 	r_sge[1].lkey = sqp_entry->sge.lkey;
3132 	r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset;
3133 
3134 	/* Create receive work request */
3135 	rwr.num_sge = 2;
3136 	rwr.sg_list = r_sge;
3137 	rwr.wr_id = tbl_idx;
3138 	rwr.next = NULL;
3139 
3140 	rc = bnxt_re_post_recv_shadow_qp(rdev, gsi_sqp, &rwr);
3141 	if (rc) {
3142 		ibdev_err(&rdev->ibdev,
3143 			  "Failed to post Rx buffers to shadow QP");
3144 		return -ENOMEM;
3145 	}
3146 
3147 	swr->num_sge = 2;
3148 	swr->sg_list = s_sge;
3149 	swr->wr_id = tbl_idx;
3150 	swr->opcode = IB_WR_SEND;
3151 	swr->next = NULL;
3152 	gsi_sah = rdev->gsi_ctx.gsi_sah;
3153 	udwr.ah = &gsi_sah->ib_ah;
3154 	udwr.remote_qpn = gsi_sqp->qplib_qp.id;
3155 	udwr.remote_qkey = gsi_sqp->qplib_qp.qkey;
3156 
3157 	/* post data received  in the send queue */
3158 	rc = bnxt_re_post_send_shadow_qp(rdev, gsi_sqp, swr);
3159 
3160 	return 0;
3161 }
3162 
3163 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc,
3164 					  struct bnxt_qplib_cqe *cqe)
3165 {
3166 	wc->opcode = IB_WC_RECV;
3167 	wc->status = __rawqp1_to_ib_wc_status(cqe->status);
3168 	wc->wc_flags |= IB_WC_GRH;
3169 }
3170 
3171 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe,
3172 				u16 *vid, u8 *sl)
3173 {
3174 	bool ret = false;
3175 	u32 metadata;
3176 	u16 tpid;
3177 
3178 	metadata = orig_cqe->raweth_qp1_metadata;
3179 	if (orig_cqe->raweth_qp1_flags2 &
3180 		CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) {
3181 		tpid = ((metadata &
3182 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >>
3183 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT);
3184 		if (tpid == ETH_P_8021Q) {
3185 			*vid = metadata &
3186 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK;
3187 			*sl = (metadata &
3188 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >>
3189 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT;
3190 			ret = true;
3191 		}
3192 	}
3193 
3194 	return ret;
3195 }
3196 
3197 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc,
3198 				      struct bnxt_qplib_cqe *cqe)
3199 {
3200 	wc->opcode = IB_WC_RECV;
3201 	wc->status = __rc_to_ib_wc_status(cqe->status);
3202 
3203 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3204 		wc->wc_flags |= IB_WC_WITH_IMM;
3205 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3206 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3207 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3208 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3209 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3210 }
3211 
3212 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *gsi_sqp,
3213 					     struct ib_wc *wc,
3214 					     struct bnxt_qplib_cqe *cqe)
3215 {
3216 	struct bnxt_re_dev *rdev = gsi_sqp->rdev;
3217 	struct bnxt_re_qp *gsi_qp = NULL;
3218 	struct bnxt_qplib_cqe *orig_cqe = NULL;
3219 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3220 	int nw_type;
3221 	u32 tbl_idx;
3222 	u16 vlan_id;
3223 	u8 sl;
3224 
3225 	tbl_idx = cqe->wr_id;
3226 
3227 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3228 	gsi_qp = sqp_entry->qp1_qp;
3229 	orig_cqe = &sqp_entry->cqe;
3230 
3231 	wc->wr_id = sqp_entry->wrid;
3232 	wc->byte_len = orig_cqe->length;
3233 	wc->qp = &gsi_qp->ib_qp;
3234 
3235 	wc->ex.imm_data = orig_cqe->immdata;
3236 	wc->src_qp = orig_cqe->src_qp;
3237 	memcpy(wc->smac, orig_cqe->smac, ETH_ALEN);
3238 	if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) {
3239 		wc->vlan_id = vlan_id;
3240 		wc->sl = sl;
3241 		wc->wc_flags |= IB_WC_WITH_VLAN;
3242 	}
3243 	wc->port_num = 1;
3244 	wc->vendor_err = orig_cqe->status;
3245 
3246 	wc->opcode = IB_WC_RECV;
3247 	wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status);
3248 	wc->wc_flags |= IB_WC_GRH;
3249 
3250 	nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags,
3251 					    orig_cqe->raweth_qp1_flags2);
3252 	if (nw_type >= 0) {
3253 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3254 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3255 	}
3256 }
3257 
3258 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp,
3259 				      struct ib_wc *wc,
3260 				      struct bnxt_qplib_cqe *cqe)
3261 {
3262 	u8 nw_type;
3263 
3264 	wc->opcode = IB_WC_RECV;
3265 	wc->status = __rc_to_ib_wc_status(cqe->status);
3266 
3267 	if (cqe->flags & CQ_RES_UD_FLAGS_IMM)
3268 		wc->wc_flags |= IB_WC_WITH_IMM;
3269 	/* report only on GSI QP for Thor */
3270 	if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) {
3271 		wc->wc_flags |= IB_WC_GRH;
3272 		memcpy(wc->smac, cqe->smac, ETH_ALEN);
3273 		wc->wc_flags |= IB_WC_WITH_SMAC;
3274 		if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) {
3275 			wc->vlan_id = (cqe->cfa_meta & 0xFFF);
3276 			if (wc->vlan_id < 0x1000)
3277 				wc->wc_flags |= IB_WC_WITH_VLAN;
3278 		}
3279 		nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >>
3280 			   CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT;
3281 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3282 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3283 	}
3284 
3285 }
3286 
3287 static int send_phantom_wqe(struct bnxt_re_qp *qp)
3288 {
3289 	struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
3290 	unsigned long flags;
3291 	int rc = 0;
3292 
3293 	spin_lock_irqsave(&qp->sq_lock, flags);
3294 
3295 	rc = bnxt_re_bind_fence_mw(lib_qp);
3296 	if (!rc) {
3297 		lib_qp->sq.phantom_wqe_cnt++;
3298 		ibdev_dbg(&qp->rdev->ibdev,
3299 			  "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
3300 			  lib_qp->id, lib_qp->sq.hwq.prod,
3301 			  HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
3302 			  lib_qp->sq.phantom_wqe_cnt);
3303 	}
3304 
3305 	spin_unlock_irqrestore(&qp->sq_lock, flags);
3306 	return rc;
3307 }
3308 
3309 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
3310 {
3311 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3312 	struct bnxt_re_qp *qp, *sh_qp;
3313 	struct bnxt_qplib_cqe *cqe;
3314 	int i, ncqe, budget;
3315 	struct bnxt_qplib_q *sq;
3316 	struct bnxt_qplib_qp *lib_qp;
3317 	u32 tbl_idx;
3318 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3319 	unsigned long flags;
3320 
3321 	spin_lock_irqsave(&cq->cq_lock, flags);
3322 	budget = min_t(u32, num_entries, cq->max_cql);
3323 	num_entries = budget;
3324 	if (!cq->cql) {
3325 		ibdev_err(&cq->rdev->ibdev, "POLL CQ : no CQL to use");
3326 		goto exit;
3327 	}
3328 	cqe = &cq->cql[0];
3329 	while (budget) {
3330 		lib_qp = NULL;
3331 		ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
3332 		if (lib_qp) {
3333 			sq = &lib_qp->sq;
3334 			if (sq->send_phantom) {
3335 				qp = container_of(lib_qp,
3336 						  struct bnxt_re_qp, qplib_qp);
3337 				if (send_phantom_wqe(qp) == -ENOMEM)
3338 					ibdev_err(&cq->rdev->ibdev,
3339 						  "Phantom failed! Scheduled to send again\n");
3340 				else
3341 					sq->send_phantom = false;
3342 			}
3343 		}
3344 		if (ncqe < budget)
3345 			ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq,
3346 							      cqe + ncqe,
3347 							      budget - ncqe);
3348 
3349 		if (!ncqe)
3350 			break;
3351 
3352 		for (i = 0; i < ncqe; i++, cqe++) {
3353 			/* Transcribe each qplib_wqe back to ib_wc */
3354 			memset(wc, 0, sizeof(*wc));
3355 
3356 			wc->wr_id = cqe->wr_id;
3357 			wc->byte_len = cqe->length;
3358 			qp = container_of
3359 				((struct bnxt_qplib_qp *)
3360 				 (unsigned long)(cqe->qp_handle),
3361 				 struct bnxt_re_qp, qplib_qp);
3362 			if (!qp) {
3363 				ibdev_err(&cq->rdev->ibdev, "POLL CQ : bad QP handle");
3364 				continue;
3365 			}
3366 			wc->qp = &qp->ib_qp;
3367 			wc->ex.imm_data = cqe->immdata;
3368 			wc->src_qp = cqe->src_qp;
3369 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
3370 			wc->port_num = 1;
3371 			wc->vendor_err = cqe->status;
3372 
3373 			switch (cqe->opcode) {
3374 			case CQ_BASE_CQE_TYPE_REQ:
3375 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3376 				if (sh_qp &&
3377 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3378 					/* Handle this completion with
3379 					 * the stored completion
3380 					 */
3381 					memset(wc, 0, sizeof(*wc));
3382 					continue;
3383 				}
3384 				bnxt_re_process_req_wc(wc, cqe);
3385 				break;
3386 			case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1:
3387 				if (!cqe->status) {
3388 					int rc = 0;
3389 
3390 					rc = bnxt_re_process_raw_qp_pkt_rx
3391 								(qp, cqe);
3392 					if (!rc) {
3393 						memset(wc, 0, sizeof(*wc));
3394 						continue;
3395 					}
3396 					cqe->status = -1;
3397 				}
3398 				/* Errors need not be looped back.
3399 				 * But change the wr_id to the one
3400 				 * stored in the table
3401 				 */
3402 				tbl_idx = cqe->wr_id;
3403 				sqp_entry = &cq->rdev->gsi_ctx.sqp_tbl[tbl_idx];
3404 				wc->wr_id = sqp_entry->wrid;
3405 				bnxt_re_process_res_rawqp1_wc(wc, cqe);
3406 				break;
3407 			case CQ_BASE_CQE_TYPE_RES_RC:
3408 				bnxt_re_process_res_rc_wc(wc, cqe);
3409 				break;
3410 			case CQ_BASE_CQE_TYPE_RES_UD:
3411 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3412 				if (sh_qp &&
3413 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3414 					/* Handle this completion with
3415 					 * the stored completion
3416 					 */
3417 					if (cqe->status) {
3418 						continue;
3419 					} else {
3420 						bnxt_re_process_res_shadow_qp_wc
3421 								(qp, wc, cqe);
3422 						break;
3423 					}
3424 				}
3425 				bnxt_re_process_res_ud_wc(qp, wc, cqe);
3426 				break;
3427 			default:
3428 				ibdev_err(&cq->rdev->ibdev,
3429 					  "POLL CQ : type 0x%x not handled",
3430 					  cqe->opcode);
3431 				continue;
3432 			}
3433 			wc++;
3434 			budget--;
3435 		}
3436 	}
3437 exit:
3438 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3439 	return num_entries - budget;
3440 }
3441 
3442 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq,
3443 			  enum ib_cq_notify_flags ib_cqn_flags)
3444 {
3445 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3446 	int type = 0, rc = 0;
3447 	unsigned long flags;
3448 
3449 	spin_lock_irqsave(&cq->cq_lock, flags);
3450 	/* Trigger on the very next completion */
3451 	if (ib_cqn_flags & IB_CQ_NEXT_COMP)
3452 		type = DBC_DBC_TYPE_CQ_ARMALL;
3453 	/* Trigger on the next solicited completion */
3454 	else if (ib_cqn_flags & IB_CQ_SOLICITED)
3455 		type = DBC_DBC_TYPE_CQ_ARMSE;
3456 
3457 	/* Poll to see if there are missed events */
3458 	if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3459 	    !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) {
3460 		rc = 1;
3461 		goto exit;
3462 	}
3463 	bnxt_qplib_req_notify_cq(&cq->qplib_cq, type);
3464 
3465 exit:
3466 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3467 	return rc;
3468 }
3469 
3470 /* Memory Regions */
3471 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags)
3472 {
3473 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3474 	struct bnxt_re_dev *rdev = pd->rdev;
3475 	struct bnxt_re_mr *mr;
3476 	u64 pbl = 0;
3477 	int rc;
3478 
3479 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3480 	if (!mr)
3481 		return ERR_PTR(-ENOMEM);
3482 
3483 	mr->rdev = rdev;
3484 	mr->qplib_mr.pd = &pd->qplib_pd;
3485 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3486 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3487 
3488 	/* Allocate and register 0 as the address */
3489 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3490 	if (rc)
3491 		goto fail;
3492 
3493 	mr->qplib_mr.hwq.level = PBL_LVL_MAX;
3494 	mr->qplib_mr.total_size = -1; /* Infinte length */
3495 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl, 0, false,
3496 			       PAGE_SIZE);
3497 	if (rc)
3498 		goto fail_mr;
3499 
3500 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3501 	if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ |
3502 			       IB_ACCESS_REMOTE_ATOMIC))
3503 		mr->ib_mr.rkey = mr->ib_mr.lkey;
3504 	atomic_inc(&rdev->mr_count);
3505 
3506 	return &mr->ib_mr;
3507 
3508 fail_mr:
3509 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3510 fail:
3511 	kfree(mr);
3512 	return ERR_PTR(rc);
3513 }
3514 
3515 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3516 {
3517 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3518 	struct bnxt_re_dev *rdev = mr->rdev;
3519 	int rc;
3520 
3521 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3522 	if (rc) {
3523 		ibdev_err(&rdev->ibdev, "Dereg MR failed: %#x\n", rc);
3524 		return rc;
3525 	}
3526 
3527 	if (mr->pages) {
3528 		rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
3529 							&mr->qplib_frpl);
3530 		kfree(mr->pages);
3531 		mr->npages = 0;
3532 		mr->pages = NULL;
3533 	}
3534 	ib_umem_release(mr->ib_umem);
3535 
3536 	kfree(mr);
3537 	atomic_dec(&rdev->mr_count);
3538 	return rc;
3539 }
3540 
3541 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr)
3542 {
3543 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3544 
3545 	if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs))
3546 		return -ENOMEM;
3547 
3548 	mr->pages[mr->npages++] = addr;
3549 	return 0;
3550 }
3551 
3552 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents,
3553 		      unsigned int *sg_offset)
3554 {
3555 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3556 
3557 	mr->npages = 0;
3558 	return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page);
3559 }
3560 
3561 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type,
3562 			       u32 max_num_sg, struct ib_udata *udata)
3563 {
3564 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3565 	struct bnxt_re_dev *rdev = pd->rdev;
3566 	struct bnxt_re_mr *mr = NULL;
3567 	int rc;
3568 
3569 	if (type != IB_MR_TYPE_MEM_REG) {
3570 		ibdev_dbg(&rdev->ibdev, "MR type 0x%x not supported", type);
3571 		return ERR_PTR(-EINVAL);
3572 	}
3573 	if (max_num_sg > MAX_PBL_LVL_1_PGS)
3574 		return ERR_PTR(-EINVAL);
3575 
3576 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3577 	if (!mr)
3578 		return ERR_PTR(-ENOMEM);
3579 
3580 	mr->rdev = rdev;
3581 	mr->qplib_mr.pd = &pd->qplib_pd;
3582 	mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR;
3583 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3584 
3585 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3586 	if (rc)
3587 		goto bail;
3588 
3589 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3590 	mr->ib_mr.rkey = mr->ib_mr.lkey;
3591 
3592 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3593 	if (!mr->pages) {
3594 		rc = -ENOMEM;
3595 		goto fail;
3596 	}
3597 	rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res,
3598 						 &mr->qplib_frpl, max_num_sg);
3599 	if (rc) {
3600 		ibdev_err(&rdev->ibdev,
3601 			  "Failed to allocate HW FR page list");
3602 		goto fail_mr;
3603 	}
3604 
3605 	atomic_inc(&rdev->mr_count);
3606 	return &mr->ib_mr;
3607 
3608 fail_mr:
3609 	kfree(mr->pages);
3610 fail:
3611 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3612 bail:
3613 	kfree(mr);
3614 	return ERR_PTR(rc);
3615 }
3616 
3617 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
3618 			       struct ib_udata *udata)
3619 {
3620 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3621 	struct bnxt_re_dev *rdev = pd->rdev;
3622 	struct bnxt_re_mw *mw;
3623 	int rc;
3624 
3625 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
3626 	if (!mw)
3627 		return ERR_PTR(-ENOMEM);
3628 	mw->rdev = rdev;
3629 	mw->qplib_mw.pd = &pd->qplib_pd;
3630 
3631 	mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
3632 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
3633 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
3634 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
3635 	if (rc) {
3636 		ibdev_err(&rdev->ibdev, "Allocate MW failed!");
3637 		goto fail;
3638 	}
3639 	mw->ib_mw.rkey = mw->qplib_mw.rkey;
3640 
3641 	atomic_inc(&rdev->mw_count);
3642 	return &mw->ib_mw;
3643 
3644 fail:
3645 	kfree(mw);
3646 	return ERR_PTR(rc);
3647 }
3648 
3649 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
3650 {
3651 	struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
3652 	struct bnxt_re_dev *rdev = mw->rdev;
3653 	int rc;
3654 
3655 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
3656 	if (rc) {
3657 		ibdev_err(&rdev->ibdev, "Free MW failed: %#x\n", rc);
3658 		return rc;
3659 	}
3660 
3661 	kfree(mw);
3662 	atomic_dec(&rdev->mw_count);
3663 	return rc;
3664 }
3665 
3666 static int bnxt_re_page_size_ok(int page_shift)
3667 {
3668 	switch (page_shift) {
3669 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4K:
3670 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_8K:
3671 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_64K:
3672 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_2M:
3673 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_256K:
3674 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1M:
3675 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4M:
3676 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1G:
3677 		return 1;
3678 	default:
3679 		return 0;
3680 	}
3681 }
3682 
3683 static int fill_umem_pbl_tbl(struct ib_umem *umem, u64 *pbl_tbl_orig,
3684 			     int page_shift)
3685 {
3686 	u64 *pbl_tbl = pbl_tbl_orig;
3687 	u64 page_size =  BIT_ULL(page_shift);
3688 	struct ib_block_iter biter;
3689 
3690 	rdma_for_each_block(umem->sg_head.sgl, &biter, umem->nmap, page_size)
3691 		*pbl_tbl++ = rdma_block_iter_dma_address(&biter);
3692 
3693 	return pbl_tbl - pbl_tbl_orig;
3694 }
3695 
3696 /* uverbs */
3697 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
3698 				  u64 virt_addr, int mr_access_flags,
3699 				  struct ib_udata *udata)
3700 {
3701 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3702 	struct bnxt_re_dev *rdev = pd->rdev;
3703 	struct bnxt_re_mr *mr;
3704 	struct ib_umem *umem;
3705 	u64 *pbl_tbl = NULL;
3706 	int umem_pgs, page_shift, rc;
3707 
3708 	if (length > BNXT_RE_MAX_MR_SIZE) {
3709 		ibdev_err(&rdev->ibdev, "MR Size: %lld > Max supported:%lld\n",
3710 			  length, BNXT_RE_MAX_MR_SIZE);
3711 		return ERR_PTR(-ENOMEM);
3712 	}
3713 
3714 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3715 	if (!mr)
3716 		return ERR_PTR(-ENOMEM);
3717 
3718 	mr->rdev = rdev;
3719 	mr->qplib_mr.pd = &pd->qplib_pd;
3720 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3721 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
3722 
3723 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3724 	if (rc) {
3725 		ibdev_err(&rdev->ibdev, "Failed to allocate MR");
3726 		goto free_mr;
3727 	}
3728 	/* The fixed portion of the rkey is the same as the lkey */
3729 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
3730 
3731 	umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags);
3732 	if (IS_ERR(umem)) {
3733 		ibdev_err(&rdev->ibdev, "Failed to get umem");
3734 		rc = -EFAULT;
3735 		goto free_mrw;
3736 	}
3737 	mr->ib_umem = umem;
3738 
3739 	mr->qplib_mr.va = virt_addr;
3740 	umem_pgs = ib_umem_page_count(umem);
3741 	if (!umem_pgs) {
3742 		ibdev_err(&rdev->ibdev, "umem is invalid!");
3743 		rc = -EINVAL;
3744 		goto free_umem;
3745 	}
3746 	mr->qplib_mr.total_size = length;
3747 
3748 	pbl_tbl = kcalloc(umem_pgs, sizeof(u64 *), GFP_KERNEL);
3749 	if (!pbl_tbl) {
3750 		rc = -ENOMEM;
3751 		goto free_umem;
3752 	}
3753 
3754 	page_shift = __ffs(ib_umem_find_best_pgsz(umem,
3755 				BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M,
3756 				virt_addr));
3757 
3758 	if (!bnxt_re_page_size_ok(page_shift)) {
3759 		ibdev_err(&rdev->ibdev, "umem page size unsupported!");
3760 		rc = -EFAULT;
3761 		goto fail;
3762 	}
3763 
3764 	if (page_shift == BNXT_RE_PAGE_SHIFT_4K &&
3765 	    length > BNXT_RE_MAX_MR_SIZE_LOW) {
3766 		ibdev_err(&rdev->ibdev, "Requested MR Sz:%llu Max sup:%llu",
3767 			  length, (u64)BNXT_RE_MAX_MR_SIZE_LOW);
3768 		rc = -EINVAL;
3769 		goto fail;
3770 	}
3771 
3772 	/* Map umem buf ptrs to the PBL */
3773 	umem_pgs = fill_umem_pbl_tbl(umem, pbl_tbl, page_shift);
3774 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, pbl_tbl,
3775 			       umem_pgs, false, 1 << page_shift);
3776 	if (rc) {
3777 		ibdev_err(&rdev->ibdev, "Failed to register user MR");
3778 		goto fail;
3779 	}
3780 
3781 	kfree(pbl_tbl);
3782 
3783 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3784 	mr->ib_mr.rkey = mr->qplib_mr.lkey;
3785 	atomic_inc(&rdev->mr_count);
3786 
3787 	return &mr->ib_mr;
3788 fail:
3789 	kfree(pbl_tbl);
3790 free_umem:
3791 	ib_umem_release(umem);
3792 free_mrw:
3793 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3794 free_mr:
3795 	kfree(mr);
3796 	return ERR_PTR(rc);
3797 }
3798 
3799 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
3800 {
3801 	struct ib_device *ibdev = ctx->device;
3802 	struct bnxt_re_ucontext *uctx =
3803 		container_of(ctx, struct bnxt_re_ucontext, ib_uctx);
3804 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
3805 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
3806 	struct bnxt_re_uctx_resp resp;
3807 	u32 chip_met_rev_num = 0;
3808 	int rc;
3809 
3810 	ibdev_dbg(ibdev, "ABI version requested %u", ibdev->ops.uverbs_abi_ver);
3811 
3812 	if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) {
3813 		ibdev_dbg(ibdev, " is different from the device %d ",
3814 			  BNXT_RE_ABI_VERSION);
3815 		return -EPERM;
3816 	}
3817 
3818 	uctx->rdev = rdev;
3819 
3820 	uctx->shpg = (void *)__get_free_page(GFP_KERNEL);
3821 	if (!uctx->shpg) {
3822 		rc = -ENOMEM;
3823 		goto fail;
3824 	}
3825 	spin_lock_init(&uctx->sh_lock);
3826 
3827 	resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX;
3828 	chip_met_rev_num = rdev->chip_ctx->chip_num;
3829 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_rev & 0xFF) <<
3830 			     BNXT_RE_CHIP_ID0_CHIP_REV_SFT;
3831 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_metal & 0xFF) <<
3832 			     BNXT_RE_CHIP_ID0_CHIP_MET_SFT;
3833 	resp.chip_id0 = chip_met_rev_num;
3834 	/* Future extension of chip info */
3835 	resp.chip_id1 = 0;
3836 	/*Temp, Use xa_alloc instead */
3837 	resp.dev_id = rdev->en_dev->pdev->devfn;
3838 	resp.max_qp = rdev->qplib_ctx.qpc_count;
3839 	resp.pg_size = PAGE_SIZE;
3840 	resp.cqe_sz = sizeof(struct cq_base);
3841 	resp.max_cqd = dev_attr->max_cq_wqes;
3842 	resp.rsvd    = 0;
3843 
3844 	rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
3845 	if (rc) {
3846 		ibdev_err(ibdev, "Failed to copy user context");
3847 		rc = -EFAULT;
3848 		goto cfail;
3849 	}
3850 
3851 	return 0;
3852 cfail:
3853 	free_page((unsigned long)uctx->shpg);
3854 	uctx->shpg = NULL;
3855 fail:
3856 	return rc;
3857 }
3858 
3859 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
3860 {
3861 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3862 						   struct bnxt_re_ucontext,
3863 						   ib_uctx);
3864 
3865 	struct bnxt_re_dev *rdev = uctx->rdev;
3866 
3867 	if (uctx->shpg)
3868 		free_page((unsigned long)uctx->shpg);
3869 
3870 	if (uctx->dpi.dbr) {
3871 		/* Free DPI only if this is the first PD allocated by the
3872 		 * application and mark the context dpi as NULL
3873 		 */
3874 		bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
3875 				       &rdev->qplib_res.dpi_tbl, &uctx->dpi);
3876 		uctx->dpi.dbr = NULL;
3877 	}
3878 }
3879 
3880 /* Helper function to mmap the virtual memory from user app */
3881 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
3882 {
3883 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3884 						   struct bnxt_re_ucontext,
3885 						   ib_uctx);
3886 	struct bnxt_re_dev *rdev = uctx->rdev;
3887 	u64 pfn;
3888 
3889 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3890 		return -EINVAL;
3891 
3892 	if (vma->vm_pgoff) {
3893 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3894 		if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
3895 				       PAGE_SIZE, vma->vm_page_prot)) {
3896 			ibdev_err(&rdev->ibdev, "Failed to map DPI");
3897 			return -EAGAIN;
3898 		}
3899 	} else {
3900 		pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT;
3901 		if (remap_pfn_range(vma, vma->vm_start,
3902 				    pfn, PAGE_SIZE, vma->vm_page_prot)) {
3903 			ibdev_err(&rdev->ibdev, "Failed to map shared page");
3904 			return -EAGAIN;
3905 		}
3906 	}
3907 
3908 	return 0;
3909 }
3910