2 * Broadcom NetXtreme-E RoCE driver.
4 * Copyright (c) 2016 - 2017, Broadcom. All rights reserved. The term
5 * Broadcom refers to Broadcom Limited and/or its subsidiaries.
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
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
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
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.
36 * Description: IB Verbs interpreter
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>
45 #include <rdma/ib_verbs.h>
46 #include <rdma/ib_user_verbs.h>
47 #include <rdma/ib_umem.h>
48 #include <rdma/ib_addr.h>
49 #include <rdma/ib_mad.h>
50 #include <rdma/ib_cache.h>
55 #include "qplib_res.h"
58 #include "qplib_rcfw.h"
62 #include <rdma/bnxt_re-abi.h>
64 static int __from_ib_access_flags(int iflags
)
68 if (iflags
& IB_ACCESS_LOCAL_WRITE
)
69 qflags
|= BNXT_QPLIB_ACCESS_LOCAL_WRITE
;
70 if (iflags
& IB_ACCESS_REMOTE_READ
)
71 qflags
|= BNXT_QPLIB_ACCESS_REMOTE_READ
;
72 if (iflags
& IB_ACCESS_REMOTE_WRITE
)
73 qflags
|= BNXT_QPLIB_ACCESS_REMOTE_WRITE
;
74 if (iflags
& IB_ACCESS_REMOTE_ATOMIC
)
75 qflags
|= BNXT_QPLIB_ACCESS_REMOTE_ATOMIC
;
76 if (iflags
& IB_ACCESS_MW_BIND
)
77 qflags
|= BNXT_QPLIB_ACCESS_MW_BIND
;
78 if (iflags
& IB_ZERO_BASED
)
79 qflags
|= BNXT_QPLIB_ACCESS_ZERO_BASED
;
80 if (iflags
& IB_ACCESS_ON_DEMAND
)
81 qflags
|= BNXT_QPLIB_ACCESS_ON_DEMAND
;
85 static enum ib_access_flags
__to_ib_access_flags(int qflags
)
87 enum ib_access_flags iflags
= 0;
89 if (qflags
& BNXT_QPLIB_ACCESS_LOCAL_WRITE
)
90 iflags
|= IB_ACCESS_LOCAL_WRITE
;
91 if (qflags
& BNXT_QPLIB_ACCESS_REMOTE_WRITE
)
92 iflags
|= IB_ACCESS_REMOTE_WRITE
;
93 if (qflags
& BNXT_QPLIB_ACCESS_REMOTE_READ
)
94 iflags
|= IB_ACCESS_REMOTE_READ
;
95 if (qflags
& BNXT_QPLIB_ACCESS_REMOTE_ATOMIC
)
96 iflags
|= IB_ACCESS_REMOTE_ATOMIC
;
97 if (qflags
& BNXT_QPLIB_ACCESS_MW_BIND
)
98 iflags
|= IB_ACCESS_MW_BIND
;
99 if (qflags
& BNXT_QPLIB_ACCESS_ZERO_BASED
)
100 iflags
|= IB_ZERO_BASED
;
101 if (qflags
& BNXT_QPLIB_ACCESS_ON_DEMAND
)
102 iflags
|= IB_ACCESS_ON_DEMAND
;
106 static int bnxt_re_build_sgl(struct ib_sge
*ib_sg_list
,
107 struct bnxt_qplib_sge
*sg_list
, int num
)
111 for (i
= 0; i
< num
; i
++) {
112 sg_list
[i
].addr
= ib_sg_list
[i
].addr
;
113 sg_list
[i
].lkey
= ib_sg_list
[i
].lkey
;
114 sg_list
[i
].size
= ib_sg_list
[i
].length
;
115 total
+= sg_list
[i
].size
;
121 struct net_device
*bnxt_re_get_netdev(struct ib_device
*ibdev
, u8 port_num
)
123 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
124 struct net_device
*netdev
= NULL
;
128 netdev
= rdev
->netdev
;
136 int bnxt_re_query_device(struct ib_device
*ibdev
,
137 struct ib_device_attr
*ib_attr
,
138 struct ib_udata
*udata
)
140 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
141 struct bnxt_qplib_dev_attr
*dev_attr
= &rdev
->dev_attr
;
143 memset(ib_attr
, 0, sizeof(*ib_attr
));
145 ib_attr
->fw_ver
= (u64
)(unsigned long)(dev_attr
->fw_ver
);
146 bnxt_qplib_get_guid(rdev
->netdev
->dev_addr
,
147 (u8
*)&ib_attr
->sys_image_guid
);
148 ib_attr
->max_mr_size
= BNXT_RE_MAX_MR_SIZE
;
149 ib_attr
->page_size_cap
= BNXT_RE_PAGE_SIZE_4K
;
151 ib_attr
->vendor_id
= rdev
->en_dev
->pdev
->vendor
;
152 ib_attr
->vendor_part_id
= rdev
->en_dev
->pdev
->device
;
153 ib_attr
->hw_ver
= rdev
->en_dev
->pdev
->subsystem_device
;
154 ib_attr
->max_qp
= dev_attr
->max_qp
;
155 ib_attr
->max_qp_wr
= dev_attr
->max_qp_wqes
;
156 ib_attr
->device_cap_flags
=
157 IB_DEVICE_CURR_QP_STATE_MOD
158 | IB_DEVICE_RC_RNR_NAK_GEN
159 | IB_DEVICE_SHUTDOWN_PORT
160 | IB_DEVICE_SYS_IMAGE_GUID
161 | IB_DEVICE_LOCAL_DMA_LKEY
162 | IB_DEVICE_RESIZE_MAX_WR
163 | IB_DEVICE_PORT_ACTIVE_EVENT
164 | IB_DEVICE_N_NOTIFY_CQ
165 | IB_DEVICE_MEM_WINDOW
166 | IB_DEVICE_MEM_WINDOW_TYPE_2B
167 | IB_DEVICE_MEM_MGT_EXTENSIONS
;
168 ib_attr
->max_sge
= dev_attr
->max_qp_sges
;
169 ib_attr
->max_sge_rd
= dev_attr
->max_qp_sges
;
170 ib_attr
->max_cq
= dev_attr
->max_cq
;
171 ib_attr
->max_cqe
= dev_attr
->max_cq_wqes
;
172 ib_attr
->max_mr
= dev_attr
->max_mr
;
173 ib_attr
->max_pd
= dev_attr
->max_pd
;
174 ib_attr
->max_qp_rd_atom
= dev_attr
->max_qp_rd_atom
;
175 ib_attr
->max_qp_init_rd_atom
= dev_attr
->max_qp_init_rd_atom
;
176 if (dev_attr
->is_atomic
) {
177 ib_attr
->atomic_cap
= IB_ATOMIC_HCA
;
178 ib_attr
->masked_atomic_cap
= IB_ATOMIC_HCA
;
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;
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
;
194 ib_attr
->max_fmr
= 0;
195 ib_attr
->max_map_per_fmr
= 0;
197 ib_attr
->max_srq
= dev_attr
->max_srq
;
198 ib_attr
->max_srq_wr
= dev_attr
->max_srq_wqes
;
199 ib_attr
->max_srq_sge
= dev_attr
->max_srq_sges
;
201 ib_attr
->max_fast_reg_page_list_len
= MAX_PBL_LVL_1_PGS
;
203 ib_attr
->max_pkeys
= 1;
204 ib_attr
->local_ca_ack_delay
= BNXT_RE_DEFAULT_ACK_DELAY
;
208 int bnxt_re_modify_device(struct ib_device
*ibdev
,
209 int device_modify_mask
,
210 struct ib_device_modify
*device_modify
)
212 switch (device_modify_mask
) {
213 case IB_DEVICE_MODIFY_SYS_IMAGE_GUID
:
214 /* Modify the GUID requires the modification of the GID table */
215 /* GUID should be made as READ-ONLY */
217 case IB_DEVICE_MODIFY_NODE_DESC
:
218 /* Node Desc should be made as READ-ONLY */
227 int bnxt_re_query_port(struct ib_device
*ibdev
, u8 port_num
,
228 struct ib_port_attr
*port_attr
)
230 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
231 struct bnxt_qplib_dev_attr
*dev_attr
= &rdev
->dev_attr
;
233 memset(port_attr
, 0, sizeof(*port_attr
));
235 if (netif_running(rdev
->netdev
) && netif_carrier_ok(rdev
->netdev
)) {
236 port_attr
->state
= IB_PORT_ACTIVE
;
237 port_attr
->phys_state
= 5;
239 port_attr
->state
= IB_PORT_DOWN
;
240 port_attr
->phys_state
= 3;
242 port_attr
->max_mtu
= IB_MTU_4096
;
243 port_attr
->active_mtu
= iboe_get_mtu(rdev
->netdev
->mtu
);
244 port_attr
->gid_tbl_len
= dev_attr
->max_sgid
;
245 port_attr
->port_cap_flags
= IB_PORT_CM_SUP
| IB_PORT_REINIT_SUP
|
246 IB_PORT_DEVICE_MGMT_SUP
|
247 IB_PORT_VENDOR_CLASS_SUP
|
248 IB_PORT_IP_BASED_GIDS
;
250 /* Max MSG size set to 2G for now */
251 port_attr
->max_msg_sz
= 0x80000000;
252 port_attr
->bad_pkey_cntr
= 0;
253 port_attr
->qkey_viol_cntr
= 0;
254 port_attr
->pkey_tbl_len
= dev_attr
->max_pkey
;
256 port_attr
->sm_lid
= 0;
258 port_attr
->max_vl_num
= 4;
259 port_attr
->sm_sl
= 0;
260 port_attr
->subnet_timeout
= 0;
261 port_attr
->init_type_reply
= 0;
262 /* call the underlying netdev's ethtool hooks to query speed settings
263 * for which we acquire rtnl_lock _only_ if it's registered with
264 * IB stack to avoid race in the NETDEV_UNREG path
266 if (test_bit(BNXT_RE_FLAG_IBDEV_REGISTERED
, &rdev
->flags
))
267 if (ib_get_eth_speed(ibdev
, port_num
, &port_attr
->active_speed
,
268 &port_attr
->active_width
))
273 int bnxt_re_get_port_immutable(struct ib_device
*ibdev
, u8 port_num
,
274 struct ib_port_immutable
*immutable
)
276 struct ib_port_attr port_attr
;
278 if (bnxt_re_query_port(ibdev
, port_num
, &port_attr
))
281 immutable
->pkey_tbl_len
= port_attr
.pkey_tbl_len
;
282 immutable
->gid_tbl_len
= port_attr
.gid_tbl_len
;
283 immutable
->core_cap_flags
= RDMA_CORE_PORT_IBA_ROCE
;
284 immutable
->core_cap_flags
|= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP
;
285 immutable
->max_mad_size
= IB_MGMT_MAD_SIZE
;
289 int bnxt_re_query_pkey(struct ib_device
*ibdev
, u8 port_num
,
290 u16 index
, u16
*pkey
)
292 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
294 /* Ignore port_num */
296 memset(pkey
, 0, sizeof(*pkey
));
297 return bnxt_qplib_get_pkey(&rdev
->qplib_res
,
298 &rdev
->qplib_res
.pkey_tbl
, index
, pkey
);
301 int bnxt_re_query_gid(struct ib_device
*ibdev
, u8 port_num
,
302 int index
, union ib_gid
*gid
)
304 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
307 /* Ignore port_num */
308 memset(gid
, 0, sizeof(*gid
));
309 rc
= bnxt_qplib_get_sgid(&rdev
->qplib_res
,
310 &rdev
->qplib_res
.sgid_tbl
, index
,
311 (struct bnxt_qplib_gid
*)gid
);
315 int bnxt_re_del_gid(struct ib_device
*ibdev
, u8 port_num
,
316 unsigned int index
, void **context
)
319 struct bnxt_re_gid_ctx
*ctx
, **ctx_tbl
;
320 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
321 struct bnxt_qplib_sgid_tbl
*sgid_tbl
= &rdev
->qplib_res
.sgid_tbl
;
323 /* Delete the entry from the hardware */
328 if (sgid_tbl
&& sgid_tbl
->active
) {
329 if (ctx
->idx
>= sgid_tbl
->max
)
333 rc
= bnxt_qplib_del_sgid(sgid_tbl
,
334 &sgid_tbl
->tbl
[ctx
->idx
],
337 dev_err(rdev_to_dev(rdev
),
338 "Failed to remove GID: %#x", rc
);
340 ctx_tbl
= sgid_tbl
->ctx
;
341 ctx_tbl
[ctx
->idx
] = NULL
;
351 int bnxt_re_add_gid(struct ib_device
*ibdev
, u8 port_num
,
352 unsigned int index
, const union ib_gid
*gid
,
353 const struct ib_gid_attr
*attr
, void **context
)
357 u16 vlan_id
= 0xFFFF;
358 struct bnxt_re_gid_ctx
*ctx
, **ctx_tbl
;
359 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
360 struct bnxt_qplib_sgid_tbl
*sgid_tbl
= &rdev
->qplib_res
.sgid_tbl
;
362 if ((attr
->ndev
) && is_vlan_dev(attr
->ndev
))
363 vlan_id
= vlan_dev_vlan_id(attr
->ndev
);
365 rc
= bnxt_qplib_add_sgid(sgid_tbl
, (struct bnxt_qplib_gid
*)gid
,
366 rdev
->qplib_res
.netdev
->dev_addr
,
367 vlan_id
, true, &tbl_idx
);
368 if (rc
== -EALREADY
) {
369 ctx_tbl
= sgid_tbl
->ctx
;
370 ctx_tbl
[tbl_idx
]->refcnt
++;
371 *context
= ctx_tbl
[tbl_idx
];
376 dev_err(rdev_to_dev(rdev
), "Failed to add GID: %#x", rc
);
380 ctx
= kmalloc(sizeof(*ctx
), GFP_KERNEL
);
383 ctx_tbl
= sgid_tbl
->ctx
;
386 ctx_tbl
[tbl_idx
] = ctx
;
391 enum rdma_link_layer
bnxt_re_get_link_layer(struct ib_device
*ibdev
,
394 return IB_LINK_LAYER_ETHERNET
;
397 #define BNXT_RE_FENCE_PBL_SIZE DIV_ROUND_UP(BNXT_RE_FENCE_BYTES, PAGE_SIZE)
399 static void bnxt_re_create_fence_wqe(struct bnxt_re_pd
*pd
)
401 struct bnxt_re_fence_data
*fence
= &pd
->fence
;
402 struct ib_mr
*ib_mr
= &fence
->mr
->ib_mr
;
403 struct bnxt_qplib_swqe
*wqe
= &fence
->bind_wqe
;
405 memset(wqe
, 0, sizeof(*wqe
));
406 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_BIND_MW
;
407 wqe
->wr_id
= BNXT_QPLIB_FENCE_WRID
;
408 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP
;
409 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE
;
410 wqe
->bind
.zero_based
= false;
411 wqe
->bind
.parent_l_key
= ib_mr
->lkey
;
412 wqe
->bind
.va
= (u64
)(unsigned long)fence
->va
;
413 wqe
->bind
.length
= fence
->size
;
414 wqe
->bind
.access_cntl
= __from_ib_access_flags(IB_ACCESS_REMOTE_READ
);
415 wqe
->bind
.mw_type
= SQ_BIND_MW_TYPE_TYPE1
;
417 /* Save the initial rkey in fence structure for now;
418 * wqe->bind.r_key will be set at (re)bind time.
420 fence
->bind_rkey
= ib_inc_rkey(fence
->mw
->rkey
);
423 static int bnxt_re_bind_fence_mw(struct bnxt_qplib_qp
*qplib_qp
)
425 struct bnxt_re_qp
*qp
= container_of(qplib_qp
, struct bnxt_re_qp
,
427 struct ib_pd
*ib_pd
= qp
->ib_qp
.pd
;
428 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
429 struct bnxt_re_fence_data
*fence
= &pd
->fence
;
430 struct bnxt_qplib_swqe
*fence_wqe
= &fence
->bind_wqe
;
431 struct bnxt_qplib_swqe wqe
;
434 memcpy(&wqe
, fence_wqe
, sizeof(wqe
));
435 wqe
.bind
.r_key
= fence
->bind_rkey
;
436 fence
->bind_rkey
= ib_inc_rkey(fence
->bind_rkey
);
438 dev_dbg(rdev_to_dev(qp
->rdev
),
439 "Posting bind fence-WQE: rkey: %#x QP: %d PD: %p\n",
440 wqe
.bind
.r_key
, qp
->qplib_qp
.id
, pd
);
441 rc
= bnxt_qplib_post_send(&qp
->qplib_qp
, &wqe
);
443 dev_err(rdev_to_dev(qp
->rdev
), "Failed to bind fence-WQE\n");
446 bnxt_qplib_post_send_db(&qp
->qplib_qp
);
451 static void bnxt_re_destroy_fence_mr(struct bnxt_re_pd
*pd
)
453 struct bnxt_re_fence_data
*fence
= &pd
->fence
;
454 struct bnxt_re_dev
*rdev
= pd
->rdev
;
455 struct device
*dev
= &rdev
->en_dev
->pdev
->dev
;
456 struct bnxt_re_mr
*mr
= fence
->mr
;
459 bnxt_re_dealloc_mw(fence
->mw
);
464 bnxt_qplib_dereg_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
,
467 bnxt_qplib_free_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
471 if (fence
->dma_addr
) {
472 dma_unmap_single(dev
, fence
->dma_addr
, BNXT_RE_FENCE_BYTES
,
478 static int bnxt_re_create_fence_mr(struct bnxt_re_pd
*pd
)
480 int mr_access_flags
= IB_ACCESS_LOCAL_WRITE
| IB_ACCESS_MW_BIND
;
481 struct bnxt_re_fence_data
*fence
= &pd
->fence
;
482 struct bnxt_re_dev
*rdev
= pd
->rdev
;
483 struct device
*dev
= &rdev
->en_dev
->pdev
->dev
;
484 struct bnxt_re_mr
*mr
= NULL
;
485 dma_addr_t dma_addr
= 0;
490 dma_addr
= dma_map_single(dev
, fence
->va
, BNXT_RE_FENCE_BYTES
,
492 rc
= dma_mapping_error(dev
, dma_addr
);
494 dev_err(rdev_to_dev(rdev
), "Failed to dma-map fence-MR-mem\n");
499 fence
->dma_addr
= dma_addr
;
502 mr
= kzalloc(sizeof(*mr
), GFP_KERNEL
);
509 mr
->qplib_mr
.pd
= &pd
->qplib_pd
;
510 mr
->qplib_mr
.type
= CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR
;
511 mr
->qplib_mr
.flags
= __from_ib_access_flags(mr_access_flags
);
512 rc
= bnxt_qplib_alloc_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
514 dev_err(rdev_to_dev(rdev
), "Failed to alloc fence-HW-MR\n");
519 mr
->ib_mr
.lkey
= mr
->qplib_mr
.lkey
;
520 mr
->qplib_mr
.va
= (u64
)(unsigned long)fence
->va
;
521 mr
->qplib_mr
.total_size
= BNXT_RE_FENCE_BYTES
;
523 rc
= bnxt_qplib_reg_mr(&rdev
->qplib_res
, &mr
->qplib_mr
, &pbl_tbl
,
524 BNXT_RE_FENCE_PBL_SIZE
, false);
526 dev_err(rdev_to_dev(rdev
), "Failed to register fence-MR\n");
529 mr
->ib_mr
.rkey
= mr
->qplib_mr
.rkey
;
531 /* Create a fence MW only for kernel consumers */
532 mw
= bnxt_re_alloc_mw(&pd
->ib_pd
, IB_MW_TYPE_1
, NULL
);
534 dev_err(rdev_to_dev(rdev
),
535 "Failed to create fence-MW for PD: %p\n", pd
);
541 bnxt_re_create_fence_wqe(pd
);
545 bnxt_re_destroy_fence_mr(pd
);
549 /* Protection Domains */
550 int bnxt_re_dealloc_pd(struct ib_pd
*ib_pd
)
552 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
553 struct bnxt_re_dev
*rdev
= pd
->rdev
;
556 bnxt_re_destroy_fence_mr(pd
);
558 if (pd
->qplib_pd
.id
) {
559 rc
= bnxt_qplib_dealloc_pd(&rdev
->qplib_res
,
560 &rdev
->qplib_res
.pd_tbl
,
563 dev_err(rdev_to_dev(rdev
), "Failed to deallocate HW PD");
570 struct ib_pd
*bnxt_re_alloc_pd(struct ib_device
*ibdev
,
571 struct ib_ucontext
*ucontext
,
572 struct ib_udata
*udata
)
574 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
575 struct bnxt_re_ucontext
*ucntx
= container_of(ucontext
,
576 struct bnxt_re_ucontext
,
578 struct bnxt_re_pd
*pd
;
581 pd
= kzalloc(sizeof(*pd
), GFP_KERNEL
);
583 return ERR_PTR(-ENOMEM
);
586 if (bnxt_qplib_alloc_pd(&rdev
->qplib_res
.pd_tbl
, &pd
->qplib_pd
)) {
587 dev_err(rdev_to_dev(rdev
), "Failed to allocate HW PD");
593 struct bnxt_re_pd_resp resp
;
595 if (!ucntx
->dpi
.dbr
) {
596 /* Allocate DPI in alloc_pd to avoid failing of
597 * ibv_devinfo and family of application when DPIs
600 if (bnxt_qplib_alloc_dpi(&rdev
->qplib_res
.dpi_tbl
,
601 &ucntx
->dpi
, ucntx
)) {
607 resp
.pdid
= pd
->qplib_pd
.id
;
608 /* Still allow mapping this DBR to the new user PD. */
609 resp
.dpi
= ucntx
->dpi
.dpi
;
610 resp
.dbr
= (u64
)ucntx
->dpi
.umdbr
;
612 rc
= ib_copy_to_udata(udata
, &resp
, sizeof(resp
));
614 dev_err(rdev_to_dev(rdev
),
615 "Failed to copy user response\n");
621 if (bnxt_re_create_fence_mr(pd
))
622 dev_warn(rdev_to_dev(rdev
),
623 "Failed to create Fence-MR\n");
626 (void)bnxt_qplib_dealloc_pd(&rdev
->qplib_res
, &rdev
->qplib_res
.pd_tbl
,
633 /* Address Handles */
634 int bnxt_re_destroy_ah(struct ib_ah
*ib_ah
)
636 struct bnxt_re_ah
*ah
= container_of(ib_ah
, struct bnxt_re_ah
, ib_ah
);
637 struct bnxt_re_dev
*rdev
= ah
->rdev
;
640 rc
= bnxt_qplib_destroy_ah(&rdev
->qplib_res
, &ah
->qplib_ah
);
642 dev_err(rdev_to_dev(rdev
), "Failed to destroy HW AH");
649 struct ib_ah
*bnxt_re_create_ah(struct ib_pd
*ib_pd
,
650 struct rdma_ah_attr
*ah_attr
,
651 struct ib_udata
*udata
)
653 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
654 struct bnxt_re_dev
*rdev
= pd
->rdev
;
655 struct bnxt_re_ah
*ah
;
656 const struct ib_global_route
*grh
= rdma_ah_read_grh(ah_attr
);
661 struct ib_gid_attr sgid_attr
;
663 if (!(rdma_ah_get_ah_flags(ah_attr
) & IB_AH_GRH
)) {
664 dev_err(rdev_to_dev(rdev
), "Failed to alloc AH: GRH not set");
665 return ERR_PTR(-EINVAL
);
667 ah
= kzalloc(sizeof(*ah
), GFP_ATOMIC
);
669 return ERR_PTR(-ENOMEM
);
672 ah
->qplib_ah
.pd
= &pd
->qplib_pd
;
674 /* Supply the configuration for the HW */
675 memcpy(ah
->qplib_ah
.dgid
.data
, grh
->dgid
.raw
,
676 sizeof(union ib_gid
));
678 * If RoCE V2 is enabled, stack will have two entries for
679 * each GID entry. Avoiding this duplicte entry in HW. Dividing
680 * the GID index by 2 for RoCE V2
682 ah
->qplib_ah
.sgid_index
= grh
->sgid_index
/ 2;
683 ah
->qplib_ah
.host_sgid_index
= grh
->sgid_index
;
684 ah
->qplib_ah
.traffic_class
= grh
->traffic_class
;
685 ah
->qplib_ah
.flow_label
= grh
->flow_label
;
686 ah
->qplib_ah
.hop_limit
= grh
->hop_limit
;
687 ah
->qplib_ah
.sl
= rdma_ah_get_sl(ah_attr
);
688 if (ib_pd
->uobject
&&
689 !rdma_is_multicast_addr((struct in6_addr
*)
691 !rdma_link_local_addr((struct in6_addr
*)
695 rc
= ib_get_cached_gid(&rdev
->ibdev
, 1,
696 grh
->sgid_index
, &sgid
,
699 dev_err(rdev_to_dev(rdev
),
700 "Failed to query gid at index %d",
704 if (sgid_attr
.ndev
) {
705 if (is_vlan_dev(sgid_attr
.ndev
))
706 vlan_tag
= vlan_dev_vlan_id(sgid_attr
.ndev
);
707 dev_put(sgid_attr
.ndev
);
709 /* Get network header type for this GID */
710 nw_type
= ib_gid_to_network_type(sgid_attr
.gid_type
, &sgid
);
712 case RDMA_NETWORK_IPV4
:
713 ah
->qplib_ah
.nw_type
= CMDQ_CREATE_AH_TYPE_V2IPV4
;
715 case RDMA_NETWORK_IPV6
:
716 ah
->qplib_ah
.nw_type
= CMDQ_CREATE_AH_TYPE_V2IPV6
;
719 ah
->qplib_ah
.nw_type
= CMDQ_CREATE_AH_TYPE_V1
;
722 rc
= rdma_addr_find_l2_eth_by_grh(&sgid
, &grh
->dgid
,
723 ah_attr
->roce
.dmac
, &vlan_tag
,
724 &sgid_attr
.ndev
->ifindex
,
727 dev_err(rdev_to_dev(rdev
), "Failed to get dmac\n");
732 memcpy(ah
->qplib_ah
.dmac
, ah_attr
->roce
.dmac
, ETH_ALEN
);
733 rc
= bnxt_qplib_create_ah(&rdev
->qplib_res
, &ah
->qplib_ah
);
735 dev_err(rdev_to_dev(rdev
), "Failed to allocate HW AH");
739 /* Write AVID to shared page. */
740 if (ib_pd
->uobject
) {
741 struct ib_ucontext
*ib_uctx
= ib_pd
->uobject
->context
;
742 struct bnxt_re_ucontext
*uctx
;
746 uctx
= container_of(ib_uctx
, struct bnxt_re_ucontext
, ib_uctx
);
747 spin_lock_irqsave(&uctx
->sh_lock
, flag
);
748 wrptr
= (u32
*)(uctx
->shpg
+ BNXT_RE_AVID_OFFT
);
749 *wrptr
= ah
->qplib_ah
.id
;
750 wmb(); /* make sure cache is updated. */
751 spin_unlock_irqrestore(&uctx
->sh_lock
, flag
);
761 int bnxt_re_modify_ah(struct ib_ah
*ib_ah
, struct rdma_ah_attr
*ah_attr
)
766 int bnxt_re_query_ah(struct ib_ah
*ib_ah
, struct rdma_ah_attr
*ah_attr
)
768 struct bnxt_re_ah
*ah
= container_of(ib_ah
, struct bnxt_re_ah
, ib_ah
);
770 ah_attr
->type
= ib_ah
->type
;
771 rdma_ah_set_sl(ah_attr
, ah
->qplib_ah
.sl
);
772 memcpy(ah_attr
->roce
.dmac
, ah
->qplib_ah
.dmac
, ETH_ALEN
);
773 rdma_ah_set_grh(ah_attr
, NULL
, 0,
774 ah
->qplib_ah
.host_sgid_index
,
775 0, ah
->qplib_ah
.traffic_class
);
776 rdma_ah_set_dgid_raw(ah_attr
, ah
->qplib_ah
.dgid
.data
);
777 rdma_ah_set_port_num(ah_attr
, 1);
778 rdma_ah_set_static_rate(ah_attr
, 0);
783 int bnxt_re_destroy_qp(struct ib_qp
*ib_qp
)
785 struct bnxt_re_qp
*qp
= container_of(ib_qp
, struct bnxt_re_qp
, ib_qp
);
786 struct bnxt_re_dev
*rdev
= qp
->rdev
;
789 bnxt_qplib_del_flush_qp(&qp
->qplib_qp
);
790 rc
= bnxt_qplib_destroy_qp(&rdev
->qplib_res
, &qp
->qplib_qp
);
792 dev_err(rdev_to_dev(rdev
), "Failed to destroy HW QP");
795 if (ib_qp
->qp_type
== IB_QPT_GSI
&& rdev
->qp1_sqp
) {
796 rc
= bnxt_qplib_destroy_ah(&rdev
->qplib_res
,
797 &rdev
->sqp_ah
->qplib_ah
);
799 dev_err(rdev_to_dev(rdev
),
800 "Failed to destroy HW AH for shadow QP");
804 bnxt_qplib_del_flush_qp(&qp
->qplib_qp
);
805 rc
= bnxt_qplib_destroy_qp(&rdev
->qplib_res
,
806 &rdev
->qp1_sqp
->qplib_qp
);
808 dev_err(rdev_to_dev(rdev
),
809 "Failed to destroy Shadow QP");
812 mutex_lock(&rdev
->qp_lock
);
813 list_del(&rdev
->qp1_sqp
->list
);
814 atomic_dec(&rdev
->qp_count
);
815 mutex_unlock(&rdev
->qp_lock
);
818 kfree(rdev
->qp1_sqp
);
821 if (!IS_ERR_OR_NULL(qp
->rumem
))
822 ib_umem_release(qp
->rumem
);
823 if (!IS_ERR_OR_NULL(qp
->sumem
))
824 ib_umem_release(qp
->sumem
);
826 mutex_lock(&rdev
->qp_lock
);
828 atomic_dec(&rdev
->qp_count
);
829 mutex_unlock(&rdev
->qp_lock
);
834 static u8
__from_ib_qp_type(enum ib_qp_type type
)
838 return CMDQ_CREATE_QP1_TYPE_GSI
;
840 return CMDQ_CREATE_QP_TYPE_RC
;
842 return CMDQ_CREATE_QP_TYPE_UD
;
848 static int bnxt_re_init_user_qp(struct bnxt_re_dev
*rdev
, struct bnxt_re_pd
*pd
,
849 struct bnxt_re_qp
*qp
, struct ib_udata
*udata
)
851 struct bnxt_re_qp_req ureq
;
852 struct bnxt_qplib_qp
*qplib_qp
= &qp
->qplib_qp
;
853 struct ib_umem
*umem
;
855 struct ib_ucontext
*context
= pd
->ib_pd
.uobject
->context
;
856 struct bnxt_re_ucontext
*cntx
= container_of(context
,
857 struct bnxt_re_ucontext
,
859 if (ib_copy_from_udata(&ureq
, udata
, sizeof(ureq
)))
862 bytes
= (qplib_qp
->sq
.max_wqe
* BNXT_QPLIB_MAX_SQE_ENTRY_SIZE
);
863 /* Consider mapping PSN search memory only for RC QPs. */
864 if (qplib_qp
->type
== CMDQ_CREATE_QP_TYPE_RC
)
865 bytes
+= (qplib_qp
->sq
.max_wqe
* sizeof(struct sq_psn_search
));
866 bytes
= PAGE_ALIGN(bytes
);
867 umem
= ib_umem_get(context
, ureq
.qpsva
, bytes
,
868 IB_ACCESS_LOCAL_WRITE
, 1);
870 return PTR_ERR(umem
);
873 qplib_qp
->sq
.sglist
= umem
->sg_head
.sgl
;
874 qplib_qp
->sq
.nmap
= umem
->nmap
;
875 qplib_qp
->qp_handle
= ureq
.qp_handle
;
877 if (!qp
->qplib_qp
.srq
) {
878 bytes
= (qplib_qp
->rq
.max_wqe
* BNXT_QPLIB_MAX_RQE_ENTRY_SIZE
);
879 bytes
= PAGE_ALIGN(bytes
);
880 umem
= ib_umem_get(context
, ureq
.qprva
, bytes
,
881 IB_ACCESS_LOCAL_WRITE
, 1);
885 qplib_qp
->rq
.sglist
= umem
->sg_head
.sgl
;
886 qplib_qp
->rq
.nmap
= umem
->nmap
;
889 qplib_qp
->dpi
= &cntx
->dpi
;
892 ib_umem_release(qp
->sumem
);
894 qplib_qp
->sq
.sglist
= NULL
;
895 qplib_qp
->sq
.nmap
= 0;
897 return PTR_ERR(umem
);
900 static struct bnxt_re_ah
*bnxt_re_create_shadow_qp_ah
901 (struct bnxt_re_pd
*pd
,
902 struct bnxt_qplib_res
*qp1_res
,
903 struct bnxt_qplib_qp
*qp1_qp
)
905 struct bnxt_re_dev
*rdev
= pd
->rdev
;
906 struct bnxt_re_ah
*ah
;
910 ah
= kzalloc(sizeof(*ah
), GFP_KERNEL
);
915 ah
->qplib_ah
.pd
= &pd
->qplib_pd
;
917 rc
= bnxt_re_query_gid(&rdev
->ibdev
, 1, 0, &sgid
);
921 /* supply the dgid data same as sgid */
922 memcpy(ah
->qplib_ah
.dgid
.data
, &sgid
.raw
,
923 sizeof(union ib_gid
));
924 ah
->qplib_ah
.sgid_index
= 0;
926 ah
->qplib_ah
.traffic_class
= 0;
927 ah
->qplib_ah
.flow_label
= 0;
928 ah
->qplib_ah
.hop_limit
= 1;
930 /* Have DMAC same as SMAC */
931 ether_addr_copy(ah
->qplib_ah
.dmac
, rdev
->netdev
->dev_addr
);
933 rc
= bnxt_qplib_create_ah(&rdev
->qplib_res
, &ah
->qplib_ah
);
935 dev_err(rdev_to_dev(rdev
),
936 "Failed to allocate HW AH for Shadow QP");
947 static struct bnxt_re_qp
*bnxt_re_create_shadow_qp
948 (struct bnxt_re_pd
*pd
,
949 struct bnxt_qplib_res
*qp1_res
,
950 struct bnxt_qplib_qp
*qp1_qp
)
952 struct bnxt_re_dev
*rdev
= pd
->rdev
;
953 struct bnxt_re_qp
*qp
;
956 qp
= kzalloc(sizeof(*qp
), GFP_KERNEL
);
962 /* Initialize the shadow QP structure from the QP1 values */
963 ether_addr_copy(qp
->qplib_qp
.smac
, rdev
->netdev
->dev_addr
);
965 qp
->qplib_qp
.pd
= &pd
->qplib_pd
;
966 qp
->qplib_qp
.qp_handle
= (u64
)(unsigned long)(&qp
->qplib_qp
);
967 qp
->qplib_qp
.type
= IB_QPT_UD
;
969 qp
->qplib_qp
.max_inline_data
= 0;
970 qp
->qplib_qp
.sig_type
= true;
972 /* Shadow QP SQ depth should be same as QP1 RQ depth */
973 qp
->qplib_qp
.sq
.max_wqe
= qp1_qp
->rq
.max_wqe
;
974 qp
->qplib_qp
.sq
.max_sge
= 2;
975 /* Q full delta can be 1 since it is internal QP */
976 qp
->qplib_qp
.sq
.q_full_delta
= 1;
978 qp
->qplib_qp
.scq
= qp1_qp
->scq
;
979 qp
->qplib_qp
.rcq
= qp1_qp
->rcq
;
981 qp
->qplib_qp
.rq
.max_wqe
= qp1_qp
->rq
.max_wqe
;
982 qp
->qplib_qp
.rq
.max_sge
= qp1_qp
->rq
.max_sge
;
983 /* Q full delta can be 1 since it is internal QP */
984 qp
->qplib_qp
.rq
.q_full_delta
= 1;
986 qp
->qplib_qp
.mtu
= qp1_qp
->mtu
;
988 qp
->qplib_qp
.sq_hdr_buf_size
= 0;
989 qp
->qplib_qp
.rq_hdr_buf_size
= BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6
;
990 qp
->qplib_qp
.dpi
= &rdev
->dpi_privileged
;
992 rc
= bnxt_qplib_create_qp(qp1_res
, &qp
->qplib_qp
);
996 rdev
->sqp_id
= qp
->qplib_qp
.id
;
998 spin_lock_init(&qp
->sq_lock
);
999 INIT_LIST_HEAD(&qp
->list
);
1000 mutex_lock(&rdev
->qp_lock
);
1001 list_add_tail(&qp
->list
, &rdev
->qp_list
);
1002 atomic_inc(&rdev
->qp_count
);
1003 mutex_unlock(&rdev
->qp_lock
);
1010 struct ib_qp
*bnxt_re_create_qp(struct ib_pd
*ib_pd
,
1011 struct ib_qp_init_attr
*qp_init_attr
,
1012 struct ib_udata
*udata
)
1014 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
1015 struct bnxt_re_dev
*rdev
= pd
->rdev
;
1016 struct bnxt_qplib_dev_attr
*dev_attr
= &rdev
->dev_attr
;
1017 struct bnxt_re_qp
*qp
;
1018 struct bnxt_re_cq
*cq
;
1021 if ((qp_init_attr
->cap
.max_send_wr
> dev_attr
->max_qp_wqes
) ||
1022 (qp_init_attr
->cap
.max_recv_wr
> dev_attr
->max_qp_wqes
) ||
1023 (qp_init_attr
->cap
.max_send_sge
> dev_attr
->max_qp_sges
) ||
1024 (qp_init_attr
->cap
.max_recv_sge
> dev_attr
->max_qp_sges
) ||
1025 (qp_init_attr
->cap
.max_inline_data
> dev_attr
->max_inline_data
))
1026 return ERR_PTR(-EINVAL
);
1028 qp
= kzalloc(sizeof(*qp
), GFP_KERNEL
);
1030 return ERR_PTR(-ENOMEM
);
1033 ether_addr_copy(qp
->qplib_qp
.smac
, rdev
->netdev
->dev_addr
);
1034 qp
->qplib_qp
.pd
= &pd
->qplib_pd
;
1035 qp
->qplib_qp
.qp_handle
= (u64
)(unsigned long)(&qp
->qplib_qp
);
1036 qp
->qplib_qp
.type
= __from_ib_qp_type(qp_init_attr
->qp_type
);
1037 if (qp
->qplib_qp
.type
== IB_QPT_MAX
) {
1038 dev_err(rdev_to_dev(rdev
), "QP type 0x%x not supported",
1043 qp
->qplib_qp
.max_inline_data
= qp_init_attr
->cap
.max_inline_data
;
1044 qp
->qplib_qp
.sig_type
= ((qp_init_attr
->sq_sig_type
==
1045 IB_SIGNAL_ALL_WR
) ? true : false);
1047 qp
->qplib_qp
.sq
.max_sge
= qp_init_attr
->cap
.max_send_sge
;
1048 if (qp
->qplib_qp
.sq
.max_sge
> dev_attr
->max_qp_sges
)
1049 qp
->qplib_qp
.sq
.max_sge
= dev_attr
->max_qp_sges
;
1051 if (qp_init_attr
->send_cq
) {
1052 cq
= container_of(qp_init_attr
->send_cq
, struct bnxt_re_cq
,
1055 dev_err(rdev_to_dev(rdev
), "Send CQ not found");
1059 qp
->qplib_qp
.scq
= &cq
->qplib_cq
;
1062 if (qp_init_attr
->recv_cq
) {
1063 cq
= container_of(qp_init_attr
->recv_cq
, struct bnxt_re_cq
,
1066 dev_err(rdev_to_dev(rdev
), "Receive CQ not found");
1070 qp
->qplib_qp
.rcq
= &cq
->qplib_cq
;
1073 if (qp_init_attr
->srq
) {
1074 dev_err(rdev_to_dev(rdev
), "SRQ not supported");
1078 /* Allocate 1 more than what's provided so posting max doesn't
1081 entries
= roundup_pow_of_two(qp_init_attr
->cap
.max_recv_wr
+ 1);
1082 qp
->qplib_qp
.rq
.max_wqe
= min_t(u32
, entries
,
1083 dev_attr
->max_qp_wqes
+ 1);
1085 qp
->qplib_qp
.rq
.q_full_delta
= qp
->qplib_qp
.rq
.max_wqe
-
1086 qp_init_attr
->cap
.max_recv_wr
;
1088 qp
->qplib_qp
.rq
.max_sge
= qp_init_attr
->cap
.max_recv_sge
;
1089 if (qp
->qplib_qp
.rq
.max_sge
> dev_attr
->max_qp_sges
)
1090 qp
->qplib_qp
.rq
.max_sge
= dev_attr
->max_qp_sges
;
1093 qp
->qplib_qp
.mtu
= ib_mtu_enum_to_int(iboe_get_mtu(rdev
->netdev
->mtu
));
1095 if (qp_init_attr
->qp_type
== IB_QPT_GSI
) {
1096 /* Allocate 1 more than what's provided */
1097 entries
= roundup_pow_of_two(qp_init_attr
->cap
.max_send_wr
+ 1);
1098 qp
->qplib_qp
.sq
.max_wqe
= min_t(u32
, entries
,
1099 dev_attr
->max_qp_wqes
+ 1);
1100 qp
->qplib_qp
.sq
.q_full_delta
= qp
->qplib_qp
.sq
.max_wqe
-
1101 qp_init_attr
->cap
.max_send_wr
;
1102 qp
->qplib_qp
.rq
.max_sge
= dev_attr
->max_qp_sges
;
1103 if (qp
->qplib_qp
.rq
.max_sge
> dev_attr
->max_qp_sges
)
1104 qp
->qplib_qp
.rq
.max_sge
= dev_attr
->max_qp_sges
;
1105 qp
->qplib_qp
.sq
.max_sge
++;
1106 if (qp
->qplib_qp
.sq
.max_sge
> dev_attr
->max_qp_sges
)
1107 qp
->qplib_qp
.sq
.max_sge
= dev_attr
->max_qp_sges
;
1109 qp
->qplib_qp
.rq_hdr_buf_size
=
1110 BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2
;
1112 qp
->qplib_qp
.sq_hdr_buf_size
=
1113 BNXT_QPLIB_MAX_QP1_SQ_HDR_SIZE_V2
;
1114 qp
->qplib_qp
.dpi
= &rdev
->dpi_privileged
;
1115 rc
= bnxt_qplib_create_qp1(&rdev
->qplib_res
, &qp
->qplib_qp
);
1117 dev_err(rdev_to_dev(rdev
), "Failed to create HW QP1");
1120 /* Create a shadow QP to handle the QP1 traffic */
1121 rdev
->qp1_sqp
= bnxt_re_create_shadow_qp(pd
, &rdev
->qplib_res
,
1123 if (!rdev
->qp1_sqp
) {
1125 dev_err(rdev_to_dev(rdev
),
1126 "Failed to create Shadow QP for QP1");
1129 rdev
->sqp_ah
= bnxt_re_create_shadow_qp_ah(pd
, &rdev
->qplib_res
,
1131 if (!rdev
->sqp_ah
) {
1132 bnxt_qplib_destroy_qp(&rdev
->qplib_res
,
1133 &rdev
->qp1_sqp
->qplib_qp
);
1135 dev_err(rdev_to_dev(rdev
),
1136 "Failed to create AH entry for ShadowQP");
1141 /* Allocate 128 + 1 more than what's provided */
1142 entries
= roundup_pow_of_two(qp_init_attr
->cap
.max_send_wr
+
1143 BNXT_QPLIB_RESERVED_QP_WRS
+ 1);
1144 qp
->qplib_qp
.sq
.max_wqe
= min_t(u32
, entries
,
1145 dev_attr
->max_qp_wqes
+
1146 BNXT_QPLIB_RESERVED_QP_WRS
+ 1);
1147 qp
->qplib_qp
.sq
.q_full_delta
= BNXT_QPLIB_RESERVED_QP_WRS
+ 1;
1150 * Reserving one slot for Phantom WQE. Application can
1151 * post one extra entry in this case. But allowing this to avoid
1152 * unexpected Queue full condition
1155 qp
->qplib_qp
.sq
.q_full_delta
-= 1;
1157 qp
->qplib_qp
.max_rd_atomic
= dev_attr
->max_qp_rd_atom
;
1158 qp
->qplib_qp
.max_dest_rd_atomic
= dev_attr
->max_qp_init_rd_atom
;
1160 rc
= bnxt_re_init_user_qp(rdev
, pd
, qp
, udata
);
1164 qp
->qplib_qp
.dpi
= &rdev
->dpi_privileged
;
1167 rc
= bnxt_qplib_create_qp(&rdev
->qplib_res
, &qp
->qplib_qp
);
1169 dev_err(rdev_to_dev(rdev
), "Failed to create HW QP");
1174 qp
->ib_qp
.qp_num
= qp
->qplib_qp
.id
;
1175 spin_lock_init(&qp
->sq_lock
);
1176 spin_lock_init(&qp
->rq_lock
);
1179 struct bnxt_re_qp_resp resp
;
1181 resp
.qpid
= qp
->ib_qp
.qp_num
;
1183 rc
= ib_copy_to_udata(udata
, &resp
, sizeof(resp
));
1185 dev_err(rdev_to_dev(rdev
), "Failed to copy QP udata");
1189 INIT_LIST_HEAD(&qp
->list
);
1190 mutex_lock(&rdev
->qp_lock
);
1191 list_add_tail(&qp
->list
, &rdev
->qp_list
);
1192 atomic_inc(&rdev
->qp_count
);
1193 mutex_unlock(&rdev
->qp_lock
);
1197 bnxt_qplib_destroy_qp(&rdev
->qplib_res
, &qp
->qplib_qp
);
1203 static u8
__from_ib_qp_state(enum ib_qp_state state
)
1207 return CMDQ_MODIFY_QP_NEW_STATE_RESET
;
1209 return CMDQ_MODIFY_QP_NEW_STATE_INIT
;
1211 return CMDQ_MODIFY_QP_NEW_STATE_RTR
;
1213 return CMDQ_MODIFY_QP_NEW_STATE_RTS
;
1215 return CMDQ_MODIFY_QP_NEW_STATE_SQD
;
1217 return CMDQ_MODIFY_QP_NEW_STATE_SQE
;
1220 return CMDQ_MODIFY_QP_NEW_STATE_ERR
;
1224 static enum ib_qp_state
__to_ib_qp_state(u8 state
)
1227 case CMDQ_MODIFY_QP_NEW_STATE_RESET
:
1228 return IB_QPS_RESET
;
1229 case CMDQ_MODIFY_QP_NEW_STATE_INIT
:
1231 case CMDQ_MODIFY_QP_NEW_STATE_RTR
:
1233 case CMDQ_MODIFY_QP_NEW_STATE_RTS
:
1235 case CMDQ_MODIFY_QP_NEW_STATE_SQD
:
1237 case CMDQ_MODIFY_QP_NEW_STATE_SQE
:
1239 case CMDQ_MODIFY_QP_NEW_STATE_ERR
:
1245 static u32
__from_ib_mtu(enum ib_mtu mtu
)
1249 return CMDQ_MODIFY_QP_PATH_MTU_MTU_256
;
1251 return CMDQ_MODIFY_QP_PATH_MTU_MTU_512
;
1253 return CMDQ_MODIFY_QP_PATH_MTU_MTU_1024
;
1255 return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048
;
1257 return CMDQ_MODIFY_QP_PATH_MTU_MTU_4096
;
1259 return CMDQ_MODIFY_QP_PATH_MTU_MTU_2048
;
1263 static enum ib_mtu
__to_ib_mtu(u32 mtu
)
1265 switch (mtu
& CREQ_QUERY_QP_RESP_SB_PATH_MTU_MASK
) {
1266 case CMDQ_MODIFY_QP_PATH_MTU_MTU_256
:
1268 case CMDQ_MODIFY_QP_PATH_MTU_MTU_512
:
1270 case CMDQ_MODIFY_QP_PATH_MTU_MTU_1024
:
1272 case CMDQ_MODIFY_QP_PATH_MTU_MTU_2048
:
1274 case CMDQ_MODIFY_QP_PATH_MTU_MTU_4096
:
1281 static int bnxt_re_modify_shadow_qp(struct bnxt_re_dev
*rdev
,
1282 struct bnxt_re_qp
*qp1_qp
,
1285 struct bnxt_re_qp
*qp
= rdev
->qp1_sqp
;
1288 if (qp_attr_mask
& IB_QP_STATE
) {
1289 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_STATE
;
1290 qp
->qplib_qp
.state
= qp1_qp
->qplib_qp
.state
;
1292 if (qp_attr_mask
& IB_QP_PKEY_INDEX
) {
1293 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY
;
1294 qp
->qplib_qp
.pkey_index
= qp1_qp
->qplib_qp
.pkey_index
;
1297 if (qp_attr_mask
& IB_QP_QKEY
) {
1298 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY
;
1299 /* Using a Random QKEY */
1300 qp
->qplib_qp
.qkey
= 0x81818181;
1302 if (qp_attr_mask
& IB_QP_SQ_PSN
) {
1303 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN
;
1304 qp
->qplib_qp
.sq
.psn
= qp1_qp
->qplib_qp
.sq
.psn
;
1307 rc
= bnxt_qplib_modify_qp(&rdev
->qplib_res
, &qp
->qplib_qp
);
1309 dev_err(rdev_to_dev(rdev
),
1310 "Failed to modify Shadow QP for QP1");
1314 int bnxt_re_modify_qp(struct ib_qp
*ib_qp
, struct ib_qp_attr
*qp_attr
,
1315 int qp_attr_mask
, struct ib_udata
*udata
)
1317 struct bnxt_re_qp
*qp
= container_of(ib_qp
, struct bnxt_re_qp
, ib_qp
);
1318 struct bnxt_re_dev
*rdev
= qp
->rdev
;
1319 struct bnxt_qplib_dev_attr
*dev_attr
= &rdev
->dev_attr
;
1320 enum ib_qp_state curr_qp_state
, new_qp_state
;
1324 struct ib_gid_attr sgid_attr
;
1327 qp
->qplib_qp
.modify_flags
= 0;
1328 if (qp_attr_mask
& IB_QP_STATE
) {
1329 curr_qp_state
= __to_ib_qp_state(qp
->qplib_qp
.cur_qp_state
);
1330 new_qp_state
= qp_attr
->qp_state
;
1331 if (!ib_modify_qp_is_ok(curr_qp_state
, new_qp_state
,
1332 ib_qp
->qp_type
, qp_attr_mask
,
1333 IB_LINK_LAYER_ETHERNET
)) {
1334 dev_err(rdev_to_dev(rdev
),
1335 "Invalid attribute mask: %#x specified ",
1337 dev_err(rdev_to_dev(rdev
),
1338 "for qpn: %#x type: %#x",
1339 ib_qp
->qp_num
, ib_qp
->qp_type
);
1340 dev_err(rdev_to_dev(rdev
),
1341 "curr_qp_state=0x%x, new_qp_state=0x%x\n",
1342 curr_qp_state
, new_qp_state
);
1345 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_STATE
;
1346 qp
->qplib_qp
.state
= __from_ib_qp_state(qp_attr
->qp_state
);
1349 qp
->qplib_qp
.state
== CMDQ_MODIFY_QP_NEW_STATE_ERR
) {
1350 dev_dbg(rdev_to_dev(rdev
),
1351 "Move QP = %p to flush list\n",
1353 bnxt_qplib_add_flush_qp(&qp
->qplib_qp
);
1356 qp
->qplib_qp
.state
== CMDQ_MODIFY_QP_NEW_STATE_RESET
) {
1357 dev_dbg(rdev_to_dev(rdev
),
1358 "Move QP = %p out of flush list\n",
1360 bnxt_qplib_del_flush_qp(&qp
->qplib_qp
);
1363 if (qp_attr_mask
& IB_QP_EN_SQD_ASYNC_NOTIFY
) {
1364 qp
->qplib_qp
.modify_flags
|=
1365 CMDQ_MODIFY_QP_MODIFY_MASK_EN_SQD_ASYNC_NOTIFY
;
1366 qp
->qplib_qp
.en_sqd_async_notify
= true;
1368 if (qp_attr_mask
& IB_QP_ACCESS_FLAGS
) {
1369 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_ACCESS
;
1370 qp
->qplib_qp
.access
=
1371 __from_ib_access_flags(qp_attr
->qp_access_flags
);
1372 /* LOCAL_WRITE access must be set to allow RC receive */
1373 qp
->qplib_qp
.access
|= BNXT_QPLIB_ACCESS_LOCAL_WRITE
;
1375 if (qp_attr_mask
& IB_QP_PKEY_INDEX
) {
1376 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_PKEY
;
1377 qp
->qplib_qp
.pkey_index
= qp_attr
->pkey_index
;
1379 if (qp_attr_mask
& IB_QP_QKEY
) {
1380 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_QKEY
;
1381 qp
->qplib_qp
.qkey
= qp_attr
->qkey
;
1383 if (qp_attr_mask
& IB_QP_AV
) {
1384 const struct ib_global_route
*grh
=
1385 rdma_ah_read_grh(&qp_attr
->ah_attr
);
1387 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_DGID
|
1388 CMDQ_MODIFY_QP_MODIFY_MASK_FLOW_LABEL
|
1389 CMDQ_MODIFY_QP_MODIFY_MASK_SGID_INDEX
|
1390 CMDQ_MODIFY_QP_MODIFY_MASK_HOP_LIMIT
|
1391 CMDQ_MODIFY_QP_MODIFY_MASK_TRAFFIC_CLASS
|
1392 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_MAC
|
1393 CMDQ_MODIFY_QP_MODIFY_MASK_VLAN_ID
;
1394 memcpy(qp
->qplib_qp
.ah
.dgid
.data
, grh
->dgid
.raw
,
1395 sizeof(qp
->qplib_qp
.ah
.dgid
.data
));
1396 qp
->qplib_qp
.ah
.flow_label
= grh
->flow_label
;
1397 /* If RoCE V2 is enabled, stack will have two entries for
1398 * each GID entry. Avoiding this duplicte entry in HW. Dividing
1399 * the GID index by 2 for RoCE V2
1401 qp
->qplib_qp
.ah
.sgid_index
= grh
->sgid_index
/ 2;
1402 qp
->qplib_qp
.ah
.host_sgid_index
= grh
->sgid_index
;
1403 qp
->qplib_qp
.ah
.hop_limit
= grh
->hop_limit
;
1404 qp
->qplib_qp
.ah
.traffic_class
= grh
->traffic_class
;
1405 qp
->qplib_qp
.ah
.sl
= rdma_ah_get_sl(&qp_attr
->ah_attr
);
1406 ether_addr_copy(qp
->qplib_qp
.ah
.dmac
,
1407 qp_attr
->ah_attr
.roce
.dmac
);
1409 status
= ib_get_cached_gid(&rdev
->ibdev
, 1,
1412 if (!status
&& sgid_attr
.ndev
) {
1413 memcpy(qp
->qplib_qp
.smac
, sgid_attr
.ndev
->dev_addr
,
1415 dev_put(sgid_attr
.ndev
);
1416 nw_type
= ib_gid_to_network_type(sgid_attr
.gid_type
,
1419 case RDMA_NETWORK_IPV4
:
1420 qp
->qplib_qp
.nw_type
=
1421 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV4
;
1423 case RDMA_NETWORK_IPV6
:
1424 qp
->qplib_qp
.nw_type
=
1425 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV2_IPV6
;
1428 qp
->qplib_qp
.nw_type
=
1429 CMDQ_MODIFY_QP_NETWORK_TYPE_ROCEV1
;
1435 if (qp_attr_mask
& IB_QP_PATH_MTU
) {
1436 qp
->qplib_qp
.modify_flags
|=
1437 CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU
;
1438 qp
->qplib_qp
.path_mtu
= __from_ib_mtu(qp_attr
->path_mtu
);
1439 } else if (qp_attr
->qp_state
== IB_QPS_RTR
) {
1440 qp
->qplib_qp
.modify_flags
|=
1441 CMDQ_MODIFY_QP_MODIFY_MASK_PATH_MTU
;
1442 qp
->qplib_qp
.path_mtu
=
1443 __from_ib_mtu(iboe_get_mtu(rdev
->netdev
->mtu
));
1446 if (qp_attr_mask
& IB_QP_TIMEOUT
) {
1447 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_TIMEOUT
;
1448 qp
->qplib_qp
.timeout
= qp_attr
->timeout
;
1450 if (qp_attr_mask
& IB_QP_RETRY_CNT
) {
1451 qp
->qplib_qp
.modify_flags
|=
1452 CMDQ_MODIFY_QP_MODIFY_MASK_RETRY_CNT
;
1453 qp
->qplib_qp
.retry_cnt
= qp_attr
->retry_cnt
;
1455 if (qp_attr_mask
& IB_QP_RNR_RETRY
) {
1456 qp
->qplib_qp
.modify_flags
|=
1457 CMDQ_MODIFY_QP_MODIFY_MASK_RNR_RETRY
;
1458 qp
->qplib_qp
.rnr_retry
= qp_attr
->rnr_retry
;
1460 if (qp_attr_mask
& IB_QP_MIN_RNR_TIMER
) {
1461 qp
->qplib_qp
.modify_flags
|=
1462 CMDQ_MODIFY_QP_MODIFY_MASK_MIN_RNR_TIMER
;
1463 qp
->qplib_qp
.min_rnr_timer
= qp_attr
->min_rnr_timer
;
1465 if (qp_attr_mask
& IB_QP_RQ_PSN
) {
1466 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_RQ_PSN
;
1467 qp
->qplib_qp
.rq
.psn
= qp_attr
->rq_psn
;
1469 if (qp_attr_mask
& IB_QP_MAX_QP_RD_ATOMIC
) {
1470 qp
->qplib_qp
.modify_flags
|=
1471 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_RD_ATOMIC
;
1472 /* Cap the max_rd_atomic to device max */
1473 qp
->qplib_qp
.max_rd_atomic
= min_t(u32
, qp_attr
->max_rd_atomic
,
1474 dev_attr
->max_qp_rd_atom
);
1476 if (qp_attr_mask
& IB_QP_SQ_PSN
) {
1477 qp
->qplib_qp
.modify_flags
|= CMDQ_MODIFY_QP_MODIFY_MASK_SQ_PSN
;
1478 qp
->qplib_qp
.sq
.psn
= qp_attr
->sq_psn
;
1480 if (qp_attr_mask
& IB_QP_MAX_DEST_RD_ATOMIC
) {
1481 if (qp_attr
->max_dest_rd_atomic
>
1482 dev_attr
->max_qp_init_rd_atom
) {
1483 dev_err(rdev_to_dev(rdev
),
1484 "max_dest_rd_atomic requested%d is > dev_max%d",
1485 qp_attr
->max_dest_rd_atomic
,
1486 dev_attr
->max_qp_init_rd_atom
);
1490 qp
->qplib_qp
.modify_flags
|=
1491 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_DEST_RD_ATOMIC
;
1492 qp
->qplib_qp
.max_dest_rd_atomic
= qp_attr
->max_dest_rd_atomic
;
1494 if (qp_attr_mask
& IB_QP_CAP
) {
1495 qp
->qplib_qp
.modify_flags
|=
1496 CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SIZE
|
1497 CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SIZE
|
1498 CMDQ_MODIFY_QP_MODIFY_MASK_SQ_SGE
|
1499 CMDQ_MODIFY_QP_MODIFY_MASK_RQ_SGE
|
1500 CMDQ_MODIFY_QP_MODIFY_MASK_MAX_INLINE_DATA
;
1501 if ((qp_attr
->cap
.max_send_wr
>= dev_attr
->max_qp_wqes
) ||
1502 (qp_attr
->cap
.max_recv_wr
>= dev_attr
->max_qp_wqes
) ||
1503 (qp_attr
->cap
.max_send_sge
>= dev_attr
->max_qp_sges
) ||
1504 (qp_attr
->cap
.max_recv_sge
>= dev_attr
->max_qp_sges
) ||
1505 (qp_attr
->cap
.max_inline_data
>=
1506 dev_attr
->max_inline_data
)) {
1507 dev_err(rdev_to_dev(rdev
),
1508 "Create QP failed - max exceeded");
1511 entries
= roundup_pow_of_two(qp_attr
->cap
.max_send_wr
);
1512 qp
->qplib_qp
.sq
.max_wqe
= min_t(u32
, entries
,
1513 dev_attr
->max_qp_wqes
+ 1);
1514 qp
->qplib_qp
.sq
.q_full_delta
= qp
->qplib_qp
.sq
.max_wqe
-
1515 qp_attr
->cap
.max_send_wr
;
1517 * Reserving one slot for Phantom WQE. Some application can
1518 * post one extra entry in this case. Allowing this to avoid
1519 * unexpected Queue full condition
1521 qp
->qplib_qp
.sq
.q_full_delta
-= 1;
1522 qp
->qplib_qp
.sq
.max_sge
= qp_attr
->cap
.max_send_sge
;
1523 if (qp
->qplib_qp
.rq
.max_wqe
) {
1524 entries
= roundup_pow_of_two(qp_attr
->cap
.max_recv_wr
);
1525 qp
->qplib_qp
.rq
.max_wqe
=
1526 min_t(u32
, entries
, dev_attr
->max_qp_wqes
+ 1);
1527 qp
->qplib_qp
.rq
.q_full_delta
= qp
->qplib_qp
.rq
.max_wqe
-
1528 qp_attr
->cap
.max_recv_wr
;
1529 qp
->qplib_qp
.rq
.max_sge
= qp_attr
->cap
.max_recv_sge
;
1531 /* SRQ was used prior, just ignore the RQ caps */
1534 if (qp_attr_mask
& IB_QP_DEST_QPN
) {
1535 qp
->qplib_qp
.modify_flags
|=
1536 CMDQ_MODIFY_QP_MODIFY_MASK_DEST_QP_ID
;
1537 qp
->qplib_qp
.dest_qpn
= qp_attr
->dest_qp_num
;
1539 rc
= bnxt_qplib_modify_qp(&rdev
->qplib_res
, &qp
->qplib_qp
);
1541 dev_err(rdev_to_dev(rdev
), "Failed to modify HW QP");
1544 if (ib_qp
->qp_type
== IB_QPT_GSI
&& rdev
->qp1_sqp
)
1545 rc
= bnxt_re_modify_shadow_qp(rdev
, qp
, qp_attr_mask
);
1549 int bnxt_re_query_qp(struct ib_qp
*ib_qp
, struct ib_qp_attr
*qp_attr
,
1550 int qp_attr_mask
, struct ib_qp_init_attr
*qp_init_attr
)
1552 struct bnxt_re_qp
*qp
= container_of(ib_qp
, struct bnxt_re_qp
, ib_qp
);
1553 struct bnxt_re_dev
*rdev
= qp
->rdev
;
1554 struct bnxt_qplib_qp qplib_qp
;
1557 memset(&qplib_qp
, 0, sizeof(struct bnxt_qplib_qp
));
1558 qplib_qp
.id
= qp
->qplib_qp
.id
;
1559 qplib_qp
.ah
.host_sgid_index
= qp
->qplib_qp
.ah
.host_sgid_index
;
1561 rc
= bnxt_qplib_query_qp(&rdev
->qplib_res
, &qplib_qp
);
1563 dev_err(rdev_to_dev(rdev
), "Failed to query HW QP");
1566 qp_attr
->qp_state
= __to_ib_qp_state(qplib_qp
.state
);
1567 qp_attr
->en_sqd_async_notify
= qplib_qp
.en_sqd_async_notify
? 1 : 0;
1568 qp_attr
->qp_access_flags
= __to_ib_access_flags(qplib_qp
.access
);
1569 qp_attr
->pkey_index
= qplib_qp
.pkey_index
;
1570 qp_attr
->qkey
= qplib_qp
.qkey
;
1571 qp_attr
->ah_attr
.type
= RDMA_AH_ATTR_TYPE_ROCE
;
1572 rdma_ah_set_grh(&qp_attr
->ah_attr
, NULL
, qplib_qp
.ah
.flow_label
,
1573 qplib_qp
.ah
.host_sgid_index
,
1574 qplib_qp
.ah
.hop_limit
,
1575 qplib_qp
.ah
.traffic_class
);
1576 rdma_ah_set_dgid_raw(&qp_attr
->ah_attr
, qplib_qp
.ah
.dgid
.data
);
1577 rdma_ah_set_sl(&qp_attr
->ah_attr
, qplib_qp
.ah
.sl
);
1578 ether_addr_copy(qp_attr
->ah_attr
.roce
.dmac
, qplib_qp
.ah
.dmac
);
1579 qp_attr
->path_mtu
= __to_ib_mtu(qplib_qp
.path_mtu
);
1580 qp_attr
->timeout
= qplib_qp
.timeout
;
1581 qp_attr
->retry_cnt
= qplib_qp
.retry_cnt
;
1582 qp_attr
->rnr_retry
= qplib_qp
.rnr_retry
;
1583 qp_attr
->min_rnr_timer
= qplib_qp
.min_rnr_timer
;
1584 qp_attr
->rq_psn
= qplib_qp
.rq
.psn
;
1585 qp_attr
->max_rd_atomic
= qplib_qp
.max_rd_atomic
;
1586 qp_attr
->sq_psn
= qplib_qp
.sq
.psn
;
1587 qp_attr
->max_dest_rd_atomic
= qplib_qp
.max_dest_rd_atomic
;
1588 qp_init_attr
->sq_sig_type
= qplib_qp
.sig_type
? IB_SIGNAL_ALL_WR
:
1590 qp_attr
->dest_qp_num
= qplib_qp
.dest_qpn
;
1592 qp_attr
->cap
.max_send_wr
= qp
->qplib_qp
.sq
.max_wqe
;
1593 qp_attr
->cap
.max_send_sge
= qp
->qplib_qp
.sq
.max_sge
;
1594 qp_attr
->cap
.max_recv_wr
= qp
->qplib_qp
.rq
.max_wqe
;
1595 qp_attr
->cap
.max_recv_sge
= qp
->qplib_qp
.rq
.max_sge
;
1596 qp_attr
->cap
.max_inline_data
= qp
->qplib_qp
.max_inline_data
;
1597 qp_init_attr
->cap
= qp_attr
->cap
;
1602 /* Routine for sending QP1 packets for RoCE V1 an V2
1604 static int bnxt_re_build_qp1_send_v2(struct bnxt_re_qp
*qp
,
1605 struct ib_send_wr
*wr
,
1606 struct bnxt_qplib_swqe
*wqe
,
1609 struct ib_device
*ibdev
= &qp
->rdev
->ibdev
;
1610 struct bnxt_re_ah
*ah
= container_of(ud_wr(wr
)->ah
, struct bnxt_re_ah
,
1612 struct bnxt_qplib_ah
*qplib_ah
= &ah
->qplib_ah
;
1613 struct bnxt_qplib_sge sge
;
1617 struct ib_gid_attr sgid_attr
;
1619 bool is_eth
= false;
1620 bool is_vlan
= false;
1621 bool is_grh
= false;
1622 bool is_udp
= false;
1624 u16 vlan_id
= 0xFFFF;
1626 int i
, rc
= 0, size
;
1628 memset(&qp
->qp1_hdr
, 0, sizeof(qp
->qp1_hdr
));
1630 rc
= ib_get_cached_gid(ibdev
, 1,
1631 qplib_ah
->host_sgid_index
, &sgid
,
1634 dev_err(rdev_to_dev(qp
->rdev
),
1635 "Failed to query gid at index %d",
1636 qplib_ah
->host_sgid_index
);
1639 if (sgid_attr
.ndev
) {
1640 if (is_vlan_dev(sgid_attr
.ndev
))
1641 vlan_id
= vlan_dev_vlan_id(sgid_attr
.ndev
);
1642 dev_put(sgid_attr
.ndev
);
1644 /* Get network header type for this GID */
1645 nw_type
= ib_gid_to_network_type(sgid_attr
.gid_type
, &sgid
);
1647 case RDMA_NETWORK_IPV4
:
1648 nw_type
= BNXT_RE_ROCEV2_IPV4_PACKET
;
1650 case RDMA_NETWORK_IPV6
:
1651 nw_type
= BNXT_RE_ROCEV2_IPV6_PACKET
;
1654 nw_type
= BNXT_RE_ROCE_V1_PACKET
;
1657 memcpy(&dgid
.raw
, &qplib_ah
->dgid
, 16);
1658 is_udp
= sgid_attr
.gid_type
== IB_GID_TYPE_ROCE_UDP_ENCAP
;
1660 if (ipv6_addr_v4mapped((struct in6_addr
*)&sgid
)) {
1662 ether_type
= ETH_P_IP
;
1665 ether_type
= ETH_P_IPV6
;
1669 ether_type
= ETH_P_IBOE
;
1674 is_vlan
= (vlan_id
&& (vlan_id
< 0x1000)) ? true : false;
1676 ib_ud_header_init(payload_size
, !is_eth
, is_eth
, is_vlan
, is_grh
,
1677 ip_version
, is_udp
, 0, &qp
->qp1_hdr
);
1680 ether_addr_copy(qp
->qp1_hdr
.eth
.dmac_h
, ah
->qplib_ah
.dmac
);
1681 ether_addr_copy(qp
->qp1_hdr
.eth
.smac_h
, qp
->qplib_qp
.smac
);
1683 /* For vlan, check the sgid for vlan existence */
1686 qp
->qp1_hdr
.eth
.type
= cpu_to_be16(ether_type
);
1688 qp
->qp1_hdr
.vlan
.type
= cpu_to_be16(ether_type
);
1689 qp
->qp1_hdr
.vlan
.tag
= cpu_to_be16(vlan_id
);
1692 if (is_grh
|| (ip_version
== 6)) {
1693 memcpy(qp
->qp1_hdr
.grh
.source_gid
.raw
, sgid
.raw
, sizeof(sgid
));
1694 memcpy(qp
->qp1_hdr
.grh
.destination_gid
.raw
, qplib_ah
->dgid
.data
,
1696 qp
->qp1_hdr
.grh
.hop_limit
= qplib_ah
->hop_limit
;
1699 if (ip_version
== 4) {
1700 qp
->qp1_hdr
.ip4
.tos
= 0;
1701 qp
->qp1_hdr
.ip4
.id
= 0;
1702 qp
->qp1_hdr
.ip4
.frag_off
= htons(IP_DF
);
1703 qp
->qp1_hdr
.ip4
.ttl
= qplib_ah
->hop_limit
;
1705 memcpy(&qp
->qp1_hdr
.ip4
.saddr
, sgid
.raw
+ 12, 4);
1706 memcpy(&qp
->qp1_hdr
.ip4
.daddr
, qplib_ah
->dgid
.data
+ 12, 4);
1707 qp
->qp1_hdr
.ip4
.check
= ib_ud_ip4_csum(&qp
->qp1_hdr
);
1711 qp
->qp1_hdr
.udp
.dport
= htons(ROCE_V2_UDP_DPORT
);
1712 qp
->qp1_hdr
.udp
.sport
= htons(0x8CD1);
1713 qp
->qp1_hdr
.udp
.csum
= 0;
1717 if (wr
->opcode
== IB_WR_SEND_WITH_IMM
) {
1718 qp
->qp1_hdr
.bth
.opcode
= IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE
;
1719 qp
->qp1_hdr
.immediate_present
= 1;
1721 qp
->qp1_hdr
.bth
.opcode
= IB_OPCODE_UD_SEND_ONLY
;
1723 if (wr
->send_flags
& IB_SEND_SOLICITED
)
1724 qp
->qp1_hdr
.bth
.solicited_event
= 1;
1726 qp
->qp1_hdr
.bth
.pad_count
= (4 - payload_size
) & 3;
1728 /* P_key for QP1 is for all members */
1729 qp
->qp1_hdr
.bth
.pkey
= cpu_to_be16(0xFFFF);
1730 qp
->qp1_hdr
.bth
.destination_qpn
= IB_QP1
;
1731 qp
->qp1_hdr
.bth
.ack_req
= 0;
1733 qp
->send_psn
&= BTH_PSN_MASK
;
1734 qp
->qp1_hdr
.bth
.psn
= cpu_to_be32(qp
->send_psn
);
1736 /* Use the priviledged Q_Key for QP1 */
1737 qp
->qp1_hdr
.deth
.qkey
= cpu_to_be32(IB_QP1_QKEY
);
1738 qp
->qp1_hdr
.deth
.source_qpn
= IB_QP1
;
1740 /* Pack the QP1 to the transmit buffer */
1741 buf
= bnxt_qplib_get_qp1_sq_buf(&qp
->qplib_qp
, &sge
);
1743 size
= ib_ud_header_pack(&qp
->qp1_hdr
, buf
);
1744 for (i
= wqe
->num_sge
; i
; i
--) {
1745 wqe
->sg_list
[i
].addr
= wqe
->sg_list
[i
- 1].addr
;
1746 wqe
->sg_list
[i
].lkey
= wqe
->sg_list
[i
- 1].lkey
;
1747 wqe
->sg_list
[i
].size
= wqe
->sg_list
[i
- 1].size
;
1751 * Max Header buf size for IPV6 RoCE V2 is 86,
1752 * which is same as the QP1 SQ header buffer.
1753 * Header buf size for IPV4 RoCE V2 can be 66.
1754 * ETH(14) + VLAN(4)+ IP(20) + UDP (8) + BTH(20).
1755 * Subtract 20 bytes from QP1 SQ header buf size
1757 if (is_udp
&& ip_version
== 4)
1760 * Max Header buf size for RoCE V1 is 78.
1761 * ETH(14) + VLAN(4) + GRH(40) + BTH(20).
1762 * Subtract 8 bytes from QP1 SQ header buf size
1767 /* Subtract 4 bytes for non vlan packets */
1771 wqe
->sg_list
[0].addr
= sge
.addr
;
1772 wqe
->sg_list
[0].lkey
= sge
.lkey
;
1773 wqe
->sg_list
[0].size
= sge
.size
;
1777 dev_err(rdev_to_dev(qp
->rdev
), "QP1 buffer is empty!");
1783 /* For the MAD layer, it only provides the recv SGE the size of
1784 * ib_grh + MAD datagram. No Ethernet headers, Ethertype, BTH, DETH,
1785 * nor RoCE iCRC. The Cu+ solution must provide buffer for the entire
1786 * receive packet (334 bytes) with no VLAN and then copy the GRH
1787 * and the MAD datagram out to the provided SGE.
1789 static int bnxt_re_build_qp1_shadow_qp_recv(struct bnxt_re_qp
*qp
,
1790 struct ib_recv_wr
*wr
,
1791 struct bnxt_qplib_swqe
*wqe
,
1794 struct bnxt_qplib_sge ref
, sge
;
1796 struct bnxt_re_sqp_entries
*sqp_entry
;
1798 rq_prod_index
= bnxt_qplib_get_rq_prod_index(&qp
->qplib_qp
);
1800 if (!bnxt_qplib_get_qp1_rq_buf(&qp
->qplib_qp
, &sge
))
1803 /* Create 1 SGE to receive the entire
1806 /* Save the reference from ULP */
1807 ref
.addr
= wqe
->sg_list
[0].addr
;
1808 ref
.lkey
= wqe
->sg_list
[0].lkey
;
1809 ref
.size
= wqe
->sg_list
[0].size
;
1811 sqp_entry
= &qp
->rdev
->sqp_tbl
[rq_prod_index
];
1814 wqe
->sg_list
[0].addr
= sge
.addr
;
1815 wqe
->sg_list
[0].lkey
= sge
.lkey
;
1816 wqe
->sg_list
[0].size
= BNXT_QPLIB_MAX_QP1_RQ_HDR_SIZE_V2
;
1817 sge
.size
-= wqe
->sg_list
[0].size
;
1819 sqp_entry
->sge
.addr
= ref
.addr
;
1820 sqp_entry
->sge
.lkey
= ref
.lkey
;
1821 sqp_entry
->sge
.size
= ref
.size
;
1822 /* Store the wrid for reporting completion */
1823 sqp_entry
->wrid
= wqe
->wr_id
;
1824 /* change the wqe->wrid to table index */
1825 wqe
->wr_id
= rq_prod_index
;
1829 static int is_ud_qp(struct bnxt_re_qp
*qp
)
1831 return qp
->qplib_qp
.type
== CMDQ_CREATE_QP_TYPE_UD
;
1834 static int bnxt_re_build_send_wqe(struct bnxt_re_qp
*qp
,
1835 struct ib_send_wr
*wr
,
1836 struct bnxt_qplib_swqe
*wqe
)
1838 struct bnxt_re_ah
*ah
= NULL
;
1841 ah
= container_of(ud_wr(wr
)->ah
, struct bnxt_re_ah
, ib_ah
);
1842 wqe
->send
.q_key
= ud_wr(wr
)->remote_qkey
;
1843 wqe
->send
.dst_qp
= ud_wr(wr
)->remote_qpn
;
1844 wqe
->send
.avid
= ah
->qplib_ah
.id
;
1846 switch (wr
->opcode
) {
1848 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_SEND
;
1850 case IB_WR_SEND_WITH_IMM
:
1851 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM
;
1852 wqe
->send
.imm_data
= wr
->ex
.imm_data
;
1854 case IB_WR_SEND_WITH_INV
:
1855 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV
;
1856 wqe
->send
.inv_key
= wr
->ex
.invalidate_rkey
;
1861 if (wr
->send_flags
& IB_SEND_SIGNALED
)
1862 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP
;
1863 if (wr
->send_flags
& IB_SEND_FENCE
)
1864 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE
;
1865 if (wr
->send_flags
& IB_SEND_SOLICITED
)
1866 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT
;
1867 if (wr
->send_flags
& IB_SEND_INLINE
)
1868 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_INLINE
;
1873 static int bnxt_re_build_rdma_wqe(struct ib_send_wr
*wr
,
1874 struct bnxt_qplib_swqe
*wqe
)
1876 switch (wr
->opcode
) {
1877 case IB_WR_RDMA_WRITE
:
1878 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE
;
1880 case IB_WR_RDMA_WRITE_WITH_IMM
:
1881 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM
;
1882 wqe
->rdma
.imm_data
= wr
->ex
.imm_data
;
1884 case IB_WR_RDMA_READ
:
1885 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_RDMA_READ
;
1886 wqe
->rdma
.inv_key
= wr
->ex
.invalidate_rkey
;
1891 wqe
->rdma
.remote_va
= rdma_wr(wr
)->remote_addr
;
1892 wqe
->rdma
.r_key
= rdma_wr(wr
)->rkey
;
1893 if (wr
->send_flags
& IB_SEND_SIGNALED
)
1894 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP
;
1895 if (wr
->send_flags
& IB_SEND_FENCE
)
1896 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE
;
1897 if (wr
->send_flags
& IB_SEND_SOLICITED
)
1898 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT
;
1899 if (wr
->send_flags
& IB_SEND_INLINE
)
1900 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_INLINE
;
1905 static int bnxt_re_build_atomic_wqe(struct ib_send_wr
*wr
,
1906 struct bnxt_qplib_swqe
*wqe
)
1908 switch (wr
->opcode
) {
1909 case IB_WR_ATOMIC_CMP_AND_SWP
:
1910 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP
;
1911 wqe
->atomic
.swap_data
= atomic_wr(wr
)->swap
;
1913 case IB_WR_ATOMIC_FETCH_AND_ADD
:
1914 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD
;
1915 wqe
->atomic
.cmp_data
= atomic_wr(wr
)->compare_add
;
1920 wqe
->atomic
.remote_va
= atomic_wr(wr
)->remote_addr
;
1921 wqe
->atomic
.r_key
= atomic_wr(wr
)->rkey
;
1922 if (wr
->send_flags
& IB_SEND_SIGNALED
)
1923 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP
;
1924 if (wr
->send_flags
& IB_SEND_FENCE
)
1925 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE
;
1926 if (wr
->send_flags
& IB_SEND_SOLICITED
)
1927 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT
;
1931 static int bnxt_re_build_inv_wqe(struct ib_send_wr
*wr
,
1932 struct bnxt_qplib_swqe
*wqe
)
1934 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_LOCAL_INV
;
1935 wqe
->local_inv
.inv_l_key
= wr
->ex
.invalidate_rkey
;
1937 if (wr
->send_flags
& IB_SEND_SIGNALED
)
1938 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP
;
1939 if (wr
->send_flags
& IB_SEND_FENCE
)
1940 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE
;
1941 if (wr
->send_flags
& IB_SEND_SOLICITED
)
1942 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SOLICIT_EVENT
;
1947 static int bnxt_re_build_reg_wqe(struct ib_reg_wr
*wr
,
1948 struct bnxt_qplib_swqe
*wqe
)
1950 struct bnxt_re_mr
*mr
= container_of(wr
->mr
, struct bnxt_re_mr
, ib_mr
);
1951 struct bnxt_qplib_frpl
*qplib_frpl
= &mr
->qplib_frpl
;
1952 int access
= wr
->access
;
1954 wqe
->frmr
.pbl_ptr
= (__le64
*)qplib_frpl
->hwq
.pbl_ptr
[0];
1955 wqe
->frmr
.pbl_dma_ptr
= qplib_frpl
->hwq
.pbl_dma_ptr
[0];
1956 wqe
->frmr
.page_list
= mr
->pages
;
1957 wqe
->frmr
.page_list_len
= mr
->npages
;
1958 wqe
->frmr
.levels
= qplib_frpl
->hwq
.level
+ 1;
1959 wqe
->type
= BNXT_QPLIB_SWQE_TYPE_REG_MR
;
1961 if (wr
->wr
.send_flags
& IB_SEND_FENCE
)
1962 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_UC_FENCE
;
1963 if (wr
->wr
.send_flags
& IB_SEND_SIGNALED
)
1964 wqe
->flags
|= BNXT_QPLIB_SWQE_FLAGS_SIGNAL_COMP
;
1966 if (access
& IB_ACCESS_LOCAL_WRITE
)
1967 wqe
->frmr
.access_cntl
|= SQ_FR_PMR_ACCESS_CNTL_LOCAL_WRITE
;
1968 if (access
& IB_ACCESS_REMOTE_READ
)
1969 wqe
->frmr
.access_cntl
|= SQ_FR_PMR_ACCESS_CNTL_REMOTE_READ
;
1970 if (access
& IB_ACCESS_REMOTE_WRITE
)
1971 wqe
->frmr
.access_cntl
|= SQ_FR_PMR_ACCESS_CNTL_REMOTE_WRITE
;
1972 if (access
& IB_ACCESS_REMOTE_ATOMIC
)
1973 wqe
->frmr
.access_cntl
|= SQ_FR_PMR_ACCESS_CNTL_REMOTE_ATOMIC
;
1974 if (access
& IB_ACCESS_MW_BIND
)
1975 wqe
->frmr
.access_cntl
|= SQ_FR_PMR_ACCESS_CNTL_WINDOW_BIND
;
1977 wqe
->frmr
.l_key
= wr
->key
;
1978 wqe
->frmr
.length
= wr
->mr
->length
;
1979 wqe
->frmr
.pbl_pg_sz_log
= (wr
->mr
->page_size
>> PAGE_SHIFT_4K
) - 1;
1980 wqe
->frmr
.va
= wr
->mr
->iova
;
1984 static int bnxt_re_copy_inline_data(struct bnxt_re_dev
*rdev
,
1985 struct ib_send_wr
*wr
,
1986 struct bnxt_qplib_swqe
*wqe
)
1988 /* Copy the inline data to the data field */
1993 in_data
= wqe
->inline_data
;
1994 for (i
= 0; i
< wr
->num_sge
; i
++) {
1995 sge_addr
= (void *)(unsigned long)
1996 wr
->sg_list
[i
].addr
;
1997 sge_len
= wr
->sg_list
[i
].length
;
1999 if ((sge_len
+ wqe
->inline_len
) >
2000 BNXT_QPLIB_SWQE_MAX_INLINE_LENGTH
) {
2001 dev_err(rdev_to_dev(rdev
),
2002 "Inline data size requested > supported value");
2005 sge_len
= wr
->sg_list
[i
].length
;
2007 memcpy(in_data
, sge_addr
, sge_len
);
2008 in_data
+= wr
->sg_list
[i
].length
;
2009 wqe
->inline_len
+= wr
->sg_list
[i
].length
;
2011 return wqe
->inline_len
;
2014 static int bnxt_re_copy_wr_payload(struct bnxt_re_dev
*rdev
,
2015 struct ib_send_wr
*wr
,
2016 struct bnxt_qplib_swqe
*wqe
)
2020 if (wr
->send_flags
& IB_SEND_INLINE
)
2021 payload_sz
= bnxt_re_copy_inline_data(rdev
, wr
, wqe
);
2023 payload_sz
= bnxt_re_build_sgl(wr
->sg_list
, wqe
->sg_list
,
2029 static void bnxt_ud_qp_hw_stall_workaround(struct bnxt_re_qp
*qp
)
2031 if ((qp
->ib_qp
.qp_type
== IB_QPT_UD
||
2032 qp
->ib_qp
.qp_type
== IB_QPT_GSI
||
2033 qp
->ib_qp
.qp_type
== IB_QPT_RAW_ETHERTYPE
) &&
2034 qp
->qplib_qp
.wqe_cnt
== BNXT_RE_UD_QP_HW_STALL
) {
2036 struct ib_qp_attr qp_attr
;
2038 qp_attr_mask
= IB_QP_STATE
;
2039 qp_attr
.qp_state
= IB_QPS_RTS
;
2040 bnxt_re_modify_qp(&qp
->ib_qp
, &qp_attr
, qp_attr_mask
, NULL
);
2041 qp
->qplib_qp
.wqe_cnt
= 0;
2045 static int bnxt_re_post_send_shadow_qp(struct bnxt_re_dev
*rdev
,
2046 struct bnxt_re_qp
*qp
,
2047 struct ib_send_wr
*wr
)
2049 struct bnxt_qplib_swqe wqe
;
2050 int rc
= 0, payload_sz
= 0;
2051 unsigned long flags
;
2053 spin_lock_irqsave(&qp
->sq_lock
, flags
);
2054 memset(&wqe
, 0, sizeof(wqe
));
2057 memset(&wqe
, 0, sizeof(wqe
));
2060 wqe
.num_sge
= wr
->num_sge
;
2061 if (wr
->num_sge
> qp
->qplib_qp
.sq
.max_sge
) {
2062 dev_err(rdev_to_dev(rdev
),
2063 "Limit exceeded for Send SGEs");
2068 payload_sz
= bnxt_re_copy_wr_payload(qp
->rdev
, wr
, &wqe
);
2069 if (payload_sz
< 0) {
2073 wqe
.wr_id
= wr
->wr_id
;
2075 wqe
.type
= BNXT_QPLIB_SWQE_TYPE_SEND
;
2077 rc
= bnxt_re_build_send_wqe(qp
, wr
, &wqe
);
2079 rc
= bnxt_qplib_post_send(&qp
->qplib_qp
, &wqe
);
2082 dev_err(rdev_to_dev(rdev
),
2083 "Post send failed opcode = %#x rc = %d",
2089 bnxt_qplib_post_send_db(&qp
->qplib_qp
);
2090 bnxt_ud_qp_hw_stall_workaround(qp
);
2091 spin_unlock_irqrestore(&qp
->sq_lock
, flags
);
2095 int bnxt_re_post_send(struct ib_qp
*ib_qp
, struct ib_send_wr
*wr
,
2096 struct ib_send_wr
**bad_wr
)
2098 struct bnxt_re_qp
*qp
= container_of(ib_qp
, struct bnxt_re_qp
, ib_qp
);
2099 struct bnxt_qplib_swqe wqe
;
2100 int rc
= 0, payload_sz
= 0;
2101 unsigned long flags
;
2103 spin_lock_irqsave(&qp
->sq_lock
, flags
);
2106 memset(&wqe
, 0, sizeof(wqe
));
2109 wqe
.num_sge
= wr
->num_sge
;
2110 if (wr
->num_sge
> qp
->qplib_qp
.sq
.max_sge
) {
2111 dev_err(rdev_to_dev(qp
->rdev
),
2112 "Limit exceeded for Send SGEs");
2117 payload_sz
= bnxt_re_copy_wr_payload(qp
->rdev
, wr
, &wqe
);
2118 if (payload_sz
< 0) {
2122 wqe
.wr_id
= wr
->wr_id
;
2124 switch (wr
->opcode
) {
2126 case IB_WR_SEND_WITH_IMM
:
2127 if (ib_qp
->qp_type
== IB_QPT_GSI
) {
2128 rc
= bnxt_re_build_qp1_send_v2(qp
, wr
, &wqe
,
2132 wqe
.rawqp1
.lflags
|=
2133 SQ_SEND_RAWETH_QP1_LFLAGS_ROCE_CRC
;
2135 switch (wr
->send_flags
) {
2136 case IB_SEND_IP_CSUM
:
2137 wqe
.rawqp1
.lflags
|=
2138 SQ_SEND_RAWETH_QP1_LFLAGS_IP_CHKSUM
;
2143 /* Fall thru to build the wqe */
2144 case IB_WR_SEND_WITH_INV
:
2145 rc
= bnxt_re_build_send_wqe(qp
, wr
, &wqe
);
2147 case IB_WR_RDMA_WRITE
:
2148 case IB_WR_RDMA_WRITE_WITH_IMM
:
2149 case IB_WR_RDMA_READ
:
2150 rc
= bnxt_re_build_rdma_wqe(wr
, &wqe
);
2152 case IB_WR_ATOMIC_CMP_AND_SWP
:
2153 case IB_WR_ATOMIC_FETCH_AND_ADD
:
2154 rc
= bnxt_re_build_atomic_wqe(wr
, &wqe
);
2156 case IB_WR_RDMA_READ_WITH_INV
:
2157 dev_err(rdev_to_dev(qp
->rdev
),
2158 "RDMA Read with Invalidate is not supported");
2161 case IB_WR_LOCAL_INV
:
2162 rc
= bnxt_re_build_inv_wqe(wr
, &wqe
);
2165 rc
= bnxt_re_build_reg_wqe(reg_wr(wr
), &wqe
);
2168 /* Unsupported WRs */
2169 dev_err(rdev_to_dev(qp
->rdev
),
2170 "WR (%#x) is not supported", wr
->opcode
);
2175 rc
= bnxt_qplib_post_send(&qp
->qplib_qp
, &wqe
);
2178 dev_err(rdev_to_dev(qp
->rdev
),
2179 "post_send failed op:%#x qps = %#x rc = %d\n",
2180 wr
->opcode
, qp
->qplib_qp
.state
, rc
);
2186 bnxt_qplib_post_send_db(&qp
->qplib_qp
);
2187 bnxt_ud_qp_hw_stall_workaround(qp
);
2188 spin_unlock_irqrestore(&qp
->sq_lock
, flags
);
2193 static int bnxt_re_post_recv_shadow_qp(struct bnxt_re_dev
*rdev
,
2194 struct bnxt_re_qp
*qp
,
2195 struct ib_recv_wr
*wr
)
2197 struct bnxt_qplib_swqe wqe
;
2198 int rc
= 0, payload_sz
= 0;
2200 memset(&wqe
, 0, sizeof(wqe
));
2203 memset(&wqe
, 0, sizeof(wqe
));
2206 wqe
.num_sge
= wr
->num_sge
;
2207 if (wr
->num_sge
> qp
->qplib_qp
.rq
.max_sge
) {
2208 dev_err(rdev_to_dev(rdev
),
2209 "Limit exceeded for Receive SGEs");
2213 payload_sz
= bnxt_re_build_sgl(wr
->sg_list
, wqe
.sg_list
,
2215 wqe
.wr_id
= wr
->wr_id
;
2216 wqe
.type
= BNXT_QPLIB_SWQE_TYPE_RECV
;
2218 rc
= bnxt_qplib_post_recv(&qp
->qplib_qp
, &wqe
);
2225 bnxt_qplib_post_recv_db(&qp
->qplib_qp
);
2229 int bnxt_re_post_recv(struct ib_qp
*ib_qp
, struct ib_recv_wr
*wr
,
2230 struct ib_recv_wr
**bad_wr
)
2232 struct bnxt_re_qp
*qp
= container_of(ib_qp
, struct bnxt_re_qp
, ib_qp
);
2233 struct bnxt_qplib_swqe wqe
;
2234 int rc
= 0, payload_sz
= 0;
2235 unsigned long flags
;
2238 spin_lock_irqsave(&qp
->rq_lock
, flags
);
2241 memset(&wqe
, 0, sizeof(wqe
));
2244 wqe
.num_sge
= wr
->num_sge
;
2245 if (wr
->num_sge
> qp
->qplib_qp
.rq
.max_sge
) {
2246 dev_err(rdev_to_dev(qp
->rdev
),
2247 "Limit exceeded for Receive SGEs");
2253 payload_sz
= bnxt_re_build_sgl(wr
->sg_list
, wqe
.sg_list
,
2255 wqe
.wr_id
= wr
->wr_id
;
2256 wqe
.type
= BNXT_QPLIB_SWQE_TYPE_RECV
;
2258 if (ib_qp
->qp_type
== IB_QPT_GSI
)
2259 rc
= bnxt_re_build_qp1_shadow_qp_recv(qp
, wr
, &wqe
,
2262 rc
= bnxt_qplib_post_recv(&qp
->qplib_qp
, &wqe
);
2268 /* Ring DB if the RQEs posted reaches a threshold value */
2269 if (++count
>= BNXT_RE_RQ_WQE_THRESHOLD
) {
2270 bnxt_qplib_post_recv_db(&qp
->qplib_qp
);
2278 bnxt_qplib_post_recv_db(&qp
->qplib_qp
);
2280 spin_unlock_irqrestore(&qp
->rq_lock
, flags
);
2285 /* Completion Queues */
2286 int bnxt_re_destroy_cq(struct ib_cq
*ib_cq
)
2288 struct bnxt_re_cq
*cq
= container_of(ib_cq
, struct bnxt_re_cq
, ib_cq
);
2289 struct bnxt_re_dev
*rdev
= cq
->rdev
;
2291 struct bnxt_qplib_nq
*nq
= cq
->qplib_cq
.nq
;
2293 rc
= bnxt_qplib_destroy_cq(&rdev
->qplib_res
, &cq
->qplib_cq
);
2295 dev_err(rdev_to_dev(rdev
), "Failed to destroy HW CQ");
2298 if (!IS_ERR_OR_NULL(cq
->umem
))
2299 ib_umem_release(cq
->umem
);
2305 atomic_dec(&rdev
->cq_count
);
2310 struct ib_cq
*bnxt_re_create_cq(struct ib_device
*ibdev
,
2311 const struct ib_cq_init_attr
*attr
,
2312 struct ib_ucontext
*context
,
2313 struct ib_udata
*udata
)
2315 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
2316 struct bnxt_qplib_dev_attr
*dev_attr
= &rdev
->dev_attr
;
2317 struct bnxt_re_cq
*cq
= NULL
;
2319 int cqe
= attr
->cqe
;
2320 struct bnxt_qplib_nq
*nq
= NULL
;
2321 unsigned int nq_alloc_cnt
;
2323 /* Validate CQ fields */
2324 if (cqe
< 1 || cqe
> dev_attr
->max_cq_wqes
) {
2325 dev_err(rdev_to_dev(rdev
), "Failed to create CQ -max exceeded");
2326 return ERR_PTR(-EINVAL
);
2328 cq
= kzalloc(sizeof(*cq
), GFP_KERNEL
);
2330 return ERR_PTR(-ENOMEM
);
2333 cq
->qplib_cq
.cq_handle
= (u64
)(unsigned long)(&cq
->qplib_cq
);
2335 entries
= roundup_pow_of_two(cqe
+ 1);
2336 if (entries
> dev_attr
->max_cq_wqes
+ 1)
2337 entries
= dev_attr
->max_cq_wqes
+ 1;
2340 struct bnxt_re_cq_req req
;
2341 struct bnxt_re_ucontext
*uctx
= container_of
2343 struct bnxt_re_ucontext
,
2345 if (ib_copy_from_udata(&req
, udata
, sizeof(req
))) {
2350 cq
->umem
= ib_umem_get(context
, req
.cq_va
,
2351 entries
* sizeof(struct cq_base
),
2352 IB_ACCESS_LOCAL_WRITE
, 1);
2353 if (IS_ERR(cq
->umem
)) {
2354 rc
= PTR_ERR(cq
->umem
);
2357 cq
->qplib_cq
.sghead
= cq
->umem
->sg_head
.sgl
;
2358 cq
->qplib_cq
.nmap
= cq
->umem
->nmap
;
2359 cq
->qplib_cq
.dpi
= &uctx
->dpi
;
2361 cq
->max_cql
= min_t(u32
, entries
, MAX_CQL_PER_POLL
);
2362 cq
->cql
= kcalloc(cq
->max_cql
, sizeof(struct bnxt_qplib_cqe
),
2369 cq
->qplib_cq
.dpi
= &rdev
->dpi_privileged
;
2370 cq
->qplib_cq
.sghead
= NULL
;
2371 cq
->qplib_cq
.nmap
= 0;
2374 * Allocating the NQ in a round robin fashion. nq_alloc_cnt is a
2375 * used for getting the NQ index.
2377 nq_alloc_cnt
= atomic_inc_return(&rdev
->nq_alloc_cnt
);
2378 nq
= &rdev
->nq
[nq_alloc_cnt
% (rdev
->num_msix
- 1)];
2379 cq
->qplib_cq
.max_wqe
= entries
;
2380 cq
->qplib_cq
.cnq_hw_ring_id
= nq
->ring_id
;
2381 cq
->qplib_cq
.nq
= nq
;
2383 rc
= bnxt_qplib_create_cq(&rdev
->qplib_res
, &cq
->qplib_cq
);
2385 dev_err(rdev_to_dev(rdev
), "Failed to create HW CQ");
2389 cq
->ib_cq
.cqe
= entries
;
2390 cq
->cq_period
= cq
->qplib_cq
.period
;
2393 atomic_inc(&rdev
->cq_count
);
2396 struct bnxt_re_cq_resp resp
;
2398 resp
.cqid
= cq
->qplib_cq
.id
;
2399 resp
.tail
= cq
->qplib_cq
.hwq
.cons
;
2400 resp
.phase
= cq
->qplib_cq
.period
;
2402 rc
= ib_copy_to_udata(udata
, &resp
, sizeof(resp
));
2404 dev_err(rdev_to_dev(rdev
), "Failed to copy CQ udata");
2405 bnxt_qplib_destroy_cq(&rdev
->qplib_res
, &cq
->qplib_cq
);
2414 ib_umem_release(cq
->umem
);
2421 static u8
__req_to_ib_wc_status(u8 qstatus
)
2424 case CQ_REQ_STATUS_OK
:
2425 return IB_WC_SUCCESS
;
2426 case CQ_REQ_STATUS_BAD_RESPONSE_ERR
:
2427 return IB_WC_BAD_RESP_ERR
;
2428 case CQ_REQ_STATUS_LOCAL_LENGTH_ERR
:
2429 return IB_WC_LOC_LEN_ERR
;
2430 case CQ_REQ_STATUS_LOCAL_QP_OPERATION_ERR
:
2431 return IB_WC_LOC_QP_OP_ERR
;
2432 case CQ_REQ_STATUS_LOCAL_PROTECTION_ERR
:
2433 return IB_WC_LOC_PROT_ERR
;
2434 case CQ_REQ_STATUS_MEMORY_MGT_OPERATION_ERR
:
2435 return IB_WC_GENERAL_ERR
;
2436 case CQ_REQ_STATUS_REMOTE_INVALID_REQUEST_ERR
:
2437 return IB_WC_REM_INV_REQ_ERR
;
2438 case CQ_REQ_STATUS_REMOTE_ACCESS_ERR
:
2439 return IB_WC_REM_ACCESS_ERR
;
2440 case CQ_REQ_STATUS_REMOTE_OPERATION_ERR
:
2441 return IB_WC_REM_OP_ERR
;
2442 case CQ_REQ_STATUS_RNR_NAK_RETRY_CNT_ERR
:
2443 return IB_WC_RNR_RETRY_EXC_ERR
;
2444 case CQ_REQ_STATUS_TRANSPORT_RETRY_CNT_ERR
:
2445 return IB_WC_RETRY_EXC_ERR
;
2446 case CQ_REQ_STATUS_WORK_REQUEST_FLUSHED_ERR
:
2447 return IB_WC_WR_FLUSH_ERR
;
2449 return IB_WC_GENERAL_ERR
;
2454 static u8
__rawqp1_to_ib_wc_status(u8 qstatus
)
2457 case CQ_RES_RAWETH_QP1_STATUS_OK
:
2458 return IB_WC_SUCCESS
;
2459 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_ACCESS_ERROR
:
2460 return IB_WC_LOC_ACCESS_ERR
;
2461 case CQ_RES_RAWETH_QP1_STATUS_HW_LOCAL_LENGTH_ERR
:
2462 return IB_WC_LOC_LEN_ERR
;
2463 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_PROTECTION_ERR
:
2464 return IB_WC_LOC_PROT_ERR
;
2465 case CQ_RES_RAWETH_QP1_STATUS_LOCAL_QP_OPERATION_ERR
:
2466 return IB_WC_LOC_QP_OP_ERR
;
2467 case CQ_RES_RAWETH_QP1_STATUS_MEMORY_MGT_OPERATION_ERR
:
2468 return IB_WC_GENERAL_ERR
;
2469 case CQ_RES_RAWETH_QP1_STATUS_WORK_REQUEST_FLUSHED_ERR
:
2470 return IB_WC_WR_FLUSH_ERR
;
2471 case CQ_RES_RAWETH_QP1_STATUS_HW_FLUSH_ERR
:
2472 return IB_WC_WR_FLUSH_ERR
;
2474 return IB_WC_GENERAL_ERR
;
2478 static u8
__rc_to_ib_wc_status(u8 qstatus
)
2481 case CQ_RES_RC_STATUS_OK
:
2482 return IB_WC_SUCCESS
;
2483 case CQ_RES_RC_STATUS_LOCAL_ACCESS_ERROR
:
2484 return IB_WC_LOC_ACCESS_ERR
;
2485 case CQ_RES_RC_STATUS_LOCAL_LENGTH_ERR
:
2486 return IB_WC_LOC_LEN_ERR
;
2487 case CQ_RES_RC_STATUS_LOCAL_PROTECTION_ERR
:
2488 return IB_WC_LOC_PROT_ERR
;
2489 case CQ_RES_RC_STATUS_LOCAL_QP_OPERATION_ERR
:
2490 return IB_WC_LOC_QP_OP_ERR
;
2491 case CQ_RES_RC_STATUS_MEMORY_MGT_OPERATION_ERR
:
2492 return IB_WC_GENERAL_ERR
;
2493 case CQ_RES_RC_STATUS_REMOTE_INVALID_REQUEST_ERR
:
2494 return IB_WC_REM_INV_REQ_ERR
;
2495 case CQ_RES_RC_STATUS_WORK_REQUEST_FLUSHED_ERR
:
2496 return IB_WC_WR_FLUSH_ERR
;
2497 case CQ_RES_RC_STATUS_HW_FLUSH_ERR
:
2498 return IB_WC_WR_FLUSH_ERR
;
2500 return IB_WC_GENERAL_ERR
;
2504 static void bnxt_re_process_req_wc(struct ib_wc
*wc
, struct bnxt_qplib_cqe
*cqe
)
2506 switch (cqe
->type
) {
2507 case BNXT_QPLIB_SWQE_TYPE_SEND
:
2508 wc
->opcode
= IB_WC_SEND
;
2510 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_IMM
:
2511 wc
->opcode
= IB_WC_SEND
;
2512 wc
->wc_flags
|= IB_WC_WITH_IMM
;
2514 case BNXT_QPLIB_SWQE_TYPE_SEND_WITH_INV
:
2515 wc
->opcode
= IB_WC_SEND
;
2516 wc
->wc_flags
|= IB_WC_WITH_INVALIDATE
;
2518 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE
:
2519 wc
->opcode
= IB_WC_RDMA_WRITE
;
2521 case BNXT_QPLIB_SWQE_TYPE_RDMA_WRITE_WITH_IMM
:
2522 wc
->opcode
= IB_WC_RDMA_WRITE
;
2523 wc
->wc_flags
|= IB_WC_WITH_IMM
;
2525 case BNXT_QPLIB_SWQE_TYPE_RDMA_READ
:
2526 wc
->opcode
= IB_WC_RDMA_READ
;
2528 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_CMP_AND_SWP
:
2529 wc
->opcode
= IB_WC_COMP_SWAP
;
2531 case BNXT_QPLIB_SWQE_TYPE_ATOMIC_FETCH_AND_ADD
:
2532 wc
->opcode
= IB_WC_FETCH_ADD
;
2534 case BNXT_QPLIB_SWQE_TYPE_LOCAL_INV
:
2535 wc
->opcode
= IB_WC_LOCAL_INV
;
2537 case BNXT_QPLIB_SWQE_TYPE_REG_MR
:
2538 wc
->opcode
= IB_WC_REG_MR
;
2541 wc
->opcode
= IB_WC_SEND
;
2545 wc
->status
= __req_to_ib_wc_status(cqe
->status
);
2548 static int bnxt_re_check_packet_type(u16 raweth_qp1_flags
,
2549 u16 raweth_qp1_flags2
)
2551 bool is_udp
= false, is_ipv6
= false, is_ipv4
= false;
2553 /* raweth_qp1_flags Bit 9-6 indicates itype */
2554 if ((raweth_qp1_flags
& CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE
)
2555 != CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS_ITYPE_ROCE
)
2558 if (raweth_qp1_flags2
&
2559 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_CS_CALC
&&
2561 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_L4_CS_CALC
) {
2563 /* raweth_qp1_flags2 Bit 8 indicates ip_type. 0-v4 1 - v6 */
2564 (raweth_qp1_flags2
&
2565 CQ_RES_RAWETH_QP1_RAWETH_QP1_FLAGS2_IP_TYPE
) ?
2566 (is_ipv6
= true) : (is_ipv4
= true);
2568 BNXT_RE_ROCEV2_IPV6_PACKET
:
2569 BNXT_RE_ROCEV2_IPV4_PACKET
);
2571 return BNXT_RE_ROCE_V1_PACKET
;
2575 static int bnxt_re_to_ib_nw_type(int nw_type
)
2577 u8 nw_hdr_type
= 0xFF;
2580 case BNXT_RE_ROCE_V1_PACKET
:
2581 nw_hdr_type
= RDMA_NETWORK_ROCE_V1
;
2583 case BNXT_RE_ROCEV2_IPV4_PACKET
:
2584 nw_hdr_type
= RDMA_NETWORK_IPV4
;
2586 case BNXT_RE_ROCEV2_IPV6_PACKET
:
2587 nw_hdr_type
= RDMA_NETWORK_IPV6
;
2593 static bool bnxt_re_is_loopback_packet(struct bnxt_re_dev
*rdev
,
2597 struct ethhdr
*eth_hdr
;
2601 tmp_buf
= (u8
*)rq_hdr_buf
;
2603 * If dest mac is not same as I/F mac, this could be a
2604 * loopback address or multicast address, check whether
2605 * it is a loopback packet
2607 if (!ether_addr_equal(tmp_buf
, rdev
->netdev
->dev_addr
)) {
2609 /* Check the ether type */
2610 eth_hdr
= (struct ethhdr
*)tmp_buf
;
2611 eth_type
= ntohs(eth_hdr
->h_proto
);
2619 struct udphdr
*udp_hdr
;
2621 len
= (eth_type
== ETH_P_IP
? sizeof(struct iphdr
) :
2622 sizeof(struct ipv6hdr
));
2623 tmp_buf
+= sizeof(struct ethhdr
) + len
;
2624 udp_hdr
= (struct udphdr
*)tmp_buf
;
2625 if (ntohs(udp_hdr
->dest
) ==
2638 static int bnxt_re_process_raw_qp_pkt_rx(struct bnxt_re_qp
*qp1_qp
,
2639 struct bnxt_qplib_cqe
*cqe
)
2641 struct bnxt_re_dev
*rdev
= qp1_qp
->rdev
;
2642 struct bnxt_re_sqp_entries
*sqp_entry
= NULL
;
2643 struct bnxt_re_qp
*qp
= rdev
->qp1_sqp
;
2644 struct ib_send_wr
*swr
;
2645 struct ib_ud_wr udwr
;
2646 struct ib_recv_wr rwr
;
2650 dma_addr_t rq_hdr_buf_map
;
2651 dma_addr_t shrq_hdr_buf_map
;
2654 struct ib_sge s_sge
[2];
2655 struct ib_sge r_sge
[2];
2658 memset(&udwr
, 0, sizeof(udwr
));
2659 memset(&rwr
, 0, sizeof(rwr
));
2660 memset(&s_sge
, 0, sizeof(s_sge
));
2661 memset(&r_sge
, 0, sizeof(r_sge
));
2664 tbl_idx
= cqe
->wr_id
;
2666 rq_hdr_buf
= qp1_qp
->qplib_qp
.rq_hdr_buf
+
2667 (tbl_idx
* qp1_qp
->qplib_qp
.rq_hdr_buf_size
);
2668 rq_hdr_buf_map
= bnxt_qplib_get_qp_buf_from_index(&qp1_qp
->qplib_qp
,
2671 /* Shadow QP header buffer */
2672 shrq_hdr_buf_map
= bnxt_qplib_get_qp_buf_from_index(&qp
->qplib_qp
,
2674 sqp_entry
= &rdev
->sqp_tbl
[tbl_idx
];
2676 /* Store this cqe */
2677 memcpy(&sqp_entry
->cqe
, cqe
, sizeof(struct bnxt_qplib_cqe
));
2678 sqp_entry
->qp1_qp
= qp1_qp
;
2680 /* Find packet type from the cqe */
2682 pkt_type
= bnxt_re_check_packet_type(cqe
->raweth_qp1_flags
,
2683 cqe
->raweth_qp1_flags2
);
2685 dev_err(rdev_to_dev(rdev
), "Invalid packet\n");
2689 /* Adjust the offset for the user buffer and post in the rq */
2691 if (pkt_type
== BNXT_RE_ROCEV2_IPV4_PACKET
)
2695 * QP1 loopback packet has 4 bytes of internal header before
2696 * ether header. Skip these four bytes.
2698 if (bnxt_re_is_loopback_packet(rdev
, rq_hdr_buf
))
2701 /* First send SGE . Skip the ether header*/
2702 s_sge
[0].addr
= rq_hdr_buf_map
+ BNXT_QPLIB_MAX_QP1_RQ_ETH_HDR_SIZE
2704 s_sge
[0].lkey
= 0xFFFFFFFF;
2705 s_sge
[0].length
= offset
? BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV4
:
2706 BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6
;
2708 /* Second Send SGE */
2709 s_sge
[1].addr
= s_sge
[0].addr
+ s_sge
[0].length
+
2710 BNXT_QPLIB_MAX_QP1_RQ_BDETH_HDR_SIZE
;
2711 if (pkt_type
!= BNXT_RE_ROCE_V1_PACKET
)
2713 s_sge
[1].lkey
= 0xFFFFFFFF;
2714 s_sge
[1].length
= 256;
2716 /* First recv SGE */
2718 r_sge
[0].addr
= shrq_hdr_buf_map
;
2719 r_sge
[0].lkey
= 0xFFFFFFFF;
2720 r_sge
[0].length
= 40;
2722 r_sge
[1].addr
= sqp_entry
->sge
.addr
+ offset
;
2723 r_sge
[1].lkey
= sqp_entry
->sge
.lkey
;
2724 r_sge
[1].length
= BNXT_QPLIB_MAX_GRH_HDR_SIZE_IPV6
+ 256 - offset
;
2726 /* Create receive work request */
2728 rwr
.sg_list
= r_sge
;
2729 rwr
.wr_id
= tbl_idx
;
2732 rc
= bnxt_re_post_recv_shadow_qp(rdev
, qp
, &rwr
);
2734 dev_err(rdev_to_dev(rdev
),
2735 "Failed to post Rx buffers to shadow QP");
2740 swr
->sg_list
= s_sge
;
2741 swr
->wr_id
= tbl_idx
;
2742 swr
->opcode
= IB_WR_SEND
;
2745 udwr
.ah
= &rdev
->sqp_ah
->ib_ah
;
2746 udwr
.remote_qpn
= rdev
->qp1_sqp
->qplib_qp
.id
;
2747 udwr
.remote_qkey
= rdev
->qp1_sqp
->qplib_qp
.qkey
;
2749 /* post data received in the send queue */
2750 rc
= bnxt_re_post_send_shadow_qp(rdev
, qp
, swr
);
2755 static void bnxt_re_process_res_rawqp1_wc(struct ib_wc
*wc
,
2756 struct bnxt_qplib_cqe
*cqe
)
2758 wc
->opcode
= IB_WC_RECV
;
2759 wc
->status
= __rawqp1_to_ib_wc_status(cqe
->status
);
2760 wc
->wc_flags
|= IB_WC_GRH
;
2763 static void bnxt_re_process_res_rc_wc(struct ib_wc
*wc
,
2764 struct bnxt_qplib_cqe
*cqe
)
2766 wc
->opcode
= IB_WC_RECV
;
2767 wc
->status
= __rc_to_ib_wc_status(cqe
->status
);
2769 if (cqe
->flags
& CQ_RES_RC_FLAGS_IMM
)
2770 wc
->wc_flags
|= IB_WC_WITH_IMM
;
2771 if (cqe
->flags
& CQ_RES_RC_FLAGS_INV
)
2772 wc
->wc_flags
|= IB_WC_WITH_INVALIDATE
;
2773 if ((cqe
->flags
& (CQ_RES_RC_FLAGS_RDMA
| CQ_RES_RC_FLAGS_IMM
)) ==
2774 (CQ_RES_RC_FLAGS_RDMA
| CQ_RES_RC_FLAGS_IMM
))
2775 wc
->opcode
= IB_WC_RECV_RDMA_WITH_IMM
;
2778 static void bnxt_re_process_res_shadow_qp_wc(struct bnxt_re_qp
*qp
,
2780 struct bnxt_qplib_cqe
*cqe
)
2783 struct bnxt_re_dev
*rdev
= qp
->rdev
;
2784 struct bnxt_re_qp
*qp1_qp
= NULL
;
2785 struct bnxt_qplib_cqe
*orig_cqe
= NULL
;
2786 struct bnxt_re_sqp_entries
*sqp_entry
= NULL
;
2789 tbl_idx
= cqe
->wr_id
;
2791 sqp_entry
= &rdev
->sqp_tbl
[tbl_idx
];
2792 qp1_qp
= sqp_entry
->qp1_qp
;
2793 orig_cqe
= &sqp_entry
->cqe
;
2795 wc
->wr_id
= sqp_entry
->wrid
;
2796 wc
->byte_len
= orig_cqe
->length
;
2797 wc
->qp
= &qp1_qp
->ib_qp
;
2799 wc
->ex
.imm_data
= orig_cqe
->immdata
;
2800 wc
->src_qp
= orig_cqe
->src_qp
;
2801 memcpy(wc
->smac
, orig_cqe
->smac
, ETH_ALEN
);
2803 wc
->vendor_err
= orig_cqe
->status
;
2805 wc
->opcode
= IB_WC_RECV
;
2806 wc
->status
= __rawqp1_to_ib_wc_status(orig_cqe
->status
);
2807 wc
->wc_flags
|= IB_WC_GRH
;
2809 nw_type
= bnxt_re_check_packet_type(orig_cqe
->raweth_qp1_flags
,
2810 orig_cqe
->raweth_qp1_flags2
);
2812 wc
->network_hdr_type
= bnxt_re_to_ib_nw_type(nw_type
);
2813 wc
->wc_flags
|= IB_WC_WITH_NETWORK_HDR_TYPE
;
2817 static void bnxt_re_process_res_ud_wc(struct ib_wc
*wc
,
2818 struct bnxt_qplib_cqe
*cqe
)
2820 wc
->opcode
= IB_WC_RECV
;
2821 wc
->status
= __rc_to_ib_wc_status(cqe
->status
);
2823 if (cqe
->flags
& CQ_RES_RC_FLAGS_IMM
)
2824 wc
->wc_flags
|= IB_WC_WITH_IMM
;
2825 if (cqe
->flags
& CQ_RES_RC_FLAGS_INV
)
2826 wc
->wc_flags
|= IB_WC_WITH_INVALIDATE
;
2827 if ((cqe
->flags
& (CQ_RES_RC_FLAGS_RDMA
| CQ_RES_RC_FLAGS_IMM
)) ==
2828 (CQ_RES_RC_FLAGS_RDMA
| CQ_RES_RC_FLAGS_IMM
))
2829 wc
->opcode
= IB_WC_RECV_RDMA_WITH_IMM
;
2832 static int send_phantom_wqe(struct bnxt_re_qp
*qp
)
2834 struct bnxt_qplib_qp
*lib_qp
= &qp
->qplib_qp
;
2835 unsigned long flags
;
2838 spin_lock_irqsave(&qp
->sq_lock
, flags
);
2840 rc
= bnxt_re_bind_fence_mw(lib_qp
);
2842 lib_qp
->sq
.phantom_wqe_cnt
++;
2843 dev_dbg(&lib_qp
->sq
.hwq
.pdev
->dev
,
2844 "qp %#x sq->prod %#x sw_prod %#x phantom_wqe_cnt %d\n",
2845 lib_qp
->id
, lib_qp
->sq
.hwq
.prod
,
2846 HWQ_CMP(lib_qp
->sq
.hwq
.prod
, &lib_qp
->sq
.hwq
),
2847 lib_qp
->sq
.phantom_wqe_cnt
);
2850 spin_unlock_irqrestore(&qp
->sq_lock
, flags
);
2854 int bnxt_re_poll_cq(struct ib_cq
*ib_cq
, int num_entries
, struct ib_wc
*wc
)
2856 struct bnxt_re_cq
*cq
= container_of(ib_cq
, struct bnxt_re_cq
, ib_cq
);
2857 struct bnxt_re_qp
*qp
;
2858 struct bnxt_qplib_cqe
*cqe
;
2859 int i
, ncqe
, budget
;
2860 struct bnxt_qplib_q
*sq
;
2861 struct bnxt_qplib_qp
*lib_qp
;
2863 struct bnxt_re_sqp_entries
*sqp_entry
= NULL
;
2864 unsigned long flags
;
2866 spin_lock_irqsave(&cq
->cq_lock
, flags
);
2867 budget
= min_t(u32
, num_entries
, cq
->max_cql
);
2868 num_entries
= budget
;
2870 dev_err(rdev_to_dev(cq
->rdev
), "POLL CQ : no CQL to use");
2876 ncqe
= bnxt_qplib_poll_cq(&cq
->qplib_cq
, cqe
, budget
, &lib_qp
);
2879 if (sq
->send_phantom
) {
2880 qp
= container_of(lib_qp
,
2881 struct bnxt_re_qp
, qplib_qp
);
2882 if (send_phantom_wqe(qp
) == -ENOMEM
)
2883 dev_err(rdev_to_dev(cq
->rdev
),
2884 "Phantom failed! Scheduled to send again\n");
2886 sq
->send_phantom
= false;
2890 ncqe
+= bnxt_qplib_process_flush_list(&cq
->qplib_cq
,
2897 for (i
= 0; i
< ncqe
; i
++, cqe
++) {
2898 /* Transcribe each qplib_wqe back to ib_wc */
2899 memset(wc
, 0, sizeof(*wc
));
2901 wc
->wr_id
= cqe
->wr_id
;
2902 wc
->byte_len
= cqe
->length
;
2904 ((struct bnxt_qplib_qp
*)
2905 (unsigned long)(cqe
->qp_handle
),
2906 struct bnxt_re_qp
, qplib_qp
);
2908 dev_err(rdev_to_dev(cq
->rdev
),
2909 "POLL CQ : bad QP handle");
2912 wc
->qp
= &qp
->ib_qp
;
2913 wc
->ex
.imm_data
= cqe
->immdata
;
2914 wc
->src_qp
= cqe
->src_qp
;
2915 memcpy(wc
->smac
, cqe
->smac
, ETH_ALEN
);
2917 wc
->vendor_err
= cqe
->status
;
2919 switch (cqe
->opcode
) {
2920 case CQ_BASE_CQE_TYPE_REQ
:
2921 if (qp
->qplib_qp
.id
==
2922 qp
->rdev
->qp1_sqp
->qplib_qp
.id
) {
2923 /* Handle this completion with
2924 * the stored completion
2926 memset(wc
, 0, sizeof(*wc
));
2929 bnxt_re_process_req_wc(wc
, cqe
);
2931 case CQ_BASE_CQE_TYPE_RES_RAWETH_QP1
:
2935 rc
= bnxt_re_process_raw_qp_pkt_rx
2938 memset(wc
, 0, sizeof(*wc
));
2943 /* Errors need not be looped back.
2944 * But change the wr_id to the one
2945 * stored in the table
2947 tbl_idx
= cqe
->wr_id
;
2948 sqp_entry
= &cq
->rdev
->sqp_tbl
[tbl_idx
];
2949 wc
->wr_id
= sqp_entry
->wrid
;
2950 bnxt_re_process_res_rawqp1_wc(wc
, cqe
);
2952 case CQ_BASE_CQE_TYPE_RES_RC
:
2953 bnxt_re_process_res_rc_wc(wc
, cqe
);
2955 case CQ_BASE_CQE_TYPE_RES_UD
:
2956 if (qp
->qplib_qp
.id
==
2957 qp
->rdev
->qp1_sqp
->qplib_qp
.id
) {
2958 /* Handle this completion with
2959 * the stored completion
2964 bnxt_re_process_res_shadow_qp_wc
2969 bnxt_re_process_res_ud_wc(wc
, cqe
);
2972 dev_err(rdev_to_dev(cq
->rdev
),
2973 "POLL CQ : type 0x%x not handled",
2982 spin_unlock_irqrestore(&cq
->cq_lock
, flags
);
2983 return num_entries
- budget
;
2986 int bnxt_re_req_notify_cq(struct ib_cq
*ib_cq
,
2987 enum ib_cq_notify_flags ib_cqn_flags
)
2989 struct bnxt_re_cq
*cq
= container_of(ib_cq
, struct bnxt_re_cq
, ib_cq
);
2992 /* Trigger on the very next completion */
2993 if (ib_cqn_flags
& IB_CQ_NEXT_COMP
)
2994 type
= DBR_DBR_TYPE_CQ_ARMALL
;
2995 /* Trigger on the next solicited completion */
2996 else if (ib_cqn_flags
& IB_CQ_SOLICITED
)
2997 type
= DBR_DBR_TYPE_CQ_ARMSE
;
2999 /* Poll to see if there are missed events */
3000 if ((ib_cqn_flags
& IB_CQ_REPORT_MISSED_EVENTS
) &&
3001 !(bnxt_qplib_is_cq_empty(&cq
->qplib_cq
)))
3004 bnxt_qplib_req_notify_cq(&cq
->qplib_cq
, type
);
3009 /* Memory Regions */
3010 struct ib_mr
*bnxt_re_get_dma_mr(struct ib_pd
*ib_pd
, int mr_access_flags
)
3012 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
3013 struct bnxt_re_dev
*rdev
= pd
->rdev
;
3014 struct bnxt_re_mr
*mr
;
3018 mr
= kzalloc(sizeof(*mr
), GFP_KERNEL
);
3020 return ERR_PTR(-ENOMEM
);
3023 mr
->qplib_mr
.pd
= &pd
->qplib_pd
;
3024 mr
->qplib_mr
.flags
= __from_ib_access_flags(mr_access_flags
);
3025 mr
->qplib_mr
.type
= CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR
;
3027 /* Allocate and register 0 as the address */
3028 rc
= bnxt_qplib_alloc_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
3032 mr
->qplib_mr
.hwq
.level
= PBL_LVL_MAX
;
3033 mr
->qplib_mr
.total_size
= -1; /* Infinte length */
3034 rc
= bnxt_qplib_reg_mr(&rdev
->qplib_res
, &mr
->qplib_mr
, &pbl
, 0, false);
3038 mr
->ib_mr
.lkey
= mr
->qplib_mr
.lkey
;
3039 if (mr_access_flags
& (IB_ACCESS_REMOTE_WRITE
| IB_ACCESS_REMOTE_READ
|
3040 IB_ACCESS_REMOTE_ATOMIC
))
3041 mr
->ib_mr
.rkey
= mr
->ib_mr
.lkey
;
3042 atomic_inc(&rdev
->mr_count
);
3047 bnxt_qplib_free_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
3053 int bnxt_re_dereg_mr(struct ib_mr
*ib_mr
)
3055 struct bnxt_re_mr
*mr
= container_of(ib_mr
, struct bnxt_re_mr
, ib_mr
);
3056 struct bnxt_re_dev
*rdev
= mr
->rdev
;
3059 rc
= bnxt_qplib_free_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
3061 dev_err(rdev_to_dev(rdev
), "Dereg MR failed: %#x\n", rc
);
3065 if (mr
->npages
&& mr
->pages
) {
3066 rc
= bnxt_qplib_free_fast_reg_page_list(&rdev
->qplib_res
,
3072 if (!IS_ERR_OR_NULL(mr
->ib_umem
))
3073 ib_umem_release(mr
->ib_umem
);
3076 atomic_dec(&rdev
->mr_count
);
3080 static int bnxt_re_set_page(struct ib_mr
*ib_mr
, u64 addr
)
3082 struct bnxt_re_mr
*mr
= container_of(ib_mr
, struct bnxt_re_mr
, ib_mr
);
3084 if (unlikely(mr
->npages
== mr
->qplib_frpl
.max_pg_ptrs
))
3087 mr
->pages
[mr
->npages
++] = addr
;
3091 int bnxt_re_map_mr_sg(struct ib_mr
*ib_mr
, struct scatterlist
*sg
, int sg_nents
,
3092 unsigned int *sg_offset
)
3094 struct bnxt_re_mr
*mr
= container_of(ib_mr
, struct bnxt_re_mr
, ib_mr
);
3097 return ib_sg_to_pages(ib_mr
, sg
, sg_nents
, sg_offset
, bnxt_re_set_page
);
3100 struct ib_mr
*bnxt_re_alloc_mr(struct ib_pd
*ib_pd
, enum ib_mr_type type
,
3103 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
3104 struct bnxt_re_dev
*rdev
= pd
->rdev
;
3105 struct bnxt_re_mr
*mr
= NULL
;
3108 if (type
!= IB_MR_TYPE_MEM_REG
) {
3109 dev_dbg(rdev_to_dev(rdev
), "MR type 0x%x not supported", type
);
3110 return ERR_PTR(-EINVAL
);
3112 if (max_num_sg
> MAX_PBL_LVL_1_PGS
)
3113 return ERR_PTR(-EINVAL
);
3115 mr
= kzalloc(sizeof(*mr
), GFP_KERNEL
);
3117 return ERR_PTR(-ENOMEM
);
3120 mr
->qplib_mr
.pd
= &pd
->qplib_pd
;
3121 mr
->qplib_mr
.flags
= BNXT_QPLIB_FR_PMR
;
3122 mr
->qplib_mr
.type
= CMDQ_ALLOCATE_MRW_MRW_FLAGS_PMR
;
3124 rc
= bnxt_qplib_alloc_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
3128 mr
->ib_mr
.lkey
= mr
->qplib_mr
.lkey
;
3129 mr
->ib_mr
.rkey
= mr
->ib_mr
.lkey
;
3131 mr
->pages
= kcalloc(max_num_sg
, sizeof(u64
), GFP_KERNEL
);
3136 rc
= bnxt_qplib_alloc_fast_reg_page_list(&rdev
->qplib_res
,
3137 &mr
->qplib_frpl
, max_num_sg
);
3139 dev_err(rdev_to_dev(rdev
),
3140 "Failed to allocate HW FR page list");
3144 atomic_inc(&rdev
->mr_count
);
3148 bnxt_qplib_free_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
3155 struct ib_mw
*bnxt_re_alloc_mw(struct ib_pd
*ib_pd
, enum ib_mw_type type
,
3156 struct ib_udata
*udata
)
3158 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
3159 struct bnxt_re_dev
*rdev
= pd
->rdev
;
3160 struct bnxt_re_mw
*mw
;
3163 mw
= kzalloc(sizeof(*mw
), GFP_KERNEL
);
3165 return ERR_PTR(-ENOMEM
);
3167 mw
->qplib_mw
.pd
= &pd
->qplib_pd
;
3169 mw
->qplib_mw
.type
= (type
== IB_MW_TYPE_1
?
3170 CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE1
:
3171 CMDQ_ALLOCATE_MRW_MRW_FLAGS_MW_TYPE2B
);
3172 rc
= bnxt_qplib_alloc_mrw(&rdev
->qplib_res
, &mw
->qplib_mw
);
3174 dev_err(rdev_to_dev(rdev
), "Allocate MW failed!");
3177 mw
->ib_mw
.rkey
= mw
->qplib_mw
.rkey
;
3179 atomic_inc(&rdev
->mw_count
);
3187 int bnxt_re_dealloc_mw(struct ib_mw
*ib_mw
)
3189 struct bnxt_re_mw
*mw
= container_of(ib_mw
, struct bnxt_re_mw
, ib_mw
);
3190 struct bnxt_re_dev
*rdev
= mw
->rdev
;
3193 rc
= bnxt_qplib_free_mrw(&rdev
->qplib_res
, &mw
->qplib_mw
);
3195 dev_err(rdev_to_dev(rdev
), "Free MW failed: %#x\n", rc
);
3200 atomic_dec(&rdev
->mw_count
);
3205 struct ib_mr
*bnxt_re_reg_user_mr(struct ib_pd
*ib_pd
, u64 start
, u64 length
,
3206 u64 virt_addr
, int mr_access_flags
,
3207 struct ib_udata
*udata
)
3209 struct bnxt_re_pd
*pd
= container_of(ib_pd
, struct bnxt_re_pd
, ib_pd
);
3210 struct bnxt_re_dev
*rdev
= pd
->rdev
;
3211 struct bnxt_re_mr
*mr
;
3212 struct ib_umem
*umem
;
3213 u64
*pbl_tbl
, *pbl_tbl_orig
;
3214 int i
, umem_pgs
, pages
, rc
;
3215 struct scatterlist
*sg
;
3218 if (length
> BNXT_RE_MAX_MR_SIZE
) {
3219 dev_err(rdev_to_dev(rdev
), "MR Size: %lld > Max supported:%ld\n",
3220 length
, BNXT_RE_MAX_MR_SIZE
);
3221 return ERR_PTR(-ENOMEM
);
3224 mr
= kzalloc(sizeof(*mr
), GFP_KERNEL
);
3226 return ERR_PTR(-ENOMEM
);
3229 mr
->qplib_mr
.pd
= &pd
->qplib_pd
;
3230 mr
->qplib_mr
.flags
= __from_ib_access_flags(mr_access_flags
);
3231 mr
->qplib_mr
.type
= CMDQ_ALLOCATE_MRW_MRW_FLAGS_MR
;
3233 umem
= ib_umem_get(ib_pd
->uobject
->context
, start
, length
,
3234 mr_access_flags
, 0);
3236 dev_err(rdev_to_dev(rdev
), "Failed to get umem");
3242 rc
= bnxt_qplib_alloc_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
3244 dev_err(rdev_to_dev(rdev
), "Failed to allocate MR");
3247 /* The fixed portion of the rkey is the same as the lkey */
3248 mr
->ib_mr
.rkey
= mr
->qplib_mr
.rkey
;
3250 mr
->qplib_mr
.va
= virt_addr
;
3251 umem_pgs
= ib_umem_page_count(umem
);
3253 dev_err(rdev_to_dev(rdev
), "umem is invalid!");
3257 mr
->qplib_mr
.total_size
= length
;
3259 pbl_tbl
= kcalloc(umem_pgs
, sizeof(u64
*), GFP_KERNEL
);
3264 pbl_tbl_orig
= pbl_tbl
;
3266 if (umem
->hugetlb
) {
3267 dev_err(rdev_to_dev(rdev
), "umem hugetlb not supported!");
3272 if (umem
->page_shift
!= PAGE_SHIFT
) {
3273 dev_err(rdev_to_dev(rdev
), "umem page shift unsupported!");
3277 /* Map umem buf ptrs to the PBL */
3278 for_each_sg(umem
->sg_head
.sgl
, sg
, umem
->nmap
, entry
) {
3279 pages
= sg_dma_len(sg
) >> umem
->page_shift
;
3280 for (i
= 0; i
< pages
; i
++, pbl_tbl
++)
3281 *pbl_tbl
= sg_dma_address(sg
) + (i
<< umem
->page_shift
);
3283 rc
= bnxt_qplib_reg_mr(&rdev
->qplib_res
, &mr
->qplib_mr
, pbl_tbl_orig
,
3286 dev_err(rdev_to_dev(rdev
), "Failed to register user MR");
3290 kfree(pbl_tbl_orig
);
3292 mr
->ib_mr
.lkey
= mr
->qplib_mr
.lkey
;
3293 mr
->ib_mr
.rkey
= mr
->qplib_mr
.lkey
;
3294 atomic_inc(&rdev
->mr_count
);
3298 kfree(pbl_tbl_orig
);
3300 bnxt_qplib_free_mrw(&rdev
->qplib_res
, &mr
->qplib_mr
);
3302 ib_umem_release(umem
);
3308 struct ib_ucontext
*bnxt_re_alloc_ucontext(struct ib_device
*ibdev
,
3309 struct ib_udata
*udata
)
3311 struct bnxt_re_dev
*rdev
= to_bnxt_re_dev(ibdev
, ibdev
);
3312 struct bnxt_re_uctx_resp resp
;
3313 struct bnxt_re_ucontext
*uctx
;
3314 struct bnxt_qplib_dev_attr
*dev_attr
= &rdev
->dev_attr
;
3317 dev_dbg(rdev_to_dev(rdev
), "ABI version requested %d",
3318 ibdev
->uverbs_abi_ver
);
3320 if (ibdev
->uverbs_abi_ver
!= BNXT_RE_ABI_VERSION
) {
3321 dev_dbg(rdev_to_dev(rdev
), " is different from the device %d ",
3322 BNXT_RE_ABI_VERSION
);
3323 return ERR_PTR(-EPERM
);
3326 uctx
= kzalloc(sizeof(*uctx
), GFP_KERNEL
);
3328 return ERR_PTR(-ENOMEM
);
3332 uctx
->shpg
= (void *)__get_free_page(GFP_KERNEL
);
3337 spin_lock_init(&uctx
->sh_lock
);
3339 resp
.dev_id
= rdev
->en_dev
->pdev
->devfn
; /*Temp, Use idr_alloc instead*/
3340 resp
.max_qp
= rdev
->qplib_ctx
.qpc_count
;
3341 resp
.pg_size
= PAGE_SIZE
;
3342 resp
.cqe_sz
= sizeof(struct cq_base
);
3343 resp
.max_cqd
= dev_attr
->max_cq_wqes
;
3346 rc
= ib_copy_to_udata(udata
, &resp
, sizeof(resp
));
3348 dev_err(rdev_to_dev(rdev
), "Failed to copy user context");
3353 return &uctx
->ib_uctx
;
3355 free_page((unsigned long)uctx
->shpg
);
3362 int bnxt_re_dealloc_ucontext(struct ib_ucontext
*ib_uctx
)
3364 struct bnxt_re_ucontext
*uctx
= container_of(ib_uctx
,
3365 struct bnxt_re_ucontext
,
3368 struct bnxt_re_dev
*rdev
= uctx
->rdev
;
3372 free_page((unsigned long)uctx
->shpg
);
3374 if (uctx
->dpi
.dbr
) {
3375 /* Free DPI only if this is the first PD allocated by the
3376 * application and mark the context dpi as NULL
3378 rc
= bnxt_qplib_dealloc_dpi(&rdev
->qplib_res
,
3379 &rdev
->qplib_res
.dpi_tbl
,
3382 dev_err(rdev_to_dev(rdev
), "Deallocate HW DPI failed!");
3383 /* Don't fail, continue*/
3384 uctx
->dpi
.dbr
= NULL
;
3391 /* Helper function to mmap the virtual memory from user app */
3392 int bnxt_re_mmap(struct ib_ucontext
*ib_uctx
, struct vm_area_struct
*vma
)
3394 struct bnxt_re_ucontext
*uctx
= container_of(ib_uctx
,
3395 struct bnxt_re_ucontext
,
3397 struct bnxt_re_dev
*rdev
= uctx
->rdev
;
3400 if (vma
->vm_end
- vma
->vm_start
!= PAGE_SIZE
)
3403 if (vma
->vm_pgoff
) {
3404 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
3405 if (io_remap_pfn_range(vma
, vma
->vm_start
, vma
->vm_pgoff
,
3406 PAGE_SIZE
, vma
->vm_page_prot
)) {
3407 dev_err(rdev_to_dev(rdev
), "Failed to map DPI");
3411 pfn
= virt_to_phys(uctx
->shpg
) >> PAGE_SHIFT
;
3412 if (remap_pfn_range(vma
, vma
->vm_start
,
3413 pfn
, PAGE_SIZE
, vma
->vm_page_prot
)) {
3414 dev_err(rdev_to_dev(rdev
),
3415 "Failed to map shared page");