2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 #include <linux/log2.h>
35 #include <linux/slab.h>
36 #include <linux/netdevice.h>
37 #include <linux/bitops.h>
38 #include <linux/rcupdate.h>
39 #include <linux/etherdevice.h>
41 #include <rdma/ib_cache.h>
42 #include <rdma/ib_pack.h>
43 #include <rdma/ib_addr.h>
44 #include <rdma/ib_mad.h>
46 #include <dev/mlx4/cmd.h>
47 #include <dev/mlx4/qp.h>
48 #include <dev/mlx4/driver.h>
52 #include <rdma/mlx4-abi.h>
54 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq,
55 struct mlx4_ib_cq *recv_cq);
56 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq,
57 struct mlx4_ib_cq *recv_cq);
60 MLX4_IB_ACK_REQ_FREQ = 8,
64 MLX4_IB_DEFAULT_SCHED_QUEUE = 0x83,
65 MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
66 MLX4_IB_LINK_TYPE_IB = 0,
67 MLX4_IB_LINK_TYPE_ETH = 1
72 * Largest possible UD header: send with GRH and immediate
73 * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
74 * tag. (LRH would only use 8 bytes, so Ethernet is the
77 MLX4_IB_UD_HEADER_SIZE = 82,
78 MLX4_IB_LSO_HEADER_SPARE = 128,
82 MLX4_IB_IBOE_ETHERTYPE = 0x8915
90 struct ib_ud_header ud_header;
91 u8 header_buf[MLX4_IB_UD_HEADER_SIZE];
92 struct ib_qp *roce_v2_gsi;
96 MLX4_IB_MIN_SQ_STRIDE = 6,
97 MLX4_IB_CACHE_LINE_SIZE = 64,
102 MLX4_RAW_QP_MSGMAX = 31,
109 static const __be32 mlx4_ib_opcode[] = {
110 [IB_WR_SEND] = cpu_to_be32(MLX4_OPCODE_SEND),
111 [IB_WR_LSO] = cpu_to_be32(MLX4_OPCODE_LSO),
112 [IB_WR_SEND_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_SEND_IMM),
113 [IB_WR_RDMA_WRITE] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
114 [IB_WR_RDMA_WRITE_WITH_IMM] = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
115 [IB_WR_RDMA_READ] = cpu_to_be32(MLX4_OPCODE_RDMA_READ),
116 [IB_WR_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
117 [IB_WR_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
118 [IB_WR_SEND_WITH_INV] = cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
119 [IB_WR_LOCAL_INV] = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
120 [IB_WR_REG_MR] = cpu_to_be32(MLX4_OPCODE_FMR),
121 [IB_WR_MASKED_ATOMIC_CMP_AND_SWP] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
122 [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD] = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
125 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
127 return container_of(mqp, struct mlx4_ib_sqp, qp);
130 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
132 if (!mlx4_is_master(dev->dev))
135 return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
136 qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
140 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
145 /* PPF or Native -- real SQP */
146 real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
147 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
148 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
151 /* VF or PF -- proxy SQP */
152 if (mlx4_is_mfunc(dev->dev)) {
153 for (i = 0; i < dev->dev->caps.num_ports; i++) {
154 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
155 qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
164 return !!(qp->flags & MLX4_IB_ROCE_V2_GSI_QP);
167 /* used for INIT/CLOSE port logic */
168 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
173 /* PPF or Native -- real QP0 */
174 real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
175 qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
176 qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
179 /* VF or PF -- proxy QP0 */
180 if (mlx4_is_mfunc(dev->dev)) {
181 for (i = 0; i < dev->dev->caps.num_ports; i++) {
182 if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
191 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
193 return mlx4_buf_offset(&qp->buf, offset);
196 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
198 return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
201 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
203 return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
207 * Stamp a SQ WQE so that it is invalid if prefetched by marking the
208 * first four bytes of every 64 byte chunk with
209 * 0x7FFFFFF | (invalid_ownership_value << 31).
211 * When the max work request size is less than or equal to the WQE
212 * basic block size, as an optimization, we can stamp all WQEs with
213 * 0xffffffff, and skip the very first chunk of each WQE.
215 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
223 struct mlx4_wqe_ctrl_seg *ctrl;
225 if (qp->sq_max_wqes_per_wr > 1) {
226 s = roundup(size, 1U << qp->sq.wqe_shift);
227 for (i = 0; i < s; i += 64) {
228 ind = (i >> qp->sq.wqe_shift) + n;
229 stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
230 cpu_to_be32(0xffffffff);
231 buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
232 wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
236 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
237 s = (ctrl->fence_size & 0x3f) << 4;
238 for (i = 64; i < s; i += 64) {
240 *wqe = cpu_to_be32(0xffffffff);
245 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
247 struct mlx4_wqe_ctrl_seg *ctrl;
248 struct mlx4_wqe_inline_seg *inl;
252 ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
253 s = sizeof(struct mlx4_wqe_ctrl_seg);
255 if (qp->ibqp.qp_type == IB_QPT_UD) {
256 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
257 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
258 memset(dgram, 0, sizeof *dgram);
259 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
260 s += sizeof(struct mlx4_wqe_datagram_seg);
263 /* Pad the remainder of the WQE with an inline data segment. */
266 inl->byte_count = cpu_to_be32(1U << 31 | (size - s - sizeof *inl));
268 ctrl->srcrb_flags = 0;
269 ctrl->fence_size = size / 16;
271 * Make sure descriptor is fully written before setting ownership bit
272 * (because HW can start executing as soon as we do).
276 ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
277 (n & qp->sq.wqe_cnt ? cpu_to_be32(1U << 31) : 0);
279 stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
282 /* Post NOP WQE to prevent wrap-around in the middle of WR */
283 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
285 unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
286 if (unlikely(s < qp->sq_max_wqes_per_wr)) {
287 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
293 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
295 struct ib_event event;
296 struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
298 if (type == MLX4_EVENT_TYPE_PATH_MIG)
299 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
301 if (ibqp->event_handler) {
302 event.device = ibqp->device;
303 event.element.qp = ibqp;
305 case MLX4_EVENT_TYPE_PATH_MIG:
306 event.event = IB_EVENT_PATH_MIG;
308 case MLX4_EVENT_TYPE_COMM_EST:
309 event.event = IB_EVENT_COMM_EST;
311 case MLX4_EVENT_TYPE_SQ_DRAINED:
312 event.event = IB_EVENT_SQ_DRAINED;
314 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
315 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
317 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
318 event.event = IB_EVENT_QP_FATAL;
320 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
321 event.event = IB_EVENT_PATH_MIG_ERR;
323 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
324 event.event = IB_EVENT_QP_REQ_ERR;
326 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
327 event.event = IB_EVENT_QP_ACCESS_ERR;
330 pr_warn("Unexpected event type %d "
331 "on QP %06x\n", type, qp->qpn);
335 ibqp->event_handler(&event, ibqp->qp_context);
339 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
342 * UD WQEs must have a datagram segment.
343 * RC and UC WQEs might have a remote address segment.
344 * MLX WQEs need two extra inline data segments (for the UD
345 * header and space for the ICRC).
349 return sizeof (struct mlx4_wqe_ctrl_seg) +
350 sizeof (struct mlx4_wqe_datagram_seg) +
351 ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
352 case MLX4_IB_QPT_PROXY_SMI_OWNER:
353 case MLX4_IB_QPT_PROXY_SMI:
354 case MLX4_IB_QPT_PROXY_GSI:
355 return sizeof (struct mlx4_wqe_ctrl_seg) +
356 sizeof (struct mlx4_wqe_datagram_seg) + 64;
357 case MLX4_IB_QPT_TUN_SMI_OWNER:
358 case MLX4_IB_QPT_TUN_GSI:
359 return sizeof (struct mlx4_wqe_ctrl_seg) +
360 sizeof (struct mlx4_wqe_datagram_seg);
363 return sizeof (struct mlx4_wqe_ctrl_seg) +
364 sizeof (struct mlx4_wqe_raddr_seg);
366 return sizeof (struct mlx4_wqe_ctrl_seg) +
367 sizeof (struct mlx4_wqe_masked_atomic_seg) +
368 sizeof (struct mlx4_wqe_raddr_seg);
369 case MLX4_IB_QPT_SMI:
370 case MLX4_IB_QPT_GSI:
371 return sizeof (struct mlx4_wqe_ctrl_seg) +
372 ALIGN(MLX4_IB_UD_HEADER_SIZE +
373 DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
375 sizeof (struct mlx4_wqe_inline_seg),
376 sizeof (struct mlx4_wqe_data_seg)) +
378 sizeof (struct mlx4_wqe_inline_seg),
379 sizeof (struct mlx4_wqe_data_seg));
381 return sizeof (struct mlx4_wqe_ctrl_seg);
385 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
386 int is_user, int has_rq, struct mlx4_ib_qp *qp)
388 /* Sanity check RQ size before proceeding */
389 if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
390 cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
394 if (cap->max_recv_wr)
397 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
399 /* HW requires >= 1 RQ entry with >= 1 gather entry */
400 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
403 qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, cap->max_recv_wr));
404 qp->rq.max_gs = roundup_pow_of_two(max(1U, cap->max_recv_sge));
405 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
408 /* leave userspace return values as they were, so as not to break ABI */
410 cap->max_recv_wr = qp->rq.max_post = qp->rq.wqe_cnt;
411 cap->max_recv_sge = qp->rq.max_gs;
413 cap->max_recv_wr = qp->rq.max_post =
414 min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
415 cap->max_recv_sge = min(qp->rq.max_gs,
416 min(dev->dev->caps.max_sq_sg,
417 dev->dev->caps.max_rq_sg));
423 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
424 enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp,
429 /* Sanity check SQ size before proceeding */
430 if (cap->max_send_wr > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
431 cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
432 cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
433 sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
437 * For MLX transport we need 2 extra S/G entries:
438 * one for the header and one for the checksum at the end
440 if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
441 type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
442 cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
445 s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
446 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
447 send_wqe_overhead(type, qp->flags);
449 if (s > dev->dev->caps.max_sq_desc_sz)
453 * Hermon supports shrinking WQEs, such that a single work
454 * request can include multiple units of 1 << wqe_shift. This
455 * way, work requests can differ in size, and do not have to
456 * be a power of 2 in size, saving memory and speeding up send
457 * WR posting. Unfortunately, if we do this then the
458 * wqe_index field in CQEs can't be used to look up the WR ID
459 * anymore, so we do this only if selective signaling is off.
461 * Further, on 32-bit platforms, we can't use vmap() to make
462 * the QP buffer virtually contiguous. Thus we have to use
463 * constant-sized WRs to make sure a WR is always fully within
464 * a single page-sized chunk.
466 * Finally, we use NOP work requests to pad the end of the
467 * work queue, to avoid wrap-around in the middle of WR. We
468 * set NEC bit to avoid getting completions with error for
469 * these NOP WRs, but since NEC is only supported starting
470 * with firmware 2.2.232, we use constant-sized WRs for older
473 * And, since MLX QPs only support SEND, we use constant-sized
476 * We look for the smallest value of wqe_shift such that the
477 * resulting number of wqes does not exceed device
480 * We set WQE size to at least 64 bytes, this way stamping
481 * invalidates each WQE.
483 if (shrink_wqe && dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
484 qp->sq_signal_bits && BITS_PER_LONG == 64 &&
485 type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
486 !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
487 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
488 qp->sq.wqe_shift = ilog2(64);
490 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
493 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
496 * We need to leave 2 KB + 1 WR of headroom in the SQ to
497 * allow HW to prefetch.
499 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
500 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
501 qp->sq_max_wqes_per_wr +
504 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
507 if (qp->sq_max_wqes_per_wr <= 1)
513 qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
514 (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
515 send_wqe_overhead(type, qp->flags)) /
516 sizeof (struct mlx4_wqe_data_seg);
518 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
519 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
520 if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
522 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
524 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
528 cap->max_send_wr = qp->sq.max_post =
529 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
530 cap->max_send_sge = min(qp->sq.max_gs,
531 min(dev->dev->caps.max_sq_sg,
532 dev->dev->caps.max_rq_sg));
533 /* We don't support inline sends for kernel QPs (yet) */
534 cap->max_inline_data = 0;
539 static int set_user_sq_size(struct mlx4_ib_dev *dev,
540 struct mlx4_ib_qp *qp,
541 struct mlx4_ib_create_qp *ucmd)
543 /* Sanity check SQ size before proceeding */
544 if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes ||
545 ucmd->log_sq_stride >
546 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
547 ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
550 qp->sq.wqe_cnt = 1 << ucmd->log_sq_bb_count;
551 qp->sq.wqe_shift = ucmd->log_sq_stride;
553 qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
554 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
559 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
564 kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
566 if (!qp->sqp_proxy_rcv)
568 for (i = 0; i < qp->rq.wqe_cnt; i++) {
569 qp->sqp_proxy_rcv[i].addr =
570 kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
572 if (!qp->sqp_proxy_rcv[i].addr)
574 qp->sqp_proxy_rcv[i].map =
575 ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
576 sizeof (struct mlx4_ib_proxy_sqp_hdr),
578 if (ib_dma_mapping_error(dev, qp->sqp_proxy_rcv[i].map)) {
579 kfree(qp->sqp_proxy_rcv[i].addr);
588 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
589 sizeof (struct mlx4_ib_proxy_sqp_hdr),
591 kfree(qp->sqp_proxy_rcv[i].addr);
593 kfree(qp->sqp_proxy_rcv);
594 qp->sqp_proxy_rcv = NULL;
598 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
602 for (i = 0; i < qp->rq.wqe_cnt; i++) {
603 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
604 sizeof (struct mlx4_ib_proxy_sqp_hdr),
606 kfree(qp->sqp_proxy_rcv[i].addr);
608 kfree(qp->sqp_proxy_rcv);
611 static int qp_has_rq(struct ib_qp_init_attr *attr)
613 if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
619 static int qp0_enabled_vf(struct mlx4_dev *dev, int qpn)
622 for (i = 0; i < dev->caps.num_ports; i++) {
623 if (qpn == dev->caps.qp0_proxy[i])
624 return !!dev->caps.qp0_qkey[i];
629 static void mlx4_ib_free_qp_counter(struct mlx4_ib_dev *dev,
630 struct mlx4_ib_qp *qp)
632 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
633 mlx4_counter_free(dev->dev, qp->counter_index->index);
634 list_del(&qp->counter_index->list);
635 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
637 kfree(qp->counter_index);
638 qp->counter_index = NULL;
641 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
642 struct ib_qp_init_attr *init_attr,
643 struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
648 struct ib_qp_cap backup_cap;
649 struct mlx4_ib_sqp *sqp;
650 struct mlx4_ib_qp *qp;
651 enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
652 struct mlx4_ib_cq *mcq;
655 /* When tunneling special qps, we use a plain UD qp */
657 if (mlx4_is_mfunc(dev->dev) &&
658 (!mlx4_is_master(dev->dev) ||
659 !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
660 if (init_attr->qp_type == IB_QPT_GSI)
661 qp_type = MLX4_IB_QPT_PROXY_GSI;
663 if (mlx4_is_master(dev->dev) ||
664 qp0_enabled_vf(dev->dev, sqpn))
665 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
667 qp_type = MLX4_IB_QPT_PROXY_SMI;
671 /* add extra sg entry for tunneling */
672 init_attr->cap.max_recv_sge++;
673 } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
674 struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
675 container_of(init_attr,
676 struct mlx4_ib_qp_tunnel_init_attr, init_attr);
677 if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
678 tnl_init->proxy_qp_type != IB_QPT_GSI) ||
679 !mlx4_is_master(dev->dev))
681 if (tnl_init->proxy_qp_type == IB_QPT_GSI)
682 qp_type = MLX4_IB_QPT_TUN_GSI;
683 else if (tnl_init->slave == mlx4_master_func_num(dev->dev) ||
684 mlx4_vf_smi_enabled(dev->dev, tnl_init->slave,
686 qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
688 qp_type = MLX4_IB_QPT_TUN_SMI;
689 /* we are definitely in the PPF here, since we are creating
690 * tunnel QPs. base_tunnel_sqpn is therefore valid. */
691 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
692 + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
697 if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
698 (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
699 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
700 sqp = kzalloc(sizeof (struct mlx4_ib_sqp), gfp);
704 qp->pri.vid = 0xFFFF;
705 qp->alt.vid = 0xFFFF;
707 qp = kzalloc(sizeof (struct mlx4_ib_qp), gfp);
710 qp->pri.vid = 0xFFFF;
711 qp->alt.vid = 0xFFFF;
716 qp->mlx4_ib_qp_type = qp_type;
718 mutex_init(&qp->mutex);
719 spin_lock_init(&qp->sq.lock);
720 spin_lock_init(&qp->rq.lock);
721 INIT_LIST_HEAD(&qp->gid_list);
722 INIT_LIST_HEAD(&qp->steering_rules);
724 qp->state = IB_QPS_RESET;
725 if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
726 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
728 err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
733 struct mlx4_ib_create_qp ucmd;
735 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
740 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
742 err = set_user_sq_size(dev, qp, &ucmd);
746 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
748 if (IS_ERR(qp->umem)) {
749 err = PTR_ERR(qp->umem);
753 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
754 ilog2(qp->umem->page_size), &qp->mtt);
758 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
762 if (qp_has_rq(init_attr)) {
763 err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
764 ucmd.db_addr, &qp->db);
769 qp->sq_no_prefetch = 0;
771 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
772 qp->flags |= MLX4_IB_QP_LSO;
774 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
775 if (dev->steering_support ==
776 MLX4_STEERING_MODE_DEVICE_MANAGED)
777 qp->flags |= MLX4_IB_QP_NETIF;
782 memcpy(&backup_cap, &init_attr->cap, sizeof(backup_cap));
783 err = set_kernel_sq_size(dev, &init_attr->cap,
788 if (qp_has_rq(init_attr)) {
789 err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
796 if (mlx4_buf_alloc(dev->dev, qp->buf_size, qp->buf_size,
798 memcpy(&init_attr->cap, &backup_cap,
800 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type,
805 if (mlx4_buf_alloc(dev->dev, qp->buf_size,
806 PAGE_SIZE * 2, &qp->buf, gfp)) {
812 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
817 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
821 qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(u64),
824 qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(u64),
825 gfp, 0 /*PAGE_KERNEL*/);
826 qp->rq.wrid = kmalloc_array(qp->rq.wqe_cnt, sizeof(u64),
829 qp->rq.wrid = __vmalloc(qp->rq.wqe_cnt * sizeof(u64),
830 gfp, 0 /*PAGE_KERNEL*/);
831 if (!qp->sq.wrid || !qp->rq.wrid) {
838 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
839 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
840 if (alloc_proxy_bufs(pd->device, qp)) {
846 /* Raw packet QPNs may not have bits 6,7 set in their qp_num;
847 * otherwise, the WQE BlueFlame setup flow wrongly causes
849 if (init_attr->qp_type == IB_QPT_RAW_PACKET)
850 err = mlx4_qp_reserve_range(dev->dev, 1, 1, &qpn,
851 (init_attr->cap.max_send_wr ?
852 MLX4_RESERVE_ETH_BF_QP : 0) |
853 (init_attr->cap.max_recv_wr ?
854 MLX4_RESERVE_A0_QP : 0));
856 if (qp->flags & MLX4_IB_QP_NETIF)
857 err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
859 err = mlx4_qp_reserve_range(dev->dev, 1, 1,
865 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
866 qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
868 err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
872 if (init_attr->qp_type == IB_QPT_XRC_TGT)
873 qp->mqp.qpn |= (1 << 23);
876 * Hardware wants QPN written in big-endian order (after
877 * shifting) for send doorbell. Precompute this value to save
878 * a little bit when posting sends.
880 qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
882 qp->mqp.event = mlx4_ib_qp_event;
886 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
887 mlx4_ib_lock_cqs(to_mcq(init_attr->send_cq),
888 to_mcq(init_attr->recv_cq));
889 /* Maintain device to QPs access, needed for further handling
892 list_add_tail(&qp->qps_list, &dev->qp_list);
893 /* Maintain CQ to QPs access, needed for further handling
896 mcq = to_mcq(init_attr->send_cq);
897 list_add_tail(&qp->cq_send_list, &mcq->send_qp_list);
898 mcq = to_mcq(init_attr->recv_cq);
899 list_add_tail(&qp->cq_recv_list, &mcq->recv_qp_list);
900 mlx4_ib_unlock_cqs(to_mcq(init_attr->send_cq),
901 to_mcq(init_attr->recv_cq));
902 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
907 if (qp->flags & MLX4_IB_QP_NETIF)
908 mlx4_ib_steer_qp_free(dev, qpn, 1);
910 mlx4_qp_release_range(dev->dev, qpn, 1);
913 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
914 free_proxy_bufs(pd->device, qp);
917 if (qp_has_rq(init_attr))
918 mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
925 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
929 ib_umem_release(qp->umem);
931 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
934 if (!pd->uobject && qp_has_rq(init_attr))
935 mlx4_db_free(dev->dev, &qp->db);
943 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
946 case IB_QPS_RESET: return MLX4_QP_STATE_RST;
947 case IB_QPS_INIT: return MLX4_QP_STATE_INIT;
948 case IB_QPS_RTR: return MLX4_QP_STATE_RTR;
949 case IB_QPS_RTS: return MLX4_QP_STATE_RTS;
950 case IB_QPS_SQD: return MLX4_QP_STATE_SQD;
951 case IB_QPS_SQE: return MLX4_QP_STATE_SQER;
952 case IB_QPS_ERR: return MLX4_QP_STATE_ERR;
957 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
958 __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
960 if (send_cq == recv_cq) {
961 spin_lock(&send_cq->lock);
962 __acquire(&recv_cq->lock);
963 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
964 spin_lock(&send_cq->lock);
965 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
967 spin_lock(&recv_cq->lock);
968 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
972 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
973 __releases(&send_cq->lock) __releases(&recv_cq->lock)
975 if (send_cq == recv_cq) {
976 __release(&recv_cq->lock);
977 spin_unlock(&send_cq->lock);
978 } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
979 spin_unlock(&recv_cq->lock);
980 spin_unlock(&send_cq->lock);
982 spin_unlock(&send_cq->lock);
983 spin_unlock(&recv_cq->lock);
987 static void del_gid_entries(struct mlx4_ib_qp *qp)
989 struct mlx4_ib_gid_entry *ge, *tmp;
991 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
997 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
999 if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
1000 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
1002 return to_mpd(qp->ibqp.pd);
1005 static void get_cqs(struct mlx4_ib_qp *qp,
1006 struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
1008 switch (qp->ibqp.qp_type) {
1009 case IB_QPT_XRC_TGT:
1010 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
1011 *recv_cq = *send_cq;
1013 case IB_QPT_XRC_INI:
1014 *send_cq = to_mcq(qp->ibqp.send_cq);
1015 *recv_cq = *send_cq;
1018 *send_cq = to_mcq(qp->ibqp.send_cq);
1019 *recv_cq = to_mcq(qp->ibqp.recv_cq);
1024 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
1027 struct mlx4_ib_cq *send_cq, *recv_cq;
1028 unsigned long flags;
1030 if (qp->state != IB_QPS_RESET) {
1031 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
1032 MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
1033 pr_warn("modify QP %06x to RESET failed.\n",
1035 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
1036 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1038 qp->pri.smac_port = 0;
1041 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1044 if (qp->pri.vid < 0x1000) {
1045 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1046 qp->pri.vid = 0xFFFF;
1047 qp->pri.candidate_vid = 0xFFFF;
1048 qp->pri.update_vid = 0;
1050 if (qp->alt.vid < 0x1000) {
1051 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1052 qp->alt.vid = 0xFFFF;
1053 qp->alt.candidate_vid = 0xFFFF;
1054 qp->alt.update_vid = 0;
1058 get_cqs(qp, &send_cq, &recv_cq);
1060 spin_lock_irqsave(&dev->reset_flow_resource_lock, flags);
1061 mlx4_ib_lock_cqs(send_cq, recv_cq);
1063 /* del from lists under both locks above to protect reset flow paths */
1064 list_del(&qp->qps_list);
1065 list_del(&qp->cq_send_list);
1066 list_del(&qp->cq_recv_list);
1068 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1069 qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
1070 if (send_cq != recv_cq)
1071 __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1074 mlx4_qp_remove(dev->dev, &qp->mqp);
1076 mlx4_ib_unlock_cqs(send_cq, recv_cq);
1077 spin_unlock_irqrestore(&dev->reset_flow_resource_lock, flags);
1079 mlx4_qp_free(dev->dev, &qp->mqp);
1081 if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
1082 if (qp->flags & MLX4_IB_QP_NETIF)
1083 mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
1085 mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
1088 mlx4_mtt_cleanup(dev->dev, &qp->mtt);
1092 mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1094 ib_umem_release(qp->umem);
1096 kvfree(qp->sq.wrid);
1097 kvfree(qp->rq.wrid);
1098 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1099 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1100 free_proxy_bufs(&dev->ib_dev, qp);
1101 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1103 mlx4_db_free(dev->dev, &qp->db);
1106 del_gid_entries(qp);
1109 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1112 if (!mlx4_is_mfunc(dev->dev) ||
1113 (mlx4_is_master(dev->dev) &&
1114 attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1115 return dev->dev->phys_caps.base_sqpn +
1116 (attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1119 /* PF or VF -- creating proxies */
1120 if (attr->qp_type == IB_QPT_SMI)
1121 return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1123 return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1126 static struct ib_qp *_mlx4_ib_create_qp(struct ib_pd *pd,
1127 struct ib_qp_init_attr *init_attr,
1128 struct ib_udata *udata)
1130 struct mlx4_ib_qp *qp = NULL;
1132 int sup_u_create_flags = MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
1136 gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1137 GFP_NOIO : GFP_KERNEL;
1139 * We only support LSO, vendor flag1, and multicast loopback blocking,
1140 * and only for kernel UD QPs.
1142 if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1143 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1144 MLX4_IB_SRIOV_TUNNEL_QP |
1147 MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1148 MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1149 return ERR_PTR(-EINVAL);
1151 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1152 if (init_attr->qp_type != IB_QPT_UD)
1153 return ERR_PTR(-EINVAL);
1156 if (init_attr->create_flags) {
1157 if (udata && init_attr->create_flags & ~(sup_u_create_flags))
1158 return ERR_PTR(-EINVAL);
1160 if ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP |
1161 MLX4_IB_QP_CREATE_USE_GFP_NOIO |
1162 MLX4_IB_QP_CREATE_ROCE_V2_GSI |
1163 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) &&
1164 init_attr->qp_type != IB_QPT_UD) ||
1165 (init_attr->create_flags & MLX4_IB_SRIOV_SQP &&
1166 init_attr->qp_type > IB_QPT_GSI) ||
1167 (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI &&
1168 init_attr->qp_type != IB_QPT_GSI))
1169 return ERR_PTR(-EINVAL);
1172 switch (init_attr->qp_type) {
1173 case IB_QPT_XRC_TGT:
1174 pd = to_mxrcd(init_attr->xrcd)->pd;
1175 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1176 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1178 case IB_QPT_XRC_INI:
1179 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1180 return ERR_PTR(-ENOSYS);
1181 init_attr->recv_cq = init_attr->send_cq;
1185 case IB_QPT_RAW_PACKET:
1186 qp = kzalloc(sizeof *qp, gfp);
1188 return ERR_PTR(-ENOMEM);
1189 qp->pri.vid = 0xFFFF;
1190 qp->alt.vid = 0xFFFF;
1194 err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1195 udata, 0, &qp, gfp);
1198 return ERR_PTR(err);
1201 qp->ibqp.qp_num = qp->mqp.qpn;
1211 /* Userspace is not allowed to create special QPs: */
1213 return ERR_PTR(-EINVAL);
1214 if (init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI) {
1215 int res = mlx4_qp_reserve_range(to_mdev(pd->device)->dev, 1, 1, &sqpn, 0);
1218 return ERR_PTR(res);
1220 sqpn = get_sqp_num(to_mdev(pd->device), init_attr);
1223 err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1227 return ERR_PTR(err);
1229 qp->port = init_attr->port_num;
1230 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 :
1231 init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn : 1;
1235 /* Don't support raw QPs */
1236 return ERR_PTR(-EINVAL);
1242 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1243 struct ib_qp_init_attr *init_attr,
1244 struct ib_udata *udata) {
1245 struct ib_device *device = pd ? pd->device : init_attr->xrcd->device;
1247 struct mlx4_ib_dev *dev = to_mdev(device);
1249 ibqp = _mlx4_ib_create_qp(pd, init_attr, udata);
1251 if (!IS_ERR(ibqp) &&
1252 (init_attr->qp_type == IB_QPT_GSI) &&
1253 !(init_attr->create_flags & MLX4_IB_QP_CREATE_ROCE_V2_GSI)) {
1254 struct mlx4_ib_sqp *sqp = to_msqp((to_mqp(ibqp)));
1255 int is_eth = rdma_cap_eth_ah(&dev->ib_dev, init_attr->port_num);
1258 dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
1259 init_attr->create_flags |= MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1260 sqp->roce_v2_gsi = ib_create_qp(pd, init_attr);
1262 if (IS_ERR(sqp->roce_v2_gsi)) {
1263 pr_err("Failed to create GSI QP for RoCEv2 (%ld)\n", PTR_ERR(sqp->roce_v2_gsi));
1264 sqp->roce_v2_gsi = NULL;
1266 sqp = to_msqp(to_mqp(sqp->roce_v2_gsi));
1267 sqp->qp.flags |= MLX4_IB_ROCE_V2_GSI_QP;
1270 init_attr->create_flags &= ~MLX4_IB_QP_CREATE_ROCE_V2_GSI;
1276 static int _mlx4_ib_destroy_qp(struct ib_qp *qp)
1278 struct mlx4_ib_dev *dev = to_mdev(qp->device);
1279 struct mlx4_ib_qp *mqp = to_mqp(qp);
1280 struct mlx4_ib_pd *pd;
1282 if (is_qp0(dev, mqp))
1283 mlx4_CLOSE_PORT(dev->dev, mqp->port);
1285 if (dev->qp1_proxy[mqp->port - 1] == mqp) {
1286 mutex_lock(&dev->qp1_proxy_lock[mqp->port - 1]);
1287 dev->qp1_proxy[mqp->port - 1] = NULL;
1288 mutex_unlock(&dev->qp1_proxy_lock[mqp->port - 1]);
1291 if (mqp->counter_index)
1292 mlx4_ib_free_qp_counter(dev, mqp);
1295 destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1297 if (is_sqp(dev, mqp))
1298 kfree(to_msqp(mqp));
1305 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1307 struct mlx4_ib_qp *mqp = to_mqp(qp);
1309 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
1310 struct mlx4_ib_sqp *sqp = to_msqp(mqp);
1312 if (sqp->roce_v2_gsi)
1313 ib_destroy_qp(sqp->roce_v2_gsi);
1316 return _mlx4_ib_destroy_qp(qp);
1319 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1322 case MLX4_IB_QPT_RC: return MLX4_QP_ST_RC;
1323 case MLX4_IB_QPT_UC: return MLX4_QP_ST_UC;
1324 case MLX4_IB_QPT_UD: return MLX4_QP_ST_UD;
1325 case MLX4_IB_QPT_XRC_INI:
1326 case MLX4_IB_QPT_XRC_TGT: return MLX4_QP_ST_XRC;
1327 case MLX4_IB_QPT_SMI:
1328 case MLX4_IB_QPT_GSI:
1329 case MLX4_IB_QPT_RAW_PACKET: return MLX4_QP_ST_MLX;
1331 case MLX4_IB_QPT_PROXY_SMI_OWNER:
1332 case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ?
1333 MLX4_QP_ST_MLX : -1);
1334 case MLX4_IB_QPT_PROXY_SMI:
1335 case MLX4_IB_QPT_TUN_SMI:
1336 case MLX4_IB_QPT_PROXY_GSI:
1337 case MLX4_IB_QPT_TUN_GSI: return (mlx4_is_mfunc(dev->dev) ?
1338 MLX4_QP_ST_UD : -1);
1343 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1348 u32 hw_access_flags = 0;
1350 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1351 dest_rd_atomic = attr->max_dest_rd_atomic;
1353 dest_rd_atomic = qp->resp_depth;
1355 if (attr_mask & IB_QP_ACCESS_FLAGS)
1356 access_flags = attr->qp_access_flags;
1358 access_flags = qp->atomic_rd_en;
1360 if (!dest_rd_atomic)
1361 access_flags &= IB_ACCESS_REMOTE_WRITE;
1363 if (access_flags & IB_ACCESS_REMOTE_READ)
1364 hw_access_flags |= MLX4_QP_BIT_RRE;
1365 if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1366 hw_access_flags |= MLX4_QP_BIT_RAE;
1367 if (access_flags & IB_ACCESS_REMOTE_WRITE)
1368 hw_access_flags |= MLX4_QP_BIT_RWE;
1370 return cpu_to_be32(hw_access_flags);
1373 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1376 if (attr_mask & IB_QP_PKEY_INDEX)
1377 sqp->pkey_index = attr->pkey_index;
1378 if (attr_mask & IB_QP_QKEY)
1379 sqp->qkey = attr->qkey;
1380 if (attr_mask & IB_QP_SQ_PSN)
1381 sqp->send_psn = attr->sq_psn;
1384 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1386 path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1389 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1390 u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1391 struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1393 int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1394 IB_LINK_LAYER_ETHERNET;
1400 path->grh_mylmc = ah->src_path_bits & 0x7f;
1401 path->rlid = cpu_to_be16(ah->dlid);
1402 if (ah->static_rate) {
1403 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1404 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1405 !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1406 --path->static_rate;
1408 path->static_rate = 0;
1410 if (ah->ah_flags & IB_AH_GRH) {
1411 int real_sgid_index = mlx4_ib_gid_index_to_real_index(dev,
1413 ah->grh.sgid_index);
1415 if (real_sgid_index >= dev->dev->caps.gid_table_len[port]) {
1416 pr_err("sgid_index (%u) too large. max is %d\n",
1417 real_sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1421 path->grh_mylmc |= 1 << 7;
1422 path->mgid_index = real_sgid_index;
1423 path->hop_limit = ah->grh.hop_limit;
1424 path->tclass_flowlabel =
1425 cpu_to_be32((ah->grh.traffic_class << 20) |
1426 (ah->grh.flow_label));
1427 memcpy(path->rgid, ah->grh.dgid.raw, 16);
1431 if (!(ah->ah_flags & IB_AH_GRH))
1434 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1435 ((port - 1) << 6) | ((ah->sl & 7) << 3);
1437 path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1438 if (vlan_tag < 0x1000) {
1439 if (smac_info->vid < 0x1000) {
1440 /* both valid vlan ids */
1441 if (smac_info->vid != vlan_tag) {
1442 /* different VIDs. unreg old and reg new */
1443 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1446 smac_info->candidate_vid = vlan_tag;
1447 smac_info->candidate_vlan_index = vidx;
1448 smac_info->candidate_vlan_port = port;
1449 smac_info->update_vid = 1;
1450 path->vlan_index = vidx;
1452 path->vlan_index = smac_info->vlan_index;
1455 /* no current vlan tag in qp */
1456 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1459 smac_info->candidate_vid = vlan_tag;
1460 smac_info->candidate_vlan_index = vidx;
1461 smac_info->candidate_vlan_port = port;
1462 smac_info->update_vid = 1;
1463 path->vlan_index = vidx;
1465 path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1468 /* have current vlan tag. unregister it at modify-qp success */
1469 if (smac_info->vid < 0x1000) {
1470 smac_info->candidate_vid = 0xFFFF;
1471 smac_info->update_vid = 1;
1475 /* get smac_index for RoCE use.
1476 * If no smac was yet assigned, register one.
1477 * If one was already assigned, but the new mac differs,
1478 * unregister the old one and register the new one.
1480 if ((!smac_info->smac && !smac_info->smac_port) ||
1481 smac_info->smac != smac) {
1482 /* register candidate now, unreg if needed, after success */
1483 smac_index = mlx4_register_mac(dev->dev, port, smac);
1484 if (smac_index >= 0) {
1485 smac_info->candidate_smac_index = smac_index;
1486 smac_info->candidate_smac = smac;
1487 smac_info->candidate_smac_port = port;
1492 smac_index = smac_info->smac_index;
1495 memcpy(path->dmac, ah->dmac, 6);
1496 path->ackto = MLX4_IB_LINK_TYPE_ETH;
1497 /* put MAC table smac index for IBoE */
1498 path->grh_mylmc = (u8) (smac_index) | 0x80;
1500 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1501 ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1507 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1508 enum ib_qp_attr_mask qp_attr_mask,
1509 struct mlx4_ib_qp *mqp,
1510 struct mlx4_qp_path *path, u8 port,
1511 u16 vlan_id, u8 *smac)
1513 return _mlx4_set_path(dev, &qp->ah_attr,
1514 mlx4_mac_to_u64(smac),
1516 path, &mqp->pri, port);
1519 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1520 const struct ib_qp_attr *qp,
1521 enum ib_qp_attr_mask qp_attr_mask,
1522 struct mlx4_ib_qp *mqp,
1523 struct mlx4_qp_path *path, u8 port)
1525 return _mlx4_set_path(dev, &qp->alt_ah_attr,
1528 path, &mqp->alt, port);
1531 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1533 struct mlx4_ib_gid_entry *ge, *tmp;
1535 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1536 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1538 ge->port = qp->port;
1543 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev,
1544 struct mlx4_ib_qp *qp,
1545 struct mlx4_qp_context *context)
1550 u64_mac = atomic64_read(&dev->iboe.mac[qp->port - 1]);
1552 context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1553 if (!qp->pri.smac && !qp->pri.smac_port) {
1554 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1555 if (smac_index >= 0) {
1556 qp->pri.candidate_smac_index = smac_index;
1557 qp->pri.candidate_smac = u64_mac;
1558 qp->pri.candidate_smac_port = qp->port;
1559 context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1567 static int create_qp_lb_counter(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1569 struct counter_index *new_counter_index;
1573 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) !=
1574 IB_LINK_LAYER_ETHERNET ||
1575 !(qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK) ||
1576 !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_LB_SRC_CHK))
1579 err = mlx4_counter_alloc(dev->dev, &tmp_idx);
1583 new_counter_index = kmalloc(sizeof(*new_counter_index), GFP_KERNEL);
1584 if (!new_counter_index) {
1585 mlx4_counter_free(dev->dev, tmp_idx);
1589 new_counter_index->index = tmp_idx;
1590 new_counter_index->allocated = 1;
1591 qp->counter_index = new_counter_index;
1593 mutex_lock(&dev->counters_table[qp->port - 1].mutex);
1594 list_add_tail(&new_counter_index->list,
1595 &dev->counters_table[qp->port - 1].counters_list);
1596 mutex_unlock(&dev->counters_table[qp->port - 1].mutex);
1602 MLX4_QPC_ROCE_MODE_1 = 0,
1603 MLX4_QPC_ROCE_MODE_2 = 2,
1604 MLX4_QPC_ROCE_MODE_UNDEFINED = 0xff
1607 static u8 gid_type_to_qpc(enum ib_gid_type gid_type)
1610 case IB_GID_TYPE_ROCE:
1611 return MLX4_QPC_ROCE_MODE_1;
1612 case IB_GID_TYPE_ROCE_UDP_ENCAP:
1613 return MLX4_QPC_ROCE_MODE_2;
1615 return MLX4_QPC_ROCE_MODE_UNDEFINED;
1619 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1620 const struct ib_qp_attr *attr, int attr_mask,
1621 enum ib_qp_state cur_state, enum ib_qp_state new_state)
1623 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1624 struct mlx4_ib_qp *qp = to_mqp(ibqp);
1625 struct mlx4_ib_pd *pd;
1626 struct mlx4_ib_cq *send_cq, *recv_cq;
1627 struct mlx4_qp_context *context;
1628 enum mlx4_qp_optpar optpar = 0;
1634 /* APM is not supported under RoCE */
1635 if (attr_mask & IB_QP_ALT_PATH &&
1636 rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1637 IB_LINK_LAYER_ETHERNET)
1640 context = kzalloc(sizeof *context, GFP_KERNEL);
1644 context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1645 (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1647 if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1648 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1650 optpar |= MLX4_QP_OPTPAR_PM_STATE;
1651 switch (attr->path_mig_state) {
1652 case IB_MIG_MIGRATED:
1653 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1656 context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1659 context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1664 if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1665 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1666 else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1667 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1668 else if (ibqp->qp_type == IB_QPT_UD) {
1669 if (qp->flags & MLX4_IB_QP_LSO)
1670 context->mtu_msgmax = (IB_MTU_4096 << 5) |
1671 ilog2(dev->dev->caps.max_gso_sz);
1673 context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1674 } else if (attr_mask & IB_QP_PATH_MTU) {
1675 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1676 pr_err("path MTU (%u) is invalid\n",
1680 context->mtu_msgmax = (attr->path_mtu << 5) |
1681 ilog2(dev->dev->caps.max_msg_sz);
1685 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1686 context->rq_size_stride |= qp->rq.wqe_shift - 4;
1689 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1690 context->sq_size_stride |= qp->sq.wqe_shift - 4;
1692 if (new_state == IB_QPS_RESET && qp->counter_index)
1693 mlx4_ib_free_qp_counter(dev, qp);
1695 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1696 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1697 context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1698 if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1699 context->param3 |= cpu_to_be32(1 << 30);
1702 if (qp->ibqp.uobject)
1703 context->usr_page = cpu_to_be32(
1704 mlx4_to_hw_uar_index(dev->dev,
1705 to_mucontext(ibqp->uobject->context)->uar.index));
1707 context->usr_page = cpu_to_be32(
1708 mlx4_to_hw_uar_index(dev->dev, dev->priv_uar.index));
1710 if (attr_mask & IB_QP_DEST_QPN)
1711 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1713 if (attr_mask & IB_QP_PORT) {
1714 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1715 !(attr_mask & IB_QP_AV)) {
1716 mlx4_set_sched(&context->pri_path, attr->port_num);
1717 optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1721 if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1722 err = create_qp_lb_counter(dev, qp);
1727 dev->counters_table[qp->port - 1].default_counter;
1728 if (qp->counter_index)
1729 counter_index = qp->counter_index->index;
1731 if (counter_index != -1) {
1732 context->pri_path.counter_index = counter_index;
1733 optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1734 if (qp->counter_index) {
1735 context->pri_path.fl |=
1736 MLX4_FL_ETH_SRC_CHECK_MC_LB;
1737 context->pri_path.vlan_control |=
1738 MLX4_CTRL_ETH_SRC_CHECK_IF_COUNTER;
1741 context->pri_path.counter_index =
1742 MLX4_SINK_COUNTER_INDEX(dev->dev);
1744 if (qp->flags & MLX4_IB_QP_NETIF) {
1745 mlx4_ib_steer_qp_reg(dev, qp, 1);
1749 if (ibqp->qp_type == IB_QPT_GSI) {
1750 enum ib_gid_type gid_type = qp->flags & MLX4_IB_ROCE_V2_GSI_QP ?
1751 IB_GID_TYPE_ROCE_UDP_ENCAP : IB_GID_TYPE_ROCE;
1752 u8 qpc_roce_mode = gid_type_to_qpc(gid_type);
1754 context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1758 if (attr_mask & IB_QP_PKEY_INDEX) {
1759 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1760 context->pri_path.disable_pkey_check = 0x40;
1761 context->pri_path.pkey_index = attr->pkey_index;
1762 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1765 if (attr_mask & IB_QP_AV) {
1766 u8 port_num = mlx4_is_bonded(to_mdev(ibqp->device)->dev) ? 1 :
1767 attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1769 struct ib_gid_attr gid_attr;
1773 int is_eth = rdma_cap_eth_ah(&dev->ib_dev, port_num) &&
1774 attr->ah_attr.ah_flags & IB_AH_GRH;
1777 int index = attr->ah_attr.grh.sgid_index;
1779 status = ib_get_cached_gid(ibqp->device, port_num,
1780 index, &gid, &gid_attr);
1781 if (!status && !memcmp(&gid, &zgid, sizeof(gid)))
1783 if (!status && gid_attr.ndev) {
1784 vlan = rdma_vlan_dev_vlan_id(gid_attr.ndev);
1785 memcpy(smac, IF_LLADDR(gid_attr.ndev), ETH_ALEN);
1786 dev_put(gid_attr.ndev);
1792 if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1793 port_num, vlan, smac))
1796 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1797 MLX4_QP_OPTPAR_SCHED_QUEUE);
1800 (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR)) {
1801 u8 qpc_roce_mode = gid_type_to_qpc(gid_attr.gid_type);
1803 if (qpc_roce_mode == MLX4_QPC_ROCE_MODE_UNDEFINED) {
1807 context->rlkey_roce_mode |= (qpc_roce_mode << 6);
1812 if (attr_mask & IB_QP_TIMEOUT) {
1813 context->pri_path.ackto |= attr->timeout << 3;
1814 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1817 if (attr_mask & IB_QP_ALT_PATH) {
1818 if (attr->alt_port_num == 0 ||
1819 attr->alt_port_num > dev->dev->caps.num_ports)
1822 if (attr->alt_pkey_index >=
1823 dev->dev->caps.pkey_table_len[attr->alt_port_num])
1826 if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1828 attr->alt_port_num))
1831 context->alt_path.pkey_index = attr->alt_pkey_index;
1832 context->alt_path.ackto = attr->alt_timeout << 3;
1833 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1837 get_cqs(qp, &send_cq, &recv_cq);
1838 context->pd = cpu_to_be32(pd->pdn);
1839 context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1840 context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1841 context->params1 = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1843 /* Set "fast registration enabled" for all kernel QPs */
1844 if (!qp->ibqp.uobject)
1845 context->params1 |= cpu_to_be32(1 << 11);
1847 if (attr_mask & IB_QP_RNR_RETRY) {
1848 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1849 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1852 if (attr_mask & IB_QP_RETRY_CNT) {
1853 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1854 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1857 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1858 if (attr->max_rd_atomic)
1860 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1861 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1864 if (attr_mask & IB_QP_SQ_PSN)
1865 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1867 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1868 if (attr->max_dest_rd_atomic)
1870 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1871 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1874 if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1875 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1876 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1880 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1882 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1883 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1884 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1886 if (attr_mask & IB_QP_RQ_PSN)
1887 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1889 /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1890 if (attr_mask & IB_QP_QKEY) {
1891 if (qp->mlx4_ib_qp_type &
1892 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1893 context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1895 if (mlx4_is_mfunc(dev->dev) &&
1896 !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1897 (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1898 MLX4_RESERVED_QKEY_BASE) {
1899 pr_err("Cannot use reserved QKEY"
1900 " 0x%x (range 0xffff0000..0xffffffff"
1901 " is reserved)\n", attr->qkey);
1905 context->qkey = cpu_to_be32(attr->qkey);
1907 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1911 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1913 if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1914 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1916 if (cur_state == IB_QPS_INIT &&
1917 new_state == IB_QPS_RTR &&
1918 (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1919 ibqp->qp_type == IB_QPT_UD ||
1920 ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1921 context->pri_path.sched_queue = (qp->port - 1) << 6;
1922 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1923 qp->mlx4_ib_qp_type &
1924 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1925 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1926 if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1927 context->pri_path.fl = 0x80;
1929 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1930 context->pri_path.fl = 0x80;
1931 context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1933 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1934 IB_LINK_LAYER_ETHERNET) {
1935 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1936 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1937 context->pri_path.feup = 1 << 7; /* don't fsm */
1938 /* handle smac_index */
1939 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1940 qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1941 qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1942 err = handle_eth_ud_smac_index(dev, qp, context);
1947 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
1948 dev->qp1_proxy[qp->port - 1] = qp;
1953 if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {
1954 context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1955 MLX4_IB_LINK_TYPE_ETH;
1956 if (dev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1957 /* set QP to receive both tunneled & non-tunneled packets */
1958 if (!(context->flags & cpu_to_be32(1 << MLX4_RSS_QPC_FLAG_OFFSET)))
1959 context->srqn = cpu_to_be32(7 << 28);
1963 if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1964 int is_eth = rdma_port_get_link_layer(
1965 &dev->ib_dev, qp->port) ==
1966 IB_LINK_LAYER_ETHERNET;
1968 context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1969 optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1974 if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD &&
1975 attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1980 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1981 context->rlkey_roce_mode |= (1 << 4);
1984 * Before passing a kernel QP to the HW, make sure that the
1985 * ownership bits of the send queue are set and the SQ
1986 * headroom is stamped so that the hardware doesn't start
1987 * processing stale work requests.
1989 if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1990 struct mlx4_wqe_ctrl_seg *ctrl;
1993 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1994 ctrl = get_send_wqe(qp, i);
1995 ctrl->owner_opcode = cpu_to_be32(1U << 31);
1996 if (qp->sq_max_wqes_per_wr == 1)
1998 1 << (qp->sq.wqe_shift - 4);
2000 stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
2004 err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
2005 to_mlx4_state(new_state), context, optpar,
2006 sqd_event, &qp->mqp);
2010 qp->state = new_state;
2012 if (attr_mask & IB_QP_ACCESS_FLAGS)
2013 qp->atomic_rd_en = attr->qp_access_flags;
2014 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
2015 qp->resp_depth = attr->max_dest_rd_atomic;
2016 if (attr_mask & IB_QP_PORT) {
2017 qp->port = attr->port_num;
2018 update_mcg_macs(dev, qp);
2020 if (attr_mask & IB_QP_ALT_PATH)
2021 qp->alt_port = attr->alt_port_num;
2023 if (is_sqp(dev, qp))
2024 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
2027 * If we moved QP0 to RTR, bring the IB link up; if we moved
2028 * QP0 to RESET or ERROR, bring the link back down.
2030 if (is_qp0(dev, qp)) {
2031 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
2032 if (mlx4_INIT_PORT(dev->dev, qp->port))
2033 pr_warn("INIT_PORT failed for port %d\n",
2036 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
2037 (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
2038 mlx4_CLOSE_PORT(dev->dev, qp->port);
2042 * If we moved a kernel QP to RESET, clean up all old CQ
2043 * entries and reinitialize the QP.
2045 if (new_state == IB_QPS_RESET) {
2046 if (!ibqp->uobject) {
2047 mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
2048 ibqp->srq ? to_msrq(ibqp->srq) : NULL);
2049 if (send_cq != recv_cq)
2050 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
2056 qp->sq_next_wqe = 0;
2060 if (qp->flags & MLX4_IB_QP_NETIF)
2061 mlx4_ib_steer_qp_reg(dev, qp, 0);
2063 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port)) {
2064 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2066 qp->pri.smac_port = 0;
2069 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2072 if (qp->pri.vid < 0x1000) {
2073 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
2074 qp->pri.vid = 0xFFFF;
2075 qp->pri.candidate_vid = 0xFFFF;
2076 qp->pri.update_vid = 0;
2079 if (qp->alt.vid < 0x1000) {
2080 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
2081 qp->alt.vid = 0xFFFF;
2082 qp->alt.candidate_vid = 0xFFFF;
2083 qp->alt.update_vid = 0;
2087 if (err && qp->counter_index)
2088 mlx4_ib_free_qp_counter(dev, qp);
2089 if (err && steer_qp)
2090 mlx4_ib_steer_qp_reg(dev, qp, 0);
2092 if (qp->pri.candidate_smac ||
2093 (!qp->pri.candidate_smac && qp->pri.candidate_smac_port)) {
2095 mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
2097 if (qp->pri.smac || (!qp->pri.smac && qp->pri.smac_port))
2098 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
2099 qp->pri.smac = qp->pri.candidate_smac;
2100 qp->pri.smac_index = qp->pri.candidate_smac_index;
2101 qp->pri.smac_port = qp->pri.candidate_smac_port;
2103 qp->pri.candidate_smac = 0;
2104 qp->pri.candidate_smac_index = 0;
2105 qp->pri.candidate_smac_port = 0;
2107 if (qp->alt.candidate_smac) {
2109 mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
2112 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
2113 qp->alt.smac = qp->alt.candidate_smac;
2114 qp->alt.smac_index = qp->alt.candidate_smac_index;
2115 qp->alt.smac_port = qp->alt.candidate_smac_port;
2117 qp->alt.candidate_smac = 0;
2118 qp->alt.candidate_smac_index = 0;
2119 qp->alt.candidate_smac_port = 0;
2122 if (qp->pri.update_vid) {
2124 if (qp->pri.candidate_vid < 0x1000)
2125 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
2126 qp->pri.candidate_vid);
2128 if (qp->pri.vid < 0x1000)
2129 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
2131 qp->pri.vid = qp->pri.candidate_vid;
2132 qp->pri.vlan_port = qp->pri.candidate_vlan_port;
2133 qp->pri.vlan_index = qp->pri.candidate_vlan_index;
2135 qp->pri.candidate_vid = 0xFFFF;
2136 qp->pri.update_vid = 0;
2139 if (qp->alt.update_vid) {
2141 if (qp->alt.candidate_vid < 0x1000)
2142 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
2143 qp->alt.candidate_vid);
2145 if (qp->alt.vid < 0x1000)
2146 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
2148 qp->alt.vid = qp->alt.candidate_vid;
2149 qp->alt.vlan_port = qp->alt.candidate_vlan_port;
2150 qp->alt.vlan_index = qp->alt.candidate_vlan_index;
2152 qp->alt.candidate_vid = 0xFFFF;
2153 qp->alt.update_vid = 0;
2159 static int _mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2160 int attr_mask, struct ib_udata *udata)
2162 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2163 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2164 enum ib_qp_state cur_state, new_state;
2167 mutex_lock(&qp->mutex);
2169 cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
2170 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
2172 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2173 ll = IB_LINK_LAYER_UNSPECIFIED;
2175 int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2176 ll = rdma_port_get_link_layer(&dev->ib_dev, port);
2179 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
2181 pr_debug("qpn 0x%x: invalid attribute mask specified "
2182 "for transition %d to %d. qp_type %d,"
2183 " attr_mask 0x%x\n",
2184 ibqp->qp_num, cur_state, new_state,
2185 ibqp->qp_type, attr_mask);
2189 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT)) {
2190 if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
2191 if ((ibqp->qp_type == IB_QPT_RC) ||
2192 (ibqp->qp_type == IB_QPT_UD) ||
2193 (ibqp->qp_type == IB_QPT_UC) ||
2194 (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
2195 (ibqp->qp_type == IB_QPT_XRC_INI)) {
2196 attr->port_num = mlx4_ib_bond_next_port(dev);
2199 /* no sense in changing port_num
2200 * when ports are bonded */
2201 attr_mask &= ~IB_QP_PORT;
2205 if ((attr_mask & IB_QP_PORT) &&
2206 (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
2207 pr_debug("qpn 0x%x: invalid port number (%d) specified "
2208 "for transition %d to %d. qp_type %d\n",
2209 ibqp->qp_num, attr->port_num, cur_state,
2210 new_state, ibqp->qp_type);
2214 if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
2215 (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
2216 IB_LINK_LAYER_ETHERNET))
2219 if (attr_mask & IB_QP_PKEY_INDEX) {
2220 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
2221 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
2222 pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
2223 "for transition %d to %d. qp_type %d\n",
2224 ibqp->qp_num, attr->pkey_index, cur_state,
2225 new_state, ibqp->qp_type);
2230 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
2231 attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
2232 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
2233 "Transition %d to %d. qp_type %d\n",
2234 ibqp->qp_num, attr->max_rd_atomic, cur_state,
2235 new_state, ibqp->qp_type);
2239 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
2240 attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
2241 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
2242 "Transition %d to %d. qp_type %d\n",
2243 ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
2244 new_state, ibqp->qp_type);
2248 if (cur_state == new_state && cur_state == IB_QPS_RESET) {
2253 err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
2255 if (mlx4_is_bonded(dev->dev) && (attr_mask & IB_QP_PORT))
2259 mutex_unlock(&qp->mutex);
2263 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2264 int attr_mask, struct ib_udata *udata)
2266 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2269 ret = _mlx4_ib_modify_qp(ibqp, attr, attr_mask, udata);
2271 if (mqp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2272 struct mlx4_ib_sqp *sqp = to_msqp(mqp);
2275 if (sqp->roce_v2_gsi)
2276 err = ib_modify_qp(sqp->roce_v2_gsi, attr, attr_mask);
2278 pr_err("Failed to modify GSI QP for RoCEv2 (%d)\n",
2284 static int vf_get_qp0_qkey(struct mlx4_dev *dev, int qpn, u32 *qkey)
2287 for (i = 0; i < dev->caps.num_ports; i++) {
2288 if (qpn == dev->caps.qp0_proxy[i] ||
2289 qpn == dev->caps.qp0_tunnel[i]) {
2290 *qkey = dev->caps.qp0_qkey[i];
2297 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
2298 struct ib_ud_wr *wr,
2299 void *wqe, unsigned *mlx_seg_len)
2301 struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
2302 struct ib_device *ib_dev = &mdev->ib_dev;
2303 struct mlx4_wqe_mlx_seg *mlx = wqe;
2304 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2305 struct mlx4_ib_ah *ah = to_mah(wr->ah);
2313 if (wr->wr.opcode != IB_WR_SEND)
2318 for (i = 0; i < wr->wr.num_sge; ++i)
2319 send_size += wr->wr.sg_list[i].length;
2321 /* for proxy-qp0 sends, need to add in size of tunnel header */
2322 /* for tunnel-qp0 sends, tunnel header is already in s/g list */
2323 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
2324 send_size += sizeof (struct mlx4_ib_tunnel_header);
2326 ib_ud_header_init(send_size, 1, 0, 0, 0, 0, 0, 0, &sqp->ud_header);
2328 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
2329 sqp->ud_header.lrh.service_level =
2330 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2331 sqp->ud_header.lrh.destination_lid =
2332 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2333 sqp->ud_header.lrh.source_lid =
2334 cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2337 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2339 /* force loopback */
2340 mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
2341 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2343 sqp->ud_header.lrh.virtual_lane = 0;
2344 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2345 ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
2346 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2347 if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
2348 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2350 sqp->ud_header.bth.destination_qpn =
2351 cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
2353 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2354 if (mlx4_is_master(mdev->dev)) {
2355 if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2358 if (vf_get_qp0_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
2361 sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
2362 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
2364 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2365 sqp->ud_header.immediate_present = 0;
2367 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2370 * Inline data segments may not cross a 64 byte boundary. If
2371 * our UD header is bigger than the space available up to the
2372 * next 64 byte boundary in the WQE, use two inline data
2373 * segments to hold the UD header.
2375 spc = MLX4_INLINE_ALIGN -
2376 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2377 if (header_size <= spc) {
2378 inl->byte_count = cpu_to_be32((1U << 31) | header_size);
2379 memcpy(inl + 1, sqp->header_buf, header_size);
2382 inl->byte_count = cpu_to_be32((1U << 31) | spc);
2383 memcpy(inl + 1, sqp->header_buf, spc);
2385 inl = (void *) (inl + 1) + spc;
2386 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2388 * Need a barrier here to make sure all the data is
2389 * visible before the byte_count field is set.
2390 * Otherwise the HCA prefetcher could grab the 64-byte
2391 * chunk with this inline segment and get a valid (!=
2392 * 0xffffffff) byte count but stale data, and end up
2393 * generating a packet with bad headers.
2395 * The first inline segment's byte_count field doesn't
2396 * need a barrier, because it comes after a
2397 * control/MLX segment and therefore is at an offset
2401 inl->byte_count = cpu_to_be32((1U << 31) | (header_size - spc));
2406 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2410 static u8 sl_to_vl(struct mlx4_ib_dev *dev, u8 sl, int port_num)
2412 union sl2vl_tbl_to_u64 tmp_vltab;
2417 tmp_vltab.sl64 = atomic64_read(&dev->sl2vl[port_num - 1]);
2418 vl = tmp_vltab.sl8[sl >> 1];
2426 #define MLX4_ROCEV2_QP1_SPORT 0xC000
2427 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_ud_wr *wr,
2428 void *wqe, unsigned *mlx_seg_len)
2430 struct ib_device *ib_dev = sqp->qp.ibqp.device;
2431 struct mlx4_wqe_mlx_seg *mlx = wqe;
2432 struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2433 struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2434 struct mlx4_ib_ah *ah = to_mah(wr->ah);
2444 bool is_vlan = false;
2446 bool is_udp = false;
2450 for (i = 0; i < wr->wr.num_sge; ++i)
2451 send_size += wr->wr.sg_list[i].length;
2453 is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2454 is_grh = mlx4_ib_ah_grh_present(ah);
2456 struct ib_gid_attr gid_attr;
2458 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2459 /* When multi-function is enabled, the ib_core gid
2460 * indexes don't necessarily match the hw ones, so
2461 * we must use our own cache */
2462 err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2463 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2464 ah->av.ib.gid_index, &sgid.raw[0]);
2468 err = ib_get_cached_gid(ib_dev,
2469 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2470 ah->av.ib.gid_index, &sgid,
2474 dev_put(gid_attr.ndev);
2475 if (!memcmp(&sgid, &zgid, sizeof(sgid)))
2479 is_udp = gid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP;
2481 if (ipv6_addr_v4mapped((struct in6_addr *)&sgid))
2491 if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2492 vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2496 err = ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh,
2497 ip_version, is_udp, 0, &sqp->ud_header);
2502 sqp->ud_header.lrh.service_level =
2503 be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2504 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2505 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2508 if (is_grh || (ip_version == 6)) {
2509 sqp->ud_header.grh.traffic_class =
2510 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2511 sqp->ud_header.grh.flow_label =
2512 ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2513 sqp->ud_header.grh.hop_limit = ah->av.ib.hop_limit;
2515 memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2517 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2518 /* When multi-function is enabled, the ib_core gid
2519 * indexes don't necessarily match the hw ones, so
2520 * we must use our own cache
2522 sqp->ud_header.grh.source_gid.global.subnet_prefix =
2523 cpu_to_be64(atomic64_read(&(to_mdev(ib_dev)->sriov.
2524 demux[sqp->qp.port - 1].
2526 sqp->ud_header.grh.source_gid.global.interface_id =
2527 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2528 guid_cache[ah->av.ib.gid_index];
2530 ib_get_cached_gid(ib_dev,
2531 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2532 ah->av.ib.gid_index,
2533 &sqp->ud_header.grh.source_gid, NULL);
2536 memcpy(sqp->ud_header.grh.destination_gid.raw,
2537 ah->av.ib.dgid, 16);
2540 if (ip_version == 4) {
2541 sqp->ud_header.ip4.tos =
2542 (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2543 sqp->ud_header.ip4.id = 0;
2544 sqp->ud_header.ip4.frag_off = htons(IP_DF);
2545 sqp->ud_header.ip4.ttl = ah->av.eth.hop_limit;
2547 memcpy(&sqp->ud_header.ip4.saddr,
2549 memcpy(&sqp->ud_header.ip4.daddr, ah->av.ib.dgid + 12, 4);
2550 sqp->ud_header.ip4.check = ib_ud_ip4_csum(&sqp->ud_header);
2554 sqp->ud_header.udp.dport = htons(ROCE_V2_UDP_DPORT);
2555 sqp->ud_header.udp.sport = htons(MLX4_ROCEV2_QP1_SPORT);
2556 sqp->ud_header.udp.csum = 0;
2559 mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2562 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2563 (sqp->ud_header.lrh.destination_lid ==
2564 IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2565 (sqp->ud_header.lrh.service_level << 8));
2566 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2567 mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2568 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2571 switch (wr->wr.opcode) {
2573 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY;
2574 sqp->ud_header.immediate_present = 0;
2576 case IB_WR_SEND_WITH_IMM:
2577 sqp->ud_header.bth.opcode = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2578 sqp->ud_header.immediate_present = 1;
2579 sqp->ud_header.immediate_data = wr->wr.ex.imm_data;
2586 struct in6_addr in6;
2588 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2590 ether_type = (!is_udp) ? MLX4_IB_IBOE_ETHERTYPE :
2591 (ip_version == 4 ? ETHERTYPE_IP : ETHERTYPE_IPV6);
2593 mlx->sched_prio = cpu_to_be16(pcp);
2595 ether_addr_copy(sqp->ud_header.eth.smac_h, ah->av.eth.s_mac);
2596 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2597 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2598 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2599 memcpy(&in6, sgid.raw, sizeof(in6));
2602 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2603 mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2605 sqp->ud_header.eth.type = cpu_to_be16(ether_type);
2607 sqp->ud_header.vlan.type = cpu_to_be16(ether_type);
2608 sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2611 sqp->ud_header.lrh.virtual_lane = !sqp->qp.ibqp.qp_num ? 15 :
2612 sl_to_vl(to_mdev(ib_dev),
2613 sqp->ud_header.lrh.service_level,
2615 if (sqp->qp.ibqp.qp_num && sqp->ud_header.lrh.virtual_lane == 15)
2617 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2618 sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2620 sqp->ud_header.bth.solicited_event = !!(wr->wr.send_flags & IB_SEND_SOLICITED);
2621 if (!sqp->qp.ibqp.qp_num)
2622 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2624 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->pkey_index, &pkey);
2625 sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2626 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->remote_qpn);
2627 sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2628 sqp->ud_header.deth.qkey = cpu_to_be32(wr->remote_qkey & 0x80000000 ?
2629 sqp->qkey : wr->remote_qkey);
2630 sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2632 header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2635 pr_err("built UD header of size %d:\n", header_size);
2636 for (i = 0; i < header_size / 4; ++i) {
2638 pr_err(" [%02x] ", i * 4);
2640 be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2641 if ((i + 1) % 8 == 0)
2648 * Inline data segments may not cross a 64 byte boundary. If
2649 * our UD header is bigger than the space available up to the
2650 * next 64 byte boundary in the WQE, use two inline data
2651 * segments to hold the UD header.
2653 spc = MLX4_INLINE_ALIGN -
2654 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2655 if (header_size <= spc) {
2656 inl->byte_count = cpu_to_be32(1U << 31 | header_size);
2657 memcpy(inl + 1, sqp->header_buf, header_size);
2660 inl->byte_count = cpu_to_be32(1U << 31 | spc);
2661 memcpy(inl + 1, sqp->header_buf, spc);
2663 inl = (void *) (inl + 1) + spc;
2664 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2666 * Need a barrier here to make sure all the data is
2667 * visible before the byte_count field is set.
2668 * Otherwise the HCA prefetcher could grab the 64-byte
2669 * chunk with this inline segment and get a valid (!=
2670 * 0xffffffff) byte count but stale data, and end up
2671 * generating a packet with bad headers.
2673 * The first inline segment's byte_count field doesn't
2674 * need a barrier, because it comes after a
2675 * control/MLX segment and therefore is at an offset
2679 inl->byte_count = cpu_to_be32(1U << 31 | (header_size - spc));
2684 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2688 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2691 struct mlx4_ib_cq *cq;
2693 cur = wq->head - wq->tail;
2694 if (likely(cur + nreq < wq->max_post))
2698 spin_lock(&cq->lock);
2699 cur = wq->head - wq->tail;
2700 spin_unlock(&cq->lock);
2702 return cur + nreq >= wq->max_post;
2705 static __be32 convert_access(int acc)
2707 return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2708 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC) : 0) |
2709 (acc & IB_ACCESS_REMOTE_WRITE ?
2710 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2711 (acc & IB_ACCESS_REMOTE_READ ?
2712 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ) : 0) |
2713 (acc & IB_ACCESS_LOCAL_WRITE ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE) : 0) |
2714 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2717 static void set_reg_seg(struct mlx4_wqe_fmr_seg *fseg,
2718 struct ib_reg_wr *wr)
2720 struct mlx4_ib_mr *mr = to_mmr(wr->mr);
2722 fseg->flags = convert_access(wr->access);
2723 fseg->mem_key = cpu_to_be32(wr->key);
2724 fseg->buf_list = cpu_to_be64(mr->page_map);
2725 fseg->start_addr = cpu_to_be64(mr->ibmr.iova);
2726 fseg->reg_len = cpu_to_be64(mr->ibmr.length);
2727 fseg->offset = 0; /* XXX -- is this just for ZBVA? */
2728 fseg->page_size = cpu_to_be32(ilog2(mr->ibmr.page_size));
2729 fseg->reserved[0] = 0;
2730 fseg->reserved[1] = 0;
2733 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2735 memset(iseg, 0, sizeof(*iseg));
2736 iseg->mem_key = cpu_to_be32(rkey);
2739 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2740 u64 remote_addr, u32 rkey)
2742 rseg->raddr = cpu_to_be64(remote_addr);
2743 rseg->rkey = cpu_to_be32(rkey);
2747 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg,
2748 struct ib_atomic_wr *wr)
2750 if (wr->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2751 aseg->swap_add = cpu_to_be64(wr->swap);
2752 aseg->compare = cpu_to_be64(wr->compare_add);
2753 } else if (wr->wr.opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2754 aseg->swap_add = cpu_to_be64(wr->compare_add);
2755 aseg->compare = cpu_to_be64(wr->compare_add_mask);
2757 aseg->swap_add = cpu_to_be64(wr->compare_add);
2763 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2764 struct ib_atomic_wr *wr)
2766 aseg->swap_add = cpu_to_be64(wr->swap);
2767 aseg->swap_add_mask = cpu_to_be64(wr->swap_mask);
2768 aseg->compare = cpu_to_be64(wr->compare_add);
2769 aseg->compare_mask = cpu_to_be64(wr->compare_add_mask);
2772 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2773 struct ib_ud_wr *wr)
2775 memcpy(dseg->av, &to_mah(wr->ah)->av, sizeof (struct mlx4_av));
2776 dseg->dqpn = cpu_to_be32(wr->remote_qpn);
2777 dseg->qkey = cpu_to_be32(wr->remote_qkey);
2778 dseg->vlan = to_mah(wr->ah)->av.eth.vlan;
2779 memcpy(dseg->mac, to_mah(wr->ah)->av.eth.mac, 6);
2782 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2783 struct mlx4_wqe_datagram_seg *dseg,
2784 struct ib_ud_wr *wr,
2785 enum mlx4_ib_qp_type qpt)
2787 union mlx4_ext_av *av = &to_mah(wr->ah)->av;
2788 struct mlx4_av sqp_av = {0};
2789 int port = *((u8 *) &av->ib.port_pd) & 0x3;
2791 /* force loopback */
2792 sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2793 sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2794 sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2795 cpu_to_be32(0xf0000000);
2797 memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2798 if (qpt == MLX4_IB_QPT_PROXY_GSI)
2799 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2801 dseg->dqpn = cpu_to_be32(dev->dev->caps.qp0_tunnel[port - 1]);
2802 /* Use QKEY from the QP context, which is set by master */
2803 dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2806 static void build_tunnel_header(struct ib_ud_wr *wr, void *wqe, unsigned *mlx_seg_len)
2808 struct mlx4_wqe_inline_seg *inl = wqe;
2809 struct mlx4_ib_tunnel_header hdr;
2810 struct mlx4_ib_ah *ah = to_mah(wr->ah);
2814 memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2815 hdr.remote_qpn = cpu_to_be32(wr->remote_qpn);
2816 hdr.pkey_index = cpu_to_be16(wr->pkey_index);
2817 hdr.qkey = cpu_to_be32(wr->remote_qkey);
2818 memcpy(hdr.mac, ah->av.eth.mac, 6);
2819 hdr.vlan = ah->av.eth.vlan;
2821 spc = MLX4_INLINE_ALIGN -
2822 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2823 if (sizeof (hdr) <= spc) {
2824 memcpy(inl + 1, &hdr, sizeof (hdr));
2826 inl->byte_count = cpu_to_be32((1U << 31) | (u32)sizeof(hdr));
2829 memcpy(inl + 1, &hdr, spc);
2831 inl->byte_count = cpu_to_be32((1U << 31) | spc);
2833 inl = (void *) (inl + 1) + spc;
2834 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2836 inl->byte_count = cpu_to_be32((1U << 31) | (u32)(sizeof (hdr) - spc));
2841 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2844 static void set_mlx_icrc_seg(void *dseg)
2847 struct mlx4_wqe_inline_seg *iseg = dseg;
2852 * Need a barrier here before writing the byte_count field to
2853 * make sure that all the data is visible before the
2854 * byte_count field is set. Otherwise, if the segment begins
2855 * a new cacheline, the HCA prefetcher could grab the 64-byte
2856 * chunk and get a valid (!= * 0xffffffff) byte count but
2857 * stale data, and end up sending the wrong data.
2861 iseg->byte_count = cpu_to_be32((1U << 31) | 4);
2864 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2866 dseg->lkey = cpu_to_be32(sg->lkey);
2867 dseg->addr = cpu_to_be64(sg->addr);
2870 * Need a barrier here before writing the byte_count field to
2871 * make sure that all the data is visible before the
2872 * byte_count field is set. Otherwise, if the segment begins
2873 * a new cacheline, the HCA prefetcher could grab the 64-byte
2874 * chunk and get a valid (!= * 0xffffffff) byte count but
2875 * stale data, and end up sending the wrong data.
2879 dseg->byte_count = cpu_to_be32(sg->length);
2882 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2884 dseg->byte_count = cpu_to_be32(sg->length);
2885 dseg->lkey = cpu_to_be32(sg->lkey);
2886 dseg->addr = cpu_to_be64(sg->addr);
2889 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_ud_wr *wr,
2890 struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2891 __be32 *lso_hdr_sz, __be32 *blh)
2893 unsigned halign = ALIGN(sizeof *wqe + wr->hlen, 16);
2895 if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2896 *blh = cpu_to_be32(1 << 6);
2898 if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2899 wr->wr.num_sge > qp->sq.max_gs - (halign >> 4)))
2902 memcpy(wqe->header, wr->header, wr->hlen);
2904 *lso_hdr_sz = cpu_to_be32(wr->mss << 16 | wr->hlen);
2905 *lso_seg_len = halign;
2909 static __be32 send_ieth(struct ib_send_wr *wr)
2911 switch (wr->opcode) {
2912 case IB_WR_SEND_WITH_IMM:
2913 case IB_WR_RDMA_WRITE_WITH_IMM:
2914 return wr->ex.imm_data;
2916 case IB_WR_SEND_WITH_INV:
2917 return cpu_to_be32(wr->ex.invalidate_rkey);
2924 static void add_zero_len_inline(void *wqe)
2926 struct mlx4_wqe_inline_seg *inl = wqe;
2928 inl->byte_count = cpu_to_be32(1U << 31);
2931 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2932 struct ib_send_wr **bad_wr)
2934 struct mlx4_ib_qp *qp = to_mqp(ibqp);
2936 struct mlx4_wqe_ctrl_seg *ctrl;
2937 struct mlx4_wqe_data_seg *dseg;
2938 unsigned long flags;
2942 int uninitialized_var(stamp);
2943 int uninitialized_var(size);
2944 unsigned uninitialized_var(seglen);
2947 __be32 lso_hdr_sz = 0;
2950 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2952 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI) {
2953 struct mlx4_ib_sqp *sqp = to_msqp(qp);
2955 if (sqp->roce_v2_gsi) {
2956 struct mlx4_ib_ah *ah = to_mah(ud_wr(wr)->ah);
2957 struct ib_gid_attr gid_attr;
2960 if (!ib_get_cached_gid(ibqp->device,
2961 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2962 ah->av.ib.gid_index, &gid,
2965 dev_put(gid_attr.ndev);
2966 qp = (gid_attr.gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ?
2967 to_mqp(sqp->roce_v2_gsi) : qp;
2969 pr_err("Failed to get gid at index %d. RoCEv2 will not work properly\n",
2970 ah->av.ib.gid_index);
2975 spin_lock_irqsave(&qp->sq.lock, flags);
2976 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
2983 ind = qp->sq_next_wqe;
2985 for (nreq = 0; wr; ++nreq, wr = wr->next) {
2989 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2995 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
3001 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
3002 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
3005 (wr->send_flags & IB_SEND_SIGNALED ?
3006 cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
3007 (wr->send_flags & IB_SEND_SOLICITED ?
3008 cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
3009 ((wr->send_flags & IB_SEND_IP_CSUM) ?
3010 cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
3011 MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
3014 ctrl->imm = send_ieth(wr);
3016 wqe += sizeof *ctrl;
3017 size = sizeof *ctrl / 16;
3019 switch (qp->mlx4_ib_qp_type) {
3020 case MLX4_IB_QPT_RC:
3021 case MLX4_IB_QPT_UC:
3022 switch (wr->opcode) {
3023 case IB_WR_ATOMIC_CMP_AND_SWP:
3024 case IB_WR_ATOMIC_FETCH_AND_ADD:
3025 case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
3026 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3027 atomic_wr(wr)->rkey);
3028 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3030 set_atomic_seg(wqe, atomic_wr(wr));
3031 wqe += sizeof (struct mlx4_wqe_atomic_seg);
3033 size += (sizeof (struct mlx4_wqe_raddr_seg) +
3034 sizeof (struct mlx4_wqe_atomic_seg)) / 16;
3038 case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
3039 set_raddr_seg(wqe, atomic_wr(wr)->remote_addr,
3040 atomic_wr(wr)->rkey);
3041 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3043 set_masked_atomic_seg(wqe, atomic_wr(wr));
3044 wqe += sizeof (struct mlx4_wqe_masked_atomic_seg);
3046 size += (sizeof (struct mlx4_wqe_raddr_seg) +
3047 sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
3051 case IB_WR_RDMA_READ:
3052 case IB_WR_RDMA_WRITE:
3053 case IB_WR_RDMA_WRITE_WITH_IMM:
3054 set_raddr_seg(wqe, rdma_wr(wr)->remote_addr,
3056 wqe += sizeof (struct mlx4_wqe_raddr_seg);
3057 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
3060 case IB_WR_LOCAL_INV:
3061 ctrl->srcrb_flags |=
3062 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3063 set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
3064 wqe += sizeof (struct mlx4_wqe_local_inval_seg);
3065 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
3069 ctrl->srcrb_flags |=
3070 cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
3071 set_reg_seg(wqe, reg_wr(wr));
3072 wqe += sizeof(struct mlx4_wqe_fmr_seg);
3073 size += sizeof(struct mlx4_wqe_fmr_seg) / 16;
3077 /* No extra segments required for sends */
3082 case MLX4_IB_QPT_TUN_SMI_OWNER:
3083 err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3085 if (unlikely(err)) {
3090 size += seglen / 16;
3092 case MLX4_IB_QPT_TUN_SMI:
3093 case MLX4_IB_QPT_TUN_GSI:
3094 /* this is a UD qp used in MAD responses to slaves. */
3095 set_datagram_seg(wqe, ud_wr(wr));
3096 /* set the forced-loopback bit in the data seg av */
3097 *(__be32 *) wqe |= cpu_to_be32(0x80000000);
3098 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3099 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3101 case MLX4_IB_QPT_UD:
3102 set_datagram_seg(wqe, ud_wr(wr));
3103 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3104 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3106 if (wr->opcode == IB_WR_LSO) {
3107 err = build_lso_seg(wqe, ud_wr(wr), qp, &seglen,
3109 if (unlikely(err)) {
3113 lso_wqe = (__be32 *) wqe;
3115 size += seglen / 16;
3119 case MLX4_IB_QPT_PROXY_SMI_OWNER:
3120 err = build_sriov_qp0_header(to_msqp(qp), ud_wr(wr),
3122 if (unlikely(err)) {
3127 size += seglen / 16;
3128 /* to start tunnel header on a cache-line boundary */
3129 add_zero_len_inline(wqe);
3132 build_tunnel_header(ud_wr(wr), wqe, &seglen);
3134 size += seglen / 16;
3136 case MLX4_IB_QPT_PROXY_SMI:
3137 case MLX4_IB_QPT_PROXY_GSI:
3138 /* If we are tunneling special qps, this is a UD qp.
3139 * In this case we first add a UD segment targeting
3140 * the tunnel qp, and then add a header with address
3142 set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe,
3144 qp->mlx4_ib_qp_type);
3145 wqe += sizeof (struct mlx4_wqe_datagram_seg);
3146 size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
3147 build_tunnel_header(ud_wr(wr), wqe, &seglen);
3149 size += seglen / 16;
3152 case MLX4_IB_QPT_SMI:
3153 case MLX4_IB_QPT_GSI:
3154 err = build_mlx_header(to_msqp(qp), ud_wr(wr), ctrl,
3156 if (unlikely(err)) {
3161 size += seglen / 16;
3169 * Write data segments in reverse order, so as to
3170 * overwrite cacheline stamp last within each
3171 * cacheline. This avoids issues with WQE
3176 dseg += wr->num_sge - 1;
3177 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
3179 /* Add one more inline data segment for ICRC for MLX sends */
3180 if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
3181 qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
3182 qp->mlx4_ib_qp_type &
3183 (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
3184 set_mlx_icrc_seg(dseg + 1);
3185 size += sizeof (struct mlx4_wqe_data_seg) / 16;
3188 for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
3189 set_data_seg(dseg, wr->sg_list + i);
3192 * Possibly overwrite stamping in cacheline with LSO
3193 * segment only after making sure all data segments
3197 *lso_wqe = lso_hdr_sz;
3199 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
3200 MLX4_WQE_CTRL_FENCE : 0) | size;
3203 * Make sure descriptor is fully written before
3204 * setting ownership bit (because HW can start
3205 * executing as soon as we do).
3209 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
3215 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
3216 (ind & qp->sq.wqe_cnt ? cpu_to_be32(1U << 31) : 0) | blh;
3218 stamp = ind + qp->sq_spare_wqes;
3219 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
3222 * We can improve latency by not stamping the last
3223 * send queue WQE until after ringing the doorbell, so
3224 * only stamp here if there are still more WQEs to post.
3226 * Same optimization applies to padding with NOP wqe
3227 * in case of WQE shrinking (used to prevent wrap-around
3228 * in the middle of WR).
3231 stamp_send_wqe(qp, stamp, size * 16);
3232 ind = pad_wraparound(qp, ind);
3238 qp->sq.head += nreq;
3241 * Make sure that descriptors are written before
3246 writel(qp->doorbell_qpn,
3247 to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
3250 * Make sure doorbells don't leak out of SQ spinlock
3251 * and reach the HCA out of order.
3255 stamp_send_wqe(qp, stamp, size * 16);
3257 ind = pad_wraparound(qp, ind);
3258 qp->sq_next_wqe = ind;
3261 spin_unlock_irqrestore(&qp->sq.lock, flags);
3266 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
3267 struct ib_recv_wr **bad_wr)
3269 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3270 struct mlx4_wqe_data_seg *scat;
3271 unsigned long flags;
3277 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
3279 max_gs = qp->rq.max_gs;
3280 spin_lock_irqsave(&qp->rq.lock, flags);
3282 if (mdev->dev->persist->state & MLX4_DEVICE_STATE_INTERNAL_ERROR) {
3289 ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
3291 for (nreq = 0; wr; ++nreq, wr = wr->next) {
3292 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
3298 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
3304 scat = get_recv_wqe(qp, ind);
3306 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
3307 MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
3308 ib_dma_sync_single_for_device(ibqp->device,
3309 qp->sqp_proxy_rcv[ind].map,
3310 sizeof (struct mlx4_ib_proxy_sqp_hdr),
3313 cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
3314 /* use dma lkey from upper layer entry */
3315 scat->lkey = cpu_to_be32(wr->sg_list->lkey);
3316 scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
3321 for (i = 0; i < wr->num_sge; ++i)
3322 __set_data_seg(scat + i, wr->sg_list + i);
3325 scat[i].byte_count = 0;
3326 scat[i].lkey = cpu_to_be32(MLX4_INVALID_LKEY);
3330 qp->rq.wrid[ind] = wr->wr_id;
3332 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
3337 qp->rq.head += nreq;
3340 * Make sure that descriptors are written before
3345 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
3348 spin_unlock_irqrestore(&qp->rq.lock, flags);
3353 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
3355 switch (mlx4_state) {
3356 case MLX4_QP_STATE_RST: return IB_QPS_RESET;
3357 case MLX4_QP_STATE_INIT: return IB_QPS_INIT;
3358 case MLX4_QP_STATE_RTR: return IB_QPS_RTR;
3359 case MLX4_QP_STATE_RTS: return IB_QPS_RTS;
3360 case MLX4_QP_STATE_SQ_DRAINING:
3361 case MLX4_QP_STATE_SQD: return IB_QPS_SQD;
3362 case MLX4_QP_STATE_SQER: return IB_QPS_SQE;
3363 case MLX4_QP_STATE_ERR: return IB_QPS_ERR;
3368 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
3370 switch (mlx4_mig_state) {
3371 case MLX4_QP_PM_ARMED: return IB_MIG_ARMED;
3372 case MLX4_QP_PM_REARM: return IB_MIG_REARM;
3373 case MLX4_QP_PM_MIGRATED: return IB_MIG_MIGRATED;
3378 static int to_ib_qp_access_flags(int mlx4_flags)
3382 if (mlx4_flags & MLX4_QP_BIT_RRE)
3383 ib_flags |= IB_ACCESS_REMOTE_READ;
3384 if (mlx4_flags & MLX4_QP_BIT_RWE)
3385 ib_flags |= IB_ACCESS_REMOTE_WRITE;
3386 if (mlx4_flags & MLX4_QP_BIT_RAE)
3387 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
3392 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
3393 struct mlx4_qp_path *path)
3395 struct mlx4_dev *dev = ibdev->dev;
3398 memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
3399 ib_ah_attr->port_num = path->sched_queue & 0x40 ? 2 : 1;
3401 if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
3404 is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
3405 IB_LINK_LAYER_ETHERNET;
3407 ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
3408 ((path->sched_queue & 4) << 1);
3410 ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
3412 ib_ah_attr->dlid = be16_to_cpu(path->rlid);
3413 ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
3414 ib_ah_attr->static_rate = path->static_rate ? path->static_rate - 5 : 0;
3415 ib_ah_attr->ah_flags = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
3416 if (ib_ah_attr->ah_flags) {
3417 ib_ah_attr->grh.sgid_index = path->mgid_index;
3418 ib_ah_attr->grh.hop_limit = path->hop_limit;
3419 ib_ah_attr->grh.traffic_class =
3420 (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
3421 ib_ah_attr->grh.flow_label =
3422 be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
3423 memcpy(ib_ah_attr->grh.dgid.raw,
3424 path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
3428 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
3429 struct ib_qp_init_attr *qp_init_attr)
3431 struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
3432 struct mlx4_ib_qp *qp = to_mqp(ibqp);
3433 struct mlx4_qp_context context;
3437 mutex_lock(&qp->mutex);
3439 if (qp->state == IB_QPS_RESET) {
3440 qp_attr->qp_state = IB_QPS_RESET;
3444 err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3450 mlx4_state = be32_to_cpu(context.flags) >> 28;
3452 qp->state = to_ib_qp_state(mlx4_state);
3453 qp_attr->qp_state = qp->state;
3454 qp_attr->path_mtu = context.mtu_msgmax >> 5;
3455 qp_attr->path_mig_state =
3456 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3457 qp_attr->qkey = be32_to_cpu(context.qkey);
3458 qp_attr->rq_psn = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3459 qp_attr->sq_psn = be32_to_cpu(context.next_send_psn) & 0xffffff;
3460 qp_attr->dest_qp_num = be32_to_cpu(context.remote_qpn) & 0xffffff;
3461 qp_attr->qp_access_flags =
3462 to_ib_qp_access_flags(be32_to_cpu(context.params2));
3464 if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3465 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3466 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3467 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3468 qp_attr->alt_port_num = qp_attr->alt_ah_attr.port_num;
3471 qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3472 if (qp_attr->qp_state == IB_QPS_INIT)
3473 qp_attr->port_num = qp->port;
3475 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3477 /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3478 qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3480 qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3482 qp_attr->max_dest_rd_atomic =
3483 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3484 qp_attr->min_rnr_timer =
3485 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3486 qp_attr->timeout = context.pri_path.ackto >> 3;
3487 qp_attr->retry_cnt = (be32_to_cpu(context.params1) >> 16) & 0x7;
3488 qp_attr->rnr_retry = (be32_to_cpu(context.params1) >> 13) & 0x7;
3489 qp_attr->alt_timeout = context.alt_path.ackto >> 3;
3492 qp_attr->cur_qp_state = qp_attr->qp_state;
3493 qp_attr->cap.max_recv_wr = qp->rq.wqe_cnt;
3494 qp_attr->cap.max_recv_sge = qp->rq.max_gs;
3496 if (!ibqp->uobject) {
3497 qp_attr->cap.max_send_wr = qp->sq.wqe_cnt;
3498 qp_attr->cap.max_send_sge = qp->sq.max_gs;
3500 qp_attr->cap.max_send_wr = 0;
3501 qp_attr->cap.max_send_sge = 0;
3505 * We don't support inline sends for kernel QPs (yet), and we
3506 * don't know what userspace's value should be.
3508 qp_attr->cap.max_inline_data = 0;
3510 qp_init_attr->cap = qp_attr->cap;
3512 qp_init_attr->create_flags = 0;
3513 if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3514 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3516 if (qp->flags & MLX4_IB_QP_LSO)
3517 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3519 if (qp->flags & MLX4_IB_QP_NETIF)
3520 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3522 qp_init_attr->sq_sig_type =
3523 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3524 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3527 mutex_unlock(&qp->mutex);