xref: /openbmc/linux/drivers/infiniband/hw/bnxt_re/ib_verbs.c (revision 45cc842d5b75ba8f9a958f2dd12b95c6dd0452bd)
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 
52 #include "bnxt_ulp.h"
53 
54 #include "roce_hsi.h"
55 #include "qplib_res.h"
56 #include "qplib_sp.h"
57 #include "qplib_fp.h"
58 #include "qplib_rcfw.h"
59 
60 #include "bnxt_re.h"
61 #include "ib_verbs.h"
62 #include <rdma/bnxt_re-abi.h>
63 
64 static int __from_ib_access_flags(int iflags)
65 {
66 	int qflags = 0;
67 
68 	if (iflags & IB_ACCESS_LOCAL_WRITE)
69 		qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
70 	if (iflags & IB_ACCESS_REMOTE_READ)
71 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ;
72 	if (iflags & IB_ACCESS_REMOTE_WRITE)
73 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE;
74 	if (iflags & IB_ACCESS_REMOTE_ATOMIC)
75 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC;
76 	if (iflags & IB_ACCESS_MW_BIND)
77 		qflags |= BNXT_QPLIB_ACCESS_MW_BIND;
78 	if (iflags & IB_ZERO_BASED)
79 		qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED;
80 	if (iflags & IB_ACCESS_ON_DEMAND)
81 		qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND;
82 	return qflags;
83 };
84 
85 static enum ib_access_flags __to_ib_access_flags(int qflags)
86 {
87 	enum ib_access_flags iflags = 0;
88 
89 	if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE)
90 		iflags |= IB_ACCESS_LOCAL_WRITE;
91 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE)
92 		iflags |= IB_ACCESS_REMOTE_WRITE;
93 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ)
94 		iflags |= IB_ACCESS_REMOTE_READ;
95 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC)
96 		iflags |= IB_ACCESS_REMOTE_ATOMIC;
97 	if (qflags & BNXT_QPLIB_ACCESS_MW_BIND)
98 		iflags |= IB_ACCESS_MW_BIND;
99 	if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED)
100 		iflags |= IB_ZERO_BASED;
101 	if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND)
102 		iflags |= IB_ACCESS_ON_DEMAND;
103 	return iflags;
104 };
105 
106 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list,
107 			     struct bnxt_qplib_sge *sg_list, int num)
108 {
109 	int i, total = 0;
110 
111 	for (i = 0; i < num; i++) {
112 		sg_list[i].addr = ib_sg_list[i].addr;
113 		sg_list[i].lkey = ib_sg_list[i].lkey;
114 		sg_list[i].size = ib_sg_list[i].length;
115 		total += sg_list[i].size;
116 	}
117 	return total;
118 }
119 
120 /* Device */
121 struct net_device *bnxt_re_get_netdev(struct ib_device *ibdev, u8 port_num)
122 {
123 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
124 	struct net_device *netdev = NULL;
125 
126 	rcu_read_lock();
127 	if (rdev)
128 		netdev = rdev->netdev;
129 	if (netdev)
130 		dev_hold(netdev);
131 
132 	rcu_read_unlock();
133 	return netdev;
134 }
135 
136 int bnxt_re_query_device(struct ib_device *ibdev,
137 			 struct ib_device_attr *ib_attr,
138 			 struct ib_udata *udata)
139 {
140 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
141 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
142 
143 	memset(ib_attr, 0, sizeof(*ib_attr));
144 	memcpy(&ib_attr->fw_ver, dev_attr->fw_ver,
145 	       min(sizeof(dev_attr->fw_ver),
146 		   sizeof(ib_attr->fw_ver)));
147 	bnxt_qplib_get_guid(rdev->netdev->dev_addr,
148 			    (u8 *)&ib_attr->sys_image_guid);
149 	ib_attr->max_mr_size = BNXT_RE_MAX_MR_SIZE;
150 	ib_attr->page_size_cap = BNXT_RE_PAGE_SIZE_4K | BNXT_RE_PAGE_SIZE_2M;
151 
152 	ib_attr->vendor_id = rdev->en_dev->pdev->vendor;
153 	ib_attr->vendor_part_id = rdev->en_dev->pdev->device;
154 	ib_attr->hw_ver = rdev->en_dev->pdev->subsystem_device;
155 	ib_attr->max_qp = dev_attr->max_qp;
156 	ib_attr->max_qp_wr = dev_attr->max_qp_wqes;
157 	ib_attr->device_cap_flags =
158 				    IB_DEVICE_CURR_QP_STATE_MOD
159 				    | IB_DEVICE_RC_RNR_NAK_GEN
160 				    | IB_DEVICE_SHUTDOWN_PORT
161 				    | IB_DEVICE_SYS_IMAGE_GUID
162 				    | IB_DEVICE_LOCAL_DMA_LKEY
163 				    | IB_DEVICE_RESIZE_MAX_WR
164 				    | IB_DEVICE_PORT_ACTIVE_EVENT
165 				    | IB_DEVICE_N_NOTIFY_CQ
166 				    | IB_DEVICE_MEM_WINDOW
167 				    | IB_DEVICE_MEM_WINDOW_TYPE_2B
168 				    | IB_DEVICE_MEM_MGT_EXTENSIONS;
169 	ib_attr->max_sge = dev_attr->max_qp_sges;
170 	ib_attr->max_sge_rd = dev_attr->max_qp_sges;
171 	ib_attr->max_cq = dev_attr->max_cq;
172 	ib_attr->max_cqe = dev_attr->max_cq_wqes;
173 	ib_attr->max_mr = dev_attr->max_mr;
174 	ib_attr->max_pd = dev_attr->max_pd;
175 	ib_attr->max_qp_rd_atom = dev_attr->max_qp_rd_atom;
176 	ib_attr->max_qp_init_rd_atom = dev_attr->max_qp_init_rd_atom;
177 	if (dev_attr->is_atomic) {
178 		ib_attr->atomic_cap = IB_ATOMIC_HCA;
179 		ib_attr->masked_atomic_cap = IB_ATOMIC_HCA;
180 	}
181 
182 	ib_attr->max_ee_rd_atom = 0;
183 	ib_attr->max_res_rd_atom = 0;
184 	ib_attr->max_ee_init_rd_atom = 0;
185 	ib_attr->max_ee = 0;
186 	ib_attr->max_rdd = 0;
187 	ib_attr->max_mw = dev_attr->max_mw;
188 	ib_attr->max_raw_ipv6_qp = 0;
189 	ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp;
190 	ib_attr->max_mcast_grp = 0;
191 	ib_attr->max_mcast_qp_attach = 0;
192 	ib_attr->max_total_mcast_qp_attach = 0;
193 	ib_attr->max_ah = dev_attr->max_ah;
194 
195 	ib_attr->max_fmr = 0;
196 	ib_attr->max_map_per_fmr = 0;
197 
198 	ib_attr->max_srq = dev_attr->max_srq;
199 	ib_attr->max_srq_wr = dev_attr->max_srq_wqes;
200 	ib_attr->max_srq_sge = dev_attr->max_srq_sges;
201 
202 	ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS;
203 
204 	ib_attr->max_pkeys = 1;
205 	ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY;
206 	return 0;
207 }
208 
209 int bnxt_re_modify_device(struct ib_device *ibdev,
210 			  int device_modify_mask,
211 			  struct ib_device_modify *device_modify)
212 {
213 	switch (device_modify_mask) {
214 	case IB_DEVICE_MODIFY_SYS_IMAGE_GUID:
215 		/* Modify the GUID requires the modification of the GID table */
216 		/* GUID should be made as READ-ONLY */
217 		break;
218 	case IB_DEVICE_MODIFY_NODE_DESC:
219 		/* Node Desc should be made as READ-ONLY */
220 		break;
221 	default:
222 		break;
223 	}
224 	return 0;
225 }
226 
227 /* Port */
228 int bnxt_re_query_port(struct ib_device *ibdev, u8 port_num,
229 		       struct ib_port_attr *port_attr)
230 {
231 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
232 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
233 
234 	memset(port_attr, 0, sizeof(*port_attr));
235 
236 	if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) {
237 		port_attr->state = IB_PORT_ACTIVE;
238 		port_attr->phys_state = 5;
239 	} else {
240 		port_attr->state = IB_PORT_DOWN;
241 		port_attr->phys_state = 3;
242 	}
243 	port_attr->max_mtu = IB_MTU_4096;
244 	port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu);
245 	port_attr->gid_tbl_len = dev_attr->max_sgid;
246 	port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
247 				    IB_PORT_DEVICE_MGMT_SUP |
248 				    IB_PORT_VENDOR_CLASS_SUP |
249 				    IB_PORT_IP_BASED_GIDS;
250 
251 	port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW;
252 	port_attr->bad_pkey_cntr = 0;
253 	port_attr->qkey_viol_cntr = 0;
254 	port_attr->pkey_tbl_len = dev_attr->max_pkey;
255 	port_attr->lid = 0;
256 	port_attr->sm_lid = 0;
257 	port_attr->lmc = 0;
258 	port_attr->max_vl_num = 4;
259 	port_attr->sm_sl = 0;
260 	port_attr->subnet_timeout = 0;
261 	port_attr->init_type_reply = 0;
262 	port_attr->active_speed = rdev->active_speed;
263 	port_attr->active_width = rdev->active_width;
264 
265 	return 0;
266 }
267 
268 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u8 port_num,
269 			       struct ib_port_immutable *immutable)
270 {
271 	struct ib_port_attr port_attr;
272 
273 	if (bnxt_re_query_port(ibdev, port_num, &port_attr))
274 		return -EINVAL;
275 
276 	immutable->pkey_tbl_len = port_attr.pkey_tbl_len;
277 	immutable->gid_tbl_len = port_attr.gid_tbl_len;
278 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
279 	immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP;
280 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
281 	return 0;
282 }
283 
284 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str)
285 {
286 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
287 
288 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d",
289 		 rdev->dev_attr.fw_ver[0], rdev->dev_attr.fw_ver[1],
290 		 rdev->dev_attr.fw_ver[2], rdev->dev_attr.fw_ver[3]);
291 }
292 
293 int bnxt_re_query_pkey(struct ib_device *ibdev, u8 port_num,
294 		       u16 index, u16 *pkey)
295 {
296 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
297 
298 	/* Ignore port_num */
299 
300 	memset(pkey, 0, sizeof(*pkey));
301 	return bnxt_qplib_get_pkey(&rdev->qplib_res,
302 				   &rdev->qplib_res.pkey_tbl, index, pkey);
303 }
304 
305 int bnxt_re_query_gid(struct ib_device *ibdev, u8 port_num,
306 		      int index, union ib_gid *gid)
307 {
308 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
309 	int rc = 0;
310 
311 	/* Ignore port_num */
312 	memset(gid, 0, sizeof(*gid));
313 	rc = bnxt_qplib_get_sgid(&rdev->qplib_res,
314 				 &rdev->qplib_res.sgid_tbl, index,
315 				 (struct bnxt_qplib_gid *)gid);
316 	return rc;
317 }
318 
319 int bnxt_re_del_gid(struct ib_device *ibdev, u8 port_num,
320 		    unsigned int index, void **context)
321 {
322 	int rc = 0;
323 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
324 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
325 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
326 	struct bnxt_qplib_gid *gid_to_del;
327 
328 	/* Delete the entry from the hardware */
329 	ctx = *context;
330 	if (!ctx)
331 		return -EINVAL;
332 
333 	if (sgid_tbl && sgid_tbl->active) {
334 		if (ctx->idx >= sgid_tbl->max)
335 			return -EINVAL;
336 		gid_to_del = &sgid_tbl->tbl[ctx->idx];
337 		/* DEL_GID is called in WQ context(netdevice_event_work_handler)
338 		 * or via the ib_unregister_device path. In the former case QP1
339 		 * may not be destroyed yet, in which case just return as FW
340 		 * needs that entry to be present and will fail it's deletion.
341 		 * We could get invoked again after QP1 is destroyed OR get an
342 		 * ADD_GID call with a different GID value for the same index
343 		 * where we issue MODIFY_GID cmd to update the GID entry -- TBD
344 		 */
345 		if (ctx->idx == 0 &&
346 		    rdma_link_local_addr((struct in6_addr *)gid_to_del) &&
347 		    ctx->refcnt == 1 && rdev->qp1_sqp) {
348 			dev_dbg(rdev_to_dev(rdev),
349 				"Trying to delete GID0 while QP1 is alive\n");
350 			return -EFAULT;
351 		}
352 		ctx->refcnt--;
353 		if (!ctx->refcnt) {
354 			rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del, true);
355 			if (rc) {
356 				dev_err(rdev_to_dev(rdev),
357 					"Failed to remove GID: %#x", rc);
358 			} else {
359 				ctx_tbl = sgid_tbl->ctx;
360 				ctx_tbl[ctx->idx] = NULL;
361 				kfree(ctx);
362 			}
363 		}
364 	} else {
365 		return -EINVAL;
366 	}
367 	return rc;
368 }
369 
370 int bnxt_re_add_gid(struct ib_device *ibdev, u8 port_num,
371 		    unsigned int index, const union ib_gid *gid,
372 		    const struct ib_gid_attr *attr, void **context)
373 {
374 	int rc;
375 	u32 tbl_idx = 0;
376 	u16 vlan_id = 0xFFFF;
377 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
378 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
379 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
380 
381 	if ((attr->ndev) && is_vlan_dev(attr->ndev))
382 		vlan_id = vlan_dev_vlan_id(attr->ndev);
383 
384 	rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)gid,
385 				 rdev->qplib_res.netdev->dev_addr,
386 				 vlan_id, true, &tbl_idx);
387 	if (rc == -EALREADY) {
388 		ctx_tbl = sgid_tbl->ctx;
389 		ctx_tbl[tbl_idx]->refcnt++;
390 		*context = ctx_tbl[tbl_idx];
391 		return 0;
392 	}
393 
394 	if (rc < 0) {
395 		dev_err(rdev_to_dev(rdev), "Failed to add GID: %#x", rc);
396 		return rc;
397 	}
398 
399 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
400 	if (!ctx)
401 		return -ENOMEM;
402 	ctx_tbl = sgid_tbl->ctx;
403 	ctx->idx = tbl_idx;
404 	ctx->refcnt = 1;
405 	ctx_tbl[tbl_idx] = ctx;
406 	*context = ctx;
407 
408 	return rc;
409 }
410 
411 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
412 					    u8 port_num)
413 {
414 	return IB_LINK_LAYER_ETHERNET;
415 }
416 
417 #define	BNXT_RE_FENCE_PBL_SIZE	DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
418 
419 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd)
420 {
421 	struct bnxt_re_fence_data *fence = &pd->fence;
422 	struct ib_mr *ib_mr = &fence->mr->ib_mr;
423 	struct bnxt_qplib_swqe *wqe = &fence->bind_wqe;
424 
425 	memset(wqe, 0, sizeof(*wqe));
426 	wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW;
427 	wqe->wr_id = BNXT_QPLIB_FENCE_WRID;
428 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
429 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
430 	wqe->bind.zero_based = false;
431 	wqe->bind.parent_l_key = ib_mr->lkey;
432 	wqe->bind.va = (u64)(unsigned long)fence->va;
433 	wqe->bind.length = fence->size;
434 	wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ);
435 	wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1;
436 
437 	/* Save the initial rkey in fence structure for now;
438 	 * wqe->bind.r_key will be set at (re)bind time.
439 	 */
440 	fence->bind_rkey = ib_inc_rkey(fence->mw->rkey);
441 }
442 
443 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp)
444 {
445 	struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp,
446 					     qplib_qp);
447 	struct ib_pd *ib_pd = qp->ib_qp.pd;
448 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
449 	struct bnxt_re_fence_data *fence = &pd->fence;
450 	struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe;
451 	struct bnxt_qplib_swqe wqe;
452 	int rc;
453 
454 	memcpy(&wqe, fence_wqe, sizeof(wqe));
455 	wqe.bind.r_key = fence->bind_rkey;
456 	fence->bind_rkey = ib_inc_rkey(fence->bind_rkey);
457 
458 	dev_dbg(rdev_to_dev(qp->rdev),
459 		"Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
460 		wqe.bind.r_key, qp->qplib_qp.id, pd);
461 	rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
462 	if (rc) {
463 		dev_err(rdev_to_dev(qp->rdev), "Failed to bind fence-WQE\n");
464 		return rc;
465 	}
466 	bnxt_qplib_post_send_db(&qp->qplib_qp);
467 
468 	return rc;
469 }
470 
471 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd)
472 {
473 	struct bnxt_re_fence_data *fence = &pd->fence;
474 	struct bnxt_re_dev *rdev = pd->rdev;
475 	struct device *dev = &rdev->en_dev->pdev->dev;
476 	struct bnxt_re_mr *mr = fence->mr;
477 
478 	if (fence->mw) {
479 		bnxt_re_dealloc_mw(fence->mw);
480 		fence->mw = NULL;
481 	}
482 	if (mr) {
483 		if (mr->ib_mr.rkey)
484 			bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr,
485 					     true);
486 		if (mr->ib_mr.lkey)
487 			bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
488 		kfree(mr);
489 		fence->mr = NULL;
490 	}
491 	if (fence->dma_addr) {
492 		dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES,
493 				 DMA_BIDIRECTIONAL);
494 		fence->dma_addr = 0;
495 	}
496 }
497 
498 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
499 {
500 	int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND;
501 	struct bnxt_re_fence_data *fence = &pd->fence;
502 	struct bnxt_re_dev *rdev = pd->rdev;
503 	struct device *dev = &rdev->en_dev->pdev->dev;
504 	struct bnxt_re_mr *mr = NULL;
505 	dma_addr_t dma_addr = 0;
506 	struct ib_mw *mw;
507 	u64 pbl_tbl;
508 	int rc;
509 
510 	dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES,
511 				  DMA_BIDIRECTIONAL);
512 	rc = dma_mapping_error(dev, dma_addr);
513 	if (rc) {
514 		dev_err(rdev_to_dev(rdev), "Failed to dma-map fence-MR-mem\n");
515 		rc = -EIO;
516 		fence->dma_addr = 0;
517 		goto fail;
518 	}
519 	fence->dma_addr = dma_addr;
520 
521 	/* Allocate a MR */
522 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
523 	if (!mr) {
524 		rc = -ENOMEM;
525 		goto fail;
526 	}
527 	fence->mr = mr;
528 	mr->rdev = rdev;
529 	mr->qplib_mr.pd = &pd->qplib_pd;
530 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
531 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
532 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
533 	if (rc) {
534 		dev_err(rdev_to_dev(rdev), "Failed to alloc fence-HW-MR\n");
535 		goto fail;
536 	}
537 
538 	/* Register MR */
539 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
540 	mr->qplib_mr.va = (u64)(unsigned long)fence->va;
541 	mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES;
542 	pbl_tbl = dma_addr;
543 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl_tbl,
544 			       BNXT_RE_FENCE_PBL_SIZE, false, PAGE_SIZE);
545 	if (rc) {
546 		dev_err(rdev_to_dev(rdev), "Failed to register fence-MR\n");
547 		goto fail;
548 	}
549 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
550 
551 	/* Create a fence MW only for kernel consumers */
552 	mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL);
553 	if (IS_ERR(mw)) {
554 		dev_err(rdev_to_dev(rdev),
555 			"Failed to create fence-MW for PD: %p\n", pd);
556 		rc = PTR_ERR(mw);
557 		goto fail;
558 	}
559 	fence->mw = mw;
560 
561 	bnxt_re_create_fence_wqe(pd);
562 	return 0;
563 
564 fail:
565 	bnxt_re_destroy_fence_mr(pd);
566 	return rc;
567 }
568 
569 /* Protection Domains */
570 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd)
571 {
572 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
573 	struct bnxt_re_dev *rdev = pd->rdev;
574 	int rc;
575 
576 	bnxt_re_destroy_fence_mr(pd);
577 
578 	if (pd->qplib_pd.id) {
579 		rc = bnxt_qplib_dealloc_pd(&rdev->qplib_res,
580 					   &rdev->qplib_res.pd_tbl,
581 					   &pd->qplib_pd);
582 		if (rc)
583 			dev_err(rdev_to_dev(rdev), "Failed to deallocate HW PD");
584 	}
585 
586 	kfree(pd);
587 	return 0;
588 }
589 
590 struct ib_pd *bnxt_re_alloc_pd(struct ib_device *ibdev,
591 			       struct ib_ucontext *ucontext,
592 			       struct ib_udata *udata)
593 {
594 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
595 	struct bnxt_re_ucontext *ucntx = container_of(ucontext,
596 						      struct bnxt_re_ucontext,
597 						      ib_uctx);
598 	struct bnxt_re_pd *pd;
599 	int rc;
600 
601 	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
602 	if (!pd)
603 		return ERR_PTR(-ENOMEM);
604 
605 	pd->rdev = rdev;
606 	if (bnxt_qplib_alloc_pd(&rdev->qplib_res.pd_tbl, &pd->qplib_pd)) {
607 		dev_err(rdev_to_dev(rdev), "Failed to allocate HW PD");
608 		rc = -ENOMEM;
609 		goto fail;
610 	}
611 
612 	if (udata) {
613 		struct bnxt_re_pd_resp resp;
614 
615 		if (!ucntx->dpi.dbr) {
616 			/* Allocate DPI in alloc_pd to avoid failing of
617 			 * ibv_devinfo and family of application when DPIs
618 			 * are depleted.
619 			 */
620 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res.dpi_tbl,
621 						 &ucntx->dpi, ucntx)) {
622 				rc = -ENOMEM;
623 				goto dbfail;
624 			}
625 		}
626 
627 		resp.pdid = pd->qplib_pd.id;
628 		/* Still allow mapping this DBR to the new user PD. */
629 		resp.dpi = ucntx->dpi.dpi;
630 		resp.dbr = (u64)ucntx->dpi.umdbr;
631 
632 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
633 		if (rc) {
634 			dev_err(rdev_to_dev(rdev),
635 				"Failed to copy user response\n");
636 			goto dbfail;
637 		}
638 	}
639 
640 	if (!udata)
641 		if (bnxt_re_create_fence_mr(pd))
642 			dev_warn(rdev_to_dev(rdev),
643 				 "Failed to create Fence-MR\n");
644 	return &pd->ib_pd;
645 dbfail:
646 	(void)bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
647 				    &pd->qplib_pd);
648 fail:
649 	kfree(pd);
650 	return ERR_PTR(rc);
651 }
652 
653 /* Address Handles */
654 int bnxt_re_destroy_ah(struct ib_ah *ib_ah)
655 {
656 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
657 	struct bnxt_re_dev *rdev = ah->rdev;
658 	int rc;
659 
660 	rc = bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah);
661 	if (rc) {
662 		dev_err(rdev_to_dev(rdev), "Failed to destroy HW AH");
663 		return rc;
664 	}
665 	kfree(ah);
666 	return 0;
667 }
668 
669 struct ib_ah *bnxt_re_create_ah(struct ib_pd *ib_pd,
670 				struct rdma_ah_attr *ah_attr,
671 				struct ib_udata *udata)
672 {
673 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
674 	struct bnxt_re_dev *rdev = pd->rdev;
675 	struct bnxt_re_ah *ah;
676 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
677 	int rc;
678 	u8 nw_type;
679 
680 	struct ib_gid_attr sgid_attr;
681 
682 	if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) {
683 		dev_err(rdev_to_dev(rdev), "Failed to alloc AH: GRH not set");
684 		return ERR_PTR(-EINVAL);
685 	}
686 	ah = kzalloc(sizeof(*ah), GFP_ATOMIC);
687 	if (!ah)
688 		return ERR_PTR(-ENOMEM);
689 
690 	ah->rdev = rdev;
691 	ah->qplib_ah.pd = &pd->qplib_pd;
692 
693 	/* Supply the configuration for the HW */
694 	memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw,
695 	       sizeof(union ib_gid));
696 	/*
697 	 * If RoCE V2 is enabled, stack will have two entries for
698 	 * each GID entry. Avoiding this duplicte entry in HW. Dividing
699 	 * the GID index by 2 for RoCE V2
700 	 */
701 	ah->qplib_ah.sgid_index = grh->sgid_index / 2;
702 	ah->qplib_ah.host_sgid_index = grh->sgid_index;
703 	ah->qplib_ah.traffic_class = grh->traffic_class;
704 	ah->qplib_ah.flow_label = grh->flow_label;
705 	ah->qplib_ah.hop_limit = grh->hop_limit;
706 	ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr);
707 	if (ib_pd->uobject &&
708 	    !rdma_is_multicast_addr((struct in6_addr *)
709 				    grh->dgid.raw) &&
710 	    !rdma_link_local_addr((struct in6_addr *)
711 				  grh->dgid.raw)) {
712 		union ib_gid sgid;
713 
714 		rc = ib_get_cached_gid(&rdev->ibdev, 1,
715 				       grh->sgid_index, &sgid,
716 				       &sgid_attr);
717 		if (rc) {
718 			dev_err(rdev_to_dev(rdev),
719 				"Failed to query gid at index %d",
720 				grh->sgid_index);
721 			goto fail;
722 		}
723 		if (sgid_attr.ndev)
724 			dev_put(sgid_attr.ndev);
725 		/* Get network header type for this GID */
726 		nw_type = ib_gid_to_network_type(sgid_attr.gid_type, &sgid);
727 		switch (nw_type) {
728 		case RDMA_NETWORK_IPV4:
729 			ah->qplib_ah.nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4;
730 			break;
731 		case RDMA_NETWORK_IPV6:
732 			ah->qplib_ah.nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6;
733 			break;
734 		default:
735 			ah->qplib_ah.nw_type = CMDQ_CREATE_AH_TYPE_V1;
736 			break;
737 		}
738 	}
739 
740 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
741 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah);
742 	if (rc) {
743 		dev_err(rdev_to_dev(rdev), "Failed to allocate HW AH");
744 		goto fail;
745 	}
746 
747 	/* Write AVID to shared page. */
748 	if (ib_pd->uobject) {
749 		struct ib_ucontext *ib_uctx = ib_pd->uobject->context;
750 		struct bnxt_re_ucontext *uctx;
751 		unsigned long flag;
752 		u32 *wrptr;
753 
754 		uctx = container_of(ib_uctx, struct bnxt_re_ucontext, ib_uctx);
755 		spin_lock_irqsave(&uctx->sh_lock, flag);
756 		wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT);
757 		*wrptr = ah->qplib_ah.id;
758 		wmb(); /* make sure cache is updated. */
759 		spin_unlock_irqrestore(&uctx->sh_lock, flag);
760 	}
761 
762 	return &ah->ib_ah;
763 
764 fail:
765 	kfree(ah);
766 	return ERR_PTR(rc);
767 }
768 
769 int bnxt_re_modify_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
770 {
771 	return 0;
772 }
773 
774 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
775 {
776 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
777 
778 	ah_attr->type = ib_ah->type;
779 	rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl);
780 	memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN);
781 	rdma_ah_set_grh(ah_attr, NULL, 0,
782 			ah->qplib_ah.host_sgid_index,
783 			0, ah->qplib_ah.traffic_class);
784 	rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data);
785 	rdma_ah_set_port_num(ah_attr, 1);
786 	rdma_ah_set_static_rate(ah_attr, 0);
787 	return 0;
788 }
789 
790 /* Queue Pairs */
791 int bnxt_re_destroy_qp(struct ib_qp *ib_qp)
792 {
793 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
794 	struct bnxt_re_dev *rdev = qp->rdev;
795 	int rc;
796 
797 	bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
798 	bnxt_qplib_del_flush_qp(&qp->qplib_qp);
799 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
800 	if (rc) {
801 		dev_err(rdev_to_dev(rdev), "Failed to destroy HW QP");
802 		return rc;
803 	}
804 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->qp1_sqp) {
805 		rc = bnxt_qplib_destroy_ah(&rdev->qplib_res,
806 					   &rdev->sqp_ah->qplib_ah);
807 		if (rc) {
808 			dev_err(rdev_to_dev(rdev),
809 				"Failed to destroy HW AH for shadow QP");
810 			return rc;
811 		}
812 
813 		bnxt_qplib_del_flush_qp(&qp->qplib_qp);
814 		rc = bnxt_qplib_destroy_qp(&rdev->qplib_res,
815 					   &rdev->qp1_sqp->qplib_qp);
816 		if (rc) {
817 			dev_err(rdev_to_dev(rdev),
818 				"Failed to destroy Shadow QP");
819 			return rc;
820 		}
821 		mutex_lock(&rdev->qp_lock);
822 		list_del(&rdev->qp1_sqp->list);
823 		atomic_dec(&rdev->qp_count);
824 		mutex_unlock(&rdev->qp_lock);
825 
826 		kfree(rdev->sqp_ah);
827 		kfree(rdev->qp1_sqp);
828 		rdev->qp1_sqp = NULL;
829 		rdev->sqp_ah = NULL;
830 	}
831 
832 	if (!IS_ERR_OR_NULL(qp->rumem))
833 		ib_umem_release(qp->rumem);
834 	if (!IS_ERR_OR_NULL(qp->sumem))
835 		ib_umem_release(qp->sumem);
836 
837 	mutex_lock(&rdev->qp_lock);
838 	list_del(&qp->list);
839 	atomic_dec(&rdev->qp_count);
840 	mutex_unlock(&rdev->qp_lock);
841 	kfree(qp);
842 	return 0;
843 }
844 
845 static u8 __from_ib_qp_type(enum ib_qp_type type)
846 {
847 	switch (type) {
848 	case IB_QPT_GSI:
849 		return CMDQ_CREATE_QP1_TYPE_GSI;
850 	case IB_QPT_RC:
851 		return CMDQ_CREATE_QP_TYPE_RC;
852 	case IB_QPT_UD:
853 		return CMDQ_CREATE_QP_TYPE_UD;
854 	default:
855 		return IB_QPT_MAX;
856 	}
857 }
858 
859 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
860 				struct bnxt_re_qp *qp, struct ib_udata *udata)
861 {
862 	struct bnxt_re_qp_req ureq;
863 	struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
864 	struct ib_umem *umem;
865 	int bytes = 0;
866 	struct ib_ucontext *context = pd->ib_pd.uobject->context;
867 	struct bnxt_re_ucontext *cntx = container_of(context,
868 						     struct bnxt_re_ucontext,
869 						     ib_uctx);
870 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
871 		return -EFAULT;
872 
873 	bytes = (qplib_qp->sq.max_wqe * BNXT_QPLIB_MAX_SQE_ENTRY_SIZE);
874 	/* Consider mapping PSN search memory only for RC QPs. */
875 	if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC)
876 		bytes += (qplib_qp->sq.max_wqe * sizeof(struct sq_psn_search));
877 	bytes = PAGE_ALIGN(bytes);
878 	umem = ib_umem_get(context, ureq.qpsva, bytes,
879 			   IB_ACCESS_LOCAL_WRITE, 1);
880 	if (IS_ERR(umem))
881 		return PTR_ERR(umem);
882 
883 	qp->sumem = umem;
884 	qplib_qp->sq.sglist = umem->sg_head.sgl;
885 	qplib_qp->sq.nmap = umem->nmap;
886 	qplib_qp->qp_handle = ureq.qp_handle;
887 
888 	if (!qp->qplib_qp.srq) {
889 		bytes = (qplib_qp->rq.max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
890 		bytes = PAGE_ALIGN(bytes);
891 		umem = ib_umem_get(context, ureq.qprva, bytes,
892 				   IB_ACCESS_LOCAL_WRITE, 1);
893 		if (IS_ERR(umem))
894 			goto rqfail;
895 		qp->rumem = umem;
896 		qplib_qp->rq.sglist = umem->sg_head.sgl;
897 		qplib_qp->rq.nmap = umem->nmap;
898 	}
899 
900 	qplib_qp->dpi = &cntx->dpi;
901 	return 0;
902 rqfail:
903 	ib_umem_release(qp->sumem);
904 	qp->sumem = NULL;
905 	qplib_qp->sq.sglist = NULL;
906 	qplib_qp->sq.nmap = 0;
907 
908 	return PTR_ERR(umem);
909 }
910 
911 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
912 				(struct bnxt_re_pd *pd,
913 				 struct bnxt_qplib_res *qp1_res,
914 				 struct bnxt_qplib_qp *qp1_qp)
915 {
916 	struct bnxt_re_dev *rdev = pd->rdev;
917 	struct bnxt_re_ah *ah;
918 	union ib_gid sgid;
919 	int rc;
920 
921 	ah = kzalloc(sizeof(*ah), GFP_KERNEL);
922 	if (!ah)
923 		return NULL;
924 
925 	ah->rdev = rdev;
926 	ah->qplib_ah.pd = &pd->qplib_pd;
927 
928 	rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid);
929 	if (rc)
930 		goto fail;
931 
932 	/* supply the dgid data same as sgid */
933 	memcpy(ah->qplib_ah.dgid.data, &sgid.raw,
934 	       sizeof(union ib_gid));
935 	ah->qplib_ah.sgid_index = 0;
936 
937 	ah->qplib_ah.traffic_class = 0;
938 	ah->qplib_ah.flow_label = 0;
939 	ah->qplib_ah.hop_limit = 1;
940 	ah->qplib_ah.sl = 0;
941 	/* Have DMAC same as SMAC */
942 	ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr);
943 
944 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah);
945 	if (rc) {
946 		dev_err(rdev_to_dev(rdev),
947 			"Failed to allocate HW AH for Shadow QP");
948 		goto fail;
949 	}
950 
951 	return ah;
952 
953 fail:
954 	kfree(ah);
955 	return NULL;
956 }
957 
958 static struct bnxt_re_qp *bnxt_re_create_shadow_qp
959 				(struct bnxt_re_pd *pd,
960 				 struct bnxt_qplib_res *qp1_res,
961 				 struct bnxt_qplib_qp *qp1_qp)
962 {
963 	struct bnxt_re_dev *rdev = pd->rdev;
964 	struct bnxt_re_qp *qp;
965 	int rc;
966 
967 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
968 	if (!qp)
969 		return NULL;
970 
971 	qp->rdev = rdev;
972 
973 	/* Initialize the shadow QP structure from the QP1 values */
974 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
975 
976 	qp->qplib_qp.pd = &pd->qplib_pd;
977 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
978 	qp->qplib_qp.type = IB_QPT_UD;
979 
980 	qp->qplib_qp.max_inline_data = 0;
981 	qp->qplib_qp.sig_type = true;
982 
983 	/* Shadow QP SQ depth should be same as QP1 RQ depth */
984 	qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
985 	qp->qplib_qp.sq.max_sge = 2;
986 	/* Q full delta can be 1 since it is internal QP */
987 	qp->qplib_qp.sq.q_full_delta = 1;
988 
989 	qp->qplib_qp.scq = qp1_qp->scq;
990 	qp->qplib_qp.rcq = qp1_qp->rcq;
991 
992 	qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
993 	qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
994 	/* Q full delta can be 1 since it is internal QP */
995 	qp->qplib_qp.rq.q_full_delta = 1;
996 
997 	qp->qplib_qp.mtu = qp1_qp->mtu;
998 
999 	qp->qplib_qp.sq_hdr_buf_size = 0;
1000 	qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
1001 	qp->qplib_qp.dpi = &rdev->dpi_privileged;
1002 
1003 	rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
1004 	if (rc)
1005 		goto fail;
1006 
1007 	rdev->sqp_id = qp->qplib_qp.id;
1008 
1009 	spin_lock_init(&qp->sq_lock);
1010 	INIT_LIST_HEAD(&qp->list);
1011 	mutex_lock(&rdev->qp_lock);
1012 	list_add_tail(&qp->list, &rdev->qp_list);
1013 	atomic_inc(&rdev->qp_count);
1014 	mutex_unlock(&rdev->qp_lock);
1015 	return qp;
1016 fail:
1017 	kfree(qp);
1018 	return NULL;
1019 }
1020 
1021 struct ib_qp *bnxt_re_create_qp(struct ib_pd *ib_pd,
1022 				struct ib_qp_init_attr *qp_init_attr,
1023 				struct ib_udata *udata)
1024 {
1025 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1026 	struct bnxt_re_dev *rdev = pd->rdev;
1027 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1028 	struct bnxt_re_qp *qp;
1029 	struct bnxt_re_cq *cq;
1030 	struct bnxt_re_srq *srq;
1031 	int rc, entries;
1032 
1033 	if ((qp_init_attr->cap.max_send_wr > dev_attr->max_qp_wqes) ||
1034 	    (qp_init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes) ||
1035 	    (qp_init_attr->cap.max_send_sge > dev_attr->max_qp_sges) ||
1036 	    (qp_init_attr->cap.max_recv_sge > dev_attr->max_qp_sges) ||
1037 	    (qp_init_attr->cap.max_inline_data > dev_attr->max_inline_data))
1038 		return ERR_PTR(-EINVAL);
1039 
1040 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1041 	if (!qp)
1042 		return ERR_PTR(-ENOMEM);
1043 
1044 	qp->rdev = rdev;
1045 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
1046 	qp->qplib_qp.pd = &pd->qplib_pd;
1047 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
1048 	qp->qplib_qp.type = __from_ib_qp_type(qp_init_attr->qp_type);
1049 	if (qp->qplib_qp.type == IB_QPT_MAX) {
1050 		dev_err(rdev_to_dev(rdev), "QP type 0x%x not supported",
1051 			qp->qplib_qp.type);
1052 		rc = -EINVAL;
1053 		goto fail;
1054 	}
1055 	qp->qplib_qp.max_inline_data = qp_init_attr->cap.max_inline_data;
1056 	qp->qplib_qp.sig_type = ((qp_init_attr->sq_sig_type ==
1057 				  IB_SIGNAL_ALL_WR) ? true : false);
1058 
1059 	qp->qplib_qp.sq.max_sge = qp_init_attr->cap.max_send_sge;
1060 	if (qp->qplib_qp.sq.max_sge > dev_attr->max_qp_sges)
1061 		qp->qplib_qp.sq.max_sge = dev_attr->max_qp_sges;
1062 
1063 	if (qp_init_attr->send_cq) {
1064 		cq = container_of(qp_init_attr->send_cq, struct bnxt_re_cq,
1065 				  ib_cq);
1066 		if (!cq) {
1067 			dev_err(rdev_to_dev(rdev), "Send CQ not found");
1068 			rc = -EINVAL;
1069 			goto fail;
1070 		}
1071 		qp->qplib_qp.scq = &cq->qplib_cq;
1072 	}
1073 
1074 	if (qp_init_attr->recv_cq) {
1075 		cq = container_of(qp_init_attr->recv_cq, struct bnxt_re_cq,
1076 				  ib_cq);
1077 		if (!cq) {
1078 			dev_err(rdev_to_dev(rdev), "Receive CQ not found");
1079 			rc = -EINVAL;
1080 			goto fail;
1081 		}
1082 		qp->qplib_qp.rcq = &cq->qplib_cq;
1083 	}
1084 
1085 	if (qp_init_attr->srq) {
1086 		srq = container_of(qp_init_attr->srq, struct bnxt_re_srq,
1087 				   ib_srq);
1088 		if (!srq) {
1089 			dev_err(rdev_to_dev(rdev), "SRQ not found");
1090 			rc = -EINVAL;
1091 			goto fail;
1092 		}
1093 		qp->qplib_qp.srq = &srq->qplib_srq;
1094 		qp->qplib_qp.rq.max_wqe = 0;
1095 	} else {
1096 		/* Allocate 1 more than what's provided so posting max doesn't
1097 		 * mean empty
1098 		 */
1099 		entries = roundup_pow_of_two(qp_init_attr->cap.max_recv_wr + 1);
1100 		qp->qplib_qp.rq.max_wqe = min_t(u32, entries,
1101 						dev_attr->max_qp_wqes + 1);
1102 
1103 		qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
1104 						qp_init_attr->cap.max_recv_wr;
1105 
1106 		qp->qplib_qp.rq.max_sge = qp_init_attr->cap.max_recv_sge;
1107 		if (qp->qplib_qp.rq.max_sge > dev_attr->max_qp_sges)
1108 			qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
1109 	}
1110 
1111 	qp->qplib_qp.mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1112 
1113 	if (qp_init_attr->qp_type == IB_QPT_GSI) {
1114 		/* Allocate 1 more than what's provided */
1115 		entries = roundup_pow_of_two(qp_init_attr->cap.max_send_wr + 1);
1116 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
1117 						dev_attr->max_qp_wqes + 1);
1118 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
1119 						qp_init_attr->cap.max_send_wr;
1120 		qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
1121 		if (qp->qplib_qp.rq.max_sge > dev_attr->max_qp_sges)
1122 			qp->qplib_qp.rq.max_sge = dev_attr->max_qp_sges;
1123 		qp->qplib_qp.sq.max_sge++;
1124 		if (qp->qplib_qp.sq.max_sge > dev_attr->max_qp_sges)
1125 			qp->qplib_qp.sq.max_sge = dev_attr->max_qp_sges;
1126 
1127 		qp->qplib_qp.rq_hdr_buf_size =
1128 					BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
1129 
1130 		qp->qplib_qp.sq_hdr_buf_size =
1131 					BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
1132 		qp->qplib_qp.dpi = &rdev->dpi_privileged;
1133 		rc = bnxt_qplib_create_qp1(&rdev->qplib_res, &qp->qplib_qp);
1134 		if (rc) {
1135 			dev_err(rdev_to_dev(rdev), "Failed to create HW QP1");
1136 			goto fail;
1137 		}
1138 		/* Create a shadow QP to handle the QP1 traffic */
1139 		rdev->qp1_sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res,
1140 							 &qp->qplib_qp);
1141 		if (!rdev->qp1_sqp) {
1142 			rc = -EINVAL;
1143 			dev_err(rdev_to_dev(rdev),
1144 				"Failed to create Shadow QP for QP1");
1145 			goto qp_destroy;
1146 		}
1147 		rdev->sqp_ah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res,
1148 							   &qp->qplib_qp);
1149 		if (!rdev->sqp_ah) {
1150 			bnxt_qplib_destroy_qp(&rdev->qplib_res,
1151 					      &rdev->qp1_sqp->qplib_qp);
1152 			rc = -EINVAL;
1153 			dev_err(rdev_to_dev(rdev),
1154 				"Failed to create AH entry for ShadowQP");
1155 			goto qp_destroy;
1156 		}
1157 
1158 	} else {
1159 		/* Allocate 128 + 1 more than what's provided */
1160 		entries = roundup_pow_of_two(qp_init_attr->cap.max_send_wr +
1161 					     BNXT_QPLIB_RESERVED_QP_WRS + 1);
1162 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
1163 						dev_attr->max_qp_wqes +
1164 						BNXT_QPLIB_RESERVED_QP_WRS + 1);
1165 		qp->qplib_qp.sq.q_full_delta = BNXT_QPLIB_RESERVED_QP_WRS + 1;
1166 
1167 		/*
1168 		 * Reserving one slot for Phantom WQE. Application can
1169 		 * post one extra entry in this case. But allowing this to avoid
1170 		 * unexpected Queue full condition
1171 		 */
1172 
1173 		qp->qplib_qp.sq.q_full_delta -= 1;
1174 
1175 		qp->qplib_qp.max_rd_atomic = dev_attr->max_qp_rd_atom;
1176 		qp->qplib_qp.max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
1177 		if (udata) {
1178 			rc = bnxt_re_init_user_qp(rdev, pd, qp, udata);
1179 			if (rc)
1180 				goto fail;
1181 		} else {
1182 			qp->qplib_qp.dpi = &rdev->dpi_privileged;
1183 		}
1184 
1185 		rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
1186 		if (rc) {
1187 			dev_err(rdev_to_dev(rdev), "Failed to create HW QP");
1188 			goto fail;
1189 		}
1190 	}
1191 
1192 	qp->ib_qp.qp_num = qp->qplib_qp.id;
1193 	spin_lock_init(&qp->sq_lock);
1194 	spin_lock_init(&qp->rq_lock);
1195 
1196 	if (udata) {
1197 		struct bnxt_re_qp_resp resp;
1198 
1199 		resp.qpid = qp->ib_qp.qp_num;
1200 		resp.rsvd = 0;
1201 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1202 		if (rc) {
1203 			dev_err(rdev_to_dev(rdev), "Failed to copy QP udata");
1204 			goto qp_destroy;
1205 		}
1206 	}
1207 	INIT_LIST_HEAD(&qp->list);
1208 	mutex_lock(&rdev->qp_lock);
1209 	list_add_tail(&qp->list, &rdev->qp_list);
1210 	atomic_inc(&rdev->qp_count);
1211 	mutex_unlock(&rdev->qp_lock);
1212 
1213 	return &qp->ib_qp;
1214 qp_destroy:
1215 	bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
1216 fail:
1217 	kfree(qp);
1218 	return ERR_PTR(rc);
1219 }
1220 
1221 static u8 __from_ib_qp_state(enum ib_qp_state state)
1222 {
1223 	switch (state) {
1224 	case IB_QPS_RESET:
1225 		return CMDQ_MODIFY_QP_NEW_STATE_RESET;
1226 	case IB_QPS_INIT:
1227 		return CMDQ_MODIFY_QP_NEW_STATE_INIT;
1228 	case IB_QPS_RTR:
1229 		return CMDQ_MODIFY_QP_NEW_STATE_RTR;
1230 	case IB_QPS_RTS:
1231 		return CMDQ_MODIFY_QP_NEW_STATE_RTS;
1232 	case IB_QPS_SQD:
1233 		return CMDQ_MODIFY_QP_NEW_STATE_SQD;
1234 	case IB_QPS_SQE:
1235 		return CMDQ_MODIFY_QP_NEW_STATE_SQE;
1236 	case IB_QPS_ERR:
1237 	default:
1238 		return CMDQ_MODIFY_QP_NEW_STATE_ERR;
1239 	}
1240 }
1241 
1242 static enum ib_qp_state __to_ib_qp_state(u8 state)
1243 {
1244 	switch (state) {
1245 	case CMDQ_MODIFY_QP_NEW_STATE_RESET:
1246 		return IB_QPS_RESET;
1247 	case CMDQ_MODIFY_QP_NEW_STATE_INIT:
1248 		return IB_QPS_INIT;
1249 	case CMDQ_MODIFY_QP_NEW_STATE_RTR:
1250 		return IB_QPS_RTR;
1251 	case CMDQ_MODIFY_QP_NEW_STATE_RTS:
1252 		return IB_QPS_RTS;
1253 	case CMDQ_MODIFY_QP_NEW_STATE_SQD:
1254 		return IB_QPS_SQD;
1255 	case CMDQ_MODIFY_QP_NEW_STATE_SQE:
1256 		return IB_QPS_SQE;
1257 	case CMDQ_MODIFY_QP_NEW_STATE_ERR:
1258 	default:
1259 		return IB_QPS_ERR;
1260 	}
1261 }
1262 
1263 static u32 __from_ib_mtu(enum ib_mtu mtu)
1264 {
1265 	switch (mtu) {
1266 	case IB_MTU_256:
1267 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_256;
1268 	case IB_MTU_512:
1269 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_512;
1270 	case IB_MTU_1024:
1271 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024;
1272 	case IB_MTU_2048:
1273 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1274 	case IB_MTU_4096:
1275 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096;
1276 	default:
1277 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1278 	}
1279 }
1280 
1281 static enum ib_mtu __to_ib_mtu(u32 mtu)
1282 {
1283 	switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) {
1284 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_256:
1285 		return IB_MTU_256;
1286 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_512:
1287 		return IB_MTU_512;
1288 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024:
1289 		return IB_MTU_1024;
1290 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048:
1291 		return IB_MTU_2048;
1292 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096:
1293 		return IB_MTU_4096;
1294 	default:
1295 		return IB_MTU_2048;
1296 	}
1297 }
1298 
1299 /* Shared Receive Queues */
1300 int bnxt_re_destroy_srq(struct ib_srq *ib_srq)
1301 {
1302 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1303 					       ib_srq);
1304 	struct bnxt_re_dev *rdev = srq->rdev;
1305 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1306 	struct bnxt_qplib_nq *nq = NULL;
1307 	int rc;
1308 
1309 	if (qplib_srq->cq)
1310 		nq = qplib_srq->cq->nq;
1311 	rc = bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
1312 	if (rc) {
1313 		dev_err(rdev_to_dev(rdev), "Destroy HW SRQ failed!");
1314 		return rc;
1315 	}
1316 
1317 	if (srq->umem)
1318 		ib_umem_release(srq->umem);
1319 	kfree(srq);
1320 	atomic_dec(&rdev->srq_count);
1321 	if (nq)
1322 		nq->budget--;
1323 	return 0;
1324 }
1325 
1326 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
1327 				 struct bnxt_re_pd *pd,
1328 				 struct bnxt_re_srq *srq,
1329 				 struct ib_udata *udata)
1330 {
1331 	struct bnxt_re_srq_req ureq;
1332 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1333 	struct ib_umem *umem;
1334 	int bytes = 0;
1335 	struct ib_ucontext *context = pd->ib_pd.uobject->context;
1336 	struct bnxt_re_ucontext *cntx = container_of(context,
1337 						     struct bnxt_re_ucontext,
1338 						     ib_uctx);
1339 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1340 		return -EFAULT;
1341 
1342 	bytes = (qplib_srq->max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
1343 	bytes = PAGE_ALIGN(bytes);
1344 	umem = ib_umem_get(context, ureq.srqva, bytes,
1345 			   IB_ACCESS_LOCAL_WRITE, 1);
1346 	if (IS_ERR(umem))
1347 		return PTR_ERR(umem);
1348 
1349 	srq->umem = umem;
1350 	qplib_srq->nmap = umem->nmap;
1351 	qplib_srq->sglist = umem->sg_head.sgl;
1352 	qplib_srq->srq_handle = ureq.srq_handle;
1353 	qplib_srq->dpi = &cntx->dpi;
1354 
1355 	return 0;
1356 }
1357 
1358 struct ib_srq *bnxt_re_create_srq(struct ib_pd *ib_pd,
1359 				  struct ib_srq_init_attr *srq_init_attr,
1360 				  struct ib_udata *udata)
1361 {
1362 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1363 	struct bnxt_re_dev *rdev = pd->rdev;
1364 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1365 	struct bnxt_re_srq *srq;
1366 	struct bnxt_qplib_nq *nq = NULL;
1367 	int rc, entries;
1368 
1369 	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
1370 		dev_err(rdev_to_dev(rdev), "Create CQ failed - max exceeded");
1371 		rc = -EINVAL;
1372 		goto exit;
1373 	}
1374 
1375 	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
1376 		rc = -ENOTSUPP;
1377 		goto exit;
1378 	}
1379 
1380 	srq = kzalloc(sizeof(*srq), GFP_KERNEL);
1381 	if (!srq) {
1382 		rc = -ENOMEM;
1383 		goto exit;
1384 	}
1385 	srq->rdev = rdev;
1386 	srq->qplib_srq.pd = &pd->qplib_pd;
1387 	srq->qplib_srq.dpi = &rdev->dpi_privileged;
1388 	/* Allocate 1 more than what's provided so posting max doesn't
1389 	 * mean empty
1390 	 */
1391 	entries = roundup_pow_of_two(srq_init_attr->attr.max_wr + 1);
1392 	if (entries > dev_attr->max_srq_wqes + 1)
1393 		entries = dev_attr->max_srq_wqes + 1;
1394 
1395 	srq->qplib_srq.max_wqe = entries;
1396 	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
1397 	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
1398 	srq->srq_limit = srq_init_attr->attr.srq_limit;
1399 	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
1400 	nq = &rdev->nq[0];
1401 
1402 	if (udata) {
1403 		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
1404 		if (rc)
1405 			goto fail;
1406 	}
1407 
1408 	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
1409 	if (rc) {
1410 		dev_err(rdev_to_dev(rdev), "Create HW SRQ failed!");
1411 		goto fail;
1412 	}
1413 
1414 	if (udata) {
1415 		struct bnxt_re_srq_resp resp;
1416 
1417 		resp.srqid = srq->qplib_srq.id;
1418 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1419 		if (rc) {
1420 			dev_err(rdev_to_dev(rdev), "SRQ copy to udata failed!");
1421 			bnxt_qplib_destroy_srq(&rdev->qplib_res,
1422 					       &srq->qplib_srq);
1423 			goto exit;
1424 		}
1425 	}
1426 	if (nq)
1427 		nq->budget++;
1428 	atomic_inc(&rdev->srq_count);
1429 
1430 	return &srq->ib_srq;
1431 
1432 fail:
1433 	if (srq->umem)
1434 		ib_umem_release(srq->umem);
1435 	kfree(srq);
1436 exit:
1437 	return ERR_PTR(rc);
1438 }
1439 
1440 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
1441 		       enum ib_srq_attr_mask srq_attr_mask,
1442 		       struct ib_udata *udata)
1443 {
1444 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1445 					       ib_srq);
1446 	struct bnxt_re_dev *rdev = srq->rdev;
1447 	int rc;
1448 
1449 	switch (srq_attr_mask) {
1450 	case IB_SRQ_MAX_WR:
1451 		/* SRQ resize is not supported */
1452 		break;
1453 	case IB_SRQ_LIMIT:
1454 		/* Change the SRQ threshold */
1455 		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
1456 			return -EINVAL;
1457 
1458 		srq->qplib_srq.threshold = srq_attr->srq_limit;
1459 		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
1460 		if (rc) {
1461 			dev_err(rdev_to_dev(rdev), "Modify HW SRQ failed!");
1462 			return rc;
1463 		}
1464 		/* On success, update the shadow */
1465 		srq->srq_limit = srq_attr->srq_limit;
1466 		/* No need to Build and send response back to udata */
1467 		break;
1468 	default:
1469 		dev_err(rdev_to_dev(rdev),
1470 			"Unsupported srq_attr_mask 0x%x", srq_attr_mask);
1471 		return -EINVAL;
1472 	}
1473 	return 0;
1474 }
1475 
1476 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
1477 {
1478 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1479 					       ib_srq);
1480 	struct bnxt_re_srq tsrq;
1481 	struct bnxt_re_dev *rdev = srq->rdev;
1482 	int rc;
1483 
1484 	/* Get live SRQ attr */
1485 	tsrq.qplib_srq.id = srq->qplib_srq.id;
1486 	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
1487 	if (rc) {
1488 		dev_err(rdev_to_dev(rdev), "Query HW SRQ failed!");
1489 		return rc;
1490 	}
1491 	srq_attr->max_wr = srq->qplib_srq.max_wqe;
1492 	srq_attr->max_sge = srq->qplib_srq.max_sge;
1493 	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
1494 
1495 	return 0;
1496 }
1497 
1498 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, struct ib_recv_wr *wr,
1499 			  struct ib_recv_wr **bad_wr)
1500 {
1501 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1502 					       ib_srq);
1503 	struct bnxt_qplib_swqe wqe;
1504 	unsigned long flags;
1505 	int rc = 0, payload_sz = 0;
1506 
1507 	spin_lock_irqsave(&srq->lock, flags);
1508 	while (wr) {
1509 		/* Transcribe each ib_recv_wr to qplib_swqe */
1510 		wqe.num_sge = wr->num_sge;
1511 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
1512 					       wr->num_sge);
1513 		wqe.wr_id = wr->wr_id;
1514 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
1515 
1516 		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
1517 		if (rc) {
1518 			*bad_wr = wr;
1519 			break;
1520 		}
1521 		wr = wr->next;
1522 	}
1523 	spin_unlock_irqrestore(&srq->lock, flags);
1524 
1525 	return rc;
1526 }
1527 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
1528 				    struct bnxt_re_qp *qp1_qp,
1529 				    int qp_attr_mask)
1530 {
1531 	struct bnxt_re_qp *qp = rdev->qp1_sqp;
1532 	int rc = 0;
1533 
1534 	if (qp_attr_mask & IB_QP_STATE) {
1535 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1536 		qp->qplib_qp.state = qp1_qp->qplib_qp.state;
1537 	}
1538 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1539 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1540 		qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index;
1541 	}
1542 
1543 	if (qp_attr_mask & IB_QP_QKEY) {
1544 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1545 		/* Using a Random  QKEY */
1546 		qp->qplib_qp.qkey = 0x81818181;
1547 	}
1548 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1549 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1550 		qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn;
1551 	}
1552 
1553 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1554 	if (rc)
1555 		dev_err(rdev_to_dev(rdev),
1556 			"Failed to modify Shadow QP for QP1");
1557 	return rc;
1558 }
1559 
1560 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1561 		      int qp_attr_mask, struct ib_udata *udata)
1562 {
1563 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1564 	struct bnxt_re_dev *rdev = qp->rdev;
1565 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1566 	enum ib_qp_state curr_qp_state, new_qp_state;
1567 	int rc, entries;
1568 	int status;
1569 	union ib_gid sgid;
1570 	struct ib_gid_attr sgid_attr;
1571 	u8 nw_type;
1572 
1573 	qp->qplib_qp.modify_flags = 0;
1574 	if (qp_attr_mask & IB_QP_STATE) {
1575 		curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state);
1576 		new_qp_state = qp_attr->qp_state;
1577 		if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state,
1578 					ib_qp->qp_type, qp_attr_mask,
1579 					IB_LINK_LAYER_ETHERNET)) {
1580 			dev_err(rdev_to_dev(rdev),
1581 				"Invalid attribute mask: %#x specified ",
1582 				qp_attr_mask);
1583 			dev_err(rdev_to_dev(rdev),
1584 				"for qpn: %#x type: %#x",
1585 				ib_qp->qp_num, ib_qp->qp_type);
1586 			dev_err(rdev_to_dev(rdev),
1587 				"curr_qp_state=0x%x, new_qp_state=0x%x\n",
1588 				curr_qp_state, new_qp_state);
1589 			return -EINVAL;
1590 		}
1591 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1592 		qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state);
1593 
1594 		if (!qp->sumem &&
1595 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1596 			dev_dbg(rdev_to_dev(rdev),
1597 				"Move QP = %p to flush list\n",
1598 				qp);
1599 			bnxt_qplib_add_flush_qp(&qp->qplib_qp);
1600 		}
1601 		if (!qp->sumem &&
1602 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) {
1603 			dev_dbg(rdev_to_dev(rdev),
1604 				"Move QP = %p out of flush list\n",
1605 				qp);
1606 			bnxt_qplib_del_flush_qp(&qp->qplib_qp);
1607 		}
1608 	}
1609 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1610 		qp->qplib_qp.modify_flags |=
1611 				CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY;
1612 		qp->qplib_qp.en_sqd_async_notify = true;
1613 	}
1614 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
1615 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS;
1616 		qp->qplib_qp.access =
1617 			__from_ib_access_flags(qp_attr->qp_access_flags);
1618 		/* LOCAL_WRITE access must be set to allow RC receive */
1619 		qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
1620 	}
1621 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1622 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1623 		qp->qplib_qp.pkey_index = qp_attr->pkey_index;
1624 	}
1625 	if (qp_attr_mask & IB_QP_QKEY) {
1626 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1627 		qp->qplib_qp.qkey = qp_attr->qkey;
1628 	}
1629 	if (qp_attr_mask & IB_QP_AV) {
1630 		const struct ib_global_route *grh =
1631 			rdma_ah_read_grh(&qp_attr->ah_attr);
1632 
1633 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID |
1634 				     CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL |
1635 				     CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX |
1636 				     CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT |
1637 				     CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS |
1638 				     CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC |
1639 				     CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID;
1640 		memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw,
1641 		       sizeof(qp->qplib_qp.ah.dgid.data));
1642 		qp->qplib_qp.ah.flow_label = grh->flow_label;
1643 		/* If RoCE V2 is enabled, stack will have two entries for
1644 		 * each GID entry. Avoiding this duplicte entry in HW. Dividing
1645 		 * the GID index by 2 for RoCE V2
1646 		 */
1647 		qp->qplib_qp.ah.sgid_index = grh->sgid_index / 2;
1648 		qp->qplib_qp.ah.host_sgid_index = grh->sgid_index;
1649 		qp->qplib_qp.ah.hop_limit = grh->hop_limit;
1650 		qp->qplib_qp.ah.traffic_class = grh->traffic_class;
1651 		qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr);
1652 		ether_addr_copy(qp->qplib_qp.ah.dmac,
1653 				qp_attr->ah_attr.roce.dmac);
1654 
1655 		status = ib_get_cached_gid(&rdev->ibdev, 1,
1656 					   grh->sgid_index,
1657 					   &sgid, &sgid_attr);
1658 		if (!status && sgid_attr.ndev) {
1659 			memcpy(qp->qplib_qp.smac, sgid_attr.ndev->dev_addr,
1660 			       ETH_ALEN);
1661 			dev_put(sgid_attr.ndev);
1662 			nw_type = ib_gid_to_network_type(sgid_attr.gid_type,
1663 							 &sgid);
1664 			switch (nw_type) {
1665 			case RDMA_NETWORK_IPV4:
1666 				qp->qplib_qp.nw_type =
1667 					CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4;
1668 				break;
1669 			case RDMA_NETWORK_IPV6:
1670 				qp->qplib_qp.nw_type =
1671 					CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6;
1672 				break;
1673 			default:
1674 				qp->qplib_qp.nw_type =
1675 					CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1;
1676 				break;
1677 			}
1678 		}
1679 	}
1680 
1681 	if (qp_attr_mask & IB_QP_PATH_MTU) {
1682 		qp->qplib_qp.modify_flags |=
1683 				CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1684 		qp->qplib_qp.path_mtu = __from_ib_mtu(qp_attr->path_mtu);
1685 		qp->qplib_qp.mtu = ib_mtu_enum_to_int(qp_attr->path_mtu);
1686 	} else if (qp_attr->qp_state == IB_QPS_RTR) {
1687 		qp->qplib_qp.modify_flags |=
1688 			CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
1689 		qp->qplib_qp.path_mtu =
1690 			__from_ib_mtu(iboe_get_mtu(rdev->netdev->mtu));
1691 		qp->qplib_qp.mtu =
1692 			ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1693 	}
1694 
1695 	if (qp_attr_mask & IB_QP_TIMEOUT) {
1696 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT;
1697 		qp->qplib_qp.timeout = qp_attr->timeout;
1698 	}
1699 	if (qp_attr_mask & IB_QP_RETRY_CNT) {
1700 		qp->qplib_qp.modify_flags |=
1701 				CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT;
1702 		qp->qplib_qp.retry_cnt = qp_attr->retry_cnt;
1703 	}
1704 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
1705 		qp->qplib_qp.modify_flags |=
1706 				CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY;
1707 		qp->qplib_qp.rnr_retry = qp_attr->rnr_retry;
1708 	}
1709 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) {
1710 		qp->qplib_qp.modify_flags |=
1711 				CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER;
1712 		qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer;
1713 	}
1714 	if (qp_attr_mask & IB_QP_RQ_PSN) {
1715 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN;
1716 		qp->qplib_qp.rq.psn = qp_attr->rq_psn;
1717 	}
1718 	if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1719 		qp->qplib_qp.modify_flags |=
1720 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC;
1721 		/* Cap the max_rd_atomic to device max */
1722 		qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic,
1723 						   dev_attr->max_qp_rd_atom);
1724 	}
1725 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1726 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1727 		qp->qplib_qp.sq.psn = qp_attr->sq_psn;
1728 	}
1729 	if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1730 		if (qp_attr->max_dest_rd_atomic >
1731 		    dev_attr->max_qp_init_rd_atom) {
1732 			dev_err(rdev_to_dev(rdev),
1733 				"max_dest_rd_atomic requested%d is > dev_max%d",
1734 				qp_attr->max_dest_rd_atomic,
1735 				dev_attr->max_qp_init_rd_atom);
1736 			return -EINVAL;
1737 		}
1738 
1739 		qp->qplib_qp.modify_flags |=
1740 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC;
1741 		qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic;
1742 	}
1743 	if (qp_attr_mask & IB_QP_CAP) {
1744 		qp->qplib_qp.modify_flags |=
1745 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE |
1746 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE |
1747 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE |
1748 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE |
1749 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA;
1750 		if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) ||
1751 		    (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) ||
1752 		    (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) ||
1753 		    (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) ||
1754 		    (qp_attr->cap.max_inline_data >=
1755 						dev_attr->max_inline_data)) {
1756 			dev_err(rdev_to_dev(rdev),
1757 				"Create QP failed - max exceeded");
1758 			return -EINVAL;
1759 		}
1760 		entries = roundup_pow_of_two(qp_attr->cap.max_send_wr);
1761 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
1762 						dev_attr->max_qp_wqes + 1);
1763 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
1764 						qp_attr->cap.max_send_wr;
1765 		/*
1766 		 * Reserving one slot for Phantom WQE. Some application can
1767 		 * post one extra entry in this case. Allowing this to avoid
1768 		 * unexpected Queue full condition
1769 		 */
1770 		qp->qplib_qp.sq.q_full_delta -= 1;
1771 		qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
1772 		if (qp->qplib_qp.rq.max_wqe) {
1773 			entries = roundup_pow_of_two(qp_attr->cap.max_recv_wr);
1774 			qp->qplib_qp.rq.max_wqe =
1775 				min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1776 			qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
1777 						       qp_attr->cap.max_recv_wr;
1778 			qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
1779 		} else {
1780 			/* SRQ was used prior, just ignore the RQ caps */
1781 		}
1782 	}
1783 	if (qp_attr_mask & IB_QP_DEST_QPN) {
1784 		qp->qplib_qp.modify_flags |=
1785 				CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID;
1786 		qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num;
1787 	}
1788 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1789 	if (rc) {
1790 		dev_err(rdev_to_dev(rdev), "Failed to modify HW QP");
1791 		return rc;
1792 	}
1793 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->qp1_sqp)
1794 		rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask);
1795 	return rc;
1796 }
1797 
1798 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1799 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
1800 {
1801 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1802 	struct bnxt_re_dev *rdev = qp->rdev;
1803 	struct bnxt_qplib_qp *qplib_qp;
1804 	int rc;
1805 
1806 	qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL);
1807 	if (!qplib_qp)
1808 		return -ENOMEM;
1809 
1810 	qplib_qp->id = qp->qplib_qp.id;
1811 	qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index;
1812 
1813 	rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp);
1814 	if (rc) {
1815 		dev_err(rdev_to_dev(rdev), "Failed to query HW QP");
1816 		goto out;
1817 	}
1818 	qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state);
1819 	qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0;
1820 	qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access);
1821 	qp_attr->pkey_index = qplib_qp->pkey_index;
1822 	qp_attr->qkey = qplib_qp->qkey;
1823 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
1824 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label,
1825 			qplib_qp->ah.host_sgid_index,
1826 			qplib_qp->ah.hop_limit,
1827 			qplib_qp->ah.traffic_class);
1828 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data);
1829 	rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl);
1830 	ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac);
1831 	qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu);
1832 	qp_attr->timeout = qplib_qp->timeout;
1833 	qp_attr->retry_cnt = qplib_qp->retry_cnt;
1834 	qp_attr->rnr_retry = qplib_qp->rnr_retry;
1835 	qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer;
1836 	qp_attr->rq_psn = qplib_qp->rq.psn;
1837 	qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic;
1838 	qp_attr->sq_psn = qplib_qp->sq.psn;
1839 	qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic;
1840 	qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR :
1841 							 IB_SIGNAL_REQ_WR;
1842 	qp_attr->dest_qp_num = qplib_qp->dest_qpn;
1843 
1844 	qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe;
1845 	qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge;
1846 	qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe;
1847 	qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge;
1848 	qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data;
1849 	qp_init_attr->cap = qp_attr->cap;
1850 
1851 out:
1852 	kfree(qplib_qp);
1853 	return rc;
1854 }
1855 
1856 /* Routine for sending QP1 packets for RoCE V1 an V2
1857  */
1858 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp,
1859 				     struct ib_send_wr *wr,
1860 				     struct bnxt_qplib_swqe *wqe,
1861 				     int payload_size)
1862 {
1863 	struct ib_device *ibdev = &qp->rdev->ibdev;
1864 	struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah,
1865 					     ib_ah);
1866 	struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah;
1867 	struct bnxt_qplib_sge sge;
1868 	union ib_gid sgid;
1869 	u8 nw_type;
1870 	u16 ether_type;
1871 	struct ib_gid_attr sgid_attr;
1872 	union ib_gid dgid;
1873 	bool is_eth = false;
1874 	bool is_vlan = false;
1875 	bool is_grh = false;
1876 	bool is_udp = false;
1877 	u8 ip_version = 0;
1878 	u16 vlan_id = 0xFFFF;
1879 	void *buf;
1880 	int i, rc = 0;
1881 
1882 	memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr));
1883 
1884 	rc = ib_get_cached_gid(ibdev, 1,
1885 			       qplib_ah->host_sgid_index, &sgid,
1886 			       &sgid_attr);
1887 	if (rc) {
1888 		dev_err(rdev_to_dev(qp->rdev),
1889 			"Failed to query gid at index %d",
1890 			qplib_ah->host_sgid_index);
1891 		return rc;
1892 	}
1893 	if (sgid_attr.ndev) {
1894 		if (is_vlan_dev(sgid_attr.ndev))
1895 			vlan_id = vlan_dev_vlan_id(sgid_attr.ndev);
1896 		dev_put(sgid_attr.ndev);
1897 	}
1898 	/* Get network header type for this GID */
1899 	nw_type = ib_gid_to_network_type(sgid_attr.gid_type, &sgid);
1900 	switch (nw_type) {
1901 	case RDMA_NETWORK_IPV4:
1902 		nw_type = BNXT_RE_ROCEV2_IPV4_PACKET;
1903 		break;
1904 	case RDMA_NETWORK_IPV6:
1905 		nw_type = BNXT_RE_ROCEV2_IPV6_PACKET;
1906 		break;
1907 	default:
1908 		nw_type = BNXT_RE_ROCE_V1_PACKET;
1909 		break;
1910 	}
1911 	memcpy(&dgid.raw, &qplib_ah->dgid, 16);
1912 	is_udp = sgid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
1913 	if (is_udp) {
1914 		if (ipv6_addr_v4mapped((struct in6_addr *)&sgid)) {
1915 			ip_version = 4;
1916 			ether_type = ETH_P_IP;
1917 		} else {
1918 			ip_version = 6;
1919 			ether_type = ETH_P_IPV6;
1920 		}
1921 		is_grh = false;
1922 	} else {
1923 		ether_type = ETH_P_IBOE;
1924 		is_grh = true;
1925 	}
1926 
1927 	is_eth = true;
1928 	is_vlan = (vlan_id && (vlan_id < 0x1000)) ? true : false;
1929 
1930 	ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh,
1931 			  ip_version, is_udp, 0, &qp->qp1_hdr);
1932 
1933 	/* ETH */
1934 	ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac);
1935 	ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac);
1936 
1937 	/* For vlan, check the sgid for vlan existence */
1938 
1939 	if (!is_vlan) {
1940 		qp->qp1_hdr.eth.type = cpu_to_be16(ether_type);
1941 	} else {
1942 		qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type);
1943 		qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id);
1944 	}
1945 
1946 	if (is_grh || (ip_version == 6)) {
1947 		memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid.raw, sizeof(sgid));
1948 		memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data,
1949 		       sizeof(sgid));
1950 		qp->qp1_hdr.grh.hop_limit     = qplib_ah->hop_limit;
1951 	}
1952 
1953 	if (ip_version == 4) {
1954 		qp->qp1_hdr.ip4.tos = 0;
1955 		qp->qp1_hdr.ip4.id = 0;
1956 		qp->qp1_hdr.ip4.frag_off = htons(IP_DF);
1957 		qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit;
1958 
1959 		memcpy(&qp->qp1_hdr.ip4.saddr, sgid.raw + 12, 4);
1960 		memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4);
1961 		qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr);
1962 	}
1963 
1964 	if (is_udp) {
1965 		qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT);
1966 		qp->qp1_hdr.udp.sport = htons(0x8CD1);
1967 		qp->qp1_hdr.udp.csum = 0;
1968 	}
1969 
1970 	/* BTH */
1971 	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
1972 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
1973 		qp->qp1_hdr.immediate_present = 1;
1974 	} else {
1975 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
1976 	}
1977 	if (wr->send_flags & IB_SEND_SOLICITED)
1978 		qp->qp1_hdr.bth.solicited_event = 1;
1979 	/* pad_count */
1980 	qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3;
1981 
1982 	/* P_key for QP1 is for all members */
1983 	qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF);
1984 	qp->qp1_hdr.bth.destination_qpn = IB_QP1;
1985 	qp->qp1_hdr.bth.ack_req = 0;
1986 	qp->send_psn++;
1987 	qp->send_psn &= BTH_PSN_MASK;
1988 	qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn);
1989 	/* DETH */
1990 	/* Use the priviledged Q_Key for QP1 */
1991 	qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY);
1992 	qp->qp1_hdr.deth.source_qpn = IB_QP1;
1993 
1994 	/* Pack the QP1 to the transmit buffer */
1995 	buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge);
1996 	if (buf) {
1997 		ib_ud_header_pack(&qp->qp1_hdr, buf);
1998 		for (i = wqe->num_sge; i; i--) {
1999 			wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr;
2000 			wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey;
2001 			wqe->sg_list[i].size = wqe->sg_list[i - 1].size;
2002 		}
2003 
2004 		/*
2005 		 * Max Header buf size for IPV6 RoCE V2 is 86,
2006 		 * which is same as the QP1 SQ header buffer.
2007 		 * Header buf size for IPV4 RoCE V2 can be 66.
2008 		 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
2009 		 * Subtract 20 bytes from QP1 SQ header buf size
2010 		 */
2011 		if (is_udp && ip_version == 4)
2012 			sge.size -= 20;
2013 		/*
2014 		 * Max Header buf size for RoCE V1 is 78.
2015 		 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
2016 		 * Subtract 8 bytes from QP1 SQ header buf size
2017 		 */
2018 		if (!is_udp)
2019 			sge.size -= 8;
2020 
2021 		/* Subtract 4 bytes for non vlan packets */
2022 		if (!is_vlan)
2023 			sge.size -= 4;
2024 
2025 		wqe->sg_list[0].addr = sge.addr;
2026 		wqe->sg_list[0].lkey = sge.lkey;
2027 		wqe->sg_list[0].size = sge.size;
2028 		wqe->num_sge++;
2029 
2030 	} else {
2031 		dev_err(rdev_to_dev(qp->rdev), "QP1 buffer is empty!");
2032 		rc = -ENOMEM;
2033 	}
2034 	return rc;
2035 }
2036 
2037 /* For the MAD layer, it only provides the recv SGE the size of
2038  * ib_grh + MAD datagram.  No Ethernet headers, Ethertype, BTH, DETH,
2039  * nor RoCE iCRC.  The Cu+ solution must provide buffer for the entire
2040  * receive packet (334 bytes) with no VLAN and then copy the GRH
2041  * and the MAD datagram out to the provided SGE.
2042  */
2043 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp,
2044 					    struct ib_recv_wr *wr,
2045 					    struct bnxt_qplib_swqe *wqe,
2046 					    int payload_size)
2047 {
2048 	struct bnxt_qplib_sge ref, sge;
2049 	u32 rq_prod_index;
2050 	struct bnxt_re_sqp_entries *sqp_entry;
2051 
2052 	rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp);
2053 
2054 	if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge))
2055 		return -ENOMEM;
2056 
2057 	/* Create 1 SGE to receive the entire
2058 	 * ethernet packet
2059 	 */
2060 	/* Save the reference from ULP */
2061 	ref.addr = wqe->sg_list[0].addr;
2062 	ref.lkey = wqe->sg_list[0].lkey;
2063 	ref.size = wqe->sg_list[0].size;
2064 
2065 	sqp_entry = &qp->rdev->sqp_tbl[rq_prod_index];
2066 
2067 	/* SGE 1 */
2068 	wqe->sg_list[0].addr = sge.addr;
2069 	wqe->sg_list[0].lkey = sge.lkey;
2070 	wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
2071 	sge.size -= wqe->sg_list[0].size;
2072 
2073 	sqp_entry->sge.addr = ref.addr;
2074 	sqp_entry->sge.lkey = ref.lkey;
2075 	sqp_entry->sge.size = ref.size;
2076 	/* Store the wrid for reporting completion */
2077 	sqp_entry->wrid = wqe->wr_id;
2078 	/* change the wqe->wrid to table index */
2079 	wqe->wr_id = rq_prod_index;
2080 	return 0;
2081 }
2082 
2083 static int is_ud_qp(struct bnxt_re_qp *qp)
2084 {
2085 	return qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD;
2086 }
2087 
2088 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp,
2089 				  struct ib_send_wr *wr,
2090 				  struct bnxt_qplib_swqe *wqe)
2091 {
2092 	struct bnxt_re_ah *ah = NULL;
2093 
2094 	if (is_ud_qp(qp)) {
2095 		ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah);
2096 		wqe->send.q_key = ud_wr(wr)->remote_qkey;
2097 		wqe->send.dst_qp = ud_wr(wr)->remote_qpn;
2098 		wqe->send.avid = ah->qplib_ah.id;
2099 	}
2100 	switch (wr->opcode) {
2101 	case IB_WR_SEND:
2102 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND;
2103 		break;
2104 	case IB_WR_SEND_WITH_IMM:
2105 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM;
2106 		wqe->send.imm_data = wr->ex.imm_data;
2107 		break;
2108 	case IB_WR_SEND_WITH_INV:
2109 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV;
2110 		wqe->send.inv_key = wr->ex.invalidate_rkey;
2111 		break;
2112 	default:
2113 		return -EINVAL;
2114 	}
2115 	if (wr->send_flags & IB_SEND_SIGNALED)
2116 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2117 	if (wr->send_flags & IB_SEND_FENCE)
2118 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2119 	if (wr->send_flags & IB_SEND_SOLICITED)
2120 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2121 	if (wr->send_flags & IB_SEND_INLINE)
2122 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2123 
2124 	return 0;
2125 }
2126 
2127 static int bnxt_re_build_rdma_wqe(struct ib_send_wr *wr,
2128 				  struct bnxt_qplib_swqe *wqe)
2129 {
2130 	switch (wr->opcode) {
2131 	case IB_WR_RDMA_WRITE:
2132 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE;
2133 		break;
2134 	case IB_WR_RDMA_WRITE_WITH_IMM:
2135 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM;
2136 		wqe->rdma.imm_data = wr->ex.imm_data;
2137 		break;
2138 	case IB_WR_RDMA_READ:
2139 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ;
2140 		wqe->rdma.inv_key = wr->ex.invalidate_rkey;
2141 		break;
2142 	default:
2143 		return -EINVAL;
2144 	}
2145 	wqe->rdma.remote_va = rdma_wr(wr)->remote_addr;
2146 	wqe->rdma.r_key = rdma_wr(wr)->rkey;
2147 	if (wr->send_flags & IB_SEND_SIGNALED)
2148 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2149 	if (wr->send_flags & IB_SEND_FENCE)
2150 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2151 	if (wr->send_flags & IB_SEND_SOLICITED)
2152 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2153 	if (wr->send_flags & IB_SEND_INLINE)
2154 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2155 
2156 	return 0;
2157 }
2158 
2159 static int bnxt_re_build_atomic_wqe(struct ib_send_wr *wr,
2160 				    struct bnxt_qplib_swqe *wqe)
2161 {
2162 	switch (wr->opcode) {
2163 	case IB_WR_ATOMIC_CMP_AND_SWP:
2164 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP;
2165 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2166 		wqe->atomic.swap_data = atomic_wr(wr)->swap;
2167 		break;
2168 	case IB_WR_ATOMIC_FETCH_AND_ADD:
2169 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD;
2170 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2171 		break;
2172 	default:
2173 		return -EINVAL;
2174 	}
2175 	wqe->atomic.remote_va = atomic_wr(wr)->remote_addr;
2176 	wqe->atomic.r_key = atomic_wr(wr)->rkey;
2177 	if (wr->send_flags & IB_SEND_SIGNALED)
2178 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2179 	if (wr->send_flags & IB_SEND_FENCE)
2180 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2181 	if (wr->send_flags & IB_SEND_SOLICITED)
2182 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2183 	return 0;
2184 }
2185 
2186 static int bnxt_re_build_inv_wqe(struct ib_send_wr *wr,
2187 				 struct bnxt_qplib_swqe *wqe)
2188 {
2189 	wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV;
2190 	wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey;
2191 
2192 	if (wr->send_flags & IB_SEND_SIGNALED)
2193 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2194 	if (wr->send_flags & IB_SEND_FENCE)
2195 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2196 	if (wr->send_flags & IB_SEND_SOLICITED)
2197 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2198 
2199 	return 0;
2200 }
2201 
2202 static int bnxt_re_build_reg_wqe(struct ib_reg_wr *wr,
2203 				 struct bnxt_qplib_swqe *wqe)
2204 {
2205 	struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr);
2206 	struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl;
2207 	int access = wr->access;
2208 
2209 	wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0];
2210 	wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0];
2211 	wqe->frmr.page_list = mr->pages;
2212 	wqe->frmr.page_list_len = mr->npages;
2213 	wqe->frmr.levels = qplib_frpl->hwq.level + 1;
2214 	wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR;
2215 
2216 	if (wr->wr.send_flags & IB_SEND_FENCE)
2217 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2218 	if (wr->wr.send_flags & IB_SEND_SIGNALED)
2219 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2220 
2221 	if (access & IB_ACCESS_LOCAL_WRITE)
2222 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE;
2223 	if (access & IB_ACCESS_REMOTE_READ)
2224 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ;
2225 	if (access & IB_ACCESS_REMOTE_WRITE)
2226 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE;
2227 	if (access & IB_ACCESS_REMOTE_ATOMIC)
2228 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC;
2229 	if (access & IB_ACCESS_MW_BIND)
2230 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND;
2231 
2232 	wqe->frmr.l_key = wr->key;
2233 	wqe->frmr.length = wr->mr->length;
2234 	wqe->frmr.pbl_pg_sz_log = (wr->mr->page_size >> PAGE_SHIFT_4K) - 1;
2235 	wqe->frmr.va = wr->mr->iova;
2236 	return 0;
2237 }
2238 
2239 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev,
2240 				    struct ib_send_wr *wr,
2241 				    struct bnxt_qplib_swqe *wqe)
2242 {
2243 	/*  Copy the inline data to the data  field */
2244 	u8 *in_data;
2245 	u32 i, sge_len;
2246 	void *sge_addr;
2247 
2248 	in_data = wqe->inline_data;
2249 	for (i = 0; i < wr->num_sge; i++) {
2250 		sge_addr = (void *)(unsigned long)
2251 				wr->sg_list[i].addr;
2252 		sge_len = wr->sg_list[i].length;
2253 
2254 		if ((sge_len + wqe->inline_len) >
2255 		    BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) {
2256 			dev_err(rdev_to_dev(rdev),
2257 				"Inline data size requested > supported value");
2258 			return -EINVAL;
2259 		}
2260 		sge_len = wr->sg_list[i].length;
2261 
2262 		memcpy(in_data, sge_addr, sge_len);
2263 		in_data += wr->sg_list[i].length;
2264 		wqe->inline_len += wr->sg_list[i].length;
2265 	}
2266 	return wqe->inline_len;
2267 }
2268 
2269 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
2270 				   struct ib_send_wr *wr,
2271 				   struct bnxt_qplib_swqe *wqe)
2272 {
2273 	int payload_sz = 0;
2274 
2275 	if (wr->send_flags & IB_SEND_INLINE)
2276 		payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe);
2277 	else
2278 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list,
2279 					       wqe->num_sge);
2280 
2281 	return payload_sz;
2282 }
2283 
2284 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
2285 {
2286 	if ((qp->ib_qp.qp_type == IB_QPT_UD ||
2287 	     qp->ib_qp.qp_type == IB_QPT_GSI ||
2288 	     qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
2289 	     qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
2290 		int qp_attr_mask;
2291 		struct ib_qp_attr qp_attr;
2292 
2293 		qp_attr_mask = IB_QP_STATE;
2294 		qp_attr.qp_state = IB_QPS_RTS;
2295 		bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
2296 		qp->qplib_qp.wqe_cnt = 0;
2297 	}
2298 }
2299 
2300 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
2301 				       struct bnxt_re_qp *qp,
2302 				struct ib_send_wr *wr)
2303 {
2304 	struct bnxt_qplib_swqe wqe;
2305 	int rc = 0, payload_sz = 0;
2306 	unsigned long flags;
2307 
2308 	spin_lock_irqsave(&qp->sq_lock, flags);
2309 	memset(&wqe, 0, sizeof(wqe));
2310 	while (wr) {
2311 		/* House keeping */
2312 		memset(&wqe, 0, sizeof(wqe));
2313 
2314 		/* Common */
2315 		wqe.num_sge = wr->num_sge;
2316 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2317 			dev_err(rdev_to_dev(rdev),
2318 				"Limit exceeded for Send SGEs");
2319 			rc = -EINVAL;
2320 			goto bad;
2321 		}
2322 
2323 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2324 		if (payload_sz < 0) {
2325 			rc = -EINVAL;
2326 			goto bad;
2327 		}
2328 		wqe.wr_id = wr->wr_id;
2329 
2330 		wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND;
2331 
2332 		rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2333 		if (!rc)
2334 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2335 bad:
2336 		if (rc) {
2337 			dev_err(rdev_to_dev(rdev),
2338 				"Post send failed opcode = %#x rc = %d",
2339 				wr->opcode, rc);
2340 			break;
2341 		}
2342 		wr = wr->next;
2343 	}
2344 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2345 	bnxt_ud_qp_hw_stall_workaround(qp);
2346 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2347 	return rc;
2348 }
2349 
2350 int bnxt_re_post_send(struct ib_qp *ib_qp, struct ib_send_wr *wr,
2351 		      struct ib_send_wr **bad_wr)
2352 {
2353 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2354 	struct bnxt_qplib_swqe wqe;
2355 	int rc = 0, payload_sz = 0;
2356 	unsigned long flags;
2357 
2358 	spin_lock_irqsave(&qp->sq_lock, flags);
2359 	while (wr) {
2360 		/* House keeping */
2361 		memset(&wqe, 0, sizeof(wqe));
2362 
2363 		/* Common */
2364 		wqe.num_sge = wr->num_sge;
2365 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2366 			dev_err(rdev_to_dev(qp->rdev),
2367 				"Limit exceeded for Send SGEs");
2368 			rc = -EINVAL;
2369 			goto bad;
2370 		}
2371 
2372 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2373 		if (payload_sz < 0) {
2374 			rc = -EINVAL;
2375 			goto bad;
2376 		}
2377 		wqe.wr_id = wr->wr_id;
2378 
2379 		switch (wr->opcode) {
2380 		case IB_WR_SEND:
2381 		case IB_WR_SEND_WITH_IMM:
2382 			if (ib_qp->qp_type == IB_QPT_GSI) {
2383 				rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe,
2384 							       payload_sz);
2385 				if (rc)
2386 					goto bad;
2387 				wqe.rawqp1.lflags |=
2388 					SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC;
2389 			}
2390 			switch (wr->send_flags) {
2391 			case IB_SEND_IP_CSUM:
2392 				wqe.rawqp1.lflags |=
2393 					SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM;
2394 				break;
2395 			default:
2396 				break;
2397 			}
2398 			/* Fall thru to build the wqe */
2399 		case IB_WR_SEND_WITH_INV:
2400 			rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2401 			break;
2402 		case IB_WR_RDMA_WRITE:
2403 		case IB_WR_RDMA_WRITE_WITH_IMM:
2404 		case IB_WR_RDMA_READ:
2405 			rc = bnxt_re_build_rdma_wqe(wr, &wqe);
2406 			break;
2407 		case IB_WR_ATOMIC_CMP_AND_SWP:
2408 		case IB_WR_ATOMIC_FETCH_AND_ADD:
2409 			rc = bnxt_re_build_atomic_wqe(wr, &wqe);
2410 			break;
2411 		case IB_WR_RDMA_READ_WITH_INV:
2412 			dev_err(rdev_to_dev(qp->rdev),
2413 				"RDMA Read with Invalidate is not supported");
2414 			rc = -EINVAL;
2415 			goto bad;
2416 		case IB_WR_LOCAL_INV:
2417 			rc = bnxt_re_build_inv_wqe(wr, &wqe);
2418 			break;
2419 		case IB_WR_REG_MR:
2420 			rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe);
2421 			break;
2422 		default:
2423 			/* Unsupported WRs */
2424 			dev_err(rdev_to_dev(qp->rdev),
2425 				"WR (%#x) is not supported", wr->opcode);
2426 			rc = -EINVAL;
2427 			goto bad;
2428 		}
2429 		if (!rc)
2430 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2431 bad:
2432 		if (rc) {
2433 			dev_err(rdev_to_dev(qp->rdev),
2434 				"post_send failed op:%#x qps = %#x rc = %d\n",
2435 				wr->opcode, qp->qplib_qp.state, rc);
2436 			*bad_wr = wr;
2437 			break;
2438 		}
2439 		wr = wr->next;
2440 	}
2441 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2442 	bnxt_ud_qp_hw_stall_workaround(qp);
2443 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2444 
2445 	return rc;
2446 }
2447 
2448 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev,
2449 				       struct bnxt_re_qp *qp,
2450 				       struct ib_recv_wr *wr)
2451 {
2452 	struct bnxt_qplib_swqe wqe;
2453 	int rc = 0;
2454 
2455 	memset(&wqe, 0, sizeof(wqe));
2456 	while (wr) {
2457 		/* House keeping */
2458 		memset(&wqe, 0, sizeof(wqe));
2459 
2460 		/* Common */
2461 		wqe.num_sge = wr->num_sge;
2462 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2463 			dev_err(rdev_to_dev(rdev),
2464 				"Limit exceeded for Receive SGEs");
2465 			rc = -EINVAL;
2466 			break;
2467 		}
2468 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
2469 		wqe.wr_id = wr->wr_id;
2470 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2471 
2472 		rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2473 		if (rc)
2474 			break;
2475 
2476 		wr = wr->next;
2477 	}
2478 	if (!rc)
2479 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2480 	return rc;
2481 }
2482 
2483 int bnxt_re_post_recv(struct ib_qp *ib_qp, struct ib_recv_wr *wr,
2484 		      struct ib_recv_wr **bad_wr)
2485 {
2486 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2487 	struct bnxt_qplib_swqe wqe;
2488 	int rc = 0, payload_sz = 0;
2489 	unsigned long flags;
2490 	u32 count = 0;
2491 
2492 	spin_lock_irqsave(&qp->rq_lock, flags);
2493 	while (wr) {
2494 		/* House keeping */
2495 		memset(&wqe, 0, sizeof(wqe));
2496 
2497 		/* Common */
2498 		wqe.num_sge = wr->num_sge;
2499 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2500 			dev_err(rdev_to_dev(qp->rdev),
2501 				"Limit exceeded for Receive SGEs");
2502 			rc = -EINVAL;
2503 			*bad_wr = wr;
2504 			break;
2505 		}
2506 
2507 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
2508 					       wr->num_sge);
2509 		wqe.wr_id = wr->wr_id;
2510 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2511 
2512 		if (ib_qp->qp_type == IB_QPT_GSI)
2513 			rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe,
2514 							      payload_sz);
2515 		if (!rc)
2516 			rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2517 		if (rc) {
2518 			*bad_wr = wr;
2519 			break;
2520 		}
2521 
2522 		/* Ring DB if the RQEs posted reaches a threshold value */
2523 		if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
2524 			bnxt_qplib_post_recv_db(&qp->qplib_qp);
2525 			count = 0;
2526 		}
2527 
2528 		wr = wr->next;
2529 	}
2530 
2531 	if (count)
2532 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2533 
2534 	spin_unlock_irqrestore(&qp->rq_lock, flags);
2535 
2536 	return rc;
2537 }
2538 
2539 /* Completion Queues */
2540 int bnxt_re_destroy_cq(struct ib_cq *ib_cq)
2541 {
2542 	int rc;
2543 	struct bnxt_re_cq *cq;
2544 	struct bnxt_qplib_nq *nq;
2545 	struct bnxt_re_dev *rdev;
2546 
2547 	cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
2548 	rdev = cq->rdev;
2549 	nq = cq->qplib_cq.nq;
2550 
2551 	rc = bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2552 	if (rc) {
2553 		dev_err(rdev_to_dev(rdev), "Failed to destroy HW CQ");
2554 		return rc;
2555 	}
2556 	if (!IS_ERR_OR_NULL(cq->umem))
2557 		ib_umem_release(cq->umem);
2558 
2559 	atomic_dec(&rdev->cq_count);
2560 	nq->budget--;
2561 	kfree(cq->cql);
2562 	kfree(cq);
2563 
2564 	return 0;
2565 }
2566 
2567 struct ib_cq *bnxt_re_create_cq(struct ib_device *ibdev,
2568 				const struct ib_cq_init_attr *attr,
2569 				struct ib_ucontext *context,
2570 				struct ib_udata *udata)
2571 {
2572 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
2573 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
2574 	struct bnxt_re_cq *cq = NULL;
2575 	int rc, entries;
2576 	int cqe = attr->cqe;
2577 	struct bnxt_qplib_nq *nq = NULL;
2578 	unsigned int nq_alloc_cnt;
2579 
2580 	/* Validate CQ fields */
2581 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
2582 		dev_err(rdev_to_dev(rdev), "Failed to create CQ -max exceeded");
2583 		return ERR_PTR(-EINVAL);
2584 	}
2585 	cq = kzalloc(sizeof(*cq), GFP_KERNEL);
2586 	if (!cq)
2587 		return ERR_PTR(-ENOMEM);
2588 
2589 	cq->rdev = rdev;
2590 	cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
2591 
2592 	entries = roundup_pow_of_two(cqe + 1);
2593 	if (entries > dev_attr->max_cq_wqes + 1)
2594 		entries = dev_attr->max_cq_wqes + 1;
2595 
2596 	if (context) {
2597 		struct bnxt_re_cq_req req;
2598 		struct bnxt_re_ucontext *uctx = container_of
2599 						(context,
2600 						 struct bnxt_re_ucontext,
2601 						 ib_uctx);
2602 		if (ib_copy_from_udata(&req, udata, sizeof(req))) {
2603 			rc = -EFAULT;
2604 			goto fail;
2605 		}
2606 
2607 		cq->umem = ib_umem_get(context, req.cq_va,
2608 				       entries * sizeof(struct cq_base),
2609 				       IB_ACCESS_LOCAL_WRITE, 1);
2610 		if (IS_ERR(cq->umem)) {
2611 			rc = PTR_ERR(cq->umem);
2612 			goto fail;
2613 		}
2614 		cq->qplib_cq.sghead = cq->umem->sg_head.sgl;
2615 		cq->qplib_cq.nmap = cq->umem->nmap;
2616 		cq->qplib_cq.dpi = &uctx->dpi;
2617 	} else {
2618 		cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL);
2619 		cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe),
2620 				  GFP_KERNEL);
2621 		if (!cq->cql) {
2622 			rc = -ENOMEM;
2623 			goto fail;
2624 		}
2625 
2626 		cq->qplib_cq.dpi = &rdev->dpi_privileged;
2627 		cq->qplib_cq.sghead = NULL;
2628 		cq->qplib_cq.nmap = 0;
2629 	}
2630 	/*
2631 	 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
2632 	 * used for getting the NQ index.
2633 	 */
2634 	nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
2635 	nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
2636 	cq->qplib_cq.max_wqe = entries;
2637 	cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
2638 	cq->qplib_cq.nq	= nq;
2639 
2640 	rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
2641 	if (rc) {
2642 		dev_err(rdev_to_dev(rdev), "Failed to create HW CQ");
2643 		goto fail;
2644 	}
2645 
2646 	cq->ib_cq.cqe = entries;
2647 	cq->cq_period = cq->qplib_cq.period;
2648 	nq->budget++;
2649 
2650 	atomic_inc(&rdev->cq_count);
2651 
2652 	if (context) {
2653 		struct bnxt_re_cq_resp resp;
2654 
2655 		resp.cqid = cq->qplib_cq.id;
2656 		resp.tail = cq->qplib_cq.hwq.cons;
2657 		resp.phase = cq->qplib_cq.period;
2658 		resp.rsvd = 0;
2659 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
2660 		if (rc) {
2661 			dev_err(rdev_to_dev(rdev), "Failed to copy CQ udata");
2662 			bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2663 			goto c2fail;
2664 		}
2665 	}
2666 
2667 	return &cq->ib_cq;
2668 
2669 c2fail:
2670 	if (context)
2671 		ib_umem_release(cq->umem);
2672 fail:
2673 	kfree(cq->cql);
2674 	kfree(cq);
2675 	return ERR_PTR(rc);
2676 }
2677 
2678 static u8 __req_to_ib_wc_status(u8 qstatus)
2679 {
2680 	switch (qstatus) {
2681 	case CQ_REQ_STATUS_OK:
2682 		return IB_WC_SUCCESS;
2683 	case CQ_REQ_STATUS_BAD_RESPONSE_ERR:
2684 		return IB_WC_BAD_RESP_ERR;
2685 	case CQ_REQ_STATUS_LOCAL_LENGTH_ERR:
2686 		return IB_WC_LOC_LEN_ERR;
2687 	case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR:
2688 		return IB_WC_LOC_QP_OP_ERR;
2689 	case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR:
2690 		return IB_WC_LOC_PROT_ERR;
2691 	case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR:
2692 		return IB_WC_GENERAL_ERR;
2693 	case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR:
2694 		return IB_WC_REM_INV_REQ_ERR;
2695 	case CQ_REQ_STATUS_REMOTE_ACCESS_ERR:
2696 		return IB_WC_REM_ACCESS_ERR;
2697 	case CQ_REQ_STATUS_REMOTE_OPERATION_ERR:
2698 		return IB_WC_REM_OP_ERR;
2699 	case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR:
2700 		return IB_WC_RNR_RETRY_EXC_ERR;
2701 	case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR:
2702 		return IB_WC_RETRY_EXC_ERR;
2703 	case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR:
2704 		return IB_WC_WR_FLUSH_ERR;
2705 	default:
2706 		return IB_WC_GENERAL_ERR;
2707 	}
2708 	return 0;
2709 }
2710 
2711 static u8 __rawqp1_to_ib_wc_status(u8 qstatus)
2712 {
2713 	switch (qstatus) {
2714 	case CQ_RES_RAWETH_QP1_STATUS_OK:
2715 		return IB_WC_SUCCESS;
2716 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR:
2717 		return IB_WC_LOC_ACCESS_ERR;
2718 	case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR:
2719 		return IB_WC_LOC_LEN_ERR;
2720 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR:
2721 		return IB_WC_LOC_PROT_ERR;
2722 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR:
2723 		return IB_WC_LOC_QP_OP_ERR;
2724 	case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR:
2725 		return IB_WC_GENERAL_ERR;
2726 	case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR:
2727 		return IB_WC_WR_FLUSH_ERR;
2728 	case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR:
2729 		return IB_WC_WR_FLUSH_ERR;
2730 	default:
2731 		return IB_WC_GENERAL_ERR;
2732 	}
2733 }
2734 
2735 static u8 __rc_to_ib_wc_status(u8 qstatus)
2736 {
2737 	switch (qstatus) {
2738 	case CQ_RES_RC_STATUS_OK:
2739 		return IB_WC_SUCCESS;
2740 	case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR:
2741 		return IB_WC_LOC_ACCESS_ERR;
2742 	case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR:
2743 		return IB_WC_LOC_LEN_ERR;
2744 	case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR:
2745 		return IB_WC_LOC_PROT_ERR;
2746 	case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR:
2747 		return IB_WC_LOC_QP_OP_ERR;
2748 	case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR:
2749 		return IB_WC_GENERAL_ERR;
2750 	case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR:
2751 		return IB_WC_REM_INV_REQ_ERR;
2752 	case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR:
2753 		return IB_WC_WR_FLUSH_ERR;
2754 	case CQ_RES_RC_STATUS_HW_FLUSH_ERR:
2755 		return IB_WC_WR_FLUSH_ERR;
2756 	default:
2757 		return IB_WC_GENERAL_ERR;
2758 	}
2759 }
2760 
2761 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe)
2762 {
2763 	switch (cqe->type) {
2764 	case BNXT_QPLIB_SWQE_TYPE_SEND:
2765 		wc->opcode = IB_WC_SEND;
2766 		break;
2767 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM:
2768 		wc->opcode = IB_WC_SEND;
2769 		wc->wc_flags |= IB_WC_WITH_IMM;
2770 		break;
2771 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV:
2772 		wc->opcode = IB_WC_SEND;
2773 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
2774 		break;
2775 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE:
2776 		wc->opcode = IB_WC_RDMA_WRITE;
2777 		break;
2778 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM:
2779 		wc->opcode = IB_WC_RDMA_WRITE;
2780 		wc->wc_flags |= IB_WC_WITH_IMM;
2781 		break;
2782 	case BNXT_QPLIB_SWQE_TYPE_RDMA_READ:
2783 		wc->opcode = IB_WC_RDMA_READ;
2784 		break;
2785 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP:
2786 		wc->opcode = IB_WC_COMP_SWAP;
2787 		break;
2788 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD:
2789 		wc->opcode = IB_WC_FETCH_ADD;
2790 		break;
2791 	case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV:
2792 		wc->opcode = IB_WC_LOCAL_INV;
2793 		break;
2794 	case BNXT_QPLIB_SWQE_TYPE_REG_MR:
2795 		wc->opcode = IB_WC_REG_MR;
2796 		break;
2797 	default:
2798 		wc->opcode = IB_WC_SEND;
2799 		break;
2800 	}
2801 
2802 	wc->status = __req_to_ib_wc_status(cqe->status);
2803 }
2804 
2805 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags,
2806 				     u16 raweth_qp1_flags2)
2807 {
2808 	bool is_ipv6 = false, is_ipv4 = false;
2809 
2810 	/* raweth_qp1_flags Bit 9-6 indicates itype */
2811 	if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
2812 	    != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
2813 		return -1;
2814 
2815 	if (raweth_qp1_flags2 &
2816 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC &&
2817 	    raweth_qp1_flags2 &
2818 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) {
2819 		/* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
2820 		(raweth_qp1_flags2 &
2821 		 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ?
2822 			(is_ipv6 = true) : (is_ipv4 = true);
2823 		return ((is_ipv6) ?
2824 			 BNXT_RE_ROCEV2_IPV6_PACKET :
2825 			 BNXT_RE_ROCEV2_IPV4_PACKET);
2826 	} else {
2827 		return BNXT_RE_ROCE_V1_PACKET;
2828 	}
2829 }
2830 
2831 static int bnxt_re_to_ib_nw_type(int nw_type)
2832 {
2833 	u8 nw_hdr_type = 0xFF;
2834 
2835 	switch (nw_type) {
2836 	case BNXT_RE_ROCE_V1_PACKET:
2837 		nw_hdr_type = RDMA_NETWORK_ROCE_V1;
2838 		break;
2839 	case BNXT_RE_ROCEV2_IPV4_PACKET:
2840 		nw_hdr_type = RDMA_NETWORK_IPV4;
2841 		break;
2842 	case BNXT_RE_ROCEV2_IPV6_PACKET:
2843 		nw_hdr_type = RDMA_NETWORK_IPV6;
2844 		break;
2845 	}
2846 	return nw_hdr_type;
2847 }
2848 
2849 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev,
2850 				       void *rq_hdr_buf)
2851 {
2852 	u8 *tmp_buf = NULL;
2853 	struct ethhdr *eth_hdr;
2854 	u16 eth_type;
2855 	bool rc = false;
2856 
2857 	tmp_buf = (u8 *)rq_hdr_buf;
2858 	/*
2859 	 * If dest mac is not same as I/F mac, this could be a
2860 	 * loopback address or multicast address, check whether
2861 	 * it is a loopback packet
2862 	 */
2863 	if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) {
2864 		tmp_buf += 4;
2865 		/* Check the  ether type */
2866 		eth_hdr = (struct ethhdr *)tmp_buf;
2867 		eth_type = ntohs(eth_hdr->h_proto);
2868 		switch (eth_type) {
2869 		case ETH_P_IBOE:
2870 			rc = true;
2871 			break;
2872 		case ETH_P_IP:
2873 		case ETH_P_IPV6: {
2874 			u32 len;
2875 			struct udphdr *udp_hdr;
2876 
2877 			len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) :
2878 						      sizeof(struct ipv6hdr));
2879 			tmp_buf += sizeof(struct ethhdr) + len;
2880 			udp_hdr = (struct udphdr *)tmp_buf;
2881 			if (ntohs(udp_hdr->dest) ==
2882 				    ROCE_V2_UDP_DPORT)
2883 				rc = true;
2884 			break;
2885 			}
2886 		default:
2887 			break;
2888 		}
2889 	}
2890 
2891 	return rc;
2892 }
2893 
2894 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *qp1_qp,
2895 					 struct bnxt_qplib_cqe *cqe)
2896 {
2897 	struct bnxt_re_dev *rdev = qp1_qp->rdev;
2898 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
2899 	struct bnxt_re_qp *qp = rdev->qp1_sqp;
2900 	struct ib_send_wr *swr;
2901 	struct ib_ud_wr udwr;
2902 	struct ib_recv_wr rwr;
2903 	int pkt_type = 0;
2904 	u32 tbl_idx;
2905 	void *rq_hdr_buf;
2906 	dma_addr_t rq_hdr_buf_map;
2907 	dma_addr_t shrq_hdr_buf_map;
2908 	u32 offset = 0;
2909 	u32 skip_bytes = 0;
2910 	struct ib_sge s_sge[2];
2911 	struct ib_sge r_sge[2];
2912 	int rc;
2913 
2914 	memset(&udwr, 0, sizeof(udwr));
2915 	memset(&rwr, 0, sizeof(rwr));
2916 	memset(&s_sge, 0, sizeof(s_sge));
2917 	memset(&r_sge, 0, sizeof(r_sge));
2918 
2919 	swr = &udwr.wr;
2920 	tbl_idx = cqe->wr_id;
2921 
2922 	rq_hdr_buf = qp1_qp->qplib_qp.rq_hdr_buf +
2923 			(tbl_idx * qp1_qp->qplib_qp.rq_hdr_buf_size);
2924 	rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&qp1_qp->qplib_qp,
2925 							  tbl_idx);
2926 
2927 	/* Shadow QP header buffer */
2928 	shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&qp->qplib_qp,
2929 							    tbl_idx);
2930 	sqp_entry = &rdev->sqp_tbl[tbl_idx];
2931 
2932 	/* Store this cqe */
2933 	memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe));
2934 	sqp_entry->qp1_qp = qp1_qp;
2935 
2936 	/* Find packet type from the cqe */
2937 
2938 	pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags,
2939 					     cqe->raweth_qp1_flags2);
2940 	if (pkt_type < 0) {
2941 		dev_err(rdev_to_dev(rdev), "Invalid packet\n");
2942 		return -EINVAL;
2943 	}
2944 
2945 	/* Adjust the offset for the user buffer and post in the rq */
2946 
2947 	if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET)
2948 		offset = 20;
2949 
2950 	/*
2951 	 * QP1 loopback packet has 4 bytes of internal header before
2952 	 * ether header. Skip these four bytes.
2953 	 */
2954 	if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf))
2955 		skip_bytes = 4;
2956 
2957 	/* First send SGE . Skip the ether header*/
2958 	s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
2959 			+ skip_bytes;
2960 	s_sge[0].lkey = 0xFFFFFFFF;
2961 	s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 :
2962 				BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
2963 
2964 	/* Second Send SGE */
2965 	s_sge[1].addr = s_sge[0].addr + s_sge[0].length +
2966 			BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE;
2967 	if (pkt_type != BNXT_RE_ROCE_V1_PACKET)
2968 		s_sge[1].addr += 8;
2969 	s_sge[1].lkey = 0xFFFFFFFF;
2970 	s_sge[1].length = 256;
2971 
2972 	/* First recv SGE */
2973 
2974 	r_sge[0].addr = shrq_hdr_buf_map;
2975 	r_sge[0].lkey = 0xFFFFFFFF;
2976 	r_sge[0].length = 40;
2977 
2978 	r_sge[1].addr = sqp_entry->sge.addr + offset;
2979 	r_sge[1].lkey = sqp_entry->sge.lkey;
2980 	r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset;
2981 
2982 	/* Create receive work request */
2983 	rwr.num_sge = 2;
2984 	rwr.sg_list = r_sge;
2985 	rwr.wr_id = tbl_idx;
2986 	rwr.next = NULL;
2987 
2988 	rc = bnxt_re_post_recv_shadow_qp(rdev, qp, &rwr);
2989 	if (rc) {
2990 		dev_err(rdev_to_dev(rdev),
2991 			"Failed to post Rx buffers to shadow QP");
2992 		return -ENOMEM;
2993 	}
2994 
2995 	swr->num_sge = 2;
2996 	swr->sg_list = s_sge;
2997 	swr->wr_id = tbl_idx;
2998 	swr->opcode = IB_WR_SEND;
2999 	swr->next = NULL;
3000 
3001 	udwr.ah = &rdev->sqp_ah->ib_ah;
3002 	udwr.remote_qpn = rdev->qp1_sqp->qplib_qp.id;
3003 	udwr.remote_qkey = rdev->qp1_sqp->qplib_qp.qkey;
3004 
3005 	/* post data received  in the send queue */
3006 	rc = bnxt_re_post_send_shadow_qp(rdev, qp, swr);
3007 
3008 	return 0;
3009 }
3010 
3011 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc,
3012 					  struct bnxt_qplib_cqe *cqe)
3013 {
3014 	wc->opcode = IB_WC_RECV;
3015 	wc->status = __rawqp1_to_ib_wc_status(cqe->status);
3016 	wc->wc_flags |= IB_WC_GRH;
3017 }
3018 
3019 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe,
3020 				u16 *vid, u8 *sl)
3021 {
3022 	bool ret = false;
3023 	u32 metadata;
3024 	u16 tpid;
3025 
3026 	metadata = orig_cqe->raweth_qp1_metadata;
3027 	if (orig_cqe->raweth_qp1_flags2 &
3028 		CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) {
3029 		tpid = ((metadata &
3030 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >>
3031 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT);
3032 		if (tpid == ETH_P_8021Q) {
3033 			*vid = metadata &
3034 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK;
3035 			*sl = (metadata &
3036 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >>
3037 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT;
3038 			ret = true;
3039 		}
3040 	}
3041 
3042 	return ret;
3043 }
3044 
3045 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc,
3046 				      struct bnxt_qplib_cqe *cqe)
3047 {
3048 	wc->opcode = IB_WC_RECV;
3049 	wc->status = __rc_to_ib_wc_status(cqe->status);
3050 
3051 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3052 		wc->wc_flags |= IB_WC_WITH_IMM;
3053 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3054 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3055 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3056 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3057 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3058 }
3059 
3060 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *qp,
3061 					     struct ib_wc *wc,
3062 					     struct bnxt_qplib_cqe *cqe)
3063 {
3064 	struct bnxt_re_dev *rdev = qp->rdev;
3065 	struct bnxt_re_qp *qp1_qp = NULL;
3066 	struct bnxt_qplib_cqe *orig_cqe = NULL;
3067 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3068 	int nw_type;
3069 	u32 tbl_idx;
3070 	u16 vlan_id;
3071 	u8 sl;
3072 
3073 	tbl_idx = cqe->wr_id;
3074 
3075 	sqp_entry = &rdev->sqp_tbl[tbl_idx];
3076 	qp1_qp = sqp_entry->qp1_qp;
3077 	orig_cqe = &sqp_entry->cqe;
3078 
3079 	wc->wr_id = sqp_entry->wrid;
3080 	wc->byte_len = orig_cqe->length;
3081 	wc->qp = &qp1_qp->ib_qp;
3082 
3083 	wc->ex.imm_data = orig_cqe->immdata;
3084 	wc->src_qp = orig_cqe->src_qp;
3085 	memcpy(wc->smac, orig_cqe->smac, ETH_ALEN);
3086 	if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) {
3087 		wc->vlan_id = vlan_id;
3088 		wc->sl = sl;
3089 		wc->wc_flags |= IB_WC_WITH_VLAN;
3090 	}
3091 	wc->port_num = 1;
3092 	wc->vendor_err = orig_cqe->status;
3093 
3094 	wc->opcode = IB_WC_RECV;
3095 	wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status);
3096 	wc->wc_flags |= IB_WC_GRH;
3097 
3098 	nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags,
3099 					    orig_cqe->raweth_qp1_flags2);
3100 	if (nw_type >= 0) {
3101 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3102 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3103 	}
3104 }
3105 
3106 static void bnxt_re_process_res_ud_wc(struct ib_wc *wc,
3107 				      struct bnxt_qplib_cqe *cqe)
3108 {
3109 	wc->opcode = IB_WC_RECV;
3110 	wc->status = __rc_to_ib_wc_status(cqe->status);
3111 
3112 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3113 		wc->wc_flags |= IB_WC_WITH_IMM;
3114 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3115 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3116 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3117 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3118 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3119 }
3120 
3121 static int send_phantom_wqe(struct bnxt_re_qp *qp)
3122 {
3123 	struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
3124 	unsigned long flags;
3125 	int rc = 0;
3126 
3127 	spin_lock_irqsave(&qp->sq_lock, flags);
3128 
3129 	rc = bnxt_re_bind_fence_mw(lib_qp);
3130 	if (!rc) {
3131 		lib_qp->sq.phantom_wqe_cnt++;
3132 		dev_dbg(&lib_qp->sq.hwq.pdev->dev,
3133 			"qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
3134 			lib_qp->id, lib_qp->sq.hwq.prod,
3135 			HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
3136 			lib_qp->sq.phantom_wqe_cnt);
3137 	}
3138 
3139 	spin_unlock_irqrestore(&qp->sq_lock, flags);
3140 	return rc;
3141 }
3142 
3143 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
3144 {
3145 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3146 	struct bnxt_re_qp *qp;
3147 	struct bnxt_qplib_cqe *cqe;
3148 	int i, ncqe, budget;
3149 	struct bnxt_qplib_q *sq;
3150 	struct bnxt_qplib_qp *lib_qp;
3151 	u32 tbl_idx;
3152 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3153 	unsigned long flags;
3154 
3155 	spin_lock_irqsave(&cq->cq_lock, flags);
3156 	budget = min_t(u32, num_entries, cq->max_cql);
3157 	num_entries = budget;
3158 	if (!cq->cql) {
3159 		dev_err(rdev_to_dev(cq->rdev), "POLL CQ : no CQL to use");
3160 		goto exit;
3161 	}
3162 	cqe = &cq->cql[0];
3163 	while (budget) {
3164 		lib_qp = NULL;
3165 		ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
3166 		if (lib_qp) {
3167 			sq = &lib_qp->sq;
3168 			if (sq->send_phantom) {
3169 				qp = container_of(lib_qp,
3170 						  struct bnxt_re_qp, qplib_qp);
3171 				if (send_phantom_wqe(qp) == -ENOMEM)
3172 					dev_err(rdev_to_dev(cq->rdev),
3173 						"Phantom failed! Scheduled to send again\n");
3174 				else
3175 					sq->send_phantom = false;
3176 			}
3177 		}
3178 		if (ncqe < budget)
3179 			ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq,
3180 							      cqe + ncqe,
3181 							      budget - ncqe);
3182 
3183 		if (!ncqe)
3184 			break;
3185 
3186 		for (i = 0; i < ncqe; i++, cqe++) {
3187 			/* Transcribe each qplib_wqe back to ib_wc */
3188 			memset(wc, 0, sizeof(*wc));
3189 
3190 			wc->wr_id = cqe->wr_id;
3191 			wc->byte_len = cqe->length;
3192 			qp = container_of
3193 				((struct bnxt_qplib_qp *)
3194 				 (unsigned long)(cqe->qp_handle),
3195 				 struct bnxt_re_qp, qplib_qp);
3196 			if (!qp) {
3197 				dev_err(rdev_to_dev(cq->rdev),
3198 					"POLL CQ : bad QP handle");
3199 				continue;
3200 			}
3201 			wc->qp = &qp->ib_qp;
3202 			wc->ex.imm_data = cqe->immdata;
3203 			wc->src_qp = cqe->src_qp;
3204 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
3205 			wc->port_num = 1;
3206 			wc->vendor_err = cqe->status;
3207 
3208 			switch (cqe->opcode) {
3209 			case CQ_BASE_CQE_TYPE_REQ:
3210 				if (qp->qplib_qp.id ==
3211 				    qp->rdev->qp1_sqp->qplib_qp.id) {
3212 					/* Handle this completion with
3213 					 * the stored completion
3214 					 */
3215 					memset(wc, 0, sizeof(*wc));
3216 					continue;
3217 				}
3218 				bnxt_re_process_req_wc(wc, cqe);
3219 				break;
3220 			case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1:
3221 				if (!cqe->status) {
3222 					int rc = 0;
3223 
3224 					rc = bnxt_re_process_raw_qp_pkt_rx
3225 								(qp, cqe);
3226 					if (!rc) {
3227 						memset(wc, 0, sizeof(*wc));
3228 						continue;
3229 					}
3230 					cqe->status = -1;
3231 				}
3232 				/* Errors need not be looped back.
3233 				 * But change the wr_id to the one
3234 				 * stored in the table
3235 				 */
3236 				tbl_idx = cqe->wr_id;
3237 				sqp_entry = &cq->rdev->sqp_tbl[tbl_idx];
3238 				wc->wr_id = sqp_entry->wrid;
3239 				bnxt_re_process_res_rawqp1_wc(wc, cqe);
3240 				break;
3241 			case CQ_BASE_CQE_TYPE_RES_RC:
3242 				bnxt_re_process_res_rc_wc(wc, cqe);
3243 				break;
3244 			case CQ_BASE_CQE_TYPE_RES_UD:
3245 				if (qp->qplib_qp.id ==
3246 				    qp->rdev->qp1_sqp->qplib_qp.id) {
3247 					/* Handle this completion with
3248 					 * the stored completion
3249 					 */
3250 					if (cqe->status) {
3251 						continue;
3252 					} else {
3253 						bnxt_re_process_res_shadow_qp_wc
3254 								(qp, wc, cqe);
3255 						break;
3256 					}
3257 				}
3258 				bnxt_re_process_res_ud_wc(wc, cqe);
3259 				break;
3260 			default:
3261 				dev_err(rdev_to_dev(cq->rdev),
3262 					"POLL CQ : type 0x%x not handled",
3263 					cqe->opcode);
3264 				continue;
3265 			}
3266 			wc++;
3267 			budget--;
3268 		}
3269 	}
3270 exit:
3271 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3272 	return num_entries - budget;
3273 }
3274 
3275 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq,
3276 			  enum ib_cq_notify_flags ib_cqn_flags)
3277 {
3278 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3279 	int type = 0, rc = 0;
3280 	unsigned long flags;
3281 
3282 	spin_lock_irqsave(&cq->cq_lock, flags);
3283 	/* Trigger on the very next completion */
3284 	if (ib_cqn_flags & IB_CQ_NEXT_COMP)
3285 		type = DBR_DBR_TYPE_CQ_ARMALL;
3286 	/* Trigger on the next solicited completion */
3287 	else if (ib_cqn_flags & IB_CQ_SOLICITED)
3288 		type = DBR_DBR_TYPE_CQ_ARMSE;
3289 
3290 	/* Poll to see if there are missed events */
3291 	if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3292 	    !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) {
3293 		rc = 1;
3294 		goto exit;
3295 	}
3296 	bnxt_qplib_req_notify_cq(&cq->qplib_cq, type);
3297 
3298 exit:
3299 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3300 	return rc;
3301 }
3302 
3303 /* Memory Regions */
3304 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags)
3305 {
3306 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3307 	struct bnxt_re_dev *rdev = pd->rdev;
3308 	struct bnxt_re_mr *mr;
3309 	u64 pbl = 0;
3310 	int rc;
3311 
3312 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3313 	if (!mr)
3314 		return ERR_PTR(-ENOMEM);
3315 
3316 	mr->rdev = rdev;
3317 	mr->qplib_mr.pd = &pd->qplib_pd;
3318 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3319 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3320 
3321 	/* Allocate and register 0 as the address */
3322 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3323 	if (rc)
3324 		goto fail;
3325 
3326 	mr->qplib_mr.hwq.level = PBL_LVL_MAX;
3327 	mr->qplib_mr.total_size = -1; /* Infinte length */
3328 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, &pbl, 0, false,
3329 			       PAGE_SIZE);
3330 	if (rc)
3331 		goto fail_mr;
3332 
3333 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3334 	if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ |
3335 			       IB_ACCESS_REMOTE_ATOMIC))
3336 		mr->ib_mr.rkey = mr->ib_mr.lkey;
3337 	atomic_inc(&rdev->mr_count);
3338 
3339 	return &mr->ib_mr;
3340 
3341 fail_mr:
3342 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3343 fail:
3344 	kfree(mr);
3345 	return ERR_PTR(rc);
3346 }
3347 
3348 int bnxt_re_dereg_mr(struct ib_mr *ib_mr)
3349 {
3350 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3351 	struct bnxt_re_dev *rdev = mr->rdev;
3352 	int rc;
3353 
3354 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3355 	if (rc)
3356 		dev_err(rdev_to_dev(rdev), "Dereg MR failed: %#x\n", rc);
3357 
3358 	if (mr->pages) {
3359 		rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
3360 							&mr->qplib_frpl);
3361 		kfree(mr->pages);
3362 		mr->npages = 0;
3363 		mr->pages = NULL;
3364 	}
3365 	if (!IS_ERR_OR_NULL(mr->ib_umem))
3366 		ib_umem_release(mr->ib_umem);
3367 
3368 	kfree(mr);
3369 	atomic_dec(&rdev->mr_count);
3370 	return rc;
3371 }
3372 
3373 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr)
3374 {
3375 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3376 
3377 	if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs))
3378 		return -ENOMEM;
3379 
3380 	mr->pages[mr->npages++] = addr;
3381 	return 0;
3382 }
3383 
3384 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents,
3385 		      unsigned int *sg_offset)
3386 {
3387 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3388 
3389 	mr->npages = 0;
3390 	return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page);
3391 }
3392 
3393 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type,
3394 			       u32 max_num_sg)
3395 {
3396 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3397 	struct bnxt_re_dev *rdev = pd->rdev;
3398 	struct bnxt_re_mr *mr = NULL;
3399 	int rc;
3400 
3401 	if (type != IB_MR_TYPE_MEM_REG) {
3402 		dev_dbg(rdev_to_dev(rdev), "MR type 0x%x not supported", type);
3403 		return ERR_PTR(-EINVAL);
3404 	}
3405 	if (max_num_sg > MAX_PBL_LVL_1_PGS)
3406 		return ERR_PTR(-EINVAL);
3407 
3408 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3409 	if (!mr)
3410 		return ERR_PTR(-ENOMEM);
3411 
3412 	mr->rdev = rdev;
3413 	mr->qplib_mr.pd = &pd->qplib_pd;
3414 	mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR;
3415 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3416 
3417 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3418 	if (rc)
3419 		goto bail;
3420 
3421 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3422 	mr->ib_mr.rkey = mr->ib_mr.lkey;
3423 
3424 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3425 	if (!mr->pages) {
3426 		rc = -ENOMEM;
3427 		goto fail;
3428 	}
3429 	rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res,
3430 						 &mr->qplib_frpl, max_num_sg);
3431 	if (rc) {
3432 		dev_err(rdev_to_dev(rdev),
3433 			"Failed to allocate HW FR page list");
3434 		goto fail_mr;
3435 	}
3436 
3437 	atomic_inc(&rdev->mr_count);
3438 	return &mr->ib_mr;
3439 
3440 fail_mr:
3441 	kfree(mr->pages);
3442 fail:
3443 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3444 bail:
3445 	kfree(mr);
3446 	return ERR_PTR(rc);
3447 }
3448 
3449 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
3450 			       struct ib_udata *udata)
3451 {
3452 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3453 	struct bnxt_re_dev *rdev = pd->rdev;
3454 	struct bnxt_re_mw *mw;
3455 	int rc;
3456 
3457 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
3458 	if (!mw)
3459 		return ERR_PTR(-ENOMEM);
3460 	mw->rdev = rdev;
3461 	mw->qplib_mw.pd = &pd->qplib_pd;
3462 
3463 	mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
3464 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
3465 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
3466 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
3467 	if (rc) {
3468 		dev_err(rdev_to_dev(rdev), "Allocate MW failed!");
3469 		goto fail;
3470 	}
3471 	mw->ib_mw.rkey = mw->qplib_mw.rkey;
3472 
3473 	atomic_inc(&rdev->mw_count);
3474 	return &mw->ib_mw;
3475 
3476 fail:
3477 	kfree(mw);
3478 	return ERR_PTR(rc);
3479 }
3480 
3481 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
3482 {
3483 	struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
3484 	struct bnxt_re_dev *rdev = mw->rdev;
3485 	int rc;
3486 
3487 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
3488 	if (rc) {
3489 		dev_err(rdev_to_dev(rdev), "Free MW failed: %#x\n", rc);
3490 		return rc;
3491 	}
3492 
3493 	kfree(mw);
3494 	atomic_dec(&rdev->mw_count);
3495 	return rc;
3496 }
3497 
3498 static int bnxt_re_page_size_ok(int page_shift)
3499 {
3500 	switch (page_shift) {
3501 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4K:
3502 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_8K:
3503 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_64K:
3504 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_2M:
3505 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_256K:
3506 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1M:
3507 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_4M:
3508 	case CMDQ_REGISTER_MR_LOG2_PBL_PG_SIZE_PG_1G:
3509 		return 1;
3510 	default:
3511 		return 0;
3512 	}
3513 }
3514 
3515 static int fill_umem_pbl_tbl(struct ib_umem *umem, u64 *pbl_tbl_orig,
3516 			     int page_shift)
3517 {
3518 	u64 *pbl_tbl = pbl_tbl_orig;
3519 	u64 paddr;
3520 	u64 page_mask = (1ULL << page_shift) - 1;
3521 	int i, pages;
3522 	struct scatterlist *sg;
3523 	int entry;
3524 
3525 	for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
3526 		pages = sg_dma_len(sg) >> PAGE_SHIFT;
3527 		for (i = 0; i < pages; i++) {
3528 			paddr = sg_dma_address(sg) + (i << PAGE_SHIFT);
3529 			if (pbl_tbl == pbl_tbl_orig)
3530 				*pbl_tbl++ = paddr & ~page_mask;
3531 			else if ((paddr & page_mask) == 0)
3532 				*pbl_tbl++ = paddr;
3533 		}
3534 	}
3535 	return pbl_tbl - pbl_tbl_orig;
3536 }
3537 
3538 /* uverbs */
3539 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
3540 				  u64 virt_addr, int mr_access_flags,
3541 				  struct ib_udata *udata)
3542 {
3543 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3544 	struct bnxt_re_dev *rdev = pd->rdev;
3545 	struct bnxt_re_mr *mr;
3546 	struct ib_umem *umem;
3547 	u64 *pbl_tbl = NULL;
3548 	int umem_pgs, page_shift, rc;
3549 
3550 	if (length > BNXT_RE_MAX_MR_SIZE) {
3551 		dev_err(rdev_to_dev(rdev), "MR Size: %lld > Max supported:%ld\n",
3552 			length, BNXT_RE_MAX_MR_SIZE);
3553 		return ERR_PTR(-ENOMEM);
3554 	}
3555 
3556 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3557 	if (!mr)
3558 		return ERR_PTR(-ENOMEM);
3559 
3560 	mr->rdev = rdev;
3561 	mr->qplib_mr.pd = &pd->qplib_pd;
3562 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3563 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
3564 
3565 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3566 	if (rc) {
3567 		dev_err(rdev_to_dev(rdev), "Failed to allocate MR");
3568 		goto free_mr;
3569 	}
3570 	/* The fixed portion of the rkey is the same as the lkey */
3571 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
3572 
3573 	umem = ib_umem_get(ib_pd->uobject->context, start, length,
3574 			   mr_access_flags, 0);
3575 	if (IS_ERR(umem)) {
3576 		dev_err(rdev_to_dev(rdev), "Failed to get umem");
3577 		rc = -EFAULT;
3578 		goto free_mrw;
3579 	}
3580 	mr->ib_umem = umem;
3581 
3582 	mr->qplib_mr.va = virt_addr;
3583 	umem_pgs = ib_umem_page_count(umem);
3584 	if (!umem_pgs) {
3585 		dev_err(rdev_to_dev(rdev), "umem is invalid!");
3586 		rc = -EINVAL;
3587 		goto free_umem;
3588 	}
3589 	mr->qplib_mr.total_size = length;
3590 
3591 	pbl_tbl = kcalloc(umem_pgs, sizeof(u64 *), GFP_KERNEL);
3592 	if (!pbl_tbl) {
3593 		rc = -ENOMEM;
3594 		goto free_umem;
3595 	}
3596 
3597 	page_shift = umem->page_shift;
3598 
3599 	if (!bnxt_re_page_size_ok(page_shift)) {
3600 		dev_err(rdev_to_dev(rdev), "umem page size unsupported!");
3601 		rc = -EFAULT;
3602 		goto fail;
3603 	}
3604 
3605 	if (!umem->hugetlb && length > BNXT_RE_MAX_MR_SIZE_LOW) {
3606 		dev_err(rdev_to_dev(rdev), "Requested MR Sz:%llu Max sup:%llu",
3607 			length,	(u64)BNXT_RE_MAX_MR_SIZE_LOW);
3608 		rc = -EINVAL;
3609 		goto fail;
3610 	}
3611 	if (umem->hugetlb && length > BNXT_RE_PAGE_SIZE_2M) {
3612 		page_shift = BNXT_RE_PAGE_SHIFT_2M;
3613 		dev_warn(rdev_to_dev(rdev), "umem hugetlb set page_size %x",
3614 			 1 << page_shift);
3615 	}
3616 
3617 	/* Map umem buf ptrs to the PBL */
3618 	umem_pgs = fill_umem_pbl_tbl(umem, pbl_tbl, page_shift);
3619 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, pbl_tbl,
3620 			       umem_pgs, false, 1 << page_shift);
3621 	if (rc) {
3622 		dev_err(rdev_to_dev(rdev), "Failed to register user MR");
3623 		goto fail;
3624 	}
3625 
3626 	kfree(pbl_tbl);
3627 
3628 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3629 	mr->ib_mr.rkey = mr->qplib_mr.lkey;
3630 	atomic_inc(&rdev->mr_count);
3631 
3632 	return &mr->ib_mr;
3633 fail:
3634 	kfree(pbl_tbl);
3635 free_umem:
3636 	ib_umem_release(umem);
3637 free_mrw:
3638 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3639 free_mr:
3640 	kfree(mr);
3641 	return ERR_PTR(rc);
3642 }
3643 
3644 struct ib_ucontext *bnxt_re_alloc_ucontext(struct ib_device *ibdev,
3645 					   struct ib_udata *udata)
3646 {
3647 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
3648 	struct bnxt_re_uctx_resp resp;
3649 	struct bnxt_re_ucontext *uctx;
3650 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
3651 	int rc;
3652 
3653 	dev_dbg(rdev_to_dev(rdev), "ABI version requested %d",
3654 		ibdev->uverbs_abi_ver);
3655 
3656 	if (ibdev->uverbs_abi_ver != BNXT_RE_ABI_VERSION) {
3657 		dev_dbg(rdev_to_dev(rdev), " is different from the device %d ",
3658 			BNXT_RE_ABI_VERSION);
3659 		return ERR_PTR(-EPERM);
3660 	}
3661 
3662 	uctx = kzalloc(sizeof(*uctx), GFP_KERNEL);
3663 	if (!uctx)
3664 		return ERR_PTR(-ENOMEM);
3665 
3666 	uctx->rdev = rdev;
3667 
3668 	uctx->shpg = (void *)__get_free_page(GFP_KERNEL);
3669 	if (!uctx->shpg) {
3670 		rc = -ENOMEM;
3671 		goto fail;
3672 	}
3673 	spin_lock_init(&uctx->sh_lock);
3674 
3675 	resp.dev_id = rdev->en_dev->pdev->devfn; /*Temp, Use idr_alloc instead*/
3676 	resp.max_qp = rdev->qplib_ctx.qpc_count;
3677 	resp.pg_size = PAGE_SIZE;
3678 	resp.cqe_sz = sizeof(struct cq_base);
3679 	resp.max_cqd = dev_attr->max_cq_wqes;
3680 	resp.rsvd    = 0;
3681 
3682 	rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
3683 	if (rc) {
3684 		dev_err(rdev_to_dev(rdev), "Failed to copy user context");
3685 		rc = -EFAULT;
3686 		goto cfail;
3687 	}
3688 
3689 	return &uctx->ib_uctx;
3690 cfail:
3691 	free_page((unsigned long)uctx->shpg);
3692 	uctx->shpg = NULL;
3693 fail:
3694 	kfree(uctx);
3695 	return ERR_PTR(rc);
3696 }
3697 
3698 int bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
3699 {
3700 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3701 						   struct bnxt_re_ucontext,
3702 						   ib_uctx);
3703 
3704 	struct bnxt_re_dev *rdev = uctx->rdev;
3705 	int rc = 0;
3706 
3707 	if (uctx->shpg)
3708 		free_page((unsigned long)uctx->shpg);
3709 
3710 	if (uctx->dpi.dbr) {
3711 		/* Free DPI only if this is the first PD allocated by the
3712 		 * application and mark the context dpi as NULL
3713 		 */
3714 		rc = bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
3715 					    &rdev->qplib_res.dpi_tbl,
3716 					    &uctx->dpi);
3717 		if (rc)
3718 			dev_err(rdev_to_dev(rdev), "Deallocate HW DPI failed!");
3719 			/* Don't fail, continue*/
3720 		uctx->dpi.dbr = NULL;
3721 	}
3722 
3723 	kfree(uctx);
3724 	return 0;
3725 }
3726 
3727 /* Helper function to mmap the virtual memory from user app */
3728 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
3729 {
3730 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
3731 						   struct bnxt_re_ucontext,
3732 						   ib_uctx);
3733 	struct bnxt_re_dev *rdev = uctx->rdev;
3734 	u64 pfn;
3735 
3736 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
3737 		return -EINVAL;
3738 
3739 	if (vma->vm_pgoff) {
3740 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
3741 		if (io_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
3742 				       PAGE_SIZE, vma->vm_page_prot)) {
3743 			dev_err(rdev_to_dev(rdev), "Failed to map DPI");
3744 			return -EAGAIN;
3745 		}
3746 	} else {
3747 		pfn = virt_to_phys(uctx->shpg) >> PAGE_SHIFT;
3748 		if (remap_pfn_range(vma, vma->vm_start,
3749 				    pfn, PAGE_SIZE, vma->vm_page_prot)) {
3750 			dev_err(rdev_to_dev(rdev),
3751 				"Failed to map shared page");
3752 			return -EAGAIN;
3753 		}
3754 	}
3755 
3756 	return 0;
3757 }
3758