xref: /openbmc/linux/drivers/infiniband/hw/bnxt_re/ib_verbs.c (revision aad29a73199b7fbccfbabea3f1ee627ad1924f52)
1 /*
2  * Broadcom NetXtreme-E RoCE driver.
3  *
4  * Copyright (c) 2016 - 2017, Broadcom. All rights reserved.  The term
5  * Broadcom refers to Broadcom Limited and/or its subsidiaries.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * BSD license below:
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  *
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in
21  *    the documentation and/or other materials provided with the
22  *    distribution.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS''
25  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
26  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
31  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
32  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
33  * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
34  * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35  *
36  * Description: IB Verbs interpreter
37  */
38 
39 #include <linux/interrupt.h>
40 #include <linux/types.h>
41 #include <linux/pci.h>
42 #include <linux/netdevice.h>
43 #include <linux/if_ether.h>
44 #include <net/addrconf.h>
45 
46 #include <rdma/ib_verbs.h>
47 #include <rdma/ib_user_verbs.h>
48 #include <rdma/ib_umem.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_mad.h>
51 #include <rdma/ib_cache.h>
52 #include <rdma/uverbs_ioctl.h>
53 
54 #include "bnxt_ulp.h"
55 
56 #include "roce_hsi.h"
57 #include "qplib_res.h"
58 #include "qplib_sp.h"
59 #include "qplib_fp.h"
60 #include "qplib_rcfw.h"
61 
62 #include "bnxt_re.h"
63 #include "ib_verbs.h"
64 
65 #include <rdma/uverbs_types.h>
66 #include <rdma/uverbs_std_types.h>
67 
68 #include <rdma/ib_user_ioctl_cmds.h>
69 
70 #define UVERBS_MODULE_NAME bnxt_re
71 #include <rdma/uverbs_named_ioctl.h>
72 
73 #include <rdma/bnxt_re-abi.h>
74 
__from_ib_access_flags(int iflags)75 static int __from_ib_access_flags(int iflags)
76 {
77 	int qflags = 0;
78 
79 	if (iflags & IB_ACCESS_LOCAL_WRITE)
80 		qflags |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
81 	if (iflags & IB_ACCESS_REMOTE_READ)
82 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_READ;
83 	if (iflags & IB_ACCESS_REMOTE_WRITE)
84 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_WRITE;
85 	if (iflags & IB_ACCESS_REMOTE_ATOMIC)
86 		qflags |= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC;
87 	if (iflags & IB_ACCESS_MW_BIND)
88 		qflags |= BNXT_QPLIB_ACCESS_MW_BIND;
89 	if (iflags & IB_ZERO_BASED)
90 		qflags |= BNXT_QPLIB_ACCESS_ZERO_BASED;
91 	if (iflags & IB_ACCESS_ON_DEMAND)
92 		qflags |= BNXT_QPLIB_ACCESS_ON_DEMAND;
93 	return qflags;
94 };
95 
__to_ib_access_flags(int qflags)96 static enum ib_access_flags __to_ib_access_flags(int qflags)
97 {
98 	enum ib_access_flags iflags = 0;
99 
100 	if (qflags & BNXT_QPLIB_ACCESS_LOCAL_WRITE)
101 		iflags |= IB_ACCESS_LOCAL_WRITE;
102 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_WRITE)
103 		iflags |= IB_ACCESS_REMOTE_WRITE;
104 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_READ)
105 		iflags |= IB_ACCESS_REMOTE_READ;
106 	if (qflags & BNXT_QPLIB_ACCESS_REMOTE_ATOMIC)
107 		iflags |= IB_ACCESS_REMOTE_ATOMIC;
108 	if (qflags & BNXT_QPLIB_ACCESS_MW_BIND)
109 		iflags |= IB_ACCESS_MW_BIND;
110 	if (qflags & BNXT_QPLIB_ACCESS_ZERO_BASED)
111 		iflags |= IB_ZERO_BASED;
112 	if (qflags & BNXT_QPLIB_ACCESS_ON_DEMAND)
113 		iflags |= IB_ACCESS_ON_DEMAND;
114 	return iflags;
115 };
116 
bnxt_re_build_sgl(struct ib_sge * ib_sg_list,struct bnxt_qplib_sge * sg_list,int num)117 static int bnxt_re_build_sgl(struct ib_sge *ib_sg_list,
118 			     struct bnxt_qplib_sge *sg_list, int num)
119 {
120 	int i, total = 0;
121 
122 	for (i = 0; i < num; i++) {
123 		sg_list[i].addr = ib_sg_list[i].addr;
124 		sg_list[i].lkey = ib_sg_list[i].lkey;
125 		sg_list[i].size = ib_sg_list[i].length;
126 		total += sg_list[i].size;
127 	}
128 	return total;
129 }
130 
131 /* Device */
bnxt_re_query_device(struct ib_device * ibdev,struct ib_device_attr * ib_attr,struct ib_udata * udata)132 int bnxt_re_query_device(struct ib_device *ibdev,
133 			 struct ib_device_attr *ib_attr,
134 			 struct ib_udata *udata)
135 {
136 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
137 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
138 
139 	memset(ib_attr, 0, sizeof(*ib_attr));
140 	memcpy(&ib_attr->fw_ver, dev_attr->fw_ver,
141 	       min(sizeof(dev_attr->fw_ver),
142 		   sizeof(ib_attr->fw_ver)));
143 	addrconf_addr_eui48((u8 *)&ib_attr->sys_image_guid,
144 			    rdev->netdev->dev_addr);
145 	ib_attr->max_mr_size = BNXT_RE_MAX_MR_SIZE;
146 	ib_attr->page_size_cap = BNXT_RE_PAGE_SIZE_SUPPORTED;
147 
148 	ib_attr->vendor_id = rdev->en_dev->pdev->vendor;
149 	ib_attr->vendor_part_id = rdev->en_dev->pdev->device;
150 	ib_attr->hw_ver = rdev->en_dev->pdev->revision;
151 	ib_attr->max_qp = dev_attr->max_qp;
152 	ib_attr->max_qp_wr = dev_attr->max_qp_wqes;
153 	ib_attr->device_cap_flags =
154 				    IB_DEVICE_CURR_QP_STATE_MOD
155 				    | IB_DEVICE_RC_RNR_NAK_GEN
156 				    | IB_DEVICE_SHUTDOWN_PORT
157 				    | IB_DEVICE_SYS_IMAGE_GUID
158 				    | IB_DEVICE_RESIZE_MAX_WR
159 				    | IB_DEVICE_PORT_ACTIVE_EVENT
160 				    | IB_DEVICE_N_NOTIFY_CQ
161 				    | IB_DEVICE_MEM_WINDOW
162 				    | IB_DEVICE_MEM_WINDOW_TYPE_2B
163 				    | IB_DEVICE_MEM_MGT_EXTENSIONS;
164 	ib_attr->kernel_cap_flags = IBK_LOCAL_DMA_LKEY;
165 	ib_attr->max_send_sge = dev_attr->max_qp_sges;
166 	ib_attr->max_recv_sge = dev_attr->max_qp_sges;
167 	ib_attr->max_sge_rd = dev_attr->max_qp_sges;
168 	ib_attr->max_cq = dev_attr->max_cq;
169 	ib_attr->max_cqe = dev_attr->max_cq_wqes;
170 	ib_attr->max_mr = dev_attr->max_mr;
171 	ib_attr->max_pd = dev_attr->max_pd;
172 	ib_attr->max_qp_rd_atom = dev_attr->max_qp_rd_atom;
173 	ib_attr->max_qp_init_rd_atom = dev_attr->max_qp_init_rd_atom;
174 	ib_attr->atomic_cap = IB_ATOMIC_NONE;
175 	ib_attr->masked_atomic_cap = IB_ATOMIC_NONE;
176 	if (dev_attr->is_atomic) {
177 		ib_attr->atomic_cap = IB_ATOMIC_GLOB;
178 		ib_attr->masked_atomic_cap = IB_ATOMIC_GLOB;
179 	}
180 
181 	ib_attr->max_ee_rd_atom = 0;
182 	ib_attr->max_res_rd_atom = 0;
183 	ib_attr->max_ee_init_rd_atom = 0;
184 	ib_attr->max_ee = 0;
185 	ib_attr->max_rdd = 0;
186 	ib_attr->max_mw = dev_attr->max_mw;
187 	ib_attr->max_raw_ipv6_qp = 0;
188 	ib_attr->max_raw_ethy_qp = dev_attr->max_raw_ethy_qp;
189 	ib_attr->max_mcast_grp = 0;
190 	ib_attr->max_mcast_qp_attach = 0;
191 	ib_attr->max_total_mcast_qp_attach = 0;
192 	ib_attr->max_ah = dev_attr->max_ah;
193 
194 	ib_attr->max_srq = dev_attr->max_srq;
195 	ib_attr->max_srq_wr = dev_attr->max_srq_wqes;
196 	ib_attr->max_srq_sge = dev_attr->max_srq_sges;
197 
198 	ib_attr->max_fast_reg_page_list_len = MAX_PBL_LVL_1_PGS;
199 
200 	ib_attr->max_pkeys = 1;
201 	ib_attr->local_ca_ack_delay = BNXT_RE_DEFAULT_ACK_DELAY;
202 	return 0;
203 }
204 
205 /* Port */
bnxt_re_query_port(struct ib_device * ibdev,u32 port_num,struct ib_port_attr * port_attr)206 int bnxt_re_query_port(struct ib_device *ibdev, u32 port_num,
207 		       struct ib_port_attr *port_attr)
208 {
209 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
210 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
211 	int rc;
212 
213 	memset(port_attr, 0, sizeof(*port_attr));
214 
215 	if (netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev)) {
216 		port_attr->state = IB_PORT_ACTIVE;
217 		port_attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
218 	} else {
219 		port_attr->state = IB_PORT_DOWN;
220 		port_attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
221 	}
222 	port_attr->max_mtu = IB_MTU_4096;
223 	port_attr->active_mtu = iboe_get_mtu(rdev->netdev->mtu);
224 	port_attr->gid_tbl_len = dev_attr->max_sgid;
225 	port_attr->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_REINIT_SUP |
226 				    IB_PORT_DEVICE_MGMT_SUP |
227 				    IB_PORT_VENDOR_CLASS_SUP;
228 	port_attr->ip_gids = true;
229 
230 	port_attr->max_msg_sz = (u32)BNXT_RE_MAX_MR_SIZE_LOW;
231 	port_attr->bad_pkey_cntr = 0;
232 	port_attr->qkey_viol_cntr = 0;
233 	port_attr->pkey_tbl_len = dev_attr->max_pkey;
234 	port_attr->lid = 0;
235 	port_attr->sm_lid = 0;
236 	port_attr->lmc = 0;
237 	port_attr->max_vl_num = 4;
238 	port_attr->sm_sl = 0;
239 	port_attr->subnet_timeout = 0;
240 	port_attr->init_type_reply = 0;
241 	rc = ib_get_eth_speed(&rdev->ibdev, port_num, &port_attr->active_speed,
242 			      &port_attr->active_width);
243 
244 	return rc;
245 }
246 
bnxt_re_get_port_immutable(struct ib_device * ibdev,u32 port_num,struct ib_port_immutable * immutable)247 int bnxt_re_get_port_immutable(struct ib_device *ibdev, u32 port_num,
248 			       struct ib_port_immutable *immutable)
249 {
250 	struct ib_port_attr port_attr;
251 
252 	if (bnxt_re_query_port(ibdev, port_num, &port_attr))
253 		return -EINVAL;
254 
255 	immutable->pkey_tbl_len = port_attr.pkey_tbl_len;
256 	immutable->gid_tbl_len = port_attr.gid_tbl_len;
257 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
258 	immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP;
259 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
260 	return 0;
261 }
262 
bnxt_re_query_fw_str(struct ib_device * ibdev,char * str)263 void bnxt_re_query_fw_str(struct ib_device *ibdev, char *str)
264 {
265 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
266 
267 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d.%d",
268 		 rdev->dev_attr.fw_ver[0], rdev->dev_attr.fw_ver[1],
269 		 rdev->dev_attr.fw_ver[2], rdev->dev_attr.fw_ver[3]);
270 }
271 
bnxt_re_query_pkey(struct ib_device * ibdev,u32 port_num,u16 index,u16 * pkey)272 int bnxt_re_query_pkey(struct ib_device *ibdev, u32 port_num,
273 		       u16 index, u16 *pkey)
274 {
275 	if (index > 0)
276 		return -EINVAL;
277 
278 	*pkey = IB_DEFAULT_PKEY_FULL;
279 
280 	return 0;
281 }
282 
bnxt_re_query_gid(struct ib_device * ibdev,u32 port_num,int index,union ib_gid * gid)283 int bnxt_re_query_gid(struct ib_device *ibdev, u32 port_num,
284 		      int index, union ib_gid *gid)
285 {
286 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
287 	int rc;
288 
289 	/* Ignore port_num */
290 	memset(gid, 0, sizeof(*gid));
291 	rc = bnxt_qplib_get_sgid(&rdev->qplib_res,
292 				 &rdev->qplib_res.sgid_tbl, index,
293 				 (struct bnxt_qplib_gid *)gid);
294 	return rc;
295 }
296 
bnxt_re_del_gid(const struct ib_gid_attr * attr,void ** context)297 int bnxt_re_del_gid(const struct ib_gid_attr *attr, void **context)
298 {
299 	int rc = 0;
300 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
301 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
302 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
303 	struct bnxt_qplib_gid *gid_to_del;
304 	u16 vlan_id = 0xFFFF;
305 
306 	/* Delete the entry from the hardware */
307 	ctx = *context;
308 	if (!ctx)
309 		return -EINVAL;
310 
311 	if (sgid_tbl && sgid_tbl->active) {
312 		if (ctx->idx >= sgid_tbl->max)
313 			return -EINVAL;
314 		gid_to_del = &sgid_tbl->tbl[ctx->idx].gid;
315 		vlan_id = sgid_tbl->tbl[ctx->idx].vlan_id;
316 		/* DEL_GID is called in WQ context(netdevice_event_work_handler)
317 		 * or via the ib_unregister_device path. In the former case QP1
318 		 * may not be destroyed yet, in which case just return as FW
319 		 * needs that entry to be present and will fail it's deletion.
320 		 * We could get invoked again after QP1 is destroyed OR get an
321 		 * ADD_GID call with a different GID value for the same index
322 		 * where we issue MODIFY_GID cmd to update the GID entry -- TBD
323 		 */
324 		if (ctx->idx == 0 &&
325 		    rdma_link_local_addr((struct in6_addr *)gid_to_del) &&
326 		    ctx->refcnt == 1 && rdev->gsi_ctx.gsi_sqp) {
327 			ibdev_dbg(&rdev->ibdev,
328 				  "Trying to delete GID0 while QP1 is alive\n");
329 			return -EFAULT;
330 		}
331 		ctx->refcnt--;
332 		if (!ctx->refcnt) {
333 			rc = bnxt_qplib_del_sgid(sgid_tbl, gid_to_del,
334 						 vlan_id,  true);
335 			if (rc) {
336 				ibdev_err(&rdev->ibdev,
337 					  "Failed to remove GID: %#x", rc);
338 			} else {
339 				ctx_tbl = sgid_tbl->ctx;
340 				ctx_tbl[ctx->idx] = NULL;
341 				kfree(ctx);
342 			}
343 		}
344 	} else {
345 		return -EINVAL;
346 	}
347 	return rc;
348 }
349 
bnxt_re_add_gid(const struct ib_gid_attr * attr,void ** context)350 int bnxt_re_add_gid(const struct ib_gid_attr *attr, void **context)
351 {
352 	int rc;
353 	u32 tbl_idx = 0;
354 	u16 vlan_id = 0xFFFF;
355 	struct bnxt_re_gid_ctx *ctx, **ctx_tbl;
356 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(attr->device, ibdev);
357 	struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
358 
359 	rc = rdma_read_gid_l2_fields(attr, &vlan_id, NULL);
360 	if (rc)
361 		return rc;
362 
363 	rc = bnxt_qplib_add_sgid(sgid_tbl, (struct bnxt_qplib_gid *)&attr->gid,
364 				 rdev->qplib_res.netdev->dev_addr,
365 				 vlan_id, true, &tbl_idx);
366 	if (rc == -EALREADY) {
367 		ctx_tbl = sgid_tbl->ctx;
368 		ctx_tbl[tbl_idx]->refcnt++;
369 		*context = ctx_tbl[tbl_idx];
370 		return 0;
371 	}
372 
373 	if (rc < 0) {
374 		ibdev_err(&rdev->ibdev, "Failed to add GID: %#x", rc);
375 		return rc;
376 	}
377 
378 	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
379 	if (!ctx)
380 		return -ENOMEM;
381 	ctx_tbl = sgid_tbl->ctx;
382 	ctx->idx = tbl_idx;
383 	ctx->refcnt = 1;
384 	ctx_tbl[tbl_idx] = ctx;
385 	*context = ctx;
386 
387 	return rc;
388 }
389 
bnxt_re_get_link_layer(struct ib_device * ibdev,u32 port_num)390 enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
391 					    u32 port_num)
392 {
393 	return IB_LINK_LAYER_ETHERNET;
394 }
395 
396 #define	BNXT_RE_FENCE_PBL_SIZE	DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
397 
bnxt_re_create_fence_wqe(struct bnxt_re_pd * pd)398 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd *pd)
399 {
400 	struct bnxt_re_fence_data *fence = &pd->fence;
401 	struct ib_mr *ib_mr = &fence->mr->ib_mr;
402 	struct bnxt_qplib_swqe *wqe = &fence->bind_wqe;
403 	struct bnxt_re_dev *rdev = pd->rdev;
404 
405 	if (bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx))
406 		return;
407 
408 	memset(wqe, 0, sizeof(*wqe));
409 	wqe->type = BNXT_QPLIB_SWQE_TYPE_BIND_MW;
410 	wqe->wr_id = BNXT_QPLIB_FENCE_WRID;
411 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
412 	wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
413 	wqe->bind.zero_based = false;
414 	wqe->bind.parent_l_key = ib_mr->lkey;
415 	wqe->bind.va = (u64)(unsigned long)fence->va;
416 	wqe->bind.length = fence->size;
417 	wqe->bind.access_cntl = __from_ib_access_flags(IB_ACCESS_REMOTE_READ);
418 	wqe->bind.mw_type = SQ_BIND_MW_TYPE_TYPE1;
419 
420 	/* Save the initial rkey in fence structure for now;
421 	 * wqe->bind.r_key will be set at (re)bind time.
422 	 */
423 	fence->bind_rkey = ib_inc_rkey(fence->mw->rkey);
424 }
425 
bnxt_re_bind_fence_mw(struct bnxt_qplib_qp * qplib_qp)426 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp *qplib_qp)
427 {
428 	struct bnxt_re_qp *qp = container_of(qplib_qp, struct bnxt_re_qp,
429 					     qplib_qp);
430 	struct ib_pd *ib_pd = qp->ib_qp.pd;
431 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
432 	struct bnxt_re_fence_data *fence = &pd->fence;
433 	struct bnxt_qplib_swqe *fence_wqe = &fence->bind_wqe;
434 	struct bnxt_qplib_swqe wqe;
435 	int rc;
436 
437 	memcpy(&wqe, fence_wqe, sizeof(wqe));
438 	wqe.bind.r_key = fence->bind_rkey;
439 	fence->bind_rkey = ib_inc_rkey(fence->bind_rkey);
440 
441 	ibdev_dbg(&qp->rdev->ibdev,
442 		  "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
443 		wqe.bind.r_key, qp->qplib_qp.id, pd);
444 	rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
445 	if (rc) {
446 		ibdev_err(&qp->rdev->ibdev, "Failed to bind fence-WQE\n");
447 		return rc;
448 	}
449 	bnxt_qplib_post_send_db(&qp->qplib_qp);
450 
451 	return rc;
452 }
453 
bnxt_re_destroy_fence_mr(struct bnxt_re_pd * pd)454 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd *pd)
455 {
456 	struct bnxt_re_fence_data *fence = &pd->fence;
457 	struct bnxt_re_dev *rdev = pd->rdev;
458 	struct device *dev = &rdev->en_dev->pdev->dev;
459 	struct bnxt_re_mr *mr = fence->mr;
460 
461 	if (bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx))
462 		return;
463 
464 	if (fence->mw) {
465 		bnxt_re_dealloc_mw(fence->mw);
466 		fence->mw = NULL;
467 	}
468 	if (mr) {
469 		if (mr->ib_mr.rkey)
470 			bnxt_qplib_dereg_mrw(&rdev->qplib_res, &mr->qplib_mr,
471 					     true);
472 		if (mr->ib_mr.lkey)
473 			bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
474 		kfree(mr);
475 		fence->mr = NULL;
476 	}
477 	if (fence->dma_addr) {
478 		dma_unmap_single(dev, fence->dma_addr, BNXT_RE_FENCE_BYTES,
479 				 DMA_BIDIRECTIONAL);
480 		fence->dma_addr = 0;
481 	}
482 }
483 
bnxt_re_create_fence_mr(struct bnxt_re_pd * pd)484 static int bnxt_re_create_fence_mr(struct bnxt_re_pd *pd)
485 {
486 	int mr_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_MW_BIND;
487 	struct bnxt_re_fence_data *fence = &pd->fence;
488 	struct bnxt_re_dev *rdev = pd->rdev;
489 	struct device *dev = &rdev->en_dev->pdev->dev;
490 	struct bnxt_re_mr *mr = NULL;
491 	dma_addr_t dma_addr = 0;
492 	struct ib_mw *mw;
493 	int rc;
494 
495 	if (bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx))
496 		return 0;
497 
498 	dma_addr = dma_map_single(dev, fence->va, BNXT_RE_FENCE_BYTES,
499 				  DMA_BIDIRECTIONAL);
500 	rc = dma_mapping_error(dev, dma_addr);
501 	if (rc) {
502 		ibdev_err(&rdev->ibdev, "Failed to dma-map fence-MR-mem\n");
503 		rc = -EIO;
504 		fence->dma_addr = 0;
505 		goto fail;
506 	}
507 	fence->dma_addr = dma_addr;
508 
509 	/* Allocate a MR */
510 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
511 	if (!mr) {
512 		rc = -ENOMEM;
513 		goto fail;
514 	}
515 	fence->mr = mr;
516 	mr->rdev = rdev;
517 	mr->qplib_mr.pd = &pd->qplib_pd;
518 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
519 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
520 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
521 	if (rc) {
522 		ibdev_err(&rdev->ibdev, "Failed to alloc fence-HW-MR\n");
523 		goto fail;
524 	}
525 
526 	/* Register MR */
527 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
528 	mr->qplib_mr.va = (u64)(unsigned long)fence->va;
529 	mr->qplib_mr.total_size = BNXT_RE_FENCE_BYTES;
530 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL,
531 			       BNXT_RE_FENCE_PBL_SIZE, PAGE_SIZE);
532 	if (rc) {
533 		ibdev_err(&rdev->ibdev, "Failed to register fence-MR\n");
534 		goto fail;
535 	}
536 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
537 
538 	/* Create a fence MW only for kernel consumers */
539 	mw = bnxt_re_alloc_mw(&pd->ib_pd, IB_MW_TYPE_1, NULL);
540 	if (IS_ERR(mw)) {
541 		ibdev_err(&rdev->ibdev,
542 			  "Failed to create fence-MW for PD: %p\n", pd);
543 		rc = PTR_ERR(mw);
544 		goto fail;
545 	}
546 	fence->mw = mw;
547 
548 	bnxt_re_create_fence_wqe(pd);
549 	return 0;
550 
551 fail:
552 	bnxt_re_destroy_fence_mr(pd);
553 	return rc;
554 }
555 
556 static struct bnxt_re_user_mmap_entry*
bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext * uctx,u64 mem_offset,enum bnxt_re_mmap_flag mmap_flag,u64 * offset)557 bnxt_re_mmap_entry_insert(struct bnxt_re_ucontext *uctx, u64 mem_offset,
558 			  enum bnxt_re_mmap_flag mmap_flag, u64 *offset)
559 {
560 	struct bnxt_re_user_mmap_entry *entry;
561 	int ret;
562 
563 	entry = kzalloc(sizeof(*entry), GFP_KERNEL);
564 	if (!entry)
565 		return NULL;
566 
567 	entry->mem_offset = mem_offset;
568 	entry->mmap_flag = mmap_flag;
569 	entry->uctx = uctx;
570 
571 	switch (mmap_flag) {
572 	case BNXT_RE_MMAP_SH_PAGE:
573 		ret = rdma_user_mmap_entry_insert_exact(&uctx->ib_uctx,
574 							&entry->rdma_entry, PAGE_SIZE, 0);
575 		break;
576 	case BNXT_RE_MMAP_UC_DB:
577 	case BNXT_RE_MMAP_WC_DB:
578 	case BNXT_RE_MMAP_DBR_BAR:
579 	case BNXT_RE_MMAP_DBR_PAGE:
580 		ret = rdma_user_mmap_entry_insert(&uctx->ib_uctx,
581 						  &entry->rdma_entry, PAGE_SIZE);
582 		break;
583 	default:
584 		ret = -EINVAL;
585 		break;
586 	}
587 
588 	if (ret) {
589 		kfree(entry);
590 		return NULL;
591 	}
592 	if (offset)
593 		*offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
594 
595 	return entry;
596 }
597 
598 /* Protection Domains */
bnxt_re_dealloc_pd(struct ib_pd * ib_pd,struct ib_udata * udata)599 int bnxt_re_dealloc_pd(struct ib_pd *ib_pd, struct ib_udata *udata)
600 {
601 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
602 	struct bnxt_re_dev *rdev = pd->rdev;
603 
604 	if (udata) {
605 		rdma_user_mmap_entry_remove(pd->pd_db_mmap);
606 		pd->pd_db_mmap = NULL;
607 	}
608 
609 	bnxt_re_destroy_fence_mr(pd);
610 
611 	if (pd->qplib_pd.id) {
612 		if (!bnxt_qplib_dealloc_pd(&rdev->qplib_res,
613 					   &rdev->qplib_res.pd_tbl,
614 					   &pd->qplib_pd))
615 			atomic_dec(&rdev->stats.res.pd_count);
616 	}
617 	return 0;
618 }
619 
bnxt_re_alloc_pd(struct ib_pd * ibpd,struct ib_udata * udata)620 int bnxt_re_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
621 {
622 	struct ib_device *ibdev = ibpd->device;
623 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
624 	struct bnxt_re_ucontext *ucntx = rdma_udata_to_drv_context(
625 		udata, struct bnxt_re_ucontext, ib_uctx);
626 	struct bnxt_re_pd *pd = container_of(ibpd, struct bnxt_re_pd, ib_pd);
627 	struct bnxt_re_user_mmap_entry *entry = NULL;
628 	u32 active_pds;
629 	int rc = 0;
630 
631 	pd->rdev = rdev;
632 	if (bnxt_qplib_alloc_pd(&rdev->qplib_res, &pd->qplib_pd)) {
633 		ibdev_err(&rdev->ibdev, "Failed to allocate HW PD");
634 		rc = -ENOMEM;
635 		goto fail;
636 	}
637 
638 	if (udata) {
639 		struct bnxt_re_pd_resp resp = {};
640 
641 		if (!ucntx->dpi.dbr) {
642 			/* Allocate DPI in alloc_pd to avoid failing of
643 			 * ibv_devinfo and family of application when DPIs
644 			 * are depleted.
645 			 */
646 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res,
647 						 &ucntx->dpi, ucntx, BNXT_QPLIB_DPI_TYPE_UC)) {
648 				rc = -ENOMEM;
649 				goto dbfail;
650 			}
651 		}
652 
653 		resp.pdid = pd->qplib_pd.id;
654 		/* Still allow mapping this DBR to the new user PD. */
655 		resp.dpi = ucntx->dpi.dpi;
656 
657 		entry = bnxt_re_mmap_entry_insert(ucntx, (u64)ucntx->dpi.umdbr,
658 						  BNXT_RE_MMAP_UC_DB, &resp.dbr);
659 
660 		if (!entry) {
661 			rc = -ENOMEM;
662 			goto dbfail;
663 		}
664 
665 		pd->pd_db_mmap = &entry->rdma_entry;
666 
667 		rc = ib_copy_to_udata(udata, &resp, min(sizeof(resp), udata->outlen));
668 		if (rc) {
669 			rdma_user_mmap_entry_remove(pd->pd_db_mmap);
670 			rc = -EFAULT;
671 			goto dbfail;
672 		}
673 	}
674 
675 	if (!udata)
676 		if (bnxt_re_create_fence_mr(pd))
677 			ibdev_warn(&rdev->ibdev,
678 				   "Failed to create Fence-MR\n");
679 	active_pds = atomic_inc_return(&rdev->stats.res.pd_count);
680 	if (active_pds > rdev->stats.res.pd_watermark)
681 		rdev->stats.res.pd_watermark = active_pds;
682 
683 	return 0;
684 dbfail:
685 	bnxt_qplib_dealloc_pd(&rdev->qplib_res, &rdev->qplib_res.pd_tbl,
686 			      &pd->qplib_pd);
687 fail:
688 	return rc;
689 }
690 
691 /* Address Handles */
bnxt_re_destroy_ah(struct ib_ah * ib_ah,u32 flags)692 int bnxt_re_destroy_ah(struct ib_ah *ib_ah, u32 flags)
693 {
694 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
695 	struct bnxt_re_dev *rdev = ah->rdev;
696 	bool block = true;
697 	int rc;
698 
699 	block = !(flags & RDMA_DESTROY_AH_SLEEPABLE);
700 	rc = bnxt_qplib_destroy_ah(&rdev->qplib_res, &ah->qplib_ah, block);
701 	if (BNXT_RE_CHECK_RC(rc)) {
702 		if (rc == -ETIMEDOUT)
703 			rc = 0;
704 		else
705 			goto fail;
706 	}
707 	atomic_dec(&rdev->stats.res.ah_count);
708 fail:
709 	return rc;
710 }
711 
bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)712 static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)
713 {
714 	u8 nw_type;
715 
716 	switch (ntype) {
717 	case RDMA_NETWORK_IPV4:
718 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV4;
719 		break;
720 	case RDMA_NETWORK_IPV6:
721 		nw_type = CMDQ_CREATE_AH_TYPE_V2IPV6;
722 		break;
723 	default:
724 		nw_type = CMDQ_CREATE_AH_TYPE_V1;
725 		break;
726 	}
727 	return nw_type;
728 }
729 
bnxt_re_create_ah(struct ib_ah * ib_ah,struct rdma_ah_init_attr * init_attr,struct ib_udata * udata)730 int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr,
731 		      struct ib_udata *udata)
732 {
733 	struct ib_pd *ib_pd = ib_ah->pd;
734 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
735 	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
736 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
737 	struct bnxt_re_dev *rdev = pd->rdev;
738 	const struct ib_gid_attr *sgid_attr;
739 	struct bnxt_re_gid_ctx *ctx;
740 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
741 	u32 active_ahs;
742 	u8 nw_type;
743 	int rc;
744 
745 	if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH)) {
746 		ibdev_err(&rdev->ibdev, "Failed to alloc AH: GRH not set");
747 		return -EINVAL;
748 	}
749 
750 	ah->rdev = rdev;
751 	ah->qplib_ah.pd = &pd->qplib_pd;
752 
753 	/* Supply the configuration for the HW */
754 	memcpy(ah->qplib_ah.dgid.data, grh->dgid.raw,
755 	       sizeof(union ib_gid));
756 	sgid_attr = grh->sgid_attr;
757 	/* Get the HW context of the GID. The reference
758 	 * of GID table entry is already taken by the caller.
759 	 */
760 	ctx = rdma_read_gid_hw_context(sgid_attr);
761 	ah->qplib_ah.sgid_index = ctx->idx;
762 	ah->qplib_ah.host_sgid_index = grh->sgid_index;
763 	ah->qplib_ah.traffic_class = grh->traffic_class;
764 	ah->qplib_ah.flow_label = grh->flow_label;
765 	ah->qplib_ah.hop_limit = grh->hop_limit;
766 	ah->qplib_ah.sl = rdma_ah_get_sl(ah_attr);
767 
768 	/* Get network header type for this GID */
769 	nw_type = rdma_gid_attr_network_type(sgid_attr);
770 	ah->qplib_ah.nw_type = bnxt_re_stack_to_dev_nw_type(nw_type);
771 
772 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
773 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah,
774 				  !(init_attr->flags &
775 				    RDMA_CREATE_AH_SLEEPABLE));
776 	if (rc) {
777 		ibdev_err(&rdev->ibdev, "Failed to allocate HW AH");
778 		return rc;
779 	}
780 
781 	/* Write AVID to shared page. */
782 	if (udata) {
783 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
784 			udata, struct bnxt_re_ucontext, ib_uctx);
785 		unsigned long flag;
786 		u32 *wrptr;
787 
788 		spin_lock_irqsave(&uctx->sh_lock, flag);
789 		wrptr = (u32 *)(uctx->shpg + BNXT_RE_AVID_OFFT);
790 		*wrptr = ah->qplib_ah.id;
791 		wmb(); /* make sure cache is updated. */
792 		spin_unlock_irqrestore(&uctx->sh_lock, flag);
793 	}
794 	active_ahs = atomic_inc_return(&rdev->stats.res.ah_count);
795 	if (active_ahs > rdev->stats.res.ah_watermark)
796 		rdev->stats.res.ah_watermark = active_ahs;
797 
798 	return 0;
799 }
800 
bnxt_re_query_ah(struct ib_ah * ib_ah,struct rdma_ah_attr * ah_attr)801 int bnxt_re_query_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
802 {
803 	struct bnxt_re_ah *ah = container_of(ib_ah, struct bnxt_re_ah, ib_ah);
804 
805 	ah_attr->type = ib_ah->type;
806 	rdma_ah_set_sl(ah_attr, ah->qplib_ah.sl);
807 	memcpy(ah_attr->roce.dmac, ah->qplib_ah.dmac, ETH_ALEN);
808 	rdma_ah_set_grh(ah_attr, NULL, 0,
809 			ah->qplib_ah.host_sgid_index,
810 			0, ah->qplib_ah.traffic_class);
811 	rdma_ah_set_dgid_raw(ah_attr, ah->qplib_ah.dgid.data);
812 	rdma_ah_set_port_num(ah_attr, 1);
813 	rdma_ah_set_static_rate(ah_attr, 0);
814 	return 0;
815 }
816 
bnxt_re_lock_cqs(struct bnxt_re_qp * qp)817 unsigned long bnxt_re_lock_cqs(struct bnxt_re_qp *qp)
818 	__acquires(&qp->scq->cq_lock) __acquires(&qp->rcq->cq_lock)
819 {
820 	unsigned long flags;
821 
822 	spin_lock_irqsave(&qp->scq->cq_lock, flags);
823 	if (qp->rcq != qp->scq)
824 		spin_lock(&qp->rcq->cq_lock);
825 	else
826 		__acquire(&qp->rcq->cq_lock);
827 
828 	return flags;
829 }
830 
bnxt_re_unlock_cqs(struct bnxt_re_qp * qp,unsigned long flags)831 void bnxt_re_unlock_cqs(struct bnxt_re_qp *qp,
832 			unsigned long flags)
833 	__releases(&qp->scq->cq_lock) __releases(&qp->rcq->cq_lock)
834 {
835 	if (qp->rcq != qp->scq)
836 		spin_unlock(&qp->rcq->cq_lock);
837 	else
838 		__release(&qp->rcq->cq_lock);
839 	spin_unlock_irqrestore(&qp->scq->cq_lock, flags);
840 }
841 
bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp * qp)842 static int bnxt_re_destroy_gsi_sqp(struct bnxt_re_qp *qp)
843 {
844 	struct bnxt_re_qp *gsi_sqp;
845 	struct bnxt_re_ah *gsi_sah;
846 	struct bnxt_re_dev *rdev;
847 	int rc;
848 
849 	rdev = qp->rdev;
850 	gsi_sqp = rdev->gsi_ctx.gsi_sqp;
851 	gsi_sah = rdev->gsi_ctx.gsi_sah;
852 
853 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow AH\n");
854 	bnxt_qplib_destroy_ah(&rdev->qplib_res,
855 			      &gsi_sah->qplib_ah,
856 			      true);
857 	atomic_dec(&rdev->stats.res.ah_count);
858 	bnxt_qplib_clean_qp(&qp->qplib_qp);
859 
860 	ibdev_dbg(&rdev->ibdev, "Destroy the shadow QP\n");
861 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &gsi_sqp->qplib_qp);
862 	if (rc) {
863 		ibdev_err(&rdev->ibdev, "Destroy Shadow QP failed");
864 		goto fail;
865 	}
866 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &gsi_sqp->qplib_qp);
867 
868 	/* remove from active qp list */
869 	mutex_lock(&rdev->qp_lock);
870 	list_del(&gsi_sqp->list);
871 	mutex_unlock(&rdev->qp_lock);
872 	atomic_dec(&rdev->stats.res.qp_count);
873 
874 	kfree(rdev->gsi_ctx.sqp_tbl);
875 	kfree(gsi_sah);
876 	kfree(gsi_sqp);
877 	rdev->gsi_ctx.gsi_sqp = NULL;
878 	rdev->gsi_ctx.gsi_sah = NULL;
879 	rdev->gsi_ctx.sqp_tbl = NULL;
880 
881 	return 0;
882 fail:
883 	return rc;
884 }
885 
886 /* Queue Pairs */
bnxt_re_destroy_qp(struct ib_qp * ib_qp,struct ib_udata * udata)887 int bnxt_re_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata)
888 {
889 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
890 	struct bnxt_qplib_qp *qplib_qp = &qp->qplib_qp;
891 	struct bnxt_re_dev *rdev = qp->rdev;
892 	struct bnxt_qplib_nq *scq_nq = NULL;
893 	struct bnxt_qplib_nq *rcq_nq = NULL;
894 	unsigned int flags;
895 	int rc;
896 
897 	bnxt_qplib_flush_cqn_wq(&qp->qplib_qp);
898 
899 	rc = bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
900 	if (rc) {
901 		ibdev_err(&rdev->ibdev, "Failed to destroy HW QP");
902 		return rc;
903 	}
904 
905 	if (rdma_is_kernel_res(&qp->ib_qp.res)) {
906 		flags = bnxt_re_lock_cqs(qp);
907 		bnxt_qplib_clean_qp(&qp->qplib_qp);
908 		bnxt_re_unlock_cqs(qp, flags);
909 	}
910 
911 	bnxt_qplib_free_qp_res(&rdev->qplib_res, &qp->qplib_qp);
912 
913 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp) {
914 		rc = bnxt_re_destroy_gsi_sqp(qp);
915 		if (rc)
916 			return rc;
917 	}
918 
919 	mutex_lock(&rdev->qp_lock);
920 	list_del(&qp->list);
921 	mutex_unlock(&rdev->qp_lock);
922 	atomic_dec(&rdev->stats.res.qp_count);
923 	if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_RC)
924 		atomic_dec(&rdev->stats.res.rc_qp_count);
925 	else if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD)
926 		atomic_dec(&rdev->stats.res.ud_qp_count);
927 
928 	ib_umem_release(qp->rumem);
929 	ib_umem_release(qp->sumem);
930 
931 	/* Flush all the entries of notification queue associated with
932 	 * given qp.
933 	 */
934 	scq_nq = qplib_qp->scq->nq;
935 	rcq_nq = qplib_qp->rcq->nq;
936 	bnxt_re_synchronize_nq(scq_nq);
937 	if (scq_nq != rcq_nq)
938 		bnxt_re_synchronize_nq(rcq_nq);
939 
940 	return 0;
941 }
942 
__from_ib_qp_type(enum ib_qp_type type)943 static u8 __from_ib_qp_type(enum ib_qp_type type)
944 {
945 	switch (type) {
946 	case IB_QPT_GSI:
947 		return CMDQ_CREATE_QP1_TYPE_GSI;
948 	case IB_QPT_RC:
949 		return CMDQ_CREATE_QP_TYPE_RC;
950 	case IB_QPT_UD:
951 		return CMDQ_CREATE_QP_TYPE_UD;
952 	default:
953 		return IB_QPT_MAX;
954 	}
955 }
956 
bnxt_re_setup_rwqe_size(struct bnxt_qplib_qp * qplqp,int rsge,int max)957 static u16 bnxt_re_setup_rwqe_size(struct bnxt_qplib_qp *qplqp,
958 				   int rsge, int max)
959 {
960 	if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC)
961 		rsge = max;
962 	return bnxt_re_get_rwqe_size(rsge);
963 }
964 
bnxt_re_get_wqe_size(int ilsize,int nsge)965 static u16 bnxt_re_get_wqe_size(int ilsize, int nsge)
966 {
967 	u16 wqe_size, calc_ils;
968 
969 	wqe_size = bnxt_re_get_swqe_size(nsge);
970 	if (ilsize) {
971 		calc_ils = sizeof(struct sq_send_hdr) + ilsize;
972 		wqe_size = max_t(u16, calc_ils, wqe_size);
973 		wqe_size = ALIGN(wqe_size, sizeof(struct sq_send_hdr));
974 	}
975 	return wqe_size;
976 }
977 
bnxt_re_setup_swqe_size(struct bnxt_re_qp * qp,struct ib_qp_init_attr * init_attr)978 static int bnxt_re_setup_swqe_size(struct bnxt_re_qp *qp,
979 				   struct ib_qp_init_attr *init_attr)
980 {
981 	struct bnxt_qplib_dev_attr *dev_attr;
982 	struct bnxt_qplib_qp *qplqp;
983 	struct bnxt_re_dev *rdev;
984 	struct bnxt_qplib_q *sq;
985 	int align, ilsize;
986 
987 	rdev = qp->rdev;
988 	qplqp = &qp->qplib_qp;
989 	sq = &qplqp->sq;
990 	dev_attr = &rdev->dev_attr;
991 
992 	align = sizeof(struct sq_send_hdr);
993 	ilsize = ALIGN(init_attr->cap.max_inline_data, align);
994 
995 	/* For gen p4 and gen p5 fixed wqe compatibility mode
996 	 * wqe size is fixed to 128 bytes - ie 6 SGEs
997 	 */
998 	if (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) {
999 		sq->wqe_size = bnxt_re_get_swqe_size(BNXT_STATIC_MAX_SGE);
1000 		sq->max_sge = BNXT_STATIC_MAX_SGE;
1001 	} else {
1002 		sq->wqe_size = bnxt_re_get_wqe_size(ilsize, sq->max_sge);
1003 		if (sq->wqe_size > bnxt_re_get_swqe_size(dev_attr->max_qp_sges))
1004 			return -EINVAL;
1005 	}
1006 
1007 	if (init_attr->cap.max_inline_data) {
1008 		qplqp->max_inline_data = sq->wqe_size -
1009 			sizeof(struct sq_send_hdr);
1010 		init_attr->cap.max_inline_data = qplqp->max_inline_data;
1011 	}
1012 
1013 	return 0;
1014 }
1015 
bnxt_re_init_user_qp(struct bnxt_re_dev * rdev,struct bnxt_re_pd * pd,struct bnxt_re_qp * qp,struct ib_udata * udata)1016 static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
1017 				struct bnxt_re_qp *qp, struct ib_udata *udata)
1018 {
1019 	struct bnxt_qplib_qp *qplib_qp;
1020 	struct bnxt_re_ucontext *cntx;
1021 	struct bnxt_re_qp_req ureq;
1022 	int bytes = 0, psn_sz;
1023 	struct ib_umem *umem;
1024 	int psn_nume;
1025 
1026 	qplib_qp = &qp->qplib_qp;
1027 	cntx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext,
1028 					 ib_uctx);
1029 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1030 		return -EFAULT;
1031 
1032 	bytes = (qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size);
1033 	/* Consider mapping PSN search memory only for RC QPs. */
1034 	if (qplib_qp->type == CMDQ_CREATE_QP_TYPE_RC) {
1035 		psn_sz = bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx) ?
1036 						   sizeof(struct sq_psn_search_ext) :
1037 						   sizeof(struct sq_psn_search);
1038 		psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ?
1039 			    qplib_qp->sq.max_wqe :
1040 			    ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) /
1041 			      sizeof(struct bnxt_qplib_sge));
1042 		bytes += (psn_nume * psn_sz);
1043 	}
1044 
1045 	bytes = PAGE_ALIGN(bytes);
1046 	umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes,
1047 			   IB_ACCESS_LOCAL_WRITE);
1048 	if (IS_ERR(umem))
1049 		return PTR_ERR(umem);
1050 
1051 	qp->sumem = umem;
1052 	qplib_qp->sq.sg_info.umem = umem;
1053 	qplib_qp->sq.sg_info.pgsize = PAGE_SIZE;
1054 	qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT;
1055 	qplib_qp->qp_handle = ureq.qp_handle;
1056 
1057 	if (!qp->qplib_qp.srq) {
1058 		bytes = (qplib_qp->rq.max_wqe * qplib_qp->rq.wqe_size);
1059 		bytes = PAGE_ALIGN(bytes);
1060 		umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes,
1061 				   IB_ACCESS_LOCAL_WRITE);
1062 		if (IS_ERR(umem))
1063 			goto rqfail;
1064 		qp->rumem = umem;
1065 		qplib_qp->rq.sg_info.umem = umem;
1066 		qplib_qp->rq.sg_info.pgsize = PAGE_SIZE;
1067 		qplib_qp->rq.sg_info.pgshft = PAGE_SHIFT;
1068 	}
1069 
1070 	qplib_qp->dpi = &cntx->dpi;
1071 	return 0;
1072 rqfail:
1073 	ib_umem_release(qp->sumem);
1074 	qp->sumem = NULL;
1075 	memset(&qplib_qp->sq.sg_info, 0, sizeof(qplib_qp->sq.sg_info));
1076 
1077 	return PTR_ERR(umem);
1078 }
1079 
bnxt_re_create_shadow_qp_ah(struct bnxt_re_pd * pd,struct bnxt_qplib_res * qp1_res,struct bnxt_qplib_qp * qp1_qp)1080 static struct bnxt_re_ah *bnxt_re_create_shadow_qp_ah
1081 				(struct bnxt_re_pd *pd,
1082 				 struct bnxt_qplib_res *qp1_res,
1083 				 struct bnxt_qplib_qp *qp1_qp)
1084 {
1085 	struct bnxt_re_dev *rdev = pd->rdev;
1086 	struct bnxt_re_ah *ah;
1087 	union ib_gid sgid;
1088 	int rc;
1089 
1090 	ah = kzalloc(sizeof(*ah), GFP_KERNEL);
1091 	if (!ah)
1092 		return NULL;
1093 
1094 	ah->rdev = rdev;
1095 	ah->qplib_ah.pd = &pd->qplib_pd;
1096 
1097 	rc = bnxt_re_query_gid(&rdev->ibdev, 1, 0, &sgid);
1098 	if (rc)
1099 		goto fail;
1100 
1101 	/* supply the dgid data same as sgid */
1102 	memcpy(ah->qplib_ah.dgid.data, &sgid.raw,
1103 	       sizeof(union ib_gid));
1104 	ah->qplib_ah.sgid_index = 0;
1105 
1106 	ah->qplib_ah.traffic_class = 0;
1107 	ah->qplib_ah.flow_label = 0;
1108 	ah->qplib_ah.hop_limit = 1;
1109 	ah->qplib_ah.sl = 0;
1110 	/* Have DMAC same as SMAC */
1111 	ether_addr_copy(ah->qplib_ah.dmac, rdev->netdev->dev_addr);
1112 
1113 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah, false);
1114 	if (rc) {
1115 		ibdev_err(&rdev->ibdev,
1116 			  "Failed to allocate HW AH for Shadow QP");
1117 		goto fail;
1118 	}
1119 	atomic_inc(&rdev->stats.res.ah_count);
1120 
1121 	return ah;
1122 
1123 fail:
1124 	kfree(ah);
1125 	return NULL;
1126 }
1127 
bnxt_re_create_shadow_qp(struct bnxt_re_pd * pd,struct bnxt_qplib_res * qp1_res,struct bnxt_qplib_qp * qp1_qp)1128 static struct bnxt_re_qp *bnxt_re_create_shadow_qp
1129 				(struct bnxt_re_pd *pd,
1130 				 struct bnxt_qplib_res *qp1_res,
1131 				 struct bnxt_qplib_qp *qp1_qp)
1132 {
1133 	struct bnxt_re_dev *rdev = pd->rdev;
1134 	struct bnxt_re_qp *qp;
1135 	int rc;
1136 
1137 	qp = kzalloc(sizeof(*qp), GFP_KERNEL);
1138 	if (!qp)
1139 		return NULL;
1140 
1141 	qp->rdev = rdev;
1142 
1143 	/* Initialize the shadow QP structure from the QP1 values */
1144 	ether_addr_copy(qp->qplib_qp.smac, rdev->netdev->dev_addr);
1145 
1146 	qp->qplib_qp.pd = &pd->qplib_pd;
1147 	qp->qplib_qp.qp_handle = (u64)(unsigned long)(&qp->qplib_qp);
1148 	qp->qplib_qp.type = IB_QPT_UD;
1149 
1150 	qp->qplib_qp.max_inline_data = 0;
1151 	qp->qplib_qp.sig_type = true;
1152 
1153 	/* Shadow QP SQ depth should be same as QP1 RQ depth */
1154 	qp->qplib_qp.sq.wqe_size = bnxt_re_get_wqe_size(0, 6);
1155 	qp->qplib_qp.sq.max_wqe = qp1_qp->rq.max_wqe;
1156 	qp->qplib_qp.sq.max_sw_wqe = qp1_qp->rq.max_wqe;
1157 	qp->qplib_qp.sq.max_sge = 2;
1158 	/* Q full delta can be 1 since it is internal QP */
1159 	qp->qplib_qp.sq.q_full_delta = 1;
1160 	qp->qplib_qp.sq.sg_info.pgsize = PAGE_SIZE;
1161 	qp->qplib_qp.sq.sg_info.pgshft = PAGE_SHIFT;
1162 
1163 	qp->qplib_qp.scq = qp1_qp->scq;
1164 	qp->qplib_qp.rcq = qp1_qp->rcq;
1165 
1166 	qp->qplib_qp.rq.wqe_size = bnxt_re_get_rwqe_size(6);
1167 	qp->qplib_qp.rq.max_wqe = qp1_qp->rq.max_wqe;
1168 	qp->qplib_qp.rq.max_sw_wqe = qp1_qp->rq.max_wqe;
1169 	qp->qplib_qp.rq.max_sge = qp1_qp->rq.max_sge;
1170 	/* Q full delta can be 1 since it is internal QP */
1171 	qp->qplib_qp.rq.q_full_delta = 1;
1172 	qp->qplib_qp.rq.sg_info.pgsize = PAGE_SIZE;
1173 	qp->qplib_qp.rq.sg_info.pgshft = PAGE_SHIFT;
1174 
1175 	qp->qplib_qp.mtu = qp1_qp->mtu;
1176 
1177 	qp->qplib_qp.sq_hdr_buf_size = 0;
1178 	qp->qplib_qp.rq_hdr_buf_size = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
1179 	qp->qplib_qp.dpi = &rdev->dpi_privileged;
1180 
1181 	rc = bnxt_qplib_create_qp(qp1_res, &qp->qplib_qp);
1182 	if (rc)
1183 		goto fail;
1184 
1185 	spin_lock_init(&qp->sq_lock);
1186 	INIT_LIST_HEAD(&qp->list);
1187 	mutex_lock(&rdev->qp_lock);
1188 	list_add_tail(&qp->list, &rdev->qp_list);
1189 	atomic_inc(&rdev->stats.res.qp_count);
1190 	mutex_unlock(&rdev->qp_lock);
1191 	return qp;
1192 fail:
1193 	kfree(qp);
1194 	return NULL;
1195 }
1196 
bnxt_re_init_rq_attr(struct bnxt_re_qp * qp,struct ib_qp_init_attr * init_attr,struct bnxt_re_ucontext * uctx)1197 static int bnxt_re_init_rq_attr(struct bnxt_re_qp *qp,
1198 				struct ib_qp_init_attr *init_attr,
1199 				struct bnxt_re_ucontext *uctx)
1200 {
1201 	struct bnxt_qplib_dev_attr *dev_attr;
1202 	struct bnxt_qplib_qp *qplqp;
1203 	struct bnxt_re_dev *rdev;
1204 	struct bnxt_qplib_q *rq;
1205 	int entries;
1206 
1207 	rdev = qp->rdev;
1208 	qplqp = &qp->qplib_qp;
1209 	rq = &qplqp->rq;
1210 	dev_attr = &rdev->dev_attr;
1211 
1212 	if (init_attr->srq) {
1213 		struct bnxt_re_srq *srq;
1214 
1215 		srq = container_of(init_attr->srq, struct bnxt_re_srq, ib_srq);
1216 		qplqp->srq = &srq->qplib_srq;
1217 		rq->max_wqe = 0;
1218 	} else {
1219 		rq->max_sge = init_attr->cap.max_recv_sge;
1220 		if (rq->max_sge > dev_attr->max_qp_sges)
1221 			rq->max_sge = dev_attr->max_qp_sges;
1222 		init_attr->cap.max_recv_sge = rq->max_sge;
1223 		rq->wqe_size = bnxt_re_setup_rwqe_size(qplqp, rq->max_sge,
1224 						       dev_attr->max_qp_sges);
1225 		/* Allocate 1 more than what's provided so posting max doesn't
1226 		 * mean empty.
1227 		 */
1228 		entries = bnxt_re_init_depth(init_attr->cap.max_recv_wr + 1, uctx);
1229 		rq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + 1);
1230 		rq->max_sw_wqe = rq->max_wqe;
1231 		rq->q_full_delta = 0;
1232 		rq->sg_info.pgsize = PAGE_SIZE;
1233 		rq->sg_info.pgshft = PAGE_SHIFT;
1234 	}
1235 
1236 	return 0;
1237 }
1238 
bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp * qp)1239 static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp)
1240 {
1241 	struct bnxt_qplib_dev_attr *dev_attr;
1242 	struct bnxt_qplib_qp *qplqp;
1243 	struct bnxt_re_dev *rdev;
1244 
1245 	rdev = qp->rdev;
1246 	qplqp = &qp->qplib_qp;
1247 	dev_attr = &rdev->dev_attr;
1248 
1249 	if (!bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx)) {
1250 		qplqp->rq.max_sge = dev_attr->max_qp_sges;
1251 		if (qplqp->rq.max_sge > dev_attr->max_qp_sges)
1252 			qplqp->rq.max_sge = dev_attr->max_qp_sges;
1253 		qplqp->rq.max_sge = 6;
1254 	}
1255 }
1256 
bnxt_re_init_sq_attr(struct bnxt_re_qp * qp,struct ib_qp_init_attr * init_attr,struct bnxt_re_ucontext * uctx)1257 static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp,
1258 				struct ib_qp_init_attr *init_attr,
1259 				struct bnxt_re_ucontext *uctx)
1260 {
1261 	struct bnxt_qplib_dev_attr *dev_attr;
1262 	struct bnxt_qplib_qp *qplqp;
1263 	struct bnxt_re_dev *rdev;
1264 	struct bnxt_qplib_q *sq;
1265 	int entries;
1266 	int diff;
1267 	int rc;
1268 
1269 	rdev = qp->rdev;
1270 	qplqp = &qp->qplib_qp;
1271 	sq = &qplqp->sq;
1272 	dev_attr = &rdev->dev_attr;
1273 
1274 	sq->max_sge = init_attr->cap.max_send_sge;
1275 	if (sq->max_sge > dev_attr->max_qp_sges) {
1276 		sq->max_sge = dev_attr->max_qp_sges;
1277 		init_attr->cap.max_send_sge = sq->max_sge;
1278 	}
1279 
1280 	rc = bnxt_re_setup_swqe_size(qp, init_attr);
1281 	if (rc)
1282 		return rc;
1283 
1284 	entries = init_attr->cap.max_send_wr;
1285 	/* Allocate 128 + 1 more than what's provided */
1286 	diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ?
1287 		0 : BNXT_QPLIB_RESERVED_QP_WRS;
1288 	entries = bnxt_re_init_depth(entries + diff + 1, uctx);
1289 	sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1);
1290 	sq->max_sw_wqe = bnxt_qplib_get_depth(sq, qplqp->wqe_mode, true);
1291 	sq->q_full_delta = diff + 1;
1292 	/*
1293 	 * Reserving one slot for Phantom WQE. Application can
1294 	 * post one extra entry in this case. But allowing this to avoid
1295 	 * unexpected Queue full condition
1296 	 */
1297 	qplqp->sq.q_full_delta -= 1;
1298 	qplqp->sq.sg_info.pgsize = PAGE_SIZE;
1299 	qplqp->sq.sg_info.pgshft = PAGE_SHIFT;
1300 
1301 	return 0;
1302 }
1303 
bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp * qp,struct ib_qp_init_attr * init_attr,struct bnxt_re_ucontext * uctx)1304 static void bnxt_re_adjust_gsi_sq_attr(struct bnxt_re_qp *qp,
1305 				       struct ib_qp_init_attr *init_attr,
1306 				       struct bnxt_re_ucontext *uctx)
1307 {
1308 	struct bnxt_qplib_dev_attr *dev_attr;
1309 	struct bnxt_qplib_qp *qplqp;
1310 	struct bnxt_re_dev *rdev;
1311 	int entries;
1312 
1313 	rdev = qp->rdev;
1314 	qplqp = &qp->qplib_qp;
1315 	dev_attr = &rdev->dev_attr;
1316 
1317 	if (!bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx)) {
1318 		entries = bnxt_re_init_depth(init_attr->cap.max_send_wr + 1, uctx);
1319 		qplqp->sq.max_wqe = min_t(u32, entries,
1320 					  dev_attr->max_qp_wqes + 1);
1321 		qplqp->sq.q_full_delta = qplqp->sq.max_wqe -
1322 			init_attr->cap.max_send_wr;
1323 		qplqp->sq.max_sge++; /* Need one extra sge to put UD header */
1324 		if (qplqp->sq.max_sge > dev_attr->max_qp_sges)
1325 			qplqp->sq.max_sge = dev_attr->max_qp_sges;
1326 	}
1327 }
1328 
bnxt_re_init_qp_type(struct bnxt_re_dev * rdev,struct ib_qp_init_attr * init_attr)1329 static int bnxt_re_init_qp_type(struct bnxt_re_dev *rdev,
1330 				struct ib_qp_init_attr *init_attr)
1331 {
1332 	struct bnxt_qplib_chip_ctx *chip_ctx;
1333 	int qptype;
1334 
1335 	chip_ctx = rdev->chip_ctx;
1336 
1337 	qptype = __from_ib_qp_type(init_attr->qp_type);
1338 	if (qptype == IB_QPT_MAX) {
1339 		ibdev_err(&rdev->ibdev, "QP type 0x%x not supported", qptype);
1340 		qptype = -EOPNOTSUPP;
1341 		goto out;
1342 	}
1343 
1344 	if (bnxt_qplib_is_chip_gen_p5_p7(chip_ctx) &&
1345 	    init_attr->qp_type == IB_QPT_GSI)
1346 		qptype = CMDQ_CREATE_QP_TYPE_GSI;
1347 out:
1348 	return qptype;
1349 }
1350 
bnxt_re_init_qp_attr(struct bnxt_re_qp * qp,struct bnxt_re_pd * pd,struct ib_qp_init_attr * init_attr,struct ib_udata * udata)1351 static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1352 				struct ib_qp_init_attr *init_attr,
1353 				struct ib_udata *udata)
1354 {
1355 	struct bnxt_qplib_dev_attr *dev_attr;
1356 	struct bnxt_re_ucontext *uctx;
1357 	struct bnxt_qplib_qp *qplqp;
1358 	struct bnxt_re_dev *rdev;
1359 	struct bnxt_re_cq *cq;
1360 	int rc = 0, qptype;
1361 
1362 	rdev = qp->rdev;
1363 	qplqp = &qp->qplib_qp;
1364 	dev_attr = &rdev->dev_attr;
1365 
1366 	uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
1367 	/* Setup misc params */
1368 	ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr);
1369 	qplqp->pd = &pd->qplib_pd;
1370 	qplqp->qp_handle = (u64)qplqp;
1371 	qplqp->max_inline_data = init_attr->cap.max_inline_data;
1372 	qplqp->sig_type = init_attr->sq_sig_type == IB_SIGNAL_ALL_WR;
1373 	qptype = bnxt_re_init_qp_type(rdev, init_attr);
1374 	if (qptype < 0) {
1375 		rc = qptype;
1376 		goto out;
1377 	}
1378 	qplqp->type = (u8)qptype;
1379 	qplqp->wqe_mode = rdev->chip_ctx->modes.wqe_mode;
1380 
1381 	if (init_attr->qp_type == IB_QPT_RC) {
1382 		qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom;
1383 		qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom;
1384 	}
1385 	qplqp->mtu = ib_mtu_enum_to_int(iboe_get_mtu(rdev->netdev->mtu));
1386 	qplqp->dpi = &rdev->dpi_privileged; /* Doorbell page */
1387 	if (init_attr->create_flags) {
1388 		ibdev_dbg(&rdev->ibdev,
1389 			  "QP create flags 0x%x not supported",
1390 			  init_attr->create_flags);
1391 		return -EOPNOTSUPP;
1392 	}
1393 
1394 	/* Setup CQs */
1395 	if (init_attr->send_cq) {
1396 		cq = container_of(init_attr->send_cq, struct bnxt_re_cq, ib_cq);
1397 		qplqp->scq = &cq->qplib_cq;
1398 		qp->scq = cq;
1399 	}
1400 
1401 	if (init_attr->recv_cq) {
1402 		cq = container_of(init_attr->recv_cq, struct bnxt_re_cq, ib_cq);
1403 		qplqp->rcq = &cq->qplib_cq;
1404 		qp->rcq = cq;
1405 	}
1406 
1407 	/* Setup RQ/SRQ */
1408 	rc = bnxt_re_init_rq_attr(qp, init_attr, uctx);
1409 	if (rc)
1410 		goto out;
1411 	if (init_attr->qp_type == IB_QPT_GSI)
1412 		bnxt_re_adjust_gsi_rq_attr(qp);
1413 
1414 	/* Setup SQ */
1415 	rc = bnxt_re_init_sq_attr(qp, init_attr, uctx);
1416 	if (rc)
1417 		goto out;
1418 	if (init_attr->qp_type == IB_QPT_GSI)
1419 		bnxt_re_adjust_gsi_sq_attr(qp, init_attr, uctx);
1420 
1421 	if (udata) /* This will update DPI and qp_handle */
1422 		rc = bnxt_re_init_user_qp(rdev, pd, qp, udata);
1423 out:
1424 	return rc;
1425 }
1426 
bnxt_re_create_shadow_gsi(struct bnxt_re_qp * qp,struct bnxt_re_pd * pd)1427 static int bnxt_re_create_shadow_gsi(struct bnxt_re_qp *qp,
1428 				     struct bnxt_re_pd *pd)
1429 {
1430 	struct bnxt_re_sqp_entries *sqp_tbl;
1431 	struct bnxt_re_dev *rdev;
1432 	struct bnxt_re_qp *sqp;
1433 	struct bnxt_re_ah *sah;
1434 	int rc = 0;
1435 
1436 	rdev = qp->rdev;
1437 	/* Create a shadow QP to handle the QP1 traffic */
1438 	sqp_tbl = kcalloc(BNXT_RE_MAX_GSI_SQP_ENTRIES, sizeof(*sqp_tbl),
1439 			  GFP_KERNEL);
1440 	if (!sqp_tbl)
1441 		return -ENOMEM;
1442 	rdev->gsi_ctx.sqp_tbl = sqp_tbl;
1443 
1444 	sqp = bnxt_re_create_shadow_qp(pd, &rdev->qplib_res, &qp->qplib_qp);
1445 	if (!sqp) {
1446 		rc = -ENODEV;
1447 		ibdev_err(&rdev->ibdev, "Failed to create Shadow QP for QP1");
1448 		goto out;
1449 	}
1450 	rdev->gsi_ctx.gsi_sqp = sqp;
1451 
1452 	sqp->rcq = qp->rcq;
1453 	sqp->scq = qp->scq;
1454 	sah = bnxt_re_create_shadow_qp_ah(pd, &rdev->qplib_res,
1455 					  &qp->qplib_qp);
1456 	if (!sah) {
1457 		bnxt_qplib_destroy_qp(&rdev->qplib_res,
1458 				      &sqp->qplib_qp);
1459 		rc = -ENODEV;
1460 		ibdev_err(&rdev->ibdev,
1461 			  "Failed to create AH entry for ShadowQP");
1462 		goto out;
1463 	}
1464 	rdev->gsi_ctx.gsi_sah = sah;
1465 
1466 	return 0;
1467 out:
1468 	kfree(sqp_tbl);
1469 	return rc;
1470 }
1471 
bnxt_re_create_gsi_qp(struct bnxt_re_qp * qp,struct bnxt_re_pd * pd,struct ib_qp_init_attr * init_attr)1472 static int bnxt_re_create_gsi_qp(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd,
1473 				 struct ib_qp_init_attr *init_attr)
1474 {
1475 	struct bnxt_re_dev *rdev;
1476 	struct bnxt_qplib_qp *qplqp;
1477 	int rc;
1478 
1479 	rdev = qp->rdev;
1480 	qplqp = &qp->qplib_qp;
1481 
1482 	qplqp->rq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
1483 	qplqp->sq_hdr_buf_size = BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2;
1484 
1485 	rc = bnxt_qplib_create_qp1(&rdev->qplib_res, qplqp);
1486 	if (rc) {
1487 		ibdev_err(&rdev->ibdev, "create HW QP1 failed!");
1488 		goto out;
1489 	}
1490 
1491 	rc = bnxt_re_create_shadow_gsi(qp, pd);
1492 out:
1493 	return rc;
1494 }
1495 
bnxt_re_test_qp_limits(struct bnxt_re_dev * rdev,struct ib_qp_init_attr * init_attr,struct bnxt_qplib_dev_attr * dev_attr)1496 static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev,
1497 				   struct ib_qp_init_attr *init_attr,
1498 				   struct bnxt_qplib_dev_attr *dev_attr)
1499 {
1500 	bool rc = true;
1501 
1502 	if (init_attr->cap.max_send_wr > dev_attr->max_qp_wqes ||
1503 	    init_attr->cap.max_recv_wr > dev_attr->max_qp_wqes ||
1504 	    init_attr->cap.max_send_sge > dev_attr->max_qp_sges ||
1505 	    init_attr->cap.max_recv_sge > dev_attr->max_qp_sges ||
1506 	    init_attr->cap.max_inline_data > dev_attr->max_inline_data) {
1507 		ibdev_err(&rdev->ibdev,
1508 			  "Create QP failed - max exceeded! 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x 0x%x/0x%x",
1509 			  init_attr->cap.max_send_wr, dev_attr->max_qp_wqes,
1510 			  init_attr->cap.max_recv_wr, dev_attr->max_qp_wqes,
1511 			  init_attr->cap.max_send_sge, dev_attr->max_qp_sges,
1512 			  init_attr->cap.max_recv_sge, dev_attr->max_qp_sges,
1513 			  init_attr->cap.max_inline_data,
1514 			  dev_attr->max_inline_data);
1515 		rc = false;
1516 	}
1517 	return rc;
1518 }
1519 
bnxt_re_create_qp(struct ib_qp * ib_qp,struct ib_qp_init_attr * qp_init_attr,struct ib_udata * udata)1520 int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr,
1521 		      struct ib_udata *udata)
1522 {
1523 	struct ib_pd *ib_pd = ib_qp->pd;
1524 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1525 	struct bnxt_re_dev *rdev = pd->rdev;
1526 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1527 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1528 	u32 active_qps;
1529 	int rc;
1530 
1531 	rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr);
1532 	if (!rc) {
1533 		rc = -EINVAL;
1534 		goto fail;
1535 	}
1536 
1537 	qp->rdev = rdev;
1538 	rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, udata);
1539 	if (rc)
1540 		goto fail;
1541 
1542 	if (qp_init_attr->qp_type == IB_QPT_GSI &&
1543 	    !(bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx))) {
1544 		rc = bnxt_re_create_gsi_qp(qp, pd, qp_init_attr);
1545 		if (rc == -ENODEV)
1546 			goto qp_destroy;
1547 		if (rc)
1548 			goto fail;
1549 	} else {
1550 		rc = bnxt_qplib_create_qp(&rdev->qplib_res, &qp->qplib_qp);
1551 		if (rc) {
1552 			ibdev_err(&rdev->ibdev, "Failed to create HW QP");
1553 			goto free_umem;
1554 		}
1555 		if (udata) {
1556 			struct bnxt_re_qp_resp resp;
1557 
1558 			resp.qpid = qp->qplib_qp.id;
1559 			resp.rsvd = 0;
1560 			rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1561 			if (rc) {
1562 				ibdev_err(&rdev->ibdev, "Failed to copy QP udata");
1563 				goto qp_destroy;
1564 			}
1565 		}
1566 	}
1567 
1568 	qp->ib_qp.qp_num = qp->qplib_qp.id;
1569 	if (qp_init_attr->qp_type == IB_QPT_GSI)
1570 		rdev->gsi_ctx.gsi_qp = qp;
1571 	spin_lock_init(&qp->sq_lock);
1572 	spin_lock_init(&qp->rq_lock);
1573 	INIT_LIST_HEAD(&qp->list);
1574 	mutex_lock(&rdev->qp_lock);
1575 	list_add_tail(&qp->list, &rdev->qp_list);
1576 	mutex_unlock(&rdev->qp_lock);
1577 	active_qps = atomic_inc_return(&rdev->stats.res.qp_count);
1578 	if (active_qps > rdev->stats.res.qp_watermark)
1579 		rdev->stats.res.qp_watermark = active_qps;
1580 	if (qp_init_attr->qp_type == IB_QPT_RC) {
1581 		active_qps = atomic_inc_return(&rdev->stats.res.rc_qp_count);
1582 		if (active_qps > rdev->stats.res.rc_qp_watermark)
1583 			rdev->stats.res.rc_qp_watermark = active_qps;
1584 	} else if (qp_init_attr->qp_type == IB_QPT_UD) {
1585 		active_qps = atomic_inc_return(&rdev->stats.res.ud_qp_count);
1586 		if (active_qps > rdev->stats.res.ud_qp_watermark)
1587 			rdev->stats.res.ud_qp_watermark = active_qps;
1588 	}
1589 
1590 	return 0;
1591 qp_destroy:
1592 	bnxt_qplib_destroy_qp(&rdev->qplib_res, &qp->qplib_qp);
1593 free_umem:
1594 	ib_umem_release(qp->rumem);
1595 	ib_umem_release(qp->sumem);
1596 fail:
1597 	return rc;
1598 }
1599 
__from_ib_qp_state(enum ib_qp_state state)1600 static u8 __from_ib_qp_state(enum ib_qp_state state)
1601 {
1602 	switch (state) {
1603 	case IB_QPS_RESET:
1604 		return CMDQ_MODIFY_QP_NEW_STATE_RESET;
1605 	case IB_QPS_INIT:
1606 		return CMDQ_MODIFY_QP_NEW_STATE_INIT;
1607 	case IB_QPS_RTR:
1608 		return CMDQ_MODIFY_QP_NEW_STATE_RTR;
1609 	case IB_QPS_RTS:
1610 		return CMDQ_MODIFY_QP_NEW_STATE_RTS;
1611 	case IB_QPS_SQD:
1612 		return CMDQ_MODIFY_QP_NEW_STATE_SQD;
1613 	case IB_QPS_SQE:
1614 		return CMDQ_MODIFY_QP_NEW_STATE_SQE;
1615 	case IB_QPS_ERR:
1616 	default:
1617 		return CMDQ_MODIFY_QP_NEW_STATE_ERR;
1618 	}
1619 }
1620 
__to_ib_qp_state(u8 state)1621 static enum ib_qp_state __to_ib_qp_state(u8 state)
1622 {
1623 	switch (state) {
1624 	case CMDQ_MODIFY_QP_NEW_STATE_RESET:
1625 		return IB_QPS_RESET;
1626 	case CMDQ_MODIFY_QP_NEW_STATE_INIT:
1627 		return IB_QPS_INIT;
1628 	case CMDQ_MODIFY_QP_NEW_STATE_RTR:
1629 		return IB_QPS_RTR;
1630 	case CMDQ_MODIFY_QP_NEW_STATE_RTS:
1631 		return IB_QPS_RTS;
1632 	case CMDQ_MODIFY_QP_NEW_STATE_SQD:
1633 		return IB_QPS_SQD;
1634 	case CMDQ_MODIFY_QP_NEW_STATE_SQE:
1635 		return IB_QPS_SQE;
1636 	case CMDQ_MODIFY_QP_NEW_STATE_ERR:
1637 	default:
1638 		return IB_QPS_ERR;
1639 	}
1640 }
1641 
__from_ib_mtu(enum ib_mtu mtu)1642 static u32 __from_ib_mtu(enum ib_mtu mtu)
1643 {
1644 	switch (mtu) {
1645 	case IB_MTU_256:
1646 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_256;
1647 	case IB_MTU_512:
1648 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_512;
1649 	case IB_MTU_1024:
1650 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024;
1651 	case IB_MTU_2048:
1652 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1653 	case IB_MTU_4096:
1654 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096;
1655 	default:
1656 		return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048;
1657 	}
1658 }
1659 
__to_ib_mtu(u32 mtu)1660 static enum ib_mtu __to_ib_mtu(u32 mtu)
1661 {
1662 	switch (mtu & CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK) {
1663 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_256:
1664 		return IB_MTU_256;
1665 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_512:
1666 		return IB_MTU_512;
1667 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024:
1668 		return IB_MTU_1024;
1669 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048:
1670 		return IB_MTU_2048;
1671 	case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096:
1672 		return IB_MTU_4096;
1673 	default:
1674 		return IB_MTU_2048;
1675 	}
1676 }
1677 
1678 /* Shared Receive Queues */
bnxt_re_destroy_srq(struct ib_srq * ib_srq,struct ib_udata * udata)1679 int bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata)
1680 {
1681 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1682 					       ib_srq);
1683 	struct bnxt_re_dev *rdev = srq->rdev;
1684 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1685 	struct bnxt_qplib_nq *nq = NULL;
1686 
1687 	if (qplib_srq->cq)
1688 		nq = qplib_srq->cq->nq;
1689 	bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq);
1690 	ib_umem_release(srq->umem);
1691 	atomic_dec(&rdev->stats.res.srq_count);
1692 	if (nq)
1693 		nq->budget--;
1694 	return 0;
1695 }
1696 
bnxt_re_init_user_srq(struct bnxt_re_dev * rdev,struct bnxt_re_pd * pd,struct bnxt_re_srq * srq,struct ib_udata * udata)1697 static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
1698 				 struct bnxt_re_pd *pd,
1699 				 struct bnxt_re_srq *srq,
1700 				 struct ib_udata *udata)
1701 {
1702 	struct bnxt_re_srq_req ureq;
1703 	struct bnxt_qplib_srq *qplib_srq = &srq->qplib_srq;
1704 	struct ib_umem *umem;
1705 	int bytes = 0;
1706 	struct bnxt_re_ucontext *cntx = rdma_udata_to_drv_context(
1707 		udata, struct bnxt_re_ucontext, ib_uctx);
1708 
1709 	if (ib_copy_from_udata(&ureq, udata, sizeof(ureq)))
1710 		return -EFAULT;
1711 
1712 	bytes = (qplib_srq->max_wqe * qplib_srq->wqe_size);
1713 	bytes = PAGE_ALIGN(bytes);
1714 	umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes,
1715 			   IB_ACCESS_LOCAL_WRITE);
1716 	if (IS_ERR(umem))
1717 		return PTR_ERR(umem);
1718 
1719 	srq->umem = umem;
1720 	qplib_srq->sg_info.umem = umem;
1721 	qplib_srq->sg_info.pgsize = PAGE_SIZE;
1722 	qplib_srq->sg_info.pgshft = PAGE_SHIFT;
1723 	qplib_srq->srq_handle = ureq.srq_handle;
1724 	qplib_srq->dpi = &cntx->dpi;
1725 
1726 	return 0;
1727 }
1728 
bnxt_re_create_srq(struct ib_srq * ib_srq,struct ib_srq_init_attr * srq_init_attr,struct ib_udata * udata)1729 int bnxt_re_create_srq(struct ib_srq *ib_srq,
1730 		       struct ib_srq_init_attr *srq_init_attr,
1731 		       struct ib_udata *udata)
1732 {
1733 	struct bnxt_qplib_dev_attr *dev_attr;
1734 	struct bnxt_qplib_nq *nq = NULL;
1735 	struct bnxt_re_ucontext *uctx;
1736 	struct bnxt_re_dev *rdev;
1737 	struct bnxt_re_srq *srq;
1738 	struct bnxt_re_pd *pd;
1739 	struct ib_pd *ib_pd;
1740 	u32 active_srqs;
1741 	int rc, entries;
1742 
1743 	ib_pd = ib_srq->pd;
1744 	pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
1745 	rdev = pd->rdev;
1746 	dev_attr = &rdev->dev_attr;
1747 	srq = container_of(ib_srq, struct bnxt_re_srq, ib_srq);
1748 
1749 	if (srq_init_attr->attr.max_wr >= dev_attr->max_srq_wqes) {
1750 		ibdev_err(&rdev->ibdev, "Create CQ failed - max exceeded");
1751 		rc = -EINVAL;
1752 		goto exit;
1753 	}
1754 
1755 	if (srq_init_attr->srq_type != IB_SRQT_BASIC) {
1756 		rc = -EOPNOTSUPP;
1757 		goto exit;
1758 	}
1759 
1760 	uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
1761 	srq->rdev = rdev;
1762 	srq->qplib_srq.pd = &pd->qplib_pd;
1763 	srq->qplib_srq.dpi = &rdev->dpi_privileged;
1764 	/* Allocate 1 more than what's provided so posting max doesn't
1765 	 * mean empty
1766 	 */
1767 	entries = bnxt_re_init_depth(srq_init_attr->attr.max_wr + 1, uctx);
1768 	if (entries > dev_attr->max_srq_wqes + 1)
1769 		entries = dev_attr->max_srq_wqes + 1;
1770 	srq->qplib_srq.max_wqe = entries;
1771 
1772 	srq->qplib_srq.max_sge = srq_init_attr->attr.max_sge;
1773 	 /* 128 byte wqe size for SRQ . So use max sges */
1774 	srq->qplib_srq.wqe_size = bnxt_re_get_rwqe_size(dev_attr->max_srq_sges);
1775 	srq->qplib_srq.threshold = srq_init_attr->attr.srq_limit;
1776 	srq->srq_limit = srq_init_attr->attr.srq_limit;
1777 	srq->qplib_srq.eventq_hw_ring_id = rdev->nq[0].ring_id;
1778 	nq = &rdev->nq[0];
1779 
1780 	if (udata) {
1781 		rc = bnxt_re_init_user_srq(rdev, pd, srq, udata);
1782 		if (rc)
1783 			goto fail;
1784 	}
1785 
1786 	rc = bnxt_qplib_create_srq(&rdev->qplib_res, &srq->qplib_srq);
1787 	if (rc) {
1788 		ibdev_err(&rdev->ibdev, "Create HW SRQ failed!");
1789 		goto fail;
1790 	}
1791 
1792 	if (udata) {
1793 		struct bnxt_re_srq_resp resp;
1794 
1795 		resp.srqid = srq->qplib_srq.id;
1796 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
1797 		if (rc) {
1798 			ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!");
1799 			bnxt_qplib_destroy_srq(&rdev->qplib_res,
1800 					       &srq->qplib_srq);
1801 			goto fail;
1802 		}
1803 	}
1804 	if (nq)
1805 		nq->budget++;
1806 	active_srqs = atomic_inc_return(&rdev->stats.res.srq_count);
1807 	if (active_srqs > rdev->stats.res.srq_watermark)
1808 		rdev->stats.res.srq_watermark = active_srqs;
1809 	spin_lock_init(&srq->lock);
1810 
1811 	return 0;
1812 
1813 fail:
1814 	ib_umem_release(srq->umem);
1815 exit:
1816 	return rc;
1817 }
1818 
bnxt_re_modify_srq(struct ib_srq * ib_srq,struct ib_srq_attr * srq_attr,enum ib_srq_attr_mask srq_attr_mask,struct ib_udata * udata)1819 int bnxt_re_modify_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr,
1820 		       enum ib_srq_attr_mask srq_attr_mask,
1821 		       struct ib_udata *udata)
1822 {
1823 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1824 					       ib_srq);
1825 	struct bnxt_re_dev *rdev = srq->rdev;
1826 	int rc;
1827 
1828 	switch (srq_attr_mask) {
1829 	case IB_SRQ_MAX_WR:
1830 		/* SRQ resize is not supported */
1831 		return -EINVAL;
1832 	case IB_SRQ_LIMIT:
1833 		/* Change the SRQ threshold */
1834 		if (srq_attr->srq_limit > srq->qplib_srq.max_wqe)
1835 			return -EINVAL;
1836 
1837 		srq->qplib_srq.threshold = srq_attr->srq_limit;
1838 		rc = bnxt_qplib_modify_srq(&rdev->qplib_res, &srq->qplib_srq);
1839 		if (rc) {
1840 			ibdev_err(&rdev->ibdev, "Modify HW SRQ failed!");
1841 			return rc;
1842 		}
1843 		/* On success, update the shadow */
1844 		srq->srq_limit = srq_attr->srq_limit;
1845 		/* No need to Build and send response back to udata */
1846 		return 0;
1847 	default:
1848 		ibdev_err(&rdev->ibdev,
1849 			  "Unsupported srq_attr_mask 0x%x", srq_attr_mask);
1850 		return -EINVAL;
1851 	}
1852 }
1853 
bnxt_re_query_srq(struct ib_srq * ib_srq,struct ib_srq_attr * srq_attr)1854 int bnxt_re_query_srq(struct ib_srq *ib_srq, struct ib_srq_attr *srq_attr)
1855 {
1856 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1857 					       ib_srq);
1858 	struct bnxt_re_srq tsrq;
1859 	struct bnxt_re_dev *rdev = srq->rdev;
1860 	int rc;
1861 
1862 	/* Get live SRQ attr */
1863 	tsrq.qplib_srq.id = srq->qplib_srq.id;
1864 	rc = bnxt_qplib_query_srq(&rdev->qplib_res, &tsrq.qplib_srq);
1865 	if (rc) {
1866 		ibdev_err(&rdev->ibdev, "Query HW SRQ failed!");
1867 		return rc;
1868 	}
1869 	srq_attr->max_wr = srq->qplib_srq.max_wqe;
1870 	srq_attr->max_sge = srq->qplib_srq.max_sge;
1871 	srq_attr->srq_limit = tsrq.qplib_srq.threshold;
1872 
1873 	return 0;
1874 }
1875 
bnxt_re_post_srq_recv(struct ib_srq * ib_srq,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)1876 int bnxt_re_post_srq_recv(struct ib_srq *ib_srq, const struct ib_recv_wr *wr,
1877 			  const struct ib_recv_wr **bad_wr)
1878 {
1879 	struct bnxt_re_srq *srq = container_of(ib_srq, struct bnxt_re_srq,
1880 					       ib_srq);
1881 	struct bnxt_qplib_swqe wqe;
1882 	unsigned long flags;
1883 	int rc = 0;
1884 
1885 	spin_lock_irqsave(&srq->lock, flags);
1886 	while (wr) {
1887 		/* Transcribe each ib_recv_wr to qplib_swqe */
1888 		wqe.num_sge = wr->num_sge;
1889 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
1890 		wqe.wr_id = wr->wr_id;
1891 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
1892 
1893 		rc = bnxt_qplib_post_srq_recv(&srq->qplib_srq, &wqe);
1894 		if (rc) {
1895 			*bad_wr = wr;
1896 			break;
1897 		}
1898 		wr = wr->next;
1899 	}
1900 	spin_unlock_irqrestore(&srq->lock, flags);
1901 
1902 	return rc;
1903 }
bnxt_re_modify_shadow_qp(struct bnxt_re_dev * rdev,struct bnxt_re_qp * qp1_qp,int qp_attr_mask)1904 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev *rdev,
1905 				    struct bnxt_re_qp *qp1_qp,
1906 				    int qp_attr_mask)
1907 {
1908 	struct bnxt_re_qp *qp = rdev->gsi_ctx.gsi_sqp;
1909 	int rc;
1910 
1911 	if (qp_attr_mask & IB_QP_STATE) {
1912 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1913 		qp->qplib_qp.state = qp1_qp->qplib_qp.state;
1914 	}
1915 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
1916 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
1917 		qp->qplib_qp.pkey_index = qp1_qp->qplib_qp.pkey_index;
1918 	}
1919 
1920 	if (qp_attr_mask & IB_QP_QKEY) {
1921 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
1922 		/* Using a Random  QKEY */
1923 		qp->qplib_qp.qkey = 0x81818181;
1924 	}
1925 	if (qp_attr_mask & IB_QP_SQ_PSN) {
1926 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
1927 		qp->qplib_qp.sq.psn = qp1_qp->qplib_qp.sq.psn;
1928 	}
1929 
1930 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
1931 	if (rc)
1932 		ibdev_err(&rdev->ibdev, "Failed to modify Shadow QP for QP1");
1933 	return rc;
1934 }
1935 
bnxt_re_modify_qp(struct ib_qp * ib_qp,struct ib_qp_attr * qp_attr,int qp_attr_mask,struct ib_udata * udata)1936 int bnxt_re_modify_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
1937 		      int qp_attr_mask, struct ib_udata *udata)
1938 {
1939 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
1940 	struct bnxt_re_dev *rdev = qp->rdev;
1941 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
1942 	enum ib_qp_state curr_qp_state, new_qp_state;
1943 	int rc, entries;
1944 	unsigned int flags;
1945 	u8 nw_type;
1946 
1947 	if (qp_attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
1948 		return -EOPNOTSUPP;
1949 
1950 	qp->qplib_qp.modify_flags = 0;
1951 	if (qp_attr_mask & IB_QP_STATE) {
1952 		curr_qp_state = __to_ib_qp_state(qp->qplib_qp.cur_qp_state);
1953 		new_qp_state = qp_attr->qp_state;
1954 		if (!ib_modify_qp_is_ok(curr_qp_state, new_qp_state,
1955 					ib_qp->qp_type, qp_attr_mask)) {
1956 			ibdev_err(&rdev->ibdev,
1957 				  "Invalid attribute mask: %#x specified ",
1958 				  qp_attr_mask);
1959 			ibdev_err(&rdev->ibdev,
1960 				  "for qpn: %#x type: %#x",
1961 				  ib_qp->qp_num, ib_qp->qp_type);
1962 			ibdev_err(&rdev->ibdev,
1963 				  "curr_qp_state=0x%x, new_qp_state=0x%x\n",
1964 				  curr_qp_state, new_qp_state);
1965 			return -EINVAL;
1966 		}
1967 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_STATE;
1968 		qp->qplib_qp.state = __from_ib_qp_state(qp_attr->qp_state);
1969 
1970 		if (!qp->sumem &&
1971 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1972 			ibdev_dbg(&rdev->ibdev,
1973 				  "Move QP = %p to flush list\n", qp);
1974 			flags = bnxt_re_lock_cqs(qp);
1975 			bnxt_qplib_add_flush_qp(&qp->qplib_qp);
1976 			bnxt_re_unlock_cqs(qp, flags);
1977 		}
1978 		if (!qp->sumem &&
1979 		    qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_RESET) {
1980 			ibdev_dbg(&rdev->ibdev,
1981 				  "Move QP = %p out of flush list\n", qp);
1982 			flags = bnxt_re_lock_cqs(qp);
1983 			bnxt_qplib_clean_qp(&qp->qplib_qp);
1984 			bnxt_re_unlock_cqs(qp, flags);
1985 		}
1986 	}
1987 	if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
1988 		qp->qplib_qp.modify_flags |=
1989 				CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY;
1990 		qp->qplib_qp.en_sqd_async_notify = true;
1991 	}
1992 	if (qp_attr_mask & IB_QP_ACCESS_FLAGS) {
1993 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS;
1994 		qp->qplib_qp.access =
1995 			__from_ib_access_flags(qp_attr->qp_access_flags);
1996 		/* LOCAL_WRITE access must be set to allow RC receive */
1997 		qp->qplib_qp.access |= BNXT_QPLIB_ACCESS_LOCAL_WRITE;
1998 		/* Temp: Set all params on QP as of now */
1999 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_WRITE;
2000 		qp->qplib_qp.access |= CMDQ_MODIFY_QP_ACCESS_REMOTE_READ;
2001 	}
2002 	if (qp_attr_mask & IB_QP_PKEY_INDEX) {
2003 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY;
2004 		qp->qplib_qp.pkey_index = qp_attr->pkey_index;
2005 	}
2006 	if (qp_attr_mask & IB_QP_QKEY) {
2007 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY;
2008 		qp->qplib_qp.qkey = qp_attr->qkey;
2009 	}
2010 	if (qp_attr_mask & IB_QP_AV) {
2011 		const struct ib_global_route *grh =
2012 			rdma_ah_read_grh(&qp_attr->ah_attr);
2013 		const struct ib_gid_attr *sgid_attr;
2014 		struct bnxt_re_gid_ctx *ctx;
2015 
2016 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_DGID |
2017 				     CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL |
2018 				     CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX |
2019 				     CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT |
2020 				     CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS |
2021 				     CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC |
2022 				     CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID;
2023 		memcpy(qp->qplib_qp.ah.dgid.data, grh->dgid.raw,
2024 		       sizeof(qp->qplib_qp.ah.dgid.data));
2025 		qp->qplib_qp.ah.flow_label = grh->flow_label;
2026 		sgid_attr = grh->sgid_attr;
2027 		/* Get the HW context of the GID. The reference
2028 		 * of GID table entry is already taken by the caller.
2029 		 */
2030 		ctx = rdma_read_gid_hw_context(sgid_attr);
2031 		qp->qplib_qp.ah.sgid_index = ctx->idx;
2032 		qp->qplib_qp.ah.host_sgid_index = grh->sgid_index;
2033 		qp->qplib_qp.ah.hop_limit = grh->hop_limit;
2034 		qp->qplib_qp.ah.traffic_class = grh->traffic_class;
2035 		qp->qplib_qp.ah.sl = rdma_ah_get_sl(&qp_attr->ah_attr);
2036 		ether_addr_copy(qp->qplib_qp.ah.dmac,
2037 				qp_attr->ah_attr.roce.dmac);
2038 
2039 		rc = rdma_read_gid_l2_fields(sgid_attr, NULL,
2040 					     &qp->qplib_qp.smac[0]);
2041 		if (rc)
2042 			return rc;
2043 
2044 		nw_type = rdma_gid_attr_network_type(sgid_attr);
2045 		switch (nw_type) {
2046 		case RDMA_NETWORK_IPV4:
2047 			qp->qplib_qp.nw_type =
2048 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4;
2049 			break;
2050 		case RDMA_NETWORK_IPV6:
2051 			qp->qplib_qp.nw_type =
2052 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6;
2053 			break;
2054 		default:
2055 			qp->qplib_qp.nw_type =
2056 				CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1;
2057 			break;
2058 		}
2059 	}
2060 
2061 	if (qp_attr->qp_state == IB_QPS_RTR) {
2062 		enum ib_mtu qpmtu;
2063 
2064 		qpmtu = iboe_get_mtu(rdev->netdev->mtu);
2065 		if (qp_attr_mask & IB_QP_PATH_MTU) {
2066 			if (ib_mtu_enum_to_int(qp_attr->path_mtu) >
2067 			    ib_mtu_enum_to_int(qpmtu))
2068 				return -EINVAL;
2069 			qpmtu = qp_attr->path_mtu;
2070 		}
2071 
2072 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU;
2073 		qp->qplib_qp.path_mtu = __from_ib_mtu(qpmtu);
2074 		qp->qplib_qp.mtu = ib_mtu_enum_to_int(qpmtu);
2075 	}
2076 
2077 	if (qp_attr_mask & IB_QP_TIMEOUT) {
2078 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT;
2079 		qp->qplib_qp.timeout = qp_attr->timeout;
2080 	}
2081 	if (qp_attr_mask & IB_QP_RETRY_CNT) {
2082 		qp->qplib_qp.modify_flags |=
2083 				CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT;
2084 		qp->qplib_qp.retry_cnt = qp_attr->retry_cnt;
2085 	}
2086 	if (qp_attr_mask & IB_QP_RNR_RETRY) {
2087 		qp->qplib_qp.modify_flags |=
2088 				CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY;
2089 		qp->qplib_qp.rnr_retry = qp_attr->rnr_retry;
2090 	}
2091 	if (qp_attr_mask & IB_QP_MIN_RNR_TIMER) {
2092 		qp->qplib_qp.modify_flags |=
2093 				CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER;
2094 		qp->qplib_qp.min_rnr_timer = qp_attr->min_rnr_timer;
2095 	}
2096 	if (qp_attr_mask & IB_QP_RQ_PSN) {
2097 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN;
2098 		qp->qplib_qp.rq.psn = qp_attr->rq_psn;
2099 	}
2100 	if (qp_attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2101 		qp->qplib_qp.modify_flags |=
2102 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC;
2103 		/* Cap the max_rd_atomic to device max */
2104 		qp->qplib_qp.max_rd_atomic = min_t(u32, qp_attr->max_rd_atomic,
2105 						   dev_attr->max_qp_rd_atom);
2106 	}
2107 	if (qp_attr_mask & IB_QP_SQ_PSN) {
2108 		qp->qplib_qp.modify_flags |= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN;
2109 		qp->qplib_qp.sq.psn = qp_attr->sq_psn;
2110 	}
2111 	if (qp_attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2112 		if (qp_attr->max_dest_rd_atomic >
2113 		    dev_attr->max_qp_init_rd_atom) {
2114 			ibdev_err(&rdev->ibdev,
2115 				  "max_dest_rd_atomic requested%d is > dev_max%d",
2116 				  qp_attr->max_dest_rd_atomic,
2117 				  dev_attr->max_qp_init_rd_atom);
2118 			return -EINVAL;
2119 		}
2120 
2121 		qp->qplib_qp.modify_flags |=
2122 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC;
2123 		qp->qplib_qp.max_dest_rd_atomic = qp_attr->max_dest_rd_atomic;
2124 	}
2125 	if (qp_attr_mask & IB_QP_CAP) {
2126 		struct bnxt_re_ucontext *uctx =
2127 			rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
2128 
2129 		qp->qplib_qp.modify_flags |=
2130 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE |
2131 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE |
2132 				CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE |
2133 				CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE |
2134 				CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA;
2135 		if ((qp_attr->cap.max_send_wr >= dev_attr->max_qp_wqes) ||
2136 		    (qp_attr->cap.max_recv_wr >= dev_attr->max_qp_wqes) ||
2137 		    (qp_attr->cap.max_send_sge >= dev_attr->max_qp_sges) ||
2138 		    (qp_attr->cap.max_recv_sge >= dev_attr->max_qp_sges) ||
2139 		    (qp_attr->cap.max_inline_data >=
2140 						dev_attr->max_inline_data)) {
2141 			ibdev_err(&rdev->ibdev,
2142 				  "Create QP failed - max exceeded");
2143 			return -EINVAL;
2144 		}
2145 		entries = bnxt_re_init_depth(qp_attr->cap.max_send_wr, uctx);
2146 		qp->qplib_qp.sq.max_wqe = min_t(u32, entries,
2147 						dev_attr->max_qp_wqes + 1);
2148 		qp->qplib_qp.sq.q_full_delta = qp->qplib_qp.sq.max_wqe -
2149 						qp_attr->cap.max_send_wr;
2150 		/*
2151 		 * Reserving one slot for Phantom WQE. Some application can
2152 		 * post one extra entry in this case. Allowing this to avoid
2153 		 * unexpected Queue full condition
2154 		 */
2155 		qp->qplib_qp.sq.q_full_delta -= 1;
2156 		qp->qplib_qp.sq.max_sge = qp_attr->cap.max_send_sge;
2157 		if (qp->qplib_qp.rq.max_wqe) {
2158 			entries = bnxt_re_init_depth(qp_attr->cap.max_recv_wr, uctx);
2159 			qp->qplib_qp.rq.max_wqe =
2160 				min_t(u32, entries, dev_attr->max_qp_wqes + 1);
2161 			qp->qplib_qp.rq.max_sw_wqe = qp->qplib_qp.rq.max_wqe;
2162 			qp->qplib_qp.rq.q_full_delta = qp->qplib_qp.rq.max_wqe -
2163 						       qp_attr->cap.max_recv_wr;
2164 			qp->qplib_qp.rq.max_sge = qp_attr->cap.max_recv_sge;
2165 		} else {
2166 			/* SRQ was used prior, just ignore the RQ caps */
2167 		}
2168 	}
2169 	if (qp_attr_mask & IB_QP_DEST_QPN) {
2170 		qp->qplib_qp.modify_flags |=
2171 				CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID;
2172 		qp->qplib_qp.dest_qpn = qp_attr->dest_qp_num;
2173 	}
2174 	rc = bnxt_qplib_modify_qp(&rdev->qplib_res, &qp->qplib_qp);
2175 	if (rc) {
2176 		ibdev_err(&rdev->ibdev, "Failed to modify HW QP");
2177 		return rc;
2178 	}
2179 	if (ib_qp->qp_type == IB_QPT_GSI && rdev->gsi_ctx.gsi_sqp)
2180 		rc = bnxt_re_modify_shadow_qp(rdev, qp, qp_attr_mask);
2181 	return rc;
2182 }
2183 
bnxt_re_query_qp(struct ib_qp * ib_qp,struct ib_qp_attr * qp_attr,int qp_attr_mask,struct ib_qp_init_attr * qp_init_attr)2184 int bnxt_re_query_qp(struct ib_qp *ib_qp, struct ib_qp_attr *qp_attr,
2185 		     int qp_attr_mask, struct ib_qp_init_attr *qp_init_attr)
2186 {
2187 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2188 	struct bnxt_re_dev *rdev = qp->rdev;
2189 	struct bnxt_qplib_qp *qplib_qp;
2190 	int rc;
2191 
2192 	qplib_qp = kzalloc(sizeof(*qplib_qp), GFP_KERNEL);
2193 	if (!qplib_qp)
2194 		return -ENOMEM;
2195 
2196 	qplib_qp->id = qp->qplib_qp.id;
2197 	qplib_qp->ah.host_sgid_index = qp->qplib_qp.ah.host_sgid_index;
2198 
2199 	rc = bnxt_qplib_query_qp(&rdev->qplib_res, qplib_qp);
2200 	if (rc) {
2201 		ibdev_err(&rdev->ibdev, "Failed to query HW QP");
2202 		goto out;
2203 	}
2204 	qp_attr->qp_state = __to_ib_qp_state(qplib_qp->state);
2205 	qp_attr->cur_qp_state = __to_ib_qp_state(qplib_qp->cur_qp_state);
2206 	qp_attr->en_sqd_async_notify = qplib_qp->en_sqd_async_notify ? 1 : 0;
2207 	qp_attr->qp_access_flags = __to_ib_access_flags(qplib_qp->access);
2208 	qp_attr->pkey_index = qplib_qp->pkey_index;
2209 	qp_attr->qkey = qplib_qp->qkey;
2210 	qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2211 	rdma_ah_set_grh(&qp_attr->ah_attr, NULL, qplib_qp->ah.flow_label,
2212 			qplib_qp->ah.host_sgid_index,
2213 			qplib_qp->ah.hop_limit,
2214 			qplib_qp->ah.traffic_class);
2215 	rdma_ah_set_dgid_raw(&qp_attr->ah_attr, qplib_qp->ah.dgid.data);
2216 	rdma_ah_set_sl(&qp_attr->ah_attr, qplib_qp->ah.sl);
2217 	ether_addr_copy(qp_attr->ah_attr.roce.dmac, qplib_qp->ah.dmac);
2218 	qp_attr->path_mtu = __to_ib_mtu(qplib_qp->path_mtu);
2219 	qp_attr->timeout = qplib_qp->timeout;
2220 	qp_attr->retry_cnt = qplib_qp->retry_cnt;
2221 	qp_attr->rnr_retry = qplib_qp->rnr_retry;
2222 	qp_attr->min_rnr_timer = qplib_qp->min_rnr_timer;
2223 	qp_attr->port_num = __to_ib_port_num(qplib_qp->port_id);
2224 	qp_attr->rq_psn = qplib_qp->rq.psn;
2225 	qp_attr->max_rd_atomic = qplib_qp->max_rd_atomic;
2226 	qp_attr->sq_psn = qplib_qp->sq.psn;
2227 	qp_attr->max_dest_rd_atomic = qplib_qp->max_dest_rd_atomic;
2228 	qp_init_attr->sq_sig_type = qplib_qp->sig_type ? IB_SIGNAL_ALL_WR :
2229 							 IB_SIGNAL_REQ_WR;
2230 	qp_attr->dest_qp_num = qplib_qp->dest_qpn;
2231 
2232 	qp_attr->cap.max_send_wr = qp->qplib_qp.sq.max_wqe;
2233 	qp_attr->cap.max_send_sge = qp->qplib_qp.sq.max_sge;
2234 	qp_attr->cap.max_recv_wr = qp->qplib_qp.rq.max_wqe;
2235 	qp_attr->cap.max_recv_sge = qp->qplib_qp.rq.max_sge;
2236 	qp_attr->cap.max_inline_data = qp->qplib_qp.max_inline_data;
2237 	qp_init_attr->cap = qp_attr->cap;
2238 
2239 out:
2240 	kfree(qplib_qp);
2241 	return rc;
2242 }
2243 
2244 /* Routine for sending QP1 packets for RoCE V1 an V2
2245  */
bnxt_re_build_qp1_send_v2(struct bnxt_re_qp * qp,const struct ib_send_wr * wr,struct bnxt_qplib_swqe * wqe,int payload_size)2246 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp *qp,
2247 				     const struct ib_send_wr *wr,
2248 				     struct bnxt_qplib_swqe *wqe,
2249 				     int payload_size)
2250 {
2251 	struct bnxt_re_ah *ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah,
2252 					     ib_ah);
2253 	struct bnxt_qplib_ah *qplib_ah = &ah->qplib_ah;
2254 	const struct ib_gid_attr *sgid_attr = ah->ib_ah.sgid_attr;
2255 	struct bnxt_qplib_sge sge;
2256 	u8 nw_type;
2257 	u16 ether_type;
2258 	union ib_gid dgid;
2259 	bool is_eth = false;
2260 	bool is_vlan = false;
2261 	bool is_grh = false;
2262 	bool is_udp = false;
2263 	u8 ip_version = 0;
2264 	u16 vlan_id = 0xFFFF;
2265 	void *buf;
2266 	int i, rc;
2267 
2268 	memset(&qp->qp1_hdr, 0, sizeof(qp->qp1_hdr));
2269 
2270 	rc = rdma_read_gid_l2_fields(sgid_attr, &vlan_id, NULL);
2271 	if (rc)
2272 		return rc;
2273 
2274 	/* Get network header type for this GID */
2275 	nw_type = rdma_gid_attr_network_type(sgid_attr);
2276 	switch (nw_type) {
2277 	case RDMA_NETWORK_IPV4:
2278 		nw_type = BNXT_RE_ROCEV2_IPV4_PACKET;
2279 		break;
2280 	case RDMA_NETWORK_IPV6:
2281 		nw_type = BNXT_RE_ROCEV2_IPV6_PACKET;
2282 		break;
2283 	default:
2284 		nw_type = BNXT_RE_ROCE_V1_PACKET;
2285 		break;
2286 	}
2287 	memcpy(&dgid.raw, &qplib_ah->dgid, 16);
2288 	is_udp = sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2289 	if (is_udp) {
2290 		if (ipv6_addr_v4mapped((struct in6_addr *)&sgid_attr->gid)) {
2291 			ip_version = 4;
2292 			ether_type = ETH_P_IP;
2293 		} else {
2294 			ip_version = 6;
2295 			ether_type = ETH_P_IPV6;
2296 		}
2297 		is_grh = false;
2298 	} else {
2299 		ether_type = ETH_P_IBOE;
2300 		is_grh = true;
2301 	}
2302 
2303 	is_eth = true;
2304 	is_vlan = vlan_id && (vlan_id < 0x1000);
2305 
2306 	ib_ud_header_init(payload_size, !is_eth, is_eth, is_vlan, is_grh,
2307 			  ip_version, is_udp, 0, &qp->qp1_hdr);
2308 
2309 	/* ETH */
2310 	ether_addr_copy(qp->qp1_hdr.eth.dmac_h, ah->qplib_ah.dmac);
2311 	ether_addr_copy(qp->qp1_hdr.eth.smac_h, qp->qplib_qp.smac);
2312 
2313 	/* For vlan, check the sgid for vlan existence */
2314 
2315 	if (!is_vlan) {
2316 		qp->qp1_hdr.eth.type = cpu_to_be16(ether_type);
2317 	} else {
2318 		qp->qp1_hdr.vlan.type = cpu_to_be16(ether_type);
2319 		qp->qp1_hdr.vlan.tag = cpu_to_be16(vlan_id);
2320 	}
2321 
2322 	if (is_grh || (ip_version == 6)) {
2323 		memcpy(qp->qp1_hdr.grh.source_gid.raw, sgid_attr->gid.raw,
2324 		       sizeof(sgid_attr->gid));
2325 		memcpy(qp->qp1_hdr.grh.destination_gid.raw, qplib_ah->dgid.data,
2326 		       sizeof(sgid_attr->gid));
2327 		qp->qp1_hdr.grh.hop_limit     = qplib_ah->hop_limit;
2328 	}
2329 
2330 	if (ip_version == 4) {
2331 		qp->qp1_hdr.ip4.tos = 0;
2332 		qp->qp1_hdr.ip4.id = 0;
2333 		qp->qp1_hdr.ip4.frag_off = htons(IP_DF);
2334 		qp->qp1_hdr.ip4.ttl = qplib_ah->hop_limit;
2335 
2336 		memcpy(&qp->qp1_hdr.ip4.saddr, sgid_attr->gid.raw + 12, 4);
2337 		memcpy(&qp->qp1_hdr.ip4.daddr, qplib_ah->dgid.data + 12, 4);
2338 		qp->qp1_hdr.ip4.check = ib_ud_ip4_csum(&qp->qp1_hdr);
2339 	}
2340 
2341 	if (is_udp) {
2342 		qp->qp1_hdr.udp.dport = htons(ROCE_V2_UDP_DPORT);
2343 		qp->qp1_hdr.udp.sport = htons(0x8CD1);
2344 		qp->qp1_hdr.udp.csum = 0;
2345 	}
2346 
2347 	/* BTH */
2348 	if (wr->opcode == IB_WR_SEND_WITH_IMM) {
2349 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2350 		qp->qp1_hdr.immediate_present = 1;
2351 	} else {
2352 		qp->qp1_hdr.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2353 	}
2354 	if (wr->send_flags & IB_SEND_SOLICITED)
2355 		qp->qp1_hdr.bth.solicited_event = 1;
2356 	/* pad_count */
2357 	qp->qp1_hdr.bth.pad_count = (4 - payload_size) & 3;
2358 
2359 	/* P_key for QP1 is for all members */
2360 	qp->qp1_hdr.bth.pkey = cpu_to_be16(0xFFFF);
2361 	qp->qp1_hdr.bth.destination_qpn = IB_QP1;
2362 	qp->qp1_hdr.bth.ack_req = 0;
2363 	qp->send_psn++;
2364 	qp->send_psn &= BTH_PSN_MASK;
2365 	qp->qp1_hdr.bth.psn = cpu_to_be32(qp->send_psn);
2366 	/* DETH */
2367 	/* Use the priviledged Q_Key for QP1 */
2368 	qp->qp1_hdr.deth.qkey = cpu_to_be32(IB_QP1_QKEY);
2369 	qp->qp1_hdr.deth.source_qpn = IB_QP1;
2370 
2371 	/* Pack the QP1 to the transmit buffer */
2372 	buf = bnxt_qplib_get_qp1_sq_buf(&qp->qplib_qp, &sge);
2373 	if (buf) {
2374 		ib_ud_header_pack(&qp->qp1_hdr, buf);
2375 		for (i = wqe->num_sge; i; i--) {
2376 			wqe->sg_list[i].addr = wqe->sg_list[i - 1].addr;
2377 			wqe->sg_list[i].lkey = wqe->sg_list[i - 1].lkey;
2378 			wqe->sg_list[i].size = wqe->sg_list[i - 1].size;
2379 		}
2380 
2381 		/*
2382 		 * Max Header buf size for IPV6 RoCE V2 is 86,
2383 		 * which is same as the QP1 SQ header buffer.
2384 		 * Header buf size for IPV4 RoCE V2 can be 66.
2385 		 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
2386 		 * Subtract 20 bytes from QP1 SQ header buf size
2387 		 */
2388 		if (is_udp && ip_version == 4)
2389 			sge.size -= 20;
2390 		/*
2391 		 * Max Header buf size for RoCE V1 is 78.
2392 		 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
2393 		 * Subtract 8 bytes from QP1 SQ header buf size
2394 		 */
2395 		if (!is_udp)
2396 			sge.size -= 8;
2397 
2398 		/* Subtract 4 bytes for non vlan packets */
2399 		if (!is_vlan)
2400 			sge.size -= 4;
2401 
2402 		wqe->sg_list[0].addr = sge.addr;
2403 		wqe->sg_list[0].lkey = sge.lkey;
2404 		wqe->sg_list[0].size = sge.size;
2405 		wqe->num_sge++;
2406 
2407 	} else {
2408 		ibdev_err(&qp->rdev->ibdev, "QP1 buffer is empty!");
2409 		rc = -ENOMEM;
2410 	}
2411 	return rc;
2412 }
2413 
2414 /* For the MAD layer, it only provides the recv SGE the size of
2415  * ib_grh + MAD datagram.  No Ethernet headers, Ethertype, BTH, DETH,
2416  * nor RoCE iCRC.  The Cu+ solution must provide buffer for the entire
2417  * receive packet (334 bytes) with no VLAN and then copy the GRH
2418  * and the MAD datagram out to the provided SGE.
2419  */
bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp * qp,const struct ib_recv_wr * wr,struct bnxt_qplib_swqe * wqe,int payload_size)2420 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp *qp,
2421 					    const struct ib_recv_wr *wr,
2422 					    struct bnxt_qplib_swqe *wqe,
2423 					    int payload_size)
2424 {
2425 	struct bnxt_re_sqp_entries *sqp_entry;
2426 	struct bnxt_qplib_sge ref, sge;
2427 	struct bnxt_re_dev *rdev;
2428 	u32 rq_prod_index;
2429 
2430 	rdev = qp->rdev;
2431 
2432 	rq_prod_index = bnxt_qplib_get_rq_prod_index(&qp->qplib_qp);
2433 
2434 	if (!bnxt_qplib_get_qp1_rq_buf(&qp->qplib_qp, &sge))
2435 		return -ENOMEM;
2436 
2437 	/* Create 1 SGE to receive the entire
2438 	 * ethernet packet
2439 	 */
2440 	/* Save the reference from ULP */
2441 	ref.addr = wqe->sg_list[0].addr;
2442 	ref.lkey = wqe->sg_list[0].lkey;
2443 	ref.size = wqe->sg_list[0].size;
2444 
2445 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[rq_prod_index];
2446 
2447 	/* SGE 1 */
2448 	wqe->sg_list[0].addr = sge.addr;
2449 	wqe->sg_list[0].lkey = sge.lkey;
2450 	wqe->sg_list[0].size = BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2;
2451 	sge.size -= wqe->sg_list[0].size;
2452 
2453 	sqp_entry->sge.addr = ref.addr;
2454 	sqp_entry->sge.lkey = ref.lkey;
2455 	sqp_entry->sge.size = ref.size;
2456 	/* Store the wrid for reporting completion */
2457 	sqp_entry->wrid = wqe->wr_id;
2458 	/* change the wqe->wrid to table index */
2459 	wqe->wr_id = rq_prod_index;
2460 	return 0;
2461 }
2462 
is_ud_qp(struct bnxt_re_qp * qp)2463 static int is_ud_qp(struct bnxt_re_qp *qp)
2464 {
2465 	return (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_UD ||
2466 		qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI);
2467 }
2468 
bnxt_re_build_send_wqe(struct bnxt_re_qp * qp,const struct ib_send_wr * wr,struct bnxt_qplib_swqe * wqe)2469 static int bnxt_re_build_send_wqe(struct bnxt_re_qp *qp,
2470 				  const struct ib_send_wr *wr,
2471 				  struct bnxt_qplib_swqe *wqe)
2472 {
2473 	struct bnxt_re_ah *ah = NULL;
2474 
2475 	if (is_ud_qp(qp)) {
2476 		ah = container_of(ud_wr(wr)->ah, struct bnxt_re_ah, ib_ah);
2477 		wqe->send.q_key = ud_wr(wr)->remote_qkey;
2478 		wqe->send.dst_qp = ud_wr(wr)->remote_qpn;
2479 		wqe->send.avid = ah->qplib_ah.id;
2480 	}
2481 	switch (wr->opcode) {
2482 	case IB_WR_SEND:
2483 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND;
2484 		break;
2485 	case IB_WR_SEND_WITH_IMM:
2486 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM;
2487 		wqe->send.imm_data = be32_to_cpu(wr->ex.imm_data);
2488 		break;
2489 	case IB_WR_SEND_WITH_INV:
2490 		wqe->type = BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV;
2491 		wqe->send.inv_key = wr->ex.invalidate_rkey;
2492 		break;
2493 	default:
2494 		return -EINVAL;
2495 	}
2496 	if (wr->send_flags & IB_SEND_SIGNALED)
2497 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2498 	if (wr->send_flags & IB_SEND_FENCE)
2499 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2500 	if (wr->send_flags & IB_SEND_SOLICITED)
2501 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2502 	if (wr->send_flags & IB_SEND_INLINE)
2503 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2504 
2505 	return 0;
2506 }
2507 
bnxt_re_build_rdma_wqe(const struct ib_send_wr * wr,struct bnxt_qplib_swqe * wqe)2508 static int bnxt_re_build_rdma_wqe(const struct ib_send_wr *wr,
2509 				  struct bnxt_qplib_swqe *wqe)
2510 {
2511 	switch (wr->opcode) {
2512 	case IB_WR_RDMA_WRITE:
2513 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE;
2514 		break;
2515 	case IB_WR_RDMA_WRITE_WITH_IMM:
2516 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM;
2517 		wqe->rdma.imm_data = be32_to_cpu(wr->ex.imm_data);
2518 		break;
2519 	case IB_WR_RDMA_READ:
2520 		wqe->type = BNXT_QPLIB_SWQE_TYPE_RDMA_READ;
2521 		wqe->rdma.inv_key = wr->ex.invalidate_rkey;
2522 		break;
2523 	default:
2524 		return -EINVAL;
2525 	}
2526 	wqe->rdma.remote_va = rdma_wr(wr)->remote_addr;
2527 	wqe->rdma.r_key = rdma_wr(wr)->rkey;
2528 	if (wr->send_flags & IB_SEND_SIGNALED)
2529 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2530 	if (wr->send_flags & IB_SEND_FENCE)
2531 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2532 	if (wr->send_flags & IB_SEND_SOLICITED)
2533 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2534 	if (wr->send_flags & IB_SEND_INLINE)
2535 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_INLINE;
2536 
2537 	return 0;
2538 }
2539 
bnxt_re_build_atomic_wqe(const struct ib_send_wr * wr,struct bnxt_qplib_swqe * wqe)2540 static int bnxt_re_build_atomic_wqe(const struct ib_send_wr *wr,
2541 				    struct bnxt_qplib_swqe *wqe)
2542 {
2543 	switch (wr->opcode) {
2544 	case IB_WR_ATOMIC_CMP_AND_SWP:
2545 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP;
2546 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2547 		wqe->atomic.swap_data = atomic_wr(wr)->swap;
2548 		break;
2549 	case IB_WR_ATOMIC_FETCH_AND_ADD:
2550 		wqe->type = BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD;
2551 		wqe->atomic.cmp_data = atomic_wr(wr)->compare_add;
2552 		break;
2553 	default:
2554 		return -EINVAL;
2555 	}
2556 	wqe->atomic.remote_va = atomic_wr(wr)->remote_addr;
2557 	wqe->atomic.r_key = atomic_wr(wr)->rkey;
2558 	if (wr->send_flags & IB_SEND_SIGNALED)
2559 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2560 	if (wr->send_flags & IB_SEND_FENCE)
2561 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2562 	if (wr->send_flags & IB_SEND_SOLICITED)
2563 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2564 	return 0;
2565 }
2566 
bnxt_re_build_inv_wqe(const struct ib_send_wr * wr,struct bnxt_qplib_swqe * wqe)2567 static int bnxt_re_build_inv_wqe(const struct ib_send_wr *wr,
2568 				 struct bnxt_qplib_swqe *wqe)
2569 {
2570 	wqe->type = BNXT_QPLIB_SWQE_TYPE_LOCAL_INV;
2571 	wqe->local_inv.inv_l_key = wr->ex.invalidate_rkey;
2572 
2573 	if (wr->send_flags & IB_SEND_SIGNALED)
2574 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2575 	if (wr->send_flags & IB_SEND_SOLICITED)
2576 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT;
2577 
2578 	return 0;
2579 }
2580 
bnxt_re_build_reg_wqe(const struct ib_reg_wr * wr,struct bnxt_qplib_swqe * wqe)2581 static int bnxt_re_build_reg_wqe(const struct ib_reg_wr *wr,
2582 				 struct bnxt_qplib_swqe *wqe)
2583 {
2584 	struct bnxt_re_mr *mr = container_of(wr->mr, struct bnxt_re_mr, ib_mr);
2585 	struct bnxt_qplib_frpl *qplib_frpl = &mr->qplib_frpl;
2586 	int access = wr->access;
2587 
2588 	wqe->frmr.pbl_ptr = (__le64 *)qplib_frpl->hwq.pbl_ptr[0];
2589 	wqe->frmr.pbl_dma_ptr = qplib_frpl->hwq.pbl_dma_ptr[0];
2590 	wqe->frmr.page_list = mr->pages;
2591 	wqe->frmr.page_list_len = mr->npages;
2592 	wqe->frmr.levels = qplib_frpl->hwq.level;
2593 	wqe->type = BNXT_QPLIB_SWQE_TYPE_REG_MR;
2594 
2595 	if (wr->wr.send_flags & IB_SEND_SIGNALED)
2596 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP;
2597 
2598 	if (access & IB_ACCESS_LOCAL_WRITE)
2599 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE;
2600 	if (access & IB_ACCESS_REMOTE_READ)
2601 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ;
2602 	if (access & IB_ACCESS_REMOTE_WRITE)
2603 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE;
2604 	if (access & IB_ACCESS_REMOTE_ATOMIC)
2605 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC;
2606 	if (access & IB_ACCESS_MW_BIND)
2607 		wqe->frmr.access_cntl |= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND;
2608 
2609 	wqe->frmr.l_key = wr->key;
2610 	wqe->frmr.length = wr->mr->length;
2611 	wqe->frmr.pbl_pg_sz_log = ilog2(PAGE_SIZE >> PAGE_SHIFT_4K);
2612 	wqe->frmr.pg_sz_log = ilog2(wr->mr->page_size >> PAGE_SHIFT_4K);
2613 	wqe->frmr.va = wr->mr->iova;
2614 	return 0;
2615 }
2616 
bnxt_re_copy_inline_data(struct bnxt_re_dev * rdev,const struct ib_send_wr * wr,struct bnxt_qplib_swqe * wqe)2617 static int bnxt_re_copy_inline_data(struct bnxt_re_dev *rdev,
2618 				    const struct ib_send_wr *wr,
2619 				    struct bnxt_qplib_swqe *wqe)
2620 {
2621 	/*  Copy the inline data to the data  field */
2622 	u8 *in_data;
2623 	u32 i, sge_len;
2624 	void *sge_addr;
2625 
2626 	in_data = wqe->inline_data;
2627 	for (i = 0; i < wr->num_sge; i++) {
2628 		sge_addr = (void *)(unsigned long)
2629 				wr->sg_list[i].addr;
2630 		sge_len = wr->sg_list[i].length;
2631 
2632 		if ((sge_len + wqe->inline_len) >
2633 		    BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH) {
2634 			ibdev_err(&rdev->ibdev,
2635 				  "Inline data size requested > supported value");
2636 			return -EINVAL;
2637 		}
2638 		sge_len = wr->sg_list[i].length;
2639 
2640 		memcpy(in_data, sge_addr, sge_len);
2641 		in_data += wr->sg_list[i].length;
2642 		wqe->inline_len += wr->sg_list[i].length;
2643 	}
2644 	return wqe->inline_len;
2645 }
2646 
bnxt_re_copy_wr_payload(struct bnxt_re_dev * rdev,const struct ib_send_wr * wr,struct bnxt_qplib_swqe * wqe)2647 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev *rdev,
2648 				   const struct ib_send_wr *wr,
2649 				   struct bnxt_qplib_swqe *wqe)
2650 {
2651 	int payload_sz = 0;
2652 
2653 	if (wr->send_flags & IB_SEND_INLINE)
2654 		payload_sz = bnxt_re_copy_inline_data(rdev, wr, wqe);
2655 	else
2656 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe->sg_list,
2657 					       wqe->num_sge);
2658 
2659 	return payload_sz;
2660 }
2661 
bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp * qp)2662 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp *qp)
2663 {
2664 	if ((qp->ib_qp.qp_type == IB_QPT_UD ||
2665 	     qp->ib_qp.qp_type == IB_QPT_GSI ||
2666 	     qp->ib_qp.qp_type == IB_QPT_RAW_ETHERTYPE) &&
2667 	     qp->qplib_qp.wqe_cnt == BNXT_RE_UD_QP_HW_STALL) {
2668 		int qp_attr_mask;
2669 		struct ib_qp_attr qp_attr;
2670 
2671 		qp_attr_mask = IB_QP_STATE;
2672 		qp_attr.qp_state = IB_QPS_RTS;
2673 		bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, qp_attr_mask, NULL);
2674 		qp->qplib_qp.wqe_cnt = 0;
2675 	}
2676 }
2677 
bnxt_re_post_send_shadow_qp(struct bnxt_re_dev * rdev,struct bnxt_re_qp * qp,const struct ib_send_wr * wr)2678 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev *rdev,
2679 				       struct bnxt_re_qp *qp,
2680 				       const struct ib_send_wr *wr)
2681 {
2682 	int rc = 0, payload_sz = 0;
2683 	unsigned long flags;
2684 
2685 	spin_lock_irqsave(&qp->sq_lock, flags);
2686 	while (wr) {
2687 		struct bnxt_qplib_swqe wqe = {};
2688 
2689 		/* Common */
2690 		wqe.num_sge = wr->num_sge;
2691 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2692 			ibdev_err(&rdev->ibdev,
2693 				  "Limit exceeded for Send SGEs");
2694 			rc = -EINVAL;
2695 			goto bad;
2696 		}
2697 
2698 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2699 		if (payload_sz < 0) {
2700 			rc = -EINVAL;
2701 			goto bad;
2702 		}
2703 		wqe.wr_id = wr->wr_id;
2704 
2705 		wqe.type = BNXT_QPLIB_SWQE_TYPE_SEND;
2706 
2707 		rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2708 		if (!rc)
2709 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2710 bad:
2711 		if (rc) {
2712 			ibdev_err(&rdev->ibdev,
2713 				  "Post send failed opcode = %#x rc = %d",
2714 				  wr->opcode, rc);
2715 			break;
2716 		}
2717 		wr = wr->next;
2718 	}
2719 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2720 	if (!bnxt_qplib_is_chip_gen_p5_p7(qp->rdev->chip_ctx))
2721 		bnxt_ud_qp_hw_stall_workaround(qp);
2722 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2723 	return rc;
2724 }
2725 
bnxt_re_legacy_set_uc_fence(struct bnxt_qplib_swqe * wqe)2726 static void bnxt_re_legacy_set_uc_fence(struct bnxt_qplib_swqe *wqe)
2727 {
2728 	/* Need unconditional fence for non-wire memory opcode
2729 	 * to work as expected.
2730 	 */
2731 	if (wqe->type == BNXT_QPLIB_SWQE_TYPE_LOCAL_INV ||
2732 	    wqe->type == BNXT_QPLIB_SWQE_TYPE_FAST_REG_MR ||
2733 	    wqe->type == BNXT_QPLIB_SWQE_TYPE_REG_MR ||
2734 	    wqe->type == BNXT_QPLIB_SWQE_TYPE_BIND_MW)
2735 		wqe->flags |= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE;
2736 }
2737 
bnxt_re_post_send(struct ib_qp * ib_qp,const struct ib_send_wr * wr,const struct ib_send_wr ** bad_wr)2738 int bnxt_re_post_send(struct ib_qp *ib_qp, const struct ib_send_wr *wr,
2739 		      const struct ib_send_wr **bad_wr)
2740 {
2741 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2742 	struct bnxt_qplib_swqe wqe;
2743 	int rc = 0, payload_sz = 0;
2744 	unsigned long flags;
2745 
2746 	spin_lock_irqsave(&qp->sq_lock, flags);
2747 	while (wr) {
2748 		/* House keeping */
2749 		memset(&wqe, 0, sizeof(wqe));
2750 
2751 		/* Common */
2752 		wqe.num_sge = wr->num_sge;
2753 		if (wr->num_sge > qp->qplib_qp.sq.max_sge) {
2754 			ibdev_err(&qp->rdev->ibdev,
2755 				  "Limit exceeded for Send SGEs");
2756 			rc = -EINVAL;
2757 			goto bad;
2758 		}
2759 
2760 		payload_sz = bnxt_re_copy_wr_payload(qp->rdev, wr, &wqe);
2761 		if (payload_sz < 0) {
2762 			rc = -EINVAL;
2763 			goto bad;
2764 		}
2765 		wqe.wr_id = wr->wr_id;
2766 
2767 		switch (wr->opcode) {
2768 		case IB_WR_SEND:
2769 		case IB_WR_SEND_WITH_IMM:
2770 			if (qp->qplib_qp.type == CMDQ_CREATE_QP1_TYPE_GSI) {
2771 				rc = bnxt_re_build_qp1_send_v2(qp, wr, &wqe,
2772 							       payload_sz);
2773 				if (rc)
2774 					goto bad;
2775 				wqe.rawqp1.lflags |=
2776 					SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC;
2777 			}
2778 			switch (wr->send_flags) {
2779 			case IB_SEND_IP_CSUM:
2780 				wqe.rawqp1.lflags |=
2781 					SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM;
2782 				break;
2783 			default:
2784 				break;
2785 			}
2786 			fallthrough;
2787 		case IB_WR_SEND_WITH_INV:
2788 			rc = bnxt_re_build_send_wqe(qp, wr, &wqe);
2789 			break;
2790 		case IB_WR_RDMA_WRITE:
2791 		case IB_WR_RDMA_WRITE_WITH_IMM:
2792 		case IB_WR_RDMA_READ:
2793 			rc = bnxt_re_build_rdma_wqe(wr, &wqe);
2794 			break;
2795 		case IB_WR_ATOMIC_CMP_AND_SWP:
2796 		case IB_WR_ATOMIC_FETCH_AND_ADD:
2797 			rc = bnxt_re_build_atomic_wqe(wr, &wqe);
2798 			break;
2799 		case IB_WR_RDMA_READ_WITH_INV:
2800 			ibdev_err(&qp->rdev->ibdev,
2801 				  "RDMA Read with Invalidate is not supported");
2802 			rc = -EINVAL;
2803 			goto bad;
2804 		case IB_WR_LOCAL_INV:
2805 			rc = bnxt_re_build_inv_wqe(wr, &wqe);
2806 			break;
2807 		case IB_WR_REG_MR:
2808 			rc = bnxt_re_build_reg_wqe(reg_wr(wr), &wqe);
2809 			break;
2810 		default:
2811 			/* Unsupported WRs */
2812 			ibdev_err(&qp->rdev->ibdev,
2813 				  "WR (%#x) is not supported", wr->opcode);
2814 			rc = -EINVAL;
2815 			goto bad;
2816 		}
2817 		if (!rc) {
2818 			if (!bnxt_qplib_is_chip_gen_p5_p7(qp->rdev->chip_ctx))
2819 				bnxt_re_legacy_set_uc_fence(&wqe);
2820 			rc = bnxt_qplib_post_send(&qp->qplib_qp, &wqe);
2821 		}
2822 bad:
2823 		if (rc) {
2824 			ibdev_err(&qp->rdev->ibdev,
2825 				  "post_send failed op:%#x qps = %#x rc = %d\n",
2826 				  wr->opcode, qp->qplib_qp.state, rc);
2827 			*bad_wr = wr;
2828 			break;
2829 		}
2830 		wr = wr->next;
2831 	}
2832 	bnxt_qplib_post_send_db(&qp->qplib_qp);
2833 	if (!bnxt_qplib_is_chip_gen_p5_p7(qp->rdev->chip_ctx))
2834 		bnxt_ud_qp_hw_stall_workaround(qp);
2835 	spin_unlock_irqrestore(&qp->sq_lock, flags);
2836 
2837 	return rc;
2838 }
2839 
bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev * rdev,struct bnxt_re_qp * qp,const struct ib_recv_wr * wr)2840 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev *rdev,
2841 				       struct bnxt_re_qp *qp,
2842 				       const struct ib_recv_wr *wr)
2843 {
2844 	struct bnxt_qplib_swqe wqe;
2845 	int rc = 0;
2846 
2847 	while (wr) {
2848 		/* House keeping */
2849 		memset(&wqe, 0, sizeof(wqe));
2850 
2851 		/* Common */
2852 		wqe.num_sge = wr->num_sge;
2853 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2854 			ibdev_err(&rdev->ibdev,
2855 				  "Limit exceeded for Receive SGEs");
2856 			rc = -EINVAL;
2857 			break;
2858 		}
2859 		bnxt_re_build_sgl(wr->sg_list, wqe.sg_list, wr->num_sge);
2860 		wqe.wr_id = wr->wr_id;
2861 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2862 
2863 		rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2864 		if (rc)
2865 			break;
2866 
2867 		wr = wr->next;
2868 	}
2869 	if (!rc)
2870 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2871 	return rc;
2872 }
2873 
bnxt_re_post_recv(struct ib_qp * ib_qp,const struct ib_recv_wr * wr,const struct ib_recv_wr ** bad_wr)2874 int bnxt_re_post_recv(struct ib_qp *ib_qp, const struct ib_recv_wr *wr,
2875 		      const struct ib_recv_wr **bad_wr)
2876 {
2877 	struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp);
2878 	struct bnxt_qplib_swqe wqe;
2879 	int rc = 0, payload_sz = 0;
2880 	unsigned long flags;
2881 	u32 count = 0;
2882 
2883 	spin_lock_irqsave(&qp->rq_lock, flags);
2884 	while (wr) {
2885 		/* House keeping */
2886 		memset(&wqe, 0, sizeof(wqe));
2887 
2888 		/* Common */
2889 		wqe.num_sge = wr->num_sge;
2890 		if (wr->num_sge > qp->qplib_qp.rq.max_sge) {
2891 			ibdev_err(&qp->rdev->ibdev,
2892 				  "Limit exceeded for Receive SGEs");
2893 			rc = -EINVAL;
2894 			*bad_wr = wr;
2895 			break;
2896 		}
2897 
2898 		payload_sz = bnxt_re_build_sgl(wr->sg_list, wqe.sg_list,
2899 					       wr->num_sge);
2900 		wqe.wr_id = wr->wr_id;
2901 		wqe.type = BNXT_QPLIB_SWQE_TYPE_RECV;
2902 
2903 		if (ib_qp->qp_type == IB_QPT_GSI &&
2904 		    qp->qplib_qp.type != CMDQ_CREATE_QP_TYPE_GSI)
2905 			rc = bnxt_re_build_qp1_shadow_qp_recv(qp, wr, &wqe,
2906 							      payload_sz);
2907 		if (!rc)
2908 			rc = bnxt_qplib_post_recv(&qp->qplib_qp, &wqe);
2909 		if (rc) {
2910 			*bad_wr = wr;
2911 			break;
2912 		}
2913 
2914 		/* Ring DB if the RQEs posted reaches a threshold value */
2915 		if (++count >= BNXT_RE_RQ_WQE_THRESHOLD) {
2916 			bnxt_qplib_post_recv_db(&qp->qplib_qp);
2917 			count = 0;
2918 		}
2919 
2920 		wr = wr->next;
2921 	}
2922 
2923 	if (count)
2924 		bnxt_qplib_post_recv_db(&qp->qplib_qp);
2925 
2926 	spin_unlock_irqrestore(&qp->rq_lock, flags);
2927 
2928 	return rc;
2929 }
2930 
2931 /* Completion Queues */
bnxt_re_destroy_cq(struct ib_cq * ib_cq,struct ib_udata * udata)2932 int bnxt_re_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata)
2933 {
2934 	struct bnxt_re_cq *cq;
2935 	struct bnxt_qplib_nq *nq;
2936 	struct bnxt_re_dev *rdev;
2937 
2938 	cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
2939 	rdev = cq->rdev;
2940 	nq = cq->qplib_cq.nq;
2941 
2942 	bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
2943 	ib_umem_release(cq->umem);
2944 
2945 	atomic_dec(&rdev->stats.res.cq_count);
2946 	nq->budget--;
2947 	kfree(cq->cql);
2948 	return 0;
2949 }
2950 
bnxt_re_create_cq(struct ib_cq * ibcq,const struct ib_cq_init_attr * attr,struct ib_udata * udata)2951 int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
2952 		      struct ib_udata *udata)
2953 {
2954 	struct bnxt_re_cq *cq = container_of(ibcq, struct bnxt_re_cq, ib_cq);
2955 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibcq->device, ibdev);
2956 	struct bnxt_re_ucontext *uctx =
2957 		rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
2958 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
2959 	int rc, entries;
2960 	int cqe = attr->cqe;
2961 	struct bnxt_qplib_nq *nq = NULL;
2962 	unsigned int nq_alloc_cnt;
2963 	u32 active_cqs;
2964 
2965 	if (attr->flags)
2966 		return -EOPNOTSUPP;
2967 
2968 	/* Validate CQ fields */
2969 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
2970 		ibdev_err(&rdev->ibdev, "Failed to create CQ -max exceeded");
2971 		return -EINVAL;
2972 	}
2973 
2974 	cq->rdev = rdev;
2975 	cq->qplib_cq.cq_handle = (u64)(unsigned long)(&cq->qplib_cq);
2976 
2977 	entries = bnxt_re_init_depth(cqe + 1, uctx);
2978 	if (entries > dev_attr->max_cq_wqes + 1)
2979 		entries = dev_attr->max_cq_wqes + 1;
2980 
2981 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
2982 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
2983 	if (udata) {
2984 		struct bnxt_re_cq_req req;
2985 		if (ib_copy_from_udata(&req, udata, sizeof(req))) {
2986 			rc = -EFAULT;
2987 			goto fail;
2988 		}
2989 
2990 		cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va,
2991 				       entries * sizeof(struct cq_base),
2992 				       IB_ACCESS_LOCAL_WRITE);
2993 		if (IS_ERR(cq->umem)) {
2994 			rc = PTR_ERR(cq->umem);
2995 			goto fail;
2996 		}
2997 		cq->qplib_cq.sg_info.umem = cq->umem;
2998 		cq->qplib_cq.dpi = &uctx->dpi;
2999 	} else {
3000 		cq->max_cql = min_t(u32, entries, MAX_CQL_PER_POLL);
3001 		cq->cql = kcalloc(cq->max_cql, sizeof(struct bnxt_qplib_cqe),
3002 				  GFP_KERNEL);
3003 		if (!cq->cql) {
3004 			rc = -ENOMEM;
3005 			goto fail;
3006 		}
3007 
3008 		cq->qplib_cq.dpi = &rdev->dpi_privileged;
3009 	}
3010 	/*
3011 	 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
3012 	 * used for getting the NQ index.
3013 	 */
3014 	nq_alloc_cnt = atomic_inc_return(&rdev->nq_alloc_cnt);
3015 	nq = &rdev->nq[nq_alloc_cnt % (rdev->num_msix - 1)];
3016 	cq->qplib_cq.max_wqe = entries;
3017 	cq->qplib_cq.cnq_hw_ring_id = nq->ring_id;
3018 	cq->qplib_cq.nq	= nq;
3019 
3020 	rc = bnxt_qplib_create_cq(&rdev->qplib_res, &cq->qplib_cq);
3021 	if (rc) {
3022 		ibdev_err(&rdev->ibdev, "Failed to create HW CQ");
3023 		goto fail;
3024 	}
3025 
3026 	cq->ib_cq.cqe = entries;
3027 	cq->cq_period = cq->qplib_cq.period;
3028 	nq->budget++;
3029 
3030 	active_cqs = atomic_inc_return(&rdev->stats.res.cq_count);
3031 	if (active_cqs > rdev->stats.res.cq_watermark)
3032 		rdev->stats.res.cq_watermark = active_cqs;
3033 	spin_lock_init(&cq->cq_lock);
3034 
3035 	if (udata) {
3036 		struct bnxt_re_cq_resp resp;
3037 
3038 		resp.cqid = cq->qplib_cq.id;
3039 		resp.tail = cq->qplib_cq.hwq.cons;
3040 		resp.phase = cq->qplib_cq.period;
3041 		resp.rsvd = 0;
3042 		rc = ib_copy_to_udata(udata, &resp, sizeof(resp));
3043 		if (rc) {
3044 			ibdev_err(&rdev->ibdev, "Failed to copy CQ udata");
3045 			bnxt_qplib_destroy_cq(&rdev->qplib_res, &cq->qplib_cq);
3046 			goto c2fail;
3047 		}
3048 	}
3049 
3050 	return 0;
3051 
3052 c2fail:
3053 	ib_umem_release(cq->umem);
3054 fail:
3055 	kfree(cq->cql);
3056 	return rc;
3057 }
3058 
bnxt_re_resize_cq_complete(struct bnxt_re_cq * cq)3059 static void bnxt_re_resize_cq_complete(struct bnxt_re_cq *cq)
3060 {
3061 	struct bnxt_re_dev *rdev = cq->rdev;
3062 
3063 	bnxt_qplib_resize_cq_complete(&rdev->qplib_res, &cq->qplib_cq);
3064 
3065 	cq->qplib_cq.max_wqe = cq->resize_cqe;
3066 	if (cq->resize_umem) {
3067 		ib_umem_release(cq->umem);
3068 		cq->umem = cq->resize_umem;
3069 		cq->resize_umem = NULL;
3070 		cq->resize_cqe = 0;
3071 	}
3072 }
3073 
bnxt_re_resize_cq(struct ib_cq * ibcq,int cqe,struct ib_udata * udata)3074 int bnxt_re_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
3075 {
3076 	struct bnxt_qplib_sg_info sg_info = {};
3077 	struct bnxt_qplib_dpi *orig_dpi = NULL;
3078 	struct bnxt_qplib_dev_attr *dev_attr;
3079 	struct bnxt_re_ucontext *uctx = NULL;
3080 	struct bnxt_re_resize_cq_req req;
3081 	struct bnxt_re_dev *rdev;
3082 	struct bnxt_re_cq *cq;
3083 	int rc, entries;
3084 
3085 	cq =  container_of(ibcq, struct bnxt_re_cq, ib_cq);
3086 	rdev = cq->rdev;
3087 	dev_attr = &rdev->dev_attr;
3088 	if (!ibcq->uobject) {
3089 		ibdev_err(&rdev->ibdev, "Kernel CQ Resize not supported");
3090 		return -EOPNOTSUPP;
3091 	}
3092 
3093 	if (cq->resize_umem) {
3094 		ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - Busy",
3095 			  cq->qplib_cq.id);
3096 		return -EBUSY;
3097 	}
3098 
3099 	/* Check the requested cq depth out of supported depth */
3100 	if (cqe < 1 || cqe > dev_attr->max_cq_wqes) {
3101 		ibdev_err(&rdev->ibdev, "Resize CQ %#x failed - out of range cqe %d",
3102 			  cq->qplib_cq.id, cqe);
3103 		return -EINVAL;
3104 	}
3105 
3106 	uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx);
3107 	entries = bnxt_re_init_depth(cqe + 1, uctx);
3108 	if (entries > dev_attr->max_cq_wqes + 1)
3109 		entries = dev_attr->max_cq_wqes + 1;
3110 
3111 	/* uverbs consumer */
3112 	if (ib_copy_from_udata(&req, udata, sizeof(req))) {
3113 		rc = -EFAULT;
3114 		goto fail;
3115 	}
3116 
3117 	cq->resize_umem = ib_umem_get(&rdev->ibdev, req.cq_va,
3118 				      entries * sizeof(struct cq_base),
3119 				      IB_ACCESS_LOCAL_WRITE);
3120 	if (IS_ERR(cq->resize_umem)) {
3121 		rc = PTR_ERR(cq->resize_umem);
3122 		cq->resize_umem = NULL;
3123 		ibdev_err(&rdev->ibdev, "%s: ib_umem_get failed! rc = %d\n",
3124 			  __func__, rc);
3125 		goto fail;
3126 	}
3127 	cq->resize_cqe = entries;
3128 	memcpy(&sg_info, &cq->qplib_cq.sg_info, sizeof(sg_info));
3129 	orig_dpi = cq->qplib_cq.dpi;
3130 
3131 	cq->qplib_cq.sg_info.umem = cq->resize_umem;
3132 	cq->qplib_cq.sg_info.pgsize = PAGE_SIZE;
3133 	cq->qplib_cq.sg_info.pgshft = PAGE_SHIFT;
3134 	cq->qplib_cq.dpi = &uctx->dpi;
3135 
3136 	rc = bnxt_qplib_resize_cq(&rdev->qplib_res, &cq->qplib_cq, entries);
3137 	if (rc) {
3138 		ibdev_err(&rdev->ibdev, "Resize HW CQ %#x failed!",
3139 			  cq->qplib_cq.id);
3140 		goto fail;
3141 	}
3142 
3143 	cq->ib_cq.cqe = cq->resize_cqe;
3144 	atomic_inc(&rdev->stats.res.resize_count);
3145 
3146 	return 0;
3147 
3148 fail:
3149 	if (cq->resize_umem) {
3150 		ib_umem_release(cq->resize_umem);
3151 		cq->resize_umem = NULL;
3152 		cq->resize_cqe = 0;
3153 		memcpy(&cq->qplib_cq.sg_info, &sg_info, sizeof(sg_info));
3154 		cq->qplib_cq.dpi = orig_dpi;
3155 	}
3156 	return rc;
3157 }
3158 
__req_to_ib_wc_status(u8 qstatus)3159 static u8 __req_to_ib_wc_status(u8 qstatus)
3160 {
3161 	switch (qstatus) {
3162 	case CQ_REQ_STATUS_OK:
3163 		return IB_WC_SUCCESS;
3164 	case CQ_REQ_STATUS_BAD_RESPONSE_ERR:
3165 		return IB_WC_BAD_RESP_ERR;
3166 	case CQ_REQ_STATUS_LOCAL_LENGTH_ERR:
3167 		return IB_WC_LOC_LEN_ERR;
3168 	case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR:
3169 		return IB_WC_LOC_QP_OP_ERR;
3170 	case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR:
3171 		return IB_WC_LOC_PROT_ERR;
3172 	case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR:
3173 		return IB_WC_GENERAL_ERR;
3174 	case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR:
3175 		return IB_WC_REM_INV_REQ_ERR;
3176 	case CQ_REQ_STATUS_REMOTE_ACCESS_ERR:
3177 		return IB_WC_REM_ACCESS_ERR;
3178 	case CQ_REQ_STATUS_REMOTE_OPERATION_ERR:
3179 		return IB_WC_REM_OP_ERR;
3180 	case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR:
3181 		return IB_WC_RNR_RETRY_EXC_ERR;
3182 	case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR:
3183 		return IB_WC_RETRY_EXC_ERR;
3184 	case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR:
3185 		return IB_WC_WR_FLUSH_ERR;
3186 	default:
3187 		return IB_WC_GENERAL_ERR;
3188 	}
3189 	return 0;
3190 }
3191 
__rawqp1_to_ib_wc_status(u8 qstatus)3192 static u8 __rawqp1_to_ib_wc_status(u8 qstatus)
3193 {
3194 	switch (qstatus) {
3195 	case CQ_RES_RAWETH_QP1_STATUS_OK:
3196 		return IB_WC_SUCCESS;
3197 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR:
3198 		return IB_WC_LOC_ACCESS_ERR;
3199 	case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR:
3200 		return IB_WC_LOC_LEN_ERR;
3201 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR:
3202 		return IB_WC_LOC_PROT_ERR;
3203 	case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR:
3204 		return IB_WC_LOC_QP_OP_ERR;
3205 	case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR:
3206 		return IB_WC_GENERAL_ERR;
3207 	case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR:
3208 		return IB_WC_WR_FLUSH_ERR;
3209 	case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR:
3210 		return IB_WC_WR_FLUSH_ERR;
3211 	default:
3212 		return IB_WC_GENERAL_ERR;
3213 	}
3214 }
3215 
__rc_to_ib_wc_status(u8 qstatus)3216 static u8 __rc_to_ib_wc_status(u8 qstatus)
3217 {
3218 	switch (qstatus) {
3219 	case CQ_RES_RC_STATUS_OK:
3220 		return IB_WC_SUCCESS;
3221 	case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR:
3222 		return IB_WC_LOC_ACCESS_ERR;
3223 	case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR:
3224 		return IB_WC_LOC_LEN_ERR;
3225 	case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR:
3226 		return IB_WC_LOC_PROT_ERR;
3227 	case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR:
3228 		return IB_WC_LOC_QP_OP_ERR;
3229 	case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR:
3230 		return IB_WC_GENERAL_ERR;
3231 	case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR:
3232 		return IB_WC_REM_INV_REQ_ERR;
3233 	case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR:
3234 		return IB_WC_WR_FLUSH_ERR;
3235 	case CQ_RES_RC_STATUS_HW_FLUSH_ERR:
3236 		return IB_WC_WR_FLUSH_ERR;
3237 	default:
3238 		return IB_WC_GENERAL_ERR;
3239 	}
3240 }
3241 
bnxt_re_process_req_wc(struct ib_wc * wc,struct bnxt_qplib_cqe * cqe)3242 static void bnxt_re_process_req_wc(struct ib_wc *wc, struct bnxt_qplib_cqe *cqe)
3243 {
3244 	switch (cqe->type) {
3245 	case BNXT_QPLIB_SWQE_TYPE_SEND:
3246 		wc->opcode = IB_WC_SEND;
3247 		break;
3248 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM:
3249 		wc->opcode = IB_WC_SEND;
3250 		wc->wc_flags |= IB_WC_WITH_IMM;
3251 		break;
3252 	case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV:
3253 		wc->opcode = IB_WC_SEND;
3254 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3255 		break;
3256 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE:
3257 		wc->opcode = IB_WC_RDMA_WRITE;
3258 		break;
3259 	case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM:
3260 		wc->opcode = IB_WC_RDMA_WRITE;
3261 		wc->wc_flags |= IB_WC_WITH_IMM;
3262 		break;
3263 	case BNXT_QPLIB_SWQE_TYPE_RDMA_READ:
3264 		wc->opcode = IB_WC_RDMA_READ;
3265 		break;
3266 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP:
3267 		wc->opcode = IB_WC_COMP_SWAP;
3268 		break;
3269 	case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD:
3270 		wc->opcode = IB_WC_FETCH_ADD;
3271 		break;
3272 	case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV:
3273 		wc->opcode = IB_WC_LOCAL_INV;
3274 		break;
3275 	case BNXT_QPLIB_SWQE_TYPE_REG_MR:
3276 		wc->opcode = IB_WC_REG_MR;
3277 		break;
3278 	default:
3279 		wc->opcode = IB_WC_SEND;
3280 		break;
3281 	}
3282 
3283 	wc->status = __req_to_ib_wc_status(cqe->status);
3284 }
3285 
bnxt_re_check_packet_type(u16 raweth_qp1_flags,u16 raweth_qp1_flags2)3286 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags,
3287 				     u16 raweth_qp1_flags2)
3288 {
3289 	bool is_ipv6 = false, is_ipv4 = false;
3290 
3291 	/* raweth_qp1_flags Bit 9-6 indicates itype */
3292 	if ((raweth_qp1_flags & CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3293 	    != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE)
3294 		return -1;
3295 
3296 	if (raweth_qp1_flags2 &
3297 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC &&
3298 	    raweth_qp1_flags2 &
3299 	    CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC) {
3300 		/* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
3301 		(raweth_qp1_flags2 &
3302 		 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE) ?
3303 			(is_ipv6 = true) : (is_ipv4 = true);
3304 		return ((is_ipv6) ?
3305 			 BNXT_RE_ROCEV2_IPV6_PACKET :
3306 			 BNXT_RE_ROCEV2_IPV4_PACKET);
3307 	} else {
3308 		return BNXT_RE_ROCE_V1_PACKET;
3309 	}
3310 }
3311 
bnxt_re_to_ib_nw_type(int nw_type)3312 static int bnxt_re_to_ib_nw_type(int nw_type)
3313 {
3314 	u8 nw_hdr_type = 0xFF;
3315 
3316 	switch (nw_type) {
3317 	case BNXT_RE_ROCE_V1_PACKET:
3318 		nw_hdr_type = RDMA_NETWORK_ROCE_V1;
3319 		break;
3320 	case BNXT_RE_ROCEV2_IPV4_PACKET:
3321 		nw_hdr_type = RDMA_NETWORK_IPV4;
3322 		break;
3323 	case BNXT_RE_ROCEV2_IPV6_PACKET:
3324 		nw_hdr_type = RDMA_NETWORK_IPV6;
3325 		break;
3326 	}
3327 	return nw_hdr_type;
3328 }
3329 
bnxt_re_is_loopback_packet(struct bnxt_re_dev * rdev,void * rq_hdr_buf)3330 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev *rdev,
3331 				       void *rq_hdr_buf)
3332 {
3333 	u8 *tmp_buf = NULL;
3334 	struct ethhdr *eth_hdr;
3335 	u16 eth_type;
3336 	bool rc = false;
3337 
3338 	tmp_buf = (u8 *)rq_hdr_buf;
3339 	/*
3340 	 * If dest mac is not same as I/F mac, this could be a
3341 	 * loopback address or multicast address, check whether
3342 	 * it is a loopback packet
3343 	 */
3344 	if (!ether_addr_equal(tmp_buf, rdev->netdev->dev_addr)) {
3345 		tmp_buf += 4;
3346 		/* Check the  ether type */
3347 		eth_hdr = (struct ethhdr *)tmp_buf;
3348 		eth_type = ntohs(eth_hdr->h_proto);
3349 		switch (eth_type) {
3350 		case ETH_P_IBOE:
3351 			rc = true;
3352 			break;
3353 		case ETH_P_IP:
3354 		case ETH_P_IPV6: {
3355 			u32 len;
3356 			struct udphdr *udp_hdr;
3357 
3358 			len = (eth_type == ETH_P_IP ? sizeof(struct iphdr) :
3359 						      sizeof(struct ipv6hdr));
3360 			tmp_buf += sizeof(struct ethhdr) + len;
3361 			udp_hdr = (struct udphdr *)tmp_buf;
3362 			if (ntohs(udp_hdr->dest) ==
3363 				    ROCE_V2_UDP_DPORT)
3364 				rc = true;
3365 			break;
3366 			}
3367 		default:
3368 			break;
3369 		}
3370 	}
3371 
3372 	return rc;
3373 }
3374 
bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp * gsi_qp,struct bnxt_qplib_cqe * cqe)3375 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp *gsi_qp,
3376 					 struct bnxt_qplib_cqe *cqe)
3377 {
3378 	struct bnxt_re_dev *rdev = gsi_qp->rdev;
3379 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3380 	struct bnxt_re_qp *gsi_sqp = rdev->gsi_ctx.gsi_sqp;
3381 	dma_addr_t shrq_hdr_buf_map;
3382 	struct ib_sge s_sge[2] = {};
3383 	struct ib_sge r_sge[2] = {};
3384 	struct bnxt_re_ah *gsi_sah;
3385 	struct ib_recv_wr rwr = {};
3386 	dma_addr_t rq_hdr_buf_map;
3387 	struct ib_ud_wr udwr = {};
3388 	struct ib_send_wr *swr;
3389 	u32 skip_bytes = 0;
3390 	int pkt_type = 0;
3391 	void *rq_hdr_buf;
3392 	u32 offset = 0;
3393 	u32 tbl_idx;
3394 	int rc;
3395 
3396 	swr = &udwr.wr;
3397 	tbl_idx = cqe->wr_id;
3398 
3399 	rq_hdr_buf = gsi_qp->qplib_qp.rq_hdr_buf +
3400 			(tbl_idx * gsi_qp->qplib_qp.rq_hdr_buf_size);
3401 	rq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3402 							  tbl_idx);
3403 
3404 	/* Shadow QP header buffer */
3405 	shrq_hdr_buf_map = bnxt_qplib_get_qp_buf_from_index(&gsi_qp->qplib_qp,
3406 							    tbl_idx);
3407 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3408 
3409 	/* Store this cqe */
3410 	memcpy(&sqp_entry->cqe, cqe, sizeof(struct bnxt_qplib_cqe));
3411 	sqp_entry->qp1_qp = gsi_qp;
3412 
3413 	/* Find packet type from the cqe */
3414 
3415 	pkt_type = bnxt_re_check_packet_type(cqe->raweth_qp1_flags,
3416 					     cqe->raweth_qp1_flags2);
3417 	if (pkt_type < 0) {
3418 		ibdev_err(&rdev->ibdev, "Invalid packet\n");
3419 		return -EINVAL;
3420 	}
3421 
3422 	/* Adjust the offset for the user buffer and post in the rq */
3423 
3424 	if (pkt_type == BNXT_RE_ROCEV2_IPV4_PACKET)
3425 		offset = 20;
3426 
3427 	/*
3428 	 * QP1 loopback packet has 4 bytes of internal header before
3429 	 * ether header. Skip these four bytes.
3430 	 */
3431 	if (bnxt_re_is_loopback_packet(rdev, rq_hdr_buf))
3432 		skip_bytes = 4;
3433 
3434 	/* First send SGE . Skip the ether header*/
3435 	s_sge[0].addr = rq_hdr_buf_map + BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
3436 			+ skip_bytes;
3437 	s_sge[0].lkey = 0xFFFFFFFF;
3438 	s_sge[0].length = offset ? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4 :
3439 				BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6;
3440 
3441 	/* Second Send SGE */
3442 	s_sge[1].addr = s_sge[0].addr + s_sge[0].length +
3443 			BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE;
3444 	if (pkt_type != BNXT_RE_ROCE_V1_PACKET)
3445 		s_sge[1].addr += 8;
3446 	s_sge[1].lkey = 0xFFFFFFFF;
3447 	s_sge[1].length = 256;
3448 
3449 	/* First recv SGE */
3450 
3451 	r_sge[0].addr = shrq_hdr_buf_map;
3452 	r_sge[0].lkey = 0xFFFFFFFF;
3453 	r_sge[0].length = 40;
3454 
3455 	r_sge[1].addr = sqp_entry->sge.addr + offset;
3456 	r_sge[1].lkey = sqp_entry->sge.lkey;
3457 	r_sge[1].length = BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6 + 256 - offset;
3458 
3459 	/* Create receive work request */
3460 	rwr.num_sge = 2;
3461 	rwr.sg_list = r_sge;
3462 	rwr.wr_id = tbl_idx;
3463 	rwr.next = NULL;
3464 
3465 	rc = bnxt_re_post_recv_shadow_qp(rdev, gsi_sqp, &rwr);
3466 	if (rc) {
3467 		ibdev_err(&rdev->ibdev,
3468 			  "Failed to post Rx buffers to shadow QP");
3469 		return -ENOMEM;
3470 	}
3471 
3472 	swr->num_sge = 2;
3473 	swr->sg_list = s_sge;
3474 	swr->wr_id = tbl_idx;
3475 	swr->opcode = IB_WR_SEND;
3476 	swr->next = NULL;
3477 	gsi_sah = rdev->gsi_ctx.gsi_sah;
3478 	udwr.ah = &gsi_sah->ib_ah;
3479 	udwr.remote_qpn = gsi_sqp->qplib_qp.id;
3480 	udwr.remote_qkey = gsi_sqp->qplib_qp.qkey;
3481 
3482 	/* post data received  in the send queue */
3483 	return bnxt_re_post_send_shadow_qp(rdev, gsi_sqp, swr);
3484 }
3485 
bnxt_re_process_res_rawqp1_wc(struct ib_wc * wc,struct bnxt_qplib_cqe * cqe)3486 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc *wc,
3487 					  struct bnxt_qplib_cqe *cqe)
3488 {
3489 	wc->opcode = IB_WC_RECV;
3490 	wc->status = __rawqp1_to_ib_wc_status(cqe->status);
3491 	wc->wc_flags |= IB_WC_GRH;
3492 }
3493 
bnxt_re_check_if_vlan_valid(struct bnxt_re_dev * rdev,u16 vlan_id)3494 static bool bnxt_re_check_if_vlan_valid(struct bnxt_re_dev *rdev,
3495 					u16 vlan_id)
3496 {
3497 	/*
3498 	 * Check if the vlan is configured in the host.  If not configured, it
3499 	 * can be a transparent VLAN. So dont report the vlan id.
3500 	 */
3501 	if (!__vlan_find_dev_deep_rcu(rdev->netdev,
3502 				      htons(ETH_P_8021Q), vlan_id))
3503 		return false;
3504 	return true;
3505 }
3506 
bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe * orig_cqe,u16 * vid,u8 * sl)3507 static bool bnxt_re_is_vlan_pkt(struct bnxt_qplib_cqe *orig_cqe,
3508 				u16 *vid, u8 *sl)
3509 {
3510 	bool ret = false;
3511 	u32 metadata;
3512 	u16 tpid;
3513 
3514 	metadata = orig_cqe->raweth_qp1_metadata;
3515 	if (orig_cqe->raweth_qp1_flags2 &
3516 		CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_META_FORMAT_VLAN) {
3517 		tpid = ((metadata &
3518 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_MASK) >>
3519 			 CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_TPID_SFT);
3520 		if (tpid == ETH_P_8021Q) {
3521 			*vid = metadata &
3522 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_VID_MASK;
3523 			*sl = (metadata &
3524 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_MASK) >>
3525 			       CQ_RES_RAWETH_QP1_RAWETH_QP1_METADATA_PRI_SFT;
3526 			ret = true;
3527 		}
3528 	}
3529 
3530 	return ret;
3531 }
3532 
bnxt_re_process_res_rc_wc(struct ib_wc * wc,struct bnxt_qplib_cqe * cqe)3533 static void bnxt_re_process_res_rc_wc(struct ib_wc *wc,
3534 				      struct bnxt_qplib_cqe *cqe)
3535 {
3536 	wc->opcode = IB_WC_RECV;
3537 	wc->status = __rc_to_ib_wc_status(cqe->status);
3538 
3539 	if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3540 		wc->wc_flags |= IB_WC_WITH_IMM;
3541 	if (cqe->flags & CQ_RES_RC_FLAGS_INV)
3542 		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
3543 	if ((cqe->flags & (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM)) ==
3544 	    (CQ_RES_RC_FLAGS_RDMA | CQ_RES_RC_FLAGS_IMM))
3545 		wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
3546 }
3547 
bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp * gsi_sqp,struct ib_wc * wc,struct bnxt_qplib_cqe * cqe)3548 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp *gsi_sqp,
3549 					     struct ib_wc *wc,
3550 					     struct bnxt_qplib_cqe *cqe)
3551 {
3552 	struct bnxt_re_dev *rdev = gsi_sqp->rdev;
3553 	struct bnxt_re_qp *gsi_qp = NULL;
3554 	struct bnxt_qplib_cqe *orig_cqe = NULL;
3555 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3556 	int nw_type;
3557 	u32 tbl_idx;
3558 	u16 vlan_id;
3559 	u8 sl;
3560 
3561 	tbl_idx = cqe->wr_id;
3562 
3563 	sqp_entry = &rdev->gsi_ctx.sqp_tbl[tbl_idx];
3564 	gsi_qp = sqp_entry->qp1_qp;
3565 	orig_cqe = &sqp_entry->cqe;
3566 
3567 	wc->wr_id = sqp_entry->wrid;
3568 	wc->byte_len = orig_cqe->length;
3569 	wc->qp = &gsi_qp->ib_qp;
3570 
3571 	wc->ex.imm_data = cpu_to_be32(orig_cqe->immdata);
3572 	wc->src_qp = orig_cqe->src_qp;
3573 	memcpy(wc->smac, orig_cqe->smac, ETH_ALEN);
3574 	if (bnxt_re_is_vlan_pkt(orig_cqe, &vlan_id, &sl)) {
3575 		if (bnxt_re_check_if_vlan_valid(rdev, vlan_id)) {
3576 			wc->vlan_id = vlan_id;
3577 			wc->sl = sl;
3578 			wc->wc_flags |= IB_WC_WITH_VLAN;
3579 		}
3580 	}
3581 	wc->port_num = 1;
3582 	wc->vendor_err = orig_cqe->status;
3583 
3584 	wc->opcode = IB_WC_RECV;
3585 	wc->status = __rawqp1_to_ib_wc_status(orig_cqe->status);
3586 	wc->wc_flags |= IB_WC_GRH;
3587 
3588 	nw_type = bnxt_re_check_packet_type(orig_cqe->raweth_qp1_flags,
3589 					    orig_cqe->raweth_qp1_flags2);
3590 	if (nw_type >= 0) {
3591 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3592 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3593 	}
3594 }
3595 
bnxt_re_process_res_ud_wc(struct bnxt_re_qp * qp,struct ib_wc * wc,struct bnxt_qplib_cqe * cqe)3596 static void bnxt_re_process_res_ud_wc(struct bnxt_re_qp *qp,
3597 				      struct ib_wc *wc,
3598 				      struct bnxt_qplib_cqe *cqe)
3599 {
3600 	struct bnxt_re_dev *rdev;
3601 	u16 vlan_id = 0;
3602 	u8 nw_type;
3603 
3604 	rdev = qp->rdev;
3605 	wc->opcode = IB_WC_RECV;
3606 	wc->status = __rc_to_ib_wc_status(cqe->status);
3607 
3608 	if (cqe->flags & CQ_RES_UD_FLAGS_IMM)
3609 		wc->wc_flags |= IB_WC_WITH_IMM;
3610 	/* report only on GSI QP for Thor */
3611 	if (qp->qplib_qp.type == CMDQ_CREATE_QP_TYPE_GSI) {
3612 		wc->wc_flags |= IB_WC_GRH;
3613 		memcpy(wc->smac, cqe->smac, ETH_ALEN);
3614 		wc->wc_flags |= IB_WC_WITH_SMAC;
3615 		if (cqe->flags & CQ_RES_UD_FLAGS_META_FORMAT_VLAN) {
3616 			vlan_id = (cqe->cfa_meta & 0xFFF);
3617 		}
3618 		/* Mark only if vlan_id is non zero */
3619 		if (vlan_id && bnxt_re_check_if_vlan_valid(rdev, vlan_id)) {
3620 			wc->vlan_id = vlan_id;
3621 			wc->wc_flags |= IB_WC_WITH_VLAN;
3622 		}
3623 		nw_type = (cqe->flags & CQ_RES_UD_FLAGS_ROCE_IP_VER_MASK) >>
3624 			   CQ_RES_UD_FLAGS_ROCE_IP_VER_SFT;
3625 		wc->network_hdr_type = bnxt_re_to_ib_nw_type(nw_type);
3626 		wc->wc_flags |= IB_WC_WITH_NETWORK_HDR_TYPE;
3627 	}
3628 
3629 }
3630 
send_phantom_wqe(struct bnxt_re_qp * qp)3631 static int send_phantom_wqe(struct bnxt_re_qp *qp)
3632 {
3633 	struct bnxt_qplib_qp *lib_qp = &qp->qplib_qp;
3634 	unsigned long flags;
3635 	int rc;
3636 
3637 	spin_lock_irqsave(&qp->sq_lock, flags);
3638 
3639 	rc = bnxt_re_bind_fence_mw(lib_qp);
3640 	if (!rc) {
3641 		lib_qp->sq.phantom_wqe_cnt++;
3642 		ibdev_dbg(&qp->rdev->ibdev,
3643 			  "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
3644 			  lib_qp->id, lib_qp->sq.hwq.prod,
3645 			  HWQ_CMP(lib_qp->sq.hwq.prod, &lib_qp->sq.hwq),
3646 			  lib_qp->sq.phantom_wqe_cnt);
3647 	}
3648 
3649 	spin_unlock_irqrestore(&qp->sq_lock, flags);
3650 	return rc;
3651 }
3652 
bnxt_re_poll_cq(struct ib_cq * ib_cq,int num_entries,struct ib_wc * wc)3653 int bnxt_re_poll_cq(struct ib_cq *ib_cq, int num_entries, struct ib_wc *wc)
3654 {
3655 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3656 	struct bnxt_re_qp *qp, *sh_qp;
3657 	struct bnxt_qplib_cqe *cqe;
3658 	int i, ncqe, budget;
3659 	struct bnxt_qplib_q *sq;
3660 	struct bnxt_qplib_qp *lib_qp;
3661 	u32 tbl_idx;
3662 	struct bnxt_re_sqp_entries *sqp_entry = NULL;
3663 	unsigned long flags;
3664 
3665 	/* User CQ; the only processing we do is to
3666 	 * complete any pending CQ resize operation.
3667 	 */
3668 	if (cq->umem) {
3669 		if (cq->resize_umem)
3670 			bnxt_re_resize_cq_complete(cq);
3671 		return 0;
3672 	}
3673 
3674 	spin_lock_irqsave(&cq->cq_lock, flags);
3675 	budget = min_t(u32, num_entries, cq->max_cql);
3676 	num_entries = budget;
3677 	if (!cq->cql) {
3678 		ibdev_err(&cq->rdev->ibdev, "POLL CQ : no CQL to use");
3679 		goto exit;
3680 	}
3681 	cqe = &cq->cql[0];
3682 	while (budget) {
3683 		lib_qp = NULL;
3684 		ncqe = bnxt_qplib_poll_cq(&cq->qplib_cq, cqe, budget, &lib_qp);
3685 		if (lib_qp) {
3686 			sq = &lib_qp->sq;
3687 			if (sq->send_phantom) {
3688 				qp = container_of(lib_qp,
3689 						  struct bnxt_re_qp, qplib_qp);
3690 				if (send_phantom_wqe(qp) == -ENOMEM)
3691 					ibdev_err(&cq->rdev->ibdev,
3692 						  "Phantom failed! Scheduled to send again\n");
3693 				else
3694 					sq->send_phantom = false;
3695 			}
3696 		}
3697 		if (ncqe < budget)
3698 			ncqe += bnxt_qplib_process_flush_list(&cq->qplib_cq,
3699 							      cqe + ncqe,
3700 							      budget - ncqe);
3701 
3702 		if (!ncqe)
3703 			break;
3704 
3705 		for (i = 0; i < ncqe; i++, cqe++) {
3706 			/* Transcribe each qplib_wqe back to ib_wc */
3707 			memset(wc, 0, sizeof(*wc));
3708 
3709 			wc->wr_id = cqe->wr_id;
3710 			wc->byte_len = cqe->length;
3711 			qp = container_of
3712 				((struct bnxt_qplib_qp *)
3713 				 (unsigned long)(cqe->qp_handle),
3714 				 struct bnxt_re_qp, qplib_qp);
3715 			wc->qp = &qp->ib_qp;
3716 			if (cqe->flags & CQ_RES_RC_FLAGS_IMM)
3717 				wc->ex.imm_data = cpu_to_be32(cqe->immdata);
3718 			else
3719 				wc->ex.invalidate_rkey = cqe->invrkey;
3720 			wc->src_qp = cqe->src_qp;
3721 			memcpy(wc->smac, cqe->smac, ETH_ALEN);
3722 			wc->port_num = 1;
3723 			wc->vendor_err = cqe->status;
3724 
3725 			switch (cqe->opcode) {
3726 			case CQ_BASE_CQE_TYPE_REQ:
3727 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3728 				if (sh_qp &&
3729 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3730 					/* Handle this completion with
3731 					 * the stored completion
3732 					 */
3733 					memset(wc, 0, sizeof(*wc));
3734 					continue;
3735 				}
3736 				bnxt_re_process_req_wc(wc, cqe);
3737 				break;
3738 			case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1:
3739 				if (!cqe->status) {
3740 					int rc = 0;
3741 
3742 					rc = bnxt_re_process_raw_qp_pkt_rx
3743 								(qp, cqe);
3744 					if (!rc) {
3745 						memset(wc, 0, sizeof(*wc));
3746 						continue;
3747 					}
3748 					cqe->status = -1;
3749 				}
3750 				/* Errors need not be looped back.
3751 				 * But change the wr_id to the one
3752 				 * stored in the table
3753 				 */
3754 				tbl_idx = cqe->wr_id;
3755 				sqp_entry = &cq->rdev->gsi_ctx.sqp_tbl[tbl_idx];
3756 				wc->wr_id = sqp_entry->wrid;
3757 				bnxt_re_process_res_rawqp1_wc(wc, cqe);
3758 				break;
3759 			case CQ_BASE_CQE_TYPE_RES_RC:
3760 				bnxt_re_process_res_rc_wc(wc, cqe);
3761 				break;
3762 			case CQ_BASE_CQE_TYPE_RES_UD:
3763 				sh_qp = qp->rdev->gsi_ctx.gsi_sqp;
3764 				if (sh_qp &&
3765 				    qp->qplib_qp.id == sh_qp->qplib_qp.id) {
3766 					/* Handle this completion with
3767 					 * the stored completion
3768 					 */
3769 					if (cqe->status) {
3770 						continue;
3771 					} else {
3772 						bnxt_re_process_res_shadow_qp_wc
3773 								(qp, wc, cqe);
3774 						break;
3775 					}
3776 				}
3777 				bnxt_re_process_res_ud_wc(qp, wc, cqe);
3778 				break;
3779 			default:
3780 				ibdev_err(&cq->rdev->ibdev,
3781 					  "POLL CQ : type 0x%x not handled",
3782 					  cqe->opcode);
3783 				continue;
3784 			}
3785 			wc++;
3786 			budget--;
3787 		}
3788 	}
3789 exit:
3790 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3791 	return num_entries - budget;
3792 }
3793 
bnxt_re_req_notify_cq(struct ib_cq * ib_cq,enum ib_cq_notify_flags ib_cqn_flags)3794 int bnxt_re_req_notify_cq(struct ib_cq *ib_cq,
3795 			  enum ib_cq_notify_flags ib_cqn_flags)
3796 {
3797 	struct bnxt_re_cq *cq = container_of(ib_cq, struct bnxt_re_cq, ib_cq);
3798 	int type = 0, rc = 0;
3799 	unsigned long flags;
3800 
3801 	spin_lock_irqsave(&cq->cq_lock, flags);
3802 	/* Trigger on the very next completion */
3803 	if (ib_cqn_flags & IB_CQ_NEXT_COMP)
3804 		type = DBC_DBC_TYPE_CQ_ARMALL;
3805 	/* Trigger on the next solicited completion */
3806 	else if (ib_cqn_flags & IB_CQ_SOLICITED)
3807 		type = DBC_DBC_TYPE_CQ_ARMSE;
3808 
3809 	/* Poll to see if there are missed events */
3810 	if ((ib_cqn_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
3811 	    !(bnxt_qplib_is_cq_empty(&cq->qplib_cq))) {
3812 		rc = 1;
3813 		goto exit;
3814 	}
3815 	bnxt_qplib_req_notify_cq(&cq->qplib_cq, type);
3816 
3817 exit:
3818 	spin_unlock_irqrestore(&cq->cq_lock, flags);
3819 	return rc;
3820 }
3821 
3822 /* Memory Regions */
bnxt_re_get_dma_mr(struct ib_pd * ib_pd,int mr_access_flags)3823 struct ib_mr *bnxt_re_get_dma_mr(struct ib_pd *ib_pd, int mr_access_flags)
3824 {
3825 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3826 	struct bnxt_re_dev *rdev = pd->rdev;
3827 	struct bnxt_re_mr *mr;
3828 	u32 active_mrs;
3829 	int rc;
3830 
3831 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3832 	if (!mr)
3833 		return ERR_PTR(-ENOMEM);
3834 
3835 	mr->rdev = rdev;
3836 	mr->qplib_mr.pd = &pd->qplib_pd;
3837 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
3838 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3839 
3840 	/* Allocate and register 0 as the address */
3841 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3842 	if (rc)
3843 		goto fail;
3844 
3845 	mr->qplib_mr.hwq.level = PBL_LVL_MAX;
3846 	mr->qplib_mr.total_size = -1; /* Infinte length */
3847 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, NULL, 0,
3848 			       PAGE_SIZE);
3849 	if (rc)
3850 		goto fail_mr;
3851 
3852 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3853 	if (mr_access_flags & (IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ |
3854 			       IB_ACCESS_REMOTE_ATOMIC))
3855 		mr->ib_mr.rkey = mr->ib_mr.lkey;
3856 	active_mrs = atomic_inc_return(&rdev->stats.res.mr_count);
3857 	if (active_mrs > rdev->stats.res.mr_watermark)
3858 		rdev->stats.res.mr_watermark = active_mrs;
3859 
3860 	return &mr->ib_mr;
3861 
3862 fail_mr:
3863 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3864 fail:
3865 	kfree(mr);
3866 	return ERR_PTR(rc);
3867 }
3868 
bnxt_re_dereg_mr(struct ib_mr * ib_mr,struct ib_udata * udata)3869 int bnxt_re_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3870 {
3871 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3872 	struct bnxt_re_dev *rdev = mr->rdev;
3873 	int rc;
3874 
3875 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3876 	if (rc) {
3877 		ibdev_err(&rdev->ibdev, "Dereg MR failed: %#x\n", rc);
3878 		return rc;
3879 	}
3880 
3881 	if (mr->pages) {
3882 		rc = bnxt_qplib_free_fast_reg_page_list(&rdev->qplib_res,
3883 							&mr->qplib_frpl);
3884 		kfree(mr->pages);
3885 		mr->npages = 0;
3886 		mr->pages = NULL;
3887 	}
3888 	ib_umem_release(mr->ib_umem);
3889 
3890 	kfree(mr);
3891 	atomic_dec(&rdev->stats.res.mr_count);
3892 	return rc;
3893 }
3894 
bnxt_re_set_page(struct ib_mr * ib_mr,u64 addr)3895 static int bnxt_re_set_page(struct ib_mr *ib_mr, u64 addr)
3896 {
3897 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3898 
3899 	if (unlikely(mr->npages == mr->qplib_frpl.max_pg_ptrs))
3900 		return -ENOMEM;
3901 
3902 	mr->pages[mr->npages++] = addr;
3903 	return 0;
3904 }
3905 
bnxt_re_map_mr_sg(struct ib_mr * ib_mr,struct scatterlist * sg,int sg_nents,unsigned int * sg_offset)3906 int bnxt_re_map_mr_sg(struct ib_mr *ib_mr, struct scatterlist *sg, int sg_nents,
3907 		      unsigned int *sg_offset)
3908 {
3909 	struct bnxt_re_mr *mr = container_of(ib_mr, struct bnxt_re_mr, ib_mr);
3910 
3911 	mr->npages = 0;
3912 	return ib_sg_to_pages(ib_mr, sg, sg_nents, sg_offset, bnxt_re_set_page);
3913 }
3914 
bnxt_re_alloc_mr(struct ib_pd * ib_pd,enum ib_mr_type type,u32 max_num_sg)3915 struct ib_mr *bnxt_re_alloc_mr(struct ib_pd *ib_pd, enum ib_mr_type type,
3916 			       u32 max_num_sg)
3917 {
3918 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3919 	struct bnxt_re_dev *rdev = pd->rdev;
3920 	struct bnxt_re_mr *mr = NULL;
3921 	u32 active_mrs;
3922 	int rc;
3923 
3924 	if (type != IB_MR_TYPE_MEM_REG) {
3925 		ibdev_dbg(&rdev->ibdev, "MR type 0x%x not supported", type);
3926 		return ERR_PTR(-EINVAL);
3927 	}
3928 	if (max_num_sg > MAX_PBL_LVL_1_PGS)
3929 		return ERR_PTR(-EINVAL);
3930 
3931 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3932 	if (!mr)
3933 		return ERR_PTR(-ENOMEM);
3934 
3935 	mr->rdev = rdev;
3936 	mr->qplib_mr.pd = &pd->qplib_pd;
3937 	mr->qplib_mr.flags = BNXT_QPLIB_FR_PMR;
3938 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR;
3939 
3940 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
3941 	if (rc)
3942 		goto bail;
3943 
3944 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
3945 	mr->ib_mr.rkey = mr->ib_mr.lkey;
3946 
3947 	mr->pages = kcalloc(max_num_sg, sizeof(u64), GFP_KERNEL);
3948 	if (!mr->pages) {
3949 		rc = -ENOMEM;
3950 		goto fail;
3951 	}
3952 	rc = bnxt_qplib_alloc_fast_reg_page_list(&rdev->qplib_res,
3953 						 &mr->qplib_frpl, max_num_sg);
3954 	if (rc) {
3955 		ibdev_err(&rdev->ibdev,
3956 			  "Failed to allocate HW FR page list");
3957 		goto fail_mr;
3958 	}
3959 
3960 	active_mrs = atomic_inc_return(&rdev->stats.res.mr_count);
3961 	if (active_mrs > rdev->stats.res.mr_watermark)
3962 		rdev->stats.res.mr_watermark = active_mrs;
3963 	return &mr->ib_mr;
3964 
3965 fail_mr:
3966 	kfree(mr->pages);
3967 fail:
3968 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
3969 bail:
3970 	kfree(mr);
3971 	return ERR_PTR(rc);
3972 }
3973 
bnxt_re_alloc_mw(struct ib_pd * ib_pd,enum ib_mw_type type,struct ib_udata * udata)3974 struct ib_mw *bnxt_re_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
3975 			       struct ib_udata *udata)
3976 {
3977 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
3978 	struct bnxt_re_dev *rdev = pd->rdev;
3979 	struct bnxt_re_mw *mw;
3980 	u32 active_mws;
3981 	int rc;
3982 
3983 	mw = kzalloc(sizeof(*mw), GFP_KERNEL);
3984 	if (!mw)
3985 		return ERR_PTR(-ENOMEM);
3986 	mw->rdev = rdev;
3987 	mw->qplib_mw.pd = &pd->qplib_pd;
3988 
3989 	mw->qplib_mw.type = (type == IB_MW_TYPE_1 ?
3990 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1 :
3991 			       CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B);
3992 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mw->qplib_mw);
3993 	if (rc) {
3994 		ibdev_err(&rdev->ibdev, "Allocate MW failed!");
3995 		goto fail;
3996 	}
3997 	mw->ib_mw.rkey = mw->qplib_mw.rkey;
3998 
3999 	active_mws = atomic_inc_return(&rdev->stats.res.mw_count);
4000 	if (active_mws > rdev->stats.res.mw_watermark)
4001 		rdev->stats.res.mw_watermark = active_mws;
4002 	return &mw->ib_mw;
4003 
4004 fail:
4005 	kfree(mw);
4006 	return ERR_PTR(rc);
4007 }
4008 
bnxt_re_dealloc_mw(struct ib_mw * ib_mw)4009 int bnxt_re_dealloc_mw(struct ib_mw *ib_mw)
4010 {
4011 	struct bnxt_re_mw *mw = container_of(ib_mw, struct bnxt_re_mw, ib_mw);
4012 	struct bnxt_re_dev *rdev = mw->rdev;
4013 	int rc;
4014 
4015 	rc = bnxt_qplib_free_mrw(&rdev->qplib_res, &mw->qplib_mw);
4016 	if (rc) {
4017 		ibdev_err(&rdev->ibdev, "Free MW failed: %#x\n", rc);
4018 		return rc;
4019 	}
4020 
4021 	kfree(mw);
4022 	atomic_dec(&rdev->stats.res.mw_count);
4023 	return rc;
4024 }
4025 
__bnxt_re_user_reg_mr(struct ib_pd * ib_pd,u64 length,u64 virt_addr,int mr_access_flags,struct ib_umem * umem)4026 static struct ib_mr *__bnxt_re_user_reg_mr(struct ib_pd *ib_pd, u64 length, u64 virt_addr,
4027 					   int mr_access_flags, struct ib_umem *umem)
4028 {
4029 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
4030 	struct bnxt_re_dev *rdev = pd->rdev;
4031 	unsigned long page_size;
4032 	struct bnxt_re_mr *mr;
4033 	int umem_pgs, rc;
4034 	u32 active_mrs;
4035 
4036 	if (length > BNXT_RE_MAX_MR_SIZE) {
4037 		ibdev_err(&rdev->ibdev, "MR Size: %lld > Max supported:%lld\n",
4038 			  length, BNXT_RE_MAX_MR_SIZE);
4039 		return ERR_PTR(-ENOMEM);
4040 	}
4041 
4042 	page_size = ib_umem_find_best_pgsz(umem, BNXT_RE_PAGE_SIZE_SUPPORTED, virt_addr);
4043 	if (!page_size) {
4044 		ibdev_err(&rdev->ibdev, "umem page size unsupported!");
4045 		return ERR_PTR(-EINVAL);
4046 	}
4047 
4048 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
4049 	if (!mr)
4050 		return ERR_PTR(-ENOMEM);
4051 
4052 	mr->rdev = rdev;
4053 	mr->qplib_mr.pd = &pd->qplib_pd;
4054 	mr->qplib_mr.flags = __from_ib_access_flags(mr_access_flags);
4055 	mr->qplib_mr.type = CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR;
4056 
4057 	rc = bnxt_qplib_alloc_mrw(&rdev->qplib_res, &mr->qplib_mr);
4058 	if (rc) {
4059 		ibdev_err(&rdev->ibdev, "Failed to allocate MR rc = %d", rc);
4060 		rc = -EIO;
4061 		goto free_mr;
4062 	}
4063 	/* The fixed portion of the rkey is the same as the lkey */
4064 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
4065 	mr->ib_umem = umem;
4066 	mr->qplib_mr.va = virt_addr;
4067 	mr->qplib_mr.total_size = length;
4068 
4069 	umem_pgs = ib_umem_num_dma_blocks(umem, page_size);
4070 	rc = bnxt_qplib_reg_mr(&rdev->qplib_res, &mr->qplib_mr, umem,
4071 			       umem_pgs, page_size);
4072 	if (rc) {
4073 		ibdev_err(&rdev->ibdev, "Failed to register user MR - rc = %d\n", rc);
4074 		rc = -EIO;
4075 		goto free_mrw;
4076 	}
4077 
4078 	mr->ib_mr.lkey = mr->qplib_mr.lkey;
4079 	mr->ib_mr.rkey = mr->qplib_mr.lkey;
4080 	active_mrs = atomic_inc_return(&rdev->stats.res.mr_count);
4081 	if (active_mrs > rdev->stats.res.mr_watermark)
4082 		rdev->stats.res.mr_watermark = active_mrs;
4083 
4084 	return &mr->ib_mr;
4085 
4086 free_mrw:
4087 	bnxt_qplib_free_mrw(&rdev->qplib_res, &mr->qplib_mr);
4088 free_mr:
4089 	kfree(mr);
4090 	return ERR_PTR(rc);
4091 }
4092 
bnxt_re_reg_user_mr(struct ib_pd * ib_pd,u64 start,u64 length,u64 virt_addr,int mr_access_flags,struct ib_udata * udata)4093 struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
4094 				  u64 virt_addr, int mr_access_flags,
4095 				  struct ib_udata *udata)
4096 {
4097 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
4098 	struct bnxt_re_dev *rdev = pd->rdev;
4099 	struct ib_umem *umem;
4100 	struct ib_mr *ib_mr;
4101 
4102 	umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags);
4103 	if (IS_ERR(umem))
4104 		return ERR_CAST(umem);
4105 
4106 	ib_mr = __bnxt_re_user_reg_mr(ib_pd, length, virt_addr, mr_access_flags, umem);
4107 	if (IS_ERR(ib_mr))
4108 		ib_umem_release(umem);
4109 	return ib_mr;
4110 }
4111 
bnxt_re_reg_user_mr_dmabuf(struct ib_pd * ib_pd,u64 start,u64 length,u64 virt_addr,int fd,int mr_access_flags,struct ib_udata * udata)4112 struct ib_mr *bnxt_re_reg_user_mr_dmabuf(struct ib_pd *ib_pd, u64 start,
4113 					 u64 length, u64 virt_addr, int fd,
4114 					 int mr_access_flags, struct ib_udata *udata)
4115 {
4116 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
4117 	struct bnxt_re_dev *rdev = pd->rdev;
4118 	struct ib_umem_dmabuf *umem_dmabuf;
4119 	struct ib_umem *umem;
4120 	struct ib_mr *ib_mr;
4121 
4122 	umem_dmabuf = ib_umem_dmabuf_get_pinned(&rdev->ibdev, start, length,
4123 						fd, mr_access_flags);
4124 	if (IS_ERR(umem_dmabuf))
4125 		return ERR_CAST(umem_dmabuf);
4126 
4127 	umem = &umem_dmabuf->umem;
4128 
4129 	ib_mr = __bnxt_re_user_reg_mr(ib_pd, length, virt_addr, mr_access_flags, umem);
4130 	if (IS_ERR(ib_mr))
4131 		ib_umem_release(umem);
4132 	return ib_mr;
4133 }
4134 
bnxt_re_alloc_ucontext(struct ib_ucontext * ctx,struct ib_udata * udata)4135 int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata)
4136 {
4137 	struct ib_device *ibdev = ctx->device;
4138 	struct bnxt_re_ucontext *uctx =
4139 		container_of(ctx, struct bnxt_re_ucontext, ib_uctx);
4140 	struct bnxt_re_dev *rdev = to_bnxt_re_dev(ibdev, ibdev);
4141 	struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr;
4142 	struct bnxt_re_user_mmap_entry *entry;
4143 	struct bnxt_re_uctx_resp resp = {};
4144 	struct bnxt_re_uctx_req ureq = {};
4145 	u32 chip_met_rev_num = 0;
4146 	int rc;
4147 
4148 	ibdev_dbg(ibdev, "ABI version requested %u", ibdev->ops.uverbs_abi_ver);
4149 
4150 	if (ibdev->ops.uverbs_abi_ver != BNXT_RE_ABI_VERSION) {
4151 		ibdev_dbg(ibdev, " is different from the device %d ",
4152 			  BNXT_RE_ABI_VERSION);
4153 		return -EPERM;
4154 	}
4155 
4156 	uctx->rdev = rdev;
4157 
4158 	uctx->shpg = (void *)__get_free_page(GFP_KERNEL);
4159 	if (!uctx->shpg) {
4160 		rc = -ENOMEM;
4161 		goto fail;
4162 	}
4163 	spin_lock_init(&uctx->sh_lock);
4164 
4165 	resp.comp_mask = BNXT_RE_UCNTX_CMASK_HAVE_CCTX;
4166 	chip_met_rev_num = rdev->chip_ctx->chip_num;
4167 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_rev & 0xFF) <<
4168 			     BNXT_RE_CHIP_ID0_CHIP_REV_SFT;
4169 	chip_met_rev_num |= ((u32)rdev->chip_ctx->chip_metal & 0xFF) <<
4170 			     BNXT_RE_CHIP_ID0_CHIP_MET_SFT;
4171 	resp.chip_id0 = chip_met_rev_num;
4172 	/*Temp, Use xa_alloc instead */
4173 	resp.dev_id = rdev->en_dev->pdev->devfn;
4174 	resp.max_qp = rdev->qplib_ctx.qpc_count;
4175 	resp.pg_size = PAGE_SIZE;
4176 	resp.cqe_sz = sizeof(struct cq_base);
4177 	resp.max_cqd = dev_attr->max_cq_wqes;
4178 
4179 	if (rdev->chip_ctx->modes.db_push)
4180 		resp.comp_mask |= BNXT_RE_UCNTX_CMASK_WC_DPI_ENABLED;
4181 
4182 	entry = bnxt_re_mmap_entry_insert(uctx, 0, BNXT_RE_MMAP_SH_PAGE, NULL);
4183 	if (!entry) {
4184 		rc = -ENOMEM;
4185 		goto cfail;
4186 	}
4187 	uctx->shpage_mmap = &entry->rdma_entry;
4188 	if (rdev->pacing.dbr_pacing)
4189 		resp.comp_mask |= BNXT_RE_UCNTX_CMASK_DBR_PACING_ENABLED;
4190 
4191 	if (udata->inlen >= sizeof(ureq)) {
4192 		rc = ib_copy_from_udata(&ureq, udata, min(udata->inlen, sizeof(ureq)));
4193 		if (rc)
4194 			goto cfail;
4195 		if (ureq.comp_mask & BNXT_RE_COMP_MASK_REQ_UCNTX_POW2_SUPPORT) {
4196 			resp.comp_mask |= BNXT_RE_UCNTX_CMASK_POW2_DISABLED;
4197 			uctx->cmask |= BNXT_RE_UCNTX_CMASK_POW2_DISABLED;
4198 		}
4199 	}
4200 
4201 	rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));
4202 	if (rc) {
4203 		ibdev_err(ibdev, "Failed to copy user context");
4204 		rc = -EFAULT;
4205 		goto cfail;
4206 	}
4207 
4208 	return 0;
4209 cfail:
4210 	free_page((unsigned long)uctx->shpg);
4211 	uctx->shpg = NULL;
4212 fail:
4213 	return rc;
4214 }
4215 
bnxt_re_dealloc_ucontext(struct ib_ucontext * ib_uctx)4216 void bnxt_re_dealloc_ucontext(struct ib_ucontext *ib_uctx)
4217 {
4218 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
4219 						   struct bnxt_re_ucontext,
4220 						   ib_uctx);
4221 
4222 	struct bnxt_re_dev *rdev = uctx->rdev;
4223 
4224 	rdma_user_mmap_entry_remove(uctx->shpage_mmap);
4225 	uctx->shpage_mmap = NULL;
4226 	if (uctx->shpg)
4227 		free_page((unsigned long)uctx->shpg);
4228 
4229 	if (uctx->dpi.dbr) {
4230 		/* Free DPI only if this is the first PD allocated by the
4231 		 * application and mark the context dpi as NULL
4232 		 */
4233 		bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->dpi);
4234 		uctx->dpi.dbr = NULL;
4235 	}
4236 }
4237 
4238 /* Helper function to mmap the virtual memory from user app */
bnxt_re_mmap(struct ib_ucontext * ib_uctx,struct vm_area_struct * vma)4239 int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma)
4240 {
4241 	struct bnxt_re_ucontext *uctx = container_of(ib_uctx,
4242 						   struct bnxt_re_ucontext,
4243 						   ib_uctx);
4244 	struct bnxt_re_user_mmap_entry *bnxt_entry;
4245 	struct rdma_user_mmap_entry *rdma_entry;
4246 	int ret = 0;
4247 	u64 pfn;
4248 
4249 	rdma_entry = rdma_user_mmap_entry_get(&uctx->ib_uctx, vma);
4250 	if (!rdma_entry)
4251 		return -EINVAL;
4252 
4253 	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
4254 				  rdma_entry);
4255 
4256 	switch (bnxt_entry->mmap_flag) {
4257 	case BNXT_RE_MMAP_WC_DB:
4258 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4259 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4260 					pgprot_writecombine(vma->vm_page_prot),
4261 					rdma_entry);
4262 		break;
4263 	case BNXT_RE_MMAP_UC_DB:
4264 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4265 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4266 					pgprot_noncached(vma->vm_page_prot),
4267 				rdma_entry);
4268 		break;
4269 	case BNXT_RE_MMAP_SH_PAGE:
4270 		ret = vm_insert_page(vma, vma->vm_start, virt_to_page(uctx->shpg));
4271 		break;
4272 	case BNXT_RE_MMAP_DBR_BAR:
4273 		pfn = bnxt_entry->mem_offset >> PAGE_SHIFT;
4274 		ret = rdma_user_mmap_io(ib_uctx, vma, pfn, PAGE_SIZE,
4275 					pgprot_noncached(vma->vm_page_prot),
4276 					rdma_entry);
4277 		break;
4278 	case BNXT_RE_MMAP_DBR_PAGE:
4279 		/* Driver doesn't expect write access for user space */
4280 		if (vma->vm_flags & VM_WRITE)
4281 			return -EFAULT;
4282 		ret = vm_insert_page(vma, vma->vm_start,
4283 				     virt_to_page((void *)bnxt_entry->mem_offset));
4284 		break;
4285 	default:
4286 		ret = -EINVAL;
4287 		break;
4288 	}
4289 
4290 	rdma_user_mmap_entry_put(rdma_entry);
4291 	return ret;
4292 }
4293 
bnxt_re_mmap_free(struct rdma_user_mmap_entry * rdma_entry)4294 void bnxt_re_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
4295 {
4296 	struct bnxt_re_user_mmap_entry *bnxt_entry;
4297 
4298 	bnxt_entry = container_of(rdma_entry, struct bnxt_re_user_mmap_entry,
4299 				  rdma_entry);
4300 
4301 	kfree(bnxt_entry);
4302 }
4303 
UVERBS_HANDLER(BNXT_RE_METHOD_NOTIFY_DRV)4304 static int UVERBS_HANDLER(BNXT_RE_METHOD_NOTIFY_DRV)(struct uverbs_attr_bundle *attrs)
4305 {
4306 	struct bnxt_re_ucontext *uctx;
4307 
4308 	uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx);
4309 	bnxt_re_pacing_alert(uctx->rdev);
4310 	return 0;
4311 }
4312 
UVERBS_HANDLER(BNXT_RE_METHOD_ALLOC_PAGE)4313 static int UVERBS_HANDLER(BNXT_RE_METHOD_ALLOC_PAGE)(struct uverbs_attr_bundle *attrs)
4314 {
4315 	struct ib_uobject *uobj = uverbs_attr_get_uobject(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
4316 	enum bnxt_re_alloc_page_type alloc_type;
4317 	struct bnxt_re_user_mmap_entry *entry;
4318 	enum bnxt_re_mmap_flag mmap_flag;
4319 	struct bnxt_qplib_chip_ctx *cctx;
4320 	struct bnxt_re_ucontext *uctx;
4321 	struct bnxt_re_dev *rdev;
4322 	u64 mmap_offset;
4323 	u32 length;
4324 	u32 dpi;
4325 	u64 addr;
4326 	int err;
4327 
4328 	uctx = container_of(ib_uverbs_get_ucontext(attrs), struct bnxt_re_ucontext, ib_uctx);
4329 	if (IS_ERR(uctx))
4330 		return PTR_ERR(uctx);
4331 
4332 	err = uverbs_get_const(&alloc_type, attrs, BNXT_RE_ALLOC_PAGE_TYPE);
4333 	if (err)
4334 		return err;
4335 
4336 	rdev = uctx->rdev;
4337 	cctx = rdev->chip_ctx;
4338 
4339 	switch (alloc_type) {
4340 	case BNXT_RE_ALLOC_WC_PAGE:
4341 		if (cctx->modes.db_push)  {
4342 			if (bnxt_qplib_alloc_dpi(&rdev->qplib_res, &uctx->wcdpi,
4343 						 uctx, BNXT_QPLIB_DPI_TYPE_WC))
4344 				return -ENOMEM;
4345 			length = PAGE_SIZE;
4346 			dpi = uctx->wcdpi.dpi;
4347 			addr = (u64)uctx->wcdpi.umdbr;
4348 			mmap_flag = BNXT_RE_MMAP_WC_DB;
4349 		} else {
4350 			return -EINVAL;
4351 		}
4352 
4353 		break;
4354 	case BNXT_RE_ALLOC_DBR_BAR_PAGE:
4355 		length = PAGE_SIZE;
4356 		addr = (u64)rdev->pacing.dbr_bar_addr;
4357 		mmap_flag = BNXT_RE_MMAP_DBR_BAR;
4358 		break;
4359 
4360 	case BNXT_RE_ALLOC_DBR_PAGE:
4361 		length = PAGE_SIZE;
4362 		addr = (u64)rdev->pacing.dbr_page;
4363 		mmap_flag = BNXT_RE_MMAP_DBR_PAGE;
4364 		break;
4365 
4366 	default:
4367 		return -EOPNOTSUPP;
4368 	}
4369 
4370 	entry = bnxt_re_mmap_entry_insert(uctx, addr, mmap_flag, &mmap_offset);
4371 	if (!entry)
4372 		return -ENOMEM;
4373 
4374 	uobj->object = entry;
4375 	uverbs_finalize_uobj_create(attrs, BNXT_RE_ALLOC_PAGE_HANDLE);
4376 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
4377 			     &mmap_offset, sizeof(mmap_offset));
4378 	if (err)
4379 		return err;
4380 
4381 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
4382 			     &length, sizeof(length));
4383 	if (err)
4384 		return err;
4385 
4386 	err = uverbs_copy_to(attrs, BNXT_RE_ALLOC_PAGE_DPI,
4387 			     &dpi, sizeof(length));
4388 	if (err)
4389 		return err;
4390 
4391 	return 0;
4392 }
4393 
alloc_page_obj_cleanup(struct ib_uobject * uobject,enum rdma_remove_reason why,struct uverbs_attr_bundle * attrs)4394 static int alloc_page_obj_cleanup(struct ib_uobject *uobject,
4395 				  enum rdma_remove_reason why,
4396 			    struct uverbs_attr_bundle *attrs)
4397 {
4398 	struct  bnxt_re_user_mmap_entry *entry = uobject->object;
4399 	struct bnxt_re_ucontext *uctx = entry->uctx;
4400 
4401 	switch (entry->mmap_flag) {
4402 	case BNXT_RE_MMAP_WC_DB:
4403 		if (uctx && uctx->wcdpi.dbr) {
4404 			struct bnxt_re_dev *rdev = uctx->rdev;
4405 
4406 			bnxt_qplib_dealloc_dpi(&rdev->qplib_res, &uctx->wcdpi);
4407 			uctx->wcdpi.dbr = NULL;
4408 		}
4409 		break;
4410 	case BNXT_RE_MMAP_DBR_BAR:
4411 	case BNXT_RE_MMAP_DBR_PAGE:
4412 		break;
4413 	default:
4414 		goto exit;
4415 	}
4416 	rdma_user_mmap_entry_remove(&entry->rdma_entry);
4417 exit:
4418 	return 0;
4419 }
4420 
4421 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_ALLOC_PAGE,
4422 			    UVERBS_ATTR_IDR(BNXT_RE_ALLOC_PAGE_HANDLE,
4423 					    BNXT_RE_OBJECT_ALLOC_PAGE,
4424 					    UVERBS_ACCESS_NEW,
4425 					    UA_MANDATORY),
4426 			    UVERBS_ATTR_CONST_IN(BNXT_RE_ALLOC_PAGE_TYPE,
4427 						 enum bnxt_re_alloc_page_type,
4428 						 UA_MANDATORY),
4429 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_OFFSET,
4430 						UVERBS_ATTR_TYPE(u64),
4431 						UA_MANDATORY),
4432 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_MMAP_LENGTH,
4433 						UVERBS_ATTR_TYPE(u32),
4434 						UA_MANDATORY),
4435 			    UVERBS_ATTR_PTR_OUT(BNXT_RE_ALLOC_PAGE_DPI,
4436 						UVERBS_ATTR_TYPE(u32),
4437 						UA_MANDATORY));
4438 
4439 DECLARE_UVERBS_NAMED_METHOD_DESTROY(BNXT_RE_METHOD_DESTROY_PAGE,
4440 				    UVERBS_ATTR_IDR(BNXT_RE_DESTROY_PAGE_HANDLE,
4441 						    BNXT_RE_OBJECT_ALLOC_PAGE,
4442 						    UVERBS_ACCESS_DESTROY,
4443 						    UA_MANDATORY));
4444 
4445 DECLARE_UVERBS_NAMED_OBJECT(BNXT_RE_OBJECT_ALLOC_PAGE,
4446 			    UVERBS_TYPE_ALLOC_IDR(alloc_page_obj_cleanup),
4447 			    &UVERBS_METHOD(BNXT_RE_METHOD_ALLOC_PAGE),
4448 			    &UVERBS_METHOD(BNXT_RE_METHOD_DESTROY_PAGE));
4449 
4450 DECLARE_UVERBS_NAMED_METHOD(BNXT_RE_METHOD_NOTIFY_DRV);
4451 
4452 DECLARE_UVERBS_GLOBAL_METHODS(BNXT_RE_OBJECT_NOTIFY_DRV,
4453 			      &UVERBS_METHOD(BNXT_RE_METHOD_NOTIFY_DRV));
4454 
4455 const struct uapi_definition bnxt_re_uapi_defs[] = {
4456 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_ALLOC_PAGE),
4457 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(BNXT_RE_OBJECT_NOTIFY_DRV),
4458 	{}
4459 };
4460