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 	if (dev_attr->is_atomic) {
167 		ib_attr->atomic_cap = IB_ATOMIC_GLOB;
168 		ib_attr->masked_atomic_cap = IB_ATOMIC_GLOB;
169 	}
170 
171 	ib_attr->max_ee_rd_atom = 0;
172 	ib_attr->max_res_rd_atom = 0;
173 	ib_attr->max_ee_init_rd_atom = 0;
174 	ib_attr->max_ee = 0;
175 	ib_attr->max_rdd = 0;
176 	ib_attr->max_mw = dev_attr->max_mw;
177 	ib_attr->max_raw_ipv6_qp = 0;
178 	ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp;
179 	ib_attr->max_mcast_grp = 0;
180 	ib_attr->max_mcast_qp_attach = 0;
181 	ib_attr->max_total_mcast_qp_attach = 0;
182 	ib_attr->max_ah = dev_attr->max_ah;
183 
184 	ib_attr->max_srq = dev_attr->max_srq;
185 	ib_attr->max_srq_wr = dev_attr->max_srq_wqes;
186 	ib_attr->max_srq_sge = dev_attr->max_srq_sges;
187 
188 	ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS;
189 
190 	ib_attr->max_pkeys = 1;
191 	ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY;
192 	return 0;
193 }
194 
195 /* Port */
196 int bnxt_re_query_port(struct ib_device *ibdev, u32 port_num,
197 		       struct ib_port_attr *port_attr)
198 {
199 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
200 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
201 
202 	memset(port_attr, 0, sizeof(*port_attr));
203 
204 	if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) {
205 		port_attr->state = IB_PORT_ACTIVE;
206 		port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
207 	} else {
208 		port_attr->state = IB_PORT_DOWN;
209 		port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
210 	}
211 	port_attr->max_mtu = IB_MTU_4096;
212 	port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu);
213 	port_attr->gid_tbl_len = dev_attr->max_sgid;
214 	port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
215 				    IB_PORT_DEVICE_MGMT_SUP |
216 				    IB_PORT_VENDOR_CLASS_SUP;
217 	port_attr->ip_gids = true;
218 
219 	port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW;
220 	port_attr->bad_pkey_cntr = 0;
221 	port_attr->qkey_viol_cntr = 0;
222 	port_attr->pkey_tbl_len = dev_attr->max_pkey;
223 	port_attr->lid = 0;
224 	port_attr->sm_lid = 0;
225 	port_attr->lmc = 0;
226 	port_attr->max_vl_num = 4;
227 	port_attr->sm_sl = 0;
228 	port_attr->subnet_timeout = 0;
229 	port_attr->init_type_reply = 0;
230 	port_attr->active_speed = rdev->active_speed;
231 	port_attr->active_width = rdev->active_width;
232 
233 	return 0;
234 }
235 
236 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u32 port_num,
237 			       struct ib_port_immutable *immutable)
238 {
239 	struct ib_port_attr port_attr;
240 
241 	if (bnxt_re_query_port(ibdev, port_num, &port_attr))
242 		return -EINVAL;
243 
244 	immutable->pkey_tbl_len = port_attr.pkey_tbl_len;
245 	immutable->gid_tbl_len = port_attr.gid_tbl_len;
246 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
247 	immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP;
248 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
249 	return 0;
250 }
251 
252 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str)
253 {
254 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
255 
256 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d",
257 		 rdev->dev_attr.fw_ver[0], rdev->dev_attr.fw_ver[1],
258 		 rdev->dev_attr.fw_ver[2], rdev->dev_attr.fw_ver[3]);
259 }
260 
261 int bnxt_re_query_pkey(struct ib_device *ibdev, u32 port_num,
262 		       u16 index, u16 *pkey)
263 {
264 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
265 
266 	/* Ignore port_num */
267 
268 	memset(pkey, 0, sizeof(*pkey));
269 	return bnxt_qplib_get_pkey(&rdev->qplib_res,
270 				   &rdev->qplib_res.pkey_tbl, index, pkey);
271 }
272 
273 int bnxt_re_query_gid(struct ib_device *ibdev, u32 port_num,
274 		      int index, union ib_gid *gid)
275 {
276 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
277 	int rc = 0;
278 
279 	/* Ignore port_num */
280 	memset(gid, 0, sizeof(*gid));
281 	rc = bnxt_qplib_get_sgid(&rdev->qplib_res,
282 				 &rdev->qplib_res.sgid_tbl, index,
283 				 (struct bnxt_qplib_gid *)gid);
284 	return rc;
285 }
286 
287 int bnxt_re_del_gid(const struct ib_gid_attr *attr, void **context)
288 {
289 	int rc = 0;
290 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
291 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
292 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
293 	struct bnxt_qplib_gid *gid_to_del;
294 	u16 vlan_id = 0xFFFF;
295 
296 	/* Delete the entry from the hardware */
297 	ctx = *context;
298 	if (!ctx)
299 		return -EINVAL;
300 
301 	if (sgid_tbl && sgid_tbl->active) {
302 		if (ctx->idx >= sgid_tbl->max)
303 			return -EINVAL;
304 		gid_to_del = &sgid_tbl->tbl[ctx->idx].gid;
305 		vlan_id = sgid_tbl->tbl[ctx->idx].vlan_id;
306 		/* DEL_GID is called in WQ context(netdevice_event_work_handler)
307 		 * or via the ib_unregister_device path. In the former case QP1
308 		 * may not be destroyed yet, in which case just return as FW
309 		 * needs that entry to be present and will fail it's deletion.
310 		 * We could get invoked again after QP1 is destroyed OR get an
311 		 * ADD_GID call with a different GID value for the same index
312 		 * where we issue MODIFY_GID cmd to update the GID entry -- TBD
313 		 */
314 		if (ctx->idx == 0 &&
315 		    rdma_link_local_addr((struct in6_addr *)gid_to_del) &&
316 		    ctx->refcnt == 1 && rdev->gsi_ctx.gsi_sqp) {
317 			ibdev_dbg(&rdev->ibdev,
318 				  "Trying to delete GID0 while QP1 is alive\n");
319 			return -EFAULT;
320 		}
321 		ctx->refcnt--;
322 		if (!ctx->refcnt) {
323 			rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del,
324 						 vlan_id,  true);
325 			if (rc) {
326 				ibdev_err(&rdev->ibdev,
327 					  "Failed to remove GID: %#x", rc);
328 			} else {
329 				ctx_tbl = sgid_tbl->ctx;
330 				ctx_tbl[ctx->idx] = NULL;
331 				kfree(ctx);
332 			}
333 		}
334 	} else {
335 		return -EINVAL;
336 	}
337 	return rc;
338 }
339 
340 int bnxt_re_add_gid(const struct ib_gid_attr *attr, void **context)
341 {
342 	int rc;
343 	u32 tbl_idx = 0;
344 	u16 vlan_id = 0xFFFF;
345 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
346 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
347 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
348 
349 	rc = rdma_read_gid_l2_fields(attr, &vlan_id, NULL);
350 	if (rc)
351 		return rc;
352 
353 	rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)&attr->gid,
354 				 rdev->qplib_res.netdev->dev_addr,
355 				 vlan_id, true, &tbl_idx);
356 	if (rc == -EALREADY) {
357 		ctx_tbl = sgid_tbl->ctx;
358 		ctx_tbl[tbl_idx]->refcnt++;
359 		*context = ctx_tbl[tbl_idx];
360 		return 0;
361 	}
362 
363 	if (rc < 0) {
364 		ibdev_err(&rdev->ibdev, "Failed to add GID: %#x", rc);
365 		return rc;
366 	}
367 
368 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
369 	if (!ctx)
370 		return -ENOMEM;
371 	ctx_tbl = sgid_tbl->ctx;
372 	ctx->idx = tbl_idx;
373 	ctx->refcnt = 1;
374 	ctx_tbl[tbl_idx] = ctx;
375 	*context = ctx;
376 
377 	return rc;
378 }
379 
380 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
381 					    u32 port_num)
382 {
383 	return IB_LINK_LAYER_ETHERNET;
384 }
385 
386 #define	BNXT_RE_FENCE_PBL_SIZE	DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
387 
388 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd)
389 {
390 	struct bnxt_re_fence_data *fence = &pd->fence;
391 	struct ib_mr *ib_mr = &fence->mr->ib_mr;
392 	struct bnxt_qplib_swqe *wqe = &fence->bind_wqe;
393 
394 	memset(wqe, 0, sizeof(*wqe));
395 	wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW;
396 	wqe->wr_id = BNXT_QPLIB_FENCE_WRID;
397 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
398 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
399 	wqe->bind.zero_based = false;
400 	wqe->bind.parent_l_key = ib_mr->lkey;
401 	wqe->bind.va = (u64)(unsigned long)fence->va;
402 	wqe->bind.length = fence->size;
403 	wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ);
404 	wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1;
405 
406 	/* Save the initial rkey in fence structure for now;
407 	 * wqe->bind.r_key will be set at (re)bind time.
408 	 */
409 	fence->bind_rkey = ib_inc_rkey(fence->mw->rkey);
410 }
411 
412 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp)
413 {
414 	struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp,
415 					     qplib_qp);
416 	struct ib_pd *ib_pd = qp->ib_qp.pd;
417 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
418 	struct bnxt_re_fence_data *fence = &pd->fence;
419 	struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe;
420 	struct bnxt_qplib_swqe wqe;
421 	int rc;
422 
423 	memcpy(&wqe, fence_wqe, sizeof(wqe));
424 	wqe.bind.r_key = fence->bind_rkey;
425 	fence->bind_rkey = ib_inc_rkey(fence->bind_rkey);
426 
427 	ibdev_dbg(&qp->rdev->ibdev,
428 		  "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
429 		wqe.bind.r_key, qp->qplib_qp.id, pd);
430 	rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
431 	if (rc) {
432 		ibdev_err(&qp->rdev->ibdev, "Failed to bind fence-WQE\n");
433 		return rc;
434 	}
435 	bnxt_qplib_post_send_db(&qp->qplib_qp);
436 
437 	return rc;
438 }
439 
440 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd)
441 {
442 	struct bnxt_re_fence_data *fence = &pd->fence;
443 	struct bnxt_re_dev *rdev = pd->rdev;
444 	struct device *dev = &rdev->en_dev->pdev->dev;
445 	struct bnxt_re_mr *mr = fence->mr;
446 
447 	if (fence->mw) {
448 		bnxt_re_dealloc_mw(fence->mw);
449 		fence->mw = NULL;
450 	}
451 	if (mr) {
452 		if (mr->ib_mr.rkey)
453 			bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr,
454 					     true);
455 		if (mr->ib_mr.lkey)
456 			bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
457 		kfree(mr);
458 		fence->mr = NULL;
459 	}
460 	if (fence->dma_addr) {
461 		dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES,
462 				 DMA_BIDIRECTIONAL);
463 		fence->dma_addr = 0;
464 	}
465 }
466 
467 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
468 {
469 	int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND;
470 	struct bnxt_re_fence_data *fence = &pd->fence;
471 	struct bnxt_re_dev *rdev = pd->rdev;
472 	struct device *dev = &rdev->en_dev->pdev->dev;
473 	struct bnxt_re_mr *mr = NULL;
474 	dma_addr_t dma_addr = 0;
475 	struct ib_mw *mw;
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 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL,
511 			       BNXT_RE_FENCE_PBL_SIZE, PAGE_SIZE);
512 	if (rc) {
513 		ibdev_err(&rdev->ibdev, "Failed to register fence-MR\n");
514 		goto fail;
515 	}
516 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
517 
518 	/* Create a fence MW only for kernel consumers */
519 	mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL);
520 	if (IS_ERR(mw)) {
521 		ibdev_err(&rdev->ibdev,
522 			  "Failed to create fence-MW for PD: %p\n", pd);
523 		rc = PTR_ERR(mw);
524 		goto fail;
525 	}
526 	fence->mw = mw;
527 
528 	bnxt_re_create_fence_wqe(pd);
529 	return 0;
530 
531 fail:
532 	bnxt_re_destroy_fence_mr(pd);
533 	return rc;
534 }
535 
536 /* Protection Domains */
537 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
538 {
539 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
540 	struct bnxt_re_dev *rdev = pd->rdev;
541 
542 	bnxt_re_destroy_fence_mr(pd);
543 
544 	if (pd->qplib_pd.id)
545 		bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
546 				      &pd->qplib_pd);
547 	return 0;
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 int 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 	return 0;
615 }
616 
617 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)
618 {
619 	u8 nw_type;
620 
621 	switch (ntype) {
622 	case RDMA_NETWORK_IPV4:
623 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4;
624 		break;
625 	case RDMA_NETWORK_IPV6:
626 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6;
627 		break;
628 	default:
629 		nw_type = CMDQ_CREATE_AH_TYPE_V1;
630 		break;
631 	}
632 	return nw_type;
633 }
634 
635 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr,
636 		      struct ib_udata *udata)
637 {
638 	struct ib_pd *ib_pd = ib_ah->pd;
639 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
640 	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
641 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
642 	struct bnxt_re_dev *rdev = pd->rdev;
643 	const struct ib_gid_attr *sgid_attr;
644 	struct bnxt_re_gid_ctx *ctx;
645 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
646 	u8 nw_type;
647 	int rc;
648 
649 	if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) {
650 		ibdev_err(&rdev->ibdev, "Failed to alloc AH: GRH not set");
651 		return -EINVAL;
652 	}
653 
654 	ah->rdev = rdev;
655 	ah->qplib_ah.pd = &pd->qplib_pd;
656 
657 	/* Supply the configuration for the HW */
658 	memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw,
659 	       sizeof(union ib_gid));
660 	sgid_attr = grh->sgid_attr;
661 	/* Get the HW context of the GID. The reference
662 	 * of GID table entry is already taken by the caller.
663 	 */
664 	ctx = rdma_read_gid_hw_context(sgid_attr);
665 	ah->qplib_ah.sgid_index = ctx->idx;
666 	ah->qplib_ah.host_sgid_index = grh->sgid_index;
667 	ah->qplib_ah.traffic_class = grh->traffic_class;
668 	ah->qplib_ah.flow_label = grh->flow_label;
669 	ah->qplib_ah.hop_limit = grh->hop_limit;
670 	ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr);
671 
672 	/* Get network header type for this GID */
673 	nw_type = rdma_gid_attr_network_type(sgid_attr);
674 	ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type);
675 
676 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
677 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah,
678 				  !(init_attr->flags &
679 				    RDMA_CREATE_AH_SLEEPABLE));
680 	if (rc) {
681 		ibdev_err(&rdev->ibdev, "Failed to allocate HW AH");
682 		return rc;
683 	}
684 
685 	/* Write AVID to shared page. */
686 	if (udata) {
687 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
688 			udata, struct bnxt_re_ucontext, ib_uctx);
689 		unsigned long flag;
690 		u32 *wrptr;
691 
692 		spin_lock_irqsave(&uctx->sh_lock, flag);
693 		wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT);
694 		*wrptr = ah->qplib_ah.id;
695 		wmb(); /* make sure cache is updated. */
696 		spin_unlock_irqrestore(&uctx->sh_lock, flag);
697 	}
698 
699 	return 0;
700 }
701 
702 int bnxt_re_modify_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
703 {
704 	return 0;
705 }
706 
707 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
708 {
709 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
710 
711 	ah_attr->type = ib_ah->type;
712 	rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl);
713 	memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN);
714 	rdma_ah_set_grh(ah_attr, NULL, 0,
715 			ah->qplib_ah.host_sgid_index,
716 			0, ah->qplib_ah.traffic_class);
717 	rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data);
718 	rdma_ah_set_port_num(ah_attr, 1);
719 	rdma_ah_set_static_rate(ah_attr, 0);
720 	return 0;
721 }
722 
723 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp)
724 	__acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock)
725 {
726 	unsigned long flags;
727 
728 	spin_lock_irqsave(&qp->scq->cq_lock, flags);
729 	if (qp->rcq != qp->scq)
730 		spin_lock(&qp->rcq->cq_lock);
731 	else
732 		__acquire(&qp->rcq->cq_lock);
733 
734 	return flags;
735 }
736 
737 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp,
738 			unsigned long flags)
739 	__releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock)
740 {
741 	if (qp->rcq != qp->scq)
742 		spin_unlock(&qp->rcq->cq_lock);
743 	else
744 		__release(&qp->rcq->cq_lock);
745 	spin_unlock_irqrestore(&qp->scq->cq_lock, flags);
746 }
747 
748 static int bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp *qp)
749 {
750 	struct bnxt_re_qp *gsi_sqp;
751 	struct bnxt_re_ah *gsi_sah;
752 	struct bnxt_re_dev *rdev;
753 	int rc = 0;
754 
755 	rdev = qp->rdev;
756 	gsi_sqp = rdev->gsi_ctx.gsi_sqp;
757 	gsi_sah = rdev->gsi_ctx.gsi_sah;
758 
759 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow AH\n");
760 	bnxt_qplib_destroy_ah(&rdev->qplib_res,
761 			      &gsi_sah->qplib_ah,
762 			      true);
763 	bnxt_qplib_clean_qp(&qp->qplib_qp);
764 
765 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow QP\n");
766 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &gsi_sqp->qplib_qp);
767 	if (rc) {
768 		ibdev_err(&rdev->ibdev, "Destroy Shadow QP failed");
769 		goto fail;
770 	}
771 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &gsi_sqp->qplib_qp);
772 
773 	/* remove from active qp list */
774 	mutex_lock(&rdev->qp_lock);
775 	list_del(&gsi_sqp->list);
776 	mutex_unlock(&rdev->qp_lock);
777 	atomic_dec(&rdev->qp_count);
778 
779 	kfree(rdev->gsi_ctx.sqp_tbl);
780 	kfree(gsi_sah);
781 	kfree(gsi_sqp);
782 	rdev->gsi_ctx.gsi_sqp = NULL;
783 	rdev->gsi_ctx.gsi_sah = NULL;
784 	rdev->gsi_ctx.sqp_tbl = NULL;
785 
786 	return 0;
787 fail:
788 	return rc;
789 }
790 
791 /* Queue Pairs */
792 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
793 {
794 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
795 	struct bnxt_re_dev *rdev = qp->rdev;
796 	unsigned int flags;
797 	int rc;
798 
799 	bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
800 
801 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
802 	if (rc) {
803 		ibdev_err(&rdev->ibdev, "Failed to destroy HW QP");
804 		return rc;
805 	}
806 
807 	if (rdma_is_kernel_res(&qp->ib_qp.res)) {
808 		flags = bnxt_re_lock_cqs(qp);
809 		bnxt_qplib_clean_qp(&qp->qplib_qp);
810 		bnxt_re_unlock_cqs(qp, flags);
811 	}
812 
813 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
814 
815 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) {
816 		rc = bnxt_re_destroy_gsi_sqp(qp);
817 		if (rc)
818 			return rc;
819 	}
820 
821 	mutex_lock(&rdev->qp_lock);
822 	list_del(&qp->list);
823 	mutex_unlock(&rdev->qp_lock);
824 	atomic_dec(&rdev->qp_count);
825 
826 	ib_umem_release(qp->rumem);
827 	ib_umem_release(qp->sumem);
828 
829 	return 0;
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 u16 bnxt_re_setup_rwqe_size(struct bnxt_qplib_qp *qplqp,
847 				   int rsge, int max)
848 {
849 	if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
850 		rsge = max;
851 	return bnxt_re_get_rwqe_size(rsge);
852 }
853 
854 static u16 bnxt_re_get_wqe_size(int ilsize, int nsge)
855 {
856 	u16 wqe_size, calc_ils;
857 
858 	wqe_size = bnxt_re_get_swqe_size(nsge);
859 	if (ilsize) {
860 		calc_ils = sizeof(struct sq_send_hdr) + ilsize;
861 		wqe_size = max_t(u16, calc_ils, wqe_size);
862 		wqe_size = ALIGN(wqe_size, sizeof(struct sq_send_hdr));
863 	}
864 	return wqe_size;
865 }
866 
867 static int bnxt_re_setup_swqe_size(struct bnxt_re_qp *qp,
868 				   struct ib_qp_init_attr *init_attr)
869 {
870 	struct bnxt_qplib_dev_attr *dev_attr;
871 	struct bnxt_qplib_qp *qplqp;
872 	struct bnxt_re_dev *rdev;
873 	struct bnxt_qplib_q *sq;
874 	int align, ilsize;
875 
876 	rdev = qp->rdev;
877 	qplqp = &qp->qplib_qp;
878 	sq = &qplqp->sq;
879 	dev_attr = &rdev->dev_attr;
880 
881 	align = sizeof(struct sq_send_hdr);
882 	ilsize = ALIGN(init_attr->cap.max_inline_data, align);
883 
884 	sq->wqe_size = bnxt_re_get_wqe_size(ilsize, sq->max_sge);
885 	if (sq->wqe_size > bnxt_re_get_swqe_size(dev_attr->max_qp_sges))
886 		return -EINVAL;
887 	/* For gen p4 and gen p5 backward compatibility mode
888 	 * wqe size is fixed to 128 bytes
889 	 */
890 	if (sq->wqe_size < bnxt_re_get_swqe_size(dev_attr->max_qp_sges) &&
891 			qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
892 		sq->wqe_size = bnxt_re_get_swqe_size(dev_attr->max_qp_sges);
893 
894 	if (init_attr->cap.max_inline_data) {
895 		qplqp->max_inline_data = sq->wqe_size -
896 			sizeof(struct sq_send_hdr);
897 		init_attr->cap.max_inline_data = qplqp->max_inline_data;
898 		if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
899 			sq->max_sge = qplqp->max_inline_data /
900 				sizeof(struct sq_sge);
901 	}
902 
903 	return 0;
904 }
905 
906 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
907 				struct bnxt_re_qp *qp, struct ib_udata *udata)
908 {
909 	struct bnxt_qplib_qp *qplib_qp;
910 	struct bnxt_re_ucontext *cntx;
911 	struct bnxt_re_qp_req ureq;
912 	int bytes = 0, psn_sz;
913 	struct ib_umem *umem;
914 	int psn_nume;
915 
916 	qplib_qp = &qp->qplib_qp;
917 	cntx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext,
918 					 ib_uctx);
919 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
920 		return -EFAULT;
921 
922 	bytes = (qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size);
923 	/* Consider mapping PSN search memory only for RC QPs. */
924 	if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) {
925 		psn_sz = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
926 						   sizeof(struct sq_psn_search_ext) :
927 						   sizeof(struct sq_psn_search);
928 		psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ?
929 			    qplib_qp->sq.max_wqe :
930 			    ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) /
931 			      sizeof(struct bnxt_qplib_sge));
932 		bytes += (psn_nume * psn_sz);
933 	}
934 
935 	bytes = PAGE_ALIGN(bytes);
936 	umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes,
937 			   IB_ACCESS_LOCAL_WRITE);
938 	if (IS_ERR(umem))
939 		return PTR_ERR(umem);
940 
941 	qp->sumem = umem;
942 	qplib_qp->sq.sg_info.umem = umem;
943 	qplib_qp->sq.sg_info.pgsize = PAGE_SIZE;
944 	qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT;
945 	qplib_qp->qp_handle = ureq.qp_handle;
946 
947 	if (!qp->qplib_qp.srq) {
948 		bytes = (qplib_qp->rq.max_wqe * qplib_qp->rq.wqe_size);
949 		bytes = PAGE_ALIGN(bytes);
950 		umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes,
951 				   IB_ACCESS_LOCAL_WRITE);
952 		if (IS_ERR(umem))
953 			goto rqfail;
954 		qp->rumem = umem;
955 		qplib_qp->rq.sg_info.umem = umem;
956 		qplib_qp->rq.sg_info.pgsize = PAGE_SIZE;
957 		qplib_qp->rq.sg_info.pgshft = PAGE_SHIFT;
958 	}
959 
960 	qplib_qp->dpi = &cntx->dpi;
961 	return 0;
962 rqfail:
963 	ib_umem_release(qp->sumem);
964 	qp->sumem = NULL;
965 	memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info));
966 
967 	return PTR_ERR(umem);
968 }
969 
970 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
971 				(struct bnxt_re_pd *pd,
972 				 struct bnxt_qplib_res *qp1_res,
973 				 struct bnxt_qplib_qp *qp1_qp)
974 {
975 	struct bnxt_re_dev *rdev = pd->rdev;
976 	struct bnxt_re_ah *ah;
977 	union ib_gid sgid;
978 	int rc;
979 
980 	ah = kzalloc(sizeof(*ah), GFP_KERNEL);
981 	if (!ah)
982 		return NULL;
983 
984 	ah->rdev = rdev;
985 	ah->qplib_ah.pd = &pd->qplib_pd;
986 
987 	rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid);
988 	if (rc)
989 		goto fail;
990 
991 	/* supply the dgid data same as sgid */
992 	memcpy(ah->qplib_ah.dgid.data, &sgid.raw,
993 	       sizeof(union ib_gid));
994 	ah->qplib_ah.sgid_index = 0;
995 
996 	ah->qplib_ah.traffic_class = 0;
997 	ah->qplib_ah.flow_label = 0;
998 	ah->qplib_ah.hop_limit = 1;
999 	ah->qplib_ah.sl = 0;
1000 	/* Have DMAC same as SMAC */
1001 	ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr);
1002 
1003 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false);
1004 	if (rc) {
1005 		ibdev_err(&rdev->ibdev,
1006 			  "Failed to allocate HW AH for Shadow QP");
1007 		goto fail;
1008 	}
1009 
1010 	return ah;
1011 
1012 fail:
1013 	kfree(ah);
1014 	return NULL;
1015 }
1016 
1017 static struct bnxt_re_qp *bnxt_re_create_shadow_qp
1018 				(struct bnxt_re_pd *pd,
1019 				 struct bnxt_qplib_res *qp1_res,
1020 				 struct bnxt_qplib_qp *qp1_qp)
1021 {
1022 	struct bnxt_re_dev *rdev = pd->rdev;
1023 	struct bnxt_re_qp *qp;
1024 	int rc;
1025 
1026 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1027 	if (!qp)
1028 		return NULL;
1029 
1030 	qp->rdev = rdev;
1031 
1032 	/* Initialize the shadow QP structure from the QP1 values */
1033 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
1034 
1035 	qp->qplib_qp.pd = &pd->qplib_pd;
1036 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
1037 	qp->qplib_qp.type = IB_QPT_UD;
1038 
1039 	qp->qplib_qp.max_inline_data = 0;
1040 	qp->qplib_qp.sig_type = true;
1041 
1042 	/* Shadow QP SQ depth should be same as QP1 RQ depth */
1043 	qp->qplib_qp.sq.wqe_size = bnxt_re_get_wqe_size(0, 6);
1044 	qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
1045 	qp->qplib_qp.sq.max_sge = 2;
1046 	/* Q full delta can be 1 since it is internal QP */
1047 	qp->qplib_qp.sq.q_full_delta = 1;
1048 	qp->qplib_qp.sq.sg_info.pgsize = PAGE_SIZE;
1049 	qp->qplib_qp.sq.sg_info.pgshft = PAGE_SHIFT;
1050 
1051 	qp->qplib_qp.scq = qp1_qp->scq;
1052 	qp->qplib_qp.rcq = qp1_qp->rcq;
1053 
1054 	qp->qplib_qp.rq.wqe_size = bnxt_re_get_rwqe_size(6);
1055 	qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
1056 	qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
1057 	/* Q full delta can be 1 since it is internal QP */
1058 	qp->qplib_qp.rq.q_full_delta = 1;
1059 	qp->qplib_qp.rq.sg_info.pgsize = PAGE_SIZE;
1060 	qp->qplib_qp.rq.sg_info.pgshft = PAGE_SHIFT;
1061 
1062 	qp->qplib_qp.mtu = qp1_qp->mtu;
1063 
1064 	qp->qplib_qp.sq_hdr_buf_size = 0;
1065 	qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
1066 	qp->qplib_qp.dpi = &rdev->dpi_privileged;
1067 
1068 	rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
1069 	if (rc)
1070 		goto fail;
1071 
1072 	spin_lock_init(&qp->sq_lock);
1073 	INIT_LIST_HEAD(&qp->list);
1074 	mutex_lock(&rdev->qp_lock);
1075 	list_add_tail(&qp->list, &rdev->qp_list);
1076 	atomic_inc(&rdev->qp_count);
1077 	mutex_unlock(&rdev->qp_lock);
1078 	return qp;
1079 fail:
1080 	kfree(qp);
1081 	return NULL;
1082 }
1083 
1084 static int bnxt_re_init_rq_attr(struct bnxt_re_qp *qp,
1085 				struct ib_qp_init_attr *init_attr)
1086 {
1087 	struct bnxt_qplib_dev_attr *dev_attr;
1088 	struct bnxt_qplib_qp *qplqp;
1089 	struct bnxt_re_dev *rdev;
1090 	struct bnxt_qplib_q *rq;
1091 	int entries;
1092 
1093 	rdev = qp->rdev;
1094 	qplqp = &qp->qplib_qp;
1095 	rq = &qplqp->rq;
1096 	dev_attr = &rdev->dev_attr;
1097 
1098 	if (init_attr->srq) {
1099 		struct bnxt_re_srq *srq;
1100 
1101 		srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq);
1102 		qplqp->srq = &srq->qplib_srq;
1103 		rq->max_wqe = 0;
1104 	} else {
1105 		rq->max_sge = init_attr->cap.max_recv_sge;
1106 		if (rq->max_sge > dev_attr->max_qp_sges)
1107 			rq->max_sge = dev_attr->max_qp_sges;
1108 		init_attr->cap.max_recv_sge = rq->max_sge;
1109 		rq->wqe_size = bnxt_re_setup_rwqe_size(qplqp, rq->max_sge,
1110 						       dev_attr->max_qp_sges);
1111 		/* Allocate 1 more than what's provided so posting max doesn't
1112 		 * mean empty.
1113 		 */
1114 		entries = roundup_pow_of_two(init_attr->cap.max_recv_wr + 1);
1115 		rq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1116 		rq->q_full_delta = 0;
1117 		rq->sg_info.pgsize = PAGE_SIZE;
1118 		rq->sg_info.pgshft = PAGE_SHIFT;
1119 	}
1120 
1121 	return 0;
1122 }
1123 
1124 static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp)
1125 {
1126 	struct bnxt_qplib_dev_attr *dev_attr;
1127 	struct bnxt_qplib_qp *qplqp;
1128 	struct bnxt_re_dev *rdev;
1129 
1130 	rdev = qp->rdev;
1131 	qplqp = &qp->qplib_qp;
1132 	dev_attr = &rdev->dev_attr;
1133 
1134 	if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) {
1135 		qplqp->rq.max_sge = dev_attr->max_qp_sges;
1136 		if (qplqp->rq.max_sge > dev_attr->max_qp_sges)
1137 			qplqp->rq.max_sge = dev_attr->max_qp_sges;
1138 		qplqp->rq.max_sge = 6;
1139 	}
1140 }
1141 
1142 static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp,
1143 				struct ib_qp_init_attr *init_attr,
1144 				struct ib_udata *udata)
1145 {
1146 	struct bnxt_qplib_dev_attr *dev_attr;
1147 	struct bnxt_qplib_qp *qplqp;
1148 	struct bnxt_re_dev *rdev;
1149 	struct bnxt_qplib_q *sq;
1150 	int entries;
1151 	int diff;
1152 	int rc;
1153 
1154 	rdev = qp->rdev;
1155 	qplqp = &qp->qplib_qp;
1156 	sq = &qplqp->sq;
1157 	dev_attr = &rdev->dev_attr;
1158 
1159 	sq->max_sge = init_attr->cap.max_send_sge;
1160 	if (sq->max_sge > dev_attr->max_qp_sges) {
1161 		sq->max_sge = dev_attr->max_qp_sges;
1162 		init_attr->cap.max_send_sge = sq->max_sge;
1163 	}
1164 
1165 	rc = bnxt_re_setup_swqe_size(qp, init_attr);
1166 	if (rc)
1167 		return rc;
1168 
1169 	entries = init_attr->cap.max_send_wr;
1170 	/* Allocate 128 + 1 more than what's provided */
1171 	diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ?
1172 		0 : BNXT_QPLIB_RESERVED_QP_WRS;
1173 	entries = roundup_pow_of_two(entries + diff + 1);
1174 	sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1);
1175 	sq->q_full_delta = diff + 1;
1176 	/*
1177 	 * Reserving one slot for Phantom WQE. Application can
1178 	 * post one extra entry in this case. But allowing this to avoid
1179 	 * unexpected Queue full condition
1180 	 */
1181 	qplqp->sq.q_full_delta -= 1;
1182 	qplqp->sq.sg_info.pgsize = PAGE_SIZE;
1183 	qplqp->sq.sg_info.pgshft = PAGE_SHIFT;
1184 
1185 	return 0;
1186 }
1187 
1188 static void bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp *qp,
1189 				       struct ib_qp_init_attr *init_attr)
1190 {
1191 	struct bnxt_qplib_dev_attr *dev_attr;
1192 	struct bnxt_qplib_qp *qplqp;
1193 	struct bnxt_re_dev *rdev;
1194 	int entries;
1195 
1196 	rdev = qp->rdev;
1197 	qplqp = &qp->qplib_qp;
1198 	dev_attr = &rdev->dev_attr;
1199 
1200 	if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) {
1201 		entries = roundup_pow_of_two(init_attr->cap.max_send_wr + 1);
1202 		qplqp->sq.max_wqe = min_t(u32, entries,
1203 					  dev_attr->max_qp_wqes + 1);
1204 		qplqp->sq.q_full_delta = qplqp->sq.max_wqe -
1205 			init_attr->cap.max_send_wr;
1206 		qplqp->sq.max_sge++; /* Need one extra sge to put UD header */
1207 		if (qplqp->sq.max_sge > dev_attr->max_qp_sges)
1208 			qplqp->sq.max_sge = dev_attr->max_qp_sges;
1209 	}
1210 }
1211 
1212 static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev,
1213 				struct ib_qp_init_attr *init_attr)
1214 {
1215 	struct bnxt_qplib_chip_ctx *chip_ctx;
1216 	int qptype;
1217 
1218 	chip_ctx = rdev->chip_ctx;
1219 
1220 	qptype = __from_ib_qp_type(init_attr->qp_type);
1221 	if (qptype == IB_QPT_MAX) {
1222 		ibdev_err(&rdev->ibdev, "QP type 0x%x not supported", qptype);
1223 		qptype = -EOPNOTSUPP;
1224 		goto out;
1225 	}
1226 
1227 	if (bnxt_qplib_is_chip_gen_p5(chip_ctx) &&
1228 	    init_attr->qp_type == IB_QPT_GSI)
1229 		qptype = CMDQ_CREATE_QP_TYPE_GSI;
1230 out:
1231 	return qptype;
1232 }
1233 
1234 static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1235 				struct ib_qp_init_attr *init_attr,
1236 				struct ib_udata *udata)
1237 {
1238 	struct bnxt_qplib_dev_attr *dev_attr;
1239 	struct bnxt_qplib_qp *qplqp;
1240 	struct bnxt_re_dev *rdev;
1241 	struct bnxt_re_cq *cq;
1242 	int rc = 0, qptype;
1243 
1244 	rdev = qp->rdev;
1245 	qplqp = &qp->qplib_qp;
1246 	dev_attr = &rdev->dev_attr;
1247 
1248 	/* Setup misc params */
1249 	ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr);
1250 	qplqp->pd = &pd->qplib_pd;
1251 	qplqp->qp_handle = (u64)qplqp;
1252 	qplqp->max_inline_data = init_attr->cap.max_inline_data;
1253 	qplqp->sig_type = ((init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ?
1254 			    true : false);
1255 	qptype = bnxt_re_init_qp_type(rdev, init_attr);
1256 	if (qptype < 0) {
1257 		rc = qptype;
1258 		goto out;
1259 	}
1260 	qplqp->type = (u8)qptype;
1261 	qplqp->wqe_mode = rdev->chip_ctx->modes.wqe_mode;
1262 
1263 	if (init_attr->qp_type == IB_QPT_RC) {
1264 		qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom;
1265 		qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
1266 	}
1267 	qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1268 	qplqp->dpi = &rdev->dpi_privileged; /* Doorbell page */
1269 	if (init_attr->create_flags) {
1270 		ibdev_dbg(&rdev->ibdev,
1271 			  "QP create flags 0x%x not supported",
1272 			  init_attr->create_flags);
1273 		return -EOPNOTSUPP;
1274 	}
1275 
1276 	/* Setup CQs */
1277 	if (init_attr->send_cq) {
1278 		cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq);
1279 		qplqp->scq = &cq->qplib_cq;
1280 		qp->scq = cq;
1281 	}
1282 
1283 	if (init_attr->recv_cq) {
1284 		cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq);
1285 		qplqp->rcq = &cq->qplib_cq;
1286 		qp->rcq = cq;
1287 	}
1288 
1289 	/* Setup RQ/SRQ */
1290 	rc = bnxt_re_init_rq_attr(qp, init_attr);
1291 	if (rc)
1292 		goto out;
1293 	if (init_attr->qp_type == IB_QPT_GSI)
1294 		bnxt_re_adjust_gsi_rq_attr(qp);
1295 
1296 	/* Setup SQ */
1297 	rc = bnxt_re_init_sq_attr(qp, init_attr, udata);
1298 	if (rc)
1299 		goto out;
1300 	if (init_attr->qp_type == IB_QPT_GSI)
1301 		bnxt_re_adjust_gsi_sq_attr(qp, init_attr);
1302 
1303 	if (udata) /* This will update DPI and qp_handle */
1304 		rc = bnxt_re_init_user_qp(rdev, pd, qp, udata);
1305 out:
1306 	return rc;
1307 }
1308 
1309 static int bnxt_re_create_shadow_gsi(struct bnxt_re_qp *qp,
1310 				     struct bnxt_re_pd *pd)
1311 {
1312 	struct bnxt_re_sqp_entries *sqp_tbl;
1313 	struct bnxt_re_dev *rdev;
1314 	struct bnxt_re_qp *sqp;
1315 	struct bnxt_re_ah *sah;
1316 	int rc = 0;
1317 
1318 	rdev = qp->rdev;
1319 	/* Create a shadow QP to handle the QP1 traffic */
1320 	sqp_tbl = kcalloc(BNXT_RE_MAX_GSI_SQP_ENTRIES, sizeof(*sqp_tbl),
1321 			  GFP_KERNEL);
1322 	if (!sqp_tbl)
1323 		return -ENOMEM;
1324 	rdev->gsi_ctx.sqp_tbl = sqp_tbl;
1325 
1326 	sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res, &qp->qplib_qp);
1327 	if (!sqp) {
1328 		rc = -ENODEV;
1329 		ibdev_err(&rdev->ibdev, "Failed to create Shadow QP for QP1");
1330 		goto out;
1331 	}
1332 	rdev->gsi_ctx.gsi_sqp = sqp;
1333 
1334 	sqp->rcq = qp->rcq;
1335 	sqp->scq = qp->scq;
1336 	sah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res,
1337 					  &qp->qplib_qp);
1338 	if (!sah) {
1339 		bnxt_qplib_destroy_qp(&rdev->qplib_res,
1340 				      &sqp->qplib_qp);
1341 		rc = -ENODEV;
1342 		ibdev_err(&rdev->ibdev,
1343 			  "Failed to create AH entry for ShadowQP");
1344 		goto out;
1345 	}
1346 	rdev->gsi_ctx.gsi_sah = sah;
1347 
1348 	return 0;
1349 out:
1350 	kfree(sqp_tbl);
1351 	return rc;
1352 }
1353 
1354 static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1355 				 struct ib_qp_init_attr *init_attr)
1356 {
1357 	struct bnxt_re_dev *rdev;
1358 	struct bnxt_qplib_qp *qplqp;
1359 	int rc = 0;
1360 
1361 	rdev = qp->rdev;
1362 	qplqp = &qp->qplib_qp;
1363 
1364 	qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
1365 	qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
1366 
1367 	rc = bnxt_qplib_create_qp1(&rdev->qplib_res, qplqp);
1368 	if (rc) {
1369 		ibdev_err(&rdev->ibdev, "create HW QP1 failed!");
1370 		goto out;
1371 	}
1372 
1373 	rc = bnxt_re_create_shadow_gsi(qp, pd);
1374 out:
1375 	return rc;
1376 }
1377 
1378 static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev,
1379 				   struct ib_qp_init_attr *init_attr,
1380 				   struct bnxt_qplib_dev_attr *dev_attr)
1381 {
1382 	bool rc = true;
1383 
1384 	if (init_attr->cap.max_send_wr > dev_attr->max_qp_wqes ||
1385 	    init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes ||
1386 	    init_attr->cap.max_send_sge > dev_attr->max_qp_sges ||
1387 	    init_attr->cap.max_recv_sge > dev_attr->max_qp_sges ||
1388 	    init_attr->cap.max_inline_data > dev_attr->max_inline_data) {
1389 		ibdev_err(&rdev->ibdev,
1390 			  "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",
1391 			  init_attr->cap.max_send_wr, dev_attr->max_qp_wqes,
1392 			  init_attr->cap.max_recv_wr, dev_attr->max_qp_wqes,
1393 			  init_attr->cap.max_send_sge, dev_attr->max_qp_sges,
1394 			  init_attr->cap.max_recv_sge, dev_attr->max_qp_sges,
1395 			  init_attr->cap.max_inline_data,
1396 			  dev_attr->max_inline_data);
1397 		rc = false;
1398 	}
1399 	return rc;
1400 }
1401 
1402 int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr,
1403 		      struct ib_udata *udata)
1404 {
1405 	struct ib_pd *ib_pd = ib_qp->pd;
1406 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1407 	struct bnxt_re_dev *rdev = pd->rdev;
1408 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1409 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1410 	int rc;
1411 
1412 	rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr);
1413 	if (!rc) {
1414 		rc = -EINVAL;
1415 		goto fail;
1416 	}
1417 
1418 	qp->rdev = rdev;
1419 	rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, udata);
1420 	if (rc)
1421 		goto fail;
1422 
1423 	if (qp_init_attr->qp_type == IB_QPT_GSI &&
1424 	    !(bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx))) {
1425 		rc = bnxt_re_create_gsi_qp(qp, pd, qp_init_attr);
1426 		if (rc == -ENODEV)
1427 			goto qp_destroy;
1428 		if (rc)
1429 			goto fail;
1430 	} else {
1431 		rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
1432 		if (rc) {
1433 			ibdev_err(&rdev->ibdev, "Failed to create HW QP");
1434 			goto free_umem;
1435 		}
1436 		if (udata) {
1437 			struct bnxt_re_qp_resp resp;
1438 
1439 			resp.qpid = qp->qplib_qp.id;
1440 			resp.rsvd = 0;
1441 			rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1442 			if (rc) {
1443 				ibdev_err(&rdev->ibdev, "Failed to copy QP udata");
1444 				goto qp_destroy;
1445 			}
1446 		}
1447 	}
1448 
1449 	qp->ib_qp.qp_num = qp->qplib_qp.id;
1450 	if (qp_init_attr->qp_type == IB_QPT_GSI)
1451 		rdev->gsi_ctx.gsi_qp = qp;
1452 	spin_lock_init(&qp->sq_lock);
1453 	spin_lock_init(&qp->rq_lock);
1454 	INIT_LIST_HEAD(&qp->list);
1455 	mutex_lock(&rdev->qp_lock);
1456 	list_add_tail(&qp->list, &rdev->qp_list);
1457 	mutex_unlock(&rdev->qp_lock);
1458 	atomic_inc(&rdev->qp_count);
1459 
1460 	return 0;
1461 qp_destroy:
1462 	bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
1463 free_umem:
1464 	ib_umem_release(qp->rumem);
1465 	ib_umem_release(qp->sumem);
1466 fail:
1467 	return rc;
1468 }
1469 
1470 static u8 __from_ib_qp_state(enum ib_qp_state state)
1471 {
1472 	switch (state) {
1473 	case IB_QPS_RESET:
1474 		return CMDQ_MODIFY_QP_NEW_STATE_RESET;
1475 	case IB_QPS_INIT:
1476 		return CMDQ_MODIFY_QP_NEW_STATE_INIT;
1477 	case IB_QPS_RTR:
1478 		return CMDQ_MODIFY_QP_NEW_STATE_RTR;
1479 	case IB_QPS_RTS:
1480 		return CMDQ_MODIFY_QP_NEW_STATE_RTS;
1481 	case IB_QPS_SQD:
1482 		return CMDQ_MODIFY_QP_NEW_STATE_SQD;
1483 	case IB_QPS_SQE:
1484 		return CMDQ_MODIFY_QP_NEW_STATE_SQE;
1485 	case IB_QPS_ERR:
1486 	default:
1487 		return CMDQ_MODIFY_QP_NEW_STATE_ERR;
1488 	}
1489 }
1490 
1491 static enum ib_qp_state __to_ib_qp_state(u8 state)
1492 {
1493 	switch (state) {
1494 	case CMDQ_MODIFY_QP_NEW_STATE_RESET:
1495 		return IB_QPS_RESET;
1496 	case CMDQ_MODIFY_QP_NEW_STATE_INIT:
1497 		return IB_QPS_INIT;
1498 	case CMDQ_MODIFY_QP_NEW_STATE_RTR:
1499 		return IB_QPS_RTR;
1500 	case CMDQ_MODIFY_QP_NEW_STATE_RTS:
1501 		return IB_QPS_RTS;
1502 	case CMDQ_MODIFY_QP_NEW_STATE_SQD:
1503 		return IB_QPS_SQD;
1504 	case CMDQ_MODIFY_QP_NEW_STATE_SQE:
1505 		return IB_QPS_SQE;
1506 	case CMDQ_MODIFY_QP_NEW_STATE_ERR:
1507 	default:
1508 		return IB_QPS_ERR;
1509 	}
1510 }
1511 
1512 static u32 __from_ib_mtu(enum ib_mtu mtu)
1513 {
1514 	switch (mtu) {
1515 	case IB_MTU_256:
1516 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_256;
1517 	case IB_MTU_512:
1518 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_512;
1519 	case IB_MTU_1024:
1520 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024;
1521 	case IB_MTU_2048:
1522 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1523 	case IB_MTU_4096:
1524 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096;
1525 	default:
1526 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1527 	}
1528 }
1529 
1530 static enum ib_mtu __to_ib_mtu(u32 mtu)
1531 {
1532 	switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) {
1533 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_256:
1534 		return IB_MTU_256;
1535 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_512:
1536 		return IB_MTU_512;
1537 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024:
1538 		return IB_MTU_1024;
1539 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048:
1540 		return IB_MTU_2048;
1541 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096:
1542 		return IB_MTU_4096;
1543 	default:
1544 		return IB_MTU_2048;
1545 	}
1546 }
1547 
1548 /* Shared Receive Queues */
1549 int bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata)
1550 {
1551 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1552 					       ib_srq);
1553 	struct bnxt_re_dev *rdev = srq->rdev;
1554 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1555 	struct bnxt_qplib_nq *nq = NULL;
1556 
1557 	if (qplib_srq->cq)
1558 		nq = qplib_srq->cq->nq;
1559 	bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
1560 	ib_umem_release(srq->umem);
1561 	atomic_dec(&rdev->srq_count);
1562 	if (nq)
1563 		nq->budget--;
1564 	return 0;
1565 }
1566 
1567 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
1568 				 struct bnxt_re_pd *pd,
1569 				 struct bnxt_re_srq *srq,
1570 				 struct ib_udata *udata)
1571 {
1572 	struct bnxt_re_srq_req ureq;
1573 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1574 	struct ib_umem *umem;
1575 	int bytes = 0;
1576 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
1577 		udata, struct bnxt_re_ucontext, ib_uctx);
1578 
1579 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1580 		return -EFAULT;
1581 
1582 	bytes = (qplib_srq->max_wqe * qplib_srq->wqe_size);
1583 	bytes = PAGE_ALIGN(bytes);
1584 	umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes,
1585 			   IB_ACCESS_LOCAL_WRITE);
1586 	if (IS_ERR(umem))
1587 		return PTR_ERR(umem);
1588 
1589 	srq->umem = umem;
1590 	qplib_srq->sg_info.umem = umem;
1591 	qplib_srq->sg_info.pgsize = PAGE_SIZE;
1592 	qplib_srq->sg_info.pgshft = PAGE_SHIFT;
1593 	qplib_srq->srq_handle = ureq.srq_handle;
1594 	qplib_srq->dpi = &cntx->dpi;
1595 
1596 	return 0;
1597 }
1598 
1599 int bnxt_re_create_srq(struct ib_srq *ib_srq,
1600 		       struct ib_srq_init_attr *srq_init_attr,
1601 		       struct ib_udata *udata)
1602 {
1603 	struct bnxt_qplib_dev_attr *dev_attr;
1604 	struct bnxt_qplib_nq *nq = NULL;
1605 	struct bnxt_re_dev *rdev;
1606 	struct bnxt_re_srq *srq;
1607 	struct bnxt_re_pd *pd;
1608 	struct ib_pd *ib_pd;
1609 	int rc, entries;
1610 
1611 	ib_pd = ib_srq->pd;
1612 	pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1613 	rdev = pd->rdev;
1614 	dev_attr = &rdev->dev_attr;
1615 	srq = container_of(ib_srq, struct bnxt_re_srq, ib_srq);
1616 
1617 	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
1618 		ibdev_err(&rdev->ibdev, "Create CQ failed - max exceeded");
1619 		rc = -EINVAL;
1620 		goto exit;
1621 	}
1622 
1623 	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
1624 		rc = -EOPNOTSUPP;
1625 		goto exit;
1626 	}
1627 
1628 	srq->rdev = rdev;
1629 	srq->qplib_srq.pd = &pd->qplib_pd;
1630 	srq->qplib_srq.dpi = &rdev->dpi_privileged;
1631 	/* Allocate 1 more than what's provided so posting max doesn't
1632 	 * mean empty
1633 	 */
1634 	entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
1635 	if (entries > dev_attr->max_srq_wqes + 1)
1636 		entries = dev_attr->max_srq_wqes + 1;
1637 	srq->qplib_srq.max_wqe = entries;
1638 
1639 	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
1640 	 /* 128 byte wqe size for SRQ . So use max sges */
1641 	srq->qplib_srq.wqe_size = bnxt_re_get_rwqe_size(dev_attr->max_srq_sges);
1642 	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
1643 	srq->srq_limit = srq_init_attr->attr.srq_limit;
1644 	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
1645 	nq = &rdev->nq[0];
1646 
1647 	if (udata) {
1648 		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
1649 		if (rc)
1650 			goto fail;
1651 	}
1652 
1653 	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
1654 	if (rc) {
1655 		ibdev_err(&rdev->ibdev, "Create HW SRQ failed!");
1656 		goto fail;
1657 	}
1658 
1659 	if (udata) {
1660 		struct bnxt_re_srq_resp resp;
1661 
1662 		resp.srqid = srq->qplib_srq.id;
1663 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1664 		if (rc) {
1665 			ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!");
1666 			bnxt_qplib_destroy_srq(&rdev->qplib_res,
1667 					       &srq->qplib_srq);
1668 			goto fail;
1669 		}
1670 	}
1671 	if (nq)
1672 		nq->budget++;
1673 	atomic_inc(&rdev->srq_count);
1674 	spin_lock_init(&srq->lock);
1675 
1676 	return 0;
1677 
1678 fail:
1679 	ib_umem_release(srq->umem);
1680 exit:
1681 	return rc;
1682 }
1683 
1684 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
1685 		       enum ib_srq_attr_mask srq_attr_mask,
1686 		       struct ib_udata *udata)
1687 {
1688 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1689 					       ib_srq);
1690 	struct bnxt_re_dev *rdev = srq->rdev;
1691 	int rc;
1692 
1693 	switch (srq_attr_mask) {
1694 	case IB_SRQ_MAX_WR:
1695 		/* SRQ resize is not supported */
1696 		break;
1697 	case IB_SRQ_LIMIT:
1698 		/* Change the SRQ threshold */
1699 		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
1700 			return -EINVAL;
1701 
1702 		srq->qplib_srq.threshold = srq_attr->srq_limit;
1703 		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
1704 		if (rc) {
1705 			ibdev_err(&rdev->ibdev, "Modify HW SRQ failed!");
1706 			return rc;
1707 		}
1708 		/* On success, update the shadow */
1709 		srq->srq_limit = srq_attr->srq_limit;
1710 		/* No need to Build and send response back to udata */
1711 		break;
1712 	default:
1713 		ibdev_err(&rdev->ibdev,
1714 			  "Unsupported srq_attr_mask 0x%x", srq_attr_mask);
1715 		return -EINVAL;
1716 	}
1717 	return 0;
1718 }
1719 
1720 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
1721 {
1722 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1723 					       ib_srq);
1724 	struct bnxt_re_srq tsrq;
1725 	struct bnxt_re_dev *rdev = srq->rdev;
1726 	int rc;
1727 
1728 	/* Get live SRQ attr */
1729 	tsrq.qplib_srq.id = srq->qplib_srq.id;
1730 	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
1731 	if (rc) {
1732 		ibdev_err(&rdev->ibdev, "Query HW SRQ failed!");
1733 		return rc;
1734 	}
1735 	srq_attr->max_wr = srq->qplib_srq.max_wqe;
1736 	srq_attr->max_sge = srq->qplib_srq.max_sge;
1737 	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
1738 
1739 	return 0;
1740 }
1741 
1742 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr,
1743 			  const struct ib_recv_wr **bad_wr)
1744 {
1745 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1746 					       ib_srq);
1747 	struct bnxt_qplib_swqe wqe;
1748 	unsigned long flags;
1749 	int rc = 0;
1750 
1751 	spin_lock_irqsave(&srq->lock, flags);
1752 	while (wr) {
1753 		/* Transcribe each ib_recv_wr to qplib_swqe */
1754 		wqe.num_sge = wr->num_sge;
1755 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
1756 		wqe.wr_id = wr->wr_id;
1757 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
1758 
1759 		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
1760 		if (rc) {
1761 			*bad_wr = wr;
1762 			break;
1763 		}
1764 		wr = wr->next;
1765 	}
1766 	spin_unlock_irqrestore(&srq->lock, flags);
1767 
1768 	return rc;
1769 }
1770 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
1771 				    struct bnxt_re_qp *qp1_qp,
1772 				    int qp_attr_mask)
1773 {
1774 	struct bnxt_re_qp *qp = rdev->gsi_ctx.gsi_sqp;
1775 	int rc = 0;
1776 
1777 	if (qp_attr_mask & IB_QP_STATE) {
1778 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1779 		qp->qplib_qp.state = qp1_qp->qplib_qp.state;
1780 	}
1781 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1782 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1783 		qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index;
1784 	}
1785 
1786 	if (qp_attr_mask & IB_QP_QKEY) {
1787 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1788 		/* Using a Random  QKEY */
1789 		qp->qplib_qp.qkey = 0x81818181;
1790 	}
1791 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1792 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1793 		qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn;
1794 	}
1795 
1796 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1797 	if (rc)
1798 		ibdev_err(&rdev->ibdev, "Failed to modify Shadow QP for QP1");
1799 	return rc;
1800 }
1801 
1802 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1803 		      int qp_attr_mask, struct ib_udata *udata)
1804 {
1805 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1806 	struct bnxt_re_dev *rdev = qp->rdev;
1807 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1808 	enum ib_qp_state curr_qp_state, new_qp_state;
1809 	int rc, entries;
1810 	unsigned int flags;
1811 	u8 nw_type;
1812 
1813 	if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1814 		return -EOPNOTSUPP;
1815 
1816 	qp->qplib_qp.modify_flags = 0;
1817 	if (qp_attr_mask & IB_QP_STATE) {
1818 		curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state);
1819 		new_qp_state = qp_attr->qp_state;
1820 		if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state,
1821 					ib_qp->qp_type, qp_attr_mask)) {
1822 			ibdev_err(&rdev->ibdev,
1823 				  "Invalid attribute mask: %#x specified ",
1824 				  qp_attr_mask);
1825 			ibdev_err(&rdev->ibdev,
1826 				  "for qpn: %#x type: %#x",
1827 				  ib_qp->qp_num, ib_qp->qp_type);
1828 			ibdev_err(&rdev->ibdev,
1829 				  "curr_qp_state=0x%x, new_qp_state=0x%x\n",
1830 				  curr_qp_state, new_qp_state);
1831 			return -EINVAL;
1832 		}
1833 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1834 		qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state);
1835 
1836 		if (!qp->sumem &&
1837 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1838 			ibdev_dbg(&rdev->ibdev,
1839 				  "Move QP = %p to flush list\n", qp);
1840 			flags = bnxt_re_lock_cqs(qp);
1841 			bnxt_qplib_add_flush_qp(&qp->qplib_qp);
1842 			bnxt_re_unlock_cqs(qp, flags);
1843 		}
1844 		if (!qp->sumem &&
1845 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) {
1846 			ibdev_dbg(&rdev->ibdev,
1847 				  "Move QP = %p out of flush list\n", qp);
1848 			flags = bnxt_re_lock_cqs(qp);
1849 			bnxt_qplib_clean_qp(&qp->qplib_qp);
1850 			bnxt_re_unlock_cqs(qp, flags);
1851 		}
1852 	}
1853 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1854 		qp->qplib_qp.modify_flags |=
1855 				CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY;
1856 		qp->qplib_qp.en_sqd_async_notify = true;
1857 	}
1858 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
1859 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS;
1860 		qp->qplib_qp.access =
1861 			__from_ib_access_flags(qp_attr->qp_access_flags);
1862 		/* LOCAL_WRITE access must be set to allow RC receive */
1863 		qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
1864 		/* Temp: Set all params on QP as of now */
1865 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE;
1866 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ;
1867 	}
1868 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1869 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1870 		qp->qplib_qp.pkey_index = qp_attr->pkey_index;
1871 	}
1872 	if (qp_attr_mask & IB_QP_QKEY) {
1873 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1874 		qp->qplib_qp.qkey = qp_attr->qkey;
1875 	}
1876 	if (qp_attr_mask & IB_QP_AV) {
1877 		const struct ib_global_route *grh =
1878 			rdma_ah_read_grh(&qp_attr->ah_attr);
1879 		const struct ib_gid_attr *sgid_attr;
1880 		struct bnxt_re_gid_ctx *ctx;
1881 
1882 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID |
1883 				     CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL |
1884 				     CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX |
1885 				     CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT |
1886 				     CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS |
1887 				     CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC |
1888 				     CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID;
1889 		memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw,
1890 		       sizeof(qp->qplib_qp.ah.dgid.data));
1891 		qp->qplib_qp.ah.flow_label = grh->flow_label;
1892 		sgid_attr = grh->sgid_attr;
1893 		/* Get the HW context of the GID. The reference
1894 		 * of GID table entry is already taken by the caller.
1895 		 */
1896 		ctx = rdma_read_gid_hw_context(sgid_attr);
1897 		qp->qplib_qp.ah.sgid_index = ctx->idx;
1898 		qp->qplib_qp.ah.host_sgid_index = grh->sgid_index;
1899 		qp->qplib_qp.ah.hop_limit = grh->hop_limit;
1900 		qp->qplib_qp.ah.traffic_class = grh->traffic_class;
1901 		qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr);
1902 		ether_addr_copy(qp->qplib_qp.ah.dmac,
1903 				qp_attr->ah_attr.roce.dmac);
1904 
1905 		rc = rdma_read_gid_l2_fields(sgid_attr, NULL,
1906 					     &qp->qplib_qp.smac[0]);
1907 		if (rc)
1908 			return rc;
1909 
1910 		nw_type = rdma_gid_attr_network_type(sgid_attr);
1911 		switch (nw_type) {
1912 		case RDMA_NETWORK_IPV4:
1913 			qp->qplib_qp.nw_type =
1914 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4;
1915 			break;
1916 		case RDMA_NETWORK_IPV6:
1917 			qp->qplib_qp.nw_type =
1918 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6;
1919 			break;
1920 		default:
1921 			qp->qplib_qp.nw_type =
1922 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1;
1923 			break;
1924 		}
1925 	}
1926 
1927 	if (qp_attr_mask & IB_QP_PATH_MTU) {
1928 		qp->qplib_qp.modify_flags |=
1929 				CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1930 		qp->qplib_qp.path_mtu = __from_ib_mtu(qp_attr->path_mtu);
1931 		qp->qplib_qp.mtu = ib_mtu_enum_to_int(qp_attr->path_mtu);
1932 	} else if (qp_attr->qp_state == IB_QPS_RTR) {
1933 		qp->qplib_qp.modify_flags |=
1934 			CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1935 		qp->qplib_qp.path_mtu =
1936 			__from_ib_mtu(iboe_get_mtu(rdev->netdev->mtu));
1937 		qp->qplib_qp.mtu =
1938 			ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1939 	}
1940 
1941 	if (qp_attr_mask & IB_QP_TIMEOUT) {
1942 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT;
1943 		qp->qplib_qp.timeout = qp_attr->timeout;
1944 	}
1945 	if (qp_attr_mask & IB_QP_RETRY_CNT) {
1946 		qp->qplib_qp.modify_flags |=
1947 				CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT;
1948 		qp->qplib_qp.retry_cnt = qp_attr->retry_cnt;
1949 	}
1950 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
1951 		qp->qplib_qp.modify_flags |=
1952 				CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY;
1953 		qp->qplib_qp.rnr_retry = qp_attr->rnr_retry;
1954 	}
1955 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) {
1956 		qp->qplib_qp.modify_flags |=
1957 				CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER;
1958 		qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer;
1959 	}
1960 	if (qp_attr_mask & IB_QP_RQ_PSN) {
1961 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN;
1962 		qp->qplib_qp.rq.psn = qp_attr->rq_psn;
1963 	}
1964 	if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1965 		qp->qplib_qp.modify_flags |=
1966 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC;
1967 		/* Cap the max_rd_atomic to device max */
1968 		qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic,
1969 						   dev_attr->max_qp_rd_atom);
1970 	}
1971 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1972 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1973 		qp->qplib_qp.sq.psn = qp_attr->sq_psn;
1974 	}
1975 	if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1976 		if (qp_attr->max_dest_rd_atomic >
1977 		    dev_attr->max_qp_init_rd_atom) {
1978 			ibdev_err(&rdev->ibdev,
1979 				  "max_dest_rd_atomic requested%d is > dev_max%d",
1980 				  qp_attr->max_dest_rd_atomic,
1981 				  dev_attr->max_qp_init_rd_atom);
1982 			return -EINVAL;
1983 		}
1984 
1985 		qp->qplib_qp.modify_flags |=
1986 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC;
1987 		qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic;
1988 	}
1989 	if (qp_attr_mask & IB_QP_CAP) {
1990 		qp->qplib_qp.modify_flags |=
1991 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE |
1992 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE |
1993 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE |
1994 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE |
1995 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA;
1996 		if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) ||
1997 		    (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) ||
1998 		    (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) ||
1999 		    (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) ||
2000 		    (qp_attr->cap.max_inline_data >=
2001 						dev_attr->max_inline_data)) {
2002 			ibdev_err(&rdev->ibdev,
2003 				  "Create QP failed - max exceeded");
2004 			return -EINVAL;
2005 		}
2006 		entries = roundup_pow_of_two(qp_attr->cap.max_send_wr);
2007 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
2008 						dev_attr->max_qp_wqes + 1);
2009 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
2010 						qp_attr->cap.max_send_wr;
2011 		/*
2012 		 * Reserving one slot for Phantom WQE. Some application can
2013 		 * post one extra entry in this case. Allowing this to avoid
2014 		 * unexpected Queue full condition
2015 		 */
2016 		qp->qplib_qp.sq.q_full_delta -= 1;
2017 		qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
2018 		if (qp->qplib_qp.rq.max_wqe) {
2019 			entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr);
2020 			qp->qplib_qp.rq.max_wqe =
2021 				min_t(u32, entries, dev_attr->max_qp_wqes + 1);
2022 			qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
2023 						       qp_attr->cap.max_recv_wr;
2024 			qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
2025 		} else {
2026 			/* SRQ was used prior, just ignore the RQ caps */
2027 		}
2028 	}
2029 	if (qp_attr_mask & IB_QP_DEST_QPN) {
2030 		qp->qplib_qp.modify_flags |=
2031 				CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID;
2032 		qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num;
2033 	}
2034 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
2035 	if (rc) {
2036 		ibdev_err(&rdev->ibdev, "Failed to modify HW QP");
2037 		return rc;
2038 	}
2039 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp)
2040 		rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask);
2041 	return rc;
2042 }
2043 
2044 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
2045 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
2046 {
2047 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2048 	struct bnxt_re_dev *rdev = qp->rdev;
2049 	struct bnxt_qplib_qp *qplib_qp;
2050 	int rc;
2051 
2052 	qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL);
2053 	if (!qplib_qp)
2054 		return -ENOMEM;
2055 
2056 	qplib_qp->id = qp->qplib_qp.id;
2057 	qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index;
2058 
2059 	rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp);
2060 	if (rc) {
2061 		ibdev_err(&rdev->ibdev, "Failed to query HW QP");
2062 		goto out;
2063 	}
2064 	qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state);
2065 	qp_attr->cur_qp_state = __to_ib_qp_state(qplib_qp->cur_qp_state);
2066 	qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0;
2067 	qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access);
2068 	qp_attr->pkey_index = qplib_qp->pkey_index;
2069 	qp_attr->qkey = qplib_qp->qkey;
2070 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2071 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label,
2072 			qplib_qp->ah.host_sgid_index,
2073 			qplib_qp->ah.hop_limit,
2074 			qplib_qp->ah.traffic_class);
2075 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data);
2076 	rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl);
2077 	ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac);
2078 	qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu);
2079 	qp_attr->timeout = qplib_qp->timeout;
2080 	qp_attr->retry_cnt = qplib_qp->retry_cnt;
2081 	qp_attr->rnr_retry = qplib_qp->rnr_retry;
2082 	qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer;
2083 	qp_attr->rq_psn = qplib_qp->rq.psn;
2084 	qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic;
2085 	qp_attr->sq_psn = qplib_qp->sq.psn;
2086 	qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic;
2087 	qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR :
2088 							 IB_SIGNAL_REQ_WR;
2089 	qp_attr->dest_qp_num = qplib_qp->dest_qpn;
2090 
2091 	qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe;
2092 	qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge;
2093 	qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe;
2094 	qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge;
2095 	qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data;
2096 	qp_init_attr->cap = qp_attr->cap;
2097 
2098 out:
2099 	kfree(qplib_qp);
2100 	return rc;
2101 }
2102 
2103 /* Routine for sending QP1 packets for RoCE V1 an V2
2104  */
2105 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp,
2106 				     const struct ib_send_wr *wr,
2107 				     struct bnxt_qplib_swqe *wqe,
2108 				     int payload_size)
2109 {
2110 	struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah,
2111 					     ib_ah);
2112 	struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah;
2113 	const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr;
2114 	struct bnxt_qplib_sge sge;
2115 	u8 nw_type;
2116 	u16 ether_type;
2117 	union ib_gid dgid;
2118 	bool is_eth = false;
2119 	bool is_vlan = false;
2120 	bool is_grh = false;
2121 	bool is_udp = false;
2122 	u8 ip_version = 0;
2123 	u16 vlan_id = 0xFFFF;
2124 	void *buf;
2125 	int i, rc = 0;
2126 
2127 	memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr));
2128 
2129 	rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL);
2130 	if (rc)
2131 		return rc;
2132 
2133 	/* Get network header type for this GID */
2134 	nw_type = rdma_gid_attr_network_type(sgid_attr);
2135 	switch (nw_type) {
2136 	case RDMA_NETWORK_IPV4:
2137 		nw_type = BNXT_RE_ROCEV2_IPV4_PACKET;
2138 		break;
2139 	case RDMA_NETWORK_IPV6:
2140 		nw_type = BNXT_RE_ROCEV2_IPV6_PACKET;
2141 		break;
2142 	default:
2143 		nw_type = BNXT_RE_ROCE_V1_PACKET;
2144 		break;
2145 	}
2146 	memcpy(&dgid.raw, &qplib_ah->dgid, 16);
2147 	is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2148 	if (is_udp) {
2149 		if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) {
2150 			ip_version = 4;
2151 			ether_type = ETH_P_IP;
2152 		} else {
2153 			ip_version = 6;
2154 			ether_type = ETH_P_IPV6;
2155 		}
2156 		is_grh = false;
2157 	} else {
2158 		ether_type = ETH_P_IBOE;
2159 		is_grh = true;
2160 	}
2161 
2162 	is_eth = true;
2163 	is_vlan = (vlan_id && (vlan_id < 0x1000)) ? true : false;
2164 
2165 	ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh,
2166 			  ip_version, is_udp, 0, &qp->qp1_hdr);
2167 
2168 	/* ETH */
2169 	ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac);
2170 	ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac);
2171 
2172 	/* For vlan, check the sgid for vlan existence */
2173 
2174 	if (!is_vlan) {
2175 		qp->qp1_hdr.eth.type = cpu_to_be16(ether_type);
2176 	} else {
2177 		qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type);
2178 		qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id);
2179 	}
2180 
2181 	if (is_grh || (ip_version == 6)) {
2182 		memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw,
2183 		       sizeof(sgid_attr->gid));
2184 		memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data,
2185 		       sizeof(sgid_attr->gid));
2186 		qp->qp1_hdr.grh.hop_limit     = qplib_ah->hop_limit;
2187 	}
2188 
2189 	if (ip_version == 4) {
2190 		qp->qp1_hdr.ip4.tos = 0;
2191 		qp->qp1_hdr.ip4.id = 0;
2192 		qp->qp1_hdr.ip4.frag_off = htons(IP_DF);
2193 		qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit;
2194 
2195 		memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4);
2196 		memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4);
2197 		qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr);
2198 	}
2199 
2200 	if (is_udp) {
2201 		qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT);
2202 		qp->qp1_hdr.udp.sport = htons(0x8CD1);
2203 		qp->qp1_hdr.udp.csum = 0;
2204 	}
2205 
2206 	/* BTH */
2207 	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
2208 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2209 		qp->qp1_hdr.immediate_present = 1;
2210 	} else {
2211 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2212 	}
2213 	if (wr->send_flags & IB_SEND_SOLICITED)
2214 		qp->qp1_hdr.bth.solicited_event = 1;
2215 	/* pad_count */
2216 	qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3;
2217 
2218 	/* P_key for QP1 is for all members */
2219 	qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF);
2220 	qp->qp1_hdr.bth.destination_qpn = IB_QP1;
2221 	qp->qp1_hdr.bth.ack_req = 0;
2222 	qp->send_psn++;
2223 	qp->send_psn &= BTH_PSN_MASK;
2224 	qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn);
2225 	/* DETH */
2226 	/* Use the priviledged Q_Key for QP1 */
2227 	qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY);
2228 	qp->qp1_hdr.deth.source_qpn = IB_QP1;
2229 
2230 	/* Pack the QP1 to the transmit buffer */
2231 	buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge);
2232 	if (buf) {
2233 		ib_ud_header_pack(&qp->qp1_hdr, buf);
2234 		for (i = wqe->num_sge; i; i--) {
2235 			wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr;
2236 			wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey;
2237 			wqe->sg_list[i].size = wqe->sg_list[i - 1].size;
2238 		}
2239 
2240 		/*
2241 		 * Max Header buf size for IPV6 RoCE V2 is 86,
2242 		 * which is same as the QP1 SQ header buffer.
2243 		 * Header buf size for IPV4 RoCE V2 can be 66.
2244 		 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
2245 		 * Subtract 20 bytes from QP1 SQ header buf size
2246 		 */
2247 		if (is_udp && ip_version == 4)
2248 			sge.size -= 20;
2249 		/*
2250 		 * Max Header buf size for RoCE V1 is 78.
2251 		 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
2252 		 * Subtract 8 bytes from QP1 SQ header buf size
2253 		 */
2254 		if (!is_udp)
2255 			sge.size -= 8;
2256 
2257 		/* Subtract 4 bytes for non vlan packets */
2258 		if (!is_vlan)
2259 			sge.size -= 4;
2260 
2261 		wqe->sg_list[0].addr = sge.addr;
2262 		wqe->sg_list[0].lkey = sge.lkey;
2263 		wqe->sg_list[0].size = sge.size;
2264 		wqe->num_sge++;
2265 
2266 	} else {
2267 		ibdev_err(&qp->rdev->ibdev, "QP1 buffer is empty!");
2268 		rc = -ENOMEM;
2269 	}
2270 	return rc;
2271 }
2272 
2273 /* For the MAD layer, it only provides the recv SGE the size of
2274  * ib_grh + MAD datagram.  No Ethernet headers, Ethertype, BTH, DETH,
2275  * nor RoCE iCRC.  The Cu+ solution must provide buffer for the entire
2276  * receive packet (334 bytes) with no VLAN and then copy the GRH
2277  * and the MAD datagram out to the provided SGE.
2278  */
2279 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp,
2280 					    const struct ib_recv_wr *wr,
2281 					    struct bnxt_qplib_swqe *wqe,
2282 					    int payload_size)
2283 {
2284 	struct bnxt_re_sqp_entries *sqp_entry;
2285 	struct bnxt_qplib_sge ref, sge;
2286 	struct bnxt_re_dev *rdev;
2287 	u32 rq_prod_index;
2288 
2289 	rdev = qp->rdev;
2290 
2291 	rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp);
2292 
2293 	if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge))
2294 		return -ENOMEM;
2295 
2296 	/* Create 1 SGE to receive the entire
2297 	 * ethernet packet
2298 	 */
2299 	/* Save the reference from ULP */
2300 	ref.addr = wqe->sg_list[0].addr;
2301 	ref.lkey = wqe->sg_list[0].lkey;
2302 	ref.size = wqe->sg_list[0].size;
2303 
2304 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[rq_prod_index];
2305 
2306 	/* SGE 1 */
2307 	wqe->sg_list[0].addr = sge.addr;
2308 	wqe->sg_list[0].lkey = sge.lkey;
2309 	wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
2310 	sge.size -= wqe->sg_list[0].size;
2311 
2312 	sqp_entry->sge.addr = ref.addr;
2313 	sqp_entry->sge.lkey = ref.lkey;
2314 	sqp_entry->sge.size = ref.size;
2315 	/* Store the wrid for reporting completion */
2316 	sqp_entry->wrid = wqe->wr_id;
2317 	/* change the wqe->wrid to table index */
2318 	wqe->wr_id = rq_prod_index;
2319 	return 0;
2320 }
2321 
2322 static int is_ud_qp(struct bnxt_re_qp *qp)
2323 {
2324 	return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD ||
2325 		qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI);
2326 }
2327 
2328 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp,
2329 				  const struct ib_send_wr *wr,
2330 				  struct bnxt_qplib_swqe *wqe)
2331 {
2332 	struct bnxt_re_ah *ah = NULL;
2333 
2334 	if (is_ud_qp(qp)) {
2335 		ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah);
2336 		wqe->send.q_key = ud_wr(wr)->remote_qkey;
2337 		wqe->send.dst_qp = ud_wr(wr)->remote_qpn;
2338 		wqe->send.avid = ah->qplib_ah.id;
2339 	}
2340 	switch (wr->opcode) {
2341 	case IB_WR_SEND:
2342 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND;
2343 		break;
2344 	case IB_WR_SEND_WITH_IMM:
2345 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM;
2346 		wqe->send.imm_data = wr->ex.imm_data;
2347 		break;
2348 	case IB_WR_SEND_WITH_INV:
2349 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV;
2350 		wqe->send.inv_key = wr->ex.invalidate_rkey;
2351 		break;
2352 	default:
2353 		return -EINVAL;
2354 	}
2355 	if (wr->send_flags & IB_SEND_SIGNALED)
2356 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2357 	if (wr->send_flags & IB_SEND_FENCE)
2358 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2359 	if (wr->send_flags & IB_SEND_SOLICITED)
2360 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2361 	if (wr->send_flags & IB_SEND_INLINE)
2362 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2363 
2364 	return 0;
2365 }
2366 
2367 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr,
2368 				  struct bnxt_qplib_swqe *wqe)
2369 {
2370 	switch (wr->opcode) {
2371 	case IB_WR_RDMA_WRITE:
2372 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE;
2373 		break;
2374 	case IB_WR_RDMA_WRITE_WITH_IMM:
2375 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM;
2376 		wqe->rdma.imm_data = wr->ex.imm_data;
2377 		break;
2378 	case IB_WR_RDMA_READ:
2379 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ;
2380 		wqe->rdma.inv_key = wr->ex.invalidate_rkey;
2381 		break;
2382 	default:
2383 		return -EINVAL;
2384 	}
2385 	wqe->rdma.remote_va = rdma_wr(wr)->remote_addr;
2386 	wqe->rdma.r_key = rdma_wr(wr)->rkey;
2387 	if (wr->send_flags & IB_SEND_SIGNALED)
2388 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2389 	if (wr->send_flags & IB_SEND_FENCE)
2390 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2391 	if (wr->send_flags & IB_SEND_SOLICITED)
2392 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2393 	if (wr->send_flags & IB_SEND_INLINE)
2394 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2395 
2396 	return 0;
2397 }
2398 
2399 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr,
2400 				    struct bnxt_qplib_swqe *wqe)
2401 {
2402 	switch (wr->opcode) {
2403 	case IB_WR_ATOMIC_CMP_AND_SWP:
2404 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP;
2405 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2406 		wqe->atomic.swap_data = atomic_wr(wr)->swap;
2407 		break;
2408 	case IB_WR_ATOMIC_FETCH_AND_ADD:
2409 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD;
2410 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2411 		break;
2412 	default:
2413 		return -EINVAL;
2414 	}
2415 	wqe->atomic.remote_va = atomic_wr(wr)->remote_addr;
2416 	wqe->atomic.r_key = atomic_wr(wr)->rkey;
2417 	if (wr->send_flags & IB_SEND_SIGNALED)
2418 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2419 	if (wr->send_flags & IB_SEND_FENCE)
2420 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2421 	if (wr->send_flags & IB_SEND_SOLICITED)
2422 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2423 	return 0;
2424 }
2425 
2426 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr,
2427 				 struct bnxt_qplib_swqe *wqe)
2428 {
2429 	wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV;
2430 	wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey;
2431 
2432 	/* Need unconditional fence for local invalidate
2433 	 * opcode to work as expected.
2434 	 */
2435 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2436 
2437 	if (wr->send_flags & IB_SEND_SIGNALED)
2438 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2439 	if (wr->send_flags & IB_SEND_SOLICITED)
2440 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2441 
2442 	return 0;
2443 }
2444 
2445 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr,
2446 				 struct bnxt_qplib_swqe *wqe)
2447 {
2448 	struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr);
2449 	struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl;
2450 	int access = wr->access;
2451 
2452 	wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0];
2453 	wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0];
2454 	wqe->frmr.page_list = mr->pages;
2455 	wqe->frmr.page_list_len = mr->npages;
2456 	wqe->frmr.levels = qplib_frpl->hwq.level;
2457 	wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR;
2458 
2459 	/* Need unconditional fence for reg_mr
2460 	 * opcode to function as expected.
2461 	 */
2462 
2463 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2464 
2465 	if (wr->wr.send_flags & IB_SEND_SIGNALED)
2466 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2467 
2468 	if (access & IB_ACCESS_LOCAL_WRITE)
2469 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE;
2470 	if (access & IB_ACCESS_REMOTE_READ)
2471 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ;
2472 	if (access & IB_ACCESS_REMOTE_WRITE)
2473 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE;
2474 	if (access & IB_ACCESS_REMOTE_ATOMIC)
2475 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC;
2476 	if (access & IB_ACCESS_MW_BIND)
2477 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND;
2478 
2479 	wqe->frmr.l_key = wr->key;
2480 	wqe->frmr.length = wr->mr->length;
2481 	wqe->frmr.pbl_pg_sz_log = (wr->mr->page_size >> PAGE_SHIFT_4K) - 1;
2482 	wqe->frmr.va = wr->mr->iova;
2483 	return 0;
2484 }
2485 
2486 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev,
2487 				    const struct ib_send_wr *wr,
2488 				    struct bnxt_qplib_swqe *wqe)
2489 {
2490 	/*  Copy the inline data to the data  field */
2491 	u8 *in_data;
2492 	u32 i, sge_len;
2493 	void *sge_addr;
2494 
2495 	in_data = wqe->inline_data;
2496 	for (i = 0; i < wr->num_sge; i++) {
2497 		sge_addr = (void *)(unsigned long)
2498 				wr->sg_list[i].addr;
2499 		sge_len = wr->sg_list[i].length;
2500 
2501 		if ((sge_len + wqe->inline_len) >
2502 		    BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) {
2503 			ibdev_err(&rdev->ibdev,
2504 				  "Inline data size requested > supported value");
2505 			return -EINVAL;
2506 		}
2507 		sge_len = wr->sg_list[i].length;
2508 
2509 		memcpy(in_data, sge_addr, sge_len);
2510 		in_data += wr->sg_list[i].length;
2511 		wqe->inline_len += wr->sg_list[i].length;
2512 	}
2513 	return wqe->inline_len;
2514 }
2515 
2516 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
2517 				   const struct ib_send_wr *wr,
2518 				   struct bnxt_qplib_swqe *wqe)
2519 {
2520 	int payload_sz = 0;
2521 
2522 	if (wr->send_flags & IB_SEND_INLINE)
2523 		payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe);
2524 	else
2525 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list,
2526 					       wqe->num_sge);
2527 
2528 	return payload_sz;
2529 }
2530 
2531 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
2532 {
2533 	if ((qp->ib_qp.qp_type == IB_QPT_UD ||
2534 	     qp->ib_qp.qp_type == IB_QPT_GSI ||
2535 	     qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
2536 	     qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
2537 		int qp_attr_mask;
2538 		struct ib_qp_attr qp_attr;
2539 
2540 		qp_attr_mask = IB_QP_STATE;
2541 		qp_attr.qp_state = IB_QPS_RTS;
2542 		bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
2543 		qp->qplib_qp.wqe_cnt = 0;
2544 	}
2545 }
2546 
2547 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
2548 				       struct bnxt_re_qp *qp,
2549 				       const struct ib_send_wr *wr)
2550 {
2551 	int rc = 0, payload_sz = 0;
2552 	unsigned long flags;
2553 
2554 	spin_lock_irqsave(&qp->sq_lock, flags);
2555 	while (wr) {
2556 		struct bnxt_qplib_swqe wqe = {};
2557 
2558 		/* Common */
2559 		wqe.num_sge = wr->num_sge;
2560 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2561 			ibdev_err(&rdev->ibdev,
2562 				  "Limit exceeded for Send SGEs");
2563 			rc = -EINVAL;
2564 			goto bad;
2565 		}
2566 
2567 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2568 		if (payload_sz < 0) {
2569 			rc = -EINVAL;
2570 			goto bad;
2571 		}
2572 		wqe.wr_id = wr->wr_id;
2573 
2574 		wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND;
2575 
2576 		rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2577 		if (!rc)
2578 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2579 bad:
2580 		if (rc) {
2581 			ibdev_err(&rdev->ibdev,
2582 				  "Post send failed opcode = %#x rc = %d",
2583 				  wr->opcode, rc);
2584 			break;
2585 		}
2586 		wr = wr->next;
2587 	}
2588 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2589 	bnxt_ud_qp_hw_stall_workaround(qp);
2590 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2591 	return rc;
2592 }
2593 
2594 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr,
2595 		      const struct ib_send_wr **bad_wr)
2596 {
2597 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2598 	struct bnxt_qplib_swqe wqe;
2599 	int rc = 0, payload_sz = 0;
2600 	unsigned long flags;
2601 
2602 	spin_lock_irqsave(&qp->sq_lock, flags);
2603 	while (wr) {
2604 		/* House keeping */
2605 		memset(&wqe, 0, sizeof(wqe));
2606 
2607 		/* Common */
2608 		wqe.num_sge = wr->num_sge;
2609 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2610 			ibdev_err(&qp->rdev->ibdev,
2611 				  "Limit exceeded for Send SGEs");
2612 			rc = -EINVAL;
2613 			goto bad;
2614 		}
2615 
2616 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2617 		if (payload_sz < 0) {
2618 			rc = -EINVAL;
2619 			goto bad;
2620 		}
2621 		wqe.wr_id = wr->wr_id;
2622 
2623 		switch (wr->opcode) {
2624 		case IB_WR_SEND:
2625 		case IB_WR_SEND_WITH_IMM:
2626 			if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) {
2627 				rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe,
2628 							       payload_sz);
2629 				if (rc)
2630 					goto bad;
2631 				wqe.rawqp1.lflags |=
2632 					SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC;
2633 			}
2634 			switch (wr->send_flags) {
2635 			case IB_SEND_IP_CSUM:
2636 				wqe.rawqp1.lflags |=
2637 					SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM;
2638 				break;
2639 			default:
2640 				break;
2641 			}
2642 			fallthrough;
2643 		case IB_WR_SEND_WITH_INV:
2644 			rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2645 			break;
2646 		case IB_WR_RDMA_WRITE:
2647 		case IB_WR_RDMA_WRITE_WITH_IMM:
2648 		case IB_WR_RDMA_READ:
2649 			rc = bnxt_re_build_rdma_wqe(wr, &wqe);
2650 			break;
2651 		case IB_WR_ATOMIC_CMP_AND_SWP:
2652 		case IB_WR_ATOMIC_FETCH_AND_ADD:
2653 			rc = bnxt_re_build_atomic_wqe(wr, &wqe);
2654 			break;
2655 		case IB_WR_RDMA_READ_WITH_INV:
2656 			ibdev_err(&qp->rdev->ibdev,
2657 				  "RDMA Read with Invalidate is not supported");
2658 			rc = -EINVAL;
2659 			goto bad;
2660 		case IB_WR_LOCAL_INV:
2661 			rc = bnxt_re_build_inv_wqe(wr, &wqe);
2662 			break;
2663 		case IB_WR_REG_MR:
2664 			rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe);
2665 			break;
2666 		default:
2667 			/* Unsupported WRs */
2668 			ibdev_err(&qp->rdev->ibdev,
2669 				  "WR (%#x) is not supported", wr->opcode);
2670 			rc = -EINVAL;
2671 			goto bad;
2672 		}
2673 		if (!rc)
2674 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2675 bad:
2676 		if (rc) {
2677 			ibdev_err(&qp->rdev->ibdev,
2678 				  "post_send failed op:%#x qps = %#x rc = %d\n",
2679 				  wr->opcode, qp->qplib_qp.state, rc);
2680 			*bad_wr = wr;
2681 			break;
2682 		}
2683 		wr = wr->next;
2684 	}
2685 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2686 	bnxt_ud_qp_hw_stall_workaround(qp);
2687 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2688 
2689 	return rc;
2690 }
2691 
2692 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev,
2693 				       struct bnxt_re_qp *qp,
2694 				       const struct ib_recv_wr *wr)
2695 {
2696 	struct bnxt_qplib_swqe wqe;
2697 	int rc = 0;
2698 
2699 	memset(&wqe, 0, sizeof(wqe));
2700 	while (wr) {
2701 		/* House keeping */
2702 		memset(&wqe, 0, sizeof(wqe));
2703 
2704 		/* Common */
2705 		wqe.num_sge = wr->num_sge;
2706 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2707 			ibdev_err(&rdev->ibdev,
2708 				  "Limit exceeded for Receive SGEs");
2709 			rc = -EINVAL;
2710 			break;
2711 		}
2712 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
2713 		wqe.wr_id = wr->wr_id;
2714 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2715 
2716 		rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2717 		if (rc)
2718 			break;
2719 
2720 		wr = wr->next;
2721 	}
2722 	if (!rc)
2723 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2724 	return rc;
2725 }
2726 
2727 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
2728 		      const struct ib_recv_wr **bad_wr)
2729 {
2730 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2731 	struct bnxt_qplib_swqe wqe;
2732 	int rc = 0, payload_sz = 0;
2733 	unsigned long flags;
2734 	u32 count = 0;
2735 
2736 	spin_lock_irqsave(&qp->rq_lock, flags);
2737 	while (wr) {
2738 		/* House keeping */
2739 		memset(&wqe, 0, sizeof(wqe));
2740 
2741 		/* Common */
2742 		wqe.num_sge = wr->num_sge;
2743 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2744 			ibdev_err(&qp->rdev->ibdev,
2745 				  "Limit exceeded for Receive SGEs");
2746 			rc = -EINVAL;
2747 			*bad_wr = wr;
2748 			break;
2749 		}
2750 
2751 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
2752 					       wr->num_sge);
2753 		wqe.wr_id = wr->wr_id;
2754 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2755 
2756 		if (ib_qp->qp_type == IB_QPT_GSI &&
2757 		    qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI)
2758 			rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe,
2759 							      payload_sz);
2760 		if (!rc)
2761 			rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2762 		if (rc) {
2763 			*bad_wr = wr;
2764 			break;
2765 		}
2766 
2767 		/* Ring DB if the RQEs posted reaches a threshold value */
2768 		if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
2769 			bnxt_qplib_post_recv_db(&qp->qplib_qp);
2770 			count = 0;
2771 		}
2772 
2773 		wr = wr->next;
2774 	}
2775 
2776 	if (count)
2777 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2778 
2779 	spin_unlock_irqrestore(&qp->rq_lock, flags);
2780 
2781 	return rc;
2782 }
2783 
2784 /* Completion Queues */
2785 int bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
2786 {
2787 	struct bnxt_re_cq *cq;
2788 	struct bnxt_qplib_nq *nq;
2789 	struct bnxt_re_dev *rdev;
2790 
2791 	cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
2792 	rdev = cq->rdev;
2793 	nq = cq->qplib_cq.nq;
2794 
2795 	bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2796 	ib_umem_release(cq->umem);
2797 
2798 	atomic_dec(&rdev->cq_count);
2799 	nq->budget--;
2800 	kfree(cq->cql);
2801 	return 0;
2802 }
2803 
2804 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
2805 		      struct ib_udata *udata)
2806 {
2807 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev);
2808 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
2809 	struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq);
2810 	int rc, entries;
2811 	int cqe = attr->cqe;
2812 	struct bnxt_qplib_nq *nq = NULL;
2813 	unsigned int nq_alloc_cnt;
2814 
2815 	if (attr->flags)
2816 		return -EOPNOTSUPP;
2817 
2818 	/* Validate CQ fields */
2819 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
2820 		ibdev_err(&rdev->ibdev, "Failed to create CQ -max exceeded");
2821 		return -EINVAL;
2822 	}
2823 
2824 	cq->rdev = rdev;
2825 	cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
2826 
2827 	entries = roundup_pow_of_two(cqe + 1);
2828 	if (entries > dev_attr->max_cq_wqes + 1)
2829 		entries = dev_attr->max_cq_wqes + 1;
2830 
2831 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
2832 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
2833 	if (udata) {
2834 		struct bnxt_re_cq_req req;
2835 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
2836 			udata, struct bnxt_re_ucontext, ib_uctx);
2837 		if (ib_copy_from_udata(&req, udata, sizeof(req))) {
2838 			rc = -EFAULT;
2839 			goto fail;
2840 		}
2841 
2842 		cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va,
2843 				       entries * sizeof(struct cq_base),
2844 				       IB_ACCESS_LOCAL_WRITE);
2845 		if (IS_ERR(cq->umem)) {
2846 			rc = PTR_ERR(cq->umem);
2847 			goto fail;
2848 		}
2849 		cq->qplib_cq.sg_info.umem = cq->umem;
2850 		cq->qplib_cq.dpi = &uctx->dpi;
2851 	} else {
2852 		cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL);
2853 		cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe),
2854 				  GFP_KERNEL);
2855 		if (!cq->cql) {
2856 			rc = -ENOMEM;
2857 			goto fail;
2858 		}
2859 
2860 		cq->qplib_cq.dpi = &rdev->dpi_privileged;
2861 	}
2862 	/*
2863 	 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
2864 	 * used for getting the NQ index.
2865 	 */
2866 	nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
2867 	nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
2868 	cq->qplib_cq.max_wqe = entries;
2869 	cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
2870 	cq->qplib_cq.nq	= nq;
2871 
2872 	rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
2873 	if (rc) {
2874 		ibdev_err(&rdev->ibdev, "Failed to create HW CQ");
2875 		goto fail;
2876 	}
2877 
2878 	cq->ib_cq.cqe = entries;
2879 	cq->cq_period = cq->qplib_cq.period;
2880 	nq->budget++;
2881 
2882 	atomic_inc(&rdev->cq_count);
2883 	spin_lock_init(&cq->cq_lock);
2884 
2885 	if (udata) {
2886 		struct bnxt_re_cq_resp resp;
2887 
2888 		resp.cqid = cq->qplib_cq.id;
2889 		resp.tail = cq->qplib_cq.hwq.cons;
2890 		resp.phase = cq->qplib_cq.period;
2891 		resp.rsvd = 0;
2892 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
2893 		if (rc) {
2894 			ibdev_err(&rdev->ibdev, "Failed to copy CQ udata");
2895 			bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2896 			goto c2fail;
2897 		}
2898 	}
2899 
2900 	return 0;
2901 
2902 c2fail:
2903 	ib_umem_release(cq->umem);
2904 fail:
2905 	kfree(cq->cql);
2906 	return rc;
2907 }
2908 
2909 static u8 __req_to_ib_wc_status(u8 qstatus)
2910 {
2911 	switch (qstatus) {
2912 	case CQ_REQ_STATUS_OK:
2913 		return IB_WC_SUCCESS;
2914 	case CQ_REQ_STATUS_BAD_RESPONSE_ERR:
2915 		return IB_WC_BAD_RESP_ERR;
2916 	case CQ_REQ_STATUS_LOCAL_LENGTH_ERR:
2917 		return IB_WC_LOC_LEN_ERR;
2918 	case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR:
2919 		return IB_WC_LOC_QP_OP_ERR;
2920 	case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR:
2921 		return IB_WC_LOC_PROT_ERR;
2922 	case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR:
2923 		return IB_WC_GENERAL_ERR;
2924 	case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR:
2925 		return IB_WC_REM_INV_REQ_ERR;
2926 	case CQ_REQ_STATUS_REMOTE_ACCESS_ERR:
2927 		return IB_WC_REM_ACCESS_ERR;
2928 	case CQ_REQ_STATUS_REMOTE_OPERATION_ERR:
2929 		return IB_WC_REM_OP_ERR;
2930 	case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR:
2931 		return IB_WC_RNR_RETRY_EXC_ERR;
2932 	case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR:
2933 		return IB_WC_RETRY_EXC_ERR;
2934 	case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR:
2935 		return IB_WC_WR_FLUSH_ERR;
2936 	default:
2937 		return IB_WC_GENERAL_ERR;
2938 	}
2939 	return 0;
2940 }
2941 
2942 static u8 __rawqp1_to_ib_wc_status(u8 qstatus)
2943 {
2944 	switch (qstatus) {
2945 	case CQ_RES_RAWETH_QP1_STATUS_OK:
2946 		return IB_WC_SUCCESS;
2947 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR:
2948 		return IB_WC_LOC_ACCESS_ERR;
2949 	case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR:
2950 		return IB_WC_LOC_LEN_ERR;
2951 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR:
2952 		return IB_WC_LOC_PROT_ERR;
2953 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR:
2954 		return IB_WC_LOC_QP_OP_ERR;
2955 	case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR:
2956 		return IB_WC_GENERAL_ERR;
2957 	case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR:
2958 		return IB_WC_WR_FLUSH_ERR;
2959 	case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR:
2960 		return IB_WC_WR_FLUSH_ERR;
2961 	default:
2962 		return IB_WC_GENERAL_ERR;
2963 	}
2964 }
2965 
2966 static u8 __rc_to_ib_wc_status(u8 qstatus)
2967 {
2968 	switch (qstatus) {
2969 	case CQ_RES_RC_STATUS_OK:
2970 		return IB_WC_SUCCESS;
2971 	case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR:
2972 		return IB_WC_LOC_ACCESS_ERR;
2973 	case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR:
2974 		return IB_WC_LOC_LEN_ERR;
2975 	case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR:
2976 		return IB_WC_LOC_PROT_ERR;
2977 	case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR:
2978 		return IB_WC_LOC_QP_OP_ERR;
2979 	case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR:
2980 		return IB_WC_GENERAL_ERR;
2981 	case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR:
2982 		return IB_WC_REM_INV_REQ_ERR;
2983 	case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR:
2984 		return IB_WC_WR_FLUSH_ERR;
2985 	case CQ_RES_RC_STATUS_HW_FLUSH_ERR:
2986 		return IB_WC_WR_FLUSH_ERR;
2987 	default:
2988 		return IB_WC_GENERAL_ERR;
2989 	}
2990 }
2991 
2992 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe)
2993 {
2994 	switch (cqe->type) {
2995 	case BNXT_QPLIB_SWQE_TYPE_SEND:
2996 		wc->opcode = IB_WC_SEND;
2997 		break;
2998 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM:
2999 		wc->opcode = IB_WC_SEND;
3000 		wc->wc_flags |= IB_WC_WITH_IMM;
3001 		break;
3002 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV:
3003 		wc->opcode = IB_WC_SEND;
3004 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3005 		break;
3006 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE:
3007 		wc->opcode = IB_WC_RDMA_WRITE;
3008 		break;
3009 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM:
3010 		wc->opcode = IB_WC_RDMA_WRITE;
3011 		wc->wc_flags |= IB_WC_WITH_IMM;
3012 		break;
3013 	case BNXT_QPLIB_SWQE_TYPE_RDMA_READ:
3014 		wc->opcode = IB_WC_RDMA_READ;
3015 		break;
3016 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP:
3017 		wc->opcode = IB_WC_COMP_SWAP;
3018 		break;
3019 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD:
3020 		wc->opcode = IB_WC_FETCH_ADD;
3021 		break;
3022 	case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV:
3023 		wc->opcode = IB_WC_LOCAL_INV;
3024 		break;
3025 	case BNXT_QPLIB_SWQE_TYPE_REG_MR:
3026 		wc->opcode = IB_WC_REG_MR;
3027 		break;
3028 	default:
3029 		wc->opcode = IB_WC_SEND;
3030 		break;
3031 	}
3032 
3033 	wc->status = __req_to_ib_wc_status(cqe->status);
3034 }
3035 
3036 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags,
3037 				     u16 raweth_qp1_flags2)
3038 {
3039 	bool is_ipv6 = false, is_ipv4 = false;
3040 
3041 	/* raweth_qp1_flags Bit 9-6 indicates itype */
3042 	if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3043 	    != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3044 		return -1;
3045 
3046 	if (raweth_qp1_flags2 &
3047 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC &&
3048 	    raweth_qp1_flags2 &
3049 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) {
3050 		/* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
3051 		(raweth_qp1_flags2 &
3052 		 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ?
3053 			(is_ipv6 = true) : (is_ipv4 = true);
3054 		return ((is_ipv6) ?
3055 			 BNXT_RE_ROCEV2_IPV6_PACKET :
3056 			 BNXT_RE_ROCEV2_IPV4_PACKET);
3057 	} else {
3058 		return BNXT_RE_ROCE_V1_PACKET;
3059 	}
3060 }
3061 
3062 static int bnxt_re_to_ib_nw_type(int nw_type)
3063 {
3064 	u8 nw_hdr_type = 0xFF;
3065 
3066 	switch (nw_type) {
3067 	case BNXT_RE_ROCE_V1_PACKET:
3068 		nw_hdr_type = RDMA_NETWORK_ROCE_V1;
3069 		break;
3070 	case BNXT_RE_ROCEV2_IPV4_PACKET:
3071 		nw_hdr_type = RDMA_NETWORK_IPV4;
3072 		break;
3073 	case BNXT_RE_ROCEV2_IPV6_PACKET:
3074 		nw_hdr_type = RDMA_NETWORK_IPV6;
3075 		break;
3076 	}
3077 	return nw_hdr_type;
3078 }
3079 
3080 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev,
3081 				       void *rq_hdr_buf)
3082 {
3083 	u8 *tmp_buf = NULL;
3084 	struct ethhdr *eth_hdr;
3085 	u16 eth_type;
3086 	bool rc = false;
3087 
3088 	tmp_buf = (u8 *)rq_hdr_buf;
3089 	/*
3090 	 * If dest mac is not same as I/F mac, this could be a
3091 	 * loopback address or multicast address, check whether
3092 	 * it is a loopback packet
3093 	 */
3094 	if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) {
3095 		tmp_buf += 4;
3096 		/* Check the  ether type */
3097 		eth_hdr = (struct ethhdr *)tmp_buf;
3098 		eth_type = ntohs(eth_hdr->h_proto);
3099 		switch (eth_type) {
3100 		case ETH_P_IBOE:
3101 			rc = true;
3102 			break;
3103 		case ETH_P_IP:
3104 		case ETH_P_IPV6: {
3105 			u32 len;
3106 			struct udphdr *udp_hdr;
3107 
3108 			len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) :
3109 						      sizeof(struct ipv6hdr));
3110 			tmp_buf += sizeof(struct ethhdr) + len;
3111 			udp_hdr = (struct udphdr *)tmp_buf;
3112 			if (ntohs(udp_hdr->dest) ==
3113 				    ROCE_V2_UDP_DPORT)
3114 				rc = true;
3115 			break;
3116 			}
3117 		default:
3118 			break;
3119 		}
3120 	}
3121 
3122 	return rc;
3123 }
3124 
3125 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *gsi_qp,
3126 					 struct bnxt_qplib_cqe *cqe)
3127 {
3128 	struct bnxt_re_dev *rdev = gsi_qp->rdev;
3129 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3130 	struct bnxt_re_qp *gsi_sqp = rdev->gsi_ctx.gsi_sqp;
3131 	struct bnxt_re_ah *gsi_sah;
3132 	struct ib_send_wr *swr;
3133 	struct ib_ud_wr udwr;
3134 	struct ib_recv_wr rwr;
3135 	int pkt_type = 0;
3136 	u32 tbl_idx;
3137 	void *rq_hdr_buf;
3138 	dma_addr_t rq_hdr_buf_map;
3139 	dma_addr_t shrq_hdr_buf_map;
3140 	u32 offset = 0;
3141 	u32 skip_bytes = 0;
3142 	struct ib_sge s_sge[2];
3143 	struct ib_sge r_sge[2];
3144 	int rc;
3145 
3146 	memset(&udwr, 0, sizeof(udwr));
3147 	memset(&rwr, 0, sizeof(rwr));
3148 	memset(&s_sge, 0, sizeof(s_sge));
3149 	memset(&r_sge, 0, sizeof(r_sge));
3150 
3151 	swr = &udwr.wr;
3152 	tbl_idx = cqe->wr_id;
3153 
3154 	rq_hdr_buf = gsi_qp->qplib_qp.rq_hdr_buf +
3155 			(tbl_idx * gsi_qp->qplib_qp.rq_hdr_buf_size);
3156 	rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3157 							  tbl_idx);
3158 
3159 	/* Shadow QP header buffer */
3160 	shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3161 							    tbl_idx);
3162 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3163 
3164 	/* Store this cqe */
3165 	memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe));
3166 	sqp_entry->qp1_qp = gsi_qp;
3167 
3168 	/* Find packet type from the cqe */
3169 
3170 	pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags,
3171 					     cqe->raweth_qp1_flags2);
3172 	if (pkt_type < 0) {
3173 		ibdev_err(&rdev->ibdev, "Invalid packet\n");
3174 		return -EINVAL;
3175 	}
3176 
3177 	/* Adjust the offset for the user buffer and post in the rq */
3178 
3179 	if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET)
3180 		offset = 20;
3181 
3182 	/*
3183 	 * QP1 loopback packet has 4 bytes of internal header before
3184 	 * ether header. Skip these four bytes.
3185 	 */
3186 	if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf))
3187 		skip_bytes = 4;
3188 
3189 	/* First send SGE . Skip the ether header*/
3190 	s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
3191 			+ skip_bytes;
3192 	s_sge[0].lkey = 0xFFFFFFFF;
3193 	s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 :
3194 				BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
3195 
3196 	/* Second Send SGE */
3197 	s_sge[1].addr = s_sge[0].addr + s_sge[0].length +
3198 			BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE;
3199 	if (pkt_type != BNXT_RE_ROCE_V1_PACKET)
3200 		s_sge[1].addr += 8;
3201 	s_sge[1].lkey = 0xFFFFFFFF;
3202 	s_sge[1].length = 256;
3203 
3204 	/* First recv SGE */
3205 
3206 	r_sge[0].addr = shrq_hdr_buf_map;
3207 	r_sge[0].lkey = 0xFFFFFFFF;
3208 	r_sge[0].length = 40;
3209 
3210 	r_sge[1].addr = sqp_entry->sge.addr + offset;
3211 	r_sge[1].lkey = sqp_entry->sge.lkey;
3212 	r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset;
3213 
3214 	/* Create receive work request */
3215 	rwr.num_sge = 2;
3216 	rwr.sg_list = r_sge;
3217 	rwr.wr_id = tbl_idx;
3218 	rwr.next = NULL;
3219 
3220 	rc = bnxt_re_post_recv_shadow_qp(rdev, gsi_sqp, &rwr);
3221 	if (rc) {
3222 		ibdev_err(&rdev->ibdev,
3223 			  "Failed to post Rx buffers to shadow QP");
3224 		return -ENOMEM;
3225 	}
3226 
3227 	swr->num_sge = 2;
3228 	swr->sg_list = s_sge;
3229 	swr->wr_id = tbl_idx;
3230 	swr->opcode = IB_WR_SEND;
3231 	swr->next = NULL;
3232 	gsi_sah = rdev->gsi_ctx.gsi_sah;
3233 	udwr.ah = &gsi_sah->ib_ah;
3234 	udwr.remote_qpn = gsi_sqp->qplib_qp.id;
3235 	udwr.remote_qkey = gsi_sqp->qplib_qp.qkey;
3236 
3237 	/* post data received  in the send queue */
3238 	rc = bnxt_re_post_send_shadow_qp(rdev, gsi_sqp, swr);
3239 
3240 	return 0;
3241 }
3242 
3243 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc,
3244 					  struct bnxt_qplib_cqe *cqe)
3245 {
3246 	wc->opcode = IB_WC_RECV;
3247 	wc->status = __rawqp1_to_ib_wc_status(cqe->status);
3248 	wc->wc_flags |= IB_WC_GRH;
3249 }
3250 
3251 static bool bnxt_re_check_if_vlan_valid(struct bnxt_re_dev *rdev,
3252 					u16 vlan_id)
3253 {
3254 	/*
3255 	 * Check if the vlan is configured in the host.  If not configured, it
3256 	 * can be a transparent VLAN. So dont report the vlan id.
3257 	 */
3258 	if (!__vlan_find_dev_deep_rcu(rdev->netdev,
3259 				      htons(ETH_P_8021Q), vlan_id))
3260 		return false;
3261 	return true;
3262 }
3263 
3264 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe,
3265 				u16 *vid, u8 *sl)
3266 {
3267 	bool ret = false;
3268 	u32 metadata;
3269 	u16 tpid;
3270 
3271 	metadata = orig_cqe->raweth_qp1_metadata;
3272 	if (orig_cqe->raweth_qp1_flags2 &
3273 		CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) {
3274 		tpid = ((metadata &
3275 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >>
3276 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT);
3277 		if (tpid == ETH_P_8021Q) {
3278 			*vid = metadata &
3279 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK;
3280 			*sl = (metadata &
3281 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >>
3282 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT;
3283 			ret = true;
3284 		}
3285 	}
3286 
3287 	return ret;
3288 }
3289 
3290 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc,
3291 				      struct bnxt_qplib_cqe *cqe)
3292 {
3293 	wc->opcode = IB_WC_RECV;
3294 	wc->status = __rc_to_ib_wc_status(cqe->status);
3295 
3296 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3297 		wc->wc_flags |= IB_WC_WITH_IMM;
3298 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3299 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3300 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3301 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3302 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3303 }
3304 
3305 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *gsi_sqp,
3306 					     struct ib_wc *wc,
3307 					     struct bnxt_qplib_cqe *cqe)
3308 {
3309 	struct bnxt_re_dev *rdev = gsi_sqp->rdev;
3310 	struct bnxt_re_qp *gsi_qp = NULL;
3311 	struct bnxt_qplib_cqe *orig_cqe = NULL;
3312 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3313 	int nw_type;
3314 	u32 tbl_idx;
3315 	u16 vlan_id;
3316 	u8 sl;
3317 
3318 	tbl_idx = cqe->wr_id;
3319 
3320 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3321 	gsi_qp = sqp_entry->qp1_qp;
3322 	orig_cqe = &sqp_entry->cqe;
3323 
3324 	wc->wr_id = sqp_entry->wrid;
3325 	wc->byte_len = orig_cqe->length;
3326 	wc->qp = &gsi_qp->ib_qp;
3327 
3328 	wc->ex.imm_data = orig_cqe->immdata;
3329 	wc->src_qp = orig_cqe->src_qp;
3330 	memcpy(wc->smac, orig_cqe->smac, ETH_ALEN);
3331 	if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) {
3332 		if (bnxt_re_check_if_vlan_valid(rdev, vlan_id)) {
3333 			wc->vlan_id = vlan_id;
3334 			wc->sl = sl;
3335 			wc->wc_flags |= IB_WC_WITH_VLAN;
3336 		}
3337 	}
3338 	wc->port_num = 1;
3339 	wc->vendor_err = orig_cqe->status;
3340 
3341 	wc->opcode = IB_WC_RECV;
3342 	wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status);
3343 	wc->wc_flags |= IB_WC_GRH;
3344 
3345 	nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags,
3346 					    orig_cqe->raweth_qp1_flags2);
3347 	if (nw_type >= 0) {
3348 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3349 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3350 	}
3351 }
3352 
3353 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp,
3354 				      struct ib_wc *wc,
3355 				      struct bnxt_qplib_cqe *cqe)
3356 {
3357 	u8 nw_type;
3358 
3359 	wc->opcode = IB_WC_RECV;
3360 	wc->status = __rc_to_ib_wc_status(cqe->status);
3361 
3362 	if (cqe->flags & CQ_RES_UD_FLAGS_IMM)
3363 		wc->wc_flags |= IB_WC_WITH_IMM;
3364 	/* report only on GSI QP for Thor */
3365 	if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) {
3366 		wc->wc_flags |= IB_WC_GRH;
3367 		memcpy(wc->smac, cqe->smac, ETH_ALEN);
3368 		wc->wc_flags |= IB_WC_WITH_SMAC;
3369 		if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) {
3370 			wc->vlan_id = (cqe->cfa_meta & 0xFFF);
3371 			if (wc->vlan_id < 0x1000)
3372 				wc->wc_flags |= IB_WC_WITH_VLAN;
3373 		}
3374 		nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >>
3375 			   CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT;
3376 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3377 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3378 	}
3379 
3380 }
3381 
3382 static int send_phantom_wqe(struct bnxt_re_qp *qp)
3383 {
3384 	struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
3385 	unsigned long flags;
3386 	int rc = 0;
3387 
3388 	spin_lock_irqsave(&qp->sq_lock, flags);
3389 
3390 	rc = bnxt_re_bind_fence_mw(lib_qp);
3391 	if (!rc) {
3392 		lib_qp->sq.phantom_wqe_cnt++;
3393 		ibdev_dbg(&qp->rdev->ibdev,
3394 			  "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
3395 			  lib_qp->id, lib_qp->sq.hwq.prod,
3396 			  HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
3397 			  lib_qp->sq.phantom_wqe_cnt);
3398 	}
3399 
3400 	spin_unlock_irqrestore(&qp->sq_lock, flags);
3401 	return rc;
3402 }
3403 
3404 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
3405 {
3406 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3407 	struct bnxt_re_qp *qp, *sh_qp;
3408 	struct bnxt_qplib_cqe *cqe;
3409 	int i, ncqe, budget;
3410 	struct bnxt_qplib_q *sq;
3411 	struct bnxt_qplib_qp *lib_qp;
3412 	u32 tbl_idx;
3413 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3414 	unsigned long flags;
3415 
3416 	spin_lock_irqsave(&cq->cq_lock, flags);
3417 	budget = min_t(u32, num_entries, cq->max_cql);
3418 	num_entries = budget;
3419 	if (!cq->cql) {
3420 		ibdev_err(&cq->rdev->ibdev, "POLL CQ : no CQL to use");
3421 		goto exit;
3422 	}
3423 	cqe = &cq->cql[0];
3424 	while (budget) {
3425 		lib_qp = NULL;
3426 		ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
3427 		if (lib_qp) {
3428 			sq = &lib_qp->sq;
3429 			if (sq->send_phantom) {
3430 				qp = container_of(lib_qp,
3431 						  struct bnxt_re_qp, qplib_qp);
3432 				if (send_phantom_wqe(qp) == -ENOMEM)
3433 					ibdev_err(&cq->rdev->ibdev,
3434 						  "Phantom failed! Scheduled to send again\n");
3435 				else
3436 					sq->send_phantom = false;
3437 			}
3438 		}
3439 		if (ncqe < budget)
3440 			ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq,
3441 							      cqe + ncqe,
3442 							      budget - ncqe);
3443 
3444 		if (!ncqe)
3445 			break;
3446 
3447 		for (i = 0; i < ncqe; i++, cqe++) {
3448 			/* Transcribe each qplib_wqe back to ib_wc */
3449 			memset(wc, 0, sizeof(*wc));
3450 
3451 			wc->wr_id = cqe->wr_id;
3452 			wc->byte_len = cqe->length;
3453 			qp = container_of
3454 				((struct bnxt_qplib_qp *)
3455 				 (unsigned long)(cqe->qp_handle),
3456 				 struct bnxt_re_qp, qplib_qp);
3457 			wc->qp = &qp->ib_qp;
3458 			wc->ex.imm_data = cqe->immdata;
3459 			wc->src_qp = cqe->src_qp;
3460 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
3461 			wc->port_num = 1;
3462 			wc->vendor_err = cqe->status;
3463 
3464 			switch (cqe->opcode) {
3465 			case CQ_BASE_CQE_TYPE_REQ:
3466 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3467 				if (sh_qp &&
3468 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3469 					/* Handle this completion with
3470 					 * the stored completion
3471 					 */
3472 					memset(wc, 0, sizeof(*wc));
3473 					continue;
3474 				}
3475 				bnxt_re_process_req_wc(wc, cqe);
3476 				break;
3477 			case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1:
3478 				if (!cqe->status) {
3479 					int rc = 0;
3480 
3481 					rc = bnxt_re_process_raw_qp_pkt_rx
3482 								(qp, cqe);
3483 					if (!rc) {
3484 						memset(wc, 0, sizeof(*wc));
3485 						continue;
3486 					}
3487 					cqe->status = -1;
3488 				}
3489 				/* Errors need not be looped back.
3490 				 * But change the wr_id to the one
3491 				 * stored in the table
3492 				 */
3493 				tbl_idx = cqe->wr_id;
3494 				sqp_entry = &cq->rdev->gsi_ctx.sqp_tbl[tbl_idx];
3495 				wc->wr_id = sqp_entry->wrid;
3496 				bnxt_re_process_res_rawqp1_wc(wc, cqe);
3497 				break;
3498 			case CQ_BASE_CQE_TYPE_RES_RC:
3499 				bnxt_re_process_res_rc_wc(wc, cqe);
3500 				break;
3501 			case CQ_BASE_CQE_TYPE_RES_UD:
3502 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3503 				if (sh_qp &&
3504 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3505 					/* Handle this completion with
3506 					 * the stored completion
3507 					 */
3508 					if (cqe->status) {
3509 						continue;
3510 					} else {
3511 						bnxt_re_process_res_shadow_qp_wc
3512 								(qp, wc, cqe);
3513 						break;
3514 					}
3515 				}
3516 				bnxt_re_process_res_ud_wc(qp, wc, cqe);
3517 				break;
3518 			default:
3519 				ibdev_err(&cq->rdev->ibdev,
3520 					  "POLL CQ : type 0x%x not handled",
3521 					  cqe->opcode);
3522 				continue;
3523 			}
3524 			wc++;
3525 			budget--;
3526 		}
3527 	}
3528 exit:
3529 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3530 	return num_entries - budget;
3531 }
3532 
3533 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq,
3534 			  enum ib_cq_notify_flags ib_cqn_flags)
3535 {
3536 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3537 	int type = 0, rc = 0;
3538 	unsigned long flags;
3539 
3540 	spin_lock_irqsave(&cq->cq_lock, flags);
3541 	/* Trigger on the very next completion */
3542 	if (ib_cqn_flags & IB_CQ_NEXT_COMP)
3543 		type = DBC_DBC_TYPE_CQ_ARMALL;
3544 	/* Trigger on the next solicited completion */
3545 	else if (ib_cqn_flags & IB_CQ_SOLICITED)
3546 		type = DBC_DBC_TYPE_CQ_ARMSE;
3547 
3548 	/* Poll to see if there are missed events */
3549 	if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3550 	    !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) {
3551 		rc = 1;
3552 		goto exit;
3553 	}
3554 	bnxt_qplib_req_notify_cq(&cq->qplib_cq, type);
3555 
3556 exit:
3557 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3558 	return rc;
3559 }
3560 
3561 /* Memory Regions */
3562 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags)
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;
3567 	int rc;
3568 
3569 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3570 	if (!mr)
3571 		return ERR_PTR(-ENOMEM);
3572 
3573 	mr->rdev = rdev;
3574 	mr->qplib_mr.pd = &pd->qplib_pd;
3575 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3576 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3577 
3578 	/* Allocate and register 0 as the address */
3579 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3580 	if (rc)
3581 		goto fail;
3582 
3583 	mr->qplib_mr.hwq.level = PBL_LVL_MAX;
3584 	mr->qplib_mr.total_size = -1; /* Infinte length */
3585 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL, 0,
3586 			       PAGE_SIZE);
3587 	if (rc)
3588 		goto fail_mr;
3589 
3590 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3591 	if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ |
3592 			       IB_ACCESS_REMOTE_ATOMIC))
3593 		mr->ib_mr.rkey = mr->ib_mr.lkey;
3594 	atomic_inc(&rdev->mr_count);
3595 
3596 	return &mr->ib_mr;
3597 
3598 fail_mr:
3599 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3600 fail:
3601 	kfree(mr);
3602 	return ERR_PTR(rc);
3603 }
3604 
3605 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3606 {
3607 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3608 	struct bnxt_re_dev *rdev = mr->rdev;
3609 	int rc;
3610 
3611 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3612 	if (rc) {
3613 		ibdev_err(&rdev->ibdev, "Dereg MR failed: %#x\n", rc);
3614 		return rc;
3615 	}
3616 
3617 	if (mr->pages) {
3618 		rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
3619 							&mr->qplib_frpl);
3620 		kfree(mr->pages);
3621 		mr->npages = 0;
3622 		mr->pages = NULL;
3623 	}
3624 	ib_umem_release(mr->ib_umem);
3625 
3626 	kfree(mr);
3627 	atomic_dec(&rdev->mr_count);
3628 	return rc;
3629 }
3630 
3631 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr)
3632 {
3633 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3634 
3635 	if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs))
3636 		return -ENOMEM;
3637 
3638 	mr->pages[mr->npages++] = addr;
3639 	return 0;
3640 }
3641 
3642 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents,
3643 		      unsigned int *sg_offset)
3644 {
3645 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3646 
3647 	mr->npages = 0;
3648 	return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page);
3649 }
3650 
3651 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type,
3652 			       u32 max_num_sg)
3653 {
3654 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3655 	struct bnxt_re_dev *rdev = pd->rdev;
3656 	struct bnxt_re_mr *mr = NULL;
3657 	int rc;
3658 
3659 	if (type != IB_MR_TYPE_MEM_REG) {
3660 		ibdev_dbg(&rdev->ibdev, "MR type 0x%x not supported", type);
3661 		return ERR_PTR(-EINVAL);
3662 	}
3663 	if (max_num_sg > MAX_PBL_LVL_1_PGS)
3664 		return ERR_PTR(-EINVAL);
3665 
3666 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3667 	if (!mr)
3668 		return ERR_PTR(-ENOMEM);
3669 
3670 	mr->rdev = rdev;
3671 	mr->qplib_mr.pd = &pd->qplib_pd;
3672 	mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR;
3673 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3674 
3675 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3676 	if (rc)
3677 		goto bail;
3678 
3679 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3680 	mr->ib_mr.rkey = mr->ib_mr.lkey;
3681 
3682 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3683 	if (!mr->pages) {
3684 		rc = -ENOMEM;
3685 		goto fail;
3686 	}
3687 	rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res,
3688 						 &mr->qplib_frpl, max_num_sg);
3689 	if (rc) {
3690 		ibdev_err(&rdev->ibdev,
3691 			  "Failed to allocate HW FR page list");
3692 		goto fail_mr;
3693 	}
3694 
3695 	atomic_inc(&rdev->mr_count);
3696 	return &mr->ib_mr;
3697 
3698 fail_mr:
3699 	kfree(mr->pages);
3700 fail:
3701 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3702 bail:
3703 	kfree(mr);
3704 	return ERR_PTR(rc);
3705 }
3706 
3707 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
3708 			       struct ib_udata *udata)
3709 {
3710 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3711 	struct bnxt_re_dev *rdev = pd->rdev;
3712 	struct bnxt_re_mw *mw;
3713 	int rc;
3714 
3715 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
3716 	if (!mw)
3717 		return ERR_PTR(-ENOMEM);
3718 	mw->rdev = rdev;
3719 	mw->qplib_mw.pd = &pd->qplib_pd;
3720 
3721 	mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
3722 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
3723 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
3724 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
3725 	if (rc) {
3726 		ibdev_err(&rdev->ibdev, "Allocate MW failed!");
3727 		goto fail;
3728 	}
3729 	mw->ib_mw.rkey = mw->qplib_mw.rkey;
3730 
3731 	atomic_inc(&rdev->mw_count);
3732 	return &mw->ib_mw;
3733 
3734 fail:
3735 	kfree(mw);
3736 	return ERR_PTR(rc);
3737 }
3738 
3739 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
3740 {
3741 	struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
3742 	struct bnxt_re_dev *rdev = mw->rdev;
3743 	int rc;
3744 
3745 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
3746 	if (rc) {
3747 		ibdev_err(&rdev->ibdev, "Free MW failed: %#x\n", rc);
3748 		return rc;
3749 	}
3750 
3751 	kfree(mw);
3752 	atomic_dec(&rdev->mw_count);
3753 	return rc;
3754 }
3755 
3756 /* uverbs */
3757 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
3758 				  u64 virt_addr, int mr_access_flags,
3759 				  struct ib_udata *udata)
3760 {
3761 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3762 	struct bnxt_re_dev *rdev = pd->rdev;
3763 	struct bnxt_re_mr *mr;
3764 	struct ib_umem *umem;
3765 	unsigned long page_size;
3766 	int umem_pgs, rc;
3767 
3768 	if (length > BNXT_RE_MAX_MR_SIZE) {
3769 		ibdev_err(&rdev->ibdev, "MR Size: %lld > Max supported:%lld\n",
3770 			  length, BNXT_RE_MAX_MR_SIZE);
3771 		return ERR_PTR(-ENOMEM);
3772 	}
3773 
3774 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3775 	if (!mr)
3776 		return ERR_PTR(-ENOMEM);
3777 
3778 	mr->rdev = rdev;
3779 	mr->qplib_mr.pd = &pd->qplib_pd;
3780 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3781 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
3782 
3783 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3784 	if (rc) {
3785 		ibdev_err(&rdev->ibdev, "Failed to allocate MR");
3786 		goto free_mr;
3787 	}
3788 	/* The fixed portion of the rkey is the same as the lkey */
3789 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
3790 
3791 	umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags);
3792 	if (IS_ERR(umem)) {
3793 		ibdev_err(&rdev->ibdev, "Failed to get umem");
3794 		rc = -EFAULT;
3795 		goto free_mrw;
3796 	}
3797 	mr->ib_umem = umem;
3798 
3799 	mr->qplib_mr.va = virt_addr;
3800 	page_size = ib_umem_find_best_pgsz(
3801 		umem, BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M, virt_addr);
3802 	if (!page_size) {
3803 		ibdev_err(&rdev->ibdev, "umem page size unsupported!");
3804 		rc = -EFAULT;
3805 		goto free_umem;
3806 	}
3807 	mr->qplib_mr.total_size = length;
3808 
3809 	umem_pgs = ib_umem_num_dma_blocks(umem, page_size);
3810 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, umem,
3811 			       umem_pgs, page_size);
3812 	if (rc) {
3813 		ibdev_err(&rdev->ibdev, "Failed to register user MR");
3814 		goto free_umem;
3815 	}
3816 
3817 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3818 	mr->ib_mr.rkey = mr->qplib_mr.lkey;
3819 	atomic_inc(&rdev->mr_count);
3820 
3821 	return &mr->ib_mr;
3822 free_umem:
3823 	ib_umem_release(umem);
3824 free_mrw:
3825 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3826 free_mr:
3827 	kfree(mr);
3828 	return ERR_PTR(rc);
3829 }
3830 
3831 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
3832 {
3833 	struct ib_device *ibdev = ctx->device;
3834 	struct bnxt_re_ucontext *uctx =
3835 		container_of(ctx, struct bnxt_re_ucontext, ib_uctx);
3836 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
3837 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
3838 	struct bnxt_re_uctx_resp resp = {};
3839 	u32 chip_met_rev_num = 0;
3840 	int rc;
3841 
3842 	ibdev_dbg(ibdev, "ABI version requested %u", ibdev->ops.uverbs_abi_ver);
3843 
3844 	if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) {
3845 		ibdev_dbg(ibdev, " is different from the device %d ",
3846 			  BNXT_RE_ABI_VERSION);
3847 		return -EPERM;
3848 	}
3849 
3850 	uctx->rdev = rdev;
3851 
3852 	uctx->shpg = (void *)__get_free_page(GFP_KERNEL);
3853 	if (!uctx->shpg) {
3854 		rc = -ENOMEM;
3855 		goto fail;
3856 	}
3857 	spin_lock_init(&uctx->sh_lock);
3858 
3859 	resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX;
3860 	chip_met_rev_num = rdev->chip_ctx->chip_num;
3861 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_rev & 0xFF) <<
3862 			     BNXT_RE_CHIP_ID0_CHIP_REV_SFT;
3863 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_metal & 0xFF) <<
3864 			     BNXT_RE_CHIP_ID0_CHIP_MET_SFT;
3865 	resp.chip_id0 = chip_met_rev_num;
3866 	/*Temp, Use xa_alloc instead */
3867 	resp.dev_id = rdev->en_dev->pdev->devfn;
3868 	resp.max_qp = rdev->qplib_ctx.qpc_count;
3869 	resp.pg_size = PAGE_SIZE;
3870 	resp.cqe_sz = sizeof(struct cq_base);
3871 	resp.max_cqd = dev_attr->max_cq_wqes;
3872 
3873 	resp.comp_mask |= BNXT_RE_UCNTX_CMASK_HAVE_MODE;
3874 	resp.mode = rdev->chip_ctx->modes.wqe_mode;
3875 
3876 	rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
3877 	if (rc) {
3878 		ibdev_err(ibdev, "Failed to copy user context");
3879 		rc = -EFAULT;
3880 		goto cfail;
3881 	}
3882 
3883 	return 0;
3884 cfail:
3885 	free_page((unsigned long)uctx->shpg);
3886 	uctx->shpg = NULL;
3887 fail:
3888 	return rc;
3889 }
3890 
3891 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
3892 {
3893 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3894 						   struct bnxt_re_ucontext,
3895 						   ib_uctx);
3896 
3897 	struct bnxt_re_dev *rdev = uctx->rdev;
3898 
3899 	if (uctx->shpg)
3900 		free_page((unsigned long)uctx->shpg);
3901 
3902 	if (uctx->dpi.dbr) {
3903 		/* Free DPI only if this is the first PD allocated by the
3904 		 * application and mark the context dpi as NULL
3905 		 */
3906 		bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
3907 				       &rdev->qplib_res.dpi_tbl, &uctx->dpi);
3908 		uctx->dpi.dbr = NULL;
3909 	}
3910 }
3911 
3912 /* Helper function to mmap the virtual memory from user app */
3913 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
3914 {
3915 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3916 						   struct bnxt_re_ucontext,
3917 						   ib_uctx);
3918 	struct bnxt_re_dev *rdev = uctx->rdev;
3919 	u64 pfn;
3920 
3921 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3922 		return -EINVAL;
3923 
3924 	if (vma->vm_pgoff) {
3925 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3926 		if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
3927 				       PAGE_SIZE, vma->vm_page_prot)) {
3928 			ibdev_err(&rdev->ibdev, "Failed to map DPI");
3929 			return -EAGAIN;
3930 		}
3931 	} else {
3932 		pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT;
3933 		if (remap_pfn_range(vma, vma->vm_start,
3934 				    pfn, PAGE_SIZE, vma->vm_page_prot)) {
3935 			ibdev_err(&rdev->ibdev, "Failed to map shared page");
3936 			return -EAGAIN;
3937 		}
3938 	}
3939 
3940 	return 0;
3941 }
3942